Browse the corpus

Introduction Dozens of trials have proven the effectiveness of collaborative care strategies at treating mood and anxiety disorders in primary care.1 These programs typically involve nonphysician care managers who promote use of evidence-based treatment protocols and monitor patients’ clinical response under the supervision of their primary care physicians (PCPs). However, challenges hinder provision of collaborative care into routine practice and at scale.2 Enabled by advances in computer technology, several computerized cognitive behavioral therapy (CCBT) programs have been developed and proven to be as effective as face-to-face therapy at treating depression and anxiety in primary care.3,4,5 These programs have the advantages of convenient 24/7 access, avoidance of stigma incurred by seeing a therapist, and greater consistency and scalability compared with traditional therapy. Still, while CCBT has been used by hundreds of thousands of patients in Europe and Australia, it remains largely unknown and little used within the United States.6 Another recent development has been the rise of internet support groups (ISGs) that offer general health and disease-specific information and enable members to share treatment information and provide peer support.7 Indeed, some ISGs have evolved into large-scale sites with thousands of members organized into numerous disease-specific groups.8,9 Yet despite indications of benefit,9,10,11,12,13 to our knowledge, their effectiveness has not been firmly established.14

rences and Effect Sizes for the 12-Item Short-Form Health Survey Mental Health Composite Scale CCBT indicates computerized cognitive behavioral therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; PCP, primary care physician; PHQ-9, 9-Item Patient Health Questionnaire; UC, usual care. Secondary Hypothesis: CCBT Alone vs UC Compared with patients in the UC arm, patients in the CCBT alone arm reported significant 6-month improvements on the PROMIS Depression and Anxiety scales (eFigure 4 in Supplement 2) but not the SF-12 MCS scale (Figure 2). However, these differences resolved 6 months later, as patients’ symptoms in the UC arm improved. Again, we observed significant treatment interactions favoring CCBT for patients aged 35 to 59 years on the SF-12 MCS (eFigure 3 in Supplement 2) and PROMIS Depression and Anxiety scales (eFigure 4 in Supplement 2), for patients living alone on the PROMIS Depression and Anxiety scales, and for nonwhite patients on the PROMIS Depression scale (eFigure 4 in Supplement 2). Moreover, patients reported improved 6-month SF-12 MCS (mean points, 0.80; 95% CI, 0.37-1.22), PROMIS Depression (mean points, 0.48; 95% CI, −0.76 to −0.19), and PROMIS Anxiety (mean points, 0.48; 95% CI, −0.79 to −0.17) scores for each additional CCBT session completed, and per-protocol analyses revealed a similar pattern (PROMIS mood symptoms: patients who completed ≥4 sessions: ES, 0.41; 95% CI, 0.17-0.65; patients who completed all 8 sessions: ES, 0.52; 95% CI, 0.26-0.78) (eTable 2 in Supplement 2).

members to share treatment information and provide peer support.7 Indeed, some ISGs have evolved into large-scale sites with thousands of members organized into numerous disease-specific groups.8,9 Yet despite indications of benefit,9,10,11,12,13 to our knowledge, their effectiveness has not been firmly established.14 Providing patients with depression and anxiety with guided access to CCBT either alone or in combination with an ISG may be an ideal method to deliver effective mental health care at scale. This report presents the main findings from the Online Treatments for Mood and Anxiety Disorders in Primary Care, the first randomized trial to evaluate the effectiveness of providing these technologies through a collaborative care program. Methods Study Setting Using a protocol approved by the University of Pittsburgh Institutional Review Board, our single-center trial recruited patients from 26 primary care offices that shared a common electronic medical record (EMR) (Epic). Informed written consent was obtained from all participants. The trial protocol can be found in Supplement 1. Participants We exposed PCPs to an EMR “Best Practice Alert” reminder about our study at the time of the clinical encounter.15 It launched automatically for all patients aged 18 to 75 years whenever anxiety, generalized anxiety, panic, or depression was entered as an encounter diagnosis. If the patient agreed to a referral, the PCP electronically “signed” the alert, which forwarded the patient’s name to a study recruiter who then called the patient by telephone to review protocol eligibility.

to 75 years whenever anxiety, generalized anxiety, panic, or depression was entered as an encounter diagnosis. If the patient agreed to a referral, the PCP electronically “signed” the alert, which forwarded the patient’s name to a study recruiter who then called the patient by telephone to review protocol eligibility. Eligible patients needed to have internet and email access; a score of 10 or greater on either the 7-Item Generalized Anxiety Disorder scale (GAD-7)16 or the 9-Item Patient Health Questionnaire (PHQ-9)17; and no alcohol dependence as determined by the Alcohol Use Disorders Identification Test,18 active suicidality, or other serious mental illness for which our interventions may be inappropriate. If confirmed, the recruiter reviewed a mailed consent form and obtained the patients’ signed consent on a recorded telephone line. Afterwards, they administered the 12-Item Short-Form Health Survey (SF-12) to determine health-related quality of life,19 the fixed-length Patient-Reported Outcomes Measurement Information System (PROMIS) depression and anxiety measures to assess mood and anxiety symptoms,20 and the Primary Care Evaluation of Mental Disorders to provide an anxiety and mood disorder diagnosis,21 and collected information on patients’ self-reported race/ethnicity, sex, and other sociodemographic characteristics.

nformation System (PROMIS) depression and anxiety measures to assess mood and anxiety symptoms,20 and the Primary Care Evaluation of Mental Disorders to provide an anxiety and mood disorder diagnosis,21 and collected information on patients’ self-reported race/ethnicity, sex, and other sociodemographic characteristics. Randomization Procedure Following the baseline assessment, we randomized patients in a 3:3:1 ratio to (1) care manager–guided CCBT (CCBT alone), (2) care manager–guided access to both CCBT and our ISG (CCBT+ISG), or (3) usual care (UC) under their PCP. We stratified randomization by practice size and age group using randomly permuted blocks according to a computer-generated assignment sequence prepared in advance by our study statistician and concealed until after the baseline assessment. Afterwards, we informed all patients of their treatment assignment and notified their referring PCP. Usual Care For ethical reasons,22 we informed patients receiving UC of their mood and anxiety symptoms and their referring PCP. However, we provided no treatment advice unless we detected suicidality or a 25% worsening of symptoms from baseline on a follow-up assessment. Interventions We employed college graduates with mental health research experience as care managers and assigned each exclusively to one intervention arm. We first prepared them in a basic understanding of mood and anxiety disorders, our pharmacotherapy algorithm,23 CCBT program, and tracking registry and later reinforced this training in our weekly case review sessions.

al health research experience as care managers and assigned each exclusively to one intervention arm. We first prepared them in a basic understanding of mood and anxiety disorders, our pharmacotherapy algorithm,23 CCBT program, and tracking registry and later reinforced this training in our weekly case review sessions. Computerized Cognitive Behavioral Therapy We used the Beating the Blues CCBT program, which has been proven to be effective.24,25 It consists of a 10-minute introductory video followed by eight 50-minute interactive sessions that our care managers encouraged patients to complete every 1 to 2 weeks. Each session used easily understood text, audiovisual clips, and “homework” assignments to impart basic CBT techniques, and patients completed the GAD-7 and PHQ-9 at the start of each CCBT session to self-track their symptoms (eMethods in Supplement 2).

care managers encouraged patients to complete every 1 to 2 weeks. Each session used easily understood text, audiovisual clips, and “homework” assignments to impart basic CBT techniques, and patients completed the GAD-7 and PHQ-9 at the start of each CCBT session to self-track their symptoms (eMethods in Supplement 2). Internet Support Group We used WordPress software to create our password-protected ISG that patients could access via computer or smartphone (eFigure 1 in Supplement 2). In addition to a variety of discussion boards created by the care manager moderator and study patients, the ISG curated links to external resources, including local $4 generic pharmacy programs, “find-a-therapist” and various crisis hotlines, and brief YouTube videos on insomnia, nutrition, exercise, and other topics, and we embedded links to our EMR’s patient portal to integrate its use into routine care. To enhance patient engagement, we featured (1) status indicators on members’ profiles and comments (eg, stars and “likes”), (2) email notifications of new ISG activities, (3) automated highlighting of recent comments on members’ home pages personalized to their ISG profile and past activities, (4) invited member-guest moderators, and (5) various contests to encourage log-ins and comments.

bers’ profiles and comments (eg, stars and “likes”), (2) email notifications of new ISG activities, (3) automated highlighting of recent comments on members’ home pages personalized to their ISG profile and past activities, (4) invited member-guest moderators, and (5) various contests to encourage log-ins and comments. To preserve confidentiality, we assigned members’ user names, encouraged them to select a representative avatar (eg, a sunrise or animal), and sent reminders not to post self-identifying information. For additional safety, an investigator logged into the ISG daily to review new posts for suicidal thoughts and other potentially inappropriate content, and we allowed members to flag comments for potential removal. Care Manager Contacts Care managers emailed their assigned patients a web link to the CCBT program and, if applicable, the ISG and requested a time to schedule an introductory telephone call to review the program(s) and establish rapport. Later, they logged into the CCBT program’s clinical helper portal to monitor their patients’ progress (eg, sessions completed, self-reported symptoms, and problems they chose to address), sent personalized feedback and encouragement via email, and contacted patients via telephone who either had not improved or failed to log in regularly.

into the CCBT program’s clinical helper portal to monitor their patients’ progress (eg, sessions completed, self-reported symptoms, and problems they chose to address), sent personalized feedback and encouragement via email, and contacted patients via telephone who either had not improved or failed to log in regularly. Case Review and Follow-up Care managers presented their patients to the study PCP, psychiatrist, and project coordinator in weekly 60-minute case review sessions split by intervention arm. To efficiently focus our time, we developed an electronic registry that could sort patients by randomization date, last contact, and highest PHQ-9 or GAD-7 score (eMethods in Supplement 2). In addition to conveying general lifestyle adjustments, including exercise and social engagement, we recommended antidepressant/anxiolytic pharmacotherapy based on patients’ treatment preferences and response to CCBT as well as referrals to mental health specialists when they did not improve or had complex psychosocial issues.23 Depending on a patient’s symptoms and level of engagement, the care manager emailed or telephoned biweekly for approximately 2 months, and these contacts lasted approximately 15 to 30 minutes. Afterwards, the patient transitioned to the continuation phase of care, during which the care manager contacted the patient approximately monthly until the end of our 6-month intervention. Given our collaborative care framework, we provided PCPs with our treatment recommendations and regular updates of their patients’ progress via EMR.

atient transitioned to the continuation phase of care, during which the care manager contacted the patient approximately monthly until the end of our 6-month intervention. Given our collaborative care framework, we provided PCPs with our treatment recommendations and regular updates of their patients’ progress via EMR. Assessments Patients, PCPs, and care managers were not blinded to their treatment assignment. Therefore, we employed several blinded assessors to determine the effectiveness of our interventions. They contacted patients by telephone to administer our assessment battery at 3-month, 6-month, and 12-month follow-up and later sent patients $15 after each completed assessment for their time (up to $60). We trained our assessors using audiotapes, manuals, and practice interviews and used a computer-assisted telephone interview system to guide them through each assessment. We digitally recorded these calls and conducted periodic spot checks to confirm responses were rated accurately and corresponded with those entered into our study database, reviewed interactions with suicidal patients, and provided staff with feedback on their performance. Later, we abstracted data from the EMR to collect information on patients’ medical conditions and health services use, our server logs to measure engagement with the CCBT and ISG programs, and our care managers’ electronic registry to document the number of email and telephone contacts.

h feedback on their performance. Later, we abstracted data from the EMR to collect information on patients’ medical conditions and health services use, our server logs to measure engagement with the CCBT and ISG programs, and our care managers’ electronic registry to document the number of email and telephone contacts. Data and Safety Monitoring We programmed our computer-assisted telephone interview system to identify patients receiving UC whose blinded PROMIS score increased by 25% or more above baseline. Following a review, we notified their PCP via EMR and offered treatment advice. Whenever our care managers or assessors encountered suicidality, either expressed spontaneously or on routine administration of our measures, our computer-assisted telephone interview system automatically launched our Suicide Risk Management Protocol that provided triage advice.26 The CCBT program also notified the care manager whenever a patient endorsed suicidality on the PHQ-9 administered at each session. Finally, an independent external data and safety monitoring board appointed by our funding agency monitored the progress and safety of our trial.

ent Protocol that provided triage advice.26 The CCBT program also notified the care manager whenever a patient endorsed suicidality on the PHQ-9 administered at each session. Finally, an independent external data and safety monitoring board appointed by our funding agency monitored the progress and safety of our trial. Statistical Analysis We powered our trial to test the primary hypothesis that patients receiving CCBT+ISG will report 0.30 or greater effect size (ES) improvement from baseline at 6 months on the SF-12 Mental Health Composite Scale (MCS) vs CCBT alone. Assuming a 2-sample t test to compare between-arm differences in 6-month improvements and 2-tailed α = .05, we needed 300 patients per arm to have 90% or greater power to detect a 0.30 ES difference in our primary outcome measure.

from baseline at 6 months on the SF-12 Mental Health Composite Scale (MCS) vs CCBT alone. Assuming a 2-sample t test to compare between-arm differences in 6-month improvements and 2-tailed α = .05, we needed 300 patients per arm to have 90% or greater power to detect a 0.30 ES difference in our primary outcome measure. We compared baseline sociodemographic and clinical characteristics by randomization status using t tests for continuous data and χ2 analyses for categorical data. Our primary intent-to-treat analyses included all randomized participants regardless of adherence to their assigned treatment. We used linear mixed models27 that included fixed effects for time, study arm, time-by-study arm, age strata, practice size, and random effects for patients. We considered time as a categorical variable because of the assumption of nonlinearity over time. To test our hypotheses, we used contrasts to estimate the adjusted mean difference between study arms in the 6-month improvement of SF-12 MCS and PROMIS measures (secondary outcomes) and 6 months later to assess treatment durability. Additionally, we calculated ESs for 6-month changes in SF-12 with 95% CIs by (1) prespecified subgroups of age group, sex, race/ethnicity, baseline symptom severity, and practice size and (2) unplanned subgroups of education and living alone status. We considered a significant 3-way interaction between time, study arm, and the potential covariate as a significant subgroup effect and used Poisson regression to compare rates of PCP contacts, emergency department visits, and hospitalizations between study arms.

nplanned subgroups of education and living alone status. We considered a significant 3-way interaction between time, study arm, and the potential covariate as a significant subgroup effect and used Poisson regression to compare rates of PCP contacts, emergency department visits, and hospitalizations between study arms. We investigated a potential dose response between the number of CCBT sessions completed within our CCBT alone and UC study arms using the same linear mixed model described earlier but parametrizing the CCBT alone arm by assigning each patient a value equal to the proportion completed of the 8-session program. Finally, we conducted exploratory post hoc per-protocol analyses restricted to those who completed 4 or more and all 8 CCBT sessions. Every effort was made to identify the mechanism of missing data. We compared participants who withdrew from study participation by baseline covariates and analyzed time until withdrawal by study arm using Kaplan-Meier curves.28 Our linear mixed models for our primary analyses assumed that data were missing at random and were robust to ignorable missingness assumptions.29 All reported P values are 2-tailed with significance levels at P ≤ .05, and all analyses were performed with SAS version 9.4 (SAS Institute).

arm using Kaplan-Meier curves.28 Our linear mixed models for our primary analyses assumed that data were missing at random and were robust to ignorable missingness assumptions.29 All reported P values are 2-tailed with significance levels at P ≤ .05, and all analyses were performed with SAS version 9.4 (SAS Institute). Results From August 2012 to September 2014, PCPs referred 2884 patients in response to our EMR prompt. Of these, 704 (24.4%) met all eligibility criteria, provided informed consent, and were randomized (Figure 1). Their baseline sociodemographic and clinical characteristics (Table 1) and completion rate of follow-up assessments at both 6 months (604 [85.8%]) and 12 months (593 [84.2%]) were similar by randomization status (Figure 1), and we found no differences in the sociodemographic and clinical characteristics between participants who withdrew and those who did not. Figure 1. Flowchart of Participants Participants were referred by primary care physicians (PCPs) between August 2012 and September 2014. CCBT indicates computerized cognitive behavioral therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; MH, mental health; MHS, mental health specialist; PHQ-9, 9-Item Patient Health Questionnaire. Table 1. Baseline Sociodemographic and Clinical Characteristics by Randomization Status Characteristic No.

Figure 1. Flowchart of Participants Participants were referred by primary care physicians (PCPs) between August 2012 and September 2014. CCBT indicates computerized cognitive behavioral therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; MH, mental health; MHS, mental health specialist; PHQ-9, 9-Item Patient Health Questionnaire. Table 1. Baseline Sociodemographic and Clinical Characteristics by Randomization Status Characteristic No. (%) Overall (N = 704) CCBT alone (n = 301) CCBT+ISG (n = 302) Usual Care (n = 101) Age, mean (SD) 42.7 (14.3) 43.0 (14.0) 42.6 (14.4) 41.7 (14.6) Age group, y 18-34 256 (36.4) 108 (35.9) 111 (36.8) 37 (36.6) 35-59 343 (48.7) 149 (49.5) 143 (47.4) 51 (50.5) 60-75 105 (14.9) 44 (14.6) 48 (15.9) 13 (12.9) Female 562 (79.8) 235 (78.1) 245 (81.1) 82 (81.2) Race/ethnicity White 576 (81.8) 257 (85.4) 242 (80.1) 77 (76.2) African American 113 (16.5) 38 (12.6) 53 (17.5) 22 (21.8) Other 15 (2.1) 6 (2.0) 7 (2.3) 2 (2.0) College degree or higher 333 (47.3) 137 (45.5) 144 (47.7) 52 (51.5) Married or living with partner 283 (40.2) 123 (40.9) 120 (39.7) 40 (39.6) Living alone 125 (17.8) 54 (17.9) 60 (19.9) 11 (10.9) Employed 492 (69.9) 217 (72.1) 204 (67.5) 71 (70.3) Practice size Large (≥6 PCPs) 433 (61.5) 185 (61.5) 186 (61.6) 62 (61.4) Small (<6 PCPs) 271 (38.5) 116 (38.5) 116 (38.4) 39 (38.6) Mental health disordera Major depression 597 (84.8) 258 (85.7) 257 (85.1) 82 (81.2) Generalized anxiety disorder 313 (44.5) 135 (44.9) 124 (41.1) 54 (53.5) Panic disorder 160 (22.7) 65 (21.6) 79 (26.2) 16 (15.8) Both depression and anxiety 499 (70.9) 219 (72.8) 207 (68.5) 73 (72.3) Depression/anxiety medication use within past year 544 (77.3) 232 (77.1) 236 (78.1) 76 (75.2) PHQ-9 score, mean (SD)b,c 13.3 (5) 13.2 (5.3) 13.4 (4.7) 13.1 (4.9) PHQ-9 score ≥ 15 281 (39.9) 119 (39.5) 122 (40.4) 40 (39.6) GAD-7 score, mean (SD)b,d 12.9 (4.4) 13.0 (4.3) 12.6 (4.5) 13.5 (4.2) GAD-7 score ≥ 15 257 (36.5) 114 (37.9) 102 (33.8) 41 (40.6) PROMIS Depression T-score, mean (SD)e 62.1 (6.3) 62.5 (6.2) 62.0 (6.3) 61.4 (6.4) PROMIS Anxiety T-score, mean (SD)f 65.8 (6) 65.9 (6) 65.8 (6.2) 65.4 (5.7) SF-12 MCS, mean (SD)g,h 31.4 (9) 31.3 (8.4) 31.7 (9.4) 31.1 (9.3) SF-12 PCS, mean (SD)g,h 51.1 (12.3) 50.7 (12.2) 51.0 (12.3) 52.2 (12.7) Abbreviations: CCBT, computerized cognitive behavioral therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; PCP, primary care physician; PHQ-9, 9-Item Patient Health Questionnaire; PROMIS, Patient-Reported Outcomes Measurement Information System; 12-Item Short-Form Health Survey Mental Health Composite Scale; SF-12 PCS, 12-Item Short-Form

therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; PCP, primary care physician; PHQ-9, 9-Item Patient Health Questionnaire; PROMIS, Patient-Reported Outcomes Measurement Information System; 12-Item Short-Form Health Survey Mental Health Composite Scale; SF-12 PCS, 12-Item Short-Form Health Survey Physical Health Composite Scale. a Determined using Primary Care Evaluation of Mental Disorders. b Higher scores indicate more severe symptoms. c Range, 0-27. d Range, 0-21. e T-score range, 37.1-81.1. f T-score range, 36.3-82.7. g Range, 0-100. h Higher scores indicate better health-related quality of life. Intervention Engagement By 6 months, 504 of 603 patients (83.6%) with CCBT access had completed at least 1 session and 221 (36.7%) had completed all 8, and the mean sessions completed was 5.4, which was similar by randomization status (Table 2), sex, race/ethnicity, and age strata (eTable 1 in Supplement 2). Overall, 228 of 302 patients (75.5%) in the CCBT+ISG arm logged into the ISG at least once, of whom 141 (61.8%) made at least 1 online comment or post (mean, 10.5; median, 3; range, 1-306) (Table 2) (eFigure 2 in Supplement 2).

domization status (Table 2), sex, race/ethnicity, and age strata (eTable 1 in Supplement 2). Overall, 228 of 302 patients (75.5%) in the CCBT+ISG arm logged into the ISG at least once, of whom 141 (61.8%) made at least 1 online comment or post (mean, 10.5; median, 3; range, 1-306) (Table 2) (eFigure 2 in Supplement 2). Table 2. 6-Month Care Processes and Health Services Use Following Randomization Characteristic CCBT Alone (n = 301) CCBT+ISG (n = 302) Usual Care (n = 101) Beating the Blues CCBT, No. (%) Participants who logged in 261 (86.7) 260 (86.1) NA CCBT sessions completed of those who completed ≥1 session, mean (SD) [denominator] 5.4 (2.8) [254] 5.5 (2.7) [250] NA No. of participants who completed all 8 sessions 112 (37.2) 109 (36.1) NA ISG Logged in, No. (%) NA 228 (75.5) NA Log-ins per user Mean NA 8.9 NA Median (range) NA 4 (1-214) NA Commented, No. (%) NA 138 (45.7) NA Comments per commenter Mean NA 9.6 NA Median (range) NA 3 (1-285) NA Posted, No. (%) NA 45 (14.9) NA Posts per poster Mean NA 3.8 NA Median (range) NA 1 (1-42) NA Commented or posted, No. (%) NA 141 (46.7) NA Comments/posts per commenter/poster Mean NA 10.5 NA Median (range) NA 3 (1-306) NA Care management, median (IQR)a No. of telephone calls 4 (3-6) 3 (2-5) NA No. of emails 9 (6-11) 12 (9-16) NA No. of total contacts 13 (10-16) 16 (12-20) NA Pharmacotherapy, No. (%)a SSRI/SNRI use at baseline 200 (66.4) 206 (68.2) 66 (65.3) SSRI/SNRI use at 6 mo, No./total No. (%) 164/253 (64.8) 166/259 (64.1) 50/92 (54) Benzodiazepine use at baseline 39 (13.0) 40 (13.2) 14 (13.9) Benzodiazepine use at 6 mo, No./total No. (%) 31/253 (12.3) 29/259 (11.2) 9/92 (10) Health care use, median (range)a PCP office visits 2 (0-12) 2 (0-16) 2 (0-7) PCP telephone contacts 0 (0-7) 0 (0-7) 0 (0-4) PCP email contacts 0 (0-7) 0 (0-11) 0 (0-6) PCP total contacts 3 (0-18) 4 (0-28) 3 (0-11) Mental health specialty visit, No./total No. (%) 45/267 (16.9) 69/271 (25.5) 17/95 (18) ED visits 0 (0-5) 0 (0-7) 0 (0-4) Hospitalizations 0 (0-4) 0 (0-2) 0 (0-3) Abbreviations: CCBT, computerized cognitive behavioral therapy; ED, emergency department; IQR, interquartile range; ISG, internet support group; NA, not applicable; PCP, primary care physician; SSRI, selective serotonin reuptake inhibitor; SNRI, serotonin norepinephrine reuptake inhibitor.

tions 0 (0-4) 0 (0-2) 0 (0-3) Abbreviations: CCBT, computerized cognitive behavioral therapy; ED, emergency department; IQR, interquartile range; ISG, internet support group; NA, not applicable; PCP, primary care physician; SSRI, selective serotonin reuptake inhibitor; SNRI, serotonin norepinephrine reuptake inhibitor. a Data from medical record abstraction. Primary Hypothesis: CCBT+ISG vs CCBT Alone At 6-month follow-up, patients in the CCBT+ISG and CCBT alone arms reported similar improvements on our primary outcome measure (SF-12 MCS: ES, 0.02; 95% CI, −0.17 to 0.13) and on the PROMIS Depression and Anxiety scales that continued 6 months later (Figure 2). We also identified a significant treatment interaction favoring CCBT+ISG for patients aged 60 to 75 years on the SF-12 MCS (Figure 3) (eFigure 3 in Supplement 2) and CCBT alone for patients aged 35 to 59 years on the PROMIS Depression and Anxiety scales (eFigure 4 in Supplement 2).

continued 6 months later (Figure 2). We also identified a significant treatment interaction favoring CCBT+ISG for patients aged 60 to 75 years on the SF-12 MCS (Figure 3) (eFigure 3 in Supplement 2) and CCBT alone for patients aged 35 to 59 years on the PROMIS Depression and Anxiety scales (eFigure 4 in Supplement 2). Figure 2. Estimated Scores by Baseline Treatment Assignment Linear mixed models adjusted for time, study arm, time-by-study arm, age strata, and clinic size. A, Estimated scores for the 12-Item Short-Form Health Survey Mental Health Composite Scale (SF-12 MCS). B, Estimated scores for the Patient-Reported Outcomes Measurement Information System (PROMIS) Depression scale. At 6 months, patients receiving computerized cognitive behavioral therapy (CCBT) alone vs usual care reported a −2.43 (95% CI, −4.16 to −0.69; P = .006) improvement. C, Estimated scores for the PROMIS Anxiety scale. At 6 months, patients receiving CCBT alone vs usual care reported a −2.30 (95% CI, −4.21 to −0.4; P = .02) improvement. The vertical line at 6 months indicates the end of care manager–led CCBT and our primary outcome point. The following 6 months were naturalistic follow-up to observe the durability of our interventions. The error bars indicate 95% CIs. ISG indicates internet support group.

(95% CI, −4.21 to −0.4; P = .02) improvement. The vertical line at 6 months indicates the end of care manager–led CCBT and our primary outcome point. The following 6 months were naturalistic follow-up to observe the durability of our interventions. The error bars indicate 95% CIs. ISG indicates internet support group. Figure 3. Forest Plot of Between-Group Differences and Effect Sizes for the 12-Item Short-Form Health Survey Mental Health Composite Scale CCBT indicates computerized cognitive behavioral therapy; GAD-7, 7-Item Generalized Anxiety Disorder Scale; ISG, internet support group; PCP, primary care physician; PHQ-9, 9-Item Patient Health Questionnaire; UC, usual care.

s, 0.48; 95% CI, −0.79 to −0.17) scores for each additional CCBT session completed, and per-protocol analyses revealed a similar pattern (PROMIS mood symptoms: patients who completed ≥4 sessions: ES, 0.41; 95% CI, 0.17-0.65; patients who completed all 8 sessions: ES, 0.52; 95% CI, 0.26-0.78) (eTable 2 in Supplement 2). 6-Month Health Services Use Primary care physicians and care manager contacts with patients via telephone and email were similar by intervention arm (Table 2). Moreover, patients receiving UC and intervention had similar rates of PCP contacts, use of antidepressant and anxiolytic pharmacotherapy, and visits to mental health specialists, emergency departments, and hospitals (Table 2). Discussion To our knowledge, this is the first trial to examine the effectiveness of incorporating either CCBT or an ISG into a collaborative care program for treating depression or anxiety in primary care. Our report confirms the effectiveness of guided CCBT, highlights the critical importance of patient engagement with online interventions, and provides high-quality evidence about the limits and potential benefits of these emerging technologies.

laborative care program for treating depression or anxiety in primary care. Our report confirms the effectiveness of guided CCBT, highlights the critical importance of patient engagement with online interventions, and provides high-quality evidence about the limits and potential benefits of these emerging technologies. Few trials have evaluated the psychologic benefits of ISGs, and none were linked to patients’ usual source of primary care as ours was.7,30,31,32 Perhaps most comparable with our trial is the trial by Griffiths et al,31 who randomized 478 adults with elevated depressive symptoms to either a moderated ISG, an online psychotherapy program, both interventions, or to a UC control. They too found their combined ISG and online psychotherapy interventions improved mood symptoms vs UC and no benefit from their ISG beyond UC. However, engagement was low, as only 62% of patients in the ISG group logged into the site, only 15% created 1 or more ISG posts, and those assigned to the ISG arm were more likely than patients in the UC arm to miss a blinded telephone assessment (52% vs 34% at 12 months).31 These findings as well as other reports33 add to our understanding of the challenges in sustaining patient engagement with online interventions to improve clinical outcomes.34,35

hose assigned to the ISG arm were more likely than patients in the UC arm to miss a blinded telephone assessment (52% vs 34% at 12 months).31 These findings as well as other reports33 add to our understanding of the challenges in sustaining patient engagement with online interventions to improve clinical outcomes.34,35 Unlike our earlier collaborative care trials, where care managers assigned patients homework lessons in printed workbooks,36,37,38 the CCBT program enabled our care managers to unobtrusively monitor their patients’ engagement with treatment while providing similarly effective care to a doubled caseload (90 to 100 patients).36,38 Although the ES improvements we obtained were smaller than those described in meta-analyses of “supported” CCBT (major depression: ES, 0.78; 95% CI, 0.59-0.964), we identified a dose effect that confirms the importance of patient engagement.39 Indeed, Gilbody et al40 reported in 2015 no differences in mood symptoms among 691 patients with depression in primary care they randomized to either Beating the Blues or MoodGYM (HealthMed) CCBT programs41 or usual care from their PCPs. Similar to our protocol, their study staff contacted patients by telephone to promote use of their programs; however, they did not monitor patients’ symptoms or send recommendations to PCPs as we did, and patient adherence with both CCBT programs was low (median sessions completed, <2).40

or usual care from their PCPs. Similar to our protocol, their study staff contacted patients by telephone to promote use of their programs; however, they did not monitor patients’ symptoms or send recommendations to PCPs as we did, and patient adherence with both CCBT programs was low (median sessions completed, <2).40 Given our findings and other recent reports,3,4,5 we anticipate more engaging and powerful CCBT programs better tailored to patients’ specific needs, sociodemographic characteristics, medical conditions, and cultural and linguistic preferences that are integrated into the EMR for documentation and billing purposes will become widely deployed over the next decade. Finally, while we were unable to demonstrate a measurable benefit from our ISG, we remain optimistic that more engaging ISGs that apply machine learning algorithms to EMR and claims data to present patients with more personalized information in real time will soon be tested by health care organizations experimenting with social media.42,43

e unable to demonstrate a measurable benefit from our ISG, we remain optimistic that more engaging ISGs that apply machine learning algorithms to EMR and claims data to present patients with more personalized information in real time will soon be tested by health care organizations experimenting with social media.42,43 Limitations Our study has limitations, several of which potentially affect the generalizability of our findings. First, our use of EMR-generated prompts to promote identification of patients for study participation is limited to settings with systems capable of generating these alerts, clinician recognition of targeted conditions, and entry of the proper diagnostic codes into the EMR. Second, we relied on 1 CCBT and ISG, and others using different programs and levels of human support to promote adherence may obtain different outcomes. Third, because we lacked information on symptom duration, patients with long-term mild mood and anxiety symptoms may have similar outcomes as those with severe acute symptoms. Fourth, given the nature of our interventions, patients knew their treatment assignment, which could have biased their responses to our blinded assessors. Finally, study sites were not cluster randomized, and the same physicians cared for patients in all study arms. Although this could have diminished outcome differences between treatment arms, we observed similar ES improvements as previous collaborative care trials.1

ave biased their responses to our blinded assessors. Finally, study sites were not cluster randomized, and the same physicians cared for patients in all study arms. Although this could have diminished outcome differences between treatment arms, we observed similar ES improvements as previous collaborative care trials.1 Conclusions In summary, although our ISG did not produce any measurable benefit over CCBT alone, providing online CCBT to patients with depression and anxiety receiving primary care via a centralized collaborative care program is an effective strategy for delivering mental health care at scale. Our study findings have important implications for transforming the way mental health care is delivered in primary care and focus further attention to the emerging field of e–mental health. Supplement 1. Trial protocol. Click here for additional data file. Supplement 2. eMethods. Statistical analysis plan (from funded grant application). eTable 1. Computerized cognitive behavior therapy sessions completed at 3 and 6 months following randomization. eTable 2. Effect size improvements by number of computerized cognitive behavior therapy (CCBT) sessions completed (CCBT alone vs usual care). eFigure 1. Screenshots of internet support group pages. eFigure 2. Boxplots of the number of logins, posts, comments, and posts or comments on internet support group. eFigure 3. Estimated scores by baseline treatment assignment for the SF-12 MCS by age. eFigure 4. Forest plots of between-group differences and effect sizes on the PROMIS Depression (top) and PROMIS Anxiety (bottom) scales.

eFigure 2. Boxplots of the number of logins, posts, comments, and posts or comments on internet support group. eFigure 3. Estimated scores by baseline treatment assignment for the SF-12 MCS by age. eFigure 4. Forest plots of between-group differences and effect sizes on the PROMIS Depression (top) and PROMIS Anxiety (bottom) scales. Click here for additional data file.

Introduction Postnatal depression (PND), also known as postpartum depression, is common, affecting approximately 10% of women in high-income countries,1 with higher estimates in low- and middle-income countries.2 Extensive research documents the association of PND with child development, including delayed cognitive and language development, higher rates of behavioral problems, insecure or disorganized attachment, lower school-leaving grades (ie, General Certificate of Secondary Education [GCSE] examinations in the United Kingdom) at 16 years, and higher rates of depression at 16 to 18 years of age.3,4,5,6 An economic analysis suggests that, in the United Kingdom alone, the long-term costs of perinatal maternal mental health disorders may reach up to £8.1 billion (US $10.8 billion as of December 2017) per year for every 1-year birth cohort, principally because of the impact on the children.7

16 to 18 years of age.3,4,5,6 An economic analysis suggests that, in the United Kingdom alone, the long-term costs of perinatal maternal mental health disorders may reach up to £8.1 billion (US $10.8 billion as of December 2017) per year for every 1-year birth cohort, principally because of the impact on the children.7 Two factors suggested to be important in determining child outcomes in the context of maternal depression are the chronicity and severity of the maternal mood disorder, but data are limited for PND specifically.8,9,10,11,12,13,14 Campbell and colleagues15 found an increased risk for the child if PND persisted to 6 months after childbirth, when infants of mothers with persistent depression were less positive during interactions than infants of women with either no depression or remitted depression. Similarly, Petterson and Albers8 reported that, in a sample of children younger than 5 years, compared with nonexposed children, those exposed to persistent maternal depression had lower cognitive scores (a decrement of 0.47 SDs for girls and 0.36 for boys). Beyond the perinatal period, the Sequenced Treatment Alternatives to Relieve Depression (STAR*D) Child studies, which examined depression in mothers of school-aged children, have shown that the risk of offspring depression increased markedly with both the chronicity and severity of maternal depression.16 In summary, evidence suggests that both persistent PND and severe PND raise the risk of adverse child outcomes.3,8,9,10

udies, which examined depression in mothers of school-aged children, have shown that the risk of offspring depression increased markedly with both the chronicity and severity of maternal depression.16 In summary, evidence suggests that both persistent PND and severe PND raise the risk of adverse child outcomes.3,8,9,10 However, PND that is both severe and persistent has not been well described in the literature, as highlighted by a recent report for the US Preventive Services Task Force.17 Studies have generally included small sample sizes, had short-term follow-up periods, and examined a limited range of outcomes. Systematic research into this question is therefore required to establish the precise duration and symptom thresholds of significance for compromised child functioning. Elucidating the relative magnitude of the consequences for children of different levels of PND severity and persistence could help with targeting intervention resources, thereby limiting the time children are exposed to maternal depression and potentially compromised parenting.

icance for compromised child functioning. Elucidating the relative magnitude of the consequences for children of different levels of PND severity and persistence could help with targeting intervention resources, thereby limiting the time children are exposed to maternal depression and potentially compromised parenting. Using participants in the Avon Longitudinal Study of Parents and Children (ALSPAC), we investigated the sequelae of PND on subsequent maternal depression and child outcome. The ALSPAC involves multiple assessments of maternal depression over the first 18 years of offspring life, including 2 postnatal assessments. Our study had 2 aims: (1) to examine the natural course of different levels of PND severity identified at 2 and 8 months after childbirth using growth curve modeling and (2) to examine the association of PND persistence and severity with child behavioral problems at 3.5 years of age, GCSE mathematics grades at 16 years of age, and self-reported depression at 18 years of age. These 3 offspring outcomes have been shown to be associated with maternal PND in this sample.18,19,20 We hypothesized that (1) women with persistent and severe PND would continue to show elevated levels of depression at the following assessments and (2) children of women with persistent and severe PND would be at an increased risk for behavioral problems, lower mathematics grades, and higher rates of depression.

,19,20 We hypothesized that (1) women with persistent and severe PND would continue to show elevated levels of depression at the following assessments and (2) children of women with persistent and severe PND would be at an increased risk for behavioral problems, lower mathematics grades, and higher rates of depression. Methods Sample Our study sample comprised participants in the ALSPAC cohort (see the eAppendix in the Supplement for full details).21 Ethical approval for the study was obtained from the ALSPAC Law and Ethics Committee and the local research ethics committees. All participants provided written informed consent. Data analysis was conducted from July 12, 2016, to February 8, 2017. We had complete data on maternal depression in the postnatal year for 9848 mothers and at 11 years after childbirth for 6182 mothers. Data were available on child behavioral problems at 3.5 years of age for 8419 children, on GCSE mathematics grades at 16 years of age for 5198 children, and on offspring depression at 18 years of age for 3613 children.

depression in the postnatal year for 9848 mothers and at 11 years after childbirth for 6182 mothers. Data were available on child behavioral problems at 3.5 years of age for 8419 children, on GCSE mathematics grades at 16 years of age for 5198 children, and on offspring depression at 18 years of age for 3613 children. Depression Measures Maternal depression was measured using the self-rated Edinburgh Postnatal Depression Scale (EPDS; score range: 0-30, with higher scores indicating more severe depressive symptoms).19,22 We defined 3 levels of PND severity: 13 to 14 points indicating moderate depression; 15 to 16 points, marked depression; and 17 or more points, severe depression. These threshold levels are consistent with those in previous analyses23 and have been shown to have high specificity and positive predictive value for major depressive disorder (for ≥13 points: 95.7% and 66.7%; for ≥15: 99.3% and 82.7%; and for ≥17 points: 99.7% and 92.4%, respectively).24 The thresholds identified 3 groups of approximately equal size at both 2 and 8 months (ie, at 8 months, 8.7% of the population scored ≥13 points; of those, 3.3% scored 13-14 points, 2.2% scored 15-16 points, and 3.2% scored ≥17 points). Depression was assessed at 2, 8, 21, 33, 61, 73, 93, and 134 months after childbirth to establish the natural course of PND.

roximately equal size at both 2 and 8 months (ie, at 8 months, 8.7% of the population scored ≥13 points; of those, 3.3% scored 13-14 points, 2.2% scored 15-16 points, and 3.2% scored ≥17 points). Depression was assessed at 2, 8, 21, 33, 61, 73, 93, and 134 months after childbirth to establish the natural course of PND. Postnatal depression was identified as persistent when an individual scored above the EPDS threshold at both the 2- and 8-month postnatal assessment. Postnatal depression was considered not persistent when someone scored above the threshold only at the 2-month postnatal assessment. In this way, we were able to examine the association of chronicity with the outcomes at different levels of PND severity. We examined the association of PND with the outcomes for the following 7 mutually exclusive groups: (1) no depression—reference group (EPDS score of <13 points in the postnatal year), (2) moderate but not persistent depression (EPDS score of 13-14 points at 2 months; EPDS score of <13 points at 8 months), (3) marked but not persistent depression (EPDS score of 15-16 points at 2 months; EPDS score of <15 points at 8 months), (4) severe but not persistent depression (EPDS score of ≥17 points at 2 months; EPDS score of <17 points at 8 months), (5) moderate and persistent depression (EPDS score of 13-14 points at 2 months; EPDS score of ≥13 points at 8 months), (6) marked and persistent depression (EPDS score of 15-16 points at 2 months; EPDS score of ≥15 points at 8 months), and (7) severe and persistent depression (EPDS score of ≥17 points at 2 and 8 months).

) moderate and persistent depression (EPDS score of 13-14 points at 2 months; EPDS score of ≥13 points at 8 months), (6) marked and persistent depression (EPDS score of 15-16 points at 2 months; EPDS score of ≥15 points at 8 months), and (7) severe and persistent depression (EPDS score of ≥17 points at 2 and 8 months). Maternal Measures Mothers participating in the ALSPAC indicated their highest education level on a questionnaire, which was sent out in their last trimester of pregnancy (eAppendix in the Supplement). Mothers with education continuing beyond 16 years of age were categorized as having high education (3841 [39%] of the sample), and those with education up to 16 years of age were categorized as having low education (6007 [61%]). Child Measures Rutter Total Problems Scale Child behavioral problems at 3.5 years of age were assessed by maternal report using the total problems scale (sum of hyperactivity and emotional and conduct problems) of the Rutter revised preschool questionnaire.25 The measure was split into quartiles. School-Leaving Mathematics Grades Grades achieved in mathematics were extracted from records of the GCSE, which are external national public examinations taken at 16 years of age at the end of high school in the United Kingdom. The accepted binary coding was the achievement of an A* to a C grade (coded as 0) or no such achievement (coded as 1).

Grades Grades achieved in mathematics were extracted from records of the GCSE, which are external national public examinations taken at 16 years of age at the end of high school in the United Kingdom. The accepted binary coding was the achievement of an A* to a C grade (coded as 0) or no such achievement (coded as 1). Offspring Depression Offspring depression was assessed at 18 years of age using the Clinical Interview Schedule–Revised, a self-administered computerized interview.26 A binary variable indicating a diagnosis of depression on the Clinical Interview Schedule–Revised or no such diagnosis was the outcome measure (eAppendix in the Supplement). Confounding Variables Owing to the relatively small sample size of mothers with PND that was persistent and marked (n = 75) or severe (n = 83), it was important to maintain the sample size and minimize the complexity of statistical models. Therefore, in the adjusted models, we controlled only for maternal education, the only variable in this sample shown to considerably influence the association between PND and child outcomes.3,19

(n = 75) or severe (n = 83), it was important to maintain the sample size and minimize the complexity of statistical models. Therefore, in the adjusted models, we controlled only for maternal education, the only variable in this sample shown to considerably influence the association between PND and child outcomes.3,19 Statistical Analysis We explored the EPDS trajectories of women across 6 repeated assessments from 21 months to 11 years after childbirth using linear growth modeling and controlling for maternal education (Stata command: xtmixed, using mixed-effects maximum likelihood regression and unstructured covariance matrix; Stata, version 13 [StataCorp LLC]). We compared both the intercepts (overall mean EPDS scores across the repeated measures) and the slopes (the extent to which the scores increased or decreased over time, achieved by exploring the interaction between the PND and time variables). We present EPDS means and SDs at 21 months, 33 months, and 11 years. We performed logistic and ordered logistic regressions, controlling for maternal education, to investigate the association between PND and the 3 child outcomes18,19,20 (child behavioral problems at 3.5 years of age, GCSE mathematics grades at 16 years of age, and offspring depression at 18 years of age). We examined different levels of severity and persistence to elucidate the contribution of these factors and to explore a potential dose-response relationship.

child outcomes18,19,20 (child behavioral problems at 3.5 years of age, GCSE mathematics grades at 16 years of age, and offspring depression at 18 years of age). We examined different levels of severity and persistence to elucidate the contribution of these factors and to explore a potential dose-response relationship. Results For the 9848 mothers in the sample, the mean (SD) age at delivery was 28.5 (4.7) years. Of the 8287 children, 4227 (51%) were boys and 4060 (49%) were girls. Long-term Course of Depression Table 1 presents means and SDs of EPDS scores for the 3 levels of PND severity (moderate, marked, and severe) at 21 months, 33 months, and 11 years for mothers with PND that did or did not persist to 8 months.

Results For the 9848 mothers in the sample, the mean (SD) age at delivery was 28.5 (4.7) years. Of the 8287 children, 4227 (51%) were boys and 4060 (49%) were girls. Long-term Course of Depression Table 1 presents means and SDs of EPDS scores for the 3 levels of PND severity (moderate, marked, and severe) at 21 months, 33 months, and 11 years for mothers with PND that did or did not persist to 8 months. Table 1. Mean EPDS Scores for Participants With Depression in the Postnatal Year Level of PND Severity 2 mo After Childbirth (n = 9848) 8 mo After Childbirth (n = 9848) 21 mo After Childbirth (n = 8679) 33 mo After Childbirth (n = 8103) 11 y After Childbirth (n = 6182) Mean (SD) Score Participants, No. (%) Mean (SD) Score Participants, No. (%) Mean (SD) Score Participants, No. (%) Mean (SD) Score Participants, No. (%) Mean (SD) Score Participants, No. (%) Below thresholda 4.71 (3.39) 8878 (90.2) 4.14 (3.27) 8878 (90.2) 4.78 (4.03) 7871 (90.7) 5.36 (4.37) 7409 (91.4) 5.03 (4.80) 5648 (91.4) Moderate but not persistentb 13.42 (0.49) 300 (3.1) 7.64 (3.25) 300 (3.1) 8.17 (4.56) 260 (3.0) 9.63 (4.72) 229 (2.8) 9.09 (5.53) 163 (2.6) Marked but not persistentc 15.44 (0.50) 158 (1.6) 8.97 (3.42) 158 (1.6) 10.75 (5.70) 130 (1.5) 10.48 (4.93) 112 (1.4) 9.08 (6.03) 87 (1.4) Severe but not persistentd 19.06 (2.30) 225 (2.3) 10.26 (4.25) 225 (2.3) 11.07 (5.60) 184 (2.1) 10.87 (5.77) 143 (1.8) 9.93 (5.84) 125 (2.0) Moderate persistente 13.50 (0.50) 129 (1.3) 15.39 (2.96) 129 (1.3) 12.52 (5.00) 108 (1.2) 13.55 (5.00) 109 (1.4) 12.45 (5.64) 78 (1.3) Marked persistentf 15.45 (0.50) 75 (0.8) 17.80 (3.14) 75 (0.8) 15.05 (5.09) 61 (0.7) 13.04 (5.34) 48 (0.6) 11.52 (5.32) 44 (0.7) Severe persistentg 20.66 (3.05) 83 (0.8) 19.95 (2.67) 83 (0.8) 16.29 (4.98) 65 (0.8) 15.30 (6.19) 53 (0.7) 14.49 (6.13) 37 (0.6) Abbreviations: EPDS, Edinburgh Postnatal Depression Scale; PND, postnatal depression.

5 (0.50) 75 (0.8) 17.80 (3.14) 75 (0.8) 15.05 (5.09) 61 (0.7) 13.04 (5.34) 48 (0.6) 11.52 (5.32) 44 (0.7) Severe persistentg 20.66 (3.05) 83 (0.8) 19.95 (2.67) 83 (0.8) 16.29 (4.98) 65 (0.8) 15.30 (6.19) 53 (0.7) 14.49 (6.13) 37 (0.6) Abbreviations: EPDS, Edinburgh Postnatal Depression Scale; PND, postnatal depression. a EPDS score of less than 13 points in the postnatal year. b EPDS score of 13 to 14 points at 2 months and less than 13 points at 8 months. c EPDS score of 15 to 16 points at 2 months and less than 15 points at 8 months. d EPDS score of 17 or more points at 2 months and less than 17 points at 8 months. e EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months. f EPDS score of 15 to 16 points at 2 months and 15 or more points at 8 months. g EPDS score of 17 or more points at 2 and 8 months. Mean scores remained relatively stable from 21 months to 11 years for women with persistent PND. The mean EPDS score of women with persistent moderate PND remained high at subsequent times: 12.52 (5.00) points at 21 months and 12.45 (5.64) points at 11 years. Similarly, the mean EPDS score for women with persistent severe PND remained high with little improvement even up to 11 years after childbirth: 16.29 (4.98) points at 21 months and 14.49 (6.13) points at 11 years.

D remained high at subsequent times: 12.52 (5.00) points at 21 months and 12.45 (5.64) points at 11 years. Similarly, the mean EPDS score for women with persistent severe PND remained high with little improvement even up to 11 years after childbirth: 16.29 (4.98) points at 21 months and 14.49 (6.13) points at 11 years. Table 2 shows the analysis of the linear growth modeling indicating the mean EPDS scores across 6 time points (measured approximately yearly from 21 months to 11 years after childbirth) and the influence of PND severity and persistence on these trajectories. For the sample as a whole, the models indicate that, after the postnatal year, there was little change in the mean EPDS scores over time; mean EPDS scores rose by 0.004 points (95% CI, −0.02 to 0.03) at each repeated assessment. Compared with women who were not above the threshold in the postnatal year (EPDS score <13 points), all other groups had consistently higher EPDS scores up to 11 years after childbirth that progressed in a stepwise function. Compared with women in the reference group, women with an EPDS score of 13 to 14 points at 2 months only (moderate but not persistent PND) had higher EPDS scores of, on average, 3.46 points (95% CI, 2.86-4.05); women whose EPDS scores were 17 or more points at 2 months only (severe but not persistent PND) had higher EPDS scores of, on average, 5.84 points (95% CI, 5.13-6.55); women whose EPDS scores were 13 to 14 points at both 2 and 8 months (moderate and persistent PND) had higher EPDS scores of, on average, 6.91 points (95% CI, 5.98-7.83); and women whose EPDS scores were 17 or more points at both 2 and 8 months (severe and persistent PND) had higher EPDS scores of, on average, 9.90 points (95% CI, 8.73-11.08). There was no evidence that the change in scores (the slope) differed between women with depression that did not persist and women with depression that did persist, suggesting that EPDS scores did not improve over time and were consistently higher compared with women with no depression (Table 2).

s (95% CI, 8.73-11.08). There was no evidence that the change in scores (the slope) differed between women with depression that did not persist and women with depression that did persist, suggesting that EPDS scores did not improve over time and were consistently higher compared with women with no depression (Table 2). Table 2. Mixed-Effects Linear Regression at Different Levels of Postnatal Depression Level of PND Severity Difference in Intercept of EPDS Scores in Postnatal Year, Coefficient (95% CI) P Value Increase in EPDS Scores at Each Assessment, Slope (95% CI) P Value Below thresholda 1 [Reference] NA 1 [Reference] NA Moderate but not persistentb 3.46 (2.86 to 4.05) <.001 0.11 (−0.01 to 0.23) .06 Marked but not persistentc 4.77 (3.93 to 5.62) <.001 −0.04 (−0.81 to 0.13) .66 Severe but not persistentd 5.84 (5.13 to 6.55) <.001 −0.10 (−0.24 to 0.045) .19 Moderate persistente 6.91 (5.98 to 7.83) <.001 0.10 (−0.08 to 0.28) .29 Marked persistentf 8.65 (7.48 to 9.83) <.001 −0.19 (−0.42 to 0.50) .12 Severe persistentg 9.90 (8.73 to 11.08) <.001 0.10 (−0.14 to 0.35) .42 Abbreviations: EPDS, Edinburgh Postnatal Depression Scale; NA, not applicable; PND, postnatal depression. a EPDS score of less than 13 points in the postnatal year. b EPDS score of 13 to 14 points at 2 months and less than 13 points at 8 months. c EPDS score of 15 to 16 points at 2 months and less than 15 points at 8 months. d EPDS score of 17 or more points at 2 months and less than 17 points at 8 months. e EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months.

a EPDS score of less than 13 points in the postnatal year. b EPDS score of 13 to 14 points at 2 months and less than 13 points at 8 months. c EPDS score of 15 to 16 points at 2 months and less than 15 points at 8 months. d EPDS score of 17 or more points at 2 months and less than 17 points at 8 months. e EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months. f EPDS score of 15 to 16 points at 2 months and 15 or more points at 8 months. g EPDS score of 17 or more points at 2 and 8 months. Child Outcomes Table 3 presents the odds ratios (ORs) of adverse child outcomes using the 3 thresholds of PND severity for women whose PND did or did not persist.

e EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months. f EPDS score of 15 to 16 points at 2 months and 15 or more points at 8 months. g EPDS score of 17 or more points at 2 and 8 months. Child Outcomes Table 3 presents the odds ratios (ORs) of adverse child outcomes using the 3 thresholds of PND severity for women whose PND did or did not persist. Table 3. Logistic and Ordinal Logistic Regressions Investigating the Association Between Postnatal Depression and Adverse Child Outcomes, Controlling for Maternal Education Level of PND Severity Behavioral Problems at 3.5 y (n = 7917)a Low GCSE Mathematics Grades at 16 y (n = 4941) Offspring Depression at 18 y (n = 3486) OR (95% CI) P Value OR (95% CI) P Value OR (95% CI) P Value Below thresholdb 1 [Reference] NA 1 [Reference] NA 1 [Reference] NA Moderate but not persistentc 2.22 (1.74-2.83) <.001 1.14 (0.77-1.68) .51 1.11 (0.51-2.44) .79 Marked but not persistentd 1.91 (1.36-2.68) <.001 1.53 (0.89-2.63) .13 2.34 (1.03-5.29) .04 Severe but not persistente 2.39 (1.78-3.22) <.001 1.40 (0.89-2.22) .15 1.72 (0.77-3.82) .18 Moderate persistentf 3.04 (2.10-4.38) <.001 1.65 (0.89-3.05) .11 1.05 (0.32-3.42) .94 Marked persistentg 2.84 (1.71-4.71) <.001 1.32 (0.60-2.90) .46 2.30 (0.67-7.90) .19 Severe persistenth 4.84 (2.94-7.98) <.001 2.65 (1.26-5.57) .01 7.44 (2.89-19.11) <.001 Abbreviations; EPDS, Edinburgh Postnatal Depression Scale; GCSE, General Certificate of Secondary Education; NA, not applicable; OR, odds ratio; PND, postnatal depression. a Using the Rutter revised total problems scale.

Table 3. Logistic and Ordinal Logistic Regressions Investigating the Association Between Postnatal Depression and Adverse Child Outcomes, Controlling for Maternal Education Level of PND Severity Behavioral Problems at 3.5 y (n = 7917)a Low GCSE Mathematics Grades at 16 y (n = 4941) Offspring Depression at 18 y (n = 3486) OR (95% CI) P Value OR (95% CI) P Value OR (95% CI) P Value Below thresholdb 1 [Reference] NA 1 [Reference] NA 1 [Reference] NA Moderate but not persistentc 2.22 (1.74-2.83) <.001 1.14 (0.77-1.68) .51 1.11 (0.51-2.44) .79 Marked but not persistentd 1.91 (1.36-2.68) <.001 1.53 (0.89-2.63) .13 2.34 (1.03-5.29) .04 Severe but not persistente 2.39 (1.78-3.22) <.001 1.40 (0.89-2.22) .15 1.72 (0.77-3.82) .18 Moderate persistentf 3.04 (2.10-4.38) <.001 1.65 (0.89-3.05) .11 1.05 (0.32-3.42) .94 Marked persistentg 2.84 (1.71-4.71) <.001 1.32 (0.60-2.90) .46 2.30 (0.67-7.90) .19 Severe persistenth 4.84 (2.94-7.98) <.001 2.65 (1.26-5.57) .01 7.44 (2.89-19.11) <.001 Abbreviations; EPDS, Edinburgh Postnatal Depression Scale; GCSE, General Certificate of Secondary Education; NA, not applicable; OR, odds ratio; PND, postnatal depression. a Using the Rutter revised total problems scale. b EPDS score of less than 13 points in the postnatal year. c EPDS score of 13 to 14 points at 2 months and less than 13 points at 8 months. d EPDS score of 15 to 16 points at 2 months and less than 15 points at 8 months. e EPDS score of 17 or more points at 2 months and less than 17 points at 8 months. f EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months.

b EPDS score of less than 13 points in the postnatal year. c EPDS score of 13 to 14 points at 2 months and less than 13 points at 8 months. d EPDS score of 15 to 16 points at 2 months and less than 15 points at 8 months. e EPDS score of 17 or more points at 2 months and less than 17 points at 8 months. f EPDS score of 13 to 14 points at 2 months and 13 or more points at 8 months. g EPDS score of 15 to 16 points at 2 months and 15 or more points at 8 months. h EPDS score of 17 or more points at 2 and 8 months. Nonpersistent PND For mothers with PND that was not persistent (ie, depression only at 2 months after childbirth), the risk of child behavioral disturbance at 3.5 years of age was somewhat raised; this risk was similar whatever PND severity threshold was applied. Thus, the OR for child behavioral disturbance for the maternal group with moderate PND was 2.22 (95% CI, 1.74-2.83), for the group with marked PND was 1.91 (95% CI, 1.36-2.68), and for the group with severe PND was 2.39 (95% CI, 1.78-3.22). For the outcomes of GCSE mathematics grades at 16 years of age and offspring depression at 18 years of age, PND that was not persistent was not associated with increased risk. Furthermore, risk did not differ substantially between levels of severity with the exception of the group of mothers with marked PND, with offspring showing higher rates of depression at 18 years of age (OR, 2.34; 95% CI, 1.03-5.29).

t 18 years of age, PND that was not persistent was not associated with increased risk. Furthermore, risk did not differ substantially between levels of severity with the exception of the group of mothers with marked PND, with offspring showing higher rates of depression at 18 years of age (OR, 2.34; 95% CI, 1.03-5.29). Persistent PND Children of women with persistent PND of moderate (OR, 3.04; 95% CI, 2.10-4.38) or marked severity (OR, 2.84; 95% CI, 1.71-4.71) were at higher risk of behavioral problems at 3.5 years of age compared with children of women with PND that was not persistent at any level of severity. The ORs for lower mathematics grades at 16 years of age and depression at 18 years of age were not substantially elevated in the context of either moderate or marked PND. Compared with children of women with an EPDS score of less than 13 points in the postnatal year (reference group), children of women with persistent and severe depression were at the highest risk for all 3 adverse child outcomes (behavioral problems OR, 4.84 [95% CI, 2.94-7.98]; lower GCSE mathematics grades OR, 2.65 [95% CI, 1.26-5.57]; higher depression rate OR, 7.44 [95% CI, 2.89-19.11]). Discussion We used a longitudinal prospective sample from a study with multiple assessments (ALSPAC) to examine the long-term course of PND and the association of PND at varying levels of severity and chronicity with child development. The adverse consequences on child development of severe and persistent PND is of particular note given the long-term follow-up of the children.

study with multiple assessments (ALSPAC) to examine the long-term course of PND and the association of PND at varying levels of severity and chronicity with child development. The adverse consequences on child development of severe and persistent PND is of particular note given the long-term follow-up of the children. Using linear growth modeling, the data indicate that depression scores from 21 months to 11 years show relative stability. The data also indicate a step function, with higher mean depression scores for women whose PND persisted from 2 to 8 months after childbirth compared with women who scored below the threshold and those whose PND was not persistent. Further analysis indicated that, although the intercepts were higher for those with persistent PND, the slopes did not differ, suggesting that women with PND did not improve over time and women with persistent PND consistently remained at relatively higher EPDS scores. We found elevated risks for adverse outcomes for children of women who had persistent PND compared with women whose PND did not persist, and this association was especially pronounced in the group with persistent and severe PND. Postnatal depression that was not persistent either at moderate or marked severity level did not increase the risk in children for lower GCSE mathematics grades or offspring depression.

ared with women whose PND did not persist, and this association was especially pronounced in the group with persistent and severe PND. Postnatal depression that was not persistent either at moderate or marked severity level did not increase the risk in children for lower GCSE mathematics grades or offspring depression. In practice, perinatal depression data are routinely collected in the United Kingdom, as recommended by the National Institute for Health and Clinical Excellence guidelines.27 Our results show that women who meet criteria for PND both early and late in the postpartum year are at an increased risk for prolonged depression. In addition, PND is most likely to raise the risk for adverse child development when PND is severe and persistent. Health care professionals should identify these women for further referral because early and effective treatment could reduce the continued exposure of the child. Owing to the frequent contact with health care professionals in the perinatal period, it is possible to identify women with persistent PND during this period. Identification of women with PND may be associated with increased treatment costs, but the overall cost to the public sector of perinatal mental health problems is 5 times more than the cost of improving services,7 further highlighting that early intervention and effective treatment of perinatal depression are a public health priority.

women with PND may be associated with increased treatment costs, but the overall cost to the public sector of perinatal mental health problems is 5 times more than the cost of improving services,7 further highlighting that early intervention and effective treatment of perinatal depression are a public health priority. To date, the literature is mixed on whether treating maternal depression leads to positive child outcomes, particularly for depression in the early years of life. Treatments for PND have been relatively brief in duration and moderate in intensity; therefore, it is perhaps unsurprising that such interventions have not shown long-term benefits for either the mother or the child.28,29,30,31 A limited number of interventions targeting the mother-child relationship have shown some short-term benefits for outcomes, such as attachment and behavior.32,33,34 This finding highlights the complex issue of treatment recommendations, which may be further compounded by persistent and severe PND, which may in itself require more intensive treatment.

s targeting the mother-child relationship have shown some short-term benefits for outcomes, such as attachment and behavior.32,33,34 This finding highlights the complex issue of treatment recommendations, which may be further compounded by persistent and severe PND, which may in itself require more intensive treatment. Strengths and Limitations A particular strength of our analysis is the large sample, which allowed us to categorize depressive symptoms in a stepwise function. However, the higher thresholds of marked and severe PND do not invariably reflect clinically severe depression, and thus further research using diagnostic instruments is needed. Other strengths of this study include the long-term follow-up and the inclusion of different outcomes at different ages, all of which confirm the increased risk for children of women with persistent and severe PND symptoms.

clinically severe depression, and thus further research using diagnostic instruments is needed. Other strengths of this study include the long-term follow-up and the inclusion of different outcomes at different ages, all of which confirm the increased risk for children of women with persistent and severe PND symptoms. Our study was also subject to some limitations, including a relatively small number of women meeting criteria for persistent and severe PND. The ALSPAC has high attrition, especially during the later time points. The patterns of missing data suggest that children most disadvantaged and more likely to have mothers with depression are overrepresented in the group with missing data20; however, previous imputation analysis of the association between PND and mathematics grades and depression revealed that these associations did not attenuate following imputation.19,20 Women experiencing significant levels of depression may be more likely to opt out of the study. Therefore, our report on the proportion of women who experience severe and persistent PND may be an underestimate. Owing to the small sample size of women with persistent and severe PND, we decided a priori to control for maternal education only, as this is the demographic variable that has consistently been shown to influence the association between PND and child outcomes. Previous analyses showed no attenuation of the association between PND and child outcomes by including other potential confounders.18,19,20

d a priori to control for maternal education only, as this is the demographic variable that has consistently been shown to influence the association between PND and child outcomes. Previous analyses showed no attenuation of the association between PND and child outcomes by including other potential confounders.18,19,20 The ALSPAC study did not collect information on whether women received psychological treatment, but the availability of such treatment at the time was severely limited. The available information indicates that less than 1% of the sample used antidepressants, which is too small a subsample for a subgroup analysis. Finally, women experiencing severe depressive symptoms may be more likely to report child behavioral problems.35 This observation is a potential limitation for 1 of our outcomes; however, we present objective and self-reported outcomes at 16 and 18 years of age indicating a similar pattern of risk when PND is persistent and severe. Conclusions The analyses we conducted highlight that women with persistent depression in the postnatal year continue to experience elevated levels of depressive symptoms until at least 11 years after childbirth. Children of women with persistent PND, especially when it is severe, are at an increased risk for a number of adverse outcomes. Screening both early and late in the first postpartum year will enable the identification of women with persistent PND and thus the offer of appropriate treatment.36 Supplement. eAppendix. Supplementary Material Click here for additional data file.

Introduction In March 1996, the US government mandated that all food manufacturers fortify enriched grain products with 140 μg of folic acid per 100 g of food by January 1, 1998.1 This intervention was implemented to provide increased fetal exposure to folic acid (a synthetic and more highly bioavailable form of naturally occurring folate) in the first month of gestation, a time critical to neural tube development but before many pregnancies are recognized. The fortification rollout rapidly doubled blood folate levels in women of childbearing age2 and substantially diminished neural tube defects in newborns.3,4 At present, 81 countries require folic acid fortification of grain products (eFigure 1 in the Supplement). Folate may play other important roles in the development of the fetal central nervous system, given its contributions to DNA synthesis, DNA and histone methylation, and gene expression. The hypothesis that prenatal exposure to folate may also influence postnatal brain development arises in part from epidemiologic studies that linked starvation during early fetal life with both neural tube defects and schizophrenia,5 and is further supported by studies that linked periconceptional folic acid supplements to lower risk of language delay and autism6,7,8,9; however, 1 study failed to find such an association.10 A critical unanswered question is whether variation in fetal exposure to folate subsequently influences brain development during the formative years preceding late adolescence and early adulthood, a period associated with heightened risk for psychiatric disorders.

however, 1 study failed to find such an association.10 A critical unanswered question is whether variation in fetal exposure to folate subsequently influences brain development during the formative years preceding late adolescence and early adulthood, a period associated with heightened risk for psychiatric disorders. The present study used the US rollout of folic acid fortification of grain products to examine the association between increased fetal exposure to folic acid and subsequent cortical development. Our primary measurement was cortical thickness obtained from magnetic resonance imaging (MRI) scans because it provides a clinically relevant developmental marker. Studies of healthy pediatric samples reveal a steady age-associated decrease in thickness across most of the cortical mantle,11 a pattern thought to reflect synaptic pruning12 and cortical myelination.13,14 Whereas the trajectory of thinning is typically linear,11 departures from this pattern can have functional consequences; delayed onset of thinning has been associated with higher intelligence,15 but accelerated loss of gray matter has been described in patients with schizophrenia16 and their unaffected relatives,17 as well as in school-aged children with autism.18

y linear,11 departures from this pattern can have functional consequences; delayed onset of thinning has been associated with higher intelligence,15 but accelerated loss of gray matter has been described in patients with schizophrenia16 and their unaffected relatives,17 as well as in school-aged children with autism.18 Using data from normative clinical brain MRI scans accessed from the Massachusetts General Hospital (MGH),19 we compared cortical thickness indices within a large cohort of youths born just before, during, or just after the rollout of folic acid fortification and who, therefore, would have received progressively greater exposure to folic acid during gestation. Although little, if any, fortification was in place by September 1996, its rapid deployment ensured that the transition was nearly complete within New England by July 1997.20 Therefore, comparison groups were predefined based on date of birth, so that no individuals in the pre-rollout group (born prior to July 1, 1996) were exposed to fortification during any part of gestation, every individual in the post-rollout group (born after June 30, 1998) was exposed during the entire pregnancy, and individuals in the rollout group (born between these dates) were intermediately exposed. We then turned to 2 additional large, US-based pediatric MRI repositories, the Philadelphia Neurodevelopmental Cohort (PNC) and the National Institutes of Health MRI Study of Normal Brain Development (NIH), to test the reliability and specificity of fortification-related associations with cortical development, and the relevance of fortification-associated MRI changes to psychopathologic characteristics.

lphia Neurodevelopmental Cohort (PNC) and the National Institutes of Health MRI Study of Normal Brain Development (NIH), to test the reliability and specificity of fortification-related associations with cortical development, and the relevance of fortification-associated MRI changes to psychopathologic characteristics. Methods MGH Cohort Patients with brain MRI scans were identified through purposeful sampling of the MGH medical record (eFigure 2A in the Supplement). The search returned 3311 radiology reports, based on both general inclusion criteria (8.0-18.0 years of age at time of scan, date of birth between January 1993 and December 2001, and MRI scans occurring between January 2005 and March 2015) and a predetermined algorithm to optimize age matching of groups. After excluding MRI scans with abnormalities that were identified in the corresponding radiology reports (eTable 1 in the Supplement), and then subjecting the remaining scans to stringent quality control procedures blinded to birthdate (eFigure 2B in the Supplement), we arrived at 292 usable, clinically normative scans, comprising 97 pre-rollout (nonexposed), 96 rollout (partially exposed), and 99 post-rollout (fully exposed) unique individuals (Table 1 and the eAppendix in the Supplement). Study procedures were approved by Partners Human Research Committee, which granted a waiver of informed consent, since this retrospective study of the medical record involved only deidentified data. Table 1. Characteristics of Massachusetts General Hospital Cohort Participants Characteristic Participants, No.

Methods MGH Cohort Patients with brain MRI scans were identified through purposeful sampling of the MGH medical record (eFigure 2A in the Supplement). The search returned 3311 radiology reports, based on both general inclusion criteria (8.0-18.0 years of age at time of scan, date of birth between January 1993 and December 2001, and MRI scans occurring between January 2005 and March 2015) and a predetermined algorithm to optimize age matching of groups. After excluding MRI scans with abnormalities that were identified in the corresponding radiology reports (eTable 1 in the Supplement), and then subjecting the remaining scans to stringent quality control procedures blinded to birthdate (eFigure 2B in the Supplement), we arrived at 292 usable, clinically normative scans, comprising 97 pre-rollout (nonexposed), 96 rollout (partially exposed), and 99 post-rollout (fully exposed) unique individuals (Table 1 and the eAppendix in the Supplement). Study procedures were approved by Partners Human Research Committee, which granted a waiver of informed consent, since this retrospective study of the medical record involved only deidentified data. Table 1. Characteristics of Massachusetts General Hospital Cohort Participants Characteristic Participants, No. (%) Statistics P Value Nonexposed (n = 97) Partially Exposed (n = 96) Fully Exposed (n = 99) Participant-Level Data (RPDR) Age, mean (SD), y 13.3 (2.1) 13.5 (2.8) 12.9 (2.0) F = 1.72 .18 Female sex 49 (50.5) 43 (44.8) 47 (47.5) χ2 = 0.67 .72 Race/ethnicity African American 6 (6.2) 3 (3.1) 6 (6.1) FE = 6.21 .81 Asian 5 (5.2) 2 (2.1) 3 (3.0) White 71 (73.2) 69 (71.9) 73 (73.7) Hispanic 9 (9.3) 11 (11.5) 7 (7.1) Not recorded 2 (2.1) 6 (6.3) 6 (6.1) Other 4 (4.1) 5 (5.2) 4 (4.0) Insurance Private 58 (59.8) 48 (50.0) 67 (67.6) FE = 6.41 .13 Public 38 (39.2) 46 (47.9) 31 (31.3) Other 1 (1.0) 2 (2.1) 1 (1.0) Scanner 1.5-T General Electric 66 (68.0) 41 (42.7) 25 (25.3) FE = 45.5 <.001 1.5-T Siemens Avanto 6 (6.2) 2 (2.1) 3 (3.0) 1.5-T Siemens Aera 0 0 1 (1.0) 3.0-T Siemens Trio 24 (24.7) 50 (52.1) 67 (67.7) 3.0-T Siemens Skyra 1 (1.0) 3 (3.1) 3 (3.0) Neighborhood Block-Level Data (ACS)a Household income, median, $ 81 252 79 368 85 361 F = 0.59 .59 Unemployment 8.6 9.7 8.8 F = 1.28 .28 Highest educational level No high school 5.4 5.3 4.1 Group: F = 1.02; group × level: F = 0.53 .36; .89 Some high school 6.0 6.0 5.5 High school graduate 26.3 26.9 26.2 Some college 15.2 14.4 14.2 Associates degree 7.2 7.7 7.3 Bachelor’s degree 22.2 21.0 23.6 Graduate degree 17.8 17.4 19.0 Vitamin consumption Households reporting use (index) 105.6 104.6 107.3 F = 0.76 .47 Spending per household, $ 76.6 76.1 82.3 F = 0.81 .45 Reason for MRI Scan Attention-deficit/hyperactivity disorder 1 (1.0) 1 (1.0) 0 FE = 1.28 .55 Altered mental status 9 (9.3) 4 (4.2) 9 (9.1) FE = 2.44 .34 Ataxia 4 (4.1) 5 (5.2) 2 (2.0) FE = 1.46 .44 Autism 3 (3.1) 9 (9.4) 7 (7.1) FE = 3.28 .18 Cognitive delay or learning disability 2 (2.1) 5 (5.2) 8 (8.1) FE = 3.60 .16 Family history of neurologic disorder 0 2 (2.1) 1 (1.0) FE = 1.86 .44 Focal neurologic finding 8 (8.2) 13 (13.5) 11 (11.1) χ2 = 1.39 .50 Head injury 8 (8.2) 7 (7.3) 9 (9.1) χ2 = 0.21 .96 Non-CNS tumor or surgery 3 (3.1) 3 (3.1) 3 (3.0) FE = 0.14 >.99 Not given 1 (1.0) 1 (1.0) 0 FE = 1.28 .55 Pituitary or endocrine 1 (1.0) 3 (3.1) 9 (9.1) FE = 7.48 .02 Psychosis 0 2 (2.1) 0 FE = 2.71 .11 Seizures or epilepsy 21 (21.6) 29 (30.2) 27 (27.3) χ2 = 1.88 .40 Somat

d injury 8 (8.2) 7 (7.3) 9 (9.1) χ2 = 0.21 .96 Non-CNS tumor or surgery 3 (3.1) 3 (3.1) 3 (3.0) FE = 0.14 >.99 Not given 1 (1.0) 1 (1.0) 0 FE = 1.28 .55 Pituitary or endocrine 1 (1.0) 3 (3.1) 9 (9.1) FE = 7.48 .02 Psychosis 0 2 (2.1) 0 FE = 2.71 .11 Seizures or epilepsy 21 (21.6) 29 (30.2) 27 (27.3) χ2 = 1.88 .40 Somat ic symptomsb 48 (49.5) 37 (38.5) 45 (45.5) χ2 = 2.39 .31 Syncope 2 (2.1) 4 (4.2) 2 (2.0) FE = 1.05 .61 Previous Medication Use Psychotropic medications Anticonvulsants 15 (15.5) 19 (19.8) 19 (19.4) χ2 = 0.74 .70 Antidepressants 11 (11.3) 12 (12.5) 17 (17.2) χ2 = 1.58 .46 Antipsychotics 8 (8.2) 7 (7.3) 7 (7.1) χ2 = 0.11 .96 Benzodiazepines 11 (11.3) 14 (14.6) 20 (20.2) χ2 = 3.03 .23 Stimulants 3 (3.1) 6 (6.3) 11 (11.1) FE = 4.81 .08 Folic acid or multivitamin 2 (2.1) 2 (2.1) 2 (2.0) FE = 0.22 >.99 Abbreviations: ACS, 2010 American Community Survey; CNS, central nervous system; FE, Fisher exact test (used in lieu of the χ2 test when <5 patients appeared in at least 1 cell); MRI, magnetic resonance imaging; RPDR, Research Patient Data Registry. a Only percentage data, not numbers of individuals, were available from the ACS. b Somatic symptoms: nausea, headache, and/or dizziness.

ic symptomsb 48 (49.5) 37 (38.5) 45 (45.5) χ2 = 2.39 .31 Syncope 2 (2.1) 4 (4.2) 2 (2.0) FE = 1.05 .61 Previous Medication Use Psychotropic medications Anticonvulsants 15 (15.5) 19 (19.8) 19 (19.4) χ2 = 0.74 .70 Antidepressants 11 (11.3) 12 (12.5) 17 (17.2) χ2 = 1.58 .46 Antipsychotics 8 (8.2) 7 (7.3) 7 (7.1) χ2 = 0.11 .96 Benzodiazepines 11 (11.3) 14 (14.6) 20 (20.2) χ2 = 3.03 .23 Stimulants 3 (3.1) 6 (6.3) 11 (11.1) FE = 4.81 .08 Folic acid or multivitamin 2 (2.1) 2 (2.1) 2 (2.0) FE = 0.22 >.99 Abbreviations: ACS, 2010 American Community Survey; CNS, central nervous system; FE, Fisher exact test (used in lieu of the χ2 test when <5 patients appeared in at least 1 cell); MRI, magnetic resonance imaging; RPDR, Research Patient Data Registry. a Only percentage data, not numbers of individuals, were available from the ACS. b Somatic symptoms: nausea, headache, and/or dizziness. Although rapid implementation of the fortification rollout and use of age-matched comparison groups diminished the risk for temporal confounding, we assessed numerous factors that could potentially influence any between-group differences in cortical thickness. To account for socioeconomic and biological diversity in the sample, we extracted from the electronic medical record demographic and socioeconomic information (age at scan, sex, race/ethnicity, and public vs private insurance); reason for MRI scan; and previous use of psychotropic medications, folic acid, or multivitamins. To provide additional measures of socioeconomic status and vitamin use, we performed geospatial analysis to tag patients’ last known addresses to block-level data from the 2010 American Community Survey.21 These data included median household income, household educational attainment, unemployment rate, vitamin intake, and vitamin-related spending (eAppendix in the Supplement).

vitamin use, we performed geospatial analysis to tag patients’ last known addresses to block-level data from the 2010 American Community Survey.21 These data included median household income, household educational attainment, unemployment rate, vitamin intake, and vitamin-related spending (eAppendix in the Supplement). PNC Cohort The PNC participants have been described elsewhere.22 In brief, participants included here comprised a subset of 861 individuals, 8.0 to 18.0 years of age, recruited from community health settings in Philadelphia, Pennsylvania. Participants underwent standardized clinical and MRI assessment using a single 3-T magnet. Clinical assessment23 characterized participants as either typically developing or exhibiting psychiatric symptoms, categorized as psychosis spectrum, attenuated psychosis, or other types of psychopathologic conditions (eAppendix in the Supplement). All included MRI scans passed stringent quality control as previously described.24 The distribution of birthdates among 8- to 18-year-olds in the PNC sample was such that nonexposed (n = 322), partially exposed (n = 189), and fully exposed (n = 350) individuals were well represented (eFigure 3 in the Supplement). Study procedures were approved by the institutional review boards of the Children’s Hospital of Philadelphia and the University of Pennsylvania. Adult participants provided written informed consent. Minors provided assent, and their parent or guardian provided written informed consent.

ed (eFigure 3 in the Supplement). Study procedures were approved by the institutional review boards of the Children’s Hospital of Philadelphia and the University of Pennsylvania. Adult participants provided written informed consent. Minors provided assent, and their parent or guardian provided written informed consent. NIH Cohort The NIH participants have been characterized elsewhere.25 In brief, healthy youths were recruited across 6 sites nationwide and underwent MRI scans on 1.5-T magnets up to 3 times at various ages. Here, we selected a subsample from this cohort whose MRI scans previously passed stringent image quality control,11 and we constrained it to our age interval of interest (8.0-18.0 years). We also excluded participants who might have been exposed to folic acid fortification based on their age at first enrollment. The final sample included 217 individual participants and 383 MRI scans. All procedures were approved by the relevant institutional review board at each of the 6 pediatric study centers, where the MRI scans took place, at a clinical coordinating center at Washington University in St Louis, and at a data coordinating center at the Montreal Neurological Institute, McGill University. Written informed consent was obtained from parents and adult participants, and minors provided assent.

atric study centers, where the MRI scans took place, at a clinical coordinating center at Washington University in St Louis, and at a data coordinating center at the Montreal Neurological Institute, McGill University. Written informed consent was obtained from parents and adult participants, and minors provided assent. Statistical Analysis Primary analyses used general linear models in FreeSurfer, version 5.0 (Martinos Center for Biomedical Imaging). Main analyses in the MGH cohort contrasted mean cortical thickness and age-associated change in thickness (linear and quadratic models) in the fully exposed vs nonexposed groups. Main analyses in the contemporaneous PNC cohort and comparison NIH cohort assessed for significant quadratic associations of age with cortical thickness across each cohort. Nuisance variables, including age, sex, total brain volume (all cohorts), scanner field strength (MGH), and site (NIH), were entered as covariates of no interest. To ensure coverage of the entire cortex, we did not limit the analysis to a priori anatomical regions of interest. Rather, we used 10 000 Monte Carlo simulations to determine whether identified clusters bounded by a vertexwise threshold of P < .05 were sufficiently large to survive control for multiple comparisons across the entire surface (clusterwise P < .05). For clusters demonstrating significant quadratic associations of age with thickness, the delay in cortical thinning (ie, time until thinning onset) was estimated using least squares analyses (MATLAB, version R2015b; Mathworks Inc). To assess the associations of cortical thinning delay with clinical outcomes in the PNC cohort, multinomial logistic regression examined the association of the local age-thickness slope with adjusted odds for diagnosis of psychosis spectrum, attenuated psychosis, and other types of psychopathologic conditions compared with typically developing participants (SPSS, version 25 [SPSS Inc]; eAppendix in the Supplement).

ltinomial logistic regression examined the association of the local age-thickness slope with adjusted odds for diagnosis of psychosis spectrum, attenuated psychosis, and other types of psychopathologic conditions compared with typically developing participants (SPSS, version 25 [SPSS Inc]; eAppendix in the Supplement). Results MGH Cohort Included and excluded patients were comparable across demographic measures (eTable 2 in the Supplement). Among included individuals, exposure groups did not differ significantly by age at MRI scan, sex, scan indication, or insurance status; a slight but nonsignificant increase in use of psychotropic medications was noted over time, consistent with previous epidemiologic studies.26 The distribution of scanner field strengths differed among groups owing to a shift from 1.5- to 3-T clinical magnets in the late 2000s, a factor taken into account in the main analyses. Tagging patients’ last known addresses to block-level data obtained through the 2010 American Community Survey, we observed no differences among fortification groups in per capita use or spending on nonprescription vitamins, or in other measures that could affect folate intake (eg, median income, household educational level, or unemployment rate) (Table 1).

to block-level data obtained through the 2010 American Community Survey, we observed no differences among fortification groups in per capita use or spending on nonprescription vitamins, or in other measures that could affect folate intake (eg, median income, household educational level, or unemployment rate) (Table 1). Group differences in cortical thickness were observed in bilateral frontal and inferior temporal regions. In each significant cluster, cortical thickness was higher in the fully exposed group compared with the nonexposed group, with intermediate effects observed in the partially exposed group (Figure 1A and B; eFigure 4 and eTable 3 in the Supplement). Sensitivity analyses revealed significant associations of scanner field strength and manufacturer with cortical thickness, but the direction of these associations varied by region, consistent with prior studies.27,28 Between-group differences remained significant after adjustment for these variables, and no significant group × field strength or group × manufacturer interactions were observed (eTable 4 in the Supplement).

ortical thickness, but the direction of these associations varied by region, consistent with prior studies.27,28 Between-group differences remained significant after adjustment for these variables, and no significant group × field strength or group × manufacturer interactions were observed (eTable 4 in the Supplement). Figure 1. Fortification-Associated Cortical Thickness Changes in the Massachusetts General Hospital Cohort A, Surface-wide maps of cortical thickness in fully exposed (n = 99) minus nonexposed (n = 97) individuals reveal higher thickness among youths who were exposed to folic acid fortification during gestation. Images are masked to show only clusters that survive correction for multiple comparisons. B, Dot plots showing cortical thickness in the left frontal cluster as a function of exposure group, suggesting intermediate effects in the partially exposed group. Horizontal lines in the boxes indicate median values, and shaded boxes indicate interquartile ranges. Cortical thickness values are z-transformed residuals after controlling for nuisance covariates. Cool colors (shades of blue) show regions for which cortical thickness is greater in the group that was not exposed to fortification, whereas hot colors (red, orange, and yellow) show regions where cortical thickness is greater in the fully exposed group. C, Age-centered regression analyses indicate clusters with significant between-group differences in thickness as a function of age at magnetic resonance imaging (MRI) scan. This analysis indicates that overall group differences largely reflect exposure-related associations within younger individuals. ITG indicates inferior temporal gyrus; L, left, and R, right.

e clusters with significant between-group differences in thickness as a function of age at magnetic resonance imaging (MRI) scan. This analysis indicates that overall group differences largely reflect exposure-related associations within younger individuals. ITG indicates inferior temporal gyrus; L, left, and R, right. To understand this pattern in the context of age-associated change in cortical thickness, we next assessed for group differences in intercept (ie, thickness means centered at 8 years of age) and slope (linear age effects), as well as for any differences in nonlinear (quadratic) age-associated change. Intercept in the bilateral frontal cortex (pars orbitalis and precentral) and the right inferior temporal gyrus was higher in the fully exposed group than in the nonexposed group; these differences diminished with age (Figure 1C; eFigure 5 and eTable 3 in the Supplement; and Video). There were no significant differences in the cortical thickness–age slope that contributed to the increased cortical thickness in the exposed groups. However, in the left inferior temporal cortex, as well as in the left inferior parietal cortex, we observed significant differences in age-squared effects, where cortical thinning in fully exposed participants was delayed compared with cortical thinning in nonexposed participants (Figure 2A and B; and eTable 5 in the Supplement). Additional modeling using least squares regression localized the onset of cortical thinning in fully exposed participants, defined by the optimal break point between flat and sloped lines, to 13.0 years of age (left inferior temporal gyrus) and 13.8 years of age (left inferior parietal lobule; eFigure 6A in the Supplement).

ional modeling using least squares regression localized the onset of cortical thinning in fully exposed participants, defined by the optimal break point between flat and sloped lines, to 13.0 years of age (left inferior temporal gyrus) and 13.8 years of age (left inferior parietal lobule; eFigure 6A in the Supplement). Video. Age-Related Group Differences in Cortical Thickness Lateral views of left and right cortex demonstrate dynamic effects of prenatal fortification exposure (fully exposed minus nonexposed) on cortical thickness between 8 and 18 years of age in the Massachusetts General Hospital cohort. In general, the most pronounced differences occurred at earlier ages. Color bar indicates effect size (Cohen d) and direction of effect.

e dynamic effects of prenatal fortification exposure (fully exposed minus nonexposed) on cortical thickness between 8 and 18 years of age in the Massachusetts General Hospital cohort. In general, the most pronounced differences occurred at earlier ages. Color bar indicates effect size (Cohen d) and direction of effect. Figure 2. Fortification-Associated Emergence of Nonlinear (Delayed) Cortical Thinning A, Surface-wide maps of age-squared associations with cortical thickness in fully exposed (n = 99) minus nonexposed (n = 97) Massachusetts General Hospital (MGH) cohort scans reveal increased age-related quadratic thinning among individuals exposed to folic acid fortification during gestation. Nonexposed greater than fully exposed indicates that cool colors show regions for which β values are greater in the group that was not exposed to fortification. More negative β values reflect stronger quadratic thinning. The cool color (ie, shades of blue) reflects the negative age-squared term in the fully exposed group. B, Age-thickness scatterplot of MGH cohort, indicating emergence of quadratic (delayed) left inferior parietal lobule thinning in participants born after fortification was implemented. C, Surface-wide maps of age-squared associations with cortical thickness in the Philadelphia Neurodevelopmental Cohort (PNC) (n = 861) indicate age-related quadratic thinning in frontal, inferior parietal lobule (IPL), and inferior temporal gyrus (ITG) regions, again driven by delayed thinning in fully exposed individuals. D, Age-thickness scatterplot of the PNC cohort demonstrating delayed thinning in left (L) IPL and right (R) ITG. E, Within the L IPL cluster that demonstrated exposure-associated differences in quadratic thinning in the MGH cohort (A and B), analysis of the nonexposed National Institutes of Health Magnetic Resonance Imaging Study of Normal Brain Development (NIH) cohort indicates linear thinning (evident at the earliest time point). F, Similarly, within the L IPL and R ITG clusters that demonstrated quadratic thinning in the PNC cohort (C and D), only linear thinning was seen in the NIH cohort (383 scans). A and B, Images are masked to show only clusters that survive correction for multiple comparisons (P < .05, clusterwise; for PNC cohort L IPL, the displayed cluster was too small to survive correction at P < .05 but is significant at P < .01). Cortical thickness values in scatterplots represent z-transformed residuals after controlling for nuisance covariates.

usters that survive correction for multiple comparisons (P < .05, clusterwise; for PNC cohort L IPL, the displayed cluster was too small to survive correction at P < .05 but is significant at P < .01). Cortical thickness values in scatterplots represent z-transformed residuals after controlling for nuisance covariates. PNC and NIH Cohorts The MGH data set relied on clinical MRI scans that were acquired through nonuniform clinical protocols, using different magnets, and within a single US city. To verify and generalize the findings from the MGH cohort, we next turned to 2 additional large US cohorts that were studied prospectively in standardized research settings (Table 2): 1 cohort with birthdates centered around the folic acid fortification rollout and MRI scans performed with a single 3-T magnet (PNC; eFigure 3 in the Supplement), and 1 cohort that included only youths who were born prior to folic acid fortification (NIH).

died prospectively in standardized research settings (Table 2): 1 cohort with birthdates centered around the folic acid fortification rollout and MRI scans performed with a single 3-T magnet (PNC; eFigure 3 in the Supplement), and 1 cohort that included only youths who were born prior to folic acid fortification (NIH). Table 2. Characteristics of PNC and NIH Cohort Participants Characteristic PNC Cohort (N = 861) NIH Cohort (N = 217) Nonexposed (n = 322) Partially Exposed (n = 189) Fully Exposed (n = 350) Statistics P Value Age, mean (SD), y 16.3 (1.0) 13.9 (0.8) 10.7 (1.4) F = 2032 <.001 13.3 (2.6)a Female sex, No. (%) 183 (56.8) 95 (50.3) 166 (47.4) χ2 = 6.10 .05 118 (54.4) Race, No. (%) African American 132 (41.0) 89 (47.1) 151 (43.1) χ2 = 5.35 .72 18 (8.3) American Indian or Alaskan 0 1 (0.5) 1 (0.3) 0 Asian 4 (1.2) 1 (0.5) 6 (1.7) 1 (0.5) White 152 (47.2) 77 (40.7) 151 (43.1) 171 (78.8) Hawaiian or Pacific Islander 0 0 0 0 >1 Race 34 (10.6) 21 (11.1) 41 (11.7) 27 (12.4) Ethnicity, No. (%) Hispanic 15 (4.7) 10 (5.3) 36 (10.3) χ2 = 9.25 .01 29 (13.4) Maternal educational level, No. (%) <High school graduate 13 (4.0) 8 (4.2) 18 (5.1) χ2 = 15.1 .23 1 (0.5) High school graduate 109 (33.9) 63 (33.3) 103 (29.4) 33 (15.2) Some college or Associates degree 60 (18.6) 56 (29.6) 87 (24.9) 61 (28.1) Bachelor’s degree 80 (24.8) 37 (19.6) 91 (26.0) 73 (33.6) Some graduate education or Master’s degree 35 (10.9) 11 (5.8) 30 (8.6) 6 (2.8) Graduate degree 13 (4.0) 8 (4.2) 13 (3.7) 41 (18.9) Data not available 12 (3.7) 6 (3.2) 8 (2.3) 2 (0.9) Paternal educational level, No. (%) <High school graduate 14 (4.3) 11 (5.8) 23 (6.6) χ2 = 17.5 .13 5 (2.3) High school graduate 117 (36.3) 74 (39.2) 136 (38.9) 42 (19.4) Some college or Associates degree 45 (14.0) 36 (19.0) 58 (16.6) 52 (24.0) Bachelor’s degree 67 (20.8) 23 (12.2) 59 (16.9) 62 (28.6) Some graduate education or Master’s degree 25 (7.8) 15 (7.9) 34 (9.7) 7 (3.2) Graduate degree 19 (5.9) 5 (2.6) 14 (4.0) 47 (21.7) Data not available 35 (10.9) 25 (13.2) 26 (7.4) 2 (0.9) Abbreviations: NIH, National Institutes of Health Magnetic Resonance Imaging Study of Normal Brain Development; PNC, Philadelphia Neurodevelopmental Cohort.

n or Master’s degree 25 (7.8) 15 (7.9) 34 (9.7) 7 (3.2) Graduate degree 19 (5.9) 5 (2.6) 14 (4.0) 47 (21.7) Data not available 35 (10.9) 25 (13.2) 26 (7.4) 2 (0.9) Abbreviations: NIH, National Institutes of Health Magnetic Resonance Imaging Study of Normal Brain Development; PNC, Philadelphia Neurodevelopmental Cohort. a Mean age of participant across all scans, including up to 3 scans per participant. Because MRI scans in the PNC data set were collected during a relatively brief time, nonexposed participants were significantly older than partially exposed or fully exposed participants. Whereas between-group cortical thickness differences would be strongly confounded by age differences, examination of age-associated thinning contours provided an opportunity to replicate findings from the MGH cohort. Specifically, younger PNC participants (who were exposed to folic acid fortification) should demonstrate delayed cortical thinning, whereas younger NIH participants (who were not exposed to folic acid fortification) should not demonstrate delayed cortical thinning.

opportunity to replicate findings from the MGH cohort. Specifically, younger PNC participants (who were exposed to folic acid fortification) should demonstrate delayed cortical thinning, whereas younger NIH participants (who were not exposed to folic acid fortification) should not demonstrate delayed cortical thinning. Within the PNC cohort, quadratic (delayed) age-related thinning was observed in 4 clusters that overlapped with those identified in the MGH analysis: left frontal, right inferior temporal, left inferior parietal, and right inferior parietal (Figure 2C and D; and eFigure 7A and B and eTable 6 in the Supplement). Least squares regression localized the onsets of cortical thinning to between 13.0 and 14.3 years of age (eFigure 6B in the Supplement). Conversely, bilateral lingual gyrus (which was not implicated in the MGH cohort) exhibited significantly accelerated cortical thinning.

igure 7A and B and eTable 6 in the Supplement). Least squares regression localized the onsets of cortical thinning to between 13.0 and 14.3 years of age (eFigure 6B in the Supplement). Conversely, bilateral lingual gyrus (which was not implicated in the MGH cohort) exhibited significantly accelerated cortical thinning. To confirm that quadratic thinning effects emerged largely after folic acid fortification, we turned to the NIH cohort, in which all included participants were born prior to the rollout. Age-related associations with cortical thickness were assessed within the same 6 clusters that showed delayed thinning in the MGH (left inferior temporal gyrus and left inferior parietal lobule) or PNC (right inferior temporal gyrus, left frontal cortex, left inferior parietal lobule, and right inferior parietal lobule) cohorts. Consistent with a previous analysis11 that included a more extended age range and that was not limited to nonexposed participants, nonlinear cortical thinning in the NIH cohort was sparse, with only 1 cluster demonstrating significant quadratic thinning (left frontal cortex; Figure 2E and F; and eFigure 7C and eTable 7 in the Supplement). Even so, within this cluster, the break point for thinning occurred at a significantly younger age when compared with the PNC break point (χ2 = 11.87; P < .001; eFigure 6C in the Supplement).

onstrating significant quadratic thinning (left frontal cortex; Figure 2E and F; and eFigure 7C and eTable 7 in the Supplement). Even so, within this cluster, the break point for thinning occurred at a significantly younger age when compared with the PNC break point (χ2 = 11.87; P < .001; eFigure 6C in the Supplement). Risk of Psychosis To gauge the association of altered cortical thinning with clinically relevant phenotypes, we again turned to the PNC cohort, which included a detailed, standardized clinical characterization of all participants. Of the 861 youths included in the MRI analysis, clinical evaluations determined that 248 were typically developing, 199 had a diagnosis of psychosis spectrum, 105 had attenuated psychotic symptoms, and the remaining 309 had various other psychopathologic conditions, as previously described (eAppendix in the Supplement).23,24,29 For each of the 4 PNC regions that demonstrated postfortification quadratic thinning, best-fit local thinning slopes were calculated for each participant, based on linear change in cortical thickness across an optimized age range (1 year) centered around that participant (Figure 3A). These local slopes were then evaluated as factors associated with participant-level diagnosis of psychosis spectrum, attenuated psychosis, or other types of psychopathologic conditions vs typical development using multinomial logistic regression, controlling for nuisance covariates. Across 3 of 4 regions, flatter (ie, less negative) local slopes were associated with significantly reduced adjusted odds of psychosis spectrum diagnosis (odds ratio, 0.37-0.59; P < .001 to P = .02, Figure 3B and C). Local slopes were not associated with risk of other types of psychopathologic conditions in any region. For participants with attenuated psychosis, local slope associations were stronger than for other types of psychopathologic conditions, but nonsignificant.

dds ratio, 0.37-0.59; P < .001 to P = .02, Figure 3B and C). Local slopes were not associated with risk of other types of psychopathologic conditions in any region. For participants with attenuated psychosis, local slope associations were stronger than for other types of psychopathologic conditions, but nonsignificant. Figure 3. Local Slope Derivation and Association of Delayed Cortical Thinning With Individual Risk for Psychopathologic Conditions in Participants in the Philadelphia Neurodevelopmental Cohort (PNC) A, In each of the 4 clusters exhibiting delayed cortical thinning in the PNC cohort, local cortical thinning slope was calculated for each individual based on the best-fit line of thickness vs age among all nearby individuals (±6 months). Insets demonstrate local slopes for an 11.0-year-old (left) and a 15.0-year-old (right) participant. B, Mean local slopes for participants in each diagnostic group, in each of the 4 clusters. For example, individuals with psychosis spectrum (PS) symptoms tended to have more negative slopes than those with other diagnoses. Local slopes reflect change in z-transformed cortical thickness scores (adjusted for nuisance covariates) during 1 year. C, Multinomial logistic regression models associated with diagnosis (PS, psychosis low [PL], or other psychopathologic condition [OP] relative to typically developing [TD]; n = 248) of each participant based on local slope, covarying for age, sex, total brain volume, and method of ascertaining diagnosis. Lower adjusted odds ratios indicate reduced odds of psychopathologic condition in the presence of flatter (less negative) local thinning slope, a pattern that was significant for PS in 3 of 4 regions tested. All P values are false discovery rate corrected. All error bars indicate 95% CIs. IPL indicates inferior parietal lobule; ITG, inferior temporal gyrus.

uced odds of psychopathologic condition in the presence of flatter (less negative) local thinning slope, a pattern that was significant for PS in 3 of 4 regions tested. All P values are false discovery rate corrected. All error bars indicate 95% CIs. IPL indicates inferior parietal lobule; ITG, inferior temporal gyrus. aP < .001. bP < .01. cP < .05. Finally, to confirm that the associations of fortification exposure with cortical thinning were not themselves confounded by the inclusion of individuals with psychosis spectrum symptoms in the PNC cohort, we repeated the original analysis of quadratic thinning using only typically developing individuals or participants with other types of psychopathologic conditions (n = 541). Significant quadratic age-associated thinning persisted in all regions after exclusion of individuals with psychosis spectrum symptoms and youths with attenuated psychosis (eTable 8 in the Supplement).

using only typically developing individuals or participants with other types of psychopathologic conditions (n = 541). Significant quadratic age-associated thinning persisted in all regions after exclusion of individuals with psychosis spectrum symptoms and youths with attenuated psychosis (eTable 8 in the Supplement). Discussion Evaluating 3 independent MRI cohorts of 8- to 18-year-old youths, we investigated the association of prenatal exposure to folic acid fortification with subsequent cortical development through adolescence. Within a large clinical cohort (MGH), we observed widespread increases in frontal and temporal cortical thickness between comparable groups of youths who gestated just after, compared with just before, the rollout of folic acid fortification. Youths who gestated during the rollout, and who therefore had partial exposure, demonstrated intermediate increases, consistent with a dose association. Exposure-associated differences were most pronounced in younger individuals, and in 2 regions (left inferior temporal and inferior parietal), we observed a delay in the onset of age-associated cortical thinning. An analogous pattern was evident in the contemporaneous PNC cohort: individuals who were exposed to folic acid fortification exhibited delays of cortical thinning of similar duration, which occurred in similar frontal, temporal, and parietal regions as those identified in the MGH cohort. Flatter age-related thinning profiles were associated with reduced risk of psychosis spectrum symptoms in this cohort. In contrast, the NIH cohort, comprising only individuals who were not exposed to fortification, exhibited earlier cortical thinning in the same regions. Collectively, these data suggest an association of prenatal exposure to folic acid fortification with increased cortical thickness through early adolescence, accompanied by delayed onset of cortical thinning and reduced risk of psychosis.

xposed to fortification, exhibited earlier cortical thinning in the same regions. Collectively, these data suggest an association of prenatal exposure to folic acid fortification with increased cortical thickness through early adolescence, accompanied by delayed onset of cortical thinning and reduced risk of psychosis. Adolescence directly precedes the period of greatest risk for psychiatric disorders, some of which are characterized by reductions in cortical thickness present at the onset of illness.30 Furthermore, some of the most severe child-onset psychiatric disorders, including autism and early-onset schizophrenia, are associated with marked accelerations in loss of gray matter during the age range that we studied.16,18 Within the PNC cohort, reductions in gray matter in multiple brain regions were associated with psychosis spectrum status in a previous analysis conducted without regard to exposure to folic acid fortification24; however, these regions differed from those demonstrating fortification-associated thinning delays herein. Rather, within these regions, shallower thinning slopes were associated with reduced risk for psychosis spectrum symptoms, suggesting a possible protective effect of fortification-associated delays in cortical thinning. This association was relatively specific for psychosis because local slopes were not associated with other psychopathologic conditions, and attenuated associations were seen in individuals with milder psychotic spectrum symptoms. The present findings are consistent with recent reports of salutary behavioral outcomes after periconceptional intake of folic acid6,7,8,9,31 and with a recent study linking maternal folate deficiency with reduced brain volume in a cohort of young European children.32

ndividuals with milder psychotic spectrum symptoms. The present findings are consistent with recent reports of salutary behavioral outcomes after periconceptional intake of folic acid6,7,8,9,31 and with a recent study linking maternal folate deficiency with reduced brain volume in a cohort of young European children.32 Limitations Although relatively large imaging cohorts, replication and temporal specificity analyses using independent samples, and linkage of imaging and clinical findings represent the strengths of our study, a number of potential limitations warrant consideration. Unrecognized temporal confounders are of particular concern in natural experimental designs. Critically, the rapid deployment of folic acid fortification in the United States allowed us to study a relatively narrow and continuous range of birthdates, which diminished the risk of potential temporal confounders, including group differences in postnatal exposure to folic acid fortification. We also sought to address a number of other potential clinical, demographic, and socioeconomic confounders through extended use of electronic medical records and block-level American Community Survey data, but we saw no substantial differences among exposure groups. Although, on the population level, the rollout of folic acid fortification rapidly doubled blood folate levels without changing folic acid supplement (vitamin) use,2,33 the lack of individual-level data on maternal folate intake represents an inevitable limitation of this experiment’s design. However, exposure groups were comparable on block-level vitamin spending and consumption.

rtification rapidly doubled blood folate levels without changing folic acid supplement (vitamin) use,2,33 the lack of individual-level data on maternal folate intake represents an inevitable limitation of this experiment’s design. However, exposure groups were comparable on block-level vitamin spending and consumption. Although exposure groups in the MGH cohort were generally well matched, we cannot rule out the potential role of magnetic field strength differences within that cohort. Field strength effects would not be expected to bias results consistently because previous work indicates that the associations of field strength with cortical thickness measurements vary in direction and magnitude across the cortex27; recognizing this heterogeneity, field strength was entered as a covariate in the surface-wide analysis. Furthermore, within regions demonstrating between-group differences, these differences remained significant in sensitivity analyses that controlled for scanner field strength and manufacturer. Perhaps more important, that similar exposure-related associations with cortical thinning were observed in the PNC cohort, which used a single 3-T magnet, suggests that scanner differences did not substantially influence the MGH cohort results. More broadly, similar findings across 2 cohorts that differed substantially in terms of population sampling strategy (retrospective and clinical vs prospective and community-based), imaging and clinical assessment (medical record–based vs standardized), and geographic location—but that were comparable in terms of exposure, sex, age, race/ethnicity, and urbanicity—suggest reliability of the findings.

ms of population sampling strategy (retrospective and clinical vs prospective and community-based), imaging and clinical assessment (medical record–based vs standardized), and geographic location—but that were comparable in terms of exposure, sex, age, race/ethnicity, and urbanicity—suggest reliability of the findings. The present findings also raise questions about the long-lasting effects of variation in the fetal methylome because folate supplies 1-carbon moieties that regulate gene expression. One possibility suggested by the present results is that programming of cortical maturation in youths is sensitive to fetal folate levels, potentially via epigenetic modification of genes that regulate cortical development,34 repair of de novo mutations,35 or mitigation of toxic exposures.36,37,38,39 Although, to our knowledge, these results are the first to link prenatal exposure to folic acid fortification to changes in subsequent cortical development, the specific mechanisms underlying these effects have yet to be elucidated. Conclusions In replicated cohorts, fetal exposure to population-wide folic acid fortification was associated with subsequent alterations in cortical development among school-aged youths. In turn, these cortical changes were associated with reduced risk of psychosis. Our findings suggest that protective effects of prenatal folic acid exposure may extend beyond prevention of neural tube defects and span neurodevelopment during childhood and adolescence. Supplement. eAppendix. Methods

Conclusions In replicated cohorts, fetal exposure to population-wide folic acid fortification was associated with subsequent alterations in cortical development among school-aged youths. In turn, these cortical changes were associated with reduced risk of psychosis. Our findings suggest that protective effects of prenatal folic acid exposure may extend beyond prevention of neural tube defects and span neurodevelopment during childhood and adolescence. Supplement. eAppendix. Methods eFigure 1. Countries and Territories With Mandatory Folic Acid Fortification of Grain Products, as of November 2017 eFigure 2. Medical Record Search Algorithm and Image Processing Pipeline for the MGH Cohort eFigure 3. Distribution of Birth Year Among MGH and PNC Cohort Individuals eFigure 4. Additional Clusters Demonstrating Exposure-Related Effects on Cortical Thickness in the MGH Cohort eFigure 5. Comparison of Significant Mean Age-Centered and Intercept (Age 8)–Centered Clusters Showing Exposure-Related Effects in the MGH Cohort eFigure 6. Estimate of Thinning Delay in MGH and PNC Clusters Showing Quadratic Age-Related Thinning eFigure 7. Additional Clusters Demonstrating Quadratic Age-Related Thinning in the PNC Cohort eTable 1. Radiologic Exclusion Criteria for MGH Cohort eTable 2. Comparison of Included And Excluded Individuals in MGH Cohort eTable 3. Clusters Demonstrating Significantly Increased Cortical Thickness in Fully Exposed Versus Non-Exposed Youths in MGH Cohort eTable 4. Effects of Scanner Strength and Manufacturer Differences on Group Cortical Thickness Differences in MGH Cohort

eTable 2. Comparison of Included And Excluded Individuals in MGH Cohort eTable 3. Clusters Demonstrating Significantly Increased Cortical Thickness in Fully Exposed Versus Non-Exposed Youths in MGH Cohort eTable 4. Effects of Scanner Strength and Manufacturer Differences on Group Cortical Thickness Differences in MGH Cohort eTable 5. Clusters Demonstrating Significantly Different Quadratic Age-Related Thinning in Fully Exposed Versus Non-Exposed Youths in MGH Cohort eTable 6. Clusters Demonstrating Quadratic Age-Related Thinning in PNC Cohort eTable 7. Tests for Quadratic Age-Related Thinning in NIH Cohort, Within Clusters Demonstrating Significant Quadratic Thinning in MGH or PNC Cohorts eTable 8. Quadratic Age-Related Thinning in PNC Cohort Before and After Exclusion of Youths With Psychotic Symptoms eReferences Click here for additional data file.

Introduction Individuals who are prodromal to schizophrenia have a higher risk for and transition rate to psychosis compared with the general population.1,2,3,4,5,6,7 Cognitive deficits are also a predictor associated with psychosis.3,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27 Cognitive impairments are the core disabling factors in psychosis and schizophrenia.28,29,30,31 Meta-analytic evidence indicates that cognitive deficits are present in individuals at ultrahigh risk (UHR) for psychosis.24,25,26,32,33,34 There is a 35% likelihood that the presence of symptoms—functional or cognitive manifestations—in high-risk, care-seeking individuals predates psychosis.6 However, systematic evidence is scarce for longitudinal cognitive trajectories in individuals at UHR for psychosis. Recent reports confirm that cognitive deficits at baseline are associated with conversion to psychosis, but the reports have not addressed the longitudinal cognitive profiles of these individuals.27 Equivocal evidence ranges from modest improvements in cognition in converters and first episode psychosis26 to suggestions that cognitive decline may be a strong factor in eventual psychosis.33,35,36 Previous reports indicate that approximately 50% of individuals at UHR for psychosis improve spontaneously within a short follow-up time frame.37

from modest improvements in cognition in converters and first episode psychosis26 to suggestions that cognitive decline may be a strong factor in eventual psychosis.33,35,36 Previous reports indicate that approximately 50% of individuals at UHR for psychosis improve spontaneously within a short follow-up time frame.37 Longitudinal schizophrenia cognitive studies may offer insights to UHR cognitive trajectories. Premorbid cognitive deficits were found to be associated with schizophrenia.33,38,39,40 Cognitive impairment can be observed also in nonpsychotic family members of psychotic patients.41,42 Progressive changes in cognition over a 30-year period were reported in children who later developed schizophrenia.33 Two aspects of cognitive trajectories may be investigated: (1) means-based change, where differential time-based cognitive changes may exist between healthy individuals and those at UHR for psychosis, and (2) covariance-based change. The latter involves changes in the cognitive component structure, as defined by cognitive tests, over time43,44 and is known as the cognitive dedifferentiation hypothesis. This dedifferentiation is associated with poorer cognitive function with increased covariation across cognitive tests, a phenomenon previously observed in aging research.43,44 Intriguingly, forms of cognitive dedifferentiation were also noted in schizophrenia,45,46 where a subtle increase in test covariation was previously reported.47,48

ion is associated with poorer cognitive function with increased covariation across cognitive tests, a phenomenon previously observed in aging research.43,44 Intriguingly, forms of cognitive dedifferentiation were also noted in schizophrenia,45,46 where a subtle increase in test covariation was previously reported.47,48 We studied the prospective cognitive trajectories of individuals at UHR for psychosis. We expected to observe the greatest decline in cognitive performance over time among individuals at UHR who converted to psychosis compared with nonconverters and healthy controls. In individuals whose UHR status did not remit during the follow-up period, we expected to observe declining cognitive performance compared with remitters and healthy controls. We hypothesized that increased test covariance would be present as a function of time for individuals whose UHR status did not remit over time. Finally, we examined how changes in cognition as a function of time affected the social and occupational functioning of individuals at UHR for psychosis. Methods Ethics approval for this study was provided by the Singapore National Healthcare Group's Domain Specific Review Board. Written informed consent was obtained from all participants, and consent from a legally acceptable representative was obtained for minors (younger than 21 years) as required by local regulations. This study was conducted from January 1, 2009, to November 11, 2012. Data analysis was conducted from June 2014 to May 2018.

Written informed consent was obtained from all participants, and consent from a legally acceptable representative was obtained for minors (younger than 21 years) as required by local regulations. This study was conducted from January 1, 2009, to November 11, 2012. Data analysis was conducted from June 2014 to May 2018. Participants This study, as part of the Longitudinal Youth at-Risk Study conducted in Singapore,49 included 384 healthy controls and 173 individuals who met the criteria for UHR for psychosis.12 After 24 months, 383 healthy controls (99.7%) and 122 individuals at UHR for psychosis (70.5%) had remained in the study. Participants either were recruited from psychiatric outpatient clinics, educational institutes, and community mental health agencies or were self-referred. Individuals with neurological causes for psychosis, current illicit substance use, or color blindness were excluded. All participants were between 14 and 29 years of age. Their UHR status was ascertained by the Comprehensive Assessment of At-Risk Mental States,12 and their psychiatric history was evaluated with the Structured Clinical Interview for DSM-IV Axis I Disorders.50 Healthy controls did not fulfill UHR criteria, had no psychiatric disorder, and had no family history of psychosis.

eir UHR status was ascertained by the Comprehensive Assessment of At-Risk Mental States,12 and their psychiatric history was evaluated with the Structured Clinical Interview for DSM-IV Axis I Disorders.50 Healthy controls did not fulfill UHR criteria, had no psychiatric disorder, and had no family history of psychosis. Follow-up assessments at 6-month intervals for 2 years or until conversion to psychosis included the Positive and Negative Syndrome Scale,51 Beck Anxiety Inventory,52 Calgary Depression Scale for Schizophrenia,53 and the Social and Occupational Functioning Assessment Scale.54 Remission status was assessed at the 12- and 24-month time points. Individuals at UHR for psychosis were categorized into converters or nonconverters and remitters or nonremitters. Individuals at UHR at baseline but who no longer fulfilled UHR criteria at the 24-month time point were categorized as remitters. Those who met UHR criteria at final assessment or had converted to psychosis were categorized as nonremitters. In subsequent analyses, 2 sets of analysis were carried out involving (1) healthy controls, converters, and nonconverters and (2) healthy controls, remitters, and nonremitters. Details of the sampling methodology and the demographic characteristics of the sample were reported elsewhere.49,55

ategorized as nonremitters. In subsequent analyses, 2 sets of analysis were carried out involving (1) healthy controls, converters, and nonconverters and (2) healthy controls, remitters, and nonremitters. Details of the sampling methodology and the demographic characteristics of the sample were reported elsewhere.49,55 Cognitive Measures The Wechsler Memory Scale-III Spatial Span56; the Brief Assessment of Cognition in Schizophrenia,57 which consists of verbal memory, digit sequencing, token motor task, verbal fluency, symbol coding, and Tower of London tests; the Binocular Depth Inversion task58,59; the Continuous Performance Test—Identical Pairs60; the High-Risk Social Challenge skills interview61; the Babble task62; and the Snakes in the Grass test63 were administered. Cognitive tests were adjusted for age, sex, age × sex, age,2 and age2 × sex via linear regression modeling,64 and standardized residual scores were used for subsequent analysis. Cognitive scores were standardized against healthy control baseline measures. (See the Supplement for eAppendixes 1 and 2 [with eFigure 1], which deal with the concept of testing factor structure changes, and eAppendix 3 for data preprocessing details.)

nd standardized residual scores were used for subsequent analysis. Cognitive scores were standardized against healthy control baseline measures. (See the Supplement for eAppendixes 1 and 2 [with eFigure 1], which deal with the concept of testing factor structure changes, and eAppendix 3 for data preprocessing details.) Statistical Analysis Ordinal logistic regression was conducted to examine between-groups baseline cognitive differences. Univariate models that were P < .05 were selected for subsequent analysis. Linear mixed models were carried out to examine cognitive changes, allowing the inclusion of all longitudinal data available for each participant and the examination of the association of maturational stage with age-related trajectory changes over time. Stuart-Maxwell Marginal Homogeneity test was used to examine the divergence of the estimated test score distributions between the baseline and the 24-month follow-up for each group; these distributions were Bonferroni corrected. A principal components analysis (PCA) was conducted on baseline and 24-month cognitive batteries to investigate the cognitive component structure changes. Component loading vectors were compared via the Kolmogorov-Smirnov test (see eAppendixes 1 and 2 in the Supplement), and the comparisons were Bonferroni corrected. Cognitive components scores were compared using 1-way and repeated measures analysis of variance to examine group-by-time interactions. Repeated-measures general linear models were used to investigate the association of cognitive component changes with the rate of functioning changes. Bonferroni corrections for all intergroup comparisons were completed; further details of the data analysis are reported in eAppendix 3 in the Supplement. Analyses were conducted using SPSS, version 22.0 (IBM), unless otherwise noted.

e the association of cognitive component changes with the rate of functioning changes. Bonferroni corrections for all intergroup comparisons were completed; further details of the data analysis are reported in eAppendix 3 in the Supplement. Analyses were conducted using SPSS, version 22.0 (IBM), unless otherwise noted. Results Demographics Prospectively we evaluated 384 healthy controls (of whom 153 [39.8%] were female and 231 [60.2%] were male with a mean [SD] age of 21.69 [3.26] years) and 173 individuals at UHR for psychosis (of whom 56 [32.4%] were female and 117 [67.6%] were male with a mean [SD] age of 21.27 [3.52] years) who were between 14 and 29 years of age. Individuals at UHR for psychosis were further studied according to their conversion status (17 converters, of whom 3 [17.6%] were female with a mean [SD] age of 20.41 [3.18] years; 156 nonconverters, of whom 53 [34.4%] were female with a mean [SD] age of 21.37 [3.55] years) and remission status (84 remitters, of whom 28 [33.3%] were female with a mean [SD] age of 21.15 [3.41] years; 89 nonremitters, of whom 28 [31.5%] were female with a mean [SD] age of 21.38 [3.64] years). Further demographic characteristics are reported in the Table.

re female with a mean [SD] age of 21.37 [3.55] years) and remission status (84 remitters, of whom 28 [33.3%] were female with a mean [SD] age of 21.15 [3.41] years; 89 nonremitters, of whom 28 [31.5%] were female with a mean [SD] age of 21.38 [3.64] years). Further demographic characteristics are reported in the Table. Table. Baseline Demographics Across Groups Variable Healthy Controls Nonconverters Convertersa Remitters Nonremittersa No. Mean (SD) No. Mean (SD) No. Mean (SD) No. Mean (SD) No. Mean (SD) Age, y 384 21.69 (3.26) 156 21.37 (3.55) 17 20.41 (3.18) 84 21.15 (3.41) 89 21.38 (3.64) Female, % 153 39.8 53 34.4 3 17.6 28 33.3 28 31.5 Male, % 231 60.2 101 65.6 14 82.4 56 66.7 61 68.5 CAARMS total score 383 1.77 (3.67) 154 24.55 (15.57) 17 24.71 (11.09) 84 23.76 (14.79) 89 25.12 (15.43) CDSS composite score NA NA 148 5.68 (4.86) 17 6.76 (5.97) 84 5.15 (4.62) 83 6.42 (5.21) BAI composite score NA NA 146 19.97 (13.29) 17 23.65 (14.32) 82 18.57 (12.89) 83 21.78 (13.81) PANSS total score NA NA 149 48.24 (11.62) 17 50.94 (12.90) 84 46.87 (11.44) 84 50.05 (11.78) PANSS positive score NA NA 149 10.68 (2.79) 17 11.29 (3.06) 84 10.49 (2.75) 84 11.01 (2.84) PANSS negative score NA NA 149 12.15 (4.24) 17 13.00 (3.61) 84 11.89 (4.37) 84 12.49 (3.97) PANSS general psychopathology NA NA 149 25.41 (6.97) 17 26.65 (7.75) 84 24.49 (6.40) 84 26.55 (7.44) Abbreviations: BAI, Beck Anxiety Inventory; CAARMS, Comprehensive Assessment of At-Risk Mental States; CDSS, Calgary Depression Scale for Schizophrenia; NA, not applicable; PANSS, Positive and Negative Syndrome Scale.

) PANSS general psychopathology NA NA 149 25.41 (6.97) 17 26.65 (7.75) 84 24.49 (6.40) 84 26.55 (7.44) Abbreviations: BAI, Beck Anxiety Inventory; CAARMS, Comprehensive Assessment of At-Risk Mental States; CDSS, Calgary Depression Scale for Schizophrenia; NA, not applicable; PANSS, Positive and Negative Syndrome Scale. a Converters are also part of the nonremitters. No statistically significant differences in sex proportions were found across healthy controls, nonconverters, and converters (χ21 = 3.74; P = .05) as well as healthy controls, remitters, and nonremitters (χ21 = 3.74; P = .05). No statistically significant differences in age were observed among healthy controls, nonconverters, and converters (F = 1.53; P = .22) and healthy controls, remitters, and nonremitters (F = 1.02; P = .36). Statistically significant higher Comprehensive Assessment of At-Risk Mental State scores were observed in individuals at UHR for psychosis compared with healthy controls (F = 766.74; P < .001; η2 = 0.581). No notable differences were observed in the Positive and Negative Syndrome Scale, Calgary Depression Scale for Schizophrenia, and Beck Anxiety Inventory measures across groups (Table).

ental State scores were observed in individuals at UHR for psychosis compared with healthy controls (F = 766.74; P < .001; η2 = 0.581). No notable differences were observed in the Positive and Negative Syndrome Scale, Calgary Depression Scale for Schizophrenia, and Beck Anxiety Inventory measures across groups (Table). Baseline Group Differences: Ordinal Logistic Regression Baseline cognitive profiles of all tests are reported in eFigure 2 in the Supplement. Statistically significant between-group differences were found in verbal memory, digit sequencing, token motor task, verbal fluency, symbol coding, and Tower of London tests; the Wechsler Memory Scale-III Spatial Span; the High-Risk Social Challenge skills interview; the Snakes in the Grass test, and the Continuous Performance Test—Identical Pairs across groups (eTable 1 in the Supplement). Post hoc independent, unpaired, 2-tailed t tests revealed differences among healthy controls, nonconverters; healthy controls, converters; and healthy controls, nonremitters (eTable 1 in the Supplement). Baseline cognitive deficits were associated with psychosis conversion (mean odds ratio [OR], 1.66; combined 95% CI, 1.08-2.83; P = .04) and nonremission of UHR status (mean OR, 1.67; combined 95% CI, 1.09-2.95; P = .04).

ers; healthy controls, converters; and healthy controls, nonremitters (eTable 1 in the Supplement). Baseline cognitive deficits were associated with psychosis conversion (mean odds ratio [OR], 1.66; combined 95% CI, 1.08-2.83; P = .04) and nonremission of UHR status (mean OR, 1.67; combined 95% CI, 1.09-2.95; P = .04). Cognitive Changes: Linear Mixed Models Verbal memory, digit sequencing, token motor task, and symbol coding tests; the High-Risk Social Challenge skills interview; the Snakes in the Grass test; and the Continuous Performance Test—Identical Pairs showed longitudinal changes across all groups. Nonlinear changes in cognitive trajectories were also found (Figure 1; eTable 2 in the Supplement). Group-level differences were expected, but few group-by-time interactions were observed across cognitive tests (eTable 2 in the Supplement). The distributions of cognitive linear mixed models estimated scores were different at baseline and 24-month follow-up for all groups (eTable 6 in the Supplement). Increasing effect sizes across groups suggested that cognitive trajectories in nonremitters and converters were most divergent, pointing to subtle underlying perturbations of test covariation. Maturation stage (median split of age at baseline) and age-related trajectory changes were unremarkable. Statistically significant model changes were mostly due to the variability within the clinical groups rather than by observed group differentials (eFigures 3-9 and eTables 3-5 in the Supplement).

f test covariation. Maturation stage (median split of age at baseline) and age-related trajectory changes were unremarkable. Statistically significant model changes were mostly due to the variability within the clinical groups rather than by observed group differentials (eFigures 3-9 and eTables 3-5 in the Supplement). Figure 1. Cognitive Trajectories of Individual Tests Over 24-Month Follow-up A, Healthy controls. B, Remitters. C, Nonconverters. D, Nonremitters. E, Converters. Each test is color-coded. Individual lines reflect estimated cognitive scores for each test computed on the basis of linear mixed model outputs for each test. Babble indicates Babble Task; BACS, Brief Assessment of Cognition for Schizophrenia; BDIT, Binocular Depth Inversion Task; CPT, Continuous Performance Task: 2d, 2Digit, 3d, 3Digit, 4d, 4 Digit subtasks; HISOC, High Risk Social Challenge; SG, Snakes in the Grass test; Acc, Accuracy; rt, reaction time; and WMSIIIss, Wechsler Memory Scale – III Spatial Span.

rief Assessment of Cognition for Schizophrenia; BDIT, Binocular Depth Inversion Task; CPT, Continuous Performance Task: 2d, 2Digit, 3d, 3Digit, 4d, 4 Digit subtasks; HISOC, High Risk Social Challenge; SG, Snakes in the Grass test; Acc, Accuracy; rt, reaction time; and WMSIIIss, Wechsler Memory Scale – III Spatial Span. Psychometric Architecture of Cognitive Constructs: Principal Components Analysis Twenty cognitive subtests with nominally significant (P < .05) baseline group differences were selected for PCA. Five orthogonal principal components were extracted using the Kaiser criterion65 (λ >1). Social cognition, attention, verbal fluency, general cognitive function (GCF), and perception were the 5 principal components that explained variances of 63.3% (healthy control), 74.1% (remitter), and 71.2% (nonremitter) at baseline and variances of 62.8% (healthy control), 75.7% (remitter), and 84.4% (nonremitter) at 24-month follow-up. The reliability of cognitive measures was comparable at baseline (overall α = .831; healthy control α = .792; remitter α = .845; nonremitter α = .809) and at 24-month follow-up (overall α = .848; healthy control α = .818; remitter α = .863; nonremitter α = .900). Component loadings by group and follow-up are represented in Figure 2.

y of cognitive measures was comparable at baseline (overall α = .831; healthy control α = .792; remitter α = .845; nonremitter α = .809) and at 24-month follow-up (overall α = .848; healthy control α = .818; remitter α = .863; nonremitter α = .900). Component loadings by group and follow-up are represented in Figure 2. Figure 2. Component Loading Plots for Baseline and 24-Month Follow-up by Healthy Controls, Remitters, and Nonremitters A-C, Component loading plots for baseline. D-F, Component loading plots for 24-month follow-up. BACS indicates Brief Assessment of Cognition in Schizophrenia (ds, digit sequencing; sc, symbol coding; sfa, verbal fluency—animals; sff, verbal fluency—fruits; sfv, verbal fluency—vegetables; tmt, token motor task; tol, Tower of London; vf, verbal fluency; vm, verbal memory); Continuous Performance Task: 2d, 2Digit, 3d, 3Digit, 4d, 4 Digit subtasks; GCF, general cognitive function; HISOC, High-Risk Social Challenge; SG, Snakes in the Grass (Distractrt, distractor reaction time; Targetrt, target reaction time); WMSIIIss, Wechsler Memory Scale-III Spatial Span.

al fluency; vm, verbal memory); Continuous Performance Task: 2d, 2Digit, 3d, 3Digit, 4d, 4 Digit subtasks; GCF, general cognitive function; HISOC, High-Risk Social Challenge; SG, Snakes in the Grass (Distractrt, distractor reaction time; Targetrt, target reaction time); WMSIIIss, Wechsler Memory Scale-III Spatial Span. Component load differences in GCF were noted among healthy controls, remitters, and nonremitters. Stark differences in component load between GCF baseline and 24-month follow-up in nonremitters were observed. Longitudinal changes for the component loadings for GCF in nonremitters were also observed (maximum vertical deviation = 0.59; χ2 = 8.03; P = .01). The observation is supported by results of the Kolmogorov-Smirnov tests that examined component load vectors across the PCA output (eTable 7 in the Supplement). Different load profiles were present among healthy controls, remitters, and nonremitters at baseline and at 24-month follow-up for the perception component, where subtler trends in social cognition load changes appeared in nonremitters but failed to survive the Bonferroni correction (eTable 7 in the Supplement).

ment). Different load profiles were present among healthy controls, remitters, and nonremitters at baseline and at 24-month follow-up for the perception component, where subtler trends in social cognition load changes appeared in nonremitters but failed to survive the Bonferroni correction (eTable 7 in the Supplement). Longitudinal Change in Cognitive Constructs Weighted and nonweighted cognitive component scores were computed on the basis of the PCA results (eAppendix 3 in the Supplement). Repeated-measures analysis of variance was conducted on the cognitive component scores. Bonferroni-corrected α level of .025 was used to handle 2 test sets that evaluated the same hypothesis. Longitudinal changes were found for attention, GCF, and perception (Figure 3; eTable 8 in the Supplement). Post hoc Bonferroni-adjusted paired sample 2-tailed t tests showed improved performance in remitters, which accounted for the overall model effects. In nonweighted component scores, only GCF was found to be significant. Between-participant analysis of variance tests at both baseline and follow-up further confirmed that, although remitters appeared more similar to nonremitters at baseline for social cognition, attention, and GCF, their performance improved spontaneously with time, and remitters were more similar to healthy controls at 24-month follow-up (Figure 3; eTable 8 in the Supplement).

eline and follow-up further confirmed that, although remitters appeared more similar to nonremitters at baseline for social cognition, attention, and GCF, their performance improved spontaneously with time, and remitters were more similar to healthy controls at 24-month follow-up (Figure 3; eTable 8 in the Supplement). Figure 3. Cognitive Component Profiles by Group and Longitudinal Time by Group Models A, Cognitive component profiles by group. B-F, Longitudinal time by group. BL indicates baseline; EM, expectation maximization; FU, follow-up; GCF, general cognitive function; NWT, nonweighted; WT, weighted; z, standardized score. Bonferroni-corrected model significance is indicated by superscript notation. aHealthy controls vs remitters. bHealthy controls vs nonremitters. cRemitters vs nonremitters. dRemitters baseline vs remitters follow-up.

Figure 3. Cognitive Component Profiles by Group and Longitudinal Time by Group Models A, Cognitive component profiles by group. B-F, Longitudinal time by group. BL indicates baseline; EM, expectation maximization; FU, follow-up; GCF, general cognitive function; NWT, nonweighted; WT, weighted; z, standardized score. Bonferroni-corrected model significance is indicated by superscript notation. aHealthy controls vs remitters. bHealthy controls vs nonremitters. cRemitters vs nonremitters. dRemitters baseline vs remitters follow-up. Relation to Functioning There was a main association of time with the range of social and occupational functioning at baseline and 24-month follow-up (Figure 4; eTable 9 in the Supplement). A statistically significant group-by-time interaction was observed, suggesting differential rates of change of functioning among healthy controls, remitters, and nonremitters. Group-by-time interaction on GCF (F = 12.23; η2 = 0.047; P < .001) and perception (F = 8.33; η2 = 0.032; P < .001) was present. Change in attention and GCF components appeared to partially mediate change in functioning (eTable 9 in the Supplement). Post hoc models revealed that change in the attention component (F = 5.65; η2 = 0.013; P = .02) partially mediated the spontaneous improvements in functioning in remitters and nonremitters compared with healthy controls (Figure 3E and Figure 4C and D). Change in GCF (F = 7.18; η2 = 0.014; P = .01) fully accounted for a differential rate of change in functioning between remitters and nonremitters. All post hoc comparisons survived Bonferroni correction.

vements in functioning in remitters and nonremitters compared with healthy controls (Figure 3E and Figure 4C and D). Change in GCF (F = 7.18; η2 = 0.014; P = .01) fully accounted for a differential rate of change in functioning between remitters and nonremitters. All post hoc comparisons survived Bonferroni correction. Figure 4. Social and Occupational Functioning and Cognitive Component Changes A, Functioning profiles by time point and group (healthy controls, remitters, and nonremitters). B, Repeated-measures schematics for time-by-group and time-by-cognitive component. Differential rate of change between baseline and follow-up: SOFAS: F = 36.85; P < .001; η2 = 0.130. Attention component: F = 5.65; P = .02; η2 = 0.013. GCF component: F = 7.18; P = .01; η2 = 0.014. C, Post hoc repeated measures. Differential rate of change between baseline and follow-up: SOFAS: F = 90.86; P < .001; η2 = 0.170. Attention component: F = 5.65; P = .02; η2 = 0.013. D, Post hoc repeated measures. Differential rate of change between baseline and follow-up: SOFAS: F = 8.67; P = .003; η2 = 0.020. Attention component: F = 15.83; P < .001; η2 = 0.036. E, Post hoc repeated measures. Differential rate of change between baseline and follow-up: SOFAS: F = 3.24; P = .07; η2 = 0.002. GCF component: F = 7.11; P < .009; η2 = 0.058. % indicates percentage difference between best and worst functioning during assessment time point; ∆, difference between baseline and follow-up; GCF, general cognitive function; and SOFAS, Social and Occupational Functioning Assessment Scale.

F = 3.24; P = .07; η2 = 0.002. GCF component: F = 7.11; P < .009; η2 = 0.058. % indicates percentage difference between best and worst functioning during assessment time point; ∆, difference between baseline and follow-up; GCF, general cognitive function; and SOFAS, Social and Occupational Functioning Assessment Scale. Discussion To our knowledge, this study is the largest prospective single-site report of a case-control sample of individuals at UHR for psychosis. Comparisons between remitters and nonremitters suggested that above baseline cognition, trajectory and component-based analyses can identify psychosis and nonremission from illness. Worsening cognitive function over time may be a prime factor in eventual, if not incipient, psychosis.33,36,66,67 Baseline Differences and Prediction Models The study results are consistent with literature that shows significant cognitive deficits in UHR samples. Participants at UHR for psychosis were differentiated from healthy controls, and converters were differentiated from nonconverters according to baseline cognitive performance. Cognitive modeling results demonstrated statistically significant differences not only among healthy controls, converters, and nonconverters but also between individuals who met UHR criteria at baseline and those whose UHR status remitted, compared with those who had no remission.

ing to baseline cognitive performance. Cognitive modeling results demonstrated statistically significant differences not only among healthy controls, converters, and nonconverters but also between individuals who met UHR criteria at baseline and those whose UHR status remitted, compared with those who had no remission. Prospective Trajectory Changes Longitudinal modeling of cognitive performance revealed that most individuals improved with repeated testing every 6 months in the 24-month follow-up. Statistically significant group differences in trajectories were observed, suggesting that baseline variations in cognitive performance interact differently with time in the different groups. These results were consistent with earlier reports indicating that some individuals at UHR for psychosis display cognitive improvements with time.67 Practice effects, pharmacological effects, and diagnostic heterogeneity67 were alternative explanations for the phenomenon, but the more fine-grained follow-up neuropsychological test data reported here may offer further clarification of the cognitive trajectories of individuals at UHR for psychosis. Gradual increases in variability of test performance over time suggest the possibility that the underlying cognitive architecture may have devolved in converters and nonremitters during follow-up. Thus, measures of dedifferentiation of cognitive components may be 1 of the most powerful factors in later conversion and nonremission in individuals at risk for psychosis. Additional analysis of the maturational stage indicated that, between age 14 and 29 years, the most cognitive trajectory changes could be associated with clinical outcomes. Improvement of cognitive performance over time seems to be associated with age, but differential age-related cognitive trajectories do not appear to be present in groups at UHR for psychosis. Nevertheless, larger samples and wider age ranges might be required to further examine differential maturational profiles.

omes. Improvement of cognitive performance over time seems to be associated with age, but differential age-related cognitive trajectories do not appear to be present in groups at UHR for psychosis. Nevertheless, larger samples and wider age ranges might be required to further examine differential maturational profiles. Cognitive Architecture and Shifts in Component Loadings on Test Performance Instead of maximizing the separation of cognitive components, we extracted them orthogonally to make apparent the cross-loading of cognitive subtests. Comparing PCA loading vectors revealed a significant shift in loading patterns between baseline and follow-up in nonremitters, implying the subtle changes in cognitive architecture over time. To our knowledge, such architectural changes have not been reported in previous studies of individuals at UHR for psychosis. We postulate that examining the prospective differential contribution of cognitive components to test performance could reveal subtle cognitive changes in at-risk states that will help differentiate between remitters and nonremitters. Covariance strength across cognitive test performance has been shown to yield vital insights into brain function in aging research43,44,68 and to be a property of deficit cognition in schizophrenia.45,47,48 Decreasing differentiation of GCF, perception, and social cognition components over time among nonremitters and converters is apparent.

ross cognitive test performance has been shown to yield vital insights into brain function in aging research43,44,68 and to be a property of deficit cognition in schizophrenia.45,47,48 Decreasing differentiation of GCF, perception, and social cognition components over time among nonremitters and converters is apparent. Investigating Cognitive Constructs Cognitive components weighted by differential component loadings revealed more sensitivity with social and occupational functioning, particularly with the attention and GCF components. These findings indicate that incorporating cognitive architecture changes appears to be essential in uncovering subtle but important cognitive fluctuations that are relevant to functioning. Neither perception nor social cognition contributes to the variance in functional change beyond the traditional neuropsychological constructs. The trend related to the lack of clear group separation within the perception component could be attributed to the psychometrics of contributing tests. The Snakes in the Grass test, a visual search paradigm, may reflect the subtler changes in lower-level cognitive processes rather than the more robust separation in more traditional neuropsychological tasks. Nevertheless, the contribution of the perception component to test covariance supports the evidence that more refined cognitive mechanisms continue to be sensitive measures in identifying UHR for psychosis in general.69,70,71,72 Social cognition was the only construct that showed decrement over time in nonremitters beyond the baseline differences between healthy controls and remitters, although the component loading analysis suggested only trend-level dedifferentiation. It confirms that these findings replicate the notion that social cognition is separable from cognition even among individuals at UHR for psychosis.73,74,75 Longer follow-up periods might be necessary to determine whether an association between functioning and social cognition might emerge, similar to those in schizophrenia, as the downward trajectory in nonremitters ensues.76,77 Although speculative, mathematical models of cognitive architecture might be more sensitive than the standard neuropsychological tests to the changing neurobiology associated with emerging psychosis in young people at risk.

erge, similar to those in schizophrenia, as the downward trajectory in nonremitters ensues.76,77 Although speculative, mathematical models of cognitive architecture might be more sensitive than the standard neuropsychological tests to the changing neurobiology associated with emerging psychosis in young people at risk. Cognition improved as a function of time, but the changes in remitters were dramatic. Remitters started at baseline with cognitive profiles that were similar to those of nonremitters, but their performance at follow-up was not different from that of healthy controls. The model that best exemplified this phenomenon included the measures comprising GCF. The correspondent longitudinal transition from dedifferentiation to differentiation of GCF that accounted for the functional recovery in remitters illuminates opportunities for follow-up work. Overall, the results point to the possibility that UHR may not be a stable clinical or cognitive construct and that the deficits observed are transient. Results indicate that cognitive deficits in nonremitters tend to be stable and impaired in nearly all components. Longitudinal changes in cognitive architecture, particularly in remitters, have an association with the social and occupational functioning in young people. The converse could be true as the cognitive architecture continues to be increasingly dedifferentiated in nonremitters. These findings suggest that the prognosis for nonremitters is poor and will require the most clinical attention and remediation in the long term.

e social and occupational functioning in young people. The converse could be true as the cognitive architecture continues to be increasingly dedifferentiated in nonremitters. These findings suggest that the prognosis for nonremitters is poor and will require the most clinical attention and remediation in the long term. Limitations This study has several limitations. First, the conversion rate is low. Of the 173 participants at UHR for psychosis followed-up during a 2-year period, 17 (9.8%) converted to psychosis, which is a lower rate than most other reported conversion rates. We speculate that the reason may be the strict drug laws in Singapore and the structured nature of its society. Low conversion rate precluded more sophisticated analysis on convertors. Second, medication use was not systematically adjusted for in the current analysis. The individuals at UHR for psychosis were not medicated with antipsychotics, although some were taking antidepressants. The association of antidepressants with cognition was found to be weak,78,79 but no differences in anxiety or depressive symptoms among UHR groups were observed. Subsequent studies of psychotropic medications and their various cognitive outcomes in at-risk mental states may be informative. Third, the subsampling between nonremitters and converters presented a challenge. Because of the limited sample sizes, we chose to use 2 analysis subsets, comparing healthy controls with remitters and nonremitters as well as healthy controls with nonconverters and converters. It would be ideal to classify samples as healthy controls, remitters, nonremitters, and converters, which is a necessary consideration for subsequent studies with larger sample sizes. Finally, following up participants at UHR for psychosis for only 24 months, although informative, limited the definition of remitters, nonremitters, and nonconverters. Cognitive dedifferentiation phenomena in nonremitters suggest the likelihood of long-lasting cognitive changes, but a longer prospective study would help clarify the degree to which these changes are detrimental to other aspects of clinical outcomes. Such a study would validate remission status (eg, if these cases slip back to being UHR for psychosis) and elucidate potential biological agent underpinnings responsible for the deficit.

longer prospective study would help clarify the degree to which these changes are detrimental to other aspects of clinical outcomes. Such a study would validate remission status (eg, if these cases slip back to being UHR for psychosis) and elucidate potential biological agent underpinnings responsible for the deficit. Conclusions To our knowledge, to date, this study had 1 of the largest single-site samples of individuals at UHR for psychosis. It replicates findings in the literature that cognition is impaired before the onset of psychosis. Baseline cognitive impairment differentiates nonremitters with more enduring symptomatology from healthy controls and individuals at UHR for psychosis whose UHR status later remits. Although predominantly a trait, cognitive architecture shows subtle changes over time in nonremitting individuals at UHR for psychosis. These cognitive architecture changes are associated with functional outcomes and may herald a conversion to psychosis and a cognitive architecture similar to schizophrenia. Supplement. eAppendix 1. Statistical Approaches for Testing Principal Component Loadings: Data Analysis to Test Cognitive Structure Differences Across Groups Prospectively eAppendix 2. Concept of Differentiation and De-Differentiation of Cognitive Factors Tested via Changes in Factor Loadings: Kolmogorov-Smirnov Test for Approximating De-Differentiation eAppendix 3. Data Analysis Workflow: Discussion of Data-Analytic Strategies Carried Out for the Current Report eTable 1. Ordinal Regression and Post-Hoc Tests for Baseline Cognitive Prediction of Group Membership

eAppendix 2. Concept of Differentiation and De-Differentiation of Cognitive Factors Tested via Changes in Factor Loadings: Kolmogorov-Smirnov Test for Approximating De-Differentiation eAppendix 3. Data Analysis Workflow: Discussion of Data-Analytic Strategies Carried Out for the Current Report eTable 1. Ordinal Regression and Post-Hoc Tests for Baseline Cognitive Prediction of Group Membership eTable 2. Linear Mixed Models Elements for Cognitive Tests eTable 3. Linear Mixed Models Elements for Maturational Stage Investigation eTable 4. Linear Mixed Model Elements for UHR Ascertainment Age and Age-Related Trajectories eTable 5. Post-Hoc Linear Mixed Model for UHR Ascertainment Age and Age-Related Trajectories eTable 6. Marginal Homogeneity Test for Effect Distribution at Baseline and Follow-Up eTable 7. Kolmogorov-Smirnov Tests for Component Loading Comparisons eTable 8. Univariate One-Way ANOVA for Cognitive Components, by Group Comparisons eTable 9. Repeated Measures ANOVA eFigure 1. Distribution of Factor Loadings Across Groups at Baseline and Follow-Up eFigure 2. Baseline Cognitive Profiles by Group eFigure 3. BACS Digit Sequencing: Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-Converters, Converters, Maturational Stage and Time eFigure 4. BACS Semantic Fluency (F): Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-converters, Converters, Maturational Stage and Time eFigure 5. BACS Symbol Coding: Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-Converters, Converters, Maturational Stage and Time

eFigure 3. BACS Digit Sequencing: Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-Converters, Converters, Maturational Stage and Time eFigure 4. BACS Semantic Fluency (F): Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-converters, Converters, Maturational Stage and Time eFigure 5. BACS Symbol Coding: Linear Mixed Model SPSS Path Plots by Healthy Controls, Non-Converters, Converters, Maturational Stage and Time eFigure 6. BACS Verbal Memory: Linear Mixed Model SPSS Path Plots by Healthy Controls, Remitters, Non-Remitters, Maturational Stage and Time eFigure 7. BACS Semantic Fluency (F): Linear Mixed Model SPSS Path Plots by Healthy Controls, Remitters, Non-Remitters, Maturational Stage and Time eFigure 8. Snakes in Grass (Reaction Time): Linear Mixed Model SPSS Path Plots of Predicted Score for Age Dependent Trajectories Across Groups eFigure 9. Snakes in Grass (Accuracy): Linear Mixed Model SPSS Path Plots of Predicted Score for Age Dependent Trajectories Across Groups Click here for additional data file.

Introduction Autism spectrum disorders (ASD) are characterized by impairments in reciprocal social interaction and by repetitive and stereotyped interests and behaviors.1 The autism spectrum is a heterogeneous construct, and its component traits are distributed across the population,2,3 with potentially distinct etiologies4,5 and outcomes. Despite increasing recognition in recent years,6 there are substantial gaps in our understanding of the outcomes of children with ASD as they transition into adulthood. Depression is disabling and is common in children with ASD, but few longitudinal population-based studies have followed the natural history of depression in ASD or its component traits.7 Because family members of children with ASD also have an increased risk of depression,8,9,10 a genetic overlap between ASD and depression is possible. However, depression in family members could also relate to difficulties associated with having a child with greater needs or behavioral difficulties.

component traits.7 Because family members of children with ASD also have an increased risk of depression,8,9,10 a genetic overlap between ASD and depression is possible. However, depression in family members could also relate to difficulties associated with having a child with greater needs or behavioral difficulties. Regardless of a genetic basis, it is possible that there are modifiable factors that could be targeted by interventions to reduce the risk of depression in individuals with autism. In clinical practice, individuals with autism seen with depression often report histories of traumatic experiences, particularly bullying. Bullying is strongly associated with depression, an effect that may endure into adulthood,11 and could thus be important in the association between autism and depression.12 For instance, bullying could be a mediator on the causal pathway between autism and depression. It is also possible that the negative effect of bullying on depression may be amplified in the context of the social impairments in autism. To our knowledge, no longitudinal studies have explored these potential mechanisms.

For instance, bullying could be a mediator on the causal pathway between autism and depression. It is also possible that the negative effect of bullying on depression may be amplified in the context of the social impairments in autism. To our knowledge, no longitudinal studies have explored these potential mechanisms. This study used data from a large population-based cohort in England. Our objectives were to (1) compare trajectories of depressive symptoms from ages 10 to 18 years for children with or without ASD or high scores on autistic trait measures, (2) assess whether children with ASD and autistic traits were at increased risk of depression at age 18 years, (3) explore the role of genetic confounding in these associations, and (4) explore the importance of bullying in any associations.

rs for children with or without ASD or high scores on autistic trait measures, (2) assess whether children with ASD and autistic traits were at increased risk of depression at age 18 years, (3) explore the role of genetic confounding in these associations, and (4) explore the importance of bullying in any associations. Methods Study Cohort The Avon Longitudinal Study of Parents and Children (ALSPAC) is a birth cohort study that enrolled mothers in early pregnancy in Bristol and surrounding areas in 1990 to 1992 in England.13,14 It has detailed information on parents and children, collected prospectively at multiple times during pregnancy and throughout childhood. Data sources include self-report questionnaires, clinical assessments, biological samples, and birth, medical, and educational records. The study website contains details of all the data available in a fully searchable format (http://www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/). We used all available data for each combination of exposure and outcome, and we imputed values for missing covariate data using multiple imputation (details are shown in eFigure 1 in the Supplement). Ethical approval for all data collected in the ALSPAC was obtained from the ALSPAC Ethics and Law Committee and the local research ethics committees. Participants provided written informed consent for all clinic assessments, and consent was implied if questionnaires were returned. Participants were followed up through age 18 years. Data analysis was conducted from January to November 2017.

the ALSPAC Ethics and Law Committee and the local research ethics committees. Participants provided written informed consent for all clinic assessments, and consent was implied if questionnaires were returned. Participants were followed up through age 18 years. Data analysis was conducted from January to November 2017. Ascertainment of Autism and Autistic Traits We identified children with ASD using a multisource approach, including a review of clinical records of all children who had multidisciplinary assessment for a developmental disorder (validated against International Statistical Classification of Diseases, 10th Revision [ICD-10] criteria by a consultant pediatrician15), educational records of special education support provided for ASD, and parental reports of an autism or Asperger syndrome diagnosis.16 The ASD cases have been cross-validated against the ASD trait measures,16,17 and as reported below in the Results section, they are associated with a polygenic risk score (PRS) for autism.

cational records of special education support provided for ASD, and parental reports of an autism or Asperger syndrome diagnosis.16 The ASD cases have been cross-validated against the ASD trait measures,16,17 and as reported below in the Results section, they are associated with a polygenic risk score (PRS) for autism. By age 11 years, the ALSPAC had collected 93 measures related to autistic features.18 Of these, the following 4 individual measures were the strongest predictors of ASD18: the Social Communication Disorders Checklist (SCDC) at 7 years, the coherence subscale of the Children’s Communication Checklist at 9 years, a repetitive behavior scale at 5 years, and the sociability subscale of the Emotionality Activity and Sociability temperament measure at 3 years.16 To define high-risk groups for these autistic traits, we dichotomized individuals closest to the worst 10% of distributions of each ASD trait measure.17 Ascertainment of Depression and Depressive Traits The Short Mood and Feelings Questionnaire (SMFQ),19 designed to measure depressive symptoms in children and adolescents, was administered at 6 time points between ages 10 and 18 years via postal questionnaires or in clinics. It has 13 items relating to low mood during the past 2 weeks, each with scores of 0 to 2. Individual item scores were summed, producing a 0 to 26 score range.20

easure depressive symptoms in children and adolescents, was administered at 6 time points between ages 10 and 18 years via postal questionnaires or in clinics. It has 13 items relating to low mood during the past 2 weeks, each with scores of 0 to 2. Individual item scores were summed, producing a 0 to 26 score range.20 The computerized version of the Clinical Interview Schedule–Revised (CIS-R)21 is a fully structured psychiatric interview widely used in community samples. It was administered at age 18 years to identify individuals with an ICD-10 diagnosis of depression. Potential Confounders We included the following variables in our models: (1) child sex, (2) parity (≤1 child vs ≥2 children), (3) maternal occupational class (manual vs nonmanual), (4) mother’s highest educational attainment, (5) financial problems (occurrence vs nonoccurrence of major financial problems), (6) maternal age at delivery (in years), (7) maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery,22 (8) maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale (EPDS score ≥13),23 and (9) accommodation type (detached house vs semidetached house vs flat). We included these variables because they are associated with both autism and depression, apart from being predictors of attrition in the ALSPAC.

and 8 months) depression measured with the Edinburgh Postnatal Depression Scale (EPDS score ≥13),23 and (9) accommodation type (detached house vs semidetached house vs flat). We included these variables because they are associated with both autism and depression, apart from being predictors of attrition in the ALSPAC. Bullying in Late Childhood and Early Adolescence Relational and overt bullying was assessed as separate yes or no items at ages 8, 10, and 13 years using the modified Bullying and Friendship Interview Schedule.24 We created a latent construct of bullying based on 6 binary measures (relational and overt bullying assessed at ages 8, 10, and 13 years) using factor analysis to identify the common variance in the items. Conceptually, this latent construct represents the tendency of children to be bullied persistently throughout childhood or adolescence and was used for the mediation analysis described below in the Statistical Analysis subsection. We also created a binary variable to capture no vs any overt or relational bullying at any time, which we used for testing interactions described below in the Statistical Analysis subsection.

oughout childhood or adolescence and was used for the mediation analysis described below in the Statistical Analysis subsection. We also created a binary variable to capture no vs any overt or relational bullying at any time, which we used for testing interactions described below in the Statistical Analysis subsection. PRS for Autism We examined potential genetic confounding of associations between ASD and depression using autism PRSs, calculated for genotyped ALSPAC children using summary data from the Psychiatric Genomics Consortium autism discovery genome-wide association study (GWAS) (eMethods 1 in the Supplement).25 We created a set of scores based on single-nucleotide polymorphisms (SNPs) that were associated with an ASD diagnosis at a range of GWAS P value thresholds (.5 to 1e−7) and used PRSs generated using SNPs meeting a 0.05 GWAS P value threshold in our main analysis because it maximally captured autism liability within our sample (eFigure 2 in the Supplement). Statistical Analysis We conducted analyses using Stata/MP (version 14; StataCorp) and Mplus (version 8; Muthén & Muthén). We examined trajectories of depressive symptoms (continuous SMFQ scores) between ages 10 and 18 years among those with or without ASD and each autistic trait using mixed-effects linear growth models. To accommodate individual differences in trends of depressive symptoms with age, we included random effects for intercept and slope coefficients and added quadratic and cubic terms to accommodate potential nonlinear trends.

mong those with or without ASD and each autistic trait using mixed-effects linear growth models. To accommodate individual differences in trends of depressive symptoms with age, we included random effects for intercept and slope coefficients and added quadratic and cubic terms to accommodate potential nonlinear trends. We then used modified Poisson regression to estimate the relative risk (RR) of an ICD-10 depression diagnosis at age 18 years in individuals with ASD and each autistic trait vs those without, with robust 95% CIs.26 We estimated crude risks, followed by adjustment for all potential confounders. We further adjusted these models for the autism PRS in the sample with genetic data. We used path analysis to assess mediation of associations between autistic traits and depression at age 18 years by the experience of being bullied in late childhood or early adolescence using latent constructs of bullying and depression. Details are provided in eMethods 2 and eMethods 3 in the Supplement.

We then used modified Poisson regression to estimate the relative risk (RR) of an ICD-10 depression diagnosis at age 18 years in individuals with ASD and each autistic trait vs those without, with robust 95% CIs.26 We estimated crude risks, followed by adjustment for all potential confounders. We further adjusted these models for the autism PRS in the sample with genetic data. We used path analysis to assess mediation of associations between autistic traits and depression at age 18 years by the experience of being bullied in late childhood or early adolescence using latent constructs of bullying and depression. Details are provided in eMethods 2 and eMethods 3 in the Supplement. Finally, to separate the association of ASD diagnosis or traits with depression within and outside the context of bullying, we created categories representing 4 groups by the presence or absence of ASD or ASD traits and the presence or absence of any experiencing of bullying; we compared trajectories of depressive traits between ages 10 and 18 years using mixed-effects linear growth models as described above in the Statistical Analysis subsection. To statistically test moderating associations of bullying, we compared models that included the ASD and bullying variables with those that included only the ASD variable; we then compared models that included the statistical interaction between the ASD and bullying variables with models that included only the main effects of these variables using likelihood ratio tests.

lying, we compared models that included the ASD and bullying variables with those that included only the ASD variable; we then compared models that included the statistical interaction between the ASD and bullying variables with models that included only the main effects of these variables using likelihood ratio tests. Missing Data Missing data in our trajectory and age 18 years analysis are listed in eTable 4 in the Supplement. We imputed missing data for covariates and outcome using multiple imputation (eFigure 1, eMethods 4 in the Supplement). The availability of extensive auxiliary socioeconomic and clinical data (including 7 measures of depression between ages 10 and 18 years) enabled us to account for factors that may explain attrition, providing support to the missing-at-random assumption.27 We repeated our analyses, estimating average associations across 100 imputed data sets and calculated standard errors using the rule by Rubin.28

g 7 measures of depression between ages 10 and 18 years) enabled us to account for factors that may explain attrition, providing support to the missing-at-random assumption.27 We repeated our analyses, estimating average associations across 100 imputed data sets and calculated standard errors using the rule by Rubin.28 Results The maximum sample available with complete data on exposures, outcomes, and covariates was 6091 for the trajectory analysis (48.8% male) and 3168 for analysis of depression diagnosis at age 18 years (44.4% male) (eFigure 1 in the Supplement). The characteristics of our study sample by the presence of ASD and autistic traits are listed in eTable 1 in the Supplement (with an abridged version in Table 1). Mothers of children scoring highest on all autistic trait measures except sociability had a greater prevalence of screening positive for depression and had higher mean anxiety scores in pregnancy and the early postnatal period than the general population, although this pattern was not observed in children with ASD. Children with ASD and those scoring highest on all the autistic traits had a higher prevalence of depressive symptoms at age 10 years, a pattern that was also observed at other time points, albeit inconsistently (eTable 2 in the Supplement). Children with ASD and those scoring highest on the autistic trait measures had a consistently greater prevalence of overt and relational bullying at ages 8, 10, and 13 years than the comparison population, although the statistical evidence for such differences varied (eTable 3 in the Supplement).

2 in the Supplement). Children with ASD and those scoring highest on the autistic trait measures had a consistently greater prevalence of overt and relational bullying at ages 8, 10, and 13 years than the comparison population, although the statistical evidence for such differences varied (eTable 3 in the Supplement). Table 1. Characteristics of the Cohort by Exposure Statusa Variable Diagnosed ASD (n = 8087)b Social Communication Difficulties (n = 5954)c No Yes P Value Nod Yese P Value No. (%) 7991 (98.8) 96 (1.2) NA 5408 (90.8) 546 (9.2) NA Male sex, No. (%) 4083 (51.1) 79 (82.3) <.001 2680 (49.6) 367 (67.2) <.001 Parity ≤1, No. (%) 6540 (81.8) 83 (86.5) .24 4511 (83.4) 436 (79.9) .03 Maternal nonmanual occupational class, No. (%) 4305 (53.9) 65 (67.7) .007 3131 (57.9) 280 (51.3) .003 Mother’s university degree attainment, No. (%) 1187 (14.9) 20 (20.8) .10 937 (17.3) 83 (15.2) .21 Maternal EPDS score ≥12 in pregnancy, No. (%) 1040 (13.0) 14 (14.6) .64 600 (11.1) 115 (21.1) <.001 Maternal EPDS score ≥12 after birth, No. (%) 1056 (13.2) 11 (11.5) .61 607 (11.2) 131 (24.0) <.001 Financial problems since pregnancy, No. (%) 858 (10.7) 9 (9.4) .67 503 (9.3) 74 (13.6) .001 Maternal age at delivery, mean (SD), y 28.0 (4.5) 29.4 (4.2) .004 28.5 (4.4) 28.1 (4.5) .02 Maternal Crown-Crisp antenatal anxiety score, mean (SD) 4.7 (3.4) 4.8 (3.3) .90 4.5 (3.3) 5.5 (3.7) <.001 Maternal Crown-Crisp postnatal anxiety score, mean (SD) 3.3 (3.2) 3.2 (3.0) .72 3.1 (3.1) 4.4 (3.7) <.001 Abbreviations: ASD, autism spectrum disorders; EPDS, Edinburgh Postnatal Depression Scale; NA, not applicable.

ternal Crown-Crisp antenatal anxiety score, mean (SD) 4.7 (3.4) 4.8 (3.3) .90 4.5 (3.3) 5.5 (3.7) <.001 Maternal Crown-Crisp postnatal anxiety score, mean (SD) 3.3 (3.2) 3.2 (3.0) .72 3.1 (3.1) 4.4 (3.7) <.001 Abbreviations: ASD, autism spectrum disorders; EPDS, Edinburgh Postnatal Depression Scale; NA, not applicable. a A more detailed version of this table is available in eTable 1 in the Supplement. P values for No. (%) are by Pearson χ2 test. P values for mean (SD) are by 2-sided t test. b Estimates based on 8087 observations with complete data on covariates and diagnosed autism. c Estimates based on 5954 observations with complete data on covariates and the Social Communication Disorders Checklist scores. d Child has score in the lower 90 percentiles. e Child has score in the upper decile. The autism PRS was associated with the ASD diagnosis and with measures of social communication and repetitive behavior (eFigure 2 in the Supplement), while being in the top decile of the autism PRS was associated with ASD and all 4 autism trait measures, with the exception of coherence (eFigure 3 in the Supplement). Results were generally consistent when using autism PRSs generated using SNP inclusion P value thresholds exceeding .001. There was no evidence of associations between the autism PRS and depression or bullying variables (eFigure 4 and eFigure 5 in the Supplement).

e exception of coherence (eFigure 3 in the Supplement). Results were generally consistent when using autism PRSs generated using SNP inclusion P value thresholds exceeding .001. There was no evidence of associations between the autism PRS and depression or bullying variables (eFigure 4 and eFigure 5 in the Supplement). Examining trajectories, children with ASD and autistic traits had higher average SMFQ depressive symptom scores than the general population at age 10 years (eg, for social communication 5.55 [95% CI, 5.16-5.95] vs 3.73 [95% CI, 3.61-3.85], for ASD 7.31 [95% CI, 6.22-8.40] vs 3.94 [95% CI, 3.83-4.05], remaining elevated in an upward trajectory until age 18 years (eg, for social communication 7.65 [95% CI, 6.92-8.37] vs 6.50 [95% CI, 6.29-6.71], for ASD 7.66 [95% CI, 5.96-9.35] vs 6.62 [95% CI, 6.43-6.81]) (Figure 1). Most pronounced were differences between those with or without social communication difficulties. Analyses using imputed data sets led to similar but more precise estimates (eFigure 6 in the Supplement).

I, 6.92-8.37] vs 6.50 [95% CI, 6.29-6.71], for ASD 7.66 [95% CI, 5.96-9.35] vs 6.62 [95% CI, 6.43-6.81]) (Figure 1). Most pronounced were differences between those with or without social communication difficulties. Analyses using imputed data sets led to similar but more precise estimates (eFigure 6 in the Supplement). Figure 1. Trajectories of Depressive Symptoms in Individuals With Autism Spectrum Disorders (ASD) and ASD Traits and the Comparison Population Shown are confounder-adjusted mean Short Mood and Feelings Questionnaire (SMFQ) scores between ages 10 and 18 years among those with or without ASD and ASD traits. Fitted means were calculated using xtmixed (Stata/MP, version 14; StataCorp) multilevel regression models with linear, quadratic, and cubic terms for time. Trajectories were adjusted for child sex, parity, maternal occupational class, mother’s highest educational attainment, financial problems, maternal age at delivery, maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery, maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale, and accommodation type. Error bars indicate 95% CIs. A, ASD estimates based on 6091 observations with complete data on autism diagnosis and covariates. B, Social communication estimates based on 5209 observations with complete data on the Social Communication Disorders Checklist scores and covariates. C, Coherence estimates based on 5204 observations with complete data on coherence scores and covariates. D, Repetitive behavior estimates based on 5299 observations with complete data on repetitive behavior scores and covariates. E, Sociability estimates based on 5677 observations with complete data on sociability scores and covariates.

ates based on 5204 observations with complete data on coherence scores and covariates. D, Repetitive behavior estimates based on 5299 observations with complete data on repetitive behavior scores and covariates. E, Sociability estimates based on 5677 observations with complete data on sociability scores and covariates. Children with social communication impairments at age 7 years were at increased risk of a diagnosis of depression at age 18 years (adjusted RR, 1.68; 95% CI, 1.05-2.70). These associations were almost unchanged after adjustment for the autism PRS (Table 2) and were estimated with greater precision in the sample without genetic data (eTable 5 in the Supplement) and after multiple imputation (eTable 6 in the Supplement). No evidence of an association between ASD and a depression diagnosis at age 18 years was observed, although the 95% CIs were wide in our main (adjusted RR, 0.50; 95% CI, 0.08-3.38) and imputed (adjusted RR, 0.80; 95% CI, 0.23-2.81) analyses.

in the Supplement) and after multiple imputation (eTable 6 in the Supplement). No evidence of an association between ASD and a depression diagnosis at age 18 years was observed, although the 95% CIs were wide in our main (adjusted RR, 0.50; 95% CI, 0.08-3.38) and imputed (adjusted RR, 0.80; 95% CI, 0.23-2.81) analyses. Table 2. Risk of Outcome of Diagnosed Depression at Age 18 Years Among Children With Autism or Autistic Traits, Including Adjustment for Autism Polygenic Risk Exposure No.a Crude Estimates Adjusted Estimatesb With Additional Adjustment for Autism Polygenic Riskc RR (95% CI) P Value RR (95% CI) P Value RR (95% CI) P Value ASD 2463 0.47 (0.07-3.24) .44 0.55 (0.09-3.50) .53 0.55 (0.09-3.49) .53 Social communication impairments 2230 1.60 (1.00-2.54) .048 1.68 (1.05-2.70) .03 1.70 (1.06-2.72) .03 Reduced speech coherence 2233 0.73 (0.38-1.41) .35 0.72 (0.37-1.37) .32 0.72 (0.38-1.38) .32 Repetitive behavior 2235 1.17 (0.65-2.10) .61 1.11 (0.61-2.00) .74 1.11 (0.61-2.00) .74 Reduced sociability temperament 2342 0.77 (0.45-1.31) .34 0.84 (0.50-1.42) .52 0.84 (0.50-1.42) .52 Abbreviations: ASD, autism spectrum disorders; RR, relative risk (estimates were calculated using modified Poisson regression). a Number of observations with complete data on exposure, covariates, diagnosis of depression at age 18 years, and autism polygenic risk scores.

Table 2. Risk of Outcome of Diagnosed Depression at Age 18 Years Among Children With Autism or Autistic Traits, Including Adjustment for Autism Polygenic Risk Exposure No.a Crude Estimates Adjusted Estimatesb With Additional Adjustment for Autism Polygenic Riskc RR (95% CI) P Value RR (95% CI) P Value RR (95% CI) P Value ASD 2463 0.47 (0.07-3.24) .44 0.55 (0.09-3.50) .53 0.55 (0.09-3.49) .53 Social communication impairments 2230 1.60 (1.00-2.54) .048 1.68 (1.05-2.70) .03 1.70 (1.06-2.72) .03 Reduced speech coherence 2233 0.73 (0.38-1.41) .35 0.72 (0.37-1.37) .32 0.72 (0.38-1.38) .32 Repetitive behavior 2235 1.17 (0.65-2.10) .61 1.11 (0.61-2.00) .74 1.11 (0.61-2.00) .74 Reduced sociability temperament 2342 0.77 (0.45-1.31) .34 0.84 (0.50-1.42) .52 0.84 (0.50-1.42) .52 Abbreviations: ASD, autism spectrum disorders; RR, relative risk (estimates were calculated using modified Poisson regression). a Number of observations with complete data on exposure, covariates, diagnosis of depression at age 18 years, and autism polygenic risk scores. b Estimates were adjusted for child sex, parity, maternal occupational class, mother’s highest educational attainment, financial problems, maternal age at delivery, maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery, maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale, and accommodation type. c Autism polygenic risk scores based on single-nucleotide polymorphisms associated with ASD at P < .05 in the discovery sample.

b Estimates were adjusted for child sex, parity, maternal occupational class, mother’s highest educational attainment, financial problems, maternal age at delivery, maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery, maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale, and accommodation type. c Autism polygenic risk scores based on single-nucleotide polymorphisms associated with ASD at P < .05 in the discovery sample. Children with ASD and autistic traits who also reported being bullied had the highest depression symptom scores at age 10 years, which remained elevated throughout adolescence (Figure 2). There was statistical evidence that the model that included the ASD and bullying variables explained the data better than one that included only ASD diagnosis (likelihood ratio χ2 = 454.75, P < .001) and that models with interaction terms for ASD and bullying explained the data better than models that included only the main effects of these variables (likelihood ratio χ2 = 5.71, P = .017). These different trajectories were most apparent for children with social communication difficulties and were least apparent for worst scores on sociability temperament. In the absence of bullying, the depressive symptom trajectories of children with or without ASD or autistic traits appeared broadly similar. Analyses using imputed data sets led to similar results (eFigure 7 in the Supplement).

munication difficulties and were least apparent for worst scores on sociability temperament. In the absence of bullying, the depressive symptom trajectories of children with or without ASD or autistic traits appeared broadly similar. Analyses using imputed data sets led to similar results (eFigure 7 in the Supplement). Figure 2. Trajectories of Depressive Symptoms in Children With or Without Autism Spectrum Disorders (ASD) and ASD Traits Within and Outside the Context of Bullying Shown are confounder-adjusted mean Short Mood and Feelings Questionnaire (SMFQ) scores among children with or without ASD and ASD traits and exposed or unexposed to bullying. Fitted means were calculated using xtmixed (Stata/MP, version 14; StataCorp) multilevel regression models with linear, quadratic, and cubic terms for time. Trajectories were adjusted for the same variables as those listed above for Figure 1. Error bars indicate 95% CIs. A, ASD estimates based on 4516 observations with complete data on autism diagnosis, bullying variables, and covariates. B, Social communication estimates based on 4041 observations with complete data on the Social Communication Disorders Checklist scores, bullying variables, and covariates. C, Coherence estimates based on 4070 observations with complete data on coherence scores, bullying variables, and covariates. D, Repetitive behavior estimates based on 4051 observations with complete data on repetitive behavior scores, bullying variables, and covariates. E, Sociability estimates based on 4268 observations with complete data on sociability scores, bullying variables, and covariates.

scores, bullying variables, and covariates. D, Repetitive behavior estimates based on 4051 observations with complete data on repetitive behavior scores, bullying variables, and covariates. E, Sociability estimates based on 4268 observations with complete data on sociability scores, bullying variables, and covariates. Finally, we assessed for mediation of associations between social communication difficulties at 7 years and diagnosed depression at age 18 years by the experience of being bullied in late childhood and early adolescence (Table 3). Both before and after adjustment for potential confounders, there was strong evidence of an indirect pathway leading from social communication difficulties in early childhood to being bullied in late childhood or early adolescence to a depression diagnosis at age 18 years. We estimated that this indirect association accounted for 50.5% (95% CI, 5.5%-95.5%) of the total association of social communication difficulties with risk of depression after accounting for potential confounders. Furthermore, there was no evidence of a direct association of social communication difficulties with depression risk after accounting for the indirect association via bullying. Repeating these analyses using imputed data led to similar results (eTable 7 in the Supplement), with a more precise estimate of the indirect association accounted for by bullying (31.5%; 95% CI, 17.3%-45.7%).

al communication difficulties with depression risk after accounting for the indirect association via bullying. Repeating these analyses using imputed data led to similar results (eTable 7 in the Supplement), with a more precise estimate of the indirect association accounted for by bullying (31.5%; 95% CI, 17.3%-45.7%). Table 3. Association Between Social Communication Impairments at Age 7 Years and a Depression Diagnosis at Age 18 Years, Mediated by the Experience of Being Bullied in Late Childhood or Early Adolescencea,b Structural Parameter Estimates β (SE) P Value β (SE) P Value Association of exposure with mediator 0.205 (0.067) .002 0.195 (0.070) .005 Association of mediator with outcome 0.490 (0.089) <.001 0.509 (0.103) <.001 Association of exposure with outcome 0.054 (0.084) .523 0.097 (0.087) .268 Indirect association 0.101 (0.042) .016 0.099 (0.045) .026 Total association 0.155 (0.076) .040 0.196 (0.080) .014 Proportion of total association mediated, % 65.2 NA 50.5 NA Abbreviations: β, unstandardized regression coefficient; CFI, Confirmatory Fit Index; NA, not applicable; RMSEA, root-mean-square error of approximation; TLI, Tucker-Lewis Index. a Depression was captured as a latent construct by means of 4 continuous measures of fatigue, concentration, sleep symptom score, and depressive symptoms. The experience of being bullied was captured as a latent construct by means of 6 binary scores capturing the child’s relational or overt status at ages 8, 10, and 13 years.

ssion was captured as a latent construct by means of 4 continuous measures of fatigue, concentration, sleep symptom score, and depressive symptoms. The experience of being bullied was captured as a latent construct by means of 6 binary scores capturing the child’s relational or overt status at ages 8, 10, and 13 years. b The model fit statistics for the unadjusted association (2152 observations) were RMSEA = 0.046, CFI = 0.937, and TLI = 0.918. The model fit statistics for the adjusted association (2152 observations) were RMSEA = 0.038, CFI = 0.924, and TLI = 0.903. Exposure-mediator and mediator-outcome associations were adjusted for child sex, mother’s highest educational attainment, maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery, maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale, and accommodation type.

ed for child sex, mother’s highest educational attainment, maternal Crown-Crisp anxiety score at 18 weeks’ gestation and 8 weeks after delivery, maternal antenatal (18 and 32 weeks’ gestation) and postnatal (8 weeks and 8 months) depression measured with the Edinburgh Postnatal Depression Scale, and accommodation type. Discussion In this detailed longitudinal study, we found that children with ASD and those with higher scores on all autistic trait measures had more depressive symptoms at age 10 years than the general population, and these remained elevated in an upward trajectory until age 18 years. Social communication impairments had the strongest association with a depression diagnosis at age 18 years. Findings were robust to adjustment for a range of confounders, including maternal depression and anxiety and the child’s polygenic risk for autism. We found evidence of a substantial role of bullying in contributing to and explaining a higher risk of depression in individuals with ASD and autistic symptoms.

e 18 years. Findings were robust to adjustment for a range of confounders, including maternal depression and anxiety and the child’s polygenic risk for autism. We found evidence of a substantial role of bullying in contributing to and explaining a higher risk of depression in individuals with ASD and autistic symptoms. Previous reports on this topic have been contradictory, with results of some studies29,30 suggesting an improvement in depressive symptoms in children with autism over time and other findings suggesting worsening,31 possibly because of selected and heterogeneous populations and different methods. The lack of a general population group in previous studies makes it difficult to conclude whether the trajectories of depressive symptoms in the autistic population differ from those of the general population,29,30,31,32 a limitation in the literature that our study attempted to address. Our findings suggesting that difficulties in social communication may have stronger associations with future depression than other autistic traits have also been reported for outcomes of suicidal thoughts and behaviors33 and are consistent with the concept of fractionation of component features of the autism spectrum.4 However, although social communication difficulties are an important feature of autism, they may occur independently in the population or within the context of other psychiatric diagnoses. Therefore, the association between social communication difficulties and depression may be important within and outside the context of ASD.

although social communication difficulties are an important feature of autism, they may occur independently in the population or within the context of other psychiatric diagnoses. Therefore, the association between social communication difficulties and depression may be important within and outside the context of ASD. We report a significant contribution of bullying as a potential environmental intermediary between childhood autistic features and later depression. Previous work has shown strong links between the experience of bullying and later depression34,35; although confounding could have a role,36 the association is considered to be at least partially causal.11 In our study, children with social communication impairments were more likely to report being bullied, and the mediation analysis suggests that this explained a substantial proportion of the variance of depression at age 18 years, possibly due to reduced self-esteem or social isolation after the bullying. The risk of depression in children with greater autistic symptoms may also be amplified in the context of bullying because of preexisting underlying vulnerabilities in children with autistic features, such as impaired social skills and decreased ability to adapt to adverse or stressful events, such as being bullied. This could explain the elevated trajectories of depressive symptoms in children with ASD and autistic traits who reported being bullied. In the absence of bullying, these children appeared to follow trajectories of depressive symptoms similar to those of the general population. However, such interactions could simply suggest that bullying and autism sometimes co-occur in causal models of depression, as might be expected for any outcome of multifactorial etiology.37

ce of bullying, these children appeared to follow trajectories of depressive symptoms similar to those of the general population. However, such interactions could simply suggest that bullying and autism sometimes co-occur in causal models of depression, as might be expected for any outcome of multifactorial etiology.37 Although it is impossible to identify the exact nature of the underlying mechanisms, our results highlight the need for further research on the role of bullying in this association and the potential for preventive interventions. Furthermore, other relevant characteristics, including comorbidities with neurodevelopmental conditions (eg, attention-deficit/hyperactivity disorder) and classroom placement could be important in this association within or outside the context of bullying and warrant future study. The main strengths of this study were the population-based design with prospectively collected data and repeated measures of depressive symptoms, reducing the possibility of selection and recall bias and allowing us to model longitudinal trajectories. The rich covariate information enabled us to minimize the possibility of confounding bias.

this study were the population-based design with prospectively collected data and repeated measures of depressive symptoms, reducing the possibility of selection and recall bias and allowing us to model longitudinal trajectories. The rich covariate information enabled us to minimize the possibility of confounding bias. Limitations This study has limitations. Like all cohort studies, there was significant attrition, and we used multiple imputation to limit any potential bias; however, selection bias related to missing data remains a possibility. While the use of PRSs was an advantage, they only capture common variation and were based on a small GWAS, so genetic confounding in the associations is still possible. We had insufficient numbers with an ASD diagnosis also meeting the diagnostic criteria for depression at 18 years, possibly due to selective attrition of individuals with autism with more severe depressive symptoms. This is likely to have led to the imprecise result because of a lack of statistical power. Furthermore, atypical presentations of depression are common in ASD, and our study has the potential for outcome measurement error because we used scales (eg, the CIS-R) that have not been adapted for autism.38 Individuals with ASD may also have difficulties in expressing and communicating their emotions and may not have sufficient verbal skills to express changes in their mood or feelings.38

r study has the potential for outcome measurement error because we used scales (eg, the CIS-R) that have not been adapted for autism.38 Individuals with ASD may also have difficulties in expressing and communicating their emotions and may not have sufficient verbal skills to express changes in their mood or feelings.38 Conclusions Our results suggest that ASD and autistic traits are associated with higher depressive symptom scores by age 10 years, which persist to age 18 years, particularly in the context of bullying. Social communication impairments are important in relation to a later diagnosis of depression, and bullying in adolescence could have an important role in this association. These findings add to the evidence highlighting a higher burden of depression, and also suggest a potentially modifiable pathway, through bullying. However, gaps remain in our understanding of the measurement and phenomenology of depression in individuals with autism, which could be a priority for future research. Further work could also focus on improvements in psychological39 and pharmacological40 management of depression in ASD. Finally, further research into the role of traumatic experiences, such as bullying, and the utility of interventions to reduce bullying or address its adverse effects could have the potential to reduce the burden of depression in this population. Supplement. eMethods 1. Supplemental Methods 1 eMethods 2. Supplemental Methods 2 eMethods 3. Supplemental Methods 3 eMethods 4. Supplemental Methods 4 eTable 1. Characteristics of the Cohort by Exposure Status

Conclusions Our results suggest that ASD and autistic traits are associated with higher depressive symptom scores by age 10 years, which persist to age 18 years, particularly in the context of bullying. Social communication impairments are important in relation to a later diagnosis of depression, and bullying in adolescence could have an important role in this association. These findings add to the evidence highlighting a higher burden of depression, and also suggest a potentially modifiable pathway, through bullying. However, gaps remain in our understanding of the measurement and phenomenology of depression in individuals with autism, which could be a priority for future research. Further work could also focus on improvements in psychological39 and pharmacological40 management of depression in ASD. Finally, further research into the role of traumatic experiences, such as bullying, and the utility of interventions to reduce bullying or address its adverse effects could have the potential to reduce the burden of depression in this population. Supplement. eMethods 1. Supplemental Methods 1 eMethods 2. Supplemental Methods 2 eMethods 3. Supplemental Methods 3 eMethods 4. Supplemental Methods 4 eTable 1. Characteristics of the Cohort by Exposure Status eTable 2. Prevalence of Depression by Exposure Status eTable 3. Prevalence of Bullying Victimization by Exposure Status eTable 4. Characteristics of Those Complete or Missing on Covariates and Depression Diagnosis or SMFQ Scores eTable 5. Risk of a Depression Diagnosis at Age 18 Years Among Children With Autism or Autistic Traits

eTable 2. Prevalence of Depression by Exposure Status eTable 3. Prevalence of Bullying Victimization by Exposure Status eTable 4. Characteristics of Those Complete or Missing on Covariates and Depression Diagnosis or SMFQ Scores eTable 5. Risk of a Depression Diagnosis at Age 18 Years Among Children With Autism or Autistic Traits eTable 6. Risk of Adulthood Depression Among Children With Autism or Autistic Traits (Missing Covariate and Outcome Data Predicted With Multiple Imputation) eTable 7. Association Between SCDC at Age 7 and a Depression Diagnosis at Age 18, Mediated by the Experience of Being Bullied in Late Childhood/Early Adolescence (Missing Covariate and Outcome Data Predicted With Multiple Imputation) eFigure 1. Flow Chart for Analysis Plan eFigure 2. Associations Between Continuous Polygenic Risk Score (PRS) for Autism and Autism Diagnosis and Autistic Traits eFigure 3. Associations Between Dichotomized Measure of the Autism PRS (Top Decile Versus Bottom 90 Percentiles) and Autism Diagnosis and Autistic Traits eFigure 4. Associations Between Continuous Polygenic Risk Score (PRS) for Autism and Diagnosed Depression at Age 18 Years (n=3,378) eFigure 5. Associations Between Continuous Polygenic Risk Score (PRS) for Autism and Bullying/Victimization at Any Time Point During Adolescence (n=5,032) eFigure 6. Confounder-Adjusted Mean MFQ Scores Between Ages 10 and 18 (Missing Covariate and Outcome Data Predicted With Multiple Imputation)

eFigure 4. Associations Between Continuous Polygenic Risk Score (PRS) for Autism and Diagnosed Depression at Age 18 Years (n=3,378) eFigure 5. Associations Between Continuous Polygenic Risk Score (PRS) for Autism and Bullying/Victimization at Any Time Point During Adolescence (n=5,032) eFigure 6. Confounder-Adjusted Mean MFQ Scores Between Ages 10 and 18 (Missing Covariate and Outcome Data Predicted With Multiple Imputation) eFigure 7. Confounder-Adjusted Mean MFQ Scores Among Children With/Without Autism or Autistic Traits and Exposed/Unexposed to Bullying (Missing Covariate and Outcome Data Predicted With Multiple Imputation) Click here for additional data file.

Introduction Anorexia nervosa (AN) is a psychiatric disorder characterized by fear of weight gain and dangerously low body weight.1 Anorexia nervosa primarily affects young girls and young women, and its mortality rate exceeds that of other psychiatric disorders.2 Self-starvation and fear of weight gain despite severe underweight and risk of death have been puzzling, and finding comprehensive brain-based models that could explain the constellation of these behaviors has been difficult.3

g girls and young women, and its mortality rate exceeds that of other psychiatric disorders.2 Self-starvation and fear of weight gain despite severe underweight and risk of death have been puzzling, and finding comprehensive brain-based models that could explain the constellation of these behaviors has been difficult.3 It has been hypothesized that anxious traits are vulnerability factors for AN; specifically, elevated harm avoidance (HA) has been found in individuals with AN.4,5,6 Harm avoidance is a temperament trait characterized by excessive worry and fearfulness that has been associated with poor AN treatment outcome and is higher in the ill state compared with the recovered state.7,8 Harm avoidance was found to be correlated with serotonin neurotransmitter receptor availability and, more recently, dopamine receptor binding in AN after recovery.3,9,10,11,12,13 Dopamine is a learning signal and is important for food approach, and animal models suggest enhanced neuronal dopamine activation following food restriction.14,15 This led to the hypothesis that brain circuits that involve dopamine are important for the pathophysiology of AN.16,17 Unexpected reward receipt and omission have been associated with brain dopamine level, the so-called prediction error (PE) response, and have been studied in AN.18,19 In adults with AN in the ill and recovered states, unexpected or randomly applied sucrose taste stimuli evoked higher insular and striatal responses,20,21 and unexpected omission or receipt of monetary reward in adolescent AN also resulted in heightened responses in those regions.22 One interpretation of such findings may lend itself to a model of brain changes in AN: an enhanced dopamine reward system response is an adaptation to starvation to stimulate motivation to approach food.14 Notably, peak onset of AN is in midadolescence,23 when sensitivity to reward and reward PE response is still developing.24,25,26 Individuals vulnerable to develop AN could be particularly sensitive to food restriction and adaptations of reward response during that developmental period.

te motivation to approach food.14 Notably, peak onset of AN is in midadolescence,23 when sensitivity to reward and reward PE response is still developing.24,25,26 Individuals vulnerable to develop AN could be particularly sensitive to food restriction and adaptations of reward response during that developmental period. In this study, we tested specific hypotheses to integrate PE reward signals with anxiety and core eating disorder signs.17 Specifically, we hypothesized that reward learning PE response would be elevated in AN and associated with HA. Second, we wanted to test the hypothesis that PE response is associated with neurocircuitry that regulates appetite and food intake. Previously, we found a pattern of effective connectivity (direction of activation) from ventral striatum to hypothalamus in adults with AN.27 We interpreted this as a possible mechanism of how the brain could override hunger signals in AN. We hypothesized that adolescents with AN show a similar pattern. Third, food restriction is a stressor and is associated with increased brain cortisol levels.28,29,30 Cortisol level affects dopamine release and postsynaptic dopamine D2 receptors,31,32,33 and we wanted to test whether cortisol level was associated with PE response and core AN behaviors.34,35,36

how a similar pattern. Third, food restriction is a stressor and is associated with increased brain cortisol levels.28,29,30 Cortisol level affects dopamine release and postsynaptic dopamine D2 receptors,31,32,33 and we wanted to test whether cortisol level was associated with PE response and core AN behaviors.34,35,36 Methods Participants Fifty-six female adolescents and young adults with AN (age range, 11-21 years) and 52 healthy matched control participants (age range, 11-21 years) were included in this study (Table 1).37,38,39,40,41 The AN group was recruited from partial hospitalization treatment, where closely supervised meal plans mitigated confounding brain effects of acute starvation or dehydration.42 Treatment involved a highly structured program aimed at weight restoration over 5 weeks, including parent training in meal support according to the family-based treatment model.43 Control participants were recruited through local advertisements. In the AN group, 53 participants were diagnosed as having pure restricting type and 3 as having infrequent purge episodes (less than once a month), and all 56 participants with AN fell below the 10th percentile for body mass index (BMI; calculated as weight in kilograms divided by height in meters squared) for age. All participants underwent functional magnetic resonance imaging (fMRI); individuals with AN were without menses and controls were in the early follicular phase to control for sex hormone effects. Participants 18 years or older, including 19 participants with AN and 11 controls, were administered the Structured Clinical Interview for DSM-5 by a doctoral-level interviewer. Those younger than 18 years completed the Mini-International Neuropsychiatric Interview.44 Participants were right-handed and had no history of head trauma, neurological disease, major medical illness, psychosis, or substance use disorders. Six participants with AN and 11 controls were taking oral contraceptives. Twenty-six participants with AN were taking antidepressants, and 7 were taking atypical antipsychotics. The Colorado Multiple institutional review board approved the study. All participants provided written informed consent.

stance use disorders. Six participants with AN and 11 controls were taking oral contraceptives. Twenty-six participants with AN were taking antidepressants, and 7 were taking atypical antipsychotics. The Colorado Multiple institutional review board approved the study. All participants provided written informed consent. Table 1. Demographic and Behavioral Variables Variable Mean (SD) t P Value Anorexia Nervosa Group (n = 56) Control Group (n = 52) Age, y 16.56 (2.47) 16.01 (2.80) −1.078 .28 BMIa 15.88 (0.86) 20.86 (2.07) 16.125 <.001 Age-adjusted BMI percentile 2.36 (2.63) 58.57 (21.94) 17.081 <.001 Drive for thinness scoreb 19.38 (7.13) 2.13 (3.05) −16.422 <.001 Body dissatisfaction scoreb 24.76 (10.20) 3.62 (4.32) −14.103 <.001 Punishment sensitivity scorec 12.57 (4.02) 5.54 (3.66) −9.481 <.001 Reward sensitivity scorec 7.25 (3.83) 6.65 (3.76) −0.815 .42 State anxiety scored 51.89 (13.85) 28.00 (6.29) −11.590 <.001 Trait anxiety scored 53.05 (13.57) 29.48 (7.05) −11.440 <.001 Harm avoidance scoree 21.98 (7.43) 10.79 (4.80) −9.362 <.001 Reward dependence scoree 14.64 (3.58) 15.77 (3.67) 1.616 .11 Depression scoref 18.16 (9.74) 2.90 (2.85) −10.008 <.001 Breakfast calories 602.857 (145.42) 568.75 (151.12) −1.195 .24 Sucrose pleasantness score 4.32 (2.41) 5.08 (2.56) 1.646 .10 Sucrose sweetness score 8.00 (1.24) 8.06 (1.07) 0.258 .80 Antidepressant use, No. (%) 26 (46.4) NA NA NA Antipsychotic use, No. (%) 7 (12.5) NA NA NA Mood disorder, No. (%) 20 (35.7) NA NA NA Anxiety disorder, No. (%) 28 (50.0) NA NA NA Abbreviations: BMI, body mass index; NA, not applicable.

.41) 5.08 (2.56) 1.646 .10 Sucrose sweetness score 8.00 (1.24) 8.06 (1.07) 0.258 .80 Antidepressant use, No. (%) 26 (46.4) NA NA NA Antipsychotic use, No. (%) 7 (12.5) NA NA NA Mood disorder, No. (%) 20 (35.7) NA NA NA Anxiety disorder, No. (%) 28 (50.0) NA NA NA Abbreviations: BMI, body mass index; NA, not applicable. a Calculated as weight in kilograms divided by height in meters squared. b Eating Disorder Inventory–3.37 c Revised Sensitivity to Punishment and Reward Questionnaire.38 d State-Trait Anxiety Inventory.39 e Temperament and Character Inventory.40 f Children’s Depression Inventory.41 Self-Assessments In addition to diagnostic interviews, participants completed a battery of self-assessments. Participants completed the Eating Disorder Inventory–3,37 Revised Sensitivity to Punishment and Reward Questionnaire,38 State-Trait Anxiety Inventory,39 Temperament and Character Inventory,40 and Children’s Depression Inventory.41

addition to diagnostic interviews, participants completed a battery of self-assessments. Participants completed the Eating Disorder Inventory–3,37 Revised Sensitivity to Punishment and Reward Questionnaire,38 State-Trait Anxiety Inventory,39 Temperament and Character Inventory,40 and Children’s Depression Inventory.41 Brain Imaging Methods fMRI Image Acquisition Between 7:00 am and 9:00 am on the study day, participants with AN ate their meal plan breakfast and controls ate a quality-matched and calorie-matched breakfast (Table 1). Brain imaging was performed between 8:00 am and 9:00 am using the 3T Signa scanner (General Electric Company) or Skyra 3T scanner (Siemens) using the following criteria: 3-plane scout scan (16 seconds), sagittally acquired, spoiled gradient sequence T1-weighted (172 slices; thickness, 1 mm; inversion time, 450 ms; repetition time, 8 ms; echo time, 4 ms; flip angle, 12°; field of view, 22 cm; scan matrix, 64 × 64), and T2-weighted echo planar scans for blood oxygen level–dependent functional activity (3.4 × 3.4 × 2.6-mm voxels; repetition time, 2100 ms; echo time, 30 ms; flip angle, 70°; 28 axial slices; thickness, 2.6 mm; gap, 1.4 mm) (eMethods 1 in the Supplement).

; flip angle, 12°; field of view, 22 cm; scan matrix, 64 × 64), and T2-weighted echo planar scans for blood oxygen level–dependent functional activity (3.4 × 3.4 × 2.6-mm voxels; repetition time, 2100 ms; echo time, 30 ms; flip angle, 70°; 28 axial slices; thickness, 2.6 mm; gap, 1.4 mm) (eMethods 1 in the Supplement). Taste Reward Task The design of this study was adapted from O’Doherty et al19 (eAppendix 1 in the Supplement). Participants learned to associate 3 unconditioned taste stimuli (1 molar sucrose solution, no solution, or artificial saliva) with paired conditioned visual stimuli. Each conditioned visual stimulus was probabilistically associated with its unconditioned taste stimulus such that 20% of sucrose and no solution conditioned visual stimulus trials were unexpectedly followed by no solution and sucrose unconditioned taste stimuli, respectively. Taste stimuli were applied using a customized programmable syringe pump (J-Kem Scientific) and E-Prime software version 2 (Psychological Software Tools).45

t 20% of sucrose and no solution conditioned visual stimulus trials were unexpectedly followed by no solution and sucrose unconditioned taste stimuli, respectively. Taste stimuli were applied using a customized programmable syringe pump (J-Kem Scientific) and E-Prime software version 2 (Psychological Software Tools).45 fMRI Analysis Image preprocessing and analysis were performed using statistical parametric mapping version 12 (Wellcome Trust Centre for Neuroimaging). Images were realigned to the first volume, normalized to the Montreal Neurological Institute template, and smoothed at 6-mm full width at half maximum gaussian kernel. Data were preprocessed with slice time correction and modeled with a hemodynamic response convolved function using the general linear model, including temporal and dispersion derivatives. A 128-second high-pass filter was applied for low-frequency blood oxygen level–dependent signal fluctuations and motion parameters as first-level analysis regressors.

time correction and modeled with a hemodynamic response convolved function using the general linear model, including temporal and dispersion derivatives. A 128-second high-pass filter was applied for low-frequency blood oxygen level–dependent signal fluctuations and motion parameters as first-level analysis regressors. PE Analysis Each participant’s PE signal was modeled based on trial sequence (absolute of positive and negative PE) and regressed with brain activation across all trials19,21,22 (eMethods 2 in the Supplement). We extracted mean parameter estimates across all voxels from 18 predefined anatomical regions of interest (ROIs) based on previous studies,22 including the bilateral dorsal anterior insula, ventral anterior insula, caudate head, orbitofrontal cortex (OFC) gyrus rectus, medial OFC, middle OFC, inferior OFC, ventral striatum,46 and nucleus accumbens47 (http://marsbar.sourceforge.net/; automated anatomical labeling Atlas48).

(ROIs) based on previous studies,22 including the bilateral dorsal anterior insula, ventral anterior insula, caudate head, orbitofrontal cortex (OFC) gyrus rectus, medial OFC, middle OFC, inferior OFC, ventral striatum,46 and nucleus accumbens47 (http://marsbar.sourceforge.net/; automated anatomical labeling Atlas48). Effective Connectivity Analysis We extracted ROI functional activation for trials of expected receipt of 1 molar sucrose solution (n = 80), with conditioned visual stimuli and unconditioned taste stimuli trial length of 6 seconds, as previously studied in adults.27 The Tetrad-V program49 was used to infer effective connectivity with independent multisample greedy equivalence search and linear nongaussian orientation, fixed structure search algorithms. This analysis aimed to understand causal associations among neuronal populations whose activity gives rise to observed fMRI signals in spatially localized ROIs27 (eMethods 3 in the Supplement). We extracted edge coefficients for ventral striatum–hypothalamus connectivity to test for correlations with behavior or PE values.

analysis aimed to understand causal associations among neuronal populations whose activity gives rise to observed fMRI signals in spatially localized ROIs27 (eMethods 3 in the Supplement). We extracted edge coefficients for ventral striatum–hypothalamus connectivity to test for correlations with behavior or PE values. Cortisol Collection and Analysis On the scan day, a subset of 20 participants with AN and 25 controls provided 0.5-mL samples of saliva (passive drool) in 5-mL Screw Cap Micro Tubes (Thermo Fisher Scientific; eAppendix 2 and eTable 1 in the Supplement). Samples were collected 30 minutes prior to breakfast, 30 minutes after breakfast, and right before brain imaging. Samples were stored at −15° C until analysis. Cortisol was assayed using commercial immunoassays (Salimetrics). The area under the curve (AUC; trapezoid method with 3 time points) was calculated and correlated with ROI PE response and across the whole brain (family-wise error rate [FWE] P < .05; eAppendix 2 in the Supplement).

were stored at −15° C until analysis. Cortisol was assayed using commercial immunoassays (Salimetrics). The area under the curve (AUC; trapezoid method with 3 time points) was calculated and correlated with ROI PE response and across the whole brain (family-wise error rate [FWE] P < .05; eAppendix 2 in the Supplement). Statistical Analysis SPSS Statistics 25 (IBM) was used for statistical analyses. Demographic and behavior data were analyzed using t test. Extracted regional brain activation parameter estimates were tested for normality with the Shapiro-Wilk test, rank-transformed when nonnormally distributed, and analyzed using multivariate analysis of variance and multivariate analysis of covariance with covariates to account for confounding factors, such as comorbidity or medication, as in previous studies.22 Spearman rank was used for correlation analyses and controlled for multiple comparisons using bootstrapping procedures (1000 samples).50 All P values were 2-tailed, and a P value less than .05 was considered significant. Results were corrected for multiple comparison. Results Demographic and Behavioral Data There were no significant group differences for age, breakfast calories, or sucrose pleasantness or sweetness ratings (Table 1). Participants with AN had significantly lower age-adjusted BMI percentile and novelty-seeking scores. Participants with AN also had elevated drive for thinness and body dissatisfaction and significantly higher HA, punishment sensitivity, state and trait anxiety, and depression scores.

ess or sweetness ratings (Table 1). Participants with AN had significantly lower age-adjusted BMI percentile and novelty-seeking scores. Participants with AN also had elevated drive for thinness and body dissatisfaction and significantly higher HA, punishment sensitivity, state and trait anxiety, and depression scores. Brain Imaging Results PE ROI Analysis Results were nonnormally distributed. Multivariate analysis of variance (no covariates) resulted in a Wilks λ of 0.642 (P < .001; partial η2 = 0.358), with associations with bilateral caudate head, ventral striatum, nucleus accumbens, right inferior OFC, right medial OFC, right gyrus rectus, right dorsal anterior insula, and right ventral anterior insula surviving Bonferroni correction (Table 2). Multivariate analysis of covariance (with age, scanner, antidepressant use, antipsychotic use, comorbid depression, and comorbid anxiety as covariates) resulted in a Wilks λ of 0.707 (P = .02; partial η2 = 0.296), with associations with right and left caudate head, right and left nucleus accumbens, and right ventral anterior insula surviving Bonferroni correction. There were no significant differences between scanners for within-group ROI comparisons, nor did results change when the scanner covariate was removed (eFigure 1 in the Supplement).

ons with right and left caudate head, right and left nucleus accumbens, and right ventral anterior insula surviving Bonferroni correction. There were no significant differences between scanners for within-group ROI comparisons, nor did results change when the scanner covariate was removed (eFigure 1 in the Supplement). Table 2. Parameter Estimate Analyses Across Groupsa Region of Interest Response, Mean (SD) MANOVA MANCOVAb Anorexia Nervosa Group (n = 56) Control Group (n = 52) F P Valuec ηp2 F P Valuec ηp2 Right caudate head 67.179 (30.158) 40.846 (26.660) 22.972 <.001 0.178 8.102 .005 0.075 Left caudate head 67.786 (30.206) 40.192 (25.917) 25.772 <.001 0.196 13.004 <.001 0.115 Right ventral striatum 57.232 (33.889) 45.134 (27.793) 8.646 .004 0.075 2.904 .09 0.028 Left ventral striatum 62.750 (32.410 45.615 (27.752) 9.695 .002 0.084 3.048 .08 0.030 Right nucleus accumbens 66.196 (30.972) 41.904 (26.678) 18.939 <.001 0.152 8.143 .005 0.075 Left nucleus accumbens 65.857 (30.064) 42.269 (28.094) 17.676 <.001 0.143 4.878 .03 0.047 Right inferior orbitofrontal cortex 61.804 (32.709) 46.635 (27.977) 6.659 .01 0.059 2.925 .09 0.028 Left inferior orbitofrontal cortex 59.161 (32.977) 49.481 (28.911) 2.614 .11 0.024 0.873 .35 0.009 Right medial orbitofrontal cortex 60.411 (33.670) 48.135 (27.494) 4.269 .04 0.039 1.149 .29 0.011 Left medial orbitofrontal cortex 59.804 (32.777) 48.788 (28.904) 3.410 .07 0.031 1.123 .29 0.011 Right middle orbitofrontal cortex 58.268 (32.302) 50.442 (30.009) 1.694 .20 0.016 0.089 .77 0.001 Left middle orbitofrontal cortex 57.661 (32.559) 51.096 (29.869) 1.186 .28 0.011 0.004 .95 0.000 Right gyrus rectus 59.393 (32.503) 49.231 (29.399) 5.953 .02 0.053 0.361 .55 0.004 Left gyrus rectus 57.607 (32.217) 51.154 (30.279) 1.152 .29 0.011 0.011 .92 0.0001 Right dorsal anterior insula 60.339 (31.840) 48.212 (29.784) 4.162 .04 0.038 2.053 .16 0.020 Left dorsal anterior insula 58.732 (32.410) 49.942 (29.742) 2.146 .15 0.020 0.256 .61 0.003 Right ventral anterior insula 62.607 (32.984) 45.769 (27.112) 8.326 .005 0.073 6.773 .01 0.063 Left ventral anterior insula 57.232 (33.889) 51.558 (28.331) 0.884 .35 0.008 0.033 .86 0.0001 Abbreviations: MANCOVA, multivariate analysis of covariance; MANOVA, multivariate analysis of variance.

6 .15 0.020 0.256 .61 0.003 Right ventral anterior insula 62.607 (32.984) 45.769 (27.112) 8.326 .005 0.073 6.773 .01 0.063 Left ventral anterior insula 57.232 (33.889) 51.558 (28.331) 0.884 .35 0.008 0.033 .86 0.0001 Abbreviations: MANCOVA, multivariate analysis of covariance; MANOVA, multivariate analysis of variance. a Data were rank-transformed. b Multivariate analysis of covariance included age, scanner, antidepressant use, antipsychotic use, comorbid depression, and comorbid anxiety. c P values are adjusted for Bonferroni multiple comparisons. Rank values remained the same for both analyses. Effective Connectivity Sucrose anticipation and receipt elicited patterns of connectivity that were similar for 50% of identified connections across groups. However, for our effective connectivity of interest, there were different patterns between groups bilaterally; the hypothalamus in controls directed activation to the ventral striatum, whereas in participants with AN, the ventral striatum directed effective connectivity to the hypothalamus (Figure 1). Figure 1. Effective Connectivity of Sucrose Receipt The effective connectivity pattern in healthy controls (A) from the right hypothalamus to the ventral striatum is the opposite pattern seen in participants with anorexia nervosa (AN) (B). The connectivity patterns indicate directionality within group but do not reflect a direct group contrast. The blue arrows indicate patterns unique to controls; the green arrows, patterns unique to participants with AN. OFC indicates orbitofrontal cortex.

ite pattern seen in participants with anorexia nervosa (AN) (B). The connectivity patterns indicate directionality within group but do not reflect a direct group contrast. The blue arrows indicate patterns unique to controls; the green arrows, patterns unique to participants with AN. OFC indicates orbitofrontal cortex. Correlation Analyses Rate of age-adjusted change of BMI percentile to reach target weight (BMI percentile change mean [SD] of 20.37 [15.12] over a mean [SD] time of 40.01 [10.94] days = 0.51 age-adjusted BMI percentile change per day) was negatively correlated with response in the inferior OFC (right ρ, −0.389; 95% CI, −0.612 to −0.100; P < .003), middle OFC (right ρ, −0.281; 95% CI, −0.522 to −0.018; P < .04), gyrus rectus (right ρ, −0.282; 95% CI, −0.534 to −0.014; P < .04; left ρ, −0.268; 95% CI, −0.509 to −0.018; P < .045), dorsal anterior insula (right ρ, −0.358; 95% CI, −0.590 to −0.092; P < .007; left ρ, −0.281; 95% CI, −0.542 to −0.004; P < .04), and ventral anterior insula (right ρ, −0.274; 95% CI, −0.512 to −0.016; P < .04). Prediction error regression weights did not significantly correlate with admission age-adjusted BMI percentiles. In participants with AN, HA was positively correlated with response in OFC gyrus rectus (right ρ, 0.317; 95% CI, 0.091 to 0.539; P < .02; left ρ, 0.336; 95% CI, 0.112 to 0.550; P < .01). Harm avoidance in participants with AN was positively correlated with drive for thinness (ρ, 0.381; 95% CI, 0.140 to 0.587; P < .004) and body dissatisfaction (ρ, 0.312; 95% CI, 0.055 to 0.541; P < .02) (eFigure 2 in the Supplement).

317; 95% CI, 0.091 to 0.539; P < .02; left ρ, 0.336; 95% CI, 0.112 to 0.550; P < .01). Harm avoidance in participants with AN was positively correlated with drive for thinness (ρ, 0.381; 95% CI, 0.140 to 0.587; P < .004) and body dissatisfaction (ρ, 0.312; 95% CI, 0.055 to 0.541; P < .02) (eFigure 2 in the Supplement). Sucrose pleasantness was negatively correlated with PE in all regions studied (ρ range, −0.451 to −0.249) in participants with AN but only in the middle OFC in controls. A comparison of regression slopes showed significantly different slopes between groups in the caudate head (right Fisher z, 2.103; P < .04), medial OFC (left Fisher z, 2.204; P < .03), and nucleus accumbens (right Fisher z, 1.958; P < .050; left Fisher z, 2.293; P < .02) (Figure 2) (eTable 2 in the Supplement). In the AN group, taste pleasantness was negatively correlated with HA (ρ, −0.294; 95% CI, −0.041 to −0.510; P < .03). In participants with AN, ventral striatum–hypothalamus edge coefficients were correlated with ipsilateral inferior OFC PE (right ρ, 0.318; 95% CI, 0.063 to 0.547; P < .02; left ρ, 0.354; 95% CI, 0.059 to 0.606; P < .007), middle OFC PE (right ρ, 0.308; 95% CI, 0.044 to 0.554; P < .02; left ρ, 0.427; 95% CI, 0.142 to 0.646; P < .001), and dorsal anterior insula PE (right ρ, 0.392; 95% CI, 0.170 to 0.602; P < .003).

ilateral inferior OFC PE (right ρ, 0.318; 95% CI, 0.063 to 0.547; P < .02; left ρ, 0.354; 95% CI, 0.059 to 0.606; P < .007), middle OFC PE (right ρ, 0.308; 95% CI, 0.044 to 0.554; P < .02; left ρ, 0.427; 95% CI, 0.142 to 0.646; P < .001), and dorsal anterior insula PE (right ρ, 0.392; 95% CI, 0.170 to 0.602; P < .003). Figure 2. Correlation of Prediction Error (PE) With Taste Pleasantness Regional PE results across groups and their correlation with taste pleasantness ratings for 1 molar sucrose solution. Prediction errors and pleasantness ratings were rank-transformed. See eTable 2 in the Supplement for full correlation results across regions and groups.

of Prediction Error (PE) With Taste Pleasantness Regional PE results across groups and their correlation with taste pleasantness ratings for 1 molar sucrose solution. Prediction errors and pleasantness ratings were rank-transformed. See eTable 2 in the Supplement for full correlation results across regions and groups. Cortisol Analysis The area under the curve for cortisol levels was elevated in participants with AN compared with controls (t29.169 = −2.515; P = .02) (eFigure 3 in the Supplement). In participants with AN, the AUC for cortisol levels was positively correlated with caudate head PE (right ρ, 0.0457; 95% CI, 0.122 to 0.692; P < .04). Whole-brain regression (FWE corrected) showed that cortisol level was significantly positively correlated with PE response in the right superior frontal gyrus (x = −18; y = 58; z = 6) in participants with AN (peak FWE P = .005; κ = 1). Subsequent small volume correction (P < .001; κ = 10) within the anatomical superior frontal gyrus ROI resulted in a significant cluster (peak FWE P < .001; κ = 53) (eAppendix 2 in the Supplement). When parameter estimates were extracted from the FWE-corrected cluster, the strength of the cortisol regression was significantly positively correlated with body dissatisfaction scores in participants with AN (ρ, 0.484; 95% CI, 0.013 to 0.815; P < .03) (eFigure 3 in the Supplement). No significant clusters or behavioral correlations were found in the control group.

e FWE-corrected cluster, the strength of the cortisol regression was significantly positively correlated with body dissatisfaction scores in participants with AN (ρ, 0.484; 95% CI, 0.013 to 0.815; P < .03) (eFigure 3 in the Supplement). No significant clusters or behavioral correlations were found in the control group. Discussion Anorexia nervosa is a perplexing psychiatric illness, and the complex biopsychosocial aspects of the illness have made it difficult to develop brain-based models for food restriction. The results from this study show (1) heightened caudate, nucleus accumbens, and insula taste PE signal in adolescents with AN; (2) ventral striatal–hypothalamic dynamic effective connectivity in participants with AN that was opposite to controls and that was positively correlated with insular and orbitofrontal PE signal; (3) a positive correlation of cortisol level with PE in a subset of participants; and (4) a positive correlation of OFC PE with HA and a negative correlation of orbitofrontal and insula PE with BMI change during treatment, as seen previously.22 The results suggest that the PE signal could be an important marker for weight gain and anxiety in adolescent AN.

level with PE in a subset of participants; and (4) a positive correlation of OFC PE with HA and a negative correlation of orbitofrontal and insula PE with BMI change during treatment, as seen previously.22 The results suggest that the PE signal could be an important marker for weight gain and anxiety in adolescent AN. The first part of this study shows heightened taste reward PE signal across the caudate, nucleus accumbens, and insula in a large sample of adolescents and young adults with AN, comparable with results in adults.21 Starvation is associated with adaptations of the body to drive food intake,51,52,53 including changes in dopamine release and receptor expression.53 We expect that in AN, the elevated PE response, which has been associated with brain dopamine activity, is an adaptation to food restriction and weight loss that normalizes with long-term recovery.54 The second major finding was a pattern of activation during sweet taste anticipation and receipt that was directed from the ventral striatum to the hypothalamus in AN and was positively correlated with OFC and insular PE signal. This was in contrast to the direction of activation in controls from the hypothalamus to the ventral striatum, a connection thought to be particularly important for feeding regulation.55 A dopamine-dependent pathway from the ventral striatum to the hypothalamus has been described that mediates fear.56 This lends itself to the hypothesis that PE signal in AN might activate this circuitry and override appetitive hypothalamic signals.

nnection thought to be particularly important for feeding regulation.55 A dopamine-dependent pathway from the ventral striatum to the hypothalamus has been described that mediates fear.56 This lends itself to the hypothesis that PE signal in AN might activate this circuitry and override appetitive hypothalamic signals. Our analyses also indicate direct associations of PE response with HA, BMI change, and taste perception. Orbitofrontal cortex gyrus rectus PE was negatively correlated with BMI increase during treatment but positively correlated with HA, which in turn was positively correlated with drive for thinness and body dissatisfaction. This indicates that PE values could be directly associated with HA and BMI change in AN when in the ill state, although OFC PE was not elevated in participants with AN compared with controls. Complex behaviors are driven by the balance between neurotransmitter systems57 and imbalance between, for instance, dopamine and serotonin neurotransmission in those with AN could make the PE signal relatively more important in its association with HA. Our data also show that HA is directly correlated with core AN behavior, such as drive for thinness and body dissatisfaction, suggesting that anxiety is an important driver of the cognitive/emotional aspects specific to AN. This study suggests that dopamine circuits via PE signaling could be involved with elevated HA. However, the specific neurotransmitter systems underlying those results need further exploration.9,10,12 Pleasant taste stimulates dopamine release to promote eating and typically activates OFC response.58,59,60 Our data raise the possibility that adolescents with AN in this study were negatively conditioned to sweet taste and may have developed an inverse association with dopamine release across the larger reward circuitry.61,62 A possible explanation could be that high HA drives low taste pleasantness, making the taste experience less pleasant.

ossibility that adolescents with AN in this study were negatively conditioned to sweet taste and may have developed an inverse association with dopamine release across the larger reward circuitry.61,62 A possible explanation could be that high HA drives low taste pleasantness, making the taste experience less pleasant. Consistent with other studies, those with AN exhibited higher AUC for cortisol levels compared with controls,29,30 which may alter appetite regulation in AN.63 Depression scores were not correlated with cortisol levels. The positive regression of AUC for cortisol level and PE signal in participants with AN suggests that stress enhances PE signals.64 Prediction error acts as a learning signal and affects value-driven attentional bias,65 and stress response may affect how individuals with AN process and form associations with salient stimuli. Although correlational, our data point to a model hypothesis for further investigation: cognitive drivers in AN, such as severe body dissatisfaction, could increase stress hormone levels, which both suppress eating and enhance PE signals.

ffect how individuals with AN process and form associations with salient stimuli. Although correlational, our data point to a model hypothesis for further investigation: cognitive drivers in AN, such as severe body dissatisfaction, could increase stress hormone levels, which both suppress eating and enhance PE signals. Taking our data together with previous research, we propose the following model to explain the paradoxical food restriction in AN (Figure 3): Food restriction and weight loss are associated with sensitization of the dopamine system14 and reflected in AN by elevated PE signal, probably to stimulate food approach.11,12 However, PE response may increase HA in AN because this biological mechanism to seek out food is inconsistent with the high drive for thinness and body dissatisfaction. Thus, there is a conflict between food approach mechanisms (PE) and cognitive-emotional processes that oppose eating (body dissatisfaction and drive for thinness). Prediction error activation may then become part of a fear-driven mechanism that includes the ventral striatum to override homeostatic signals from the hypothalamus, which would normally trigger food intake. Future studies will have to test the validity of this model and test whether this circuitry indeed activates or involves the previously described fear-mediating pathway from the ventral striatum to the hypothalamus.56

iatum to override homeostatic signals from the hypothalamus, which would normally trigger food intake. Future studies will have to test the validity of this model and test whether this circuitry indeed activates or involves the previously described fear-mediating pathway from the ventral striatum to the hypothalamus.56 Figure 3. Model of Association of Reward Learning Prediction Error (PE) With Harm Avoidance, Weight Change, and Effective Connectivity in Adolescents With Anorexia Nervosa Based on the correlational results, we propose that weight loss is associated with expected sensitization of the dopamine system, reflected in elevated PE to stimulate food approach. However, increased PE may elevate anxiety (harm avoidance) in anorexia nervosa because this is in conflict with a high drive for thinness and body dissatisfaction. Prediction error activation may then become part of a fear-driven mechanism to override homeostatic signals from the hypothalamus, signals that would normally trigger eating. BMI indicates body mass index; OFC, orbitofrontal cortex.

sa because this is in conflict with a high drive for thinness and body dissatisfaction. Prediction error activation may then become part of a fear-driven mechanism to override homeostatic signals from the hypothalamus, signals that would normally trigger eating. BMI indicates body mass index; OFC, orbitofrontal cortex. Limitations This study has limitations. Functional magnetic resonance imaging does not directly measure dopaminergic signaling; the biologically based computational model used in this study provides strong evidence of altered dopamine-related taste reward processing in adolescent AN, but specific pharmacological challenge studies are needed to further support this model. The dynamic connectivity analysis included conditioned visual stimulus and unconditioned taste stimulus response; future studies will be required to separate these over the full hemodynamic response time. Power for the cortisol level analysis was limited because we were unable to collect salivary cortisol samples from the entire cohort. The analysis of 20 participants with AN and 25 controls still resulted in robust, multiple comparison–corrected findings, although replication is needed. The data were collected on 2 separate MRI scanners. However, comparison results did not indicate within-group differences. We did not assess learning rates; however, a subset of the study participants showed similar learning rates across groups in a monetary PE paradigm.22 Structural brain alterations could affect brain function, but there were no differences across groups (eFigure 4 in the Supplement). The age range up to 21 years was within child range by National Institute of Mental Health standards at the start of the study. However, there is no evidence that extending the current definition by 3 years would confound the results.

function, but there were no differences across groups (eFigure 4 in the Supplement). The age range up to 21 years was within child range by National Institute of Mental Health standards at the start of the study. However, there is no evidence that extending the current definition by 3 years would confound the results. Conclusions Prediction error brain response may have a central role in adolescent AN and illness behaviors. However, longitudinal studies and neurotransmitter challenge studies are needed to further understand how brain circuits are disrupted or altered in AN. This will help to identify neurobiological systems that are involved in AN pathophysiology and to develop targeted biological interventions. Another goal will be to identify demographic, behavioral, and biological variables that can predict PE signal and can be measured without a brain scan to make this mechanism clinically more accessible and useful. Supplement. eAppendix 1. Taste reward task paradigm. eAppendix 2. Cortisol-brain activation regression. eMethods 1. 3T GE Signa and Siemens Skyra 3T scanners. eMethods 2. Temporal difference learning algorithm. eMethods 3. Effective connectivity analysis methods. eTable 1. Cortisol sample demographics. eTable 2. Taste pleasantness 1 molar sucrose solution and ROI PE response correlations. eFigure 1. PE analysis by scanner. eFigure 2. Behavioral correlations. eFigure 3. Cortisol analysis. eFigure 4. Brain volume measures. Click here for additional data file.

Introduction International migration is an established risk factor for psychosis,1 and the risk appears greatest for those migrating at earlier ages.2,3 Some studies have suggested that internal migration (known as residential mobility) may also increase psychosis risk.4,5 Evidence from Denmark has suggested that long-distance childhood residential mobility increases the subsequent risk of schizophrenia and other nonaffective psychoses,4,5,6 with some evidence of a stronger effect of residential instability during adolescence than in childhood.4 However, to our knowledge, none of the studies have examined the effect of residential mobility beyond midadolescence, which may be particularly important given that some have suggested that higher rates of psychotic disorders in more deprived, socially fragmented urban environments7,8,9 are a consequence of social drift during the prodromal phases of disorder, as people may move into cheaper, more socially isolated environments.10 Moreover, to date, the geographical distances people move have been crudely treated as moves between large administrative areas, potentially obscuring the nuanced effects of moving over smaller or larger geographical distances.

phases of disorder, as people may move into cheaper, more socially isolated environments.10 Moreover, to date, the geographical distances people move have been crudely treated as moves between large administrative areas, potentially obscuring the nuanced effects of moving over smaller or larger geographical distances. In this study, we used data from individuals within a large population-based cohort, whose residential moves over their entire early life course (up to age 29 years) could be identified to small area neighborhood resolution, to examine the risk of developing nonaffective psychotic disorders associated with residential mobility during childhood, adolescence, and early adulthood. We focused on nonaffective psychoses, given stronger evidence that these psychiatric disorders are more strongly associated with urbanicity and migration than other disorders, such as bipolar disorder or unipolar depression.11,12 Given previous evidence, we hypothesized that having more frequent residential moves in childhood, adolescence, and early adulthood would be associated with an increased psychosis risk and that this would be highest for individuals who moved during adolescence, which is a key period for social development.4,13 We also hypothesized that the risk would increase with greater geographical distances moved in childhood and adolescence but in a nonlinear fashion, representing a “threshold” effect at which most moves were likely to result in a breakup of social networks (eg, due to an enforced change of school).

ial development.4,13 We also hypothesized that the risk would increase with greater geographical distances moved in childhood and adolescence but in a nonlinear fashion, representing a “threshold” effect at which most moves were likely to result in a breakup of social networks (eg, due to an enforced change of school). Methods Sample We identified all individuals born in Sweden between January 1, 1982, and December 31, 1995, who resided in Sweden on their 16th birthday from the Total Population Register. This study received ethical approval through Psychiatry Sweden from the Stockholm Regional Ethical Review Board and consent was waived. Persons were followed up from their 16th birthday until receiving a first diagnosis of an International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) nonaffective psychotic disorder, censorship due to emigration, death, or December 31, 2011, whichever was sooner. We excluded first-generation immigrants because information on their residential mobility before moving to Sweden was unavailable. From our initial cohort (N = 1 472 446), our final analytical sample included 1 440 383 participants with complete residential mobility data ( eMethods in the Supplement).

hever was sooner. We excluded first-generation immigrants because information on their residential mobility before moving to Sweden was unavailable. From our initial cohort (N = 1 472 446), our final analytical sample included 1 440 383 participants with complete residential mobility data ( eMethods in the Supplement). Outcome Measure Our main outcome measure was a clinical diagnosis of nonaffective psychotic disorder (ICD-10: F20-29), including schizophrenia (F20) and other nonaffective psychoses (F21-29) as recorded in the Swedish National Patient Register. For this study, the coverage for all inpatient admissions was complete over the follow-up period and for outpatient admissions from 2001.14,15

f nonaffective psychotic disorder (ICD-10: F20-29), including schizophrenia (F20) and other nonaffective psychoses (F21-29) as recorded in the Swedish National Patient Register. For this study, the coverage for all inpatient admissions was complete over the follow-up period and for outpatient admissions from 2001.14,15 Exposure Variables We investigated whether the number of moves over discrete periods of the life course and the cumulative distances moved were associated with subsequent psychosis risk. Age periods (0-6 years, 7-15 years, 16-19 years, and 20-29 years) were determined a priori to coincide with the transition through the Swedish public education system. We estimated the number of moves from the Total Population Register, which records the residential location of all people each year to one of 9200 “Small Area for Market Statistics” (SAMS) areas (median population size in 2011, 726 [interquartile range, 312-1378]). Small Area for Market Statistics are designed to be internally socioeconomically homogenous but differ according to the characteristics of the social environment, including deprivation and population density.16 For each participant and age period, we calculated the total number of times a change in SAMS residence occurred from year to year as: no moves (reference), 1 move, 2 moves, 3 moves, or 4 or more moves (eMethods in the Supplement). For each age period, we also estimated the cumulative distance moved by each participant (in kilometers) (eMethods in the Supplement).

ated the total number of times a change in SAMS residence occurred from year to year as: no moves (reference), 1 move, 2 moves, 3 moves, or 4 or more moves (eMethods in the Supplement). For each age period, we also estimated the cumulative distance moved by each participant (in kilometers) (eMethods in the Supplement). Covariates We included confounder data on continuous age; sex; parental migration status (both parents Swedish-born vs at least 1 parent being foreign-born); biological parental history of severe mental illness (SMI), including nonaffective psychosis and bipolar disorders or mania with or without psychotic symptoms; the biological mother’s age at participant birth (as a proxy for paternal age); parental (biological or adoptive) death in any age period before the participant’s 16th birthday; SAMS population density at birth (eFigure 1 in the Supplement); participant compulsory school educational attainment; family disposable income at cohort entry; and university attendance (yes or no) (eMethods in the Supplement).

tal (biological or adoptive) death in any age period before the participant’s 16th birthday; SAMS population density at birth (eFigure 1 in the Supplement); participant compulsory school educational attainment; family disposable income at cohort entry; and university attendance (yes or no) (eMethods in the Supplement). Statistical Methods We fitted discrete time proportional hazards models using complementary log-log models on the attained age scale (eMethods in the Supplement). Modeling proceeded as follows: (1) we modeled the crude association of the number of residential moves and cumulative distance moved in each age period with psychosis risk, (2) we added all the covariates to these models (adjustment 1) except for educational attainment at age 15 to 16 years (see Results), (3) we adjusted the models for residential move data in previous age periods (adjustment 2), and (4) because educational attainment at age 15 to 16 years may have been on the causal pathway between earlier moves and future psychosis risk, we restricted the adjustment for this variable to models of residential moves made after age 16 years (adjustment 3). When modeling residential moves after age 20 years, we excluded participants who had not reached this age by the end of the follow-up period or who were otherwise censored between age 16 to 19 years(n = 441 416; 30.1%). We included university attendance as a potential confounder (adjustment 4) and effect modifier of the association between residential moves and nonaffective psychosis risk in adulthood. To examine possible threshold effects in the geographical distances of residential moves, we inspected nonlinear distance functions using an inverse power (square root) transformation and compared this with a model that was fitted with a linear distance function via an inspection of Akaike Information Criterion scores in which lower scores indicated a better fit. We predicted and graphed marginal hazards over the cumulative distance moved in each age period. In a subgroup analysis, we investigated whether any associations of the geographical distances of residential moves with nonaffective psychosis risk were upheld among those who moved only once in each age period compared with those who never moved, with moving distance categorized as never moved, less than 5 km, 5 to 30 km, 30 to 100 km, 100 to 500 km, and 500 or more km. We reported hazard ratios (HRs) and 95% confidence intervals. Statistical significance was set at P < .05.

among those who moved only once in each age period compared with those who never moved, with moving distance categorized as never moved, less than 5 km, 5 to 30 km, 30 to 100 km, 100 to 500 km, and 500 or more km. We reported hazard ratios (HRs) and 95% confidence intervals. Statistical significance was set at P < .05. Results Sample Characteristics Of 1 440 383 included participants (97.8% of cohort; eTable 1 and eResults in the Supplement), 4537 (0.31%; 95% CI, 0.30-0.33) received an ICD-10 diagnosis of nonaffective psychotic disorder in Sweden during the follow-up period. The median age when receiving the first diagnosis was 20.9 years (interquartile range, 19.0-23.3). Participants with nonaffective psychosis were more likely to be men, come from a lower-income quintile, and have a foreign background, parental history of SMI, death of a parent before age 16 years, and lower educational attainment and less likelihood of attending university than the remainder of the cohort (Table 1). The distribution of the number of residential moves (Figure 1) and the cumulative distance moved differed for participants with nonaffective psychosis compared with the remainder of the cohort (Table 1). Thus, before age 20 years, case participants were more likely to have moved at least once and have had a longer cumulative distance moved (all P < .001); this pattern was reversed after age 20 years. The correlations within and between the number of moves and the cumulative distance moved were moderate (eTable 2 and eResults in the Supplement).

case participants were more likely to have moved at least once and have had a longer cumulative distance moved (all P < .001); this pattern was reversed after age 20 years. The correlations within and between the number of moves and the cumulative distance moved were moderate (eTable 2 and eResults in the Supplement). Table 1. Cohort Characteristics by Outcome Status Variable Outcome Status (Nonaffective Psychotic Disorder) χ2 df P Value Yes No Participants (%)a 4537 (0.31) 1 435 846 (99.69) NA NA NA Median age (IQR) 20.9 (19.0-23.3) 22.4 (19.4-25.8) 21.6 NA <.001b Men (%) 2746 (60.5) 737 105 (51.3) 152.9 1 <.001 Foreign background (%) 851 (18.8) 181 462 (12.6) 153.2 1 <.001 Death of a parent before age 16 y (%) 173 (3.8) 32 279 (2.3) 50.3 1 <.001 Parental history of SMI (%) 467 (10.3) 42 695 (3.0) 833.6 1 <.001 Median maternal age at birth (IQR) 28.2 (24.4-32.3) 28.3 (24.9-32.0) 1.5 .12b Income quintile (%) Highest 67 (1.5) 307 727 (2.1) 244.4 4 <.001 2 128 (2.8) 55 056 (3.8) 3 305 (6.7) 139 489 (9.7) 4 783 (17.3) 343 889 (24.0) Lowest 3254 (71.7) 866 685 (60.4) Population density at birth (pp km2) 1255.9 (158.7-3889.6) 739.3 (48.7-2850.6) −13.3 NA <.001b Educational attainment at age 16 (%) Fail 650 (14.3) 64 934 (4.5) 2566.1 4 <.001 D or E grades 2622 (57.8) 863 511 (60.1) C grade 484 (10.7) 285 199 (19.9) A or B grades 198 (4.4) 173 691 (12.1) Missing 583 (12.9) 48 511 (3.4) University attendancec No 2312 (82.7) 621 033 (62.3) 499.3 1 <.001 Yes 483 (17.3) 375 139 (37.7) Moved ≥1 times (%), yd 0-6 2467 (54.4) 646 645 (45.0) 159.3 1 <.001 7-15 2148 (47.3) 504 245 (35.1) 295.8 1 <.001 16-19 1762 (38.8) 382 908 (26.7) 342.1 1 <.001 20-29c 1726 (61.8) 690 316 (69.3) 74.5 1 <.001 Median cumulative distance moved (km) (10-90th percentile), ye 0-6 1.2 (0-133.5) 0 (0-45.5) −14.4 NA <.001b 7-15 0 (0-86.3) 0 (0-27.4) −19.0 NA <.001b 16-19 0 (0-64.7) 0 (0-23.3) −18.7 NA <.001b 20-29c 0 (0-332.6) 0 (0-416.8) 9.9 NA <.001b Abbreviations: df, degrees of freedom; IQR, interquartile range; NA, not applicable; SMI, severe mental illness.

(km) (10-90th percentile), ye 0-6 1.2 (0-133.5) 0 (0-45.5) −14.4 NA <.001b 7-15 0 (0-86.3) 0 (0-27.4) −19.0 NA <.001b 16-19 0 (0-64.7) 0 (0-23.3) −18.7 NA <.001b 20-29c 0 (0-332.6) 0 (0-416.8) 9.9 NA <.001b Abbreviations: df, degrees of freedom; IQR, interquartile range; NA, not applicable; SMI, severe mental illness. a Row percentage. b Mann-Whitney U test for nonnormally distributed data. c Among those who did not exit the cohort before age 20 years (n = 998 967 [69.4%]). d For descriptive purposes, the number and proportion of people who moved 1 or more times in each period are displayed. A categorical variable (0,1,2,3, ≥4) was used for modeling purposes. e The 10th-90th percentile is reported in favor of the interquartile range, given substantial skew in the distribution of the exposures. Figure 1. Residential Mobility by Number of Moves and Outcome Status by Age Period At ages 0 to 6 (A), 7 to 15 (B), and 16 to 19 years (C) the proportion of cases who moved once or more was greater than in the population at risk. By contrast, at 20 years and older (D), case participants were more likely to have never moved than the population at risk. Percentages in the 20 years or older group were estimated from participants who were not censored before this age (n = 998 967 [69.9%]). Percentages at all other age ranges were based on the full sample (N = 1 440 383). At age 16 to 19 years, the maximum number of possible moves during this period is 4.

lation at risk. Percentages in the 20 years or older group were estimated from participants who were not censored before this age (n = 998 967 [69.9%]). Percentages at all other age ranges were based on the full sample (N = 1 440 383). At age 16 to 19 years, the maximum number of possible moves during this period is 4. Association Between Residential Mobility and Nonaffective Psychosis We observed dose-response associations between greater moves at age 0 to 6 years, 7 to 15 years, and 16 to 19 years and the risk of nonaffective psychotic disorder in unadjusted survival models (Table 2) that persisted after adjusting for covariates (adjustment 1), including moves at previous ages (adjustment 2). Thus, compared with never moving, 1, 2, 3, or 4 or more moves between birth and age 6 years were associated with HRs of 1.13, 1.47, 1.46, and 1.83, respectively (adjustment 1; all P < .001) (Table 2). We observed similar associations for moves between age 7 to 15 years and stronger associations at age 16 to 19 years (adjustment 2; Table 2), with those moving in each year of this period having a 2.88-fold (95% CI, 1.89-4.40) increased risk compared with those who never moved. Further adjustment for educational attainment at age 15 to 16 years attenuated risks between age 16 to 19 years (adjustment 3, Table 2), but strong dose-response patterns remained (ie, moving 4 times; HR, 1.99; 95% CI, 1.30-3.05).

aving a 2.88-fold (95% CI, 1.89-4.40) increased risk compared with those who never moved. Further adjustment for educational attainment at age 15 to 16 years attenuated risks between age 16 to 19 years (adjustment 3, Table 2), but strong dose-response patterns remained (ie, moving 4 times; HR, 1.99; 95% CI, 1.30-3.05). Table 2. Hazard Ratios (HR) for Nonaffective Psychosis by Number and Distance of Moves in Each Age Period Exposures Cases, No. (%) HR (95% CI) Crude Adjustment 1a Adjustment 2b Adjustment 3c Adjustment 4d Age 0-6, ye 1 move 1416 (31.2) 1.29 (1.21-1.38) 1.13 (1.04-1.21) NA NA NA 2 moves 696 (15.3) 1.81 (1.66-1.97) 1.47 (1.32-1.62) NA NA NA 3 moves 241 (5.4) 1.92 (1.68-2.20) 1.46 (1.25-1.70) NA NA NA ≥4 moves 114 (2.5) 2.51 (2.08-3.03) 1.83 (1.48-2.26) NA NA NA Distance (square root)f NA 1.37 (1.31-1.43) 1.13 (1.06-1.19) NA NA NA Age 7-15, ye 1 move 1062 (23.4) 1.44 (1.34-1.55) 1.24 (1.15-1.34) 1.22 (1.13-1.32) NA NA 2 moves 580 (12.8) 1.94 (1.77-2.13) 1.58 (1.43-1.75) 1.51 (1.36-1.68) NA NA 3 moves 288 (6.4) 2.41 (2.13-2.72) 1.86 (1.62-2.13) 1.74 (1.52-2.01) NA NA ≥4 moves 218 (4.8) 2.99 (2.60-3.44) 2.14 (1.82-2.53) 1.95 (1.65-2.31) NA NA Distance (square root)f NA 1.52 (1.46-1.59) 1.16 (1.09-1.23) 1.11 (1.05-1.19) NA NA Age 16-19, ye 1 move 1125 (24.8) 1.45 (1.35-1.55) 1.47 (1.36-1.59) 1.35 (1.25-1.47) 1.28 (1.18-1.39) NA 2 moves 497 (11.0) 2.47 (2.25-2.72) 2.45 (2.20-2.74) 2.08 (1.85-2.33) 1.79 (1.60-2.01) NA 3 moves 117 (2.6) 2.61 (2.17-3.14) 2.55 (2.09-3.11) 2.00 (1.63-2.45) 1.57 (1.28-1.92) NA ≥4 moves 23 (0.5) 3.96 (2.62-5.97) 3.87 (2.54-5.90) 2.88 (1.89-4.40) 1.99 (1.30-3.05) NA Distance (square root)f NA 1.32 (1.26-1.40) 0.98 (0.92-1.06) 0.95 (0.88-1.02) 0.99 (0.98-1.03) NA Age ≥20, ye,g 1 move 787 (28.2) 0.82 (0.74-0.89) 1.11 (1.01-1.22) 1.04 (0.94-1.14) 1.05 (0.96-1.16) 1.04 (0.94-1.14) 2 moves 491 (17.6) 0.69 (0.62-0.78) 1.18 (1.04-1.33) 1.05 (0.93-1.19) 1.07 (0.95-1.21) 1.05 (0.93-1.18) 3 moves 254 (9.1) 0.70 (0.61-0.81) 1.46 (1.25-1.71) 1.25 (1.07-1.47) 1.27 (1.08-1.49) 1.23 (1.05-1.44) ≥4 moves 194 (6.9) 0.88 (0.75-1.05) 2.35 (1.95-2.84) 1.91 (1.58-2.31) 1.91 (1.58-2.30) 1.82 (1.51-2.20) Distance (square root)f NA 0.60 (0.57-0.63) 0.58 (0.54-0.61) 0.56 (0.53-0.60) 0.60 (0.56-0.63) 0.67 (0.63-0.71) Abbreviation: NA, not applicable.

1) 1.46 (1.25-1.71) 1.25 (1.07-1.47) 1.27 (1.08-1.49) 1.23 (1.05-1.44) ≥4 moves 194 (6.9) 0.88 (0.75-1.05) 2.35 (1.95-2.84) 1.91 (1.58-2.31) 1.91 (1.58-2.30) 1.82 (1.51-2.20) Distance (square root)f NA 0.60 (0.57-0.63) 0.58 (0.54-0.61) 0.56 (0.53-0.60) 0.60 (0.56-0.63) 0.67 (0.63-0.71) Abbreviation: NA, not applicable. a Adjustment 1: age, quadratic age, sex, foreign background, parental history of severe mental illness, parental death before age 16 years, disposable income quintile, mother’s age at participant birth, population density at birth (log transformed people per square kilometer), and distance moved in age period. b Adjustment 2: adjustment 1 + number of and distance moved at previous ages. c Adjustment 3: adjustment 2 + educational attainment at age 15 to 16 years. d Adjustment 4: adjustment 3 + university attendance. e Reference group for number of moves: 0 moves. f A nonlinear distance function (square root transformation) was provided to better fit to the data than a linear term; assessed via Akaike Information Criterion (Figure 2; eTable 3 in the Supplement). g After age 20 years, model was restricted to cohort not censored before this point (n = 998 967).

e Reference group for number of moves: 0 moves. f A nonlinear distance function (square root transformation) was provided to better fit to the data than a linear term; assessed via Akaike Information Criterion (Figure 2; eTable 3 in the Supplement). g After age 20 years, model was restricted to cohort not censored before this point (n = 998 967). There was weaker evidence that moving after age 20 years was associated with psychosis risk, with little variation in risk for those who moved fewer than 3 times in early adulthood, including after adjustment for educational attainment and university attendance (adjustment 4, Table 2). Nonetheless, those who moved more frequently (4 or more times) remained at a substantially elevated risk (HR, 1.82; 95% CI, 1.51-2.20). We found moderate evidence that this relative association was stronger in those who attended university (HR, 2.56; 95% CI, 1.55-3.54) than those who did not (HR, 1.65; 95% CI, 1.34-2.02; likelihood ratio test P = .02; eTable 3 in the Supplement), although marginal (ie, absolute) changes in the predicted probabilities of nonaffective psychosis for each additional move were similar in both groups (eFigure 2 in the Supplement).

2.56; 95% CI, 1.55-3.54) than those who did not (HR, 1.65; 95% CI, 1.34-2.02; likelihood ratio test P = .02; eTable 3 in the Supplement), although marginal (ie, absolute) changes in the predicted probabilities of nonaffective psychosis for each additional move were similar in both groups (eFigure 2 in the Supplement). The cumulative distances moved at all ages were better modeled as nonlinear functions with respect to psychosis risk (eTable 4 in the Supplement). Independent of the number of moves, greater moving distances before age 16 years increased risk (Table 2), most sharply over shorter (ie, less than 30 km) distances moved (Figure 2A and B). Between age 16 to 19 years, we observed no evidence of any statistically significant association with distance. After age 20 years, greater cumulative distances moved were associated with decreased psychosis risk, with similar evidence of threshold effects at shorter distances (Figure 2D). These patterns were replicated in subgroup analyses that were restricted to participants who moved only once during each period compared with those who never moved (Table 3).

ative distances moved were associated with decreased psychosis risk, with similar evidence of threshold effects at shorter distances (Figure 2D). These patterns were replicated in subgroup analyses that were restricted to participants who moved only once during each period compared with those who never moved (Table 3). Figure 2. Predicted Hazard of Nonaffective Psychotic Disorder by Cumulative Distance Moved in Each Age Period Relative hazard of nonaffective psychotic disorder by cumulative distance between ages 0 to 6 (A), 7 to 15(B), 16 to 19 (C), and 20 or more years (D). Distances are displayed per 100 km up to a total of 1000 km. The shading denotes 95% CIs. Each model is based on the predicted relative hazard following modeling that was adjusted for the covariates listed in adjustment 2 (Table 2). Distances moved before age 16 years displayed a strong nonlinear trend, such that the relative hazard of nonaffective psychosis increased most quickly over shorter move distances (ie, within 30 km) before increasing at a slower rate over longer distances (with less certainty around point estimates). Distance moved between age 16 to 19 years was best modeled as a linear predictor, with no significant differences in the relative hazard of nonaffective psychosis observed by distance (Table 2). Cumulative distances moved after age 20 years were associated with a strong, nonlinear reduction in the relative hazard of nonaffective psychosis, particularly for moves up to approximately 30 km.

predictor, with no significant differences in the relative hazard of nonaffective psychosis observed by distance (Table 2). Cumulative distances moved after age 20 years were associated with a strong, nonlinear reduction in the relative hazard of nonaffective psychosis, particularly for moves up to approximately 30 km. Table 3. Risk of Developing Nonaffective Psychosis by Distance Moved Among Those Who Moved Once vs Never Having Moved in Each Age Period Distance Moved Adjusted HRa (95% CI) Age 0-6, y No moves 1 [Reference] <5 km 1.11 (1.01-1.22) 5-29 km 1.14 (1.03-1.26) 30-99 km 1.47 (1.25-1.74) 100-499 km 1.37 (1.04-1.50) >500 km 1.58 (1.08-2.34) Age 7-15, y No moves 1 [Reference] <5 km 1.16 (1.05-1.28) 5-29 km 1.27 (1.14-1.42) 30-99 km 1.61 (1.32-1.95) 100-499 km 1.37 (1.10-1.70) >500 km 1.58 (0.99-2.53) Age, 16-19, y No moves 1 [Reference] <5 km 1.10 (0.99-1.23) 5-29 km 1.10 (0.99-1.22) 30-99 km 1.09 (0.92-1.30) 100-499 km 1.03 (0.87-1.22) >500 km 0.82 (0.53-1.27) Age, ≥20, y No moves 1 [Reference] <5 km 0.41 (0.36-0.46) 5-29 km 0.24 (0.21-0.27) 30-99 km 0.20 (0.16-0.24) 100-499 km 0.16 (0.13-0.19) >500 km 0.08 (0.05-0.12) Abbreviation: HR, hazard ratio. a Adjusted for age, quadratic age, sex, foreign background, family history of severe mental illness, parental death before age 16 years, disposable income quintile, mother’s age at participant birth, population density at birth (log transformed people per square kilometer,) and distance and number of moves (categorical) in previous age periods (except age 0-6 years).

ackground, family history of severe mental illness, parental death before age 16 years, disposable income quintile, mother’s age at participant birth, population density at birth (log transformed people per square kilometer,) and distance and number of moves (categorical) in previous age periods (except age 0-6 years). In a sensitivity analysis (eTable 5 in the Supplement), we presented results from a fully mutually adjusted model of the number and distances of moves at each period of the life course to facilitate comparability with earlier studies.4 In this model, the association between the number of moves and psychosis risk was most substantially attenuated at ages 0 to 6 years and 7 to 15 years; the number of moves between age 16 to 19 years continued to exhibit a dose-response association with later psychosis risk.

cilitate comparability with earlier studies.4 In this model, the association between the number of moves and psychosis risk was most substantially attenuated at ages 0 to 6 years and 7 to 15 years; the number of moves between age 16 to 19 years continued to exhibit a dose-response association with later psychosis risk. Discussion In this study we show that greater residential mobility during childhood and adolescence is associated with a dose-response increase in risk of developing nonaffective psychosis. These patterns were impervious to adjustment for psychiatric family history and sociodemographic indicators, including family disposable income, and could not be explained by moves at previous ages nor, when relevant, educational attainment at age 15 to 16 years or university attendance. The larger effect sizes for moves between age 16 to 19 years is consistent with the thesis that residential mobility is associated with nonaffective psychosis through a mechanism that is at its most sensitive during adolescence, in line with earlier observations.4 We also found that longer geographical distances of residential moves during childhood and early adolescence were associated with increased risk, independently of the number of moves, particularly for moves more than approximately 30 km, which was consistent with the distances at which the disruption of school-based or other social networks were more likely.

istances of residential moves during childhood and early adolescence were associated with increased risk, independently of the number of moves, particularly for moves more than approximately 30 km, which was consistent with the distances at which the disruption of school-based or other social networks were more likely. The association between residential mobility and nonaffective psychosis was different in young adults. While there was some evidence that moving frequently (3 or more times) between age 20 to 29 years was associated with increased risk, no differences emerged for individuals who moved fewer times in early adulthood. Moreover, moving longer distances in adulthood was strongly associated with a reduced risk of nonaffective psychosis. Taken together, these results suggest that residential stability in early life and some geographical mobility in adulthood do not increase and may confer protection against psychosis risk. Potential Mechanisms The most supported explanation as to how residential mobility could have an association with nonaffective psychosis is that a change of residence disrupts an individual’s ability to form and maintain friendships or fit within a peer group. Social isolation is likely to increase one’s vulnerability to the effects of life stressors. For example, exposure to stressful life events could have a greater impact on negative schemata, low self-esteem, and cognitive biases that are associated with psychosis17,18,19 without the buffering effect of stable friendships.

cial isolation is likely to increase one’s vulnerability to the effects of life stressors. For example, exposure to stressful life events could have a greater impact on negative schemata, low self-esteem, and cognitive biases that are associated with psychosis17,18,19 without the buffering effect of stable friendships. Some studies suggest that part of the association of residential mobility with psychosis is mediated via having to change schools, and that loss of peer relationships and increased social isolation may be involved in the pathway to risk.20,21 Residential mobility may disrupt social relationships and be associated with subsequent psychosis risk if it necessitates a change in schools, and if it occurs at a time when relationships with peers become as or more important than family-based ones. Our finding that the greatest risk was observed for residential moves during late adolescence, independent of academic ability at age 15 to 16 years, is consistent with this thesis, as is our finding that longer moves predicted greater psychosis risk. Nonetheless, not all studies have observed associations between school mobility and psychosis risk,22 suggesting that beyond the school context, other peer group relationships, including family, kinship, and wider neighborhood ties, may also be relevant.

s our finding that longer moves predicted greater psychosis risk. Nonetheless, not all studies have observed associations between school mobility and psychosis risk,22 suggesting that beyond the school context, other peer group relationships, including family, kinship, and wider neighborhood ties, may also be relevant. We have previously shown that the characteristics that mark someone out as different from most of their peers, whether at a school level or neighborhood level, are associated with an increased risk of psychosis,7,14 findings that are often conceptualized within the concept of social defeat. It has been hypothesized that social defeat contributes causally to psychosis risk via the sensitization of the mesolimbic dopamine system,19 the disruption of which is a widely supported biological theory of schizophrenia.23 Support for this theory is evident from animal model studies,24,25 and such a mechanism might explain how greater residential mobility, especially during adolescence, increases psychosis risk if it is subsequently accompanied by changes in the propensity to experience social adversities, such as social isolation and/or exposure to stressful life events.26 It is also possible that the association between residential mobility and psychosis is, at least in part, mediated by factors other than disrupted social relationships (eg, an earlier initiation of drug use27,28,29 or reduced engagement with health, social, and education services30,31,32).

e to stressful life events.26 It is also possible that the association between residential mobility and psychosis is, at least in part, mediated by factors other than disrupted social relationships (eg, an earlier initiation of drug use27,28,29 or reduced engagement with health, social, and education services30,31,32). Our findings with respect to early adulthood somewhat contrast those for mobility at earlier ages. Moving once or twice during this period did not alter risk, and those who moved longer distances were substantially less likely to subsequently develop psychotic disorder; cumulative distances moved accounted for the change in the direction of the unadjusted, protective association between the number of moves and nonaffective psychosis risk to a risk factor in adjusted models (data available from the authors). These findings were not substantially confounded by university attendance. Changing residence after age 20 years, the age at which students in Sweden complete their university entrance examinations, is likely to reflect the onset of independence for an individual, be it through university attendance or entry into the labor market, and may explain why there were weaker associations with the number of moves and a strong negative association between moving distance and psychosis risk during this period. Although less consistently than for greater cognitive ability,33,34 higher levels of education has been associated with reduced risk of nonaffective psychosis,35,36 Nonetheless, frequent moves (particularly 4 or more) in early adulthood remained associated with increased psychosis risk irrespective of university attendance, and we hypothesize that this reflects the more chaotic lifestyle of individuals who are at higher risk of developing psychosis (eg, as a result of substance misuse or the presence of financial, social, or other mental health difficulties).

associated with increased psychosis risk irrespective of university attendance, and we hypothesize that this reflects the more chaotic lifestyle of individuals who are at higher risk of developing psychosis (eg, as a result of substance misuse or the presence of financial, social, or other mental health difficulties). Strengths and Limitations This study has several strengths, including the longitudinal design and large sample size that is highly representative of the entire (Swedish-born) population. The prospective measurement of our exposure and the use of register data eliminate the possibility of recall bias. Our outcome measure is known to have good concurrent validity for diagnoses of schizophrenia in this register,37,38 and psychiatric care in Sweden is both accessible and free. Using geographical information systems data to estimate the number and distance of small area moves over the life course is a further strength of this study, although we acknowledge that the distances were based on “as the crow flies” estimates. We controlled for several confounders that may have precipitated residential mobility, including parental death, a parental history of SMI, urban birth, income, younger maternal age at birth, and, with respect to moves made after age 16 years, educational attainment. We acknowledge that we did not have data on other potential confounders, including other adverse childhood experiences such as family discord or parental separation. Nor did we have data on measures such as quality of friendships and peer problems, such as bullying, to test potential mediating pathways. We also lacked direct data on school changes as an index of disruption to peer relationships. Selection bias might be present from the small amount of incomplete geographical data in this study (eTable 1 in the Supplement), although this might be expected to have underestimated our associations, given that reasons for missingness include homelessness and being in prison, which are further markers of residential mobility and are associated with nonaffective psychosis. While reverse causation is unlikely to explain our findings, it remains feasible that subthreshold or prodromal symptoms during childhood or early adolescence led some families to change residence in the hope that a different school, neighborhood, or proximity to specialist health care clinicians might improve their child’s well-being.

sation is unlikely to explain our findings, it remains feasible that subthreshold or prodromal symptoms during childhood or early adolescence led some families to change residence in the hope that a different school, neighborhood, or proximity to specialist health care clinicians might improve their child’s well-being. Finally, in other studies, residential mobility has been associated with bipolar disorder and substance use disorders, suggesting that residential mobility may be a nonspecific risk factor for several psychiatric disorders.4,5 Conclusions Accumulating evidence supports childhood and adolescent residential mobility as an independent risk factor for psychosis and other mental health outcomes. Efforts are now required to examine the reasons for this, which may include precipitating factors such as family discord, as well as the effect such moves are likely to have on peer group formation and social support during critical periods of development; these findings will also have implications for informing the development of child health services and social policy. It is important that health, social, and educational practitioners ensure that children and adolescents who are newly resident to their neighborhoods receive adequate support to minimize the risks of adverse outcomes during adulthood, and every effort should be made to ensure the effective transfer of care for highly mobile children who are already in contact with health and social services. Supplement. eMethods. eResults. eFigure 1. Positively skewed distribution of population density at birth in cohort

Conclusions Accumulating evidence supports childhood and adolescent residential mobility as an independent risk factor for psychosis and other mental health outcomes. Efforts are now required to examine the reasons for this, which may include precipitating factors such as family discord, as well as the effect such moves are likely to have on peer group formation and social support during critical periods of development; these findings will also have implications for informing the development of child health services and social policy. It is important that health, social, and educational practitioners ensure that children and adolescents who are newly resident to their neighborhoods receive adequate support to minimize the risks of adverse outcomes during adulthood, and every effort should be made to ensure the effective transfer of care for highly mobile children who are already in contact with health and social services. Supplement. eMethods. eResults. eFigure 1. Positively skewed distribution of population density at birth in cohort eFigure 2. Predicted probability of non-affective psychosis by number of moves aged 20 years and older, by university attandence eTable 1. Cohort characteristics by missing data type eTable 2. Correlation between number and cumulative distance of moves at different age periods eTable 3. Effect modification of number of moves after 20 years old on non-affective psychosis risk, by university attendance eTable 4. Model fit comparisons for linear versus non-linear distance functions eTable 5. Risk of non-affective psychoses after mutual adjustment for moves in all periods

eTable 2. Correlation between number and cumulative distance of moves at different age periods eTable 3. Effect modification of number of moves after 20 years old on non-affective psychosis risk, by university attendance eTable 4. Model fit comparisons for linear versus non-linear distance functions eTable 5. Risk of non-affective psychoses after mutual adjustment for moves in all periods Click here for additional data file.

Introduction Psychiatric disorders are impairing and associated with genetic factors.1 Although clinical practice and most genetic studies follow a case-control conceptualization of these disorders, subclinical traits of these disorders are common among unaffected individuals and are as heritable as the disorders themselves based on twin methods.2 One common theory is that the genetic risks for psychiatric disorders are associated with these milder traits and that psychiatric disorders arise after particularly strong exposure to the same genetic risks associated with these traits.2 Preliminary support for this theory comes from twin and molecular genetic studies. Twin studies indicate consistent heritability of traits at varying levels of severity3 for autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), and learning difficulties.4 A more recent UK-based twin study5 that used a contemporary analytic approach reported a genetic correlation of 0.70 between ASD and autistic traits. To our knowledge, this approach has not been applied to other psychiatric phenotypes, meaning that the degree of genetic correlation between most psychiatric disorders and related traits has been largely understudied using twin methods. Few twin studies have focused on the links between anxiety and major depressive disorder (MDD) with the traits of these disorders despite these being 2 of the most common psychiatric diagnoses.

of genetic correlation between most psychiatric disorders and related traits has been largely understudied using twin methods. Few twin studies have focused on the links between anxiety and major depressive disorder (MDD) with the traits of these disorders despite these being 2 of the most common psychiatric diagnoses. Molecular genetic methods also allow estimation of genetic correlations between psychiatric disorders and subclinical traits of these disorders based on additive, common genome-wide variants. Such studies report strong genetic correlations across disorders and traits for ADHD and MDD, with a moderate estimate for ASD.4 This approach requires large genome-wide data sets, however, which are lacking for psychiatric traits beyond symptoms of ADHD, MDD, and ASD.6,7 An alternative approach is to calculate polygenic risk scores (PRSs) based on discovery genome-wide association studies (GWASs) of psychiatric disorders; PRSs capture an individual’s common variant risk for a phenotype.8 A review4 of preliminary studies reported that psychiatric disorder PRSs are associated with corresponding population traits of ASD, ADHD, obsessive-compulsive disorder (OCD), and MDD, with null or mixed findings for schizophrenia and bipolar disorder (BD).

capture an individual’s common variant risk for a phenotype.8 A review4 of preliminary studies reported that psychiatric disorder PRSs are associated with corresponding population traits of ASD, ADHD, obsessive-compulsive disorder (OCD), and MDD, with null or mixed findings for schizophrenia and bipolar disorder (BD). Although some preliminary evidence supports shared genetic risks across disorders and traits for some psychiatric phenotypes, evidence is weak, mixed, or entirely lacking for many phenotypes. We aimed to assess the degree of shared genetic risks between disorders and traits for multiple psychiatric disorders. Leveraging data from a unique twin sample, with comprehensive clinical diagnostic, trait measurement, and genetic data available, enabled us to perform twin modeling and molecular genetic methods in the same sample. We used a novel twin method5 to estimate the genetic correlation between psychiatric diagnoses and continuous traits of these disorders. We then calculated PRSs based on recent, large-scale GWASs and tested their associations with continuous variation in related phenotypes.

et depression trajectory. Phenotypic childhood neurodevelopmental problems were markedly increased in the early-adolescence–onset group (by 5- to 7-fold) compared with the typical depression trajectory. Studies with follow-up further into adult life will help to clarify the adult mental health outcomes of these groups. Strengths and Limitations Strengths of this study include the repeated-measures longitudinal design where depression was assessed using the same measure and informant. Typically, longitudinal studies include changes in measurement and informant, in particular, as children age, the informant tends to change from the parent to the young person. This variation provides a challenge to studies seeking to examine the development of symptoms over time because changes of measurement and informant can affect results. This invariance of measurement over time is a strength.

molecular genetic methods in the same sample. We used a novel twin method5 to estimate the genetic correlation between psychiatric diagnoses and continuous traits of these disorders. We then calculated PRSs based on recent, large-scale GWASs and tested their associations with continuous variation in related phenotypes. Methods Participants Families of all twins born in Sweden beginning in 1992 were contacted in connection with the twins’ ninth birthday (earlier cohorts included individuals aged 12 years) and invited to participate in the Child and Adolescent Twin Study in Sweden (CATSS).9 The response rate was 75%. Follow-ups were conducted when the twins were 15 years of age (response rate, 61%) and 18 years of age (response rate, 59%). Exclusion criteria were brain injuries (n = 207 pairs), chromosomal syndromes (n = 35 pairs), death (n = 29 pairs), and migration (n = 100 pairs). Phenotypic data were available for 13 923 pairs at 9 years of age (1983 monozygotic male [MZM], 2641 dizygotic male [DZM], 2108 monozygotic female [MZF], 2304 dizygotic female [DZF], and 4887 dizygotic opposite-sex [DZOS]), 5165 pairs at 15 years of age (649 MZM, 854 DZM, 831 MZF, 924 DZF, and 1907 DZOS), and 4273 pairs at 18 years of age (553 MZM, 693 DZM, 722 MZF, 747 DZF, and 1558 DZOS). Zygosity was ascertained using a panel of 48 single-nucleotide polymorphisms (SNPs) or 5 questions concerning twin similarity. The latter method was only used in cases with a 95% probability of correct classification. Zygosity was reconfirmed for pairs with genotype data. All families provided written informed consent before participation, and all data were deidentified. This study received ethical approval from the Karolinska Institutet Ethical Review Board.

d was only used in cases with a 95% probability of correct classification. Zygosity was reconfirmed for pairs with genotype data. All families provided written informed consent before participation, and all data were deidentified. This study received ethical approval from the Karolinska Institutet Ethical Review Board. DNA samples (from saliva) were obtained from the CATSS participants at study enrollment. A total of 11 551 individuals with available DNA were genotyped using the Illumina PsychChip. Standard quality control and imputation procedures were performed in the sample; for details, see the article by Brikell et al.10 A total of 11 081 samples passed quality control assessment; MZ twins were then imputed, resulting in 13 576 samples and 6 981 993 imputed SNPs that passed all quality control assessments. After individual-level exclusions (described above), 13 412 children (50.2% females) were included in genetic analyses.

ll et al.10 A total of 11 081 samples passed quality control assessment; MZ twins were then imputed, resulting in 13 576 samples and 6 981 993 imputed SNPs that passed all quality control assessments. After individual-level exclusions (described above), 13 412 children (50.2% females) were included in genetic analyses. Phenotypic Measures Clinical Diagnoses The CATSS is linked with the Swedish National Patient Register (NPR).11 The NPR contains International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes for diagnoses from all visits to specialist inpatient and outpatient care in Sweden. Inpatient data were available for January 1, 1987, to December 31, 2014, and outpatient data for January 1, 2001, to December 31, 2013. Diagnoses of ASD, ADHD, intellectual disability (ID), tic disorders (TDs), OCD, anxiety disorders (ADs), and MDD were extracted. Diagnostic codes and the numbers of individuals with each diagnosis are given in Table 1. At the end of follow-up (December 31, 2014), the individuals in this study were between 9 and 22 years of age. Data analysis was performed from January 1, 2017, to September 30, 2017. Table 1. Description of the CATSS Sample and Measuresa Phenotype NPR Diagnosis Study Measures ICD-10 Diagnostic Codes Affected, No. (%) Description of Measure No.

Phenotypic Measures Clinical Diagnoses The CATSS is linked with the Swedish National Patient Register (NPR).11 The NPR contains International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes for diagnoses from all visits to specialist inpatient and outpatient care in Sweden. Inpatient data were available for January 1, 1987, to December 31, 2014, and outpatient data for January 1, 2001, to December 31, 2013. Diagnoses of ASD, ADHD, intellectual disability (ID), tic disorders (TDs), OCD, anxiety disorders (ADs), and MDD were extracted. Diagnostic codes and the numbers of individuals with each diagnosis are given in Table 1. At the end of follow-up (December 31, 2014), the individuals in this study were between 9 and 22 years of age. Data analysis was performed from January 1, 2017, to September 30, 2017. Table 1. Description of the CATSS Sample and Measuresa Phenotype NPR Diagnosis Study Measures ICD-10 Diagnostic Codes Affected, No. (%) Description of Measure No. of Individuals With Data Twin Analyses Genetic Analyses Twin Analyses Genetic Analyses ASD F84 253 (0.9) 142 (1.1) Ages of 9 and 12 y: parent-rated A-TAC ASD module (17 items) 27 780 13 396 ADHD F90 824 (3.0) 440 (3.3) Ages of 9 and 12 y: parent-rated A-TAC ADHD module (19 items) 27 759 13 391 ID F70-F73 166 (0.6) 77 (0.6) Ages of 9 and 12 y: parent-rated A-TAC learning module (3 items) 27 804 13 400 TDs F95 103 (0.4) 43 (0.3) Ages of 9 and 12 y: parent-rated A-TAC tics module (3 items) 27 791 13 396 OCD F42 118 (0.4) 67 (0.5) Ages of 9 and 12 y: parent-rated A-TAC compulsions module (2 items) 27 802 13 400 Age of 18 y: self-rated BOCS (15 items) 5757 3982 ADs F40-F41 474 (1.7) 251 (1.9) Ages of 9 and 12 y: parent-rated SCARED (41 items) 15 589 6806 Age of 15 y: parent-rated SDQ-E (5 items) 7663 5703 Age of 15 y: self-rated SDQ-E (5 items) 8150 5917 Age of 18 y: parent-rated ABCL DSM-IV anxiety subscale (6 items) 5160 3755 Self-rated SCARED (38 items) 5801 4007 MDD F32-F34 414 (1.5) 222 (1.7) Ages of 9 and 12 y: parent-rated SMFQ (13 items) 15 873 6826 Age of 15 y: parent-rated SDQ-E (5 items); 7663 5703 Age of 15 y: self-rated SDQ-E (5 items); 8150 5917 Age of 18 y: parent-rated ABCL DSM-IV depression subscale (15 items); 5215 3791 Self-rated CES-D (11 items) 5518 3813 Mania NA NA NA Age of 18 y: parent-rated MDQ (13 items) NA 3808 Age of 18 y: self-rated MDQ (13 items) NA 4128 Psychosis NA NA NA Age of 18 y: parent-rated APSS (7 items) NA 5368 Age of 18 y: self-rated APSS (7 items) NA 5518 Abbreviations: ABCL, Adult Behavior Checklist12; ADs, anxiety disorders; ADHD, attention-deficit/hyperactivity disorder; APSS, Adolescent Psychotic-Like Symptom Screener13; ASD, autism spectrum disorder; A-TAC, Autism-Tics, AD/HD, and Other Comorbidities Inventory14,15; BOCS, Brief Obsessive-Compulsive Scale16; CATSS, Child and Adolescent Twin Study in Sweden; CES-D, Center for Epidemiologic Studies Depression Scale17; ICD-10, International Statistical Classification of Diseases and Related Health Problems, Tenth Revision; ID, intellectual disability; MDD, major depressive disorder; MDQ, Mood Disorder Questionnaire18; NA, not applicable; NPR, National Patient Register; OCD,

CES-D, Center for Epidemiologic Studies Depression Scale17; ICD-10, International Statistical Classification of Diseases and Related Health Problems, Tenth Revision; ID, intellectual disability; MDD, major depressive disorder; MDQ, Mood Disorder Questionnaire18; NA, not applicable; NPR, National Patient Register; OCD, obsessive-compulsive disorder; SCARED, Screen for Child Anxiety Related Emotional Disorders19; SDQ-E, Strengths and Difficulties Questionnaire20; SMFQ, Short Mood and Feelings Questionnaire21; TDs, tic disorders. a Diagnoses of bipolar disorder and schizophrenia were not included because of small numbers of participants with these diagnoses. There were too few items to divide the SDQ-E subscale into anxiety and depression separately. Continuous Measures Traits of ASD, ADHD, ID, TDs, OCD, ADs, and MDD were measured using continuous scales at 9 and 12 years of age. Internalizing problems (related to ADs and MDD) were then measured at 15 years of age. Traits of OCD, ADs, MDD, mania, and psychotic-like experiences were assessed at 18 years of age. Details of these measures and sample sizes are provided in Table 1, with additional details provided in eTable 1 in the Supplement.

age. Internalizing problems (related to ADs and MDD) were then measured at 15 years of age. Traits of OCD, ADs, MDD, mania, and psychotic-like experiences were assessed at 18 years of age. Details of these measures and sample sizes are provided in Table 1, with additional details provided in eTable 1 in the Supplement. Analyses Twin Analyses We used joint categorical/continuous twin models to estimate the degree of etiologic overlap between continuous traits and categorical diagnoses. These models assume a normal distribution of continuous liability underlying psychiatric disorders, whereas questionnaires were treated continuously. The model partitions variance in each phenotype into additive genetic (A), nonadditive genetic (D), shared environmental (C), and nonshared environmental (E, which encompasses measurement error) components. The correlations among these components are then estimated between 2 phenotypes. The phenotypic correlations were decomposed into genetic and environmental factors to assess which factors explain the correlation between psychiatric diagnoses and traits. On the basis of twin correlations, we tested ACE or ADE models for each disorder-trait pairing. We included a sibling interaction term in ADE models (ADE-s) because these interactions can mimic the effects of D on the twin correlations.22 The principles of the twin design are described at length elsewhere.23

nd traits. On the basis of twin correlations, we tested ACE or ADE models for each disorder-trait pairing. We included a sibling interaction term in ADE models (ADE-s) because these interactions can mimic the effects of D on the twin correlations.22 The principles of the twin design are described at length elsewhere.23 The ACE or ADE-s model was compared with a saturated model of the observed data. If the pattern of twin correlations differed between the continuous trait and the categorical diagnoses, both models were fitted and the best-fitting model was chosen on the basis of the lowest Bayesian Information Criteria value. More parsimonious models were tested by reducing each model by constraining certain components to equal zero and comparing these models to the ACE or ADE-s model using the likelihood ratio test; if the model fit did not deteriorate significantly, the reduced model was favored. Continuous scales were standardized by sex, whereas the association of sex with the thresholds were included in the models. Models were fitted in OpenMx.24 Opposite-sex twins were included, but the study was underpowered to test for sex differences. Obsessive-compulsive disorder was omitted from the twin analyses because of the small sample and low heritability.

whereas the association of sex with the thresholds were included in the models. Models were fitted in OpenMx.24 Opposite-sex twins were included, but the study was underpowered to test for sex differences. Obsessive-compulsive disorder was omitted from the twin analyses because of the small sample and low heritability. PRS Analyses Publicly available GWAS summary statistics for 8 psychiatric disorders (ie, ASD, ADHD, TDs, OCD, ADs, MDD, BD, and schizophrenia) and 3 continuously distributed psychiatric or cognitive traits (ie, ADHD symptoms, cognitive ability, and depressive symptoms) were used to derive PRSs in the CATSS individuals.25,26,27,28,29,30,31,32,33,34 eTable 2 in the Supplement lists these discovery data sets along with sample sizes. Discovery and target data were independent (details of PRS calculations are given in the eAppendix in the Supplement). In brief, PRSs were calculated in imputed CATSS data for each individual by scoring the number of effect alleles (weighted by the SNP effect size) across each discovery set of clumped SNPs in PLINK, version 1.9, for a range of P value thresholds used for SNP selection. The primary analyses are based on the threshold P < .50. The PRSs were standardized using z-score transformations; effect sizes can be interpreted as increase in risk of the outcome per SD increase in PRS. Principal component analysis was used to derive covariates to account for population stratification (eAppendix in the Supplement).

yses are based on the threshold P < .50. The PRSs were standardized using z-score transformations; effect sizes can be interpreted as increase in risk of the outcome per SD increase in PRS. Principal component analysis was used to derive covariates to account for population stratification (eAppendix in the Supplement). Analyses of PRSs were performed using generalized estimating equations using the R package drgee, with robust SEs, based on clustering related individuals to account for twins in the data. The principal components, sex, and age (for measures that were assessed at 9 or 12 years of age) were included as covariates. First, we tested for association between PRSs for each of the 8 discovery GWASs of psychiatric disorders and the corresponding continuously distributed trait(s). Second, these analyses were repeated after excluding individuals diagnosed with the relevant psychiatric disorder based on available information on ICD-10 diagnoses to determine whether effects were driven primarily by individuals with clinically recognized problems. Third, we tested for associations between PRSs for each of the 3 discovery GWASs of continuously distributed population traits and the corresponding psychiatric diagnosis in the target sample. Fourth, all PRS analyses were repeated using PRSs derived on the basis of different P value selection thresholds to assess sensitivity. False discovery rate corrections were applied in R (using the fdr method in the function p.adjust) (R Foundation for Statistical Computing) to account for multiple testing.

get sample. Fourth, all PRS analyses were repeated using PRSs derived on the basis of different P value selection thresholds to assess sensitivity. False discovery rate corrections were applied in R (using the fdr method in the function p.adjust) (R Foundation for Statistical Computing) to account for multiple testing. Supplemental Analyses Because there is some evidence of genetic specificity within ASD and ADHD trait domains,35,36 we reran all twin and PRS analyses for 3 specific DSM-IV ASD dimensions (social problems, language impairment, and behavioral inflexibility) and 2 DSM-IV ADHD dimensions (hyperactivity/impulsivity and inattention). These domains were assessed by dividing the Autism-Tics, AD/HD, and Other Comorbidities Inventory ASD and ADHD subscales based on prior work.9

r 3 specific DSM-IV ASD dimensions (social problems, language impairment, and behavioral inflexibility) and 2 DSM-IV ADHD dimensions (hyperactivity/impulsivity and inattention). These domains were assessed by dividing the Autism-Tics, AD/HD, and Other Comorbidities Inventory ASD and ADHD subscales based on prior work.9 Results Twin Analyses Phenotype data were available for 13 923 twin pairs (35.1% opposite sex and 31.7% same-sex females) at 9 years of age, 5165 pairs (36.9% opposite sex and 34.0% same-sex females) at 15 years of age, and 4273 pairs (36.5% opposite sex and 34.4% same-sex females) at 18 years of age. Genetic data were available for 13 412 individuals (50.2% females). Probandwise concordances for each diagnosis, which represent the probability of co-twins of probands also receiving a given diagnosis, are given in eTable 3 in the Supplement. The MZ probandwise concordances all exceeded the DZ estimates, indicating genetic associations with each diagnosis. Phenotypic correlations between disorders and related traits ranged from 0.22 for MDD to 0.66 for ID (mean estimate, 0.40) (Table 2). The MZ twin correlations were higher than the DZ correlations within each disorder and trait and across disorder-trait pairs, suggesting that each phenotype and the covariance between was associated with genetic factors. Table 2.

Results Twin Analyses Phenotype data were available for 13 923 twin pairs (35.1% opposite sex and 31.7% same-sex females) at 9 years of age, 5165 pairs (36.9% opposite sex and 34.0% same-sex females) at 15 years of age, and 4273 pairs (36.5% opposite sex and 34.4% same-sex females) at 18 years of age. Genetic data were available for 13 412 individuals (50.2% females). Probandwise concordances for each diagnosis, which represent the probability of co-twins of probands also receiving a given diagnosis, are given in eTable 3 in the Supplement. The MZ probandwise concordances all exceeded the DZ estimates, indicating genetic associations with each diagnosis. Phenotypic correlations between disorders and related traits ranged from 0.22 for MDD to 0.66 for ID (mean estimate, 0.40) (Table 2). The MZ twin correlations were higher than the DZ correlations within each disorder and trait and across disorder-trait pairs, suggesting that each phenotype and the covariance between was associated with genetic factors. Table 2. Phenotypic and Twin Correlations Across Disorders and Continuous Traitsa Disorder, Outcome (Age, y) Correlation Coefficient (95% CI) rPH Continuous Scale Categorical Diagnosis Cross-Trait MZ DZ MZ DZ MZ DZ ASD A-TAC ASD (9 and 12) 0.45 (0.42 to 0.48) 0.74 (0.72 to 0.75) 0.27 (0.25 to 0.28) 0.81 (0.67 to 0.90) 0.31 (0.18 to 0.43) 0.39 (0.33 to 0.45) 0.14 (0.08 to 0.21) ADHD A-TAC ADHD (9 and 12) 0.52 (0.50 to 0.54) 0.69 (0.68 to 0.71) 0.23 (0.21 to 0.25) 0.88 (0.83 to 0.92) 0.44 (0.37 to 0.50) 0.47 (0.43 to 0.51) 0.17 (0.13 to 0.21) ID A-TAC learning (9 and 12) 0.66 (0.63 to 0.69) 0.72 (0.70 to 0.73) 0.13 (0.11 to 0.15) 0.93 (0.84 to 0.98) 0.30 (0.15 to 0.44) 0.54 (0.45 to 0.62) 0.12 (0.03 to 0.21) TDs A-TAC tics (9 and 12) 0.48 (0.44 to 0.52) 0.44 (0.41 to 0.46) 0.11 (0.09 to 0.13) 0.64 (0.39 to 0.82) −0.26 (−0.68 to 0.28) 0.32 (0.20 to 0.43) 0.09 (−0.02 to 0.20) ADs SCARED (9) 0.30 (0.24 to 0.36) 0.66 (0.64 to 0.68) 0.37 (0.35 to 0.39) 0.67 (0.55 to 0.77) 0.30 (0.18 to 0.42) 0.39 (0.22 to 0.52) 0.14 (0.04 to 0.25) SDQ-E parent-rated (15) 0.36 (0.32 to 0.41) 0.48 (0.44 to 0.52) 0.22 (0.18 to 0.25) 0.66 (0.53 to 0.76) 0.30 (0.18 to 0.41) 0.33 (0.23 to 0.43) 0.15 (0.06 to 0.23) SDQ-E self-rated (15) 0.26 (0.21 to 0.31) 0.44 (0.39 to 0.48) 0.18 (0.15 to 0.22) 0.66 (0.53 to 0.76) 0.30 (0.18 to 0.42) 0.23 (0.11 to 0.35) 0.07 (−0.02 to 0.16) ABCL anxiety (18) 0.40 (0.35 to 0.45) 0.57 (0.53 to 0.61) 0.32 (0.28 to 0.36) 0.65 (0.52 to 0.76) 0.30 (0.18 to 0.41) 0.30 (0.19 to 0.40) 0.16 (0.06 to 0.25) SCARED (18) 0.42 (0.36 to 0.47) 0.49 (0.44 to 0.54) 0.17 (0.12 to 0.22) 0.65 (0.53 to 0.76) 0.30 (0.18 to 0.41) 0.42 (0.30 to 0.52) 0.17 (0.08 to 0.27) MDD SMFQ (9) 0.22 (0.13 to 0.30) 0.55 (0.52 to 0.57) 0.29 (0.27 to 0.31) 0.68 (0.56 to 0.77) 0.37 (0.24 to 0.48) 0.56 (0.33 to 0.70) 0.03 (−0.10 to 0.17) SDQ-E parent-rated (15) 0.37 (0.32 to 0.42) 0.48 (0.44 to 0.52) 0.22 (0.18 to 0.25) 0.69 (0.57 to 0.78) 0.34 (0.21 to 0.46) 0.22 (0.12 to 0.32) 0.15 (0.06 to 0.24) SDQ-E self-rated (15) 0.28 (0.22 to 0.33) 0.45 (0.40 to 0.49) 0.18 (0.15 to 0.22) 0.68 (0.56 to 0.77) 0.36 (0.23 to 0.48) 0.30 (0.18 to 0.41) 0.09 (−0.02 to 0.19) ABCL depression (18) 0.44 (0.40 to 0.49) 0.52 (0.48

0.52) 0.22 (0.18 to 0.25) 0.69 (0.57 to 0.78) 0.34 (0.21 to 0.46) 0.22 (0.12 to 0.32) 0.15 (0.06 to 0.24) SDQ-E self-rated (15) 0.28 (0.22 to 0.33) 0.45 (0.40 to 0.49) 0.18 (0.15 to 0.22) 0.68 (0.56 to 0.77) 0.36 (0.23 to 0.48) 0.30 (0.18 to 0.41) 0.09 (−0.02 to 0.19) ABCL depression (18) 0.44 (0.40 to 0.49) 0.52 (0.48 to 0.56) 0.26 (0.21 to 0.30) 0.66 (0.54 to 0.77) 0.35 (0.22 to 0.47) 0.34 (0.25 to 0.43) 0.14 (0.04 to 0.23) CES-D (18) 0.44 (0.39 to 0.49) 0.47 (0.41 to 0.51) 0.19 (0.14 to 0.23) 0.68 (0.56 to 0.78) 0.36 (0.23 to 0.47) 0.32 (0.20 to 0.44) 0.17 (0.07 to 0.26) Abbreviations: ABCL, Adult Behavior Checklist; ADs, anxiety disorders; ASD, autism spectrum disorder; ADHD, attention-deficit/hyperactivity disorder; A-TAC, Autism-Tics, AD/HD, and Other Comorbidities Inventory; CES-D, Center for Epidemiologic Studies Depression Scale; DZ, dizygotic; ID, intellectual disability; MDD, major depressive disorder; MZ, monozygotic; rPH, phenotypic correlation between continuous and categorical diagnosis; SCARED, Screen for Child Anxiety Related Emotional Disorders; SDQ-E, Strengths and Difficulties Questionnaire, emotional problems subscale; SMFQ, Short Mood and Feelings Questionnaire; TDs, tic disorders. a The table gives the cross-twin correlations for the continuous scale and categorical diagnosis followed by the cross-trait cross-twin correlations, with ranges in parentheses. All correlations given in the table were estimated from a saturated model with constraints on the means and variances for continuous traits and on the thresholds for categorical diagnoses.

for the continuous scale and categorical diagnosis followed by the cross-trait cross-twin correlations, with ranges in parentheses. All correlations given in the table were estimated from a saturated model with constraints on the means and variances for continuous traits and on the thresholds for categorical diagnoses. For all traits, AE-s or AE models fit best except for ADs at 9 and 12 years of age, where there was a small, significant C estimate (eTable 4 in the Supplement). Variance components and etiologic correlations from each model are given in eTable 5 in the Supplement. All measures and diagnoses except for the Strengths and Difficulties Questionnaire at 15 years of age and self-reported MDD at 18 years of age were under strong genetic influence. The genetic and nonshared environmental correlations are shown in the Figure. Point estimates of genetic correlation varied: 0.48 for ASD (95% CI, 0.44-0.53), 0.56 for ADHD (95% CI, 0.53-0.59), 0.69 for ID (95% CI, 0.64-0.73), 0.61 for TDs (95% CI, 0.51-0.80), 0.46 to 0.57 for ADs (95% CI, 0.34-0.58 and 0.46-0.68), and 0.33 to 0.58 for MDD (95% CI, 0.19-0.47 and 0.52-0.63). For anxiety and MDD, higher genetic correlations were estimated at 18 years of age. However, most of the CIs around the genetic correlations overlapped. Shared genetic factors explained 60% to 100% (mean estimate, 80.7%) of the phenotypic covariance between each trait and diagnosis (eTable 5 in the Supplement).

-0.63). For anxiety and MDD, higher genetic correlations were estimated at 18 years of age. However, most of the CIs around the genetic correlations overlapped. Shared genetic factors explained 60% to 100% (mean estimate, 80.7%) of the phenotypic covariance between each trait and diagnosis (eTable 5 in the Supplement). Figure. Genetic and Nonshared Environmental Correlations Between Psychiatric Diagnoses and Traits From the Best-Fitting Twin Models ABCL, Adult Behavior Checklist; ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; A-TAC: Autism-Tics, AD/HD, and Other Comorbidities Inventory; CES-D, Center for Epidemiologic Studies Depression Scale; MDD, major depressive disorder; SCARED, Screen for Child Anxiety Related Emotional Disorders; SDQ-E, Strengths and Difficulties Questionnaire, emotional problems subscale; SMFQ, Short Mood and Feelings Questionnaire. Analyses of specific ASD and ADHD domains are given in eTables 6 and 7 in the Supplement. All ASD trait domains displayed moderate phenotypic (mean estimate, 0.41; range, 0.43-0.48) and genetic correlations (mean estimate, 0.46; range, 0.43-0.47) with ASD, whereas both ADHD dimensions displayed moderate phenotypic (mean estimate, 0.49; range, 0.45-0.53) and genetic (mean estimate, 0.53; range, 0.49-0.57) correlations with ADHD.

displayed moderate phenotypic (mean estimate, 0.41; range, 0.43-0.48) and genetic correlations (mean estimate, 0.46; range, 0.43-0.47) with ASD, whereas both ADHD dimensions displayed moderate phenotypic (mean estimate, 0.49; range, 0.45-0.53) and genetic (mean estimate, 0.53; range, 0.49-0.57) correlations with ADHD. PRS Analyses The PRSs for psychiatric disorders were associated with related traits for all disorders except BD (Table 3). At 9 years of age, ASD PRSs were associated with autistic traits (β [SE] = 0.04 [0.01]), ADHD PRSs were associated with ADHD traits (β [SE] = 0.27 [0.03]), and TD PRSs were associated with tic problems (β [SE] = 0.02 [0.004]). The OCD PRSs were associated with obsessive-compulsive symptoms at 18 years of age (β [SE] = 0.13 [0.05]) but not at 9 and 12 years of age (β [SE] = 0.002 [0.002]). The AD PRSs were associated with anxiety traits at 9 years of age (β [SE] = 0.18 [0.08]), parent-rated internalizing traits at 15 years of age (β [SE] = 0.07 [0.02]), and self-rated traits at 18 years of age (β [SE] = 0.40 [0.17]) but not with self-rated internalizing traits at 15 years of age (β [SE] = 0.06 [0.03]) and parent-rated symptoms at 18 years of age (β [SE] = 0.04 [0.03]). The MDD PRSs were associated with all measures of depressive symptoms (β [SE] = 0.10 [0.03] at 9 years of age, β [SE] = 0.11 [0.02] at 15 years of age parent-rated, β [SE] = 0.11 [0.03] at 15 years of age self-rated, β [SE] = 0.25 [0.06] at 18 years of age parent-rated; β [SE] = 0.41 [0.10] at 18 years of age self-rated). Schizophrenia PRSs were associated with psychotic traits at 18 years of age (β [SE] = 0.02 [0.01]).

of age, β [SE] = 0.11 [0.02] at 15 years of age parent-rated, β [SE] = 0.11 [0.03] at 15 years of age self-rated, β [SE] = 0.25 [0.06] at 18 years of age parent-rated; β [SE] = 0.41 [0.10] at 18 years of age self-rated). Schizophrenia PRSs were associated with psychotic traits at 18 years of age (β [SE] = 0.02 [0.01]). Table 3.

of age, β [SE] = 0.11 [0.02] at 15 years of age parent-rated, β [SE] = 0.11 [0.03] at 15 years of age self-rated, β [SE] = 0.25 [0.06] at 18 years of age parent-rated; β [SE] = 0.41 [0.10] at 18 years of age self-rated). Schizophrenia PRSs were associated with psychotic traits at 18 years of age (β [SE] = 0.02 [0.01]). Table 3. Association of PRSs With Related Continuous Outcomes Discovery PRS, Outcome (Age, y) Full Sample Excluding Those With ICD-10 Diagnoses β (SE) P Value R2 β (SE) P Value R2 ASD A-TAC ASD (9 and 12) 0.043 (0.014) 5.4 × 10−3a 9.5 × 10−4 0.036 (0.013) 6.7 × 10−3a 8.2 × 10−4 ADHD A-TAC ADHD (9 and 12) 0.268 (0.029) 5.9 × 10−19b 8.4 × 10−3 0.205 (0.027) 2.2 × 10−13b 6.2 × 10−3 TDs A-TAC tics (9 and 12) 0.015 (0.004) 6.6 × 10−4b 1.2 × 10−3 0.016 (0.004) 5.3 × 10−4b 1.3 × 10−3 OCD A-TAC OC traits (9 and 12) 0.002 (0.002) 0.333 1.2 × 10−4 0.002 (0.002) 0.262 1.4 × 10−4 BOCS (18) 0.126 (0.047) 0.014c 2.3 × 10−3 0.132 (0.046) 6.7 × 10−3a 2.6 × 10−3 ADs SCARED (9) 0.180 (0.078) 0.033c 9.1 × 10−4 0.186 (0.078) 0.023c 1.0 × 10−3 SDQ-E parent-rated (15) 0.069 (0.023) 5.4 × 10−3a 1.9 × 10−3 0.054 (0.022) 0.018c 1.3 × 10−3 SDQ-E self-rated (15) 0.061 (0.030) 0.060 7.3 × 10−4 0.055 (0.030) 0.077 6.1 × 10−4 ABCL anxiety (18) 0.043 (0.030) 0.186 6.6 × 10−4 0.032 (0.029) 0.262 4.2 × 10−4 SCARED (18) 0.404 (0.171) 0.031c 1.5 × 10−3 0.344 (0.166) 0.047c 1.2 × 10−3 MDD SMFQ (9) 0.095 (0.028) 2.1 × 10−3a 2.0 × 10−3 0.095 (0.028) 1.7 × 10−3a 2.0 × 10−3 SDQ-E parent-rated (15) 0.105 (0.023) 5.3 × 10−5b 4.4 × 10−3 0.079 (0.022) 1.3 × 10−3a 2.7 × 10−3 SDQ-E self-rated (15) 0.105 (0.030) 1.3 × 10−3a 2.1 × 10−3 0.087 (0.029) 6.5 × 10−3a 1.5 × 10−3 ABCL depression (18) 0.249 (0.055) 5.3 × 10−5b 7.3 × 10−3 0.186 (0.049) 6.3 × 10−4b 5.1 × 10−3 CES-D (18) 0.408 (0.100) 2.3 × 10−4b 4.8 × 10−3 0.343 (0.097) 1.3 × 10−3a 3.6 × 10−3 BD MDQ parent-rated (18) -0.045 (0.061) 0.478 2.3 × 10−4 NAd NAd NAd MDQ self-rated (18) -0.054 (0.058) 0.388 2.4 × 10−4 NAd NAd NAd Schizophrenia APSS parent-rated (18) 0.024 (0.010) 0.031c 1.5 × 10−3 NAd NAd NAd APSS self-rated (18) 0.062 (0.021) 8.4 × 10−3a 1.7 × 10−3 NAd NAd NAd Abbreviations: ABCL, Adult Behavior Checklist; ADs, anxiety disorders; ADHD, attention-deficit/hyperactivity disorder; APSS, Adolescent Psychotic-Like Symptom Screener; ASD, autism spectrum disorder; A-TAC, Autism-Tics, AD/HD, and Other Comorbidities Inventory; BD, bipolar disorder; BOCS, Brief Obsessive-Compulsive Scale; CES-D, Center for Epidemiologic Studies Depression Scale; ICD-10, Internat

ttention-deficit/hyperactivity disorder; APSS, Adolescent Psychotic-Like Symptom Screener; ASD, autism spectrum disorder; A-TAC, Autism-Tics, AD/HD, and Other Comorbidities Inventory; BD, bipolar disorder; BOCS, Brief Obsessive-Compulsive Scale; CES-D, Center for Epidemiologic Studies Depression Scale; ICD-10, Internat ional Statistical Classification of Diseases and Related Health Problems, Tenth Revision; MDD, major depressive disorder; MDQ, Mood Disorder Questionnaire; NA, not applicable; OCD, obsessive-compulsive disorder; PRS, polygenic risk score; SCARED, Screen for Child Anxiety Related Emotional Disorders; SDQ-E, Strengths and Difficulties Questionnaire, emotional problems subscale; SMFQ, Short Mood and Feelings Questionnaire; TDs, tic disorders. a False discovery rate P < .01. b False discovery rate P < .001. c False discovery rate P < .05. d Sample too young to exclude individuals diagnosed with BD or schizophrenia. The PRSs were derived using common variants with P < .50 in the discovery data. After removing individuals diagnosed with the relevant psychiatric disorder, the results remained significant for ASD, ADHD, TDs, OCD, ADs, and MDD, although the effect sizes decreased (Table 3). All estimates of variance explained were modest (mean, 0.23%; range, 0.01%-0.84%).

d Sample too young to exclude individuals diagnosed with BD or schizophrenia. The PRSs were derived using common variants with P < .50 in the discovery data. After removing individuals diagnosed with the relevant psychiatric disorder, the results remained significant for ASD, ADHD, TDs, OCD, ADs, and MDD, although the effect sizes decreased (Table 3). All estimates of variance explained were modest (mean, 0.23%; range, 0.01%-0.84%). In the analysis of PRSs for quantitative traits associated with psychiatric diagnoses (Table 4), PRSs for depressive symptoms were associated with MDD diagnosis (odds ratio [OR], 1.16; 95% CI, 1.02-1.32); PRSs for traits of ADHD (OR, 1.09; 95% CI, 0.98-1.22) and general cognitive ability (OR, 0.97; 95% CI, 0.76-1.23) were not associated with related diagnoses (ADHD and ID, respectively). Table 4. Association of Continuous Trait PRSs With Related Binary Outcomes Discovery PRS Outcome (Diagnosis) No. of Individuals OR (95% CI) P Value R2 ADHD traits ICD-10 (ADHD) 13 412 1.09 (0.98-1.22) .13 9.9 × 10−4 General cognition ICD-10 (ID) 13 412 0.97 (0.76-1.23) .78 9.6 × 10−5 Depressive traits ICD-10 (MDD) 13 412 1.16 (1.02-1.32) .04a 2.3 × 10−3 Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ICD-10, International Statistical Classification of Diseases and Related Health Problems, Tenth Revision; ID, intellectual disability; MDD, major depressive disorder; OR, odds ratio; PRS, polygenic risk score. a False discovery rate P < .05.

Table 4. Association of Continuous Trait PRSs With Related Binary Outcomes Discovery PRS Outcome (Diagnosis) No. of Individuals OR (95% CI) P Value R2 ADHD traits ICD-10 (ADHD) 13 412 1.09 (0.98-1.22) .13 9.9 × 10−4 General cognition ICD-10 (ID) 13 412 0.97 (0.76-1.23) .78 9.6 × 10−5 Depressive traits ICD-10 (MDD) 13 412 1.16 (1.02-1.32) .04a 2.3 × 10−3 Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ICD-10, International Statistical Classification of Diseases and Related Health Problems, Tenth Revision; ID, intellectual disability; MDD, major depressive disorder; OR, odds ratio; PRS, polygenic risk score. a False discovery rate P < .05. Secondary analyses for ASD and ADHD PRSs associated with specific trait domains are given in eTable 8 in the Supplement. The analyses of these subdomains were consistent with the analyses of total ASD (social: β [SE] = 0.011 [0.006], language: β [SE] = 0.011 [0.006], flexibility: β [SE] = 0.022 [0.005]) and ADHD (hyperactivity/impulsivity: β [SE] = 0.14 [0.015], inattention: β [SE] = 0.130 [0.016]) trait scores except that, for ASD, only the estimate for flexibility remained significant after excluding individuals with ASD diagnoses. All analyses were repeated using PRSs derived on the basis of different P value thresholds for SNP inclusion (eFigure in the Supplement). The pattern of results was consistent in these sensitivity analyses, but 1 result (ie, ADs at 18 years of age) was not statistically significant after false discovery rate correction.

All analyses were repeated using PRSs derived on the basis of different P value thresholds for SNP inclusion (eFigure in the Supplement). The pattern of results was consistent in these sensitivity analyses, but 1 result (ie, ADs at 18 years of age) was not statistically significant after false discovery rate correction. Discussion Using a unique, large, genotyped twin sample, we tested for genetic associations between clinical psychiatric diagnoses and related traits. All disorders analyzed using novel twin models (ASD, ADHD, TDs, ID, ADs, and MDD) showed modest to strong genetic correlations with related traits. Squaring these correlations gives the proportion of genetic variance shared between 2 traits; thus, our findings suggest that at least a modest proportion of genetic factors associated with clinical psychiatric disorders are associated with continuous variation in milder traits of these disorders. This finding replicated the results of an earlier study5 of ASD and extended the method to many other disorders. Common variant PRS analyses supported these results, revealing an association of shared risks between ASD, ADHD, TDs, OCD, ADs, MDD, and schizophrenia and related traits, even after excluding individuals who had received a diagnosis, where possible. These converging results using 2 contemporary methods revealed that many psychiatric disorders may share genetic risks with continuous symptom dimensions in the population.

ASD, ADHD, TDs, OCD, ADs, MDD, and schizophrenia and related traits, even after excluding individuals who had received a diagnosis, where possible. These converging results using 2 contemporary methods revealed that many psychiatric disorders may share genetic risks with continuous symptom dimensions in the population. Our study went beyond traditional twin studies by directly estimating the genetic correlation between psychiatric disorders and continuous traits; we also assessed the association between disorder PRSs and continuous traits in the same sample. Dichotomous definitions of psychiatric disorders may not be optimal for all studies of these phenotypes. Our results indicate that moving beyond dichotomous definitions of psychiatric disorders to joint analyses of disorders and traits may increase statistical power and yield insights into the biology of these traits. The value of such an approach was demonstrated by a recent ADHD GWAS.26 Studies of traits in community-based samples may also be more representative than clinical samples while generating results that, to a degree, generalize to clinical populations.

tatistical power and yield insights into the biology of these traits. The value of such an approach was demonstrated by a recent ADHD GWAS.26 Studies of traits in community-based samples may also be more representative than clinical samples while generating results that, to a degree, generalize to clinical populations. However, genetic correlations in the twin analyses were less than 1. Associations between PRSs and traits had small effect sizes (which is typical of PRS studies). Thus, only a proportion of genetic risks were shared across disorders and traits. Although twin methods capture all sources of inherited genetic risk, PRSs are limited to additive common effects. Rare genetic variants may have a more deleterious effect than common variants; however, several studies have demonstrated genetic overlap from rare variants across disorders and continuous measures of ASD37 and ID.38 Correlations between environmental factors associated with psychiatric disorders and traits were also lower than the genetic correlations, suggesting that environmental factors associated with psychiatric disorders may be more unique to psychiatric disorders than genetic factors. Future work identifying risk factors that are not shared between psychiatric traits and disorders may thus help elucidate why some individuals present with mild traits, whereas others manifest clinically severe problems.

particular, as children age, the informant tends to change from the parent to the young person. This variation provides a challenge to studies seeking to examine the development of symptoms over time because changes of measurement and informant can affect results. This invariance of measurement over time is a strength. One limitation of our investigation is that, like many longitudinal studies, nonrandom attrition occurs in ALSPAC over time (eAppendix 2 in the Supplement). This nonrandom loss of participants is likely to result in conservative estimates of the prevalence of the elevated depression trajectory groups. We used a number of approaches to deal with missing data, including full information maximum likelihood in trajectory modeling and inverse probability weighting for tests of association. The pattern of results was replicated using inverse probability weighting. Nonetheless, the missing data assumption made in our analyses is that there are not systematic differences between participants who do and do not provide trajectory data and membership in the sample after conditioning on the other variables in the model (eg, PRS and variables included in the inverse probability weighting analysis).

ciated with psychiatric disorders may be more unique to psychiatric disorders than genetic factors. Future work identifying risk factors that are not shared between psychiatric traits and disorders may thus help elucidate why some individuals present with mild traits, whereas others manifest clinically severe problems. Strengths and Limitations A unique strength of our study was that we were able to perform both twin and molecular genetic analyses in 1 cohort. Linkage with nationwide patient records enabled us to focus on clinical diagnoses of psychiatric disorders as opposed to percentile-based cutoffs or screening diagnoses. This study design allowed the exclusion of individuals diagnosed with psychiatric disorders from PRS analyses, thus ruling out the possibility that observed effects were driven by individuals with clinically recognized problems. Assessments from multiple ages and raters led to a wealth of information on psychiatric phenotypes. Nonetheless, the sample was young. Although many disorders develop during childhood and adolescence, disorders such as schizophrenia, BD, and severe depression become more common with age. Because few of the participants had passed through the periods of high risk for these disorders, we could not perform twin analyses of schizophrenia or BD or exclude those individuals from PRS analyses. Studies of older individuals are thus needed as a next step. In addition, the NPR covers specialist care; thus, diagnoses ascribed through primary care were likely missed.

periods of high risk for these disorders, we could not perform twin analyses of schizophrenia or BD or exclude those individuals from PRS analyses. Studies of older individuals are thus needed as a next step. In addition, the NPR covers specialist care; thus, diagnoses ascribed through primary care were likely missed. Specific limitations of the genetic analyses include modest sample sizes and associated low power of several of the discovery GWAS analyses used to derive PRSs; for certain phenotypes, these limitations may have led to less robust results (eg, for ADs and BD). Nonrandom attrition, previously reported to affect genetic studies,39 may have decreased the observed effect sizes, particularly at later ages. In addition, the estimates of variance explained were low although typical of PRS analyses (eg, PRSs explain only 5.5% of the variance in ADHD clinical case status26) because PRSs capture only the most strongly associated common variants and rely on discovery GWAS power to accurately estimate SNP effects. Although the PRS results showed the presence of associations between genetic risk for disorders and traits consistent with the twin analyses, the degree to which common genetic risks are shared is unclear from our study. Future studies that use larger GWAS data sets and other methods are needed to estimate genetic correlations from molecular genetic data.

resence of associations between genetic risk for disorders and traits consistent with the twin analyses, the degree to which common genetic risks are shared is unclear from our study. Future studies that use larger GWAS data sets and other methods are needed to estimate genetic correlations from molecular genetic data. We did not have the statistical power to divide the cases by severity or diagnostic subtype. Thus, we cannot conclude that all levels of disorder severity share genetic risks with milder traits. This topic will be an important focus in future research because there is some evidence that severe ID is genetically independent from cognitive abilities and milder ID.40 In addition, all the continuous measures used in this study were designed to assess potentially problematic behaviors. As such, we cannot extrapolate our results to the very low positive end of each trait distribution. Studies that use measures that are sensitive to lower scores are needed.41 Conclusions Although our results do not rule out the possibility that some genetic factors are not shared between psychiatric disorders and milder traits of these disorders, they suggest that a proportion of genetic risks associated with psychiatric disorders are also associated with milder traits of these disorders. Future studies are needed to replicate our findings in older individuals and to test whether more severe forms of psychiatric disorders also share genetic risks with milder traits. Supplement. eAppendix. Supplemental Methods eTable 1. Additional Description of the Study Measures

Conclusions Although our results do not rule out the possibility that some genetic factors are not shared between psychiatric disorders and milder traits of these disorders, they suggest that a proportion of genetic risks associated with psychiatric disorders are also associated with milder traits of these disorders. Future studies are needed to replicate our findings in older individuals and to test whether more severe forms of psychiatric disorders also share genetic risks with milder traits. Supplement. eAppendix. Supplemental Methods eTable 1. Additional Description of the Study Measures eTable 2. Summary of Discovery Data Sets Used to Derive Polygenic Risk Scores eTable 3. Probandwise Concordances eTable 4. Twin Model Fit Statistics for Different Models Comparing the Associations Between Clinical Diagnosis and Continuous Traits eTable 5. Parameter Estimates From the Best-Fitting Twin Models of the Associations Between Continuous Traits and Clinical Diagnoses eTable 6. Twin Fit Statistics for Diagnostic Subdomains eTable 7. Twin Model Estimates for Diagnostic Subdomains eTable 8. Association of Polygenic Risk Scores for ADHD and ASD With Traits of Diagnostic Subdomains eFigure. Results of Sensitivity Analyses Click here for additional data file.

Introduction Major depressive disorder (MDD) is the most common mental disorder and a leading cause of disability1; even subthreshold depressive symptoms are associated with functional impairment and future mental health problems.2,3 Depression often first manifests in adolescence4,5,6 and, thereafter, individual trajectories of depressive symptoms vary substantially.7 A family history of depression and an early age at onset are each associated with a more chronic symptom course in adults with MDD,8,9,10 but it is not known what shapes early depression trajectories in youth. Depression has a complex multifactorial set of causes, including a moderate heritable component.4,11,12 Longitudinal and family studies show continuity between both adolescence-onset depressive disorder and symptoms with depression in adult life, but there are also developmental differences between depression in children, adolescents, and adults.4 For instance, clinical follow-up studies of very early-onset depression (average age at onset, 10.7 years) report high rates of heterotypic continuity, where depression is often followed by a different type of clinical disorder.13,14,15 Twin studies also show differences in the genetic set of causes of very early-onset depressive symptoms compared with those arising in mid- to late adolescence.16,17,18

at onset, 10.7 years) report high rates of heterotypic continuity, where depression is often followed by a different type of clinical disorder.13,14,15 Twin studies also show differences in the genetic set of causes of very early-onset depressive symptoms compared with those arising in mid- to late adolescence.16,17,18 At the molecular level, a recent genome-wide association study of adults with MDD found evidence of differences in the genetic architecture of depression where a relatively early age at onset (before the median age at onset of 27 years) was associated with genetic liability to schizophrenia, an association not seen for later-onset depression, which was instead associated with MDD risk alleles.19 Similar findings have been reported for emotional problems (symptoms of depression and anxiety) in that emotional problems in childhood were associated with schizophrenia risk alleles, but in adult life they were additionally associated with MDD genetic risk.20 The association of schizophrenia risk alleles with childhood emotional problems was particularly pronounced in those with emotional problems in both childhood and adulthood, suggesting that persistent emotional symptoms beginning early may drive the association with schizophrenia risk alleles. As schizophrenia genetic risk is thought to involve an early neurodevelopmental component,21,22 the role of genetic risk for other neurodevelopmental disorders in early-onset depression may be important to consider. In particular, genetic risk for ADHD, a common childhood-onset neurodevelopmental disorder, may be important in early-onset depression because cross-sectional and longitudinal cohort studies show heightened rates of depression in children with ADHD,23,24,25 which may be partly due to the strong genetic correlation between ADHD and depression (rg = 0.424).26,27

mmon childhood-onset neurodevelopmental disorder, may be important in early-onset depression because cross-sectional and longitudinal cohort studies show heightened rates of depression in children with ADHD,23,24,25 which may be partly due to the strong genetic correlation between ADHD and depression (rg = 0.424).26,27 Herein, we test the contribution of neuropsychiatric disorder genetic risk variants, specifically genetic liability to MDD, schizophrenia, and ADHD, to early depression trajectories. Schizophrenia and ADHD were selected in addition to MDD as they show moderate to high genetic correlations with major depression,27 there is evidence linking schizophrenia polygenic risk score (PRS) to early-onset depression,19,20 and epidemiologic and clinical evidence15,23,24,25 that ADHD may be an important antecedent of depression.

ADHD were selected in addition to MDD as they show moderate to high genetic correlations with major depression,27 there is evidence linking schizophrenia polygenic risk score (PRS) to early-onset depression,19,20 and epidemiologic and clinical evidence15,23,24,25 that ADHD may be an important antecedent of depression. Estimates of genetic liability to the disorders in the form of PRSs were derived from risk alleles defined in the largest available genome-wide association study of those disorders.26,28,29 We did not have a specific hypothesis for bipolar disorder genetic risk because existing studies reporting conflicting results about the phenotypic association between early-onset depression and bipolar disorder,13,15,30 with little evidence that this association is stronger for early-onset depression. Bipolar disorder also differs from ADHD and schizophrenia in that evidence suggests that bipolar disorder is less neurodevelopmental in origin.21,22 However, for completeness, we included bipolar PRSs in our analyses (eTable 2A in the Supplement). We hypothesized that ADHD and schizophrenia genetic risk would show an association with early-onset depression and that depression genetic risk would be associated with depression with an onset later in adolescence.

21,22 However, for completeness, we included bipolar PRSs in our analyses (eTable 2A in the Supplement). We hypothesized that ADHD and schizophrenia genetic risk would show an association with early-onset depression and that depression genetic risk would be associated with depression with an onset later in adolescence. Methods The Avon Longitudinal Study of Parents and Children (ALSPAC) is an ongoing, population-based, prospective, longitudinal UK birth cohort.31,32 Data collection began September 6, 1990. The enrolled core sample consisted of 14 541 pregnant women living in Avon, England, with expected delivery dates between April 1, 1991, and December 31, 1992. Of these births, 13 988 children were alive at 1 year. An additional 713 children who would have been eligible but whose mothers did not choose to participate during pregnancy were enrolled after age 7 years, giving a total sample of 14 701 children alive at 1 year. The present study was conducted between November 10, 2017, and August 14, 2018. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the local research ethics committees. All participants provided written informed consent; there was no financial compensation.

Methods The Avon Longitudinal Study of Parents and Children (ALSPAC) is an ongoing, population-based, prospective, longitudinal UK birth cohort.31,32 Data collection began September 6, 1990. The enrolled core sample consisted of 14 541 pregnant women living in Avon, England, with expected delivery dates between April 1, 1991, and December 31, 1992. Of these births, 13 988 children were alive at 1 year. An additional 713 children who would have been eligible but whose mothers did not choose to participate during pregnancy were enrolled after age 7 years, giving a total sample of 14 701 children alive at 1 year. The present study was conducted between November 10, 2017, and August 14, 2018. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the local research ethics committees. All participants provided written informed consent; there was no financial compensation. The study website contains details of all of the data that are available through a fully searchable data dictionary.33 For families with multiple births, we included the oldest sibling. Individuals were included in analyses when data on the primary outcome of depressive symptoms were available for at least 2 time points (n = 7543). The sample mean age (SD) was 10.64 (0.25) years at baseline and 18.65 (0.49) years at the final time point. The numbers of individuals with data available at different times are shown in the eFigure in the Supplement.

ta on the primary outcome of depressive symptoms were available for at least 2 time points (n = 7543). The sample mean age (SD) was 10.64 (0.25) years at baseline and 18.65 (0.49) years at the final time point. The numbers of individuals with data available at different times are shown in the eFigure in the Supplement. Depressive symptoms were reported by the young person at 6 time points (ages 10.5, 12.5, 13.5, 16.5, 17.5, and 18.5 years) on the short Mood and Feelings Questionnaire. This is a well-validated symptom checklist34,35,36 that includes 13 items about mood symptoms during the past 2 weeks (rated 0, not true; 1, sometimes true; or 2, true; score range, 0-26). Scores above 11 represent clinically significant symptoms,34,36 and we analyzed individuals scoring above and below this level to examine trajectories of clinically significant symptoms.

includes 13 items about mood symptoms during the past 2 weeks (rated 0, not true; 1, sometimes true; or 2, true; score range, 0-26). Scores above 11 represent clinically significant symptoms,34,36 and we analyzed individuals scoring above and below this level to examine trajectories of clinically significant symptoms. Polygenic risk scores for MDD, schizophrenia, and ADHD were generated in study individuals as the standardized mean number of disorder risk alleles in approximate linkage equilibrium (R2<0.20), weighted by genome-wide association study allele effect size derived from data of imputed autosomal single-nucleotide polymorphisms. All analyses were performed using Stata, version 13.0 (StataCorp) to implement the PLINK toolset (http://zzz.bwh.harvard.edu/plink/; code is available at https://github.com/ricanney/stata). In brief, best-guess genotype underwent additional marker and individual quality control. Individuals were excluded on the basis of excessive heterozygosity (>4 SDs from sample mean), relatedness (>3 SDs from sample mean), and genotype missingness (>2%). Markers were excluded if they were rare (minor allele count <5), had high levels of missingness (>2%), or deviated from Hardy-Weinberg equilibrium (P ≤ 10−10) or from reference minor allele frequency (>10%) (eMethods in the Supplement).

ample mean), relatedness (>3 SDs from sample mean), and genotype missingness (>2%). Markers were excluded if they were rare (minor allele count <5), had high levels of missingness (>2%), or deviated from Hardy-Weinberg equilibrium (P ≤ 10−10) or from reference minor allele frequency (>10%) (eMethods in the Supplement). Scores were derived from MDD, ADHD, and schizophrenia weights for 152 536, 103 041, and 27 336 single-nucleotide polymorphisms, respectively. Risk alleles were defined as those associated with case status in the most recent Psychiatric Genomics Consortium analyses of MDD, ADHD, and schizophrenia at a threshold of P < .50 for depression and ADHD and P < .05 for schizophrenia, as these thresholds maximally capture phenotypic variance.26,27,28,29,37 Genome-wide association study discovery sample sizes were 130 664 cases and 330 470 controls for MDD, 20 183 cases and 35 191 controls for ADHD, and 35 476 cases and 46 839 controls for schizophrenia. All PRSs were standardized prior to analysis so odds ratios (ORs) represent 1 SD change (eTable 2A in the Supplement for bipolar PRSs). Phenotypic measures of neurodevelopmental problems (DSM-IV38 diagnoses of childhood ADHD, social communication problems, and pragmatic language difficulties at age 7 years), psychotic experiences (ages 12 and 17 years), family history of severe depression and schizophrenia, and maternal educational level were used (eMethods in the Supplement).

urodevelopmental problems (DSM-IV38 diagnoses of childhood ADHD, social communication problems, and pragmatic language difficulties at age 7 years), psychotic experiences (ages 12 and 17 years), family history of severe depression and schizophrenia, and maternal educational level were used (eMethods in the Supplement). Statistical Analysis We characterized depression trajectories of symptoms dichotomized by clinical cutpoint (n = 7543) using latent class growth analysis in Mplus, version 8.39 This analysis is a probability-based technique used to identify an optimum number of distinct patterns (classes) of growth (change) in the longitudinal depression scores of individuals.40 Models were run with increasing numbers of classes, starting with a 1-class solution specifying both linear and quadratic change with 500 random starting values and 50 optimizations. Residual variances were allowed to vary across measurement points. A maximum likelihood parameter estimator for which SEs are robust to nonnormality was used.

increasing numbers of classes, starting with a 1-class solution specifying both linear and quadratic change with 500 random starting values and 50 optimizations. Residual variances were allowed to vary across measurement points. A maximum likelihood parameter estimator for which SEs are robust to nonnormality was used. To examine associations with categorical variables (eg, sex), the DCAT auxiliary option in MPlus was used. A bias-free, 3-step approach in MPlus (R3STEP) estimated the associations between continuous hypothesized predictor variables (PRSs) and trajectory class.41,42,43 Model selection was informed by model fit indices and interpretability as recommended.44 Full information maximum likelihood estimation was used in MPlus and included all individuals with more than 1 depression assessment in analyses (eTable 1 in the Supplement). For tests of PRS association with trajectory class, we reran analyses using inverse probability weighting45 to address any potential bias caused by participant dropout. The pattern of results was similar (eTable 3 in the Supplement).

duals with more than 1 depression assessment in analyses (eTable 1 in the Supplement). For tests of PRS association with trajectory class, we reran analyses using inverse probability weighting45 to address any potential bias caused by participant dropout. The pattern of results was similar (eTable 3 in the Supplement). Results Depression Symptom Trajectories A 3-class trajectory model provided the best fit to the data and results that were most readily interpretable (eTable 1 and eAppendix 1 in the Supplement). The Figure shows the 3 distinct trajectory classes: a persistently low class (73.7%), a later-adolescence–onset class (17.3%), and an early-adolescence–onset class (9.0%). In the early-adolescence–onset class, the probability of clinically significant depression was first elevated (as indicated by a probability of clinically significant depression symptoms of 0.44) at age 12.5 years, which rose to 0.52 at 13.5 years. In the later-adolescence–onset class, the probability of clinically significant depression (probability, 0.47) was first elevated at age 16.5 years and rose at 17.5 years (probability, 0.57). Both elevated trajectories were associated with a diagnosis of MDD (assessed by the Clinical Interview Schedule–Revised46 at age 17.5 years) providing validation of the trajectory classes (later-adolescence onset, 34.4%; early-adolescence onset, 22.8%; low level, 1.5%; overall difference, χ22 = 193.70; P = .001). The estimated proportion of females was 45.8% in the low-level class and was higher, but did not differ significantly, between the early-adolescence– (74.3%) and later-adolescence– (73.2%) onset classes (Table 1).

nset, 34.4%; early-adolescence onset, 22.8%; low level, 1.5%; overall difference, χ22 = 193.70; P = .001). The estimated proportion of females was 45.8% in the low-level class and was higher, but did not differ significantly, between the early-adolescence– (74.3%) and later-adolescence– (73.2%) onset classes (Table 1). Figure. Developmental Trajectories of Depressive Symptoms Depression trajectories identified by latent class growth analyses. Table 1. Phenotypic Associations With Trajectory Classa Variable Onset, OR (95% CI) Difference Between Early- and Later-Adolescence–Onset Classesb Early Adolescence P Value Later Adolescence P Value χ21 or OR (95% CI) P Value Sex, % 74.3 <.001 73.2 <.001 χ21 = 0.015 .90 Maternal education, completed A-levels, %c 39.1 .01 34.9 .001 χ21 = 0.707 .44 Childhood ADHD, % 6.3 .008 0.9 .37 χ21 = 6.837 .009 Pragmatic language difficultiesd 0.63 (0.55-0.71) <.001 0.82 (0.72-0.94) .006 OR, 1.31 .004 χ21 = 11.709 .001 (for Cutpoint) Social communication difficultiese 1.50 (1.34-1.68) <.001 1.01 (0.87-1.18) .86 OR, 0.68 <.001 χ21 = 18.819 .001 (for Cutpoint) Psychotic experiences 12 y 1.47 (1.35-1.61) <.001 0.89 (0.64-1.22) .46 OR, 0.60 .003 17 y 1.57 (1.36-1.80) <.001 1.54 (1.33-1.79) <.001 OR, 0.99 .74 Abbreviations: ADHD, attention deficit/hyperactivity disorder; OR, odds ratio. a Continuous scores are standardized so that ORs are for 1-SD increase. Low-risk group was the reference group except for tests of comparison between early-adolescence– and later-adolescence–onset groups where the early-adolescence–onset group was the reference group.

Table 1. Phenotypic Associations With Trajectory Classa Variable Onset, OR (95% CI) Difference Between Early- and Later-Adolescence–Onset Classesb Early Adolescence P Value Later Adolescence P Value χ21 or OR (95% CI) P Value Sex, % 74.3 <.001 73.2 <.001 χ21 = 0.015 .90 Maternal education, completed A-levels, %c 39.1 .01 34.9 .001 χ21 = 0.707 .44 Childhood ADHD, % 6.3 .008 0.9 .37 χ21 = 6.837 .009 Pragmatic language difficultiesd 0.63 (0.55-0.71) <.001 0.82 (0.72-0.94) .006 OR, 1.31 .004 χ21 = 11.709 .001 (for Cutpoint) Social communication difficultiese 1.50 (1.34-1.68) <.001 1.01 (0.87-1.18) .86 OR, 0.68 <.001 χ21 = 18.819 .001 (for Cutpoint) Psychotic experiences 12 y 1.47 (1.35-1.61) <.001 0.89 (0.64-1.22) .46 OR, 0.60 .003 17 y 1.57 (1.36-1.80) <.001 1.54 (1.33-1.79) <.001 OR, 0.99 .74 Abbreviations: ADHD, attention deficit/hyperactivity disorder; OR, odds ratio. a Continuous scores are standardized so that ORs are for 1-SD increase. Low-risk group was the reference group except for tests of comparison between early-adolescence– and later-adolescence–onset groups where the early-adolescence–onset group was the reference group. b χ2 Tests of difference for social communication and pragmatic language difficulties used the established clinical cut-points for identifying problems (eAppendix in the Supplement). The OR values represent the difference between the ORs in the preceding columns for later-adolescence onset vs early-adolescence onset. c A-level education is equivalent to high school diploma in the United States d Lower scores represent more difficulties.

b χ2 Tests of difference for social communication and pragmatic language difficulties used the established clinical cut-points for identifying problems (eAppendix in the Supplement). The OR values represent the difference between the ORs in the preceding columns for later-adolescence onset vs early-adolescence onset. c A-level education is equivalent to high school diploma in the United States d Lower scores represent more difficulties. e Higher scores represent more problems. Neuropsychiatric PRS and Trajectory Class As reported in Table 2, the later-adolescence–onset class was associated only with higher MDD PRS (OR, 1.27; 95% CI, 1.09-1.48; P = .003). The early-adolescence–onset class was associated with higher PRSs for ADHD (OR, 1.32; 95% CI, 1.13-1.54; P < .001), schizophrenia (OR, 1.22; 95% CI, 1.04-1.43; P = .01), and MDD (OR, 1.24; 95% CI, 1.06-1.46; P = .007). Post hoc, we examined the association with all 3 psychiatric PRSs and trajectory class to examine which PRS contributed most strongly (Table 2). As expected, the PRSs were correlated (eTable 2B and C in the Supplement).

hizophrenia (OR, 1.22; 95% CI, 1.04-1.43; P = .01), and MDD (OR, 1.24; 95% CI, 1.06-1.46; P = .007). Post hoc, we examined the association with all 3 psychiatric PRSs and trajectory class to examine which PRS contributed most strongly (Table 2). As expected, the PRSs were correlated (eTable 2B and C in the Supplement). Table 2. Associations of Polygenic Risk Scores With Trajectory Classes Association Onset, OR (95% CI) Early Adolescence P Value Later Adolescence P Value Univariate MDD PRS 1.24 (1.06-1.46) .007 1.27 (1.09-1.48) .003 Schizophrenia PRS 1.22 (1.04-1.43) .01 .95 (0.82-1.11) .56 ADHD PRS 1.32 (1.13-1.54) <.001 .94 (0.80-1.11) .48 Multivariate MDD PRS 1.16 (0.98-1.36) .09 1.31 (1.12-1.53) .001 Schizophrenia PRS 1.19 (1.01-1.41) .04 .93 (0.79-1.10) .39 ADHD PRS 1.27 (1.08-1.50) .003 .90 (0.76-1.07) .23 Abbreviations: ADHD, attention deficit/hyperactivity disorder; MDD, major depressive disorder; OR, odds ratio; PRS, polygenic risk score.

.11) .48 Multivariate MDD PRS 1.16 (0.98-1.36) .09 1.31 (1.12-1.53) .001 Schizophrenia PRS 1.19 (1.01-1.41) .04 .93 (0.79-1.10) .39 ADHD PRS 1.27 (1.08-1.50) .003 .90 (0.76-1.07) .23 Abbreviations: ADHD, attention deficit/hyperactivity disorder; MDD, major depressive disorder; OR, odds ratio; PRS, polygenic risk score. Multivariate analysis showed that the strongest association with the early-adolescence–onset class was observed for ADHD PRS, the association with schizophrenia PRS was retained, and the association with MDD PRS became nonsignificant (Table 2). Results for the later-adolescence–onset class remained the same. Comparing the early- and later-onset classes showed significant differences (MDD PRS: OR, 1.13; 95% CI, 0.88-1.46; P = .33; schizophrenia PRS: OR, 0.78; 95% CI, 0.60-1.01; P = .08; and ADHD PRS: OR, 0.71; 95% CI, 0.55-0.92; P = .009). Bipolar PRS was not associated with trajectory classes (eTable 2A in the Supplement). Including ancestry-derived principal components did not alter the results (eAppendix 2 in the Supplement).

8-1.46; P = .33; schizophrenia PRS: OR, 0.78; 95% CI, 0.60-1.01; P = .08; and ADHD PRS: OR, 0.71; 95% CI, 0.55-0.92; P = .009). Bipolar PRS was not associated with trajectory classes (eTable 2A in the Supplement). Including ancestry-derived principal components did not alter the results (eAppendix 2 in the Supplement). We tested whether the trajectory classes differed phenotypically on traits conceptually related to ADHD PRS (childhood ADHD and neurodevelopmental traits) and schizophrenia PRS (psychotic experiences). For childhood neurodevelopmental traits, there is evidence that these traits are associated with both ADHD and ADHD PRS47,48; for psychotic experiences, there is inconsistent evidence that these experiences are linked with psychosis and schizophrenia PRSs49,50 (Table 1). Individuals in the early-adolescence–onset class had higher rates of childhood ADHD (6.3%) than the later-adolescence–onset (0.9%) and low (1.7%) classes and more social communication and pragmatic language problems (Table 1). Proportions scoring above the standard cut points were early onset, 20.7%; later onset, 4.2%; and low level, 5.8% for social communication and early onset, 13.3%; later onset, 2.1%; and low level, 1.4% for pragmatic language. These differences distinguished the early-adolescence– and later-adolescence–onset classes (Table 2). For psychotic experiences, these differences distinguished the early-adolescence– and later-adolescence–onset classes only at age 12 years.

y onset, 13.3%; later onset, 2.1%; and low level, 1.4% for pragmatic language. These differences distinguished the early-adolescence– and later-adolescence–onset classes (Table 2). For psychotic experiences, these differences distinguished the early-adolescence– and later-adolescence–onset classes only at age 12 years. Discussion This study identified variation in the developmental trajectories of depression from childhood to early adult life, and moreover, that this variation is partly attributable to MDD, schizophrenia, and ADHD risk alleles. We found evidence of distinct depressive trajectories primarily distinguished by age at onset. We found that the more common, typical developmental trajectory, with onset after puberty and persistence into early adulthood,6,51 was associated with elevated genetic risk for depression indexed by MDD PRS. In contrast, we found that depressive symptoms defined by a very early onset (by age 12 years) were associated with all neuropsychiatric genetic risk scores assessed, with multivariate analysis showing that the association was strongest for ADHD PRS.

ociated with elevated genetic risk for depression indexed by MDD PRS. In contrast, we found that depressive symptoms defined by a very early onset (by age 12 years) were associated with all neuropsychiatric genetic risk scores assessed, with multivariate analysis showing that the association was strongest for ADHD PRS. Phenotypically, childhood neurodevelopmental difficulties (ADHD, pragmatic language, and social communication difficulties) differentiated the depression trajectories that were elevated only in the early-adolescence–onset group with rates increased by 5- to 7-fold in the early-adolescence–onset group. Psychotic experiences differentiated the groups only at age 12 years. This discrepancy may be driven by depressive symptom differences between the groups at age 12 years (Figure) given the reported association between psychotic experiences and depression and an inconsistent association with psychotic experiences and schizophrenia PRS.49,50 The findings are consistent with a growing body of literature showing that depression has heterogeneous causes partly indexed by age at onset. In particular, studies of adult MDD and symptoms measured continuously in population-based samples illustrate that a relatively earlier onset is more strongly associated with schizophrenia polygenic risk.19,20,52 We found an additional contribution from ADHD PRSs.

as heterogeneous causes partly indexed by age at onset. In particular, studies of adult MDD and symptoms measured continuously in population-based samples illustrate that a relatively earlier onset is more strongly associated with schizophrenia polygenic risk.19,20,52 We found an additional contribution from ADHD PRSs. The implication of those results is that early- and later-adolescence-onset depression differ to some extent with respect to the risk factors involved and that the earlier-onset disorder is more strongly influenced by neurodevelopmental factors than depression with a more typical onset in later adolescence or early adulthood. This finding is consistent with a number of observations from epidemiologic, family, and clinical studies. First, several family and clinical follow-up studies suggest that childhood-onset depression might differ etiologically from adolescent-onset depression.53,54,55,56 Second, the epidemiologic factors associated with very early-onset depression differ from those of depression with onset in midadolescence to late adolescence in the sex ratio of affected individuals and long-term psychiatric outcomes.13,57 Third, neurodevelopmental difficulties, including speech abnormalities and poor motor skills, are particularly associated with early-onset rather than adolescent- or adult-onset depression.15,58,59 Fourth, substantial clinical evidence shows that children with ADHD, a common neurodevelopmental disorder, are at elevated risk of subsequent depressive symptoms, suicide attempt, and emotional problems in adult life.25,60,61,62,63

y associated with early-onset rather than adolescent- or adult-onset depression.15,58,59 Fourth, substantial clinical evidence shows that children with ADHD, a common neurodevelopmental disorder, are at elevated risk of subsequent depressive symptoms, suicide attempt, and emotional problems in adult life.25,60,61,62,63 Theory suggests neurodevelopmental difficulties as one route to emotional disturbance through the repeated experience of academic failure and peer rejection,64 although ADHD and depression may also be associated owing to common risk factors.65 A clinical issue is that the response to antidepressant medication66,67,68,69 in youth is not as good as it is in adults and evidence suggests the response to tricyclic antidepressants may differ in prepubertal vs postpubertal depression. One possibility is that more neurodevelopmental depression shows a different type of treatment response.

t the response to antidepressant medication66,67,68,69 in youth is not as good as it is in adults and evidence suggests the response to tricyclic antidepressants may differ in prepubertal vs postpubertal depression. One possibility is that more neurodevelopmental depression shows a different type of treatment response. The present study indicates that genetic risk for ADHD and schizophrenia in the general population is associated with a persistent, early-onset trajectory of depressive symptoms. Such effects could operate through overlapping biological pathways as well as evocative gene-environment correlation where genetic factors influence traits that then affect environmental exposures (eg, social exclusion) associated with depression. Irritability, which is common in children with ADHD and other neurodevelopmental disorders, is indexed by genetic risk for ADHD in youth70 and has been shown to increase risk for later depression,71,72 may be a potential route through which ADHD genetic risk increases the likelihood of mood problems.

d with depression. Irritability, which is common in children with ADHD and other neurodevelopmental disorders, is indexed by genetic risk for ADHD in youth70 and has been shown to increase risk for later depression,71,72 may be a potential route through which ADHD genetic risk increases the likelihood of mood problems. Among those with early-onset depression, we did not identify the equal sex ratio of affected males and females that has often been reported when depression onset is very early.4,73 This finding was somewhat surprising, and several factors may have contributed to it. First, some research suggests that depression is particularly likely in females with neurodevelopmental disorders, which may imply that high neurodevelopmental risk is more likely to manifest as mood disorder in females.24,48,74 Second, while it is generally accepted that self-reports of adolescent mood (as used in the present study) are reliable, children with neurodevelopmental disorders, predominately boys, may underreport their mood symptoms compared with typically developing children.75 This reporting difference raises the possibility that the reliance on self-reported mood necessary in the present study due to repeated longitudinal assessments (see below) may mean that some individuals at high neurodevelopmental risk may have been misclassified. Finally, PRSs alone are unlikely to be able to reliably classify children’s risk of developing different types of depression trajectories. However, collectively, results converge to suggest that neurodevelopmental phenotypes (ADHD, as well as social communication and pragmatic language difficulties) and neurodevelopmental genetic risk indicate a greater probability of an early-onset depression trajectory. Phenotypic childhood neurodevelopmental problems were markedly increased in the early-adolescence–onset group (by 5- to 7-fold) compared with the typical depression trajectory. Studies with follow-up further into adult life will help to clarify the adult mental health outcomes of these groups.

the missing data assumption made in our analyses is that there are not systematic differences between participants who do and do not provide trajectory data and membership in the sample after conditioning on the other variables in the model (eg, PRS and variables included in the inverse probability weighting analysis). Depression was assessed using self-reported questionnaires rather than clinical assessment. Nonetheless, subthreshold symptoms are associated with impairment and subsequent MDD.2,3,4 It was not possible to investigate rates of mania or bipolar disorder in the trajectory groups. However, evidence is inconsistent on the link with early-onset depression and bipolar disorder.13,15,30 The follow-up period in this study was limited to early adult life. A final limitation is that PRSs are weak predictors and explain only a small to modest proportion of phenotypic variance as they do in the present article. However, they provide a useful biological indicator of genetic liability.76 Conclusions The findings of this study suggest etiologically distinct trajectories of depressive symptoms in youth dependent on age at onset. The findings also show that neurodevelopmental genetic risk contributes to very early-onset depression. Supplement. eMethods. Other Assessments eFigure. Number of Individuals With Data Available at Each Measurement Point eTable 1. Model Fit Indices for Latent Class Growth Models of Self-reported Depression eAppendix 1. Additional Analyses eTable 2A. Associations Between Bipolar PRS and Trajectory Class eTable 2B. Correlations Between Polygenic Risk Scores

Supplement. eMethods. Other Assessments eFigure. Number of Individuals With Data Available at Each Measurement Point eTable 1. Model Fit Indices for Latent Class Growth Models of Self-reported Depression eAppendix 1. Additional Analyses eTable 2A. Associations Between Bipolar PRS and Trajectory Class eTable 2B. Correlations Between Polygenic Risk Scores eTable 2C. Correlations Between Polygenic Risk Scores and Parent-Reported Family History of Psychiatric Disorder eAppendix 2. Additional Analyses eTable 3. Associations of Polygenic Risk Scores With Trajectory Classes Using Inverse Probability Weighting Click here for additional data file.

Introduction Millions of adults now entering middle age were exposed to high levels of lead as children,1 a phenomenon that accompanied the peak use of lead in gasoline worldwide from the 1940s through the early 1990s.2 From 1976 to 1980, the average child living in the United States had blood lead levels (BLLs) 3 times higher (>15 μg/dL)1 than the current reference value for clinical attention (5 μg/dL).3 Lead-exposed children experience disrupted cognitive and behavioral development,4 with childhood lead exposure linked to lower child IQ,5 poorer academic achievement,6 and greater rates of child behavior problems, particularly inattention, hyperactivity, and antisocial behavior.7,8,9 Meanwhile, adults exposed to lead are at increased risk of developing some psychiatric conditions.10,11,12,13 Although follow-up studies14,15 of lead-tested children have reported the persistence of lead-related cognitive deficits well into adulthood, apart from antisocial outcomes, the long-term mental and behavioral health consequences of early-life lead exposure have not been fully characterized.

tric conditions.10,11,12,13 Although follow-up studies14,15 of lead-tested children have reported the persistence of lead-related cognitive deficits well into adulthood, apart from antisocial outcomes, the long-term mental and behavioral health consequences of early-life lead exposure have not been fully characterized. To our knowledge, 2 studies16,17 have undertaken long-term follow-up in lead-exposed children to determine whether early behavior problems persist or evolve into adult mental health concerns (a larger number of studies18,19,20,21,22 have examined whether adolescents and young adults exposed to lead as children display more antisocial and criminal behaviors, with most studies, although not all, suggesting that they do). A US study16 that used linked health records and clinical interviews to identify cases of psychosis in 2 lead-tested child cohorts born in the late 1960s (N = 200; age range, 30-35 years at follow-up) reported a 2-fold increased risk of schizophrenia spectrum disorder in adulthood for individuals with high BLLs as children (approximately >15 μg/dL). Another study, which is to our knowledge the only comprehensive adult psychiatric follow-up study17 conducted in a lead-tested child cohort (N = 210; cohort born in the early 1980s), reported greater social phobia, anxiety, and substance abuse problems in adulthood (mean age at follow-up, 26.3 years) for Australian women who had greater BLLs as children, although all associations were attenuated to nonsignificance by statistical adjustment for study covariates, including parental educational and occupational attainment.

l phobia, anxiety, and substance abuse problems in adulthood (mean age at follow-up, 26.3 years) for Australian women who had greater BLLs as children, although all associations were attenuated to nonsignificance by statistical adjustment for study covariates, including parental educational and occupational attainment. This existing evidence base has limitations. First, because of small sample sizes, these studies had limited power to detect effects. Second, because they considered only specific disorders (eg, schizophrenia) or relied on right-hand censored, single–time point clinical interviews to assess psychiatric problems, these studies likely underdetected episodes of illness and overlooked disorders that have a pattern of reoccurrence.23 Third, it is now appreciated that most psychiatric disorders are dimensional constructs, not discrete categorical entities.24 Individuals who meet criteria for one disorder typically also meet criteria for others both cross-sectionally and across the life course.25,26,27 Empirical evidence suggests that psychiatric illnesses can be represented by 3 higher-order dimensions—internalizing, externalizing, and thought disorders (eg, psychotic experiences)28—that are intercorrelated and may reflect a common liability toward psychopathology in general, labeled the p factor.29,30,31 The p factor may be a particularly appropriate outcome for studies that link environmental toxins to mental disorder because (1) the few previous studies10,11,13,16,17 of lead and psychopathology suggest that lead exposure may increase the risk of internalizing, externalizing, and thought disorders, without particular specificity, and (2) the continuous and omnibus nature of the p factor allows investigators to easily test for dose-effect relationships.

previous studies10,11,13,16,17 of lead and psychopathology suggest that lead exposure may increase the risk of internalizing, externalizing, and thought disorders, without particular specificity, and (2) the continuous and omnibus nature of the p factor allows investigators to easily test for dose-effect relationships. Although epidemiologists have hypothesized a relationship between child lead exposure and adult psychopathology, another way of considering the link between lead and behavioral dysfunction focuses on personality features that may impair an individual's capacity to lead a happy, successful life. Decades of research using the Big Five framework to represent the 5 broadest factors of personality32 have identified a combination of traits, including poor impulse control, high antagonism, and a tendency toward negative emotionality, that have a detrimental effect on love, work, and health and that appear to predispose individuals to psychiatric illness.33,34,35 Few studies have examined personality traits in association with lead exposure, but adults occupationally exposed to lead have reported feeling angrier and more tired, tense, and depressed than their less exposed peers36,37; these emotional symptoms seem to improve with the abatement of lead exposure.38 In the one cohort of lead-tested children who received personality testing (born in Cincinnati, Ohio, in the early 1980s; aged 19-24 years at follow-up), young adults with greater childhood lead exposure tested higher, on average, than cohort peers on a self-report inventory of psychopathic traits, such as impulsivity and egocentricity.39 Alterations in emotion regulation and adult personality have consequently been proposed as explanatory mechanisms for the reported link between childhood BLLs and adolescent delinquency40 and young adult criminal arrests41 also observed in this Cincinnati cohort.

opathic traits, such as impulsivity and egocentricity.39 Alterations in emotion regulation and adult personality have consequently been proposed as explanatory mechanisms for the reported link between childhood BLLs and adolescent delinquency40 and young adult criminal arrests41 also observed in this Cincinnati cohort. With this study, we undertook, to our knowledge, the longest and largest psychiatric follow-up to date in a cohort of adults who were lead exposed and lead tested as children, as well as the only follow-up to use (1) repeated clinical interviews assessing psychopathology symptoms across adulthood up to 38 years of age; (2) comprehensive, dimensional measures of psychopathology that account for severity, comorbidity, and reoccurrence; and (3) a broad measure of adult personality (Big Five Personality Inventory)32 that did not rely on self-report. We conducted these follow-ups in a sample in which the extent of children's exposure to lead was unrelated to their socioeconomic origins15 (Figure 1), removing a potentially important confounder that is present in most studies of children and lead.42

(Big Five Personality Inventory)32 that did not rely on self-report. We conducted these follow-ups in a sample in which the extent of children's exposure to lead was unrelated to their socioeconomic origins15 (Figure 1), removing a potentially important confounder that is present in most studies of children and lead.42 Figure 1. Association of Childhood Blood Lead Level at 11 Years of Age With Child Family Socioeconomic Status There was no significant social gradient in lead exposure in the Dunedin study children (r = −0.01; 95% CI, −0.08 to 0.07; P = .86); high blood lead levels were observed among children from all socioeconomic status groups. Childhood socioeconomic status was determined through the 6-point Elley-Irving scale (categories 1 and 2 [low status], 3 and 4 [middle status], and 5 and 6 [high status]), which codes the occupations and associated income and educational levels of members’ parents. The orange line depicts the nonsignificant association between child blood lead levels and childhood socioeconomic status. The dotted line depicts the current Centers for Disease Control and Prevention child blood lead level reference value for clinical attention (5 μg/dL).3 A total of 554 (94.0%) tested study members had blood lead levels above the current reference value.

etween child blood lead levels and childhood socioeconomic status. The dotted line depicts the current Centers for Disease Control and Prevention child blood lead level reference value for clinical attention (5 μg/dL).3 A total of 554 (94.0%) tested study members had blood lead levels above the current reference value. Methods Study Design and Population Members were members of the Dunedin Multidisciplinary Health and Development Study, a longitudinal investigation of health and behavior in a birth cohort. The full cohort comprised all individuals born between April 1, 1972, and March 31, 1973, in Dunedin, New Zealand, who were eligible based on residence in the province and who participated in the first assessment at 3 years of age. The cohort represented the full range of socioeconomic status in the general population of New Zealand's South Island.43 With regard to adult health, the cohort matched the New Zealand National Health and Nutrition Survey on key indicators (eg, body mass index, smoking, and visits to a physician).43 The cohort was primarily white; less than 7% self-identified as having nonwhite ancestry, matching the demographics of the South Island.43 Assessments were performed at birth and 3, 5, 7, 9, 11, 13, 15, 18, 21, 26, and 32 years of age, and the most recent data collection was completed in December 2012, when members were 38 years of age. Data analysis was performed from March 14, 2018, to October 24, 2018. Written informed consent was obtained from parents and cohort members, and data were deidentified. Study protocols were approved by the Southern Health and Disability Ethics Committee at the New Zealand Ministry of Health and The Duke University Health System Institutional Review Board for Clinical Investigations.

018. Written informed consent was obtained from parents and cohort members, and data were deidentified. Study protocols were approved by the Southern Health and Disability Ethics Committee at the New Zealand Ministry of Health and The Duke University Health System Institutional Review Board for Clinical Investigations. Measures Childhood BLLs Approximately 30-mL venous blood samples were obtained at 11 years of age from 579 of the 803 children (72.1%) who participated in the assessment performed at the Dunedin Multidisciplinary Health and Development Research Unit and who freely agreed to provide a blood sample. An additional 122 children were assessed in their schools, where blood samples could not be obtained. Whole-blood samples were analyzed through graphite fumance atomic absorption spectrophotometry. Details on the method of blood collection, storage, and analysis have been described previously.9

provide a blood sample. An additional 122 children were assessed in their schools, where blood samples could not be obtained. Whole-blood samples were analyzed through graphite fumance atomic absorption spectrophotometry. Details on the method of blood collection, storage, and analysis have been described previously.9 Assessment of Symptoms of Mental Disorder The Dunedin study longitudinally ascertains mental disorders every 2 to 6 years, interviewing members about past-year symptoms. We also used life-history calendar interviews to ascertain indicators of mental disorder that occur in the gaps between assessments, including inpatient treatment, outpatient treatment, or spells taking prescribed psychiatric medication (indicators that are salient and recalled more reliably than individual symptoms). Life-history calendar data indicate that virtually all members with a disorder consequential enough to be associated with treatment have been detected in our net of past-year diagnoses made at 18, 21, 26, 32, and 38 years. Specifically, we identified only 11 people who reported treatment but had not been captured in our net of diagnoses from 18 to 38 years of age (most of whom experienced brief postnatal depression).

to be associated with treatment have been detected in our net of past-year diagnoses made at 18, 21, 26, 32, and 38 years. Specifically, we identified only 11 people who reported treatment but had not been captured in our net of diagnoses from 18 to 38 years of age (most of whom experienced brief postnatal depression). Psychopathology symptoms were assessed through private structured interviews using the Diagnostic Interview Schedule44 at 18, 21, 26, 32, and 38 years of age. Interviewers were health care professionals, had completed a 2-week training course to criterion, and were retrained periodically as needed throughout data collection. We studied Diagnostic and Statistical Manual of Mental Disorders (DSM)–defined symptoms of the following disorders that were repeatedly assessed in our longitudinal study: alcohol dependence, cannabis dependence, dependence on hard drugs, tobacco dependence (assessed with the Fagerström Test for Nicotine Dependence),45 conduct disorder, major depression, generalized anxiety disorder, fears and/or phobias, obsessive compulsive disorder, mania, and positive and negative schizophrenia symptoms. Ordinal measures represented the number of the 7 (eg, mania, generalized anxiety disorder) to 10 (eg, alcohol dependence, cannabis dependence) possible DSM-defined symptoms associated with each disorder. Fears and/or phobias were assessed as the count of diagnoses for simple phobia, social phobia, agoraphobia, and panic disorder that a study member reported at each assessment. Symptoms were assessed without regard for hierarchical exclusionary rules to facilitate the examination of comorbidity. Each of the 11 disorders were assessed at least 3 times. The past-year prevalence rates of psychiatric disorders in the Dunedin cohort are similar to prevalence rates in nationwide surveys of the United States and New Zealand.23,46

regard for hierarchical exclusionary rules to facilitate the examination of comorbidity. Each of the 11 disorders were assessed at least 3 times. The past-year prevalence rates of psychiatric disorders in the Dunedin cohort are similar to prevalence rates in nationwide surveys of the United States and New Zealand.23,46 Structure of Psychopathology From 18 to 38 Years of Age The methods used to compute the hierarchical measures of psychopathology in the Dunedin cohort have been described previously.29 In brief, we used confirmatory factor analysis to calculate factor scores that represent internalizing (with loadings from depression, anxiety, and fear and/or phobia symptoms), externalizing (with loadings from substance dependence and conduct disorder symptoms), and thought disorders (with loading from obsessive-compulsive, manic, and psychotic symptoms), as well as general psychopathology (ie, the p factor; with loadings from all 11 assessed disorders). Fit indexes met criteria for good model fit. For expository purposes, we scaled study members’ scores on all factors to a mean (SD) of 100 (15). These measures are further described in the eMethods and eFigure in the Supplement.

s general psychopathology (ie, the p factor; with loadings from all 11 assessed disorders). Fit indexes met criteria for good model fit. For expository purposes, we scaled study members’ scores on all factors to a mean (SD) of 100 (15). These measures are further described in the eMethods and eFigure in the Supplement. Adult Personality At 26, 32, and 38 years of age, study members nominated people who knew them well. These informants were mailed questionnaires and asked to describe each study member using a 25-item version of the Big Five Personality Inventory, which measured the personality traits of neuroticism, extraversion, openness to experience, agreeableness, and conscientiousness.47 We created cross-age composites for each of the traits. In the analysis sample, these measures correlated with study members’ scores for general psychopathology (r = 0.38, P < .001 for neuroticism; r = 0.07, P = .10 for extraversion; r = 0.07, P = .12 for openness to experience; r = −0.27, P < .001 for agreeableness; and r = −0.29, P < .001 for conscientiousness).

e traits. In the analysis sample, these measures correlated with study members’ scores for general psychopathology (r = 0.38, P < .001 for neuroticism; r = 0.07, P = .10 for extraversion; r = 0.07, P = .12 for openness to experience; r = −0.27, P < .001 for agreeableness; and r = −0.29, P < .001 for conscientiousness). Child Externalizing and Internalizing Problems At 11 years of age, parents and teachers completed the Rutter Child Scale,48 a questionnaire that inquires about the major areas of behavioral and emotional functioning during the past year. Parents and teachers rated each behavior on the Rutter Child Scale as “does not apply” (score of 0), “applies somewhat” (score of 1), or “certainly applies” (score of 2). Child externalizing problems were assessed using scores for the 8-item antisocial scale and scores for 4 items that address hyperactivity. Items on the antisocial scale describe children who frequently fight, bully other children, lie, disobey, steal, destroy belongings, and have irritable tempers. Items that contribute to the measurement of hyperactivity describe children who are “very restless,” “hardly ever still,” “squirmy,” “fidgety,” and unable to “settle into anything.” Child internalizing problems were assessed using scores on 6 items that describe children who “worry about many things” and “often appear miserable,” “unhappy,” and “tearful.” Details about the reliability and validity of the parent and teacher versions of the scale have been described previously.49,50 Parent and teacher scores were averaged.

lems were assessed using scores on 6 items that describe children who “worry about many things” and “often appear miserable,” “unhappy,” and “tearful.” Details about the reliability and validity of the parent and teacher versions of the scale have been described previously.49,50 Parent and teacher scores were averaged. Covariates Study covariates included family-level risk factors known to relate to childhood lead exposure or adult psychopathology and personality, including family socioeconomic status, maternal IQ, and family history of mental illness. These measures are described in the eMethods and eFigure in the Supplement. Comparison of Members Who Were Tested for Lead Exposure at 11 Years of Age vs Those Not Tested A total of 579 study members had been tested for lead exposure during childhood (55.8% of the full cohort). Study members with and without (n = 458 [44.2%]) BLL data were similar on all study covariates, including their social class origins, their mother's IQ scores, and their family history of mental illness. However, as a group, those without BLL data had greater internalizing problems at 11 years of age (mean of −0.06 z-scored units for children with BLL data; mean of 0.10 z-scored units for children without BLL data; difference of 0.16; P = .02).

gins, their mother's IQ scores, and their family history of mental illness. However, as a group, those without BLL data had greater internalizing problems at 11 years of age (mean of −0.06 z-scored units for children with BLL data; mean of 0.10 z-scored units for children without BLL data; difference of 0.16; P = .02). Statistical Analysis First, we tested the association between childhood BLLs and adult general psychopathology using ordinary least-squares multiple linear regression. We also tested for specificity in the association between childhood lead exposure and adult psychopathology by examining whether BLLs were associated with scores on the internalizing, externalizing, and thought disorder factors. Each outcome was examined using 2 models: (1) a sex-adjusted model in which the outcome was regressed on childhood BLL and sex and (2) a fully adjusted model that included all covariates. We used these same models to test associations between childhood BLLs and scores on informant-reported measures of adult personality. Post hoc sensitivity tests were also conducted to examine possible sex differences in the association between childhood BLLs and the adult outcome measures.51 The models specified above were rerun with a sex × lead interaction term included; these terms were nonsignificant in all models (P values ranged from .13 to .66).

st hoc sensitivity tests were also conducted to examine possible sex differences in the association between childhood BLLs and the adult outcome measures.51 The models specified above were rerun with a sex × lead interaction term included; these terms were nonsignificant in all models (P values ranged from .13 to .66). Second, we repeated these analyses using measures of childhood externalizing and internalizing problems (ie, antisocial behavior, hyperactivity, and internalizing problems) in place of adult outcomes. These models allowed us to test whether the association between lead exposure and psychopathology could be seen as early as 11 years of age. Only members who had complete data on all covariates for each outcome were included in each model; no data were imputed. For adult psychopathology, 551 members (95.2%) were analyzed; for adult personality, 539 members (93.1%) were analyzed; and for childhood externalizing and internalizing problems, 552 members (95.3%) were analyzed. Lead level was analyzed as a continuous measure and is presented here in 5-μg/dL units with 95% CIs, which correspond to approximately 1 SD of BLL in the cohort. Association of lead exposure in childhood with adult personality differences is presented with regression coefficients and 97% CIs. P values were generated from t tests for the null hypothesis. A 2-tailed P < .05 was considered to be statistically significant.

hich correspond to approximately 1 SD of BLL in the cohort. Association of lead exposure in childhood with adult personality differences is presented with regression coefficients and 97% CIs. P values were generated from t tests for the null hypothesis. A 2-tailed P < .05 was considered to be statistically significant. Results Association of Lead Exposure in Childhood With Psychopathologic Measures In Adulthood Of 1037 original study members, 579 (55.8%) were tested for lead exposure at 11 years of age (311 [53.7%] male). Child BLLs ranged from 4 to 50 μg/dL (mean [SD], 11.08 [4.96] μg/dL; to convert to micromoles per liter, multiply by 0.0483). A total of 544 study members (94.0%) had BLLs above the current reference value for clinical attention (5 μg/dL).3 Figure 2 depicts the mean adult general psychopathology scores of members at each childhood BLL. Members with childhood BLLs above the historical international level of concern for clinical attention (>10 μg/dL)3 tested a mean of 2.52 points higher (95% CI, 0.14-4.90; P = .04) on general psychopathology than their peers with lower BLLs (after adjusting for covariates, 2.30 points higher; 95% CI, −0.02 to 4.62; P = .05).

s with childhood BLLs above the historical international level of concern for clinical attention (>10 μg/dL)3 tested a mean of 2.52 points higher (95% CI, 0.14-4.90; P = .04) on general psychopathology than their peers with lower BLLs (after adjusting for covariates, 2.30 points higher; 95% CI, −0.02 to 4.62; P = .05). Figure 2. Association of Childhood Blood Lead Level at 11 Years of Age With Adult General Psychopathology at 38 Years of Age (Unadjusted for Covariates) The mean general psychopathology scores (circles) in adulthood with 95% CIs (error bars) by childhood blood lead level are shown. Each 5-μg/dL–higher level of blood lead in childhood was associated with an additional 1.49-point higher score (95% CI, 0.22-2.77; P = .02) in adult general psychopathology on a scale standardized to a mean (SD) of 100 (15) (horizontal line). Of the 579 study members with childhood blood lead level measured, 551 (95.2%) also had present data on all the covariates and the psychopathology outcome measures.

1.49-point higher score (95% CI, 0.22-2.77; P = .02) in adult general psychopathology on a scale standardized to a mean (SD) of 100 (15) (horizontal line). Of the 579 study members with childhood blood lead level measured, 551 (95.2%) also had present data on all the covariates and the psychopathology outcome measures. Results from the multiple linear regressions testing associations between BLL at 11 years of age and psychopathology from 18 to 38 years of age are given in the Table. After adjusting for covariates, each 5-μg/dL increase in childhood BLL was associated with a 1.34-point increase (95% CI, 0.11-2.57; P = .03) in general psychopathology. Examination of models testing associations between BLLs and factor scores for internalizing, externalizing, and thought disorder symptoms indicated that the association between BLLs and general psychopathology was driven primarily by associations between childhood BLL and internalizing and thought disorder symptoms. After adjusting for covariates, each 5-μg/dL increase in childhood BLL was associated with a 1.41-point increase (95% CI, 0.19-2.62; P = .02) in internalizing and a 1.30-point increase (95% CI, 0.06-2.54; P = .04) in thought disorder symptoms.

tions between childhood BLL and internalizing and thought disorder symptoms. After adjusting for covariates, each 5-μg/dL increase in childhood BLL was associated with a 1.41-point increase (95% CI, 0.19-2.62; P = .02) in internalizing and a 1.30-point increase (95% CI, 0.06-2.54; P = .04) in thought disorder symptoms. Table. Association Between Childhood Blood Lead Level and Adult Psychopathology, Adult Personality Traits, and Childhood Externalizing and Internalizing Problemsa Variable Sex Adjusted Fully Adjusted b (95% CI) P Value b (95% CI) P Value Adult psychopathologyb General psychopathology 1.49 (0.22 to 2.77) .02 1.34 (0.11 to 2.57) .03 Externalizing symptoms 0.80 (−0.47 to 2.06) .21 0.73 (−0.52 to 1.97) .25 Internalizing symptoms 1.57 (0.30 to 2.83) .02 1.41 (0.19 to 2.62) .02 Thought disorder symptoms 1.44 (0.16 to 2.72) .03 1.30 (0.06 to 2.54) .04 Adult personality traits (Big Five Personality Inventory)c Neuroticism 0.10 (0.02 to 0.19) .01 0.10 (0.02 to 0.18) .02 Extraversion −0.08 (−0.17 to 0.01) .09 −0.09 (−0.17 to 0.004) .06 Openness to experience −0.07 (−0.16 to 0.03) .17 −0.07 (−017 to 0.03) .15 Agreeableness −0.09 (−0.17 to −0.003) .04 −0.09 (−0.18 to −0.01) .03 Conscientiousness −0.14 (−0.25 to −0.03) .01 −0.14 (−0.25 to −0.03) .01 Childhood externalizing and internalizing problemsd Antisocial behavior 0.11 (0.03 to 0.19) .01 0.10 (0.02 to 0.18) .02 Hyperactivity 0.17 (0.08 to 026) <.001 0.16 (0.07 to 0.25) <.001 Internalizing problems 0.12 (0.03 to 0.20) .01 0.11 (0.02 to 0.20) .01 a Covariates in the fully adjusted model were sex, childhood socioeconomic status, maternal IQ, and family history of mental illness. Of the 579 study members with childhood blood lead level measured, 551 (95.2%) had present data on all the covariates and the psychopathology outcome measures, 539 (93.1%) had present data on all the covariates and the personality outcome measures, and 552 (95.3%) had present data on all the covariates and the childhood emotion and behavior outcome measures. Regression coefficients indicate change in outcome per 5-μg/dL increase in childhood blood lead level.

utcome measures, 539 (93.1%) had present data on all the covariates and the personality outcome measures, and 552 (95.3%) had present data on all the covariates and the childhood emotion and behavior outcome measures. Regression coefficients indicate change in outcome per 5-μg/dL increase in childhood blood lead level. b General psychopathology and the constituent psychiatric spectra are standardized to a mean (SD) of 100 (15). c The Big Five Personality Inventory traits scores are standardized to a mean (SD) of 0 (1). d Childhood antisocial behavior, hyperactivity, and internalizing problem scores are standardized to a mean (SD) of 0 (1). Association of Lead Exposure in Childhood With Adult Personality Differences Results from the multiple linear regressions testing associations between childhood BLL and the informant-reported measures of adult personality are given in the Table. Consistent with the adult psychopathology results, after adjustment for covariates, study members with higher BLLs at 11 years of age were viewed in adulthood by their informants as more neurotic (b = 0.10; 95% CI, 0.02-0.08; P = .02), less agreeable (b = −0.09; 95% CI, −0.18 to −0.01; P = .03), and less conscientious (b = −0.14; 95% CI, −0.25 to −0.03; P = .01). There were no statistically significant associations with informant-rated extraversion (b = −0.09; 95% CI, −0.17 to 0.004; P = .06) and openness to experience (b = −0.07; 95% CI, −0.17 to 0.03; P = .15).

e (b = −0.09; 95% CI, −0.18 to −0.01; P = .03), and less conscientious (b = −0.14; 95% CI, −0.25 to −0.03; P = .01). There were no statistically significant associations with informant-rated extraversion (b = −0.09; 95% CI, −0.17 to 0.004; P = .06) and openness to experience (b = −0.07; 95% CI, −0.17 to 0.03; P = .15). Early Detection of Lead-Related Psychiatric Differences After the detection of significant associations between child BLLs and both adult psychopathology symptoms and difficult adult personality traits, we tested whether psychiatric problems related to lead exposure could be detected as early as 11 years of age, when BLLs were assessed. The Dunedin study reported in 1988 that children with higher BLLs at 11 years of age scored higher on concurrent parent-report measures of hyperactivity and inattention symptoms.9 We tested whether study members with higher BLLs at 11 years of age also scored higher on measures at 11 years of age that assessed a broader suite of early-life externalizing and internalizing problems, including parent- and teacher-report measures of antisocial behavior, hyperactivity, and internalizing problems. We found that they did score higher (Table), suggesting that the association between lead exposure and psychopathology may begin to manifest broadly well before adulthood.

alizing and internalizing problems, including parent- and teacher-report measures of antisocial behavior, hyperactivity, and internalizing problems. We found that they did score higher (Table), suggesting that the association between lead exposure and psychopathology may begin to manifest broadly well before adulthood. Discussion This multidecade, longitudinal analysis of the association between childhood BLLs and adult mental health and personality generated 3 findings. First, across nearly 3 decades of follow-up, childhood BLLs were associated with higher levels of general psychopathology, driven primarily by greater rates of internalizing and thought disorder symptoms. Second, childhood BLLs were associated with higher neuroticism, lower agreeableness, and lower conscientiousness. Third, childhood BLLs were associated with greater externalizing and internalizing symptoms assessed contemporaneously with BLL measurement at 11 years of age. Each of these findings remained significant after adjusting for members’ social class backgrounds, their mothers’ IQs, and their family histories of mental illness.

d, childhood BLLs were associated with greater externalizing and internalizing symptoms assessed contemporaneously with BLL measurement at 11 years of age. Each of these findings remained significant after adjusting for members’ social class backgrounds, their mothers’ IQs, and their family histories of mental illness. These results suggest that early-life lead exposure in the era of leaded gasoline experienced by individuals who are currently adults may have contributed to subtle, lifelong differences in emotion and behavior that are detectable at least up to 38 years of age. Are these differences clinically or practically meaningful? On the one hand, the effect sizes reflecting the associations between childhood lead exposure and adult psychopathology and personality difficulties are small (approximately r = 0.08). This size is approximately one-third the size of the association seen in the Dunedin study between psychopathology and other modifiable (eg, childhood maltreatment, r = 0.21) and nonmodifiable (eg, family history of mental illness, r = 0.23) risk factors.29 Childhood lead exposure may not be a major etiologic factor in adult psychiatric disease today. On the other hand, compared with other findings from this sample, the associations reported herein are similar to those reported for lead and IQ15 and are stronger than those reported for lead and criminal offending.20 On a population basis, even modest alterations in risk can lead to significant shifts in the overall burden of disease.

compared with other findings from this sample, the associations reported herein are similar to those reported for lead and IQ15 and are stronger than those reported for lead and criminal offending.20 On a population basis, even modest alterations in risk can lead to significant shifts in the overall burden of disease. The finding that associations between childhood BLLs and psychopathology symptoms were observable as early as the age of BLL testing suggests that lead-related alterations in emotion and behavior, however modest, likely emerge early and persist across the life course. Of note, in childhood, these psychopathology symptoms tended to involve more externalizing symptoms, particularly hyperactivity, whereas in adulthood, they tended to involve more internalizing symptoms. This finding suggests that lead-related alterations in emotion and behavior may demonstrate heterotypic continuity in their psychiatric presentation,52 with either one class of psychiatric disorders creating conditions that lead to another class (eg, when hyperactivity elicits harsh parenting, it may lead to anxiety and depression) or else the same underlying condition (eg, a general liability to psychopathology) presenting differently across different developmental windows.29,53

one class of psychiatric disorders creating conditions that lead to another class (eg, when hyperactivity elicits harsh parenting, it may lead to anxiety and depression) or else the same underlying condition (eg, a general liability to psychopathology) presenting differently across different developmental windows.29,53 The association between childhood lead exposure and adult personality traits also suggests that lead-related differences in adult emotion and behavior can be detected not only by asking individuals to self-report on their mental health symptoms but also by simply asking informants who know them well to describe their behavior. Childhood lead exposure may alter how people behave toward or are perceived by others across their lives. In other studies,33,35,54,55 the blend of personality traits seen in adults exposed to lead as children has been associated with a number of poor life outcomes, including more psychopathology, worse physical health, less job satisfaction, and troubled interpersonal relationships.

perceived by others across their lives. In other studies,33,35,54,55 the blend of personality traits seen in adults exposed to lead as children has been associated with a number of poor life outcomes, including more psychopathology, worse physical health, less job satisfaction, and troubled interpersonal relationships. The present study has implications for future research, public policy, and clinical practice. For researchers, these findings add further evidence to the suggestion that environmental toxins may affect important life outcomes through subtle changes in the way that individuals feel and behave. Future toxicologic studies should consider assessing these subjective outcomes alongside more objective ones, such as physical health. For policymakers and practitioners, the findings suggest that the generation of adult patients with a history of childhood lead exposure may benefit from increased screening and access to mental health services.56 As the generation of lead-exposed individuals age, it is also possible that bone loss during menopause and osteoporosis may result in childhood lead stored in bone being recirculated throughout the body, suggesting the testable hypothesis that the long-term consequences of childhood lead exposure may evolve or expand over time.57 It is possible that the pediatric challenges of the past may represent emerging concerns for geriatric psychiatry.

sult in childhood lead stored in bone being recirculated throughout the body, suggesting the testable hypothesis that the long-term consequences of childhood lead exposure may evolve or expand over time.57 It is possible that the pediatric challenges of the past may represent emerging concerns for geriatric psychiatry. Limitations This study has limitations. First, it used a single, predominantly white cohort born in the 1970s; therefore, its results will require replication in other samples from other countries. Second, although child BLLs in this New Zealand cohort are similar to those recorded in other developed countries at the time of testing,58,59 the high historical levels of lead exposure experienced by the Dunedin study members may not generalize to the relatively lower levels of exposure that are more common for children in developed countries today. Nevertheless, children in many developed and developing countries still experience high lead exposure from contaminated water, soil, paints, and pipes.60,61 Third, there was only one time point of lead testing, which precluded evaluation of sensitive periods for associations of lead with behavior or of the effects of cumulative lead dose by adulthood. Fourth, this study was observational and does not establish a causal relationship between lead and the tested outcomes.

,61 Third, there was only one time point of lead testing, which precluded evaluation of sensitive periods for associations of lead with behavior or of the effects of cumulative lead dose by adulthood. Fourth, this study was observational and does not establish a causal relationship between lead and the tested outcomes. Conclusions In this multidecade, longitudinal study of lead-exposed children, higher childhood BLLs were associated with more psychopathology across the life course and difficult adult personality traits. Childhood lead exposure may have long-term psychiatric and behavioral consequences. Supplement. eMethods. Supplemental Material eFigure. Two Alternative Correlational Structures of Early-Adult Psychopathology Click here for additional data file.

Introduction Antipsychotic polypharmacy is used among up to 30% of patients with schizophrenia.1 The use of antipsychotic polypharmacy has raised concern owing to the lack of evidence for its efficacy and safety as well as variable justifications and practice patterns.2,3,4,5,6,7 Meta-analyses of randomized clinical trials (RCTs) have shown mixed results,8,9,10,11,12 possibly because of shortcomings owing to low number of participants and lacking separations of high- vs low-quality studies. The most-recent meta-analysis without these limitations concluded that data from high-quality studies show beneficial outcomes only for negative symptom reduction with aripiprazole augmentation.12 Short-term symptom reduction, used as the primary outcome in RCTs, is an important measure for effectiveness of antipsychotic treatment. However, schizophrenia is a lifelong illness, and long-term outcome, including relapse prevention and avoidance of adverse physical morbidity and mortality effects owing to long-term antipsychotic load, is an even more important issue for the patients’ well-being.13,14 Conducting an RCT on these outcomes would require several thousands of patient-years, which is probably the reason why no such studies have been done.

d avoidance of adverse physical morbidity and mortality effects owing to long-term antipsychotic load, is an even more important issue for the patients’ well-being.13,14 Conducting an RCT on these outcomes would require several thousands of patient-years, which is probably the reason why no such studies have been done. Observational studies can overcome this problem by using large electronic databases. Results from 1 large observational study have shown that any antipsychotic polypharmacy was associated with an approximately 40% lower risk of rehospitalization and death compared with any monotherapy,15 but the major problem in observational studies is residual confounding related to selection bias. This limitation can be eliminated by using within-individual analyses in which each patient is used as his or her own control.16,17,18 To our knowledge, no observational study has been published on the comparative effectiveness of antipsychotic combinations vs monotherapies using this method. We aimed to study this issue, using within-individual analyses, in a nationwide cohort including all patients with schizophrenia in Finland.

own control.16,17,18 To our knowledge, no observational study has been published on the comparative effectiveness of antipsychotic combinations vs monotherapies using this method. We aimed to study this issue, using within-individual analyses, in a nationwide cohort including all patients with schizophrenia in Finland. Methods Study Population This study was based on a cohort of all persons with schizophrenia treated in the inpatient setting during 1972-2014 in Finland, identified from the Hospital Discharge register maintained by the National Institute of Health and Welfare. Data were also retrieved from the National Prescription register (maintained by Social Insurance Institution, 1995-2015) and National Death register (1972-2015). The Hospital Discharge register includes all inpatient hospital stays in Finland, recorded for all residents. The detailed information on the study population is given in the eMethods in the Supplement. The follow-up started January 1, 1996, for the prevalent cohort and at the first discharge from inpatient care for the incident cases. The follow-up time ended at death or December 31, 2015, whichever occurred first. We conducted analysis of the data from April 24 to June 15, 2018.

en in the eMethods in the Supplement. The follow-up started January 1, 1996, for the prevalent cohort and at the first discharge from inpatient care for the incident cases. The follow-up time ended at death or December 31, 2015, whichever occurred first. We conducted analysis of the data from April 24 to June 15, 2018. The research project was approved by the ethics committee of the Finnish National Institute for Health and Welfare. Further permissions were granted by pertinent institutional authorities at the National Institute for Health and Welfare of Finland, the Social Insurance Institution of Finland, and Statistics Finland. The study was registry based and no contact was made with the participants of the study, and therefore according to Finnish legislation informed consents were not needed. Exposure Antipsychotic dispensations were derived from the National Prescription register, defined by Anatomical Therapeutic Chemical classification code N05A, excluding lithium.19 The register includes reimbursed drug dispensations for the entire Finnish population but does not include drugs used during hospital stays. Data on dispensed drug and amount, also recorded in defined daily doses (World Health Organization19), were used for this study. Drug dispensations were modeled with PRE2DUP modeling, a method used to define drug use periods (ie, when drug use started and ended).20 The detailed indication of the drug use modeling is shown in the eMethods in the Supplement.

o recorded in defined daily doses (World Health Organization19), were used for this study. Drug dispensations were modeled with PRE2DUP modeling, a method used to define drug use periods (ie, when drug use started and ended).20 The detailed indication of the drug use modeling is shown in the eMethods in the Supplement. Outcomes Psychiatric rehospitalization (International Classification of Diseases, Tenth Revision codes F20-F29 as main diagnoses) and all-cause hospitalization were the primary outcome measures. Sensitivity analyses were conducted among incident (first-episode of schizophrenia) patients and for antipsychotic polypharmacy periods excluding the first 90 days of overlap. Hospitalization owing to physical illness and mortality were included as secondary outcomes to take into account that antipsychotic use may have adverse effects on physical health.

among incident (first-episode of schizophrenia) patients and for antipsychotic polypharmacy periods excluding the first 90 days of overlap. Hospitalization owing to physical illness and mortality were included as secondary outcomes to take into account that antipsychotic use may have adverse effects on physical health. Covariates The within-individual study design is based on the comparison of different time periods for the same person. Thus, all time-invariant covariates, such as sex, age, time since illness onset, comorbidities, and number of previous psychiatric hospitalizations at cohort entry, are controlled for in the design, and only time-varying covariates are adjusted for in the statistical analysis. Time-varying covariates were the order of antipsychotic exposures, time since cohort entry, and use of other psychotropic drugs (ie, antidepressants, benzodiazepines, lithium, mood stabilizers, sedatives). In this study, grouping of antipsychotics was not identical with the previous study,18 resulting in minimally different hazard ratios (HRs) for monotherapies. The traditional Cox proportional hazards regression models (between-individual analysis) were adjusted for sex, age at cohort entry, year of cohort entry, time since diagnosis, number of prior psychiatric hospitalizations, comorbidities, and drug use, as described in eTable 1 in the Supplement.

Covariates The within-individual study design is based on the comparison of different time periods for the same person. Thus, all time-invariant covariates, such as sex, age, time since illness onset, comorbidities, and number of previous psychiatric hospitalizations at cohort entry, are controlled for in the design, and only time-varying covariates are adjusted for in the statistical analysis. Time-varying covariates were the order of antipsychotic exposures, time since cohort entry, and use of other psychotropic drugs (ie, antidepressants, benzodiazepines, lithium, mood stabilizers, sedatives). In this study, grouping of antipsychotics was not identical with the previous study,18 resulting in minimally different hazard ratios (HRs) for monotherapies. The traditional Cox proportional hazards regression models (between-individual analysis) were adjusted for sex, age at cohort entry, year of cohort entry, time since diagnosis, number of prior psychiatric hospitalizations, comorbidities, and drug use, as described in eTable 1 in the Supplement. Statistical Analysis Hospitalization-based outcomes (psychiatric and all-cause hospitalization) were treated as recurrent events and analyzed with a stratified Cox proportional hazards regression model.16 In this within-individual design, each patient formed his or her own stratum, and follow-up time was reset to 0 after each outcome event (eFigure 1 in the Supplement). Persons who had both variation in exposure and experienced an outcome event during the follow-up contributed to within-individual analysis. The main analysis compared use of the following drugs in monotherapy and as 2-drug combinations with time when no antipsychotic was used: oral risperidone, quetiapine, clozapine, olanzapine, aripiprazole, and other oral formulations, as well as any long-acting injectable agent. These analyses were conducted in the prevalent cohort (including all patients) and in the incident cohort, including only patients with first-episode, and for both psychiatric and all-cause hospitalization as an outcome event.

pine, aripiprazole, and other oral formulations, as well as any long-acting injectable agent. These analyses were conducted in the prevalent cohort (including all patients) and in the incident cohort, including only patients with first-episode, and for both psychiatric and all-cause hospitalization as an outcome event. Sensitivity analyses were conducted by censoring the first 90 days from antipsychotic use to retrieve time of conservatively defined polypharmacy (ie, excluding switches between monotherapies). This censoring was also conducted for monotherapy periods in an identical way. Also, traditional multivariate-adjusted Cox proportional hazards regression analyses were conducted for all outcome events, and these models were adjusted for covariates provided in eTable 1 in the Supplement. The level of statistical significance was set at P < .0017 according to Bonferroni correction (0.05/29 = .0017). We used the traditional multivariate-adjusted Cox proportional hazards regression as secondary between-individual analyses. In within-individual analysis, only individuals with variation in the exposure (monotherapy, polypharmacy, no antipsychotic) and outcome (rehospitalization) contributed to the model, whereas, for between-individual analysis, all patients contributed to the model. The analyses were conducted using R, version 3.1.1 (R Foundation).

individual analysis, only individuals with variation in the exposure (monotherapy, polypharmacy, no antipsychotic) and outcome (rehospitalization) contributed to the model, whereas, for between-individual analysis, all patients contributed to the model. The analyses were conducted using R, version 3.1.1 (R Foundation). Results In the total cohort, including 62 250 patients, 31 257 individuals (50.2%) were men, and the median age was 45.6 (interquartile range, 34.6-57.9) years. The baseline characteristics of the cohort are reported in the Table. The follow-up time in this study was 20 years or less, with a median time of 14.1 years (interquartile range [IQR], 6.9-20.0 years) in the prevalent cohort and 10.1 years (IQR, 5.0-14.3) in the incident cohort (Table). During the follow-up, 58.8% (n = 36 631) of the prevalent cohort and 57.9% (n = 5045) of the incident cohort were readmitted for psychiatric inpatient care. Concerning rehospitalization rates after discontinuation of medication, the median time was 211 (IQR, 68-566) days.

(IQR, 5.0-14.3) in the incident cohort (Table). During the follow-up, 58.8% (n = 36 631) of the prevalent cohort and 57.9% (n = 5045) of the incident cohort were readmitted for psychiatric inpatient care. Concerning rehospitalization rates after discontinuation of medication, the median time was 211 (IQR, 68-566) days. Table. Characteristics of the Prevalent and Incident Cohorts and Hospitalizations During Follow-up Characteristic Cohort, No. (%) Prevalent (n = 62 250) Incident (n = 8719) Age at baseline, y ≤24 5368 (8.6) 1844 (21.2) 25-34 10 748 (17.3) 2297 (26.3) 35-44 13 996 (22.5) 1417(16.3) 45-54 13 767 (22.1) 1266 (14.5) 55-64 8833 (14.2) 763 (8.8) ≥65 9538 (15.3) 1132 (13.0) Median age (IQR), y 45.6 (34.6-57.9) 36.2 (26.2-52.3) Men 31 257 (50.2) 4898 (56.2) No. of all-cause hospitalizations 0 8617 (13.8) 1748 (20.0) 1 7948 (12.8) 1443 (16.6) 2-4 17 194 (27.6) 2603 (29.9) 5-8 12 423 (20.0) 1520 (17.4) ≥9 16 068 (25.8) 1405 (16.1) No. of all-cause hospitalizations per person, median (IQR) 4 (1-9) 3 (1-6) No. of psychiatric hospitalizations 0 25 619 (41.2) 3674 (42.1) 1 10 233 (16.4) 1615 (18.5) 2-4 13 490 (21.7) 1980 (22.7) 5-8 6273 (10.1) 805 (9.2) ≥9 6635 (10.7) 645 (7.4) No. of psychiatric hospitalizations per person, median (IQR) 1 (0-4) 1 (0-3) Follow-up time, median (IQR), y 14.1 (6.9-20.0) 10.1 (5.0-14.3) Abbreviation: IQR, interquartile range.

ns 0 25 619 (41.2) 3674 (42.1) 1 10 233 (16.4) 1615 (18.5) 2-4 13 490 (21.7) 1980 (22.7) 5-8 6273 (10.1) 805 (9.2) ≥9 6635 (10.7) 645 (7.4) No. of psychiatric hospitalizations per person, median (IQR) 1 (0-4) 1 (0-3) Follow-up time, median (IQR), y 14.1 (6.9-20.0) 10.1 (5.0-14.3) Abbreviation: IQR, interquartile range. Corresponding values for all-cause hospitalization were 86.2% (n = 53 633) of the prevalent cohort and 80.0% (n = 6971) of the incident cohort. A total of 67.2% (n = 41 812) of the patients in the total cohort and 54.1% (n = 4717) of those in the incident cohort used antipsychotic polypharmacy during the follow-up, and 57.5% (n = 35 793) of the prevalent cohort and 41.6% (n = 3627) of the incident cohort were exposed to antipsychotic polypharmacy for at least 90 days. Median doses of specific antipsychotics used in the prevalent and incident cohorts are described in eTable 2 in the Supplement.

c polypharmacy during the follow-up, and 57.5% (n = 35 793) of the prevalent cohort and 41.6% (n = 3627) of the incident cohort were exposed to antipsychotic polypharmacy for at least 90 days. Median doses of specific antipsychotics used in the prevalent and incident cohorts are described in eTable 2 in the Supplement. The overall view on the risks of psychiatric rehospitalization for specific treatments in the total cohort are presented in Figure 1 and eTable 3 in the Supplement. The corresponding results using monotherapies as the reference are shown in Figure 2 and Figure 3. The lowest risk of rehospitalization was observed for clozapine plus aripiprazole polypharmacy (Figure 2), being 14% lower (HR, 0.86; 95% CI, 0.79-0.94; P < .001) than that for clozapine, the monotherapy associated with the best outcomes. Among individuals who used clozapine both as monotherapy and polypharmacy, the mean clozapine dose was 426 mg/d during monotherapy and 399 mg/d during polypharmacy. When other specific polypharmacy combinations, excluding the composite group of any long-acting injectable agents, were compared with the better antipsychotic component of each combination, no combination was statistically superior to monotherapy when strict Bonferroni correction was applied. eTable 4 in the Supplement reports the outcome noted when any other antipsychotic was added to aripiprazole, clozapine, olanzapine, quetiapine, and risperidone.

the better antipsychotic component of each combination, no combination was statistically superior to monotherapy when strict Bonferroni correction was applied. eTable 4 in the Supplement reports the outcome noted when any other antipsychotic was added to aripiprazole, clozapine, olanzapine, quetiapine, and risperidone. Figure 1. Risk of Psychiatric Rehospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis) HR indicates hazard ratio; LAI, long-acting injectable agent. Orange markers indicate monotherapies. Figure 2. Risk of Psychiatric Rehospitalization in the Total Cohort, Compared With Clozapine, Aripiprazole, and Olanzapine Monotherapy (Within-Individual Analysis) HR indicates hazard ratio; LAI, long-acting injectable agent. Figure 3. Risk of Psychiatric Rehospitalization in the Total Cohort, Compared With Risperidone, Quetiapine, and Any Long-Acting Injectable Agent (LAI) Monotherapy (Within-Individual Analysis) HR indicates hazard ratio; LAI, long-acting injectable agent.

Figure 2. Risk of Psychiatric Rehospitalization in the Total Cohort, Compared With Clozapine, Aripiprazole, and Olanzapine Monotherapy (Within-Individual Analysis) HR indicates hazard ratio; LAI, long-acting injectable agent. Figure 3. Risk of Psychiatric Rehospitalization in the Total Cohort, Compared With Risperidone, Quetiapine, and Any Long-Acting Injectable Agent (LAI) Monotherapy (Within-Individual Analysis) HR indicates hazard ratio; LAI, long-acting injectable agent. The risk of psychiatric hospitalization, all-cause hospitalization, or death was not decreased significantly by adding any miscellaneous antipsychotic (most of those other than aripiprazole) to clozapine. However, including any other agent with quetiapine, which had the worst monotherapy response, resulted in a better outcome. At an aggregate level, the risk of psychiatric rehospitalization was 7% lower during any polypharmacy than any monotherapy period without censoring the first 90-day periods of antipsychotic treatment (HR, 0.93; 95% CI, 0.91-0.95; P < .001). The HR for all-cause hospitalization was 0.91 (95% CI, 0.89-0.92; P < .001), and for death, 0.76 (95% CI, 0.73-0.79; P < .001).

ehospitalization was 7% lower during any polypharmacy than any monotherapy period without censoring the first 90-day periods of antipsychotic treatment (HR, 0.93; 95% CI, 0.91-0.95; P < .001). The HR for all-cause hospitalization was 0.91 (95% CI, 0.89-0.92; P < .001), and for death, 0.76 (95% CI, 0.73-0.79; P < .001). eFigure 2 and eTable 5 in the Supplement provide the results for psychiatric rehospitalization when the first 90 days were censored from all treatment periods to eliminate potential artificial polypharmacy periods that may have occurred when one antipsychotic was switched to another. The superiority of clozapine plus aripiprazole (which had HR, 0.55; 95% CI, 0.51-0.61 vs no antipsychotic) over clozapine monotherapy (which had HR, 0.68; 95% CI, 0.66-0.70 vs no antipsychotic) was even greater (difference, 18%; HR, 0.82; 95% CI, 0.75-0.89; P < .001) in this conservatively defined polypharmacy analysis. In this analysis, the risk of rehospitalization was 13% lower during any polypharmacy than any monotherapy treatment (HR, 0.87; 95% CI, 0.85-0.88; P < .001).

CI, 0.66-0.70 vs no antipsychotic) was even greater (difference, 18%; HR, 0.82; 95% CI, 0.75-0.89; P < .001) in this conservatively defined polypharmacy analysis. In this analysis, the risk of rehospitalization was 13% lower during any polypharmacy than any monotherapy treatment (HR, 0.87; 95% CI, 0.85-0.88; P < .001). The risk of all-cause hospitalization is shown in eFigure 3 and eTable 6 in the Supplement. Again, clozapine plus aripiprazole was associated with a substantially better outcome than any other treatment. Figure 4 and eTable 7 in the Supplement show the results for the first-episode group with no antipsychotic use as reference. The superiority of the clozapine plus aripiprazole combination over any other treatments was even more robust than in the total cohort. When clozapine, the monotherapy associated with the best outcomes, was used as reference, the HR for clozapine plus aripiprazole combination was 0.78 (95% CI, 0.63-0.96) in the analysis including all polypharmacy periods, and 0.77 (95% CI, 0.63-0.95) in conservatively defined polypharmacy analysis. The risk of hospitalization owing to physical illness is given in eFigure 4 and eTable 8 in the Supplement, showing the lowest risks for polypharmacy and long-acting injectable agent monotherapy. Figure 4. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Incident Cohort (Within-Individual Analysis) HR indicates hazard ratio. LAI, long-acting injectable agent. Orange markers indicate monotherapies.

The risk of all-cause hospitalization is shown in eFigure 3 and eTable 6 in the Supplement. Again, clozapine plus aripiprazole was associated with a substantially better outcome than any other treatment. Figure 4 and eTable 7 in the Supplement show the results for the first-episode group with no antipsychotic use as reference. The superiority of the clozapine plus aripiprazole combination over any other treatments was even more robust than in the total cohort. When clozapine, the monotherapy associated with the best outcomes, was used as reference, the HR for clozapine plus aripiprazole combination was 0.78 (95% CI, 0.63-0.96) in the analysis including all polypharmacy periods, and 0.77 (95% CI, 0.63-0.95) in conservatively defined polypharmacy analysis. The risk of hospitalization owing to physical illness is given in eFigure 4 and eTable 8 in the Supplement, showing the lowest risks for polypharmacy and long-acting injectable agent monotherapy. Figure 4. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Incident Cohort (Within-Individual Analysis) HR indicates hazard ratio. LAI, long-acting injectable agent. Orange markers indicate monotherapies. The results for all-cause mortality are presented in eFigure 5 and eTable 9 in the Supplement. Among 29 different treatments, all monotherapies except clozapine were among the 10 worst, although all treatments were associated with a 50% or more lower risk of death compared with no antipsychotic use. The results for psychiatric hospitalizations with between-individual analysis including all patients, as well as also those without polypharmacy or hospitalizations, are presented in eFigure 6, eFigure 7, eTable 10, and eTable 11 in the Supplement. Clozapine and clozapine plus aripiprazole were again among the most favorable treatments.

ychiatric hospitalizations with between-individual analysis including all patients, as well as also those without polypharmacy or hospitalizations, are presented in eFigure 6, eFigure 7, eTable 10, and eTable 11 in the Supplement. Clozapine and clozapine plus aripiprazole were again among the most favorable treatments. Discussion To our knowledge, this is the first study on the long-term use of antipsychotic polypharmacy in schizophrenia. It is nearly impossible to conduct an RCT including tens of thousands of patient-years to achieve sufficient statistical power. Therefore, observational studies are the only way to investigate long-term comparative outcomes. The major shortcoming in observational studies is selection bias, because treatments are not chosen on a random basis. In this study, we used within-individual analyses in which each patient is used as his or her own control to minimize selection bias. We observed that when treatment for the same patient was switched back and forth between monotherapy and polypharmacy, the use of aripiprazole plus clozapine was associated with a 14% to 23% lower risk of psychiatric or all-cause hospitalization compared with clozapine monotherapy. Clozapine was the monotherapy associated with the best outcomes, and clozapine plus aripiprazole was associated with significantly better outcome than any other antipsychotic treatment, either as monotherapy or polypharmacy. Quetiapine was the least successful monotherapy, as has been observed also in previous Swedish and Finnish studies.17,18

rapy associated with the best outcomes, and clozapine plus aripiprazole was associated with significantly better outcome than any other antipsychotic treatment, either as monotherapy or polypharmacy. Quetiapine was the least successful monotherapy, as has been observed also in previous Swedish and Finnish studies.17,18 One possible explanation for the superiority of polypharmacy is that, in the real-world setting, treatment adherence is poor,21,22,23 and, if the patient has prescriptions for 2 antipsychotics, he or she may use at least 1 of them. However, other data suggest that the more medications or doses used, the more difficult it is for patients to adhere to the treatment regimen.22 Regarding the clozapine plus aripiprazole combination, the clozapine dose was only slightly lower during polypharmacy than during monotherapy, which suggests that reduction of the dose was not the major explanation for the better outcome. However, it is plausible that the different types of receptor profiles result in beneficial effects. For example, in a meta-analysis, addition of the partial dopamine D2 receptor agonist aripiprazole to clozapine therapy improved negative symptoms and reduced several adverse effects, such as weight gain and increased prolactin level, whereas a combination of 2 dopamine D2 antagonists was associated with greater prolactin elevation but less insomnia.12 Such a reduced adverse effect burden could also increase adherence.

clozapine therapy improved negative symptoms and reduced several adverse effects, such as weight gain and increased prolactin level, whereas a combination of 2 dopamine D2 antagonists was associated with greater prolactin elevation but less insomnia.12 Such a reduced adverse effect burden could also increase adherence. The mean daily doses of risperidone and quetiapine were low in our cohort, showing what happens in a total, nationwide cohort of patients with schizophrenia. It is probable that the prescribed doses had been higher, but the observed consumed doses (calculated on the basis of successively filled prescriptions in pharmacies) indicated the doses that patients actually used, and it is plausible that they titrated the doses to the level that they could tolerate. Therefore, patients may be willing to use 2 antipsychotics in relatively low doses but refuse to use high or moderate doses of monotherapy even though the total defined daily dose would be lower during monotherapy. Therefore, it would be misleading to presume that effectiveness (efficacy plus tolerability) would correlate with the total antipsychotic dose expressed as defined daily doses or chlorpromazine equivalents.

high or moderate doses of monotherapy even though the total defined daily dose would be lower during monotherapy. Therefore, it would be misleading to presume that effectiveness (efficacy plus tolerability) would correlate with the total antipsychotic dose expressed as defined daily doses or chlorpromazine equivalents. Our results revealed that, in general, antipsychotic polypharmacy was associated with an approximately 10% lower relative risk of psychiatric rehospitalization (corresponding to an approximately 6% lower absolute risk with an approximately 60% rehospitalization rate in the cohort) compared with antipsychotic monotherapy, translating into a number needed to treat of 10 to 20, which is generally considered clinically meaningful. This effect size is larger than the effect size of statin treatments for the prevention of cardiovascular incidents.24 To reach statistical significance for a result of this magnitude, the minimum number of patients to compare 2 groups of treatments is approximately 1000. This sample size requirement for a long-term study is probably a major reason why RCTs are difficult to conduct to answer questions about the relative effectiveness of several active treatments in the maintenance management of schizophrenia.

um number of patients to compare 2 groups of treatments is approximately 1000. This sample size requirement for a long-term study is probably a major reason why RCTs are difficult to conduct to answer questions about the relative effectiveness of several active treatments in the maintenance management of schizophrenia. In addition, secondary analyses on hospitalization owing to physical illness and mortality showed better outcomes for antipsychotic combinations than for monotherapy. Because add-on treatments are started when monotherapy is no longer sufficient to control for worsening of symptoms, it is likely that the effect sizes for the superiority of antipsychotic polypharmacy over monotherapy are underestimates. The results obtained with 90-day omission in each treatment period (conservatively defined polypharmacy analysis excluding switch periods) were more robust than those without 90-day omission and are probably more accurate estimates on the results of polypharmacy. The results from between-individual analyses, including all patients, showed similar rank order as within-individual analyses. This finding indicates that the clozapine plus aripiprazole combination is associated with the best outcome in the entire population of Finnish patients with schizophrenia. The results may generalize to other high-income countries with a majority white population but not necessarily to other societies. Only 1 other study has compared multiple antipsychotic combinations in relapse prevention.15 Our results are in line with those of Katona et al,15 showing a lower risk of rehospitalization during antipsychotic polypharmacy compared with monotherapy. However, our effect sizes were smaller, which may be explained by use of within-individual analysis, which minimizes selection bias.

tions in relapse prevention.15 Our results are in line with those of Katona et al,15 showing a lower risk of rehospitalization during antipsychotic polypharmacy compared with monotherapy. However, our effect sizes were smaller, which may be explained by use of within-individual analysis, which minimizes selection bias. Our results on mortality are in line with those of previous cohort studies, showing lower mortality during antipsychotic polypharmacy than monotherapy.15,25,26 Two small studies have reported a positive correlation between antipsychotic polypharmacy and mortality: one investigated the maximum number of concomitant antipsychotics used during a 10-year follow-up (n = 88)27 and the other studied the number of antipsychotics used (n = 99) at baseline of a 17-year follow-up.28 Because these studies were based on case records, their results may reveal associations between mortality and the number of antipsychotic prescriptions given to patients several years before death rather than actual past or current use based on filling of prescriptions. Our outcomes did not include functioning or quality of life, but these outcomes are reflected to some extent in mortality and psychiatric and somatic hospitalization rates.

antipsychotic prescriptions given to patients several years before death rather than actual past or current use based on filling of prescriptions. Our outcomes did not include functioning or quality of life, but these outcomes are reflected to some extent in mortality and psychiatric and somatic hospitalization rates. Strengths and Limitations This study has both strengths and limitations. We included all hospital-treated patients in Finland with up to 20-year follow-up and used within-individual analysis to minimize selection bias. In addition, time-varying covariates, such as time since cohort entry and order of exposures (ie, switch from monotherapy to polypharmacy vs switch from polypharmacy to monotherapy), were adjusted in the analysis. One limitation is that our main outcomes were risk of rehospitalization owing to psychiatric or somatic reasons, and our database did not include information on symptoms, reason for polypharmacy, quality of life, and level of functioning. We did not have information on concomitant psychosocial treatments, but it is unlikely that there would be any systematic differences between polypharmacy and monotherapy in this regard. In addition, clozapine and long-acting injectable agent treatments require regular contact with health care staff, which may be associated with better outcomes, but it is unlikely that the increased contact would explain the difference between polypharmacy and monotherapy. Another limitation of the study was that the results may generalize to only high-income countries with a majority white population. An additional limitation was protopathic-type bias (ie, the fact that add-on treatments are started when symptoms become worse). Therefore, the effect sizes for polypharmacy are probably underestimated.

tion of the study was that the results may generalize to only high-income countries with a majority white population. An additional limitation was protopathic-type bias (ie, the fact that add-on treatments are started when symptoms become worse). Therefore, the effect sizes for polypharmacy are probably underestimated. Conclusions Our results suggest that patients had the lowest risk of psychiatric or all-cause hospitalization when they received combination therapy with clozapine plus aripiprazole, which was significantly superior to clozapine, which was the monotherapy associated with the best outcomes. These results indicate that rational antipsychotic polypharmacy seems to be feasible by using 2 particular antipsychotics with different types of receptor profiles.

ation therapy with clozapine plus aripiprazole, which was significantly superior to clozapine, which was the monotherapy associated with the best outcomes. These results indicate that rational antipsychotic polypharmacy seems to be feasible by using 2 particular antipsychotics with different types of receptor profiles. Current treatment guidelines state that antipsychotic monotherapy should be preferred and polypharmacy should be avoided if possible. These recommendations reflect the recent evidence in high-quality studies on the acute-phase treatment. However, results from our study suggest that antipsychotic polypharmacy may be superior to monotherapy for maintenance treatment, which has not been examined with RCTs. Therefore, it should be acknowledged that statements about a preferential use of antipsychotic monotherapy for maintenance treatment of schizophrenia lack evidence, and that currently available evidence—although gathered with few nonrandomized cohort studies that have their own limitations—indicates the opposite. Therefore, the current treatment guidelines should modify their categorical recommendations discouraging all antipsychotic polypharmacy in the maintenance treatment of schizophrenia. Supplement. eMethods. Population and Exposure eTable 1. Covariate Definitions eTable 2. Median Doses of Antipsychotics in Defined Daily Doses (DDDs) and Milligrams (mg), With Interquartile Ranges (IQRs) eTable 3. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis)

Supplement. eMethods. Population and Exposure eTable 1. Covariate Definitions eTable 2. Median Doses of Antipsychotics in Defined Daily Doses (DDDs) and Milligrams (mg), With Interquartile Ranges (IQRs) eTable 3. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis) eTable 4. The Effect Of Adding Any Other Antipsychotic on Top of Aripiprazole, Clozapine, Olanzapine, Quetiapine, and Risperidone Monotherapy eTable 5. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis Conservatively Defined Polypharmacy) eTable 6. Risk of All-Cause Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis) eTable 7. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Incident Cohort (Within-Individual Analysis) eTable 8. Risk of Somatic Hospitalization During Specific Treatments Compared With No Antipsychotic Use (Within-Individual Analysis) eTable 9. Risk Of Death During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis) eTable 10. Risk of Psychiatric Hospitalization During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis)

eTable 9. Risk Of Death During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis) eTable 10. Risk of Psychiatric Hospitalization During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis) eTable 11. Risk of Psychiatric Re-Hospitalization During Specific Treatments Compared With No Antipsychotic Treatment in the Incident Cohort (Between-Individual Analyses) eFigure 1. Restructuring of Data for Within Individual Analyses eFigure 2. Risk of Psychiatric Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis) eFigure 3. Risk of All-Cause Hospitalization During Specific Treatments Compared With No Antipsychotic Use in the Prevalent Cohort (Within-Individual Analysis) eFigure 4. Risk of Somatic Hospitalization During Specific Treatments Compared With No Antipsychotic Use (Within-Individual Analysis) eFigure 5. Risk of Death During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis) eFigure 6. Risk of Psychiatric Hospitalization During Specific Antipsychotic Treatments Compared With No Use of Antipsychotic in the Total Cohort (Between-Individual Analysis) eFigure 7. Risk of Psychiatric Re-Hospitalization During Specific Treatments Compared With No Antipsychotic Treatment in the Incident Cohort (Between-Individual Analyses) Click here for additional data file.

ociated with genetic risk factors shared with hyperactivity-impulsivity and up to 16% with inattention (Figure 2). The models’ fit statistics and estimates are presented in eTables 7-9 in the Supplement. Analyses using the reduced hypomania scales yielded comparable findings (eTable 10 and eTable 11 in the Supplement). Discussion To our knowledge, this was the first twin study to explore the shared genetic and environmental factors associated with ADHD and hypomania symptoms in youths. Traits of ADHD across childhood and adolescence were associated with adolescent hypomania. More than a quarter of the variance for hypomania was associated with shared genetic risk factors for ADHD traits (range, 13%-29%). Hypomania-specific genetic risk factors accounted for 27% to 46% of its variance. Environmental factors played a negligible role in the ADHD-hypomania symptom association. The genetic overlap with hypomania was larger with hyperactivity-impulsivity (range, 10%-25%) compared with inattention (range, 6%-16%).

The continuous measures of ADHD and hypomanic symptoms were used for the remaining analyses. Twin correlations for ADHD and hypomanic traits are given in Table 3; the univariate twin analyses and fit statistics are presented in eTable 3 and eTable 4 in the Supplement. The assumption of equal means and variances across zygosity was not met for ADHD, suggesting sibling contrast pathways (eTable 3 in the Supplement). Moderate to strong heritability was found for ADHD (heritability, 0.51-0.74) and hypomania traits (heritability, 0.50-0.64). Table 3. Phenotypic, Twin, and Cross-Trait Cross-Twin Correlations for ADHD and Hypomania Variable rph Male Female MZM DZM MZF DZF DZOS Cross-Twin Correlations Age at ADHD symptoms onset, y 9 NA NA 0.68 (0.66 to 0.70) 0.24 (0.21 to 0.28) 0.64 (0.61 to 0.66) 0.24 (0.20 to 0.27) 0.32 (0.29 to 0.34) 15 NA NA 0.55 (0.50 to 0.59) 0.12 (0.05 to 0.18) 0.56 (0.52 to 0.60) 0.07 (0.01 to 0.13) 0.14 (0.09 to 0.19) 18 NA NA 0.53 (0.45 to 0.50) 0.24 (0.14 to 0.33) 0.52 (0.45 to 0.59) 0.17 (0.08 to 0.26) 0.22 (0.15 to 0.29) Age at hyperactivity-impulsivity onset, y 9 or 12 NA NA 0.69 (0.66 to 0.71) 0.19 (0.15 to 0.22) 0.65 (0.62 to 0.67) 0.19 (0.16 to 0.23) 0.24 (0.22 to 0.27) 18 NA NA 0.39 (0.28 to 0.48) 0.14 (0.04 to 0.23) 0.49 (0.41 to 0.56) 0.08 (to 0.02 to 0.17) 0.12 (0.04 to 0.19) Age at inattention onset, y 9 or 12 NA NA 0.60 (0.57 to 0.63) 0.14 (0.10 to 0.17) 0.54 (0.51 to 0.56) 0.12 (0.08 to 0.15) 0.20 (0.17 to 0.23) 18 NA NA 0.52 (0.44 to 0.59) 0.19 (0.09 to 0.29) 0.48 (0.40 to 0.55) 0.05 (to 0.05 to 0.14) 0.19 (0.12 to 0.26) 15 NA NA 0.78 (0.75 to 0.81) 0.55 (0.49 to 0.60) 0.77 (0.74 to 0.80) 0.51 (0.45 to 0.57) 0.52 (0.48 to 0.56) 18 0.61 (0.54 to 0.67) 0.29 (0.19 to 0.37) 0.66 (0.60 to 0.70) 0.42 (0.33 to 0.50) 0.30 (0.23 to 0.37) Cross-Trait Cross-Twin Correlations Between ADHD and Hypomania at Age 15 y Age at ADHD symptom onset, y 9 or 12 0.28 (0.24 to 0.31) 0.32 (0.29 to 0.36) 0.26 (0.22 to 0.30) 0.07 (0.02 to 0.12) 0.26 (0.22 to 0.30) 0.13 (0.08 to 0.18) 0.14 (0.10 to 0.17) 15 0.43 (0.40 to 0.43) 0.45 (0.42 to 0.47) 0.36 (0.31 to 0.39) 0.16 (0.15 to 0.21) 0.36 (0.32 to 0.39) 0.19 (0.14 to 0.24) 0.17 (0.14 to 0.21) 18 0.39 (0.34 to 0.44) 0.38 (0.33 to 0.42) 0.31 (0.24 to 0.37) 0.20 (0.13 to 0.28) 0.36 (0.30 to 0.41) 0.21 (0.13 to 0.28) 0.19 (0.14 to 0.25) Age at hyperactivity-impulsivity onset, y 9 or 12 0.28 (0.25 to 0.32) 0.30 (0.27 to 0.34) 0.28 (0.23 to 0.32) 0.06 (0.05 to 0.11) 0.26 (0.22 to 0.30) 0.11 (0.06 to 0.16) 0.12 (0.09 to 0.15) 18 0.38 (0.33 to 0.43) 0.39 (0.34 to 0.44) 0.29 (0.21 to 0.36) 0.19 (0.12 to 0.27) 0.33 (0.28 to 0.39) 0.15 (0.07 to 0.23) 0.17 (0.11 to 0.23) Age at inattention onset, y 9 0.23 (0.19 to 0.26) 0.27 (0.24 to 0.31) 0.20 (0.15 to 0.25

to 0.32) 0.06 (0.05 to 0.11) 0.26 (0.22 to 0.30) 0.11 (0.06 to 0.16) 0.12 (0.09 to 0.15) 18 0.38 (0.33 to 0.43) 0.39 (0.34 to 0.44) 0.29 (0.21 to 0.36) 0.19 (0.12 to 0.27) 0.33 (0.28 to 0.39) 0.15 (0.07 to 0.23) 0.17 (0.11 to 0.23) Age at inattention onset, y 9 0.23 (0.19 to 0.26) 0.27 (0.24 to 0.31) 0.20 (0.15 to 0.25 ) 0.10 (0.04 to 0.15) 0.20 (0.16 to 0.25) 0.13 (0.08 to 0.18) 0.11 (0.07 to 0.15) 18 0.33 (0.27 to 0.38) 0.30 (0.25 to 0.35) 0.27 (0.20 to 0.34) 0.19 (0.11 to 0.27) 0.31 (0.25 to 0.37) 0.19 (0.10 to 0.27) 0.17 (0.11 to 0.22) Cross-Trait Cross-Twin Correlations Between ADHD and Hypomania at Age 18 y Age at ADHD symptom onset, y 9 or 12 0.26 (0.21 to 0.30) 0.31 (0.27 to 0.35) 0.18 (0.13 to 0.24) 0.17 (0.10 to 0.23) 0.27 (0.22 to 0.32) 0.08 (0.02 to 0.15) 0.12 (0.07 to 0.17) 15 0.29 (0.24 to 0.34) 0.35 (0.31 to 0.40) 0.18 (0.11 to 0.25) 0.15 (0.06 to 0.22) 0.31 (0.25 to 0.37) 0.08 (0.00 to 0.15) 0.15 (0.09 to 0.22) 18 0.45 (0.41 to 0.48) 0.51 (0.48 to 0.54) 0.30 (0.23 to 0.35) 0.23 (0.15 to 0.30) 0.42 (0.37 to 0.46) 0.21 (0.13 to 0.28) 0.19 (0.14 to 0.24) Age at hyperactivity-impulsivity onset, y 9 or 12 0.28 (0.24 to 0.33) 0.32 (0.27 to 0.36) 0.23 (0.17 to 0.28) 0.15 (0.08 to 0.22) 0.29 (0.24 to 0.34) 0.09 (0.02 to 0.16) 0.12 (0.06 to 0.17) 18 0.48 (0.45 to 0.51) 0.52 (0.49 to 0.56) 0.32 (0.25 to 0.39) 0.18 (0.10 to 0.25) 0.41 (0.36 to 0.46) 0.20 (0.12 to 0.27) 0.16 (0.11 to 0.22) Age at inattention onset, y 9 or 12 0.22 (0.18 to 0.26) 0.25 (0.20 to 0.29) 0.14 (0.08 to 0.20) 0.15 (0.09 to 0.22) 0.20 (0.15 to 0.26) 0.06 (−0.01 to 0.12) 0.10 (0.05 to 0.15) 18 0.38 (0.34 to 0.42) 0.43 (0.40 to 0.47) 0.24 (0.18 to 0.30) 0.21 (0.14 to 0.28) 0.36 (0.31 to 0.41) 0.16 (0.09 to 0.23) 0.17 (0.12 to 0.22) Abbreviations: ADHD, attention-deficit/hyperactivity disorder; DZF, dizygotic female twin pairs; DZM, dizygotic male twin pairs; DZOS, dizygotic opposite-sex twin pairs; MZF, monozygotic female twin pairs; MZM, monozygotic male twin pairs; NA, not applicable; rph, phenotypic correlation.

Introduction Early identification of individuals at high risk of bipolar disorder (BD) is essential for prevention and intervention and can be aided by investigating hypomanic symptoms in youths. Bipolar disorder is characterized by hypomanic or manic episodes. Hypomania is common in youths, with up to 10% identified as being at high risk of BD based on the clustering, duration, and impairment of symptoms. Subsyndromal hypomanic symptoms or traits have been linked to subsequent manic or hypomanic episodes and BD onset. There is evidence that BD is preceded by childhood ADHD; a population-based study found higher incidence rates of BD among those with a history of ADHD (incidence rate, 23.86) compared with those without (incidence rate, 2.17). The comorbidity rates of BD-ADHD are higher than expected by chance; the weighted mean prevalence of ADHD in pediatric bipolar samples is 48%. A study that used modified BD diagnostic criteria characterized by nonepisodic irritability and ultrarapid cycling in pediatric samples reported high ADHD-BD comorbidity rates of 74% to 98%. The focus on nonepisodic irritability, which also covers temper outbursts and emotion dysregulation, may have increased these comorbidity rates because such symptoms are associated features of ADHD. After much debate, there is some consensus that nonepisodic irritability is more characteristic of severe mood dysregulation or disruptive dysregulation disorder rather than BD. Thus, care needs to be taken to differentiate between BD and these syndromes.

ity rates because such symptoms are associated features of ADHD. After much debate, there is some consensus that nonepisodic irritability is more characteristic of severe mood dysregulation or disruptive dysregulation disorder rather than BD. Thus, care needs to be taken to differentiate between BD and these syndromes. Significant, modest correlations between adolescent hypomanic and hyperactivity symptoms have also been reported. The ADHD symptom domains of hyperactivity-impulsivity and inattention may be differentially associated with BD. One study found that BD was associated with inattentive and combined but not hyperactive-impulsive ADHD presentations. Others have reported similar levels of inattention and hyperactivity-impulsivity in patients with BD. These disparate findings need to be addressed with further research. The co-occurrence of ADHD and BD is associated with worse outcomes, including higher rates of comorbidity and suicide attempts, compared with BD or ADHD alone. It is crucial to determine the origins of the ADHD-BD overlap, distinguishing inattention from hyperactivity-impulsivity, to avoid such outcomes.

ther research. The co-occurrence of ADHD and BD is associated with worse outcomes, including higher rates of comorbidity and suicide attempts, compared with BD or ADHD alone. It is crucial to determine the origins of the ADHD-BD overlap, distinguishing inattention from hyperactivity-impulsivity, to avoid such outcomes. Shared genetic risk factors are postulated to be partly responsible for the ADHD-BD association and their related symptoms. A moderate genetic correlation between ADHD and BD II using family data has been reported. That study focused on BD not hypomania and did not distinguish between the ADHD presentations. The sample had a broad age range; thus, it is unclear whether the results apply to different age groups. Research focused on childhood and adolescence would be useful because the initial emergence of psychopathologic symptoms occurs at these developmental stages. This was the first twin study, to our knowledge, to explore the extent to which genetic and environmental risk factors for hypomanic traits are associated with ADHD symptoms in youths, examining inattention and hyperactivity-impulsivity separately.

ence of psychopathologic symptoms occurs at these developmental stages. This was the first twin study, to our knowledge, to explore the extent to which genetic and environmental risk factors for hypomanic traits are associated with ADHD symptoms in youths, examining inattention and hyperactivity-impulsivity separately. Methods Participants This twin study used data from 13 532 twin pairs who participated in the Child and Adolescent Twin Study in Sweden, a longitudinal, prospective twin study performed from December 20, 2017, to December 5, 2018. Their parents provided ADHD data when children were 9 or 12 years of age (response rate, 75%; 3951 monozygotic [MZ] twins and 9581 dizygotic [DZ] twins). Of those who reached 15 years of age, 3784 participated (response rate of those eligible, 61%; 1115 MZ twins and 2669 DZ twins). Of those who reached 18 years or older, 3013 participated (response rate of those eligible, 59%; 983 MZ twins and 2030 DZ twins). Pairs were excluded if either twin had a known brain injury or chromosomal disorder (n = 207). Data analysis was performed at the Department of Medical Epidemiology & Biostatistics, Karolinska Institutet, Stockholm, Sweden, from March 1, 2018, to October 31, 2018. Parents provided consent for themselves and their children to participate at 9 or 12 years of age, and the twins and their parents gave separate written informed consent at subsequent waves after receiving the study description. The study was approved by the Karolinska Institutet Ethical Review Board. All data were deidentified.

ovided consent for themselves and their children to participate at 9 or 12 years of age, and the twins and their parents gave separate written informed consent at subsequent waves after receiving the study description. The study was approved by the Karolinska Institutet Ethical Review Board. All data were deidentified. Measures Hypomania was assessed at 15 years of age using the parent-rated Child Mania Rating Scale (CMRS), which distinguishes children with BD from children with ADHD and healthy control individuals with high sensitivity (sensitivity for children with ADHD, 0.84; sensitivity for controls, 0.90) and specificity (specificity for children with ADHD, 0.92; specificity for controls, 0.96). The parent-rated Mood Disorders Questionnaire was used to assess hypomanic symptoms at 18 years of age, with high sensitivity (sensitivity, 0.72) and specificity (specificity, 0.81) in identifying adolescent BD. Both instruments cover symptoms that are more specific to mania (eg, hypersexuality and grandiosity) compared with other forms of psychopathology (ADHD). Further details of all measures are presented in eTable 1 in the Supplement.

ity (sensitivity, 0.72) and specificity (specificity, 0.81) in identifying adolescent BD. Both instruments cover symptoms that are more specific to mania (eg, hypersexuality and grandiosity) compared with other forms of psychopathology (ADHD). Further details of all measures are presented in eTable 1 in the Supplement. The ADHD symptoms at 9 and 12 years of age were assessed using the Autism-Tics, ADHD, and Other Comorbidities Inventory (A-TAC), a structured telephone interview completed by parents. The A-TAC ADHD domain consists of impulsivity and activity as well as concentration and attention modules, corresponding to the DSM-IV ADHD criteria. Parents rated ADHD symptoms when the twins were 15 years of age using the Strengths and Difficulties Questionnaire hyperactivity subscale, and the ADHD DSM-IV subscale of the Adult Behavior Checklist was used when twins were 18 years of age. The Adult Behavior Checklist was divided into hyperactivity-impulsivity (6 items) and inattention (7 items) domains based on similarity with DSM-5 criteria. Participants who had received an ADHD (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) code, F90) and/or BD diagnosis (ICD-10 codes, F30-F31) were identified using the Swedish National Patient Register, which records all specialist inpatient and outpatient care given to residents of Sweden. Cases of BD were also identified through lithium prescriptions using the Prescribed Drug Register, which covers all medications prescribed to residents of Sweden since 2005.

e identified using the Swedish National Patient Register, which records all specialist inpatient and outpatient care given to residents of Sweden. Cases of BD were also identified through lithium prescriptions using the Prescribed Drug Register, which covers all medications prescribed to residents of Sweden since 2005. Statistical Analysis Positively skewed variables were log-transformed. The birth year associations were included as a covariate in all twin analyses, and means were permitted to differ by sex to account for mean sex differences. Participants were split into groups based on published cutoffs for the hypomania (CMRS: score of ≥10; Mood Disorders Questionnaire: minimum of ≥7 symptoms clustered in same period with at least moderate impairment) and ADHD (A-TAC: broad cutoff: ≥6; strict cutoff: ≥12) instruments (eTable 1 in the Supplement). In separate analyses, participants who received an ADHD and/or BD diagnosis in the Swedish National Patient Register (and/or a prescription of lithium in the Prescribed Drug Register for BD) were compared with those without such diagnoses. To test whether ADHD traits at each age (number of symptoms, screening diagnoses, and clinical diagnoses) were associated with hypomanic traits at 15 and 18 years of age, linear regressions within generalized estimating equations (GEEs) were performed with ADHD as the exposure and hypomania the outcome. This approach allows for clustering of related individuals and calculates robust SEs. To assess the associations between each ADHD diagnosis and high risk of BD, we implemented logistic regressions that calculated odds ratios within a GEE framework, adjusting for sex and birth year. The GEEs were implemented in the drgee package of R. A 1-sided P < .05 was considered to be statistically significant.

09 to 0.23) 0.17 (0.12 to 0.22) Abbreviations: ADHD, attention-deficit/hyperactivity disorder; DZF, dizygotic female twin pairs; DZM, dizygotic male twin pairs; DZOS, dizygotic opposite-sex twin pairs; MZF, monozygotic female twin pairs; MZM, monozygotic male twin pairs; NA, not applicable; rph, phenotypic correlation. Phenotypic and cross-trait cross-twin correlations are presented in Table 3. An A, C, and E model was chosen as best fitting, with sibling interaction paths for ADHD. The proportions of variation in each hypomania scale that were associated with genetic and environmental risk factors unique to hypomania and shared with ADHD traits are shown in Figure 2. In total, 21% to 22% of the variance in hypomania at 15 years of age and 13% to 29% at 18 years of age was associated with genetic factors shared with ADHD at any age. Nonshared environmental factors associated with ADHD played a negligible role in hypomania. Hypomania-specific genetic factors accounted for 25% to 42% of its variance (eTable 5 in the Supplement). Similar results were found with the reduced hypomania scales (eTable 6 in the Supplement).

Es. To assess the associations between each ADHD diagnosis and high risk of BD, we implemented logistic regressions that calculated odds ratios within a GEE framework, adjusting for sex and birth year. The GEEs were implemented in the drgee package of R. A 1-sided P < .05 was considered to be statistically significant. Twin Analyses The classic twin method was used to investigate the degree to which genetic and environmental risk factors for ADHD traits were associated with hypomanic symptoms. This method relies on comparing the correlations between MZ twins who share all their segregating DNA code and DZ twins who share approximately 50% of their segregating DNA code. On the basis of this information, variance in and among phenotypes can be decomposed into additive genetic risk factors (A), nonadditive genetic risk factors (D), common or shared environmental risk factors (C; common to both twins and increase their similarity), and nonshared or unique environmental risk factors (E; related to environmental factors that differ across twins, including measurement error). The general principles of the twin design are described in detail elsewhere. Cross-trait cross-twin correlations involve correlating one twin’s ADHD score with their co-twin’s hypomania score; by calculating these correlations separately for MZ and DZ twins, the degree to which genetic and environmental factors affect the covariance between ADHD and hypomania can be estimated. The eMethods in the Supplement give the analytic codes used in the analyses.

twin’s ADHD score with their co-twin’s hypomania score; by calculating these correlations separately for MZ and DZ twins, the degree to which genetic and environmental factors affect the covariance between ADHD and hypomania can be estimated. The eMethods in the Supplement give the analytic codes used in the analyses. Univariate models were used to assess the relative contribution of A, D, C, and E to each measure and to test assumptions of the twin design. We assessed whether genetic and environmental risk factors for ADHD traits are associated with hypomania by fitting a multivariate Cholesky decomposition to the data (Figure 1). The proportion of variance in each trait that was associated with A, D, C, and E was estimated. Variance in hypomania that is associated with genetic risk factors for ADHD was also estimated (the pathways from latent variables A1, A2, and A4 to the hypomania scales in Figure 1). The pathway from variable A3 to hypomania at 15 years of age represents the proportion of genetic variance in hypomania at 15 years of age that was independent of ADHD traits. Hypomania at 18 years of age is associated with genetic risk factors for ADHD traits (A1, A2, and A4), genetic risk factors are associated with hypomania at 15 years of age (A3), and genetic risk factors unique to hypomania at 18 years of age (A5). Equivalent pathways are included for C, D, and E. Squaring these pathways gives the proportion of variance in each trait accounted for by each pathway.

aits (A1, A2, and A4), genetic risk factors are associated with hypomania at 15 years of age (A3), and genetic risk factors unique to hypomania at 18 years of age (A5). Equivalent pathways are included for C, D, and E. Squaring these pathways gives the proportion of variance in each trait accounted for by each pathway. Figure 1. Path Diagram for Cholesky Decomposition The pathways from the latent variables (enclosed in circles) labeled A1, A2, and A4 to the 2 hypomania scales estimate the proportion of variation in hypomania that is associated with genetic risk factors for attention-deficit/hyperactivity disorder (ADHD) traits at 9, 12, 15, and 18 years of age. The pathway from A3 to hypomania at 15 years of age represents the proportion of genetic variance in hypomania at 15 years of age that is independent of ADHD traits. Hypomania at 18 years of age was associated with genetic risk factors for ADHD traits at each age (A1, A2, and A4), genetic risk factors for hypomania at 15 years of age (A3), and genetic factors that are unique to hypomania at 18 years of age (A5). Equivalent pathways are included for C, D, and E.

dent of ADHD traits. Hypomania at 18 years of age was associated with genetic risk factors for ADHD traits at each age (A1, A2, and A4), genetic risk factors for hypomania at 15 years of age (A3), and genetic factors that are unique to hypomania at 18 years of age (A5). Equivalent pathways are included for C, D, and E. Because univariate analysis implicated C for hypomania and not ADHD, we fitted a model in which A, D, and E were associated with ADHD and A, C, and E were associated with hypomania (C and D did not contribute to the covariance among phenotypes in this model), as well as the A, C, and E and A, D, and E models. Whenever D was estimated, sibling interaction paths were also included in the model because these factors can mimic that of D with twin correlations and are implicated when the assumption of equal variances across zygosity is not met. Each model was fitted with separate variance and covariance components by sex (quantitative sex limitation). The significance of these sex differences was tested by constraining all pathways to be equal by sex. We tested further nested models to assess the significance of individual groups of variance and covariance components. Model fit was assessed using Bayesian information criteria, which outperforms alternative fit statistics when fitting multivariate models to large samples. Lower Bayesian information criteria values indicate better fitting models. All analyses were repeated for both ADHD dimensions. The Strengths and Difficulties Questionnaire hyperactivity scale was omitted from these analyses because it only covers hyperactivity.

rs shared with ADHD at any age. Nonshared environmental factors associated with ADHD played a negligible role in hypomania. Hypomania-specific genetic factors accounted for 25% to 42% of its variance (eTable 5 in the Supplement). Similar results were found with the reduced hypomania scales (eTable 6 in the Supplement). Figure 2. Proportion of the Genetic and Nonshared Environmental Risk Factors for Symptoms of Attention-Deficit/Hyperactivity Disorder (ADHD), Hyperactivity-Impulsivity, and Inattention Across Childhood and Adolescence That Are Associated With the Variation in Adolescent Hypomania by Sex Hyperactivity-impulsivity was more strongly associated with hypomanic symptoms compared with inattention (Table 2). Up to 25% of the variance in hypomania was associated with genetic risk factors shared with hyperactivity-impulsivity and up to 16% with inattention (Figure 2). The models’ fit statistics and estimates are presented in eTables 7-9 in the Supplement. Analyses using the reduced hypomania scales yielded comparable findings (eTable 10 and eTable 11 in the Supplement).

istics when fitting multivariate models to large samples. Lower Bayesian information criteria values indicate better fitting models. All analyses were repeated for both ADHD dimensions. The Strengths and Difficulties Questionnaire hyperactivity scale was omitted from these analyses because it only covers hyperactivity. Results A total of 13 532 twin pairs (3951 MZ twins, 9581 DZ twins, 2031 female MZ pairs, 1920 male MZ pairs, 22231 female DZ pairs, 2582 male DZ pairs, and 4778 opposite-sex DZ pairs) participated in this study at 9 or 12 years of age, with 3784 followed up at 15 years of age and 3013 at 18 years or older. The descriptive statistics by sex and zygosity are presented in Table 1. The results of the GEEs testing the association between ADHD and hypomania traits are presented in Table 2. Symptoms of ADHD were significantly associated with hypomania at 15 years of age (β = 0.30; 95% CI, 0.26-0.34) and 18 years of age (β = 0.19; 95% CI, 0.16-0.22) after adjustment for sex and birth year. Removal of hypomania items that were similar to ADHD items (2 of 10 CMRS items and 5 of 13 Mood Disorders Questionnaire items) did not affect the results (eTable 2 in the Supplement). All diagnostic definitions of ADHD were significantly associated with being at high risk of BD (Table 2). The rates of clinically diagnosed ADHD among the 52 individuals with a formal diagnosis of BD and/or a prescription of lithium (37%) were significantly higher than that among controls (4%) (odds ratio, 15.41; 95% CI, 8.55-27.76; P < .001).

s of ADHD were significantly associated with being at high risk of BD (Table 2). The rates of clinically diagnosed ADHD among the 52 individuals with a formal diagnosis of BD and/or a prescription of lithium (37%) were significantly higher than that among controls (4%) (odds ratio, 15.41; 95% CI, 8.55-27.76; P < .001). Table 1. ADHD and Hypomania Measures by Sex and Zygosity Variable Mean (SD) Score Range Total Male Female MZM DZM MZF DZF DZOS Male Female ADHD Assessments Age, y 9 or 12 2.02 (2.97) 2.47 (3.42) 1.56 (2.61) 2.24 (3.21) 2.55 (3.51) 1.44 (2.43) 1.79 (2.90) 2.56 (3.48) 1.46 (2.47) 0-19 15 1.83 (1.84) 2.11 (1.94) 1.57 (1.71) 1.99 (1.82) 2.08 (1.95) 1.45 (1.62) 1.75 (1.85) 2.23 (2.01) 1.51 (1.64) 0-8 18 2.15 (2.97) 2.43 (3.17) 1.88 (2.74) 2.14 (2.78) 2.41 (3.20) 1.72 (2.50) 2.12 (3.05) 2.75 (3.46) 1.84 (2.68) 0-24 ADHD Subscales Age of 9 or 12 y Hyperactivity-impulsivity 0.97 (1.67) 1.17 (1.85) 0.77 (1.44) 1.09 (1.74) 1.21 (1.91) 0.70 (1.32) 0.87 (1.60) 1.19 (1.86) 0.74 (1.39) 0-10 Inattention 1.05 (1.75) 1.30 (1.93) 0.79 (1.50) 1.16 (1.79) 1.34 (1.98) 0.74 (1.43) 0.92 (1.63) 1.37 (1.98) 0.72 (1.41) 0-9 Age of 18 y Hyperactivity-impulsivity 0.73 (1.40) 0.76 (1.43) 0.70 (1.37) 0.65 (1.24) 0.81 (1.50) 0.60 (1.22) 0.82 (1.54) 0.80 (1.52) 0.69 (1.35) 0-12 Inattention 1.42 (1.91) 1.68 (2.07) 1.19 (1.72) 1.49 (1.86) 1.60 (2.03) 1.11 (1.61) 1.31 (1.86) 1.95 (2.27) 1.15 (1.68) 0-13 Hypomania Measures Age, y 15 1.80 (2.43) 1.60 (2.32) 2.00 (2.53) 1.47 (2.16) 1.66 (2.43) 1.93 (2.37) 2.07 (2.69) 1.66 (2.32) 1.99 (2.52) 0-24 18 0.93 (1.85) 0.91 (1.86) 0.95 (1.85) 0.79 (1.68) 0.93 (1.90) 0.90 (1.75) 0.95 (1.90) 0.99 (1.97) 1.01 (1.92) 0-13 Abbreviations: ADHD, attention-deficit hyperactivity disorder; DZF, dizygotic female twin pairs; DZM, dizygotic male twin pairs; DZOS, dizygotic opposite-sex twin pairs; MZF, monozygotic female twin pairs; MZM, monozygotic male twin pairs.

85) 0.79 (1.68) 0.93 (1.90) 0.90 (1.75) 0.95 (1.90) 0.99 (1.97) 1.01 (1.92) 0-13 Abbreviations: ADHD, attention-deficit hyperactivity disorder; DZF, dizygotic female twin pairs; DZM, dizygotic male twin pairs; DZOS, dizygotic opposite-sex twin pairs; MZF, monozygotic female twin pairs; MZM, monozygotic male twin pairs. Table 2. Associations Between ADHD and Hypomania Variable β (95% CI)a SE 1-Sided P Value Association Among ADHD Traits, Diagnoses, and Hypomania Hypomania age 15 y ADHD symptoms 0.30 (0.24-0.34) 0.02 <.001 Hyperactive-impulsive symptoms 0.53 (0.46-0.60) 0.04 <.001 Inattentive symptoms 0.40 (0.34-0.47) 0.03 <.001 Broad cutoff ADHDb 2.26 (1.90-2.61) 0.18 <.001 Strict cutoff ADHDc 3.68 (2.42-4.95) 0.65 <.001 ADHD diagnosisd 3.18 (2.55-3.81) 0.32 <.001 Hypomania at age 18 y ADHD symptoms 0.19 (0.16-0.22) 0.02 <.001 Hyperactive-impulsive symptoms 0.36 (0.30-0.42) 0.03 <.001 Inattentive symptoms 0.24 (0.19-0.29) 0.03 <.001 Broad cutoff ADHDb 1.57 (1.24-1.90) 0.17 <.001 Strict cutoff ADHDc 2.54 (1.68-3.39) 0.44 <.001 ADHD diagnosisd 2.88 (2.19-3.56) 0.35 <.001 Association Between Hypomanic Symptoms and ADHD ADHD at age 15 y Hypomanic symptoms 1.49 (1.34-1.64) 0.08 <.001 ADHD at age 18 y Hypomanic symptoms 1.98 (1.79-2.17) 0.10 <.001 Association Between ADHD Diagnoses and High Risk of Bipolar Disordere Variable Comparison Group/High-Risk Group, No. (%) Adjusted OR (95% CI)a 1-Sided P Value Hypomania at age 15 y Broad cutoff ADHDa 429 (0.07)/36 (38) 8.27 (5.43-12.58) <.001 Strict cutoff ADHDb 56 (0.009)/12 (13) 15.66 (8.02-30.59) <.001 ADHD diagnosisc 165 (0.03)/28 (29) 15.75 (9.81-25.29) <.001 Hypomania at age 18 y Broad cutoff ADHDa 273 (0.06)/27 (43) 11.95 (7.00-20.40) <.001 Strict cutoff ADHDb 36 (0.008)/7 (11) 15.46 (6.50-36.73) <.001 ADHD diagnosisc 78 (0.02)/20 (32) 28.53 (15.95-51.04) <.001 Abbreviations: ADHD, attention-deficit hyperactivity disorder; BD, bipolar disorder; OR, odds ratio.

5.29) <.001 Hypomania at age 18 y Broad cutoff ADHDa 273 (0.06)/27 (43) 11.95 (7.00-20.40) <.001 Strict cutoff ADHDb 36 (0.008)/7 (11) 15.46 (6.50-36.73) <.001 ADHD diagnosisc 78 (0.02)/20 (32) 28.53 (15.95-51.04) <.001 Abbreviations: ADHD, attention-deficit hyperactivity disorder; BD, bipolar disorder; OR, odds ratio. a Adjusted for sex and age. b Broad ADHD: score of 6 or more on the Autism-Tics, ADHD, and Other Comorbidities Inventory ADHD module at 9 and 12 years of age (n = 465 [7%] at 15 years of age; n = 300 [7%] at 18 years of age). c Strict ADHD: score of 12.5 or more on the Autism-Tics, ADHD, and Other Comorbidities Inventory ADHD module at 9 and 12 years of age (n = 68 [1%] at 15 years of age; n = 43 [1%] at 18 years of age). d ADHD diagnosis: at least 1 recorded diagnosis of ADHD in the National Patient Register (n = 193 [3%] at 15 years of age; n = 98 [2%] at 18 years of age). e These data are given as number (percentage) of comparison group (n = 6301)/high-risk BD group (n = 96). High risk of BD was defined as a cutoff of 10 or more on the parent-rated Child Mania Rating Scale at 15 years of age or a parent-rated Mood Disorders Questionnaire score of at least 7 with symptoms clustered together in the same period and moderate to severe problems (eg, work and legal problems).

k BD group (n = 96). High risk of BD was defined as a cutoff of 10 or more on the parent-rated Child Mania Rating Scale at 15 years of age or a parent-rated Mood Disorders Questionnaire score of at least 7 with symptoms clustered together in the same period and moderate to severe problems (eg, work and legal problems). The continuous measures of ADHD and hypomanic symptoms were used for the remaining analyses. Twin correlations for ADHD and hypomanic traits are given in Table 3; the univariate twin analyses and fit statistics are presented in eTable 3 and eTable 4 in the Supplement. The assumption of equal means and variances across zygosity was not met for ADHD, suggesting sibling contrast pathways (eTable 3 in the Supplement). Moderate to strong heritability was found for ADHD (heritability, 0.51-0.74) and hypomania traits (heritability, 0.50-0.64). Table 3.

, 13%-29%). Hypomania-specific genetic risk factors accounted for 27% to 46% of its variance. Environmental factors played a negligible role in the ADHD-hypomania symptom association. The genetic overlap with hypomania was larger with hyperactivity-impulsivity (range, 10%-25%) compared with inattention (range, 6%-16%). The associations between ADHD and hypomanic symptoms observed here are similar to those reported in other adolescent samples. For example, 1 study found that among 98 adolescent patients with BD, 37.8% presented with ADHD, similar to the comorbidity rate of 37% in our study. Both ADHD dimensions were significantly associated with hypomania in this study, with a stronger association with hyperactivity-impulsivity. Previous research suggests that both ADHD domains are associated with BD to a similar degree or more strongly with inattention using data from adults and outpatients. We used a community sample of youths because the association between hypomania and the ADHD presentations may vary with age and from population to service level and assessment method. This investigation provides a novel contribution by exploring the shared genetic and environmental factors associated with the ADHD-hypomania overlap in youths, examining hyperactivity-impulsivity and inattention separately. The percentage of variance in hypomania that could be associated with shared genetic factors with ADHD symptoms (range, 13%-29%) concurs with genetic correlations reported by a family study on BD II and ADHD (correlation, 0.33) and a molecular genetic study (correlation, 0.26).

y-impulsivity and inattention separately. The percentage of variance in hypomania that could be associated with shared genetic factors with ADHD symptoms (range, 13%-29%) concurs with genetic correlations reported by a family study on BD II and ADHD (correlation, 0.33) and a molecular genetic study (correlation, 0.26). Our other novel finding is that up to 25% of the variance in hypomania was associated with genetic factors related to hyperactivity-impulsivity compared with up to 16% for inattention. This finding is consistent with results showing that the degree to which ADHD symptom domains share genetic risk factors with other psychopathological dimensions in youths varies.

the variance in hypomania was associated with genetic factors related to hyperactivity-impulsivity compared with up to 16% for inattention. This finding is consistent with results showing that the degree to which ADHD symptom domains share genetic risk factors with other psychopathological dimensions in youths varies. Our findings have important clinical and research implications. First, our results provide additional evidence of the ADHD-hypomania symptom overlap in youths, highlighting the need for early identification and recognition among practitioners, especially given the care needed to avoid the negative outcomes associated with ADHD-BD comorbidity (eg, suicide attempts) compared with when these conditions occur alone. Future studies should follow up youths identified as high risk for BD and who exhibit ADHD symptoms to ascertain whether they develop specific forms of psychopathology. Second, significant genetic factors play a role in adolescent hypomania that are distinct from ADHD, suggesting that these phenotypes are not an extension of one another. Our results build on the ADHD and hypomania etiologic models, indicating shared genetic factors associated with these disorders. Uncovering the specific nature of the genetic overlap between these phenotypes should be the focus of further research.

, suggesting that these phenotypes are not an extension of one another. Our results build on the ADHD and hypomania etiologic models, indicating shared genetic factors associated with these disorders. Uncovering the specific nature of the genetic overlap between these phenotypes should be the focus of further research. Strengths and Limitations Strengths of this study include the use of a large, longitudinal, genetically informative sample; use of nationwide registries; and the distinction between the ADHD symptom domains in association with adolescent hypomanic traits; however, there are some limitations. First, hypomania was not measured in childhood; thus, the association between earlier hypomanic symptoms and ADHD traits across childhood and adolescence could not be assessed. Second, hypomanic and ADHD symptoms were measured using different instruments at each age, which may have affected the results. Both hypomania measures are considered among the best-validated and most discriminating adolescent BD instruments. The ADHD measures have good psychometric properties and are widely used. Also, there has been much debate surrounding the overlap between ADHD and hypomania, particularly concerning their symptom similarity. In the current study, significant correlations between hypomania and ADHD symptoms were detected even when overlapping symptoms were removed from the hypomania scales, but various factors may still confound these associations. For instance, shared method variance may have inflated the associations because both phenotypes were measured using questionnaires.

elations between hypomania and ADHD symptoms were detected even when overlapping symptoms were removed from the hypomania scales, but various factors may still confound these associations. For instance, shared method variance may have inflated the associations because both phenotypes were measured using questionnaires. Several associated features of ADHD (eg, irritability and emotional liability) not included in its diagnostic criteria overlap with hypomania. These features were not accounted for here, which may have affected the ADHD-hypomania associations reported and should be considered in future research. Questionnaires are a practical data collection method for large samples needed to undertake twin research. However, they have various limitations, such as the reliance on a restricted number of items that sometimes lack context. Given that ADHD and hypomania were assessed using a restricted number of items, only 2 of 10 items for the CMRS and 5 of 13 for the Mood Disorders Questionnaire overlapped between these phenotypes and were removed to account for symptom similarity. It is possible that some overlapping symptoms were not measured. Whether the presentation of overlapping symptoms is chronic or episodic is crucial to determining whether the symptoms are characteristic of ADHD or hypomania but was not clarified in all instruments used in this study (eg, CMRS). Standardized diagnostic interviews are a more comprehensive approach to establish the specific nature of the symptom presentation and should be adopted by future studies.

l to determining whether the symptoms are characteristic of ADHD or hypomania but was not clarified in all instruments used in this study (eg, CMRS). Standardized diagnostic interviews are a more comprehensive approach to establish the specific nature of the symptom presentation and should be adopted by future studies. The ADHD-BD comorbidity was assessed using official diagnoses in the Swedish National Patient Register, but this method has several shortcomings that can inflate the comorbidity rate, although it was comparable to that in another study. These shortcomings include low agreement with standardized diagnostic interviews and questionable interrater reliability, particularly for BD. When BD is misdiagnosed, it is frequently misdiagnosed as ADHD, which affects the accuracy of the ADHD-BD comorbidity rate detected.

it was comparable to that in another study. These shortcomings include low agreement with standardized diagnostic interviews and questionable interrater reliability, particularly for BD. When BD is misdiagnosed, it is frequently misdiagnosed as ADHD, which affects the accuracy of the ADHD-BD comorbidity rate detected. Conclusions To our knowledge, this was the first study to examine the shared genetic and environmental factors associated with ADHD traits across childhood and adolescence and adolescent hypomania in a representative, longitudinal twin cohort. The collective genetic risk factors for ADHD across childhood and adolescence may also be associated with hypomanic symptoms. The observed associations were stronger between hypomania and hyperactivity-impulsivity compared with inattention, but the association between inattention and hypomania was moderate and significant. This finding suggests that the overlap between ADHD and hypomania traits is likely to reflect a genetic link between these phenotypes. Nevertheless, a substantial amount of the variance for hypomania was associated with genetic risk factors that were not shared with ADHD. Supplement. eTable 1. Instrument Information eTable 2. Correlations Between ADHD Measures and Full and Reduced Hypomania Questionnaires eTable 3. Univariate Assumptions Testing eTable 4. Univariate Model Fit Statistics and Parameter Estimates eTable 5. Squared Standardized Path Coefficients From the Multivariate Cholesky Decomposition of ADHD Traits and Hypomania

Supplement. eTable 1. Instrument Information eTable 2. Correlations Between ADHD Measures and Full and Reduced Hypomania Questionnaires eTable 3. Univariate Assumptions Testing eTable 4. Univariate Model Fit Statistics and Parameter Estimates eTable 5. Squared Standardized Path Coefficients From the Multivariate Cholesky Decomposition of ADHD Traits and Hypomania eTable 6. Squared Standardized Path Coefficients From the Multivariate Cholesky Decomposition of ADHD Traits and Reduced Hypomania Scales, Along With Estimates of Shared and Unique Environmental and Genetic Influences on Hypomania and ADHD eTable 7. Twin Model Fit Statistics eTable 8. Squared Path Coefficients From the Multivariate Cholesky Decomposition of Hyperactivity/Impulsivity and Hypomania eTable 9. Squared Path Coefficients From the Multivariate Cholesky Decomposition of Inattention and Hypomania eTable 10. Squared Standardized Path Coefficients From the Multivariate Cholesky Decomposition of Hyperactivity/Impulsivity and Reduced Hypomania Scales, Along With Estimates of Shared and Unique Environmental and Genetic Influences on Hypomania and Hyperactivity/Impulsivity eTable 11. Squared Standardized Path Coefficients From the Multivariate Cholesky Decomposition of Inattention and Reduced Hypomania Scales, Along With Estimates of Shared and Unique Environmental and Genetic Influences on Hypomania and Inattention eMethods. Analytical Code for Analysis Click here for additional data file.

Introduction Iron deficiency and iron deficiency anemia are common during pregnancy, with an estimated prevalence of 30% to 50% for iron deficiency and 15% to 20% for iron deficiency anemia. Iron demands increase in pregnancy to support the growing fetus and placenta and expand the maternal red blood cell mass. Severe maternal iron shortage can lead to fetal and neonatal iron deficiency. Children with neonatal anemia experience cognitive and behavioral deficits, whereas previous animal studies indicate irreversible neurologic effects of prenatal iron deficiency. Studies of maternal supplemental iron and offspring risk of neurodevelopmental disorders, such as autism spectrum disorder (ASD), have been mixed, with 1 study indicating a protective association of high intakes of supplemental iron (>86 mg/d) compared with lower levels of iron supplementation (<30 mg/d) and 1 study reporting no consistent association between iron supplementation and risk of ASD. Often ASD co-occurs with attention-deficit/hyperactivity disorder (ADHD) and intellectual disability (ID). This comorbid presentation could be the result of shared causes that involve heritable and nonheritable factors occurring during neurodevelopmentally relevant windows. The aim of this study was to examine the association between prenatal anemia diagnoses in mothers and offspring risk of ASD, ADHD, and ID. To assess critical windows of development, we considered the timing of anemia diagnosis.

e heritable and nonheritable factors occurring during neurodevelopmentally relevant windows. The aim of this study was to examine the association between prenatal anemia diagnoses in mothers and offspring risk of ASD, ADHD, and ID. To assess critical windows of development, we considered the timing of anemia diagnosis. Methods Study Population This cohort study used data from the Stockholm Youth Cohort, a prospective, register-based cohort of individuals born from January 1, 1984, to December 31, 2011, residing in Stockholm County at any point from January 1, 2001, to December 31, 2011. Data were derived from registers that contain routinely collected health and sociodemographic data cross-linked via each resident’s national identification number. We included all nonadopted individuals born from January 1, 1987, to December 31, 2010, in Sweden, with a complete record in the Medical Birth Register who were residing in Stockholm County for more than 4 years (Figure 1A). We excluded individuals affected by a study outcome who were also affected by a congenital disorder known to be associated with ID (eg, Down syndrome). Excluded individuals had a lower socioeconomic status compared with included individuals, likely because of the large proportion of migrants in the excluded group (eTable 1 in the Supplement). Data analysis was performed from January 15, 2018, to June 20, 2018. Ethical approval was obtained from the Stockholm Regional Ethical Review Committee, which determined that informed consent was not required for the analysis of anonymized register data.

n of migrants in the excluded group (eTable 1 in the Supplement). Data analysis was performed from January 15, 2018, to June 20, 2018. Ethical approval was obtained from the Stockholm Regional Ethical Review Committee, which determined that informed consent was not required for the analysis of anonymized register data. Figure 1. Selection of the Study Population and Characterization of Outcomes in the Stockholm Youth Cohort Less prevalent outcomes (ie, individuals diagnosed with intellectual disability [ID] and attention-deficit/hyperactivity disorder [ADHD] but not autism spectrum disorder [ASD] and individuals diagnosed with all 3 diagnoses) were not considered individually because of insufficient power in these groups. Numbers in panels B and C are cases per 10 000 in the study cohort. aLess than 1% of mothers were diagnosed with anemia from 1984 to 1986, and we excluded people born before 1987 as a result. bRecords in the medical birth register (MBR) for individuals born in 2011 were not included in the data linkage for this cohort, and thus we excluded these individuals. cDiagnosed with a known congenital disorder or an inborn error of metabolism that has been associated with intellectual disability.

aLess than 1% of mothers were diagnosed with anemia from 1984 to 1986, and we excluded people born before 1987 as a result. bRecords in the medical birth register (MBR) for individuals born in 2011 were not included in the data linkage for this cohort, and thus we excluded these individuals. cDiagnosed with a known congenital disorder or an inborn error of metabolism that has been associated with intellectual disability. Case Ascertainment International Classification of Diseases, Ninth Revision, International Classification of Diseases, 10th Revision, and DSM-IV codes and information from the Prescription Drug Register (for ADHD medication) were used in previously described case ascertainment procedures that covered all inpatient and outpatient pathways to care in Stockholm County (eTable 2 in the Supplement), with follow-up until December 31, 2016. We considered 3 potentially overlapping outcomes: any ASD, any ADHD, and any ID (Figure 1B). We also considered 5 mutually exclusive outcomes: ASD only (no ADHD or ID), ADHD only (no ASD or ID), ID without ASD (no ASD, not excluding ADHD), ASD with ID (not excluding ADHD), and ASD with ADHD (no ID) (Figure 1C).

, 2016. We considered 3 potentially overlapping outcomes: any ASD, any ADHD, and any ID (Figure 1B). We also considered 5 mutually exclusive outcomes: ASD only (no ADHD or ID), ADHD only (no ASD or ID), ID without ASD (no ASD, not excluding ADHD), ASD with ID (not excluding ADHD), and ASD with ADHD (no ID) (Figure 1C). Anemia Maternal anemia was defined as an International Classification of Diseases–coded diagnosis of anemia complicating pregnancy or iron deficiency anemia (eTable 2 in the Supplement) registered up to 1 calendar year before the birth of the index person, recorded in the Medical Birth Register and the National Patient Register. Anemia diagnosis during the periconceptual period was included because it likely indicates exposure to iron deficiency during early gestation. Hemoglobin level is screened a minimum of 3 times throughout pregnancy (at approximately gestational weeks 10, 25, and 37), with additional measurements if indicated. To determine critical windows of exposure, the earliest date of anemia diagnosis was considered relative to the gestational day of pregnancy. Gestational week at diagnosis could not be established for 1286 women (4.2% of women who received an anemia diagnosis). Maternal and child characteristics were similar for those missing a diagnosis date compared with those for whom a date was established (eTable 3 in the Supplement).

o the gestational day of pregnancy. Gestational week at diagnosis could not be established for 1286 women (4.2% of women who received an anemia diagnosis). Maternal and child characteristics were similar for those missing a diagnosis date compared with those for whom a date was established (eTable 3 in the Supplement). Covariates Covariates were chosen based on prior evidence of an association with the outcomes and were evaluated in our cohort for their association with the exposure (Table 1) and outcomes (eFigure 1 in the Supplement). Disposable income at birth was divided into quintiles, using the complete distribution of disposable income in Sweden and accounting for inflation and family size. The highest level of education obtained by either parent was categorized as 9 or less, 10 to 12, or more than 12 years of schooling. Maternal country of origin was dichotomized as born in Sweden or not. Table 1. Selected Characteristics by Maternal Anemia Exposure Category in the Stockholm Youth Cohort (Born 1987-2010) Characteristic No.

Covariates Covariates were chosen based on prior evidence of an association with the outcomes and were evaluated in our cohort for their association with the exposure (Table 1) and outcomes (eFigure 1 in the Supplement). Disposable income at birth was divided into quintiles, using the complete distribution of disposable income in Sweden and accounting for inflation and family size. The highest level of education obtained by either parent was categorized as 9 or less, 10 to 12, or more than 12 years of schooling. Maternal country of origin was dichotomized as born in Sweden or not. Table 1. Selected Characteristics by Maternal Anemia Exposure Category in the Stockholm Youth Cohort (Born 1987-2010) Characteristic No. (%) of Participants Maternal Anemia Diagnosis, wk Yes (n = 31 018) No (n = 501 214) ≤30 (n = 1534) >30 (n = 28 198) Sex Male 16 122 (52.0) 256 762 (51.2) 788 (51.4) 14 626 (51.9) Female 14 896 (48.0) 244 452 (48.8) 746 (48.6) 13 572 (48.1) Maternal BMI Normal (18.5-25) 15 685 (50.6) 252 748 (50.4) 725 (47.3) 14 319 (50.8) Underweight (<18.5) 800 (2.6) 13 399 (2.7) 52 (3.4) 716 (2.5) Overweight (25-30) 5579 (18.0) 71 039 (14.2) 250 (16.3) 5101 (18.1) Obese (>30) 2170 (7.0) 23 675 (4.7) 103 (6.7) 1978 (7.0) Missing 6784 (21.9) 140 353 (28.0) 404 (26.3) 6084 (21.6) Maternal age, y <25 3933 (12.7) 75 180 (15.0) 251 (16.4) 3503 (12.4) 25-29 8245 (26.6) 148 240 (29.6) 403 (26.3) 7485 (26.5) 30-34 11 188 (36.1) 173 187 (34.6) 501 (32.7) 10 262 (36.4) 35-39 6135 (19.8) 86 675 (17.3) 304 (19.8) 5584 (19.8) ≥40 1517 (4.9) 17 932 (3.6) 75 (4.9) 1364 (4.8) Disposable income at IP’s birth First quintile 4212 (13.6) 71 856 (14.3) 315 (20.5) 3697 (13.1) Second quintile 6606 (21.3) 105 514 (21.0) 429 (28.0) 5889 (20.9) Third quintile 6553 (21.1) 109 096 (21.8) 324 (21.1) 5963 (21.2) Fourth quintile 6734 (21.7) 107 993 (21.6) 247 (16.1) 6228 (22.1) Fifth quintile 6913 (22.3) 106 755 (21.3) 219 (14.3) 6421 (22.8) Highest parental education level, y ≤9 1599 (5.2) 29 014 (5.8) 129 (8.4) 1381 (4.9) 10-12 10 999 (35.5) 191 079 (38.1) 643 (41.9) 9903 (35.1) >12 17 760 (57.3) 272 670 (54.4) 727 (47.4) 16 325 (57.9) Missing 660 (2.1) 8451 (1.7) 35 (2.3) 589 (2.1) Maternal psychiatric history before IP’s birth (any diagnosis) Not present 20 544 (66.2) 355 244 (67.8) 908 (59.2) 18 817 (66.7) Present 10 474 (33.8) 161 733 (32.3) 626 (40.8) 9381 (33.3) Single or multiple birth Single 28 699 (92.5) 488 961 (97.6) 1 360 (88.7) 26 237 (93.0) Multiple 2 319 (7.45) 12 253 (2.4) 174 (11.3) 1961 (7.0) Birth order (parity) First child 17 320 (55.8) 224 443 (44.8) 575 (37.5) 15 982 (56.7) Second child 8 934 (28.8) 183 821 (36.7) 540 (35.2) 8074 (28.6) Third or later child 4 764 (15.4) 92 950 (18.5) 419 (27.3) 4142 (14.7) Mother hospitalized for infection during pregnancy No 28 619 (92.3) 483 099 (96.6) 1 336 (87.1) 26 139 (92.7) Yes 2373 (7.7) 17 229 (3.4) 198 (12.9) 2059 (7.3) Mother born outside S

.7) Second child 8 934 (28.8) 183 821 (36.7) 540 (35.2) 8074 (28.6) Third or later child 4 764 (15.4) 92 950 (18.5) 419 (27.3) 4142 (14.7) Mother hospitalized for infection during pregnancy No 28 619 (92.3) 483 099 (96.6) 1 336 (87.1) 26 139 (92.7) Yes 2373 (7.7) 17 229 (3.4) 198 (12.9) 2059 (7.3) Mother born outside S weden No 22 141 (71.34) 377 498 (75.3) 862 (56.2) 20 397 (72.3) Yes 8877 (28.6) 123 716 (24.7) 672 (43.8) 7801 (27.7) Interpregnancy interval, y First born 17 320 (55.8) 224 443 (44.8) 575 (37.5) 15 982 (56.7) <1 1941 (6.23) 42 557(8.5) 157 (10.2) 1709 (6.1) 1-2 3579 (11.5) 81 891 (16.3) 194 (12.6) 3246 (11.5) 2-5 4500 (14.5) 91 657 (18.3) 307 (20.0) 4037 (14.3) 5-10 1871 (6.0) 33 499 (6.7) 141 (9.2) 1652 (5.9) >10 539 (1.7) 8465 (1.7) 40 (2.6) 472 (1.7) Missing 1268 (4.1) 18 702 (3.7) 120 (7.8) 1100 (3.9) Size for gestational age Small for gestational age 684 (2.2) 11 761 (2.4) 92 (6.0) 556 (2.0) Normal 26 320 (84.8) 460 425 (91.9) 1212 (79.0) 24 121 (85.5) Large for gestational age 1 547 (5.0) 14 317 (2.9) 47 (3.1) 1447 (5.1) Missing size for gestational age 148 (0.5) 2458 (0.5) 9 (0.6) 113 (0.4) Missing because of multiple birth 2319 (7.5) 12 253 (2.4) 174 (11.3) 1961 (7.0) Low Apgar score (<7) No 30 228 (98.2) 492 997 (99.1) 1453 (95.7) 27 543 (98.4) Yes 547 (1.8) 4363 (0.9) 65 (4.3) 442 (1.6) Missing 243 (1.8) 3854 (0.8) 16 (1.0) 213 (0.8) Cesarean delivery No 20 585 (66.4) 422 980 (84.4) 956 (62.3) 18 782 (66.6) Yes 10 433 (33.6) 78 225 (15.6) 578 (37.7) 9416 (33.4) Missing 0 9 (<0.1) 0 0 Gestational age at birth Preterm (induced) 1879 (6.1) 11 948 (2.4) 355 (23.1) 1397 (5.0) Preterm (spontaneous) 852 (2.8) 14 898 (3.0) 140 (9.1) 678 (2.4) Term 25 186 (81.3) 437 864 (87.5) 973 (63.4) 23 281 (82.6) Post term 3075 (9.9) 35 618 (7.1) 66 (4.3) 2842 (10.1) Missing 26 (0.1) 886 (0.2) 0 0 Preeclampsia Yes 2 541 (8.2) 15 459 (3.1) 109 (7.1) 2284 (8.1) No 28 477 (91.8) 485 755 (96.9) 1425 (92.9) 25 914 (91.9) Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); IP, index person.

5 (9.9) 35 618 (7.1) 66 (4.3) 2842 (10.1) Missing 26 (0.1) 886 (0.2) 0 0 Preeclampsia Yes 2 541 (8.2) 15 459 (3.1) 109 (7.1) 2284 (8.1) No 28 477 (91.8) 485 755 (96.9) 1425 (92.9) 25 914 (91.9) Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); IP, index person. Maternal body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared from height and weight recorded by midwives at the first antenatal visit. The interpregnancy interval (IPI; ie, the time between the birth of a child and the conception of the next child) was categorized as less than 1, 1 to 2, 2 to 5, 5 to 10, or more than 10 years and being first born (because first-born children by definition have no IPI). Births were categorized as multiple or singleton. Maternal hospitalization for infection during pregnancy (yes or no) and maternal inpatient or outpatient psychiatric history before birth of the child (eTable 2 in the Supplement) were extracted from the National Patient Register and regional psychiatry registers. Statistical Analysis Statistical analyses were performed using Stata/SE, version 13.1 (StataCorp). Odds ratios (ORs) were calculated with generalized estimating equation models with logit link clustered on maternal identification number to account for the clustering of siblings born to the same mother in the data set.

Maternal body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared from height and weight recorded by midwives at the first antenatal visit. The interpregnancy interval (IPI; ie, the time between the birth of a child and the conception of the next child) was categorized as less than 1, 1 to 2, 2 to 5, 5 to 10, or more than 10 years and being first born (because first-born children by definition have no IPI). Births were categorized as multiple or singleton. Maternal hospitalization for infection during pregnancy (yes or no) and maternal inpatient or outpatient psychiatric history before birth of the child (eTable 2 in the Supplement) were extracted from the National Patient Register and regional psychiatry registers. Statistical Analysis Statistical analyses were performed using Stata/SE, version 13.1 (StataCorp). Odds ratios (ORs) were calculated with generalized estimating equation models with logit link clustered on maternal identification number to account for the clustering of siblings born to the same mother in the data set. To examine the association between maternal anemia and offspring risk for ASD, ADHD, and ID, model 1 adjusted only for sex and birth year. Model 2 accounted for sex; birth year; parental educational level and disposable income; maternal country of origin, BMI, age, psychiatric history, and infection during pregnancy; multiple birth; and IPI, as specified above. This analysis was repeated after categorizing the exposure by time of maternal diagnosis of anemia: earlier diagnosis of anemia (≤30 weeks), later diagnosis of anemia (>30 weeks), or no anemia (referent). We selected the earliest cutoff that included adequate cases of ASD, ADHD, and ID to allow for adjusted analyses. To further examine the role of timing of anemia onset, we used restricted cubic spline models with 3 knots to flexibly fit associations between gestational week at anemia diagnosis and odds of each outcome among the 29 732 women with a dated anemia diagnosis, using week 40 as the referent.

allow for adjusted analyses. To further examine the role of timing of anemia onset, we used restricted cubic spline models with 3 knots to flexibly fit associations between gestational week at anemia diagnosis and odds of each outcome among the 29 732 women with a dated anemia diagnosis, using week 40 as the referent. Sensitivity Analysis The prevalence of maternal anemia was substantially lower before 1997 (eFigure 2 in the Supplement), indicating potential underascertainment before 1997. We repeated our main statistical analyses after stratification on birth year (<1997 or ≥1997). Sibling Analysis Associations between maternal anemia and offspring risk of neurodevelopmental disorders may be confounded by unobserved factors, such as shared genetic liability. To evaluate the possibility of such unobserved confounding, we used conditional logistic regression models to compare full siblings exposed to anemia (any, ≤30 weeks, or >30 weeks) with nonexposed siblings in terms of risk of any ASD, any ADHD, and any ID diagnoses, adjusted for factors that are often not shared by siblings: sex, birth year, and IPI.

uch unobserved confounding, we used conditional logistic regression models to compare full siblings exposed to anemia (any, ≤30 weeks, or >30 weeks) with nonexposed siblings in terms of risk of any ASD, any ADHD, and any ID diagnoses, adjusted for factors that are often not shared by siblings: sex, birth year, and IPI. Mediation Analysis Maternal anemia can lead to adverse obstetric outcomes, which in turn may be associated with increased risk of ASD, ADHD, and ID. We examined mediation by the following covariates: size for gestational age (small for gestational age or z score ≤−2, large for gestational age or z score ≥2, or normal), low Apgar score (<7) 5 minutes after birth (yes or no), cesarean delivery (yes or no), and gestational age at birth (<37 weeks, 37-42 weeks, or >42 weeks). Preterm birth at less than 37 weeks was further categorized to indicate whether labor started spontaneously or was medically indicated (ie, the result of induction or cesarean delivery). We performed a mediation analysis using the counterfactual framework method to explore mechanisms by which anemia may be associated with increased risk of ASD, ADHD, and ID, adjusted as in model 2, to estimate direct and indirect relationships (via the mediator) in the Stata module PARAMED. The proportion mediated was calculated as log(natural indirect relationship)/log(total relationship).

od to explore mechanisms by which anemia may be associated with increased risk of ASD, ADHD, and ID, adjusted as in model 2, to estimate direct and indirect relationships (via the mediator) in the Stata module PARAMED. The proportion mediated was calculated as log(natural indirect relationship)/log(total relationship). Because we suspected underascertainment of anemia before 1997 and because nondifferential measurement error of the exposure can bias estimates, we repeated the mediation analysis excluding individuals born before 1997. Since a complex association between multiple births and adverse birth outcomes exists, multiple births were excluded from the mediation analysis. Results Study Sample A total of 532 232 children (272 884 [51.3%] male) between 6 and 29 years of age at the end of follow-up (mean [SD] age, 17.6 [7.1] years) and their 299 768 mothers were included in our study. During 31 018 pregnancies (5.8%), mothers were diagnosed with anemia. Of these diagnoses, 1534 (5.0%) occurred before 30 weeks of pregnancy and 28 198 (90.9%) occurred after 30 weeks of pregnancy (eFigure 3 in the Supplement).

ow-up (mean [SD] age, 17.6 [7.1] years) and their 299 768 mothers were included in our study. During 31 018 pregnancies (5.8%), mothers were diagnosed with anemia. Of these diagnoses, 1534 (5.0%) occurred before 30 weeks of pregnancy and 28 198 (90.9%) occurred after 30 weeks of pregnancy (eFigure 3 in the Supplement). Anemia diagnosis was more common among overweight (OR, 1.16; 95% CI, 1.12-1.19) and obese (OR, 1.28; 95% CI, 1.22-1.34) mothers compared with normal-weight mothers, among mothers older than 40 years compared with mothers younger than 25 years (OR, 1.17; 95% CI, 1.10-1.25), among mothers with a psychiatric history (OR, 1.12; 95% CI, 1.09-1.15), in families in the highest income quintile compared with the lowest (OR, 1.04; 95% CI, 1.00-1.08), in primiparous women (OR, 1.57; 95% CI, 1.52-1.63), and in mothers with an IPI longer than 5 years (OR, 1.15; 95% CI, 1.09-1.21) compared with multiparous women with 2 to 5 years between pregnancies, in multiple births (OR, 3.09; 95% CI, 2.94-3.24), and among mothers who were hospitalized for infection during pregnancy (OR, 2.04; 95% CI, 1.95-2.14) (eTable 4 in the Supplement). In contrast, earlier-onset anemia occurred more often in less educated parents (OR, 0.49; 95%, CI 0.40-0.60, comparing parents with >12 years of education with those with <9 years), in less wealthy families (OR, 0.46; 95% CI, 0.39-0.55, comparing those in the highest income quintile with those in the lowest), among underweight mothers compared with normal-weight mothers (OR, 1.46; 95% CI, 1.09-1.96), and among younger mothers (OR, 0.75; 95% CI, 0.64-0.88, comparing mothers 30-34 years old with those <25 years old).

milies (OR, 0.46; 95% CI, 0.39-0.55, comparing those in the highest income quintile with those in the lowest), among underweight mothers compared with normal-weight mothers (OR, 1.46; 95% CI, 1.09-1.96), and among younger mothers (OR, 0.75; 95% CI, 0.64-0.88, comparing mothers 30-34 years old with those <25 years old). Children born to mothers with anemia diagnosed at 30 weeks or less were more likely to be born preterm (OR, 7.10; 95% CI, 6.28-8.03) or small for gestational age (OR, 2.81; 95% CI, 2.26-3.50) compared with children whose mothers were not diagnosed with anemia, whereas children whose mothers were diagnosed with anemia at greater than 30 weeks’ gestation were more likely to be born post term (OR, 1.56; 95% CI, 1.49-1.62) and large for gestational age (OR, 1.76; 95% CI, 1.66-1.87) (eFigure 4 in the Supplement). Primary Analyses We observed a small increase in risk of ASD (OR, 1.11; 95% CI, 1.04-1.18), ADHD (OR, 1.06; 95% CI, 1.01-1.11), and ID (OR, 1.12; 95% CI, 1.00-1.24) in offspring of mothers diagnosed with anemia in crude models (Table 2). Risks were also elevated in crude models when considering the mutually exclusive diagnostic groups, although the association was not statistically significant for the diagnoses of ID without ASD, ASD with ID, and ASD with ADHD (Table 2). Table 2. Prevalence of and ORs for ASD, ADHD, and ID in Offspring of Mothers Diagnosed With Anemia During Pregnancy, With Consideration of the Timing of Diagnosis Outcome Cases, No.

Primary Analyses We observed a small increase in risk of ASD (OR, 1.11; 95% CI, 1.04-1.18), ADHD (OR, 1.06; 95% CI, 1.01-1.11), and ID (OR, 1.12; 95% CI, 1.00-1.24) in offspring of mothers diagnosed with anemia in crude models (Table 2). Risks were also elevated in crude models when considering the mutually exclusive diagnostic groups, although the association was not statistically significant for the diagnoses of ID without ASD, ASD with ID, and ASD with ADHD (Table 2). Table 2. Prevalence of and ORs for ASD, ADHD, and ID in Offspring of Mothers Diagnosed With Anemia During Pregnancy, With Consideration of the Timing of Diagnosis Outcome Cases, No. (%) [95% CI]a OR (95% CI) No .of Siblings (Discordant Pairs, %)d Model 3, OR (95%CI)e No Maternal Anemia Maternal Anemia Model 1b Model 2c Anemia (All Prenatal Diagnoses) Any ASDf 16 523 (3.30) [3.25-3.35] 1147 (3.70) [3.49-3.91] 1.11 (1.04-1.18) 1.05 (0.99-1.12) 21 949 (4.92/4.34) 1.00 (0.87-1.14) Any ADHDf 34 890 (6.96) [6.89-7.03] 2252 (7.26) [6.97-7.55] 1.06 (1.01-1.11) 1.04 (0.99-1.09) 44 563 (4.64/4.18) 1.01 (0.91-1.11) Any IDf 5904 (1.18) [1.15-1.21] 361 (1.16) [1.05-1.29] 1.12 (1.00-1.24) 1.05 (0.94-1.17) 8 512 (4.18/3.59) 1.10 (0.87-1.38) ASDg 6732 (1.34) [1.31-1.38] 481 (1.56) [1.42-1.69] 1.14 (1.04-1.25) 1.08 (0.98-1.19) NA NA ADHDg 26 958 (5.38) [5.32-5.44] 1681 (5.42) [5.17-5.68] 1.05 (1.00-1.11) 1.04 (0.99-1.10) NA NA ID without ASDg 3083 (0.62) [0.59-0.64] 180 (0.58) [0.50-0.67] 1.14 (0.98-1.32) 1.06 (0.91-1.24) NA NA ASD with IDg 2821 (0.56) [0.54-0.58] 181 (0.58) [0.50-0.67] 1.10 (0.95-1.28) 1.03 (0.88-1.20) NA NA ASD with ADHDg 6970 (1.39) [1.36-1.42] 485 (1.56) [1.43-1.71] 1.08 (0.99-1.19) 1.04 (0.94-1.14) NA NA Anemia Diagnosis ≤30 wkh Any ASDf 16 523 (3.30) [3.25-3.35] 69 (4.49) [3.52-5.66] 1.54 (1.21-1.96) 1.44 (1.13-1.84) 21 835 (0.29/0.06) 2.25 (1.24-4.11) Any ADHDf 34 890 (6.96) [6.89-7.03] 138 (9.00) [7.61–10.54] 1.48 (1.24-1.77) 1.37 (1.14-1.64) 44 352 (0.20/0.19) 1.18 (0.79-1.76) Any IDf 5904 (1.18) [1.15-1.21] 43 (2.80) [2.04-3.76] 2.85 (2.09-3.89) 2.20 (1.61-3.01) 8 479 (0.22/0.13) 2.59 (1.08-6.22) ASDg 6732 (1.34) [1.31-1.38] 26 (1.69) [1.11-2.47] 1.38 (0.94-2.03) 1.35 (0.92-2.00) NA NA ADHDg 26 958 (5.38) [5.32-5.44] 91 (5.93) [4.80-7.23] 1.34 (1.08-1.66) 1.23 (0.99-1.53) NA NA ID without ASDg 3083 (0.62) [0.59-0.64] 27 (1.76) [1.16-2.55] 3.57 (2.41-5.27) 2.72 (1.84-4.01) NA NA ASD with IDg 2821 (0.56) [0.54-0.58] 16 (1.04) [0.60-1.69] 2.23 (1.36-3.64) 1.74 (1.06-2.86) NA NA ASD with ADHDg 6970 (1.39) [1.36-1.42] 27 (1.76) [1.16-2.55] 1.44 (0.99-2.11) 1.38 (0.94-2.03) NA NA Anemia Diagnosis >30 wkh Any ASDf 16 523 (3.30) [3.25-3.35] 1014 (3.60) [3.38-3.82] 1.07 (1.00-1.14) 1.02 (0.95-1.09) 21 835 (4.44/4.15) 0.93 (0.81-1.08) Any ADHDf 34 890 (6.96) [6.89-7.03] 1997 (7.08) [6.79-7.39] 1.03 (0.98-1.08) 1.01 (0.96-1.06) 44 352 (4.19/

(1.76) [1.16-2.55] 1.44 (0.99-2.11) 1.38 (0.94-2.03) NA NA Anemia Diagnosis >30 wkh Any ASDf 16 523 (3.30) [3.25-3.35] 1014 (3.60) [3.38-3.82] 1.07 (1.00-1.14) 1.02 (0.95-1.09) 21 835 (4.44/4.15) 0.93 (0.81-1.08) Any ADHDf 34 890 (6.96) [6.89-7.03] 1997 (7.08) [6.79-7.39] 1.03 (0.98-1.08) 1.01 (0.96-1.06) 44 352 (4.19/ 3.83) 0.97 (0.88-1.08) Any IDf 5904 (1.18) [1.15-1.21] 296 (1.05) [0.93-1.18] 1.01 (0.90-1.14) 0.96 (0.85-1.09) 8 479 (4.44/4.15) 1.01 (0.79-1.29) ASDg 6732 (1.34) [1.31-1.38] 435 (1.54) [1.40-1.69] 1.13 (1.02-1.25) 1.07 (0.97-1.18) NA NA ADHDg 26 958 (5.38) [5.32-5.44] 1478 (5.24) [4.98-5.51] 1.04 (0.98-1.09) 1.03 (0.97-1.09) NA NA ID without ASDg 3083 (0.62) [0.59-0.64] 142 (0.50) [0.42-0.59] 0.99 (0.84-1.18) 0.95 (0.80-1.12) NA NA ASD with IDg 2821 (0.56) [0.54-0.58] 154 (0.55) [0.46-0.64] 1.03 (0.87-1.21) 0.98 (0.83-1.15) NA NA ASD with ADHDg 6970 (1.39) [1.36-1.42] 425 (1.51) [1.37-1.66] 1.04 (0.94-1.15) 0.99 (0.90-1.10) NA NA Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; GEE, generalized estimating equation; ID, intellectual disability; NA, not applicable; OR, odds ratio. a The frequency of each diagnostic outcome, followed by the prevalence (proportion) and 95% confidence interval for the prevalence. b Model 1: GEE model, clustered on maternal identifier, adjusted only for birth year and sex.

3.83) 0.97 (0.88-1.08) Any IDf 5904 (1.18) [1.15-1.21] 296 (1.05) [0.93-1.18] 1.01 (0.90-1.14) 0.96 (0.85-1.09) 8 479 (4.44/4.15) 1.01 (0.79-1.29) ASDg 6732 (1.34) [1.31-1.38] 435 (1.54) [1.40-1.69] 1.13 (1.02-1.25) 1.07 (0.97-1.18) NA NA ADHDg 26 958 (5.38) [5.32-5.44] 1478 (5.24) [4.98-5.51] 1.04 (0.98-1.09) 1.03 (0.97-1.09) NA NA ID without ASDg 3083 (0.62) [0.59-0.64] 142 (0.50) [0.42-0.59] 0.99 (0.84-1.18) 0.95 (0.80-1.12) NA NA ASD with IDg 2821 (0.56) [0.54-0.58] 154 (0.55) [0.46-0.64] 1.03 (0.87-1.21) 0.98 (0.83-1.15) NA NA ASD with ADHDg 6970 (1.39) [1.36-1.42] 425 (1.51) [1.37-1.66] 1.04 (0.94-1.15) 0.99 (0.90-1.10) NA NA Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; GEE, generalized estimating equation; ID, intellectual disability; NA, not applicable; OR, odds ratio. a The frequency of each diagnostic outcome, followed by the prevalence (proportion) and 95% confidence interval for the prevalence. b Model 1: GEE model, clustered on maternal identifier, adjusted only for birth year and sex. c Model 2: GEE model, clustered on maternal identifier, adjusted for: birth year, sex, educational level, disposable income, mother born outside Sweden, body mass index, maternal age, maternal psychiatric history, multiple birth, interpregnancy interval, and maternal infection during pregnancy.

b Model 1: GEE model, clustered on maternal identifier, adjusted only for birth year and sex. c Model 2: GEE model, clustered on maternal identifier, adjusted for: birth year, sex, educational level, disposable income, mother born outside Sweden, body mass index, maternal age, maternal psychiatric history, multiple birth, interpregnancy interval, and maternal infection during pregnancy. d Proportion of discordant sibling pairs: the proportion of sibling pairs with the affected sibling exposed and the unaffected sibling unexposed is reported followed by the proportion of sibling pairs with the unaffected sibling exposed and the affected sibling unexposed. e Model 3: conditional logistic regression model. f Potentially overlapping diagnostic groups. Individuals included in one diagnostic group are potentially included in another diagnostic group (Figure 1B). g Mutually exclusive diagnostic groups. Individuals are included in only a single diagnostic group (Figure 1C). h Pregnancies affected by anemia but with unknown gestational week of diagnosis were excluded from this analysis.

f Potentially overlapping diagnostic groups. Individuals included in one diagnostic group are potentially included in another diagnostic group (Figure 1B). g Mutually exclusive diagnostic groups. Individuals are included in only a single diagnostic group (Figure 1C). h Pregnancies affected by anemia but with unknown gestational week of diagnosis were excluded from this analysis. After stratification of the exposure by the timing of anemia diagnosis, risk of ADHD (OR, 1.48; 95% CI, 1.24-1.77) and ID (OR, 2.85; 95% CI, 2.09-3.89) was increased among offspring of mothers with an early (≤30 weeks) diagnosis of anemia in crude models (Table 2). In contrast, risk of ADHD (OR, 1.03; 95% CI, 0.98-1.08) and ID (OR, 1.01; 95% CI, 0.90-1.14) were not increased among offspring of mothers with a later (>30 weeks) diagnosis of anemia (Table 2). Risk of offspring ASD was associated with both earlier (OR, 1.54; 95% CI, 1.21-1.96) and later (OR, 1.07; 95% CI, 1.00-1.14) anemia diagnoses, although the association with earlier anemia was more pronounced (Table 2). When the mutually exclusive diagnostic groups were considered, an early maternal diagnosis of anemia was associated with increased risk of ADHD (without comorbidities) (OR, 1.34; 95% CI, 1.08-1.66), ID without ASD (OR, 3.57; 95% CI, 2.41-5.27), and ASD with ID (OR, 2.23; 95% CI, 1.36-3.64) (Table 2). Risk for ASD (without comorbidities) was increased among offspring to mothers with later diagnosed anemia (OR, 1.13; 95% CI, 1.02-1.25). In continuous analysis, a pattern of decreasing risk with later gestational week at anemia diagnosis was observed for all outcomes except ASD with ADHD (Figure 2).

36-3.64) (Table 2). Risk for ASD (without comorbidities) was increased among offspring to mothers with later diagnosed anemia (OR, 1.13; 95% CI, 1.02-1.25). In continuous analysis, a pattern of decreasing risk with later gestational week at anemia diagnosis was observed for all outcomes except ASD with ADHD (Figure 2). Figure 2. Association Between Gestational Week of Maternal Anemia Diagnosis and Offspring Odds of Neurodevelopmental Outcomes Among the 29 732 Women With a Dated Anemia Diagnosis The odds of each outcome according to gestational week at anemia diagnosis were flexibly fit using a restricted cubic spline model with 3 knots and gestational week 40 set as the referent. The solid line represents the odds ratio (OR) estimated from the fully adjusted generalized estimating equation model, clustered on maternal identifier, and adjusted for birth year, sex, educational level, disposable income, mother born outside Sweden, body mass index, maternal age, maternal psychiatric history, multiple birth, interpregnancy interval, and maternal infection during pregnancy. The dotted lines represent the 95% CI for the fully adjusted model. Results are shown for the potentially overlapping diagnostic outcomes (Figure 1B) in panels A to C and for the mutually exclusive diagnostic categories (Figure 1C) in panels D to H.

rth, interpregnancy interval, and maternal infection during pregnancy. The dotted lines represent the 95% CI for the fully adjusted model. Results are shown for the potentially overlapping diagnostic outcomes (Figure 1B) in panels A to C and for the mutually exclusive diagnostic categories (Figure 1C) in panels D to H. In adjusted models, adding socioeconomic and maternal- and pregnancy-related factors in combination attenuated the associations between maternal anemia and the outcomes (Table 2). After adjustment, we observed the strongest association between early (≤30 weeks) diagnosis of anemia and ID without co-occurring ASD (OR, 2.72; 95% CI, 1.84-4.01). Although the distribution of the covariates varied among the exposed and unexposed groups (Table 1) and all were associated with risk of neurodevelopmental disorders (eFigure 1), the modulating effect of any individual covariate was not extensive (eFigure 5 in the Supplement). Sensitivity Analysis After stratification by birth year, the associations remained largely similar (eTables 5-7 in the Supplement), with ORs that were often lower among those born before 1997. An exception to this was ASD with ADHD, which was associated with maternal anemia only among those born before 1997 (OR, 1.28; 95% CI, 1.04-1.57) (eTables 5-7 in the Supplement).

year, the associations remained largely similar (eTables 5-7 in the Supplement), with ORs that were often lower among those born before 1997. An exception to this was ASD with ADHD, which was associated with maternal anemia only among those born before 1997 (OR, 1.28; 95% CI, 1.04-1.57) (eTables 5-7 in the Supplement). Sibling Analysis Although there were a limited number of siblings exposed to maternal anemia diagnosed early in pregnancy, early anemia diagnosis was associated with increased risk for ASD and ID in the sibling analysis, with a higher OR for ASD (OR, 2.25; 95% CI, 1.24-4.11) compared with the primary analysis and a similar OR for ID (OR, 2.59; 95% CI, 1.08-6.22) (Table 2). The risk estimate for ADHD was similar to the primary analysis but with a wide CI (OR, 1.18; 95% CI, 0.79-1.76).

creased risk for ASD and ID in the sibling analysis, with a higher OR for ASD (OR, 2.25; 95% CI, 1.24-4.11) compared with the primary analysis and a similar OR for ID (OR, 2.59; 95% CI, 1.08-6.22) (Table 2). The risk estimate for ADHD was similar to the primary analysis but with a wide CI (OR, 1.18; 95% CI, 0.79-1.76). Mediation Analysis We evaluated as potential mediators obstetric complications that were associated with the outcomes of ASD, ADHD, or ID and with anemia diagnosed earlier in pregnancy (Table 1 and eFigure 4 and eFigure 6 in the Supplement) because only anemia diagnosed at 30 weeks or less was consistently associated with the outcomes. Adverse obstetric outcomes accounted for a modest proportion (2.2%-43.1%) of the association between maternal anemia and risk of ASD, ADHD, and ID, although the natural indirect relationship estimates were not statistically significant in most cases (Table 3). Excluding individuals born before 1997 yielded similar results. Preterm birth was the strongest mediator for all outcomes, particularly induced preterm birth, which accounted for approximately one-third of the association between anemia diagnosed at 30 weeks or less and risk of ASD (proportion mediated, 28.21%), ADHD (proportion mediated, 32.93%), and ID (proportion mediated, 32.03%). Table 3.

Mediation Analysis We evaluated as potential mediators obstetric complications that were associated with the outcomes of ASD, ADHD, or ID and with anemia diagnosed earlier in pregnancy (Table 1 and eFigure 4 and eFigure 6 in the Supplement) because only anemia diagnosed at 30 weeks or less was consistently associated with the outcomes. Adverse obstetric outcomes accounted for a modest proportion (2.2%-43.1%) of the association between maternal anemia and risk of ASD, ADHD, and ID, although the natural indirect relationship estimates were not statistically significant in most cases (Table 3). Excluding individuals born before 1997 yielded similar results. Preterm birth was the strongest mediator for all outcomes, particularly induced preterm birth, which accounted for approximately one-third of the association between anemia diagnosed at 30 weeks or less and risk of ASD (proportion mediated, 28.21%), ADHD (proportion mediated, 32.93%), and ID (proportion mediated, 32.03%). Table 3. Mediation Analysis With Adverse Obstetric Outcomes as Potential Mediators Between Earlier Maternal Anemia (≤30 Weeks) and Offspring Risk of ASD, ADHD, and ID Variable Offspring, Odds Ratio (95% CI)a Any ASD Any ADHD Any ID Early Anemia Diagnosis (≤30 wk), 1987-2010 Cesarean delivery Natural direct relationship 1.35 (1.00-1.80) 1.28 (1.02-1.55) 1.95 (1.29-2.77) Natural indirect relationship 1.03 (0.95-1.14) 1.04 (0.98-1.12) 1.02 (0.91-1.16) Total relationship 1.39 (1.06-1.82) 1.33 (1.07-1.58) 1.98 (1.32-2.77) Proportion, %b 9.52 12.78 2.21 Low Apgar score Natural direct relationship 1.39 (1.05-1.82) 1.33 (1.07-1.60) 1.81 (1.22-2.55) Natural indirect relationship 1.00 (0.98-1.04) 1.00 (0.98-1.04) 1.11 (1.00-1.31) Total relationship 1.39 (1.05-1.81) 1.34 (1.08-1.61) 2.00 (1.31-2.82) Proportion, %b NRc NR c 14.81 Small for gestational age Natural direct relationship 1.45 (1.10-1.90) 1.35 (1.09-1.62) 1.92 (1.26-2.73) Natural indirect relationship 1.00 (0.97-1.05) 1.01 (0.98-1.04) 1.07 (1.00-1.21) Total relationship 1.45 (1.10-1.90) 1.36 (1.08-1.61) 2.05 (1.35-2.89) Proportion, %b NRc 1.90 9.27 Preterm (induced) Natural direct relationship 1.15 (0.75-1.53) 1.18 (0.93-1.48) 1.75 (1.08-2.58) Natural indirect relationship 1.11 (0.77-1.35) 1.08 (0.99-1.21) 1.09 (0.93-1.43) Total relationship 1.29 (0.94-1.67) 1.27 (1.03-1.56) 1.90 (1.22-2.75) Proportion, %b 43.10 32.79 13.05 Preterm (spontaneous) Natural direct relationship 1.17 (0.80-1.55) 1.18 (0.93-1.47) 1.78 (1.08-2.57) Natural indirect relationship 1.06 (0.98-1.21) 1.02 (0.97-1.11) 1.07 (0.97-1.27) Total relationship 1.24 (0.88-1.65) 1.20 (0.95-1.50) 1.90 (1.17-2.75) Proportion, %b 28.13 11.10 10.21 Preeclampsia Natural direct relationship 1.30 (0.95 -1.64) 1.27 (1.06-1.54) 1.86 (1.27-2.62) Natural indirect relationship 1.04 (0.99-1.12) 1.01 (0.98-1.05) 1.07 (0.99-1.23) Total relationship 1.35 (1.02-1.72) 1.27 (1.06-1.52) 1.99 (1.32-2.80) Proportion, %b 13.1 2.45 9.39 Early Anemia Diagnosis (≤30 wk), 1997-2010 Cesarean delivery Natural direct relationship 1.55 (1.06-2.09) 1.32 (0.99-1.66) 2.05 (1.15-3.01) Natural indirect relationship 1.08 (0.96-1.28) 1.09 (0.98-1.21) 1.07 (0.92-1.32) Total relationship 1.68 (1.22-2.2

) 1.27 (1.06-1.52) 1.99 (1.32-2.80) Proportion, %b 13.1 2.45 9.39 Early Anemia Diagnosis (≤30 wk), 1997-2010 Cesarean delivery Natural direct relationship 1.55 (1.06-2.09) 1.32 (0.99-1.66) 2.05 (1.15-3.01) Natural indirect relationship 1.08 (0.96-1.28) 1.09 (0.98-1.21) 1.07 (0.92-1.32) Total relationship 1.68 (1.22-2.2 2) 1.44 (1.09-1.77) 2.21 (1.31-3.19) Proportion, %b 15.56 22.86 9.11 Low Apgar score Natural direct relationship 1.64 (1.14-2.18) 1.43 (1.07-1.77) 1.89 (1.07-2.83) Natural indirect relationship 1.00 (0.97-1.04) 1.01 (0.98-1.07) 1.19 (1.02-1.53) Total relationship 1.63 (1.17-2.15) 1.44 (1.07-1.76) 2.24 (1.29-3.23) Proportion, %b NRc 3.43 21.20 Small for gestational age Natural direct relationship 1.70 (1.23-2.28) 1.39 (1.05-1.72) 2.04 (1.18-3.01) Natural indirect relationship 0.99 (0.97-1.04) 1.02 (0.98-1.08) 1.08 (1.00-1.27) Total relationship 1.69 (1.22-2.23) 1.42 (1.06-1.74) 2.21 (1.31-3.18) Proportion, %b NRc 5.55 10.12 Preterm (induced) Natural direct relationship 1.43 (0.95-2.02) 1.27 (0.92-1.65) 1.65 (0.80-2.70) Natural indirect relationship 1.15 (0.98-1.45) 1.13 (0.99-1.34) 1.27 (0.96-1.87) Total relationship 1.64 (1.17-2.24) 1.43 (1.09-1.79) 2.10 (1.22-3.23) Proportion, %b 28.21 32.93 32.03 Preterm (spontaneous) Natural direct relationship 1.45 (0.98-2.05) 1.28 (0.93-1.66) 1.74 (0.90-2.81) Natural indirect relationship 1.07 (0.98-1.26) 1.05 (0.98-1.17) 1.20 (1.00-1.70) Total relationship 1.55 (1.07-2.16) 1.34 (0.98-1.73) 2.09 (1.19-3.39) Proportion, %b 15.90 15.22 24.53 Preeclampsia Natural direct relationship 1.63 (1.16-2.17) 1.46 (1.12-1.83) 2.05 (1.28.3.09) Natural indirect relationship 1.03 (0.97-1.12) 1.01 (0.97-1.08) 1.12 (0.99-1.36) Total relationship 1.68 (1.21-2.22) 1.47 (1.13-1.81) 2.39 (1.48-3.43) Proportion, %b 6.30 2.00 13.68 Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; ID, intellectual disability; NR, not reported.

tural indirect relationship 1.03 (0.97-1.12) 1.01 (0.97-1.08) 1.12 (0.99-1.36) Total relationship 1.68 (1.21-2.22) 1.47 (1.13-1.81) 2.39 (1.48-3.43) Proportion, %b 6.30 2.00 13.68 Abbreviations: ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; ID, intellectual disability; NR, not reported. a Adjusted for sex, birth year, educational level, disposable income, mother born outside Sweden, body mass index, maternal age, maternal psychiatric history, interpregnancy interval, and maternal infection during pregnancy. b Proportion mediated was calculated as log(natural indirect relationship)/log(total relationship). c The natural direct and natural indirect relationship were not in the same direction; therefore, the proportion mediated is not a logical value. Discussion Although less prevalent, earlier diagnosed anemia was associated with a greater risk of ASD, ADHD, and ID compared with no diagnosis of anemia, even in models accounting for potentially confounding socioeconomic, maternal, and pregnancy-related factors. Anemia that occurred earlier in pregnancy was most strongly associated with the outcome of ID. The associations between anemia earlier in pregnancy and risk of ASD and ID were also apparent in matched sibling analyses. Obstetric complications previously associated with maternal anemia mediated a modest proportion of the risk of the neurodevelopmental disorders associated with earlier maternal anemia diagnosis.

. The associations between anemia earlier in pregnancy and risk of ASD and ID were also apparent in matched sibling analyses. Obstetric complications previously associated with maternal anemia mediated a modest proportion of the risk of the neurodevelopmental disorders associated with earlier maternal anemia diagnosis. Agreement With Other Studies Long-term cognitive and behavioral effects of maternal anemia have not been well studied in humans, especially regarding ASD and ADHD. Leonard et al found a 5-fold increased risk of severe ID in offspring of anemic mothers, although no increased risk for mild to moderate ID or ID with ASD was observed. In a secondary analysis of the Collaborative Perinatal Project, a dose-response relationship between maternal hematocrit during pregnancy and children’s IQ at 4 and 7 years of age was found. Women with moderate anemia had a 59% greater chance of having a child with an IQ below 70 at 7 years of age. To our knowledge, no other studies have been performed regarding maternal anemia during pregnancy and offspring risk of ASD, ADHD, and ID. A few small studies have found an increased risk of ASD associated with anemia in the infant, which can be the result of maternal anemia.

hild with an IQ below 70 at 7 years of age. To our knowledge, no other studies have been performed regarding maternal anemia during pregnancy and offspring risk of ASD, ADHD, and ID. A few small studies have found an increased risk of ASD associated with anemia in the infant, which can be the result of maternal anemia. Comparing the results of previous supplementation studies with our results warrants caution. Anemia status cannot be extrapolated from the supplementation studies, and we have not studied the effects of supplemental iron intake. Our results would support a potentially protective role of iron supplementation in pregnant women with regard to offspring risk of neurodevelopmental disorders because iron supplementation can prevent iron deficiency anemia. Regardless of neurodevelopmental outcomes, iron supplementation in pregnant women is associated with reduced risk of low birth weight and preterm birth. However, excessive iron intake can be toxic.

regard to offspring risk of neurodevelopmental disorders because iron supplementation can prevent iron deficiency anemia. Regardless of neurodevelopmental outcomes, iron supplementation in pregnant women is associated with reduced risk of low birth weight and preterm birth. However, excessive iron intake can be toxic. Interpretation and Potential Mechanisms There is evidence that fetal needs for iron are prioritized compared with the mother’s. The fetus is not necessarily exposed to iron deficiency if the mother is affected by anemia but may be affected only when a threshold is crossed and the maternal shortage is more severe and long lasting. Although later anemia diagnoses may result from the higher iron demands of larger fetuses, early nutritional deficiencies represent a distinct phenomenon, leading to growth restriction and increased risk of being small for gestational age. Earlier anemia diagnoses were associated with infants being born small for gestational age, and later anemia diagnoses were associated with infants being born large for gestational age. The highest rates of fetal iron uptake occur from 30 weeks onward, with the neonate’s total iron endowment directly proportional to its birth weight. An alternative explanation could be that earlier diagnosed anemia occurs during a more critical time window with regard to neurodevelopmental disorders.

gestational age. The highest rates of fetal iron uptake occur from 30 weeks onward, with the neonate’s total iron endowment directly proportional to its birth weight. An alternative explanation could be that earlier diagnosed anemia occurs during a more critical time window with regard to neurodevelopmental disorders. The associations reported here could be the result of iron deficiency in the developing brain. Iron is necessary for a number of developmental processes, such as myelination and dendrite arborization, and for the synthesis of monoamine neurotransmitters, which are implicated in the etiology of ASD and ADHD. Because erythrocyte hemoglobin is essential for the transport of oxygen, oxygen supply to the developing fetus might be limited in anemic mothers and may be associated with an increased risk of hypoxia. Alternatively, adverse obstetric outcomes caused by maternal anemia might mediate the association. Mediation analysis indicated that some adverse obstetric outcomes, particularly being born preterm, could explain a portion of the associations reported here.

thers and may be associated with an increased risk of hypoxia. Alternatively, adverse obstetric outcomes caused by maternal anemia might mediate the association. Mediation analysis indicated that some adverse obstetric outcomes, particularly being born preterm, could explain a portion of the associations reported here. Strengths and Limitations Strengths of our study include data that were prospectively collected in high-quality population-based registers from a setting with universal access to comprehensive health care, reducing bias in ascertainment of the exposure and outcomes. However, generalizability may be limited by the inclusion of only Swedish-born individuals. Another strength of this study is the consideration of multiple commonly co-occurring neurodevelopmental disorders as outcomes. For some diagnostic groups, we had a limited number of cases, especially when considering earlier anemia and when performing the mediation and sibling analysis. Also, this study integrated evidence from the sibling comparison with evidence from the multivariable regression models to address the issue of residual confounding. The sibling analysis accounted for factors that are largely shared by siblings, such as socioeconomic status and genetic background.

and sibling analysis. Also, this study integrated evidence from the sibling comparison with evidence from the multivariable regression models to address the issue of residual confounding. The sibling analysis accounted for factors that are largely shared by siblings, such as socioeconomic status and genetic background. This study has limitations. Although we evaluated multiple covariates and include a sibling comparison, a possibility exists for residual confounding, for instance by other dietary deficiencies associated with anemia. Both obesity and being underweight may indicate a suboptimal diet. Adjusting for maternal BMI did not modulate the association between maternal anemia and offspring risk of neurodevelopmental disorders. Diet quality is associated with socioeconomic factors, which were also investigated as potential confounders. We were limited in our ascertainment of infections during pregnancy. Most infections do not require hospitalization, and we were not able to specifically examine infections known to be detrimental to the developing nervous system (eg, TORCH [Toxoplasma gondii, other, rubella virus, cytomegalovirus, and herpes simplex virus]). In addition, anemia caused by iron deficiency could not be disentangled from other causes. Given that most anemia is caused by iron deficiency, iron deficiency during pregnancy is the most plausible explanation of our results, although anemia regardless of its cause may affect neurodevelopment. Timing of anemia was determined by the first date of anemia diagnosis during or shortly before pregnancy. The true onset of micronutrient deficiencies and subsequent duration of anemia cannot be identified. Treatment of anemia caused by micronutrient deficiencies consists of supplying the necessary micronutrient, often via oral supplements but sometimes intravenously. Oral iron therapy can improve hemoglobin levels, although such therapy may not fully replete iron stores in pregnant women. Timing and effectiveness of treatment were unknown in our study.

cronutrient deficiencies consists of supplying the necessary micronutrient, often via oral supplements but sometimes intravenously. Oral iron therapy can improve hemoglobin levels, although such therapy may not fully replete iron stores in pregnant women. Timing and effectiveness of treatment were unknown in our study. Conclusions In this study, anemia diagnosed at 30 weeks or less of pregnancy was associated with modestly increased offspring risk of ASD and ADHD and greater risk of ID, suggesting that exposure to anemia earlier in gestation may be negatively associated with neurodevelopment in the child. Given that iron deficiency and anemia are common among women of childbearing age, our findings appear to emphasize the importance of early screening for iron status and nutritional counseling in antenatal care. Supplement. eTable 1. A Comparison of Individuals Included in the Final Study Sample to Those Excluded From the Study Sample eTable 2. Diagnostic Codes and Register Databases Used to Ascertain Diagnoses in the Stockholm Youth Cohort (SYC) eTable 3. A Comparison of Maternal and Child Characteristics for 1286 Women for Whom Gestational Age at Anemia Diagnosis Could Be Determined to 29732 Women for Whom Gestational Age Anemia Diagnosis Could Be Determined eTable 4. Prevalence and Odds Ratios (+95% Confidence Intervals) of Maternal Anemia Diagnosed During Pregnancy (in General and Diagnosed ≤30 Weeks or Diagnosed >30 Weeks) by Selected Characteristics

eTable 3. A Comparison of Maternal and Child Characteristics for 1286 Women for Whom Gestational Age at Anemia Diagnosis Could Be Determined to 29732 Women for Whom Gestational Age Anemia Diagnosis Could Be Determined eTable 4. Prevalence and Odds Ratios (+95% Confidence Intervals) of Maternal Anemia Diagnosed During Pregnancy (in General and Diagnosed ≤30 Weeks or Diagnosed >30 Weeks) by Selected Characteristics eTable 5. Prevalence of Selected Characteristics and Pregnancy Outcomes of the Child by Anemia in a Cohort of Non-Adoptive Births in Sweden Between 1997 & 2010 eTable 6. Prevalence of Selected Characteristics and Pregnancy Outcomes of the Child by Anemia in a Cohort of Non-Adoptive Births in Sweden Between 1987 & 1996 eTable 7. Odds Ratios for ASD, ADHD, and ID in Offspring of Mothers Diagnosed With Anemia During Pregnancy After Stratification on Birth Years Before and After 1997 eFigure 1. The Association Between Potentially Confounding Factors and Risk of ASD, ADHD, or ID eFigure 2. Prevalence of Maternal Anemia per Birth Year eFigure 3. Prevalence of Maternal Anemia per Gestational Week eFigure 4. Risk for Pregnancy Outcomes in Relation to Diagnosis of Maternal Anemia, Comparing Mothers Diagnosed With Anemia During Pregnancy (at Any Point, Diagnosed ≤30 Weeks, or Diagnosed >30 Weeks) to Mothers Not Diagnosed With Anemia eFigure 5. An Exploration of the Influence of Different Potentially Confounding Factors of the Risk for Diagnostic Outcomes Related to Any Anemia Diagnosis, Anemia Diagnosed ≤30 Weeks and Anemia Diagnosed >30 Weeks

eFigure 4. Risk for Pregnancy Outcomes in Relation to Diagnosis of Maternal Anemia, Comparing Mothers Diagnosed With Anemia During Pregnancy (at Any Point, Diagnosed ≤30 Weeks, or Diagnosed >30 Weeks) to Mothers Not Diagnosed With Anemia eFigure 5. An Exploration of the Influence of Different Potentially Confounding Factors of the Risk for Diagnostic Outcomes Related to Any Anemia Diagnosis, Anemia Diagnosed ≤30 Weeks and Anemia Diagnosed >30 Weeks eFigure 6. Odds Ratios (+95% Confidence Intervals) for Neurodevelopmental Disorders (ASD, ADHD, and ID) in Relation to Potential Mediators Click here for additional data file.

ears who exhibit MPFC-DLPFC anticorrelations may have the capacity to toggle between internal and external foci of attention more readily than those who do not. The failure to decouple these networks may be an early indicator of attentional problems or may preclude the development of age-appropriate attentional skills. That stronger sgACC-left DLPFC anticorrelations predicted a future worsening of internalization, characteristic of MDD, is consistent with the MDD and at-risk literature. One study found a reduction of left DLPFC-sgACC rs-fMRI connectivity in children at familial risk for MDD, for which the at-risk group had significant anticorrelations while the not-at-risk group had positive correlations. Furthermore, left DLPFC-sgACC anticorrelations have been used to identify individually specific targets for TMS in patients with MDD. Stronger sgACC-left DLPFC anticorrelations at this young age may already reflect an attenuation or failure of top-down control mechanisms that are evident in adult MDD. Thus, the functional connectivity of specific neural systems in middle childhood forecasts individuals’ vulnerability or resilience in cognition and emotion over the ensuing 4 years of development.

Introduction The regulation of cognition and emotion is thought to depend on the top-down modulation of multiple neural circuits by the prefrontal cortex and, in particular, the dorsolateral prefrontal cortex (DLPFC). Prefrontal-dependent cognitive control mechanisms that regulate attention and mood likely play a key role in mental health. There is ample evidence of attenuation or failure of top-down control mechanisms in adults with depression, anxiety, and attention-deficit/hyperactivity disorder (ADHD). Given that these prevalent mental health problems often emerge during childhood and adolescence, it is important to know whether dysregulated top-down control can be detected even before behavioral symptoms are evident. The strength of coupling between regions involved in top-down control and their targets can be measured with resting-state functional magnetic resonance imaging (rs-fMRI). Brain regions that are highly temporally correlated during rest-form resting-state networks (RSNs), which are intrinsic, spontaneous, low-frequency fluctuations in the fMRI blood oxygen level–dependent signal that define specific networks of the brain in the absence of any task. They reveal great heterogeneity in the functional organization of the brain. In fact, they may be considered “fingerprints” of the human brain, as they can accurately identify an individual from a large group (N = 126). Furthermore, RSN profiles are known to be robust and reliable.

the brain in the absence of any task. They reveal great heterogeneity in the functional organization of the brain. In fact, they may be considered “fingerprints” of the human brain, as they can accurately identify an individual from a large group (N = 126). Furthermore, RSN profiles are known to be robust and reliable. Resting-state networks are particularly relevant for studying psychiatric and pediatric populations because they are (1) task-independent, so individual differences in task performance cannot explain differences observed in the blood oxygen level–dependent data, (2) easy and fast to acquire, which make them accessible to many people, including young children and various clinical populations, and (3) and plastic, been shown to change during typical development, and can be modulated by behavioral or pharmacological interventions. An RSN that is particularly relevant for mental health is the Central Executive Network (CEN), of which the DLPFC is a key node. The CEN has been associated with externally focused attention and goal-directed behavior. In neurotypical adults, the CEN is negatively correlated (ie, anticorrelated) with the default mode network (DMN), an RSN associated with internal mentation and self-referential processing, whose key nodes include the medial prefrontal cortex (MPFC).

ociated with externally focused attention and goal-directed behavior. In neurotypical adults, the CEN is negatively correlated (ie, anticorrelated) with the default mode network (DMN), an RSN associated with internal mentation and self-referential processing, whose key nodes include the medial prefrontal cortex (MPFC). The decoupling of these RSNs has been found to be adaptive: stronger MPFC-DLPFC anticorrelations are associated with superior cognitive control and cognitive performance in adults, such as greater working memory capacity. In addition, there is an increase with age in the magnitude of anticorrelations between the MPFC and DLPFC in typically developing children, which is consistent with the findings that top-down control mechanisms improve markedly over childhood and adolescence. Resting-state fMRI studies have also shown an association between diminished MPFC-DLPFC anticorrelations and cognitive impairment in ADHD.

between the MPFC and DLPFC in typically developing children, which is consistent with the findings that top-down control mechanisms improve markedly over childhood and adolescence. Resting-state fMRI studies have also shown an association between diminished MPFC-DLPFC anticorrelations and cognitive impairment in ADHD. The CEN also plays a role in regulating mood through its interactions with the subgenual anterior cingulate cortex (sgACC). The sgACC is part of the affective network, which is involved in emotion processing and has anatomical connections to the hypothalamus, amygdala, entorhinal cortex, nucleus accumbens, and other limbic structures. There are several lines of evidence showing that top-down modulation of the sgACC is dysregulated in adults with major depressive disorder (MDD). Neuroimaging studies have reported decreased metabolisms and decreased gray matter volumes and a decreased number of glia in sgACC in patients with MDD. Furthermore, deep brain stimulation of the sgACC results in an attenuation of hyperactivation in sgACC and increased activation in previously underactive DLPFC in adults with MDD. In addition, the left DLPFC region that shows maximal anticorrelation with the sgACC in rs-fMRI has been identified as an optimal target for transcranial magnetic stimulation (TMS) of MDD. The sgACC has also been shown to exhibit decreased connectivity with cognitive control regions in children with a history of preschool depression. Finally, left DLPFC and sgACC exhibit anticorrelation in children at familial risk for MDD.

entified as an optimal target for transcranial magnetic stimulation (TMS) of MDD. The sgACC has also been shown to exhibit decreased connectivity with cognitive control regions in children with a history of preschool depression. Finally, left DLPFC and sgACC exhibit anticorrelation in children at familial risk for MDD. In sum, prior research on patient and familial high-risk populations reveals that atypically strong functional connectivity between DLPFC and MPFC is characteristic of ADHD, whereas atypically weak connectivity between DLPFC and the sgACC is a characteristic of MDD. Here, we build on this prior work by asking whether the strength of the connectivity between these regions can predict a progression toward attentional or mood disorders in a longitudinal study of a community pediatric sample not selected for risk of ADHD or MDD. Specifically, we tested whether DLPFC-MPFC and DLPFC-sgACC connectivity at age 7 years predict scores at age 11 years on a questionnaire used to screen children for behavioral problems, the Child Behavior Checklist (CBCL). The goals of this research were 2-fold: first, to better understand how changes in brain connectivity over childhood are associated with cognitive and affective development, and second, to evaluate the predictive validity of DLPFC-MPFC and DLPFC-sgACC connectivity for future mental health problems in children who have not been identified previously as being at elevated risk for developing a psychiatric disorder.

tivity over childhood are associated with cognitive and affective development, and second, to evaluate the predictive validity of DLPFC-MPFC and DLPFC-sgACC connectivity for future mental health problems in children who have not been identified previously as being at elevated risk for developing a psychiatric disorder. Numerous studies have demonstrated high reliability between the CBCL scales and actual psychiatric diagnosis. For example, CBCL attention problem scores are used for the screening and prediction of ADHD. A subthreshold elevation on the anxiety/depression subscale of the CBCL in preadolescence predicts future development of MDD. However, in conjunction with behavioral measures, neuroimaging measures may identify children at the greatest risk for developing psychiatric disorders with greater confidence and at an earlier age.

bthreshold elevation on the anxiety/depression subscale of the CBCL in preadolescence predicts future development of MDD. However, in conjunction with behavioral measures, neuroimaging measures may identify children at the greatest risk for developing psychiatric disorders with greater confidence and at an earlier age. Therefore, in this study, we investigated whether rs-fMRI data could predict future CBCL scores in a community sample of 54 children. Specifically, we tested whether the individual differences in MPFC-DLPFC connectivity at age 7 years predict subsequent changes in attention 4 years later, as measured by the CBCL attentional problems measure at age 11 years. Additionally, we performed an exploratory analysis to investigate whether individual differences in sgACC-DLPFC connectivity at age 7 years predict subsequent changes in anxiety/depression 4 years later, as measured by the CBCL “internalization” and anxiety/depressed subscale at age 11 years. We preregistered our hypotheses through the Open Science Framework (OSF; https://osf.io/6cgbs/). Because of space limitations, we report only major results here; the Supplement reports additional findings, as well as a null result based on the preregistered hypotheses.

n” and anxiety/depressed subscale at age 11 years. We preregistered our hypotheses through the Open Science Framework (OSF; https://osf.io/6cgbs/). Because of space limitations, we report only major results here; the Supplement reports additional findings, as well as a null result based on the preregistered hypotheses. Methods Participants Ninety-four participants were included who were enrolled in a developmental longitudinal study, “Predicting Late-Emerging Reading Disability” (Vanderbilt University; principal investigator, L.C.). In this sample, 77 children (82%) met behavioral criteria for typical development; 17 children (18%) were identified as being at risk for a late-emerging reading disability. Time 1 (or baseline) data were collected from participants at age 7 years (n = 94; 41 girls [43.6]) and subsequently at 1-year intervals for 4 years. Data at time 4 were available for 54 of the original participants (57.4%) (see eMethods in the Supplement for exclusion criteria). The CBCL subscale scores at baseline did not differ significantly between those who did and did not complete the study (attentional problems, t91 = 1.0; P = .33; internalization, t91 = 0.51; P = .61; anxiety/depression, t91 = 0.41; P = .68; Table). The study was approved by the institutional review board at Vanderbilt University and written informed consent was received from all participants.

hose who did and did not complete the study (attentional problems, t91 = 1.0; P = .33; internalization, t91 = 0.51; P = .61; anxiety/depression, t91 = 0.41; P = .68; Table). The study was approved by the institutional review board at Vanderbilt University and written informed consent was received from all participants. Table. Child Behavior Checklist Measures for Time 1 at Age 7 Years and Time 4 at Age 11 Yearsa Time Attention Internalization Anxiety/Depression Withdrawn Somatic Time 1 Mean (SD) 56.29 (8.13) 160.02 (13.02) 53.27 (5.3) 53.59 (5.48) 53.37 (5.48) Subclinical (>60), No. (%) 24 (25) 10 (11) 13 (14) 14 (15) 9 (10) Time 4 Mean (SD) 54 (7.46) 160.13 (14.02) 53.11 (5.54) 53.15 (6.32) 53.87 (4.86) Subclinical (>60), No. (%) 10 (20) 6 (11) 9 (17) 8 (15) 9 (17) P (time 1/time 4) 0.09 0.9 0.87 0.88 0.89 Mean Change −1.4 1.27 0.5 −0.17 1.17 Range Change [−16 to 12] = 28 [−41 to 32] = 73 [−17 to 12] = 29 [−12 to 15] = 27 [−17 to 12] = 29 a Higher scores indicate worse problems. A Child Behavior Checklist score of 60 to 70 (>1 SD, <2 SD) is generally considered to represent a medium level of symptoms (subclinical or subthreshold).

e −1.4 1.27 0.5 −0.17 1.17 Range Change [−16 to 12] = 28 [−41 to 32] = 73 [−17 to 12] = 29 [−12 to 15] = 27 [−17 to 12] = 29 a Higher scores indicate worse problems. A Child Behavior Checklist score of 60 to 70 (>1 SD, <2 SD) is generally considered to represent a medium level of symptoms (subclinical or subthreshold). CBCL Scoring and Data Acquisition The CBCL assesses behavioral problems and competencies of children ages 6 to 18 years based on parental reports (eMethods in the Supplement). Data were acquired at Vanderbilt University Institute of Imaging Science on a 3-T Philips Achieva magnetic resonance spectroscopy scanner with a 32-channel head coil. One 5.9-minute resting-state echoplanar imaging scan was collected with the following parameters: TR = 2200 milliseconds, TE = 30 milliseconds, 35 slices, 3-mm isotropic voxels. rs-fMRI Analyses The rs-fMRI data were analyzed in CONN (NeuroImaging Tools & Resources Collaboratory), which incorporates methods to minimize the association of head motion artifacts and allow for valid identification of correlated and anticorrelated networks (see eMethods in the Supplement for a complete description of image preprocessing/denoising, seed selection, bivariate correlation, and independent component analysis).

incorporates methods to minimize the association of head motion artifacts and allow for valid identification of correlated and anticorrelated networks (see eMethods in the Supplement for a complete description of image preprocessing/denoising, seed selection, bivariate correlation, and independent component analysis). Longitudinal Analyses Fisher-transformed r-maps from the MPFC seed were submitted to a second-level analysis of covariance regressing the changes in the CBCL measures (time 4–time 1) onto brain responses, controlling for the effect of initial severity (baseline CBCL). To create a robust prediction model that could be generalized to new cases, we performed leave-1-out cross-validation, which minimizes potential biases due to voxel selection in our predictive models (eMethods in the Supplement).

e 1) onto brain responses, controlling for the effect of initial severity (baseline CBCL). To create a robust prediction model that could be generalized to new cases, we performed leave-1-out cross-validation, which minimizes potential biases due to voxel selection in our predictive models (eMethods in the Supplement). We previously found that the magnitude of MPFC-DLPFC anticorrelations grow during typical development, as does executive function, so we first implemented a whole-brain MPFC seed-based approach to determine whether MPFC-DLPFC correlations at time 1 were associated with future change in CBCL attentional symptoms after controlling for the baseline attentional score. Second, we implemented a more targeted approach by testing whether the baseline connectivity between the MPFC and the DLPFC mask derived from the previous study predicts future attentional symptoms. Finally, we ran an exploratory analysis using an independent component analysis–defined DLPFC-sgACC component to test whether the connectivity of this component predicted change of internalization symptoms and subsequently examined the internalization subscales separately: (1) anxiety/depression, (2) withdrawn behavior, and (3) somatic complaints.

y analysis using an independent component analysis–defined DLPFC-sgACC component to test whether the connectivity of this component predicted change of internalization symptoms and subsequently examined the internalization subscales separately: (1) anxiety/depression, (2) withdrawn behavior, and (3) somatic complaints. Logistic Regression for CBCL Internalization (and Anxiety/Depression Subscale) As per CBCL diagnostic category definitions, we subdivided participants into a “subclinical” category for individuals with a CBCL internalization (and anxiety/depression) score of 55 or greater and a “typical” category for those whose scores on this subscale fell below this cutoff based on the literature. We used a logistic regression of initial severity (baseline CBCL scores) and baseline resting-state measures combined with leave-1-out cross-validation. We did not have enough participants with subclinical scores for the CBCL attentional problems at time 4 to perform this logistic regression for that CBCL scale. Finally, we correlated the changes in connectivity with changes in CBCL measures over 4 years (from age 7 years to age 11 years). Conceptual Replication/Clinical Extension We tested the prediction model on an independent sample of 25 youths between ages 8 to 14 years identified as being at familial risk for MDD as well as 18 age-matched children without familial risk for MDD. We used baseline sgACC-DLPFC connectivity to predict the progression of CBCL anxiety/depression 3 years later (eTable in the Supplement).

model on an independent sample of 25 youths between ages 8 to 14 years identified as being at familial risk for MDD as well as 18 age-matched children without familial risk for MDD. We used baseline sgACC-DLPFC connectivity to predict the progression of CBCL anxiety/depression 3 years later (eTable in the Supplement). Results Behavioral Results and Head Motion Between ages 7 and 11 years, 14 children (26%) had significant changes in internalizing scores (9 [17%] showing more internalizing problems at age 11 years and 5 [9%] showing fewer) and 8 children (15%) had significant changes in attentional problem scores (1 [2%] showing more attentional problems at age 11 years and 17 [3%] showing fewer). The mean (SD) number of outliers across all points was 17 of 160 (21) points. Excluding these points preserved enough data to achieve a stable estimate of RSNs (eMethods in the Supplement).

ficant changes in attentional problem scores (1 [2%] showing more attentional problems at age 11 years and 17 [3%] showing fewer). The mean (SD) number of outliers across all points was 17 of 160 (21) points. Excluding these points preserved enough data to achieve a stable estimate of RSNs (eMethods in the Supplement). Neuroimaging Results Cross-sectional analyses at time 1 (N = 94) revealed that, on average, children age 7 years did not exhibit the significant negative MPFC-DLPFC correlations that are evident in adults but rather exhibited positive MPFC-DLPFC correlations on the whole, consistent with prior findings from children ages 8 to 12-years (eFigure 1 in the Supplement). We had predicted in our preregistration that there would be insufficient variance in the CBCL attentional scores to establish a significant brain-CBCL association in this sample. Indeed, we did not observe any significant correlations between the MPFC-DLPFC connectivity and CBCL attentional scores at a height threshold of P < .001 (t92 > 3.40) uncorrected (or even at a liberal threshold of P < . 01 uncorrected).

ttentional scores to establish a significant brain-CBCL association in this sample. Indeed, we did not observe any significant correlations between the MPFC-DLPFC connectivity and CBCL attentional scores at a height threshold of P < .001 (t92 > 3.40) uncorrected (or even at a liberal threshold of P < . 01 uncorrected). Although there was minimal average change in CBCL scores, there was considerable interparticipant variability in the change of CBCL scores across 4 years. Here, we used baseline neuroimaging data at age 7 years to predict CBCL change over 4 years. Less positive MPFC-DLFPC correlations at time 1 were associated with improvement of attentional problems 4 years later (t49 = 2.38, P = .01, controlling for medication; t49 = 1.02, P = .03, controlling for those children who received a diagnosis of ADHD; t50 = 2.36, P = .01 without controlling for participants with ADHD participants; reported P values are 1-sided because of our a priori and preregistered hypotheses) (eFigure 2 in the Supplement). Because we implemented this analysis using leave-1-out cross-validation, this is a prediction as opposed to a simple correlation, a distinction that is frequently lost in the neuroimaging literature. Furthermore, we found that less positive MPFC-DLPFC (a priori mask) correlations at time 1 were associated with an improvement of attentional problems 4 years later (r = 0.3; P = .04; Figure 1).

his is a prediction as opposed to a simple correlation, a distinction that is frequently lost in the neuroimaging literature. Furthermore, we found that less positive MPFC-DLPFC (a priori mask) correlations at time 1 were associated with an improvement of attentional problems 4 years later (r = 0.3; P = .04; Figure 1). Figure 1. Longitudinal Prediction of Change in Attentional Problems From Ages 7 to 11 Years Baseline medial prefrontal cortex and dorsolateral prefrontal cortex (MPFC-DLPFC) (a priori mask) anticorrelations were associated with changes in attentional problems 4 years later. Negative change scores indicate improvement and positive change scores indicate decline over 4 years. The peak MPFC coordinates are −1, 47, −4 and the peak coordinates for the DLPFC mask are 46, 46, 6. R indicates right; time 1, t 1; time 4, t 4.

tions were associated with changes in attentional problems 4 years later. Negative change scores indicate improvement and positive change scores indicate decline over 4 years. The peak MPFC coordinates are −1, 47, −4 and the peak coordinates for the DLPFC mask are 46, 46, 6. R indicates right; time 1, t 1; time 4, t 4. Weaker left DLPFC-sgACC connectivity at baseline predicted a greater worsening of internalization CBCL scores by time 4 (t49 = −2.4; P = .01; controlling for medication; Figure 2) and (t49 = −2.15; P = .02, controlling for ADHD) and (t50 = −2.61; P = .01, not controlling for ADHD or medication). Specifically, weaker left DLPFC-sgACC connectivity (or stronger anticorrelations) predicted greater worsening on the internalization subscales of anxiety/depression (t49 = −2.64; P = .005, controlling for medication) and withdrawn (t49 = −2.38; P = .01, controlling for medication). By contrast, left DLPFC-sgACC connectivity was not associated with changes in somatic complaints (t49 = −0.88; P = .19, controlling for medication). Based on our previous work, we had hypothesized that the sgACC-DMN connectivity would predict a worsening of internalization in our preregistration; however, this analysis did not reach current statistical threshold standards (eMethods in the Supplement).

c complaints (t49 = −0.88; P = .19, controlling for medication). Based on our previous work, we had hypothesized that the sgACC-DMN connectivity would predict a worsening of internalization in our preregistration; however, this analysis did not reach current statistical threshold standards (eMethods in the Supplement). Figure 2. Longitudinal Prediction of Change in Internalization Problems From Ages 7 to 11 Years A, Left dorsolateral prefrontal cortex (DLPFC) and subgenual anterior cingulate cortex (sgACC) predicted change in internalization (and anxiety/depression and withdrawn subscales) such that a greater anticorrelation at time 1 (age 7 years) predicted a worsening of internalization 4 years later (age 11 years). B, Scatterplot of longitudinal prediction. Negative change scores indicate improvement and positive change scores indicate decline over 4 years. C, Logistic regression using sgACC-DLPFC to predict internalization problems at time 4 (≥55), controlling for internalization scores at time 1 (t 1). The histograms represent the distribution of the risk of internalization problems at time 4 (t 4) (as predicted by sgACC-DLPFC connectivity at time 1) displayed separately for those participants with low (left) vs high (right) internalization problem scores at time 4. Logistic Regression Logistic regression analyses revealed that sgACC-DLPFC connectivity was a more accurate predictor than baseline CBCL measures for progression to a subclinical score on internalization (t50 = −2.61; P = .01; Figure 2). This analysis yielded 77% accuracy, 87% sensitivity, and 74% specificity.

Figure 2. Longitudinal Prediction of Change in Internalization Problems From Ages 7 to 11 Years A, Left dorsolateral prefrontal cortex (DLPFC) and subgenual anterior cingulate cortex (sgACC) predicted change in internalization (and anxiety/depression and withdrawn subscales) such that a greater anticorrelation at time 1 (age 7 years) predicted a worsening of internalization 4 years later (age 11 years). B, Scatterplot of longitudinal prediction. Negative change scores indicate improvement and positive change scores indicate decline over 4 years. C, Logistic regression using sgACC-DLPFC to predict internalization problems at time 4 (≥55), controlling for internalization scores at time 1 (t 1). The histograms represent the distribution of the risk of internalization problems at time 4 (t 4) (as predicted by sgACC-DLPFC connectivity at time 1) displayed separately for those participants with low (left) vs high (right) internalization problem scores at time 4. Logistic Regression Logistic regression analyses revealed that sgACC-DLPFC connectivity was a more accurate predictor than baseline CBCL measures for progression to a subclinical score on internalization (t50 = −2.61; P = .01; Figure 2). This analysis yielded 77% accuracy, 87% sensitivity, and 74% specificity. Association of Brain Connectivity Changes With Changes in the CBCL and Conceptual Replication/Clinical Extension An increase in MPFC-DLPFC anticorrelations correlated with an improvement of CBCL attentional scores. and an increase in sgACC-DLPFC anticorrelations correlated with a worsening of CBCL anxiety/depression scores over 4 years (eFigures 3 and 4 in the Supplement). Weaker DLPFC-sgACC connectivity (or stronger anticorrelations) at baseline predicted worsening on the internalization subscale of anxiety/depression 3 years later for children at familial risk for MDD as well as a new sample of typically developing children (at-risk: r = −0.75; P < .001; controls: r = −0.81; P = .01; Figure 3 and eFigure 3 in the Supplement).

ger anticorrelations) at baseline predicted worsening on the internalization subscale of anxiety/depression 3 years later for children at familial risk for MDD as well as a new sample of typically developing children (at-risk: r = −0.75; P < .001; controls: r = −0.81; P = .01; Figure 3 and eFigure 3 in the Supplement). Figure 3. Conceptual Replication/Clinical Extension A, Longitudinal prediction of change in Child Behavior Checklist (CBCL) anxiety/depression symptoms over 3 years in children with (blue) or without (red) familial risk for depression. Baseline resting-state subgenual anterior cingulate cortex–dorsolateral prefrontal cortex (sgACC-DLPFC) connectivity predicted a change of anxiety/depression, such that less positive correlations at time 1 predicted a worsening of anxiety/depression scores 3 years later. B, Scatterplot of longitudinal prediction. Negative change scores indicate improvement, and positive change scores indicate decline over 3 years.

DLPFC) connectivity predicted a change of anxiety/depression, such that less positive correlations at time 1 predicted a worsening of anxiety/depression scores 3 years later. B, Scatterplot of longitudinal prediction. Negative change scores indicate improvement, and positive change scores indicate decline over 3 years. Discussion The RSNs at age 7 years in a community sample of children predicted the developmental trajectory of symptoms associated with ADHD and MDD at age 11 years. The variations in functional connectivity occurred in neural systems that are known to be salient for attention or mood. Weaker positive MPFC-DLPFC correlations at age 7 years predicted improved attention scores at age 11 years, whereas weaker positive sgACC-left DLPFC correlations at age 7 years predicted a worsening of MDD symptoms (internalization) at age 11 years. It is noteworthy that most children with attentional problems at age 7 years exhibited reduced symptoms at age 11 years, whereas most children with internalizing symptoms at age 7 years exhibited more symptoms at age 11 years. Thus, our functional connectivity measures appear to be sensitive to resilience and vulnerability.

noteworthy that most children with attentional problems at age 7 years exhibited reduced symptoms at age 11 years, whereas most children with internalizing symptoms at age 7 years exhibited more symptoms at age 11 years. Thus, our functional connectivity measures appear to be sensitive to resilience and vulnerability. The associations between specific neural networks and specific longitudinal declines are consistent with prior findings. That a stronger positive MPFC-DLPFC coupling was associated with a worse developmental trajectory for attention is consistent with the hypothesis that anticorrelated MPFC-DLPFC activation is associated with the ability to selectively focus attention on internal thoughts vs external stimuli. Weaker anticorrelations between MPFC and DLPFC, which are core nodes of DMN and CEN, respectively, may reflect an attenuation of top-down control mechanisms and an inability to allocate resources away from internal thoughts and feelings and toward external stimuli to adaptively perform difficult tasks. Thus, children age 7 years who exhibit MPFC-DLPFC anticorrelations may have the capacity to toggle between internal and external foci of attention more readily than those who do not. The failure to decouple these networks may be an early indicator of attentional problems or may preclude the development of age-appropriate attentional skills.

this young age may already reflect an attenuation or failure of top-down control mechanisms that are evident in adult MDD. Thus, the functional connectivity of specific neural systems in middle childhood forecasts individuals’ vulnerability or resilience in cognition and emotion over the ensuing 4 years of development. These findings extend the use of neuroimaging to identify childhood neuromarkers of risk for psychopathology from highly selected children, such as those with identified familial risk, to a sample of children more representative of the population as a whole. Although children with parents who have had depression are at an elevated risk for developing depression, most children who develop depression do not come from families with an identified history of depression. Further, the longitudinal nature of this study supports the validity of using RSNs to predict the worsening of psychiatric symptoms in childhood. Although variation in RSNs forecasts the developmental trajectory of attentional and emotional symptoms, there is strong evidence that such networks are plastic, and thus may be altered by supportive interventions. Resting-state functional connectivity is thought to reflect habitual network activations that can be remodeled by various long-term and even brief behavioral interventions and pharmacological interventions.

ms, there is strong evidence that such networks are plastic, and thus may be altered by supportive interventions. Resting-state functional connectivity is thought to reflect habitual network activations that can be remodeled by various long-term and even brief behavioral interventions and pharmacological interventions. Limitations First, although some children developed subclinical scores on CBCL measures by age 11 years, we do not know which children have subsequently converted to psychiatric diagnoses. However, elevated scores on the CBCL measures, such as internalization (including anxiety/depression), are highly predictive of near-term conversion to psychiatric diagnoses. Second, the current sample size was too small to make any meaningful interpretations for the subset of participants who moved between clinical categories over time. Third, our targeted, hypothesized-driven approach could be viewed by some readers as a limitation of the study. Conclusions These findings not only further our understanding of the neurobiological vulnerabilities that foster the deterioration of mental health, but also could inform early identification and preventative treatment. Identification of risk at the level of individual children may be strengthened by large multilevel data sets that integrate multimodal neuroimaging, genetics, and social factors with new statistical tools. These findings illustrate the idea that the neurobiological seeds of future psychopathology are becoming visible and measurable in children. Supplement. eMethods.

Conclusions These findings not only further our understanding of the neurobiological vulnerabilities that foster the deterioration of mental health, but also could inform early identification and preventative treatment. Identification of risk at the level of individual children may be strengthened by large multilevel data sets that integrate multimodal neuroimaging, genetics, and social factors with new statistical tools. These findings illustrate the idea that the neurobiological seeds of future psychopathology are becoming visible and measurable in children. Supplement. eMethods. eTable. Table of cluster statistics and locations from Figure 3 eFigure 1. On average, children 7 years of age exhibit positive MPFC-DLPFC resting state connectivity eFigure 2. Longitudinal prediction of progression of attentional problems over four years (ages 7-11) eFigure 3. Increase in MPFC-DLPFC anticorrelations correlates with improvement of CBCL attentional scores over 4 years. eFigure 4. Increase sgACC-DLPFC anticorrelations correlates with a worsening of CBCL anxiety/depression scores over 4 years. Click here for additional data file.

Introduction Individuals diagnosed with psychiatric disorders may experience a range of adverse outcomes, with elevated risks of premature mortality, suicide, unemployment, and homelessness. In addition, these individuals have more contact with the criminal justice system and an increased risk of engaging in violent crime compared with the general population and with their siblings without psychiatric disorders. The evidence base regarding perpetration risks needs to be interpreted in the context of subjection to violence among individuals with psychiatric disorders. Reviews of previous research, expert opinion, and advocacy groups report that the rate of subjection to violence is considerably higher than the rate of perpetration of violence, and it is commonly suggested that this rate is elevated approximately 10-fold.

tion to violence among individuals with psychiatric disorders. Reviews of previous research, expert opinion, and advocacy groups report that the rate of subjection to violence is considerably higher than the rate of perpetration of violence, and it is commonly suggested that this rate is elevated approximately 10-fold. However, evidence for increased rates of subjection to violence in individuals with psychiatric disorders is limited. First, systematic reviews have reported large but imprecise relative risks, ranging from a factor of 2 to 140, for subjection to violence in individuals with any psychiatric disorder compared with the general population. These reviews have primarily been based on cross-sectional studies using small and selected clinical samples that have relied on retrospective self-reports. Some of these limitations have been addressed by 2 population-based studies. However, these investigations did not adequately control for premorbid subjection to violence; hence, they were unable to exclude the possibility of reverse causation (ie, subjection to violence as the cause of the psychiatric disorder rather than vice versa) given the evidence indicating an association between subjection to violence in early life (and its related trauma) and psychiatric morbidity in adulthood.

hence, they were unable to exclude the possibility of reverse causation (ie, subjection to violence as the cause of the psychiatric disorder rather than vice versa) given the evidence indicating an association between subjection to violence in early life (and its related trauma) and psychiatric morbidity in adulthood. Second, twin and family studies have reported that psychiatric disorders, violent crime, and subjection to violence tend to aggregate in families, but their etiological associations remain poorly understood. The literature suggests, however, that the estimates in the studies examining associations between psychiatric disorders and subjection to violence may have been biased upwards because of substantial unmeasured familial confounding. Third, only a few studies have explicitly considered the co-occurrence of subjection to violence and perpetration of violence. This point is notable because previous studies have suggested that the co-occurrence may represent a distinct subgroup that is differentially associated with psychiatric disorders. To address these gaps in knowledge, we conducted a study of the entire Swedish population born between January 1, 1973, and December 31, 1993. This approach allowed us to assess the potential associations of a wide range of psychiatric disorders with the risks of subjection to and perpetration of violence while accounting for unmeasured familial confounding.

onducted a study of the entire Swedish population born between January 1, 1973, and December 31, 1993. This approach allowed us to assess the potential associations of a wide range of psychiatric disorders with the risks of subjection to and perpetration of violence while accounting for unmeasured familial confounding. Methods Data Collection All Swedish residents are assigned a unique 10-digit civic registration number, which is used in different nationwide registers and provides accurate linkage. We received deidentified data from Statistics Sweden after the study was approved by the regional research ethics committee of Karolinska Institutet. Informed consent is not a requirement for nationwide register-based studies in Sweden.

ber, which is used in different nationwide registers and provides accurate linkage. We received deidentified data from Statistics Sweden after the study was approved by the regional research ethics committee of Karolinska Institutet. Informed consent is not a requirement for nationwide register-based studies in Sweden. The Multi-Generation Register provided data on all individuals born in Sweden and their biological parents, which enabled identification of full biological siblings. The National Patient Register provided data on all inpatient hospitalization episodes (International Classification of Diseases, Revision 8 [ICD-8], International Classification of Diseases, Ninth Revision [ICD-9], and International Statistical Classification of Diseases and Related Health Problems, Tenth Revision [ICD-10] data from 1973-2013) and specialist outpatient care visits (ICD-10 data from 2001-2013) and is comprehensive of Swedish universal health care coverage. Violent crime convictions were derived from the National Crime Register, which includes information on criminal convictions beginning in 1973. Data on sociodemographic factors were gathered from census registers. The Migration Register and the Causes of Death Register, respectively, provided emigration and mortality dates. This study followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for cohort studies (eTable 1 in the Supplement).

aphic factors were gathered from census registers. The Migration Register and the Causes of Death Register, respectively, provided emigration and mortality dates. This study followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for cohort studies (eTable 1 in the Supplement). Subjection to violence was defined as an outpatient visit (excluding primary care), inpatient care episode, or death related to any diagnosis of an injury purposefully inflicted by other persons (ICD codes and validation information in eMethods and eTable 2 in the Supplement). Violent perpetration was defined as a conviction for homicide, assault, robbery, violence against an officer, arson, or sexual offenses (excluding prostitution, solicitation of prostitution, or possession of child pornography). Individuals are convicted in Swedish courts irrespective of psychiatric disorder, although sentencing may be informed by such conditions. The patient data were not used to determine perpetration of violence status, and the conviction data were not used to determine subjection to violence status.

of child pornography). Individuals are convicted in Swedish courts irrespective of psychiatric disorder, although sentencing may be informed by such conditions. The patient data were not used to determine perpetration of violence status, and the conviction data were not used to determine subjection to violence status. From a population sample of all individuals born in Sweden between January 1, 1973, and December 31, 1993 (eMethods in the Supplement), we identified all patients diagnosed with a psychiatric disorder older than 15 years (n = 250 419). Premorbid subjection to violence was measured since birth. We adopted a hierarchical approach to differentiate between schizophrenia, bipolar disorder, depression, and anxiety disorder (ICD codes in eTable 2 in the Supplement). We also examined patients with personality disorders, alcohol use disorders, and drug use disorders.

d subjection to violence was measured since birth. We adopted a hierarchical approach to differentiate between schizophrenia, bipolar disorder, depression, and anxiety disorder (ICD codes in eTable 2 in the Supplement). We also examined patients with personality disorders, alcohol use disorders, and drug use disorders. Control Groups We individually matched each patient by sex and birth year with 10 individuals in the general population who did not have that particular psychiatric disorder. Participants in the general population control group had to be alive and Swedish residents at the date of matching (eg, when the index person first received the psychiatric diagnosis). Patients could be matched with multiple individuals because the analyses only considered the associations within each cluster of patients and individuals from the general population. We also matched the patients with their full biological siblings who did not have psychiatric disorders to assess the role of unmeasured familial confounding. To maintain a high degree of statistical power, we analyzed all potential sibling pairs in the main analyses, with covariate adjustments for age and sex. The sibling comparison approach allowed us to account for all time-constant unmeasured familial confounding factors shared between siblings (eg, half of their cosegregating genes and their shared childhood environments). The extent to which the sibling comparisons were attenuated compared with the population estimates indicated the influence of unmeasured familial confounding.

ll time-constant unmeasured familial confounding factors shared between siblings (eg, half of their cosegregating genes and their shared childhood environments). The extent to which the sibling comparisons were attenuated compared with the population estimates indicated the influence of unmeasured familial confounding. The start date for the patients and control groups was defined as the discharge date of the first psychiatric episode. The participants were censored either when they migrated, died, experienced the outcome of interest, or reached the end of the study period on December 31, 2013. Statistical Analysis We quantified the associations between psychiatric disorders and subjection to and perpetration of violence by fitting stratified Cox models that estimated adjusted hazard ratios (aHRs). Because each person diagnosed with a psychiatric disorder and their matches in the general population and sibling control groups were separately defined as unique strata, the model used this information to estimate varying baseline hazard rates across each combination of patients and individuals in the control groups. This approach implies that the comparisons were made within each stratum.

es in the general population and sibling control groups were separately defined as unique strata, the model used this information to estimate varying baseline hazard rates across each combination of patients and individuals in the control groups. This approach implies that the comparisons were made within each stratum. We initially fitted a crude model that accounted for sex and birth year (model 1). We subsequently further adjusted the model for birth order and parental background factors (model 2) as well as the individual’s history of subjection to violence and perpetration of violence (model 3). In model 4, we adjusted for unmeasured familial risks by refitting model 3 on the subsamples of differentially affected siblings. Model 4 was then refitted to each sex separately to assess moderation effects by sex. We examined the associations between specific psychiatric disorders and outcomes by testing each of them individually (model 4) and jointly adjusting for them (model 5).

g model 3 on the subsamples of differentially affected siblings. Model 4 was then refitted to each sex separately to assess moderation effects by sex. We examined the associations between specific psychiatric disorders and outcomes by testing each of them individually (model 4) and jointly adjusting for them (model 5). Conditional multinomial logistic regression models estimating adjusted odds ratios (aORs) were used to examine the associations between diagnosis with any psychiatric disorder and subjection to and perpetration of violence status, which was defined as an unordered categorical variable with the following categories: (1) neither subjected to violence nor perpetrated violence, (2) subjected to violence only, (3) perpetrated violence only, and (4) both subjected to violence and perpetrated violence. Sensitivity tests for alternative measurement definitions and model specifications were also performed (eMethods in the Supplement). Data were analyzed from January 15 to September 14, 2019. We accounted for measured confounders, including birth order, parental background factors (eg, low income, low educational level, immigrant background, and history of psychiatric disorders and violent criminality), and the individual’s history of subjection to and perpetration of violence (definitions in eMethods in the Supplement).

or measured confounders, including birth order, parental background factors (eg, low income, low educational level, immigrant background, and history of psychiatric disorders and violent criminality), and the individual’s history of subjection to and perpetration of violence (definitions in eMethods in the Supplement). Results The patient sample comprised 250 419 individuals, of which 138 622 (55.4%) were women and 111 797 (44.6%) were men. The patients were individually matched with 10 people in the general population without psychiatric disorders (n = 2 504 190) and with their siblings without psychiatric disorders (n = 194 788). The largest patient groups included those diagnosed with depression (n = 103 814) or alcohol use disorder (n = 69 116; Table). The median (interquartile range) age at first diagnosis ranged from 20.0 (17.4-24.0) years for alcohol use disorder to 23.7 (19.9-28.8) years for anxiety disorder (Table). The participants had mean (SD) of 7.3 (4.5) years of postdischarge data available. Table. Participant Characteristics by Psychiatric Diagnosis Characteristic No. (%) Participants Without Psychiatric Disorder Diagnosis Participants With Psychiatric Disorder Diagnosis Anxiety Depression Bipolar Disorder Schizophrenia Personality Disorder Alcohol Use Disorder Drug Use Disorder Total, No.

Results The patient sample comprised 250 419 individuals, of which 138 622 (55.4%) were women and 111 797 (44.6%) were men. The patients were individually matched with 10 people in the general population without psychiatric disorders (n = 2 504 190) and with their siblings without psychiatric disorders (n = 194 788). The largest patient groups included those diagnosed with depression (n = 103 814) or alcohol use disorder (n = 69 116; Table). The median (interquartile range) age at first diagnosis ranged from 20.0 (17.4-24.0) years for alcohol use disorder to 23.7 (19.9-28.8) years for anxiety disorder (Table). The participants had mean (SD) of 7.3 (4.5) years of postdischarge data available. Table. Participant Characteristics by Psychiatric Diagnosis Characteristic No. (%) Participants Without Psychiatric Disorder Diagnosis Participants With Psychiatric Disorder Diagnosis Anxiety Depression Bipolar Disorder Schizophrenia Personality Disorder Alcohol Use Disorder Drug Use Disorder Total, No. 2 504 190 68 244 103 814 17 309 4153 29 713 69 116 37 039 Age at first diagnosis, median (IQR), y NA 23.7 (19.9-28.8) 23.1 (19.2-28.2 ) 22.9 (19.3-27.9) 22.5 (19.4-26.6) 21.8 (18.9-26.1) 20.0 (17.4-24.0) 21.3 (18.7-24.9) Sex Male 1 117 970 (44.6) 27 787 (40.7) 39 448 (38.0) 5844 (33.8) 2753 (66.3) 10 676 (35.9) 38 987 (56.4) 24 161 (65.2) Female 1 386 220 (55.4) 40 457 (59.3) 64 366 (62.0) 11 465 (66.2) 1400 (33.7) 19 037 (64.1) 30 129 (43.6) 12 878 (34.8) Birth order 1 1 036 966 (41.4) 28 405 (41.6) 42 922 (41.3) 7438 (43.0) 1696 (40.8) 12 571 (42.3) 27 383 (39.6) 15 270 (41.2) 2 923 286 (36.9) 24 094 (35.3) 36 417 (35.1) 5926 (34.2) 1557 (37.5) 9926 (33.4) 24 931 (36.1) 12 924 (34.9) 3 394 206 (15.7) 10 680 (15.6) 16 712 (16.1) 2720 (15.7) 575 (13.8) 4767 (16.0) 11 415 (16.5) 5899 (15.9) ≥4 149 732 (6.0) 5065 (7.4) 7763 (7.5) 1225 (7.1) 325 (7.8) 2449 (8.2) 5387 (7.8) 2946 (8.0) Immigrant background No 2 280 324 (91.1) 61 158 (89.6) 93 885 (90.4) 15683 (90.6) 3551 (85.5) 26 590 (89.5) 62 708 (90.7) 31 955 (86.3) Yes 223 866 (8.9) 7086 (10.4) 9929 (9.6) 1626 (9.4) 602 (14.5) 3123 (10.5) 6408 (9.3) 5084 (13.7) Parental income in bottom decile No 2 273 931 (90.8) 58 879 (86.3) 89 318 (86.0) 14 708 (85.0) 3247 (78.2) 24 236 (81.6) 58 960 (85.3) 29 659 (80.1) Yes 230 259 (9.2) 9365 (13.7) 14496 (14.0) 2601 (15.0) 906 (21.8) 5477 (18.4) 10 156 (14.7) 7380 (19.9) Low parental education level No 2 323 875 (92.8) 62 116 (91.0) 95 145 (91.6) 16 020 (92.6) 3727 (89.7) 26 820 (90.3) 62 947 (91.1) 33 018 (89.1) Yes 180 315 (7.2) 6128 (9.0) 8669 (8.4) 1289 (7.4) 426 (10.3) 2893 (9.7) 6169 (8.9) 4021 (10.9) Parental lifetime violent crime conviction No 2 349 118 (93.8) 60 514 (88.7) 91 719 (88.3) 15 272 (88.2) 3595 (86.6) 25 026 (84.2) 58 678 (84.9) 29 158 (78.7) Yes 155 072 (6.2) 7730 (11.3) 12 095 (11.7) 2037 (11.8) 558 (13.4) 4687 (15.8) 10 438 (15.1) 7881 (21.3) Parental lifetime psychiatric morbidity No 1 920 673 (76.7) 42 344 (62.0) 62 145 (59.9) 9606 (55.5) 2305 (55.5) 16 226 (54.6) 41 781 (60.5) 19 132 (51.7) Yes 583 517 (23.3) 25 900 (38.0) 41 669 (40.1) 7703 (44.5) 1848 (44.5) 13 487 (45.4) 27 335 (39.5) 17 907 (48.3) History of violence None 2 442 123 (97.5) 62 98

21.3) Parental lifetime psychiatric morbidity No 1 920 673 (76.7) 42 344 (62.0) 62 145 (59.9) 9606 (55.5) 2305 (55.5) 16 226 (54.6) 41 781 (60.5) 19 132 (51.7) Yes 583 517 (23.3) 25 900 (38.0) 41 669 (40.1) 7703 (44.5) 1848 (44.5) 13 487 (45.4) 27 335 (39.5) 17 907 (48.3) History of violence None 2 442 123 (97.5) 62 98 7 (92.3) 96 599 (93.1) 16 123 (93.1) 3724 (89.7) 26 444 (89.0) 60 643 (87.7) 28 351 (76.5) Subjected to violence only 28 305 (1.1) 1966 (2.9) 2856 (2.8) 513 (3.0) 77 (1.9) 959 (3.2) 2960 (4.3) 1934 (5.2) Perpetrated violence only 30 279 (1.2) 2774 (4.1) 3714 (3.6) 578 (3.3) 317 (7.6) 1982 (6.7) 4518 (6.5) 5677 (15.3) Both subjected to violence and perpetrated violence 3483 (0.1) 517 (0.8) 645 (0.6) 95 (0.5) 35 (0.8) 328 (1.1) 995 (1.4) 1077 (2.9) Abbreviation: IQR, interquartile range.

) 959 (3.2) 2960 (4.3) 1934 (5.2) Perpetrated violence only 30 279 (1.2) 2774 (4.1) 3714 (3.6) 578 (3.3) 317 (7.6) 1982 (6.7) 4518 (6.5) 5677 (15.3) Both subjected to violence and perpetrated violence 3483 (0.1) 517 (0.8) 645 (0.6) 95 (0.5) 35 (0.8) 328 (1.1) 995 (1.4) 1077 (2.9) Abbreviation: IQR, interquartile range. Less than half of the individuals who were diagnosed with any psychiatric disorder were either subjected to violence to the extent that they required medical treatment or were convicted of a violent crime after the onset of their condition. The unadjusted incidence rates were similar between both outcomes (7.1 [95% CI, 6.9-7.2] vs 7.5 [95% CI, 7.4-7.6] per 1000 person-years among patients diagnosed with psychiatric disorders and 1.0 [95% CI, 0.9-1.0] vs 0.7 [95% CI, 0.7-0.7] per 1000 person-years among individuals without psychiatric disorders) (Figure 1A). The 10-year cumulative incidence rates adjusted for sex and birth year were also similar, ranging from 6.4% to 6.5% among patients with psychiatric disorders and from 0.9% to 0.6% among individuals without psychiatric disorders (eFigures 1 and 2 in the Supplement). In addition, we found that the outcomes co-occurred to a moderate extent among patients with psychiatric disorders (r = 0.43 [95% CI, 0.41-0.45]).

nging from 6.4% to 6.5% among patients with psychiatric disorders and from 0.9% to 0.6% among individuals without psychiatric disorders (eFigures 1 and 2 in the Supplement). In addition, we found that the outcomes co-occurred to a moderate extent among patients with psychiatric disorders (r = 0.43 [95% CI, 0.41-0.45]). Figure 1. Risk of Subjection to Violence and Perpetration of Violence Among Individuals Diagnosed With Any Psychiatric Disorder Compared With Individuals Without a Psychiatric Disorder Model 1 included matches by sex and birth year. Model 2 was adjusted for birth order and parental characteristics (immigrant background, low income, low educational level, lifetime violent crime conviction, and psychiatric history). Model 3 was further adjusted for the individual’s history of subjection to and perpetration of violence. Model 4 included within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families.

uded within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families. We initially found that persons who were diagnosed with any psychiatric disorder were more than 7 times as likely as those without psychiatric disorders to be subjected to violence (aHR, 7.4 [95% CI, 7.2-7.5]; model 1 in Figure 1B). Further adjustments for birth order and parental confounders (model 2) and the individual’s history of subjection to violence and perpetration of violence (model 3) attenuated those estimates to a nearly 6-fold risk increase (aHR, 5.8 [95% CI, 5.6-5.9]). Unmeasured familial confounders were important because the sibling comparison estimate (model 4) further attenuated the association to an approximately 3-fold risk increase (aHR, 3.4 [95% CI, 3.2-3.6]). We observed a similar pattern of associations for violent perpetration as the outcome, ranging from an 11-fold risk increase in the crude model (aHR, 11.2 [95% CI, 10.9-11.5]) to a 4-fold risk increase in the fully adjusted model (aHR, 4.2 [95% CI, 3.9-4.4]). We found the model 4 estimates to be robust to most model specifications, with effect sizes typically ranging from a 3- to 4-fold risk increase across both of the outcomes (eFigure 3 in the Supplement).

l (aHR, 11.2 [95% CI, 10.9-11.5]) to a 4-fold risk increase in the fully adjusted model (aHR, 4.2 [95% CI, 3.9-4.4]). We found the model 4 estimates to be robust to most model specifications, with effect sizes typically ranging from a 3- to 4-fold risk increase across both of the outcomes (eFigure 3 in the Supplement). We observed sex differences in the distribution of the violence outcomes (Figure 2A). Men with any psychiatric disorder were approximately 3 times more likely to be subjected to violence (aHR, 2.8 [95% CI, 2.5-3.0]) and approximately 4 times more likely to perpetrate violence (aHR, 3.8 [95% CI, 3.5-4.1]; Figure 2B) than their siblings without psychiatric disorders. We observed minor differences for the equivalent estimates in women among those who were subjected to violence (aHR, 4.3 [95% CI, 3.8-5.0]) and among those who perpetrated violence (aHR, 4.6 [95% CI, 3.7-5.7]) (Figure 2B). Assuming that the violence outcomes were directly comparable, we observed that both men and women with any psychiatric disorder were more likely than their siblings without psychiatric disorders to be both subjected to violence and to perpetrate violence (aOR, men, 8.6 [95% CI, 6.8-10.8]; aOR, women, 19.8 [95% CI, 6.4-61.7]) than to have solely experienced subjection to violence (aOR, men, 2.5 [95% CI, 2.3-2.8]; aOR, women, 4.3 [95% CI, 3.7-4.9]) or to have solely perpetrated violence (aOR, men, 3.8 [95% CI, 3.4-4.2]; aOR, women, 4.5 [95% CI, 3.7-4.9]; Figure 3).

lence (aOR, men, 8.6 [95% CI, 6.8-10.8]; aOR, women, 19.8 [95% CI, 6.4-61.7]) than to have solely experienced subjection to violence (aOR, men, 2.5 [95% CI, 2.3-2.8]; aOR, women, 4.3 [95% CI, 3.7-4.9]) or to have solely perpetrated violence (aOR, men, 3.8 [95% CI, 3.4-4.2]; aOR, women, 4.5 [95% CI, 3.7-4.9]; Figure 3). Figure 2. Sex-Stratified Risk of Subjection to Violence and Perpetration of Violence Among Individuals Diagnosed With Any Psychiatric Disorder Compared With Siblings Without Psychiatric Disorders The adjusted hazard ratios refer to within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families. Figure 3. Adjusted Odds Ratios for Subjection to Violence Only, Perpetration of Violence Only, and Both Subjection to and Perpetration of Violence Among Men and Women With Psychiatric Disorders The adjusted odds ratios refer to within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families.

r to within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families. Further stratification by psychiatric diagnoses indicated that the outcomes were more common among persons diagnosed with drug and alcohol use disorders (12-17 subjection to violence events and 13-27 perpetration of violence events per 1000 person-years) than in other diagnostic groups (Figure 4A). We initially found that persons with any of the specific psychiatric disorders were more likely than their siblings without psychiatric disorders to be subjected to violence (range between aHR, 2.3 [95% CI, 1.5-3.5] to aHR, 5.3 [95% CI, 4.3-6.7], respectively) and to perpetrate violence against others (range between aHR, 3.1 [95% CI, 2.7-3.6] to aHR, 9.6 [95% CI, 5.6-16.6], respectively; model 4 in eFigure 4 in the Supplement). However, when we jointly adjusted for all of the conditions, we found that these estimates were attenuated but remained statistically significant for both outcomes, with the sole exception of persons diagnosed with schizophrenia, who did not have a higher risk of being subjected to violence when compared with their siblings without psychiatric disorders (aHR, 0.9 [95% CI, 0.5-1.6]; Figure 4).

these estimates were attenuated but remained statistically significant for both outcomes, with the sole exception of persons diagnosed with schizophrenia, who did not have a higher risk of being subjected to violence when compared with their siblings without psychiatric disorders (aHR, 0.9 [95% CI, 0.5-1.6]; Figure 4). Figure 4. Risk of Subjection to Violence and Perpetration of Violence Among Individuals Diagnosed With Specific Psychiatric Disorders Compared With Siblings Without Psychiatric Disorders The adjusted hazard ratios refer to within-family estimates comparing differentially exposed siblings and adjusted for sex, birth year, birth order, and the individual’s history of subjection to and perpetration of violence. The estimates were further jointly adjusted for all of the psychiatric disorders and substance use disorders. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families.

order, and the individual’s history of subjection to and perpetration of violence. The estimates were further jointly adjusted for all of the psychiatric disorders and substance use disorders. Because the comparisons were made within families, there was no need to adjust for factors that were constant within families. Discussion In this nationwide study of 250 419 individuals born between 1973 and 1993 in Sweden, we examined the associations between psychiatric disorders and the later risk of subjection to violence and perpetration of violence. The patients were matched by age and sex to a general population control group and to their full biological siblings without psychiatric disorders. To our knowledge, this is the first study to have examined these associations using a sibling comparison approach, which enabled us to account for important shared unmeasured familial (eg, genetic and environmental) confounders. Our study had 4 principal findings. First, we estimated that the 10-year cumulative incidence rate of being subjected to violence was less than 7% in persons diagnosed with any psychiatric disorder. This estimate is therefore considerably smaller in magnitude, even when compared with the previous annual rates for subjection to violence reported in studies from Sweden, the Netherlands, the United Kingdom, and the United States, which typically range between 20% and 60%. This discrepancy is expected, as the earlier research relied on broad and self-reported measures of subjection to violence and were based on selected samples.

or subjection to violence reported in studies from Sweden, the Netherlands, the United Kingdom, and the United States, which typically range between 20% and 60%. This discrepancy is expected, as the earlier research relied on broad and self-reported measures of subjection to violence and were based on selected samples. Second, associations between psychiatric morbidity and a later risk of subjection to violence were considerably attenuated once we accounted for the individual’s history of subjection to and perpetration of violence as well as for unmeasured familial confounding by comparing patients with psychiatric conditions with their siblings without psychiatric disorders. The estimated risk increase of subjection to violence among people with psychiatric diagnoses was reduced from a factor of 6.5 to 3.4. These findings suggest that the estimates reported in 2 previous population studies, which examined alternative subjection to violence outcomes (eg, police-reported events and homicidal deaths), may have been substantially overestimated because of the lack of adjustment for these factors.

d from a factor of 6.5 to 3.4. These findings suggest that the estimates reported in 2 previous population studies, which examined alternative subjection to violence outcomes (eg, police-reported events and homicidal deaths), may have been substantially overestimated because of the lack of adjustment for these factors. Third, we found that the risks of subjection to and perpetration of violence varied across specific psychiatric disorders and were highest in persons with substance use disorders. In contrast, after adjusting for comorbid substance use disorders and personality disorders, we observed that persons diagnosed with schizophrenia were no more likely than their siblings without psychiatric disorders to be subjected to violence. One explanation for this finding is that patients with schizophrenia who do not have the comorbid conditions are more socially isolated and therefore less likely to be in environments where the risk of subjection to violence is increased.

ly than their siblings without psychiatric disorders to be subjected to violence. One explanation for this finding is that patients with schizophrenia who do not have the comorbid conditions are more socially isolated and therefore less likely to be in environments where the risk of subjection to violence is increased. Fourth, consistent with the literature, we found overlap between the risk of subjection to violence and perpetration of violence in individuals with psychiatric disorders. Although direct comparisons of the outcome measures require cautious interpretation, we note that this overlap may be important because it may offer etiologic and treatment targets. By separately considering each outcome, the dynamic interplay between them is overlooked. To take one example, subjection to violence is a trigger for subsequent perpetration of violence in patients diagnosed with psychotic disorders and individuals without psychiatric disorders.

offer etiologic and treatment targets. By separately considering each outcome, the dynamic interplay between them is overlooked. To take one example, subjection to violence is a trigger for subsequent perpetration of violence in patients diagnosed with psychotic disorders and individuals without psychiatric disorders. Our findings are largely consistent with those of the MacArthur risk assessment study, which found elevated postdischarge rates of violence among patients with psychiatric illnesses and comorbid substance use disorders as well as among individuals with early experiences of physical abuse and violence perpetration. Our findings diverge with regard to the association between certain psychiatric disorders and violence perpetration; we found higher rates of perpetration among persons with schizophrenia than depression. These observed differences could potentially be associated with contextual differences between Sweden and the United States, but we note that our study had sufficient statistical power to estimate differences between the conditions with a high degree of precision (250 419 vs 951 patients). Our study also benefited from almost no selection bias, unlike clinical studies in which the nonconsenting patients may have had different background risks. In the MacArthur study, for example, 44% of nonconsenting patients had schizophrenia diagnoses and were more likely than consenting patients to have histories of violence.

y also benefited from almost no selection bias, unlike clinical studies in which the nonconsenting patients may have had different background risks. In the MacArthur study, for example, 44% of nonconsenting patients had schizophrenia diagnoses and were more likely than consenting patients to have histories of violence. Strengths and Limitations Our study has several strengths. The use of Swedish national registers allowed us to study more than 250 000 patients diagnosed with psychiatric disorders, each individually matched with 10 people in the general population, while keeping selection bias to a minimum, as more than 95% of the overall sample was retained. We defined subjection to violence, using objective and validated measures of assault, as being events that either required hospital care (including specialist outpatient visits but excluding primary care visits) or resulted in death, in a country with universal health care. By adopting the sibling comparison approach, we were able to control for unmeasured familial confounding for the first time, to our knowledge.

ents that either required hospital care (including specialist outpatient visits but excluding primary care visits) or resulted in death, in a country with universal health care. By adopting the sibling comparison approach, we were able to control for unmeasured familial confounding for the first time, to our knowledge. However, the study had important limitations. First, we did not include incidents of less severe subjection to violence that did not result in hospitalization or death, which suggests that the reported absolute risk estimates should be interpreted as capturing the most severe violence subjection events. At the same time, our measures had the advantage of focusing on patients for whom implications for clinical services existed and interventions were potentially available (as these patients would have had service contacts). The extent to which the magnitude of the examined associations differs between severity levels remains relatively unknown and needs to be addressed in future studies. We would, however, expect that the inclusion of violence subjection events with a lower severity level would attenuate the reported relative risk estimates. This expectation is consistent with our sensitivity analysis, which observed a dose-response association between psychiatric morbidity and the severity of violence subjection events, a finding also reported in previous studies. Furthermore, it has been reported that approaches to measuring less severe violence subjection events (eg, self-reports and police reports) are associated with substantial measurement error, particularly when studying adolescents and individuals with elevated levels of psychiatric symptoms. Notably, the combination of using self-reports and sibling comparisons would further inflate the measurement error, causing an artificial bias of the estimates toward the null.

ciated with substantial measurement error, particularly when studying adolescents and individuals with elevated levels of psychiatric symptoms. Notably, the combination of using self-reports and sibling comparisons would further inflate the measurement error, causing an artificial bias of the estimates toward the null. Second, although sibling comparisons offer a powerful approach that accounts for genetic confounding, they account for approximately half of the genetic influences. Given the large reductions of the estimates in the sibling comparison models, we have likely overestimated the true associations.

ciated with substantial measurement error, particularly when studying adolescents and individuals with elevated levels of psychiatric symptoms. Notably, the combination of using self-reports and sibling comparisons would further inflate the measurement error, causing an artificial bias of the estimates toward the null. Second, although sibling comparisons offer a powerful approach that accounts for genetic confounding, they account for approximately half of the genetic influences. Given the large reductions of the estimates in the sibling comparison models, we have likely overestimated the true associations. Third, nationwide registries lack sufficient detail to fully ascertain the timing of our measures. Future studies may benefit from combining developmental life-course approaches with quasi-experimental designs to assess the relative importance of timing effects (eg, changes to diagnoses over time) and etiologic mechanisms (eg, mediation and moderation effects). Fourth, we had an mean of 7.3 years of postdischarge data available per participant, which captured a limited portion of their lives. Although our data were similar in magnitude to those of related Scandinavian population-based studies that had a maximum of 8 to 13 years of follow-up data, a need exists for studies with longer follow-up data to improve understanding of the long-term developmental trajectories of violent outcomes in people with psychiatric disorders. Fifth, although we used similar definitions for our outcome measures, they were derived from different data sources, which implies that comparisons between them should be interpreted with caution. Subjection to violence measures potentially represent a higher threshold because they require individuals to access health care services, although this assertion requires more empirical evidence to test. We note that the correlation between the outcomes in our patient sample (r = 0.43 [95% CI, 0.41-0.45]) replicated that of the MacArthur study (r = 0.44 [95% CI, 0.32-0.56]). Their data have been widely used to examine the co-occurrence between the outcomes despite heterogeneous definitions and data collection strategies.

note that the correlation between the outcomes in our patient sample (r = 0.43 [95% CI, 0.41-0.45]) replicated that of the MacArthur study (r = 0.44 [95% CI, 0.32-0.56]). Their data have been widely used to examine the co-occurrence between the outcomes despite heterogeneous definitions and data collection strategies. Sixth, the generalizability of our findings is unclear. Internationally comparable surveys of individuals subjected to violent crime reported that the annual rate of subjection to violence in Sweden (3.5%) was comparable with the global average (3.1%). Furthermore, a 2010 systematic review did not observe any clear differences in the rates of psychiatric disorders across Western European countries. However, associations between psychiatric disorders and violence outcomes may vary in other contexts, particularly in countries with different base rates of violence. Future studies should therefore test for this variance by using large-scale population-based data with adjustments for unmeasured familial confounders and early experiences of violence.

ychiatric disorders and violence outcomes may vary in other contexts, particularly in countries with different base rates of violence. Future studies should therefore test for this variance by using large-scale population-based data with adjustments for unmeasured familial confounders and early experiences of violence. Conclusions In this large longitudinal cohort study, we found that individuals diagnosed with psychiatric disorders in Sweden were more likely than 2 comparison groups without psychiatric disorders—siblings and individuals of similar age and gender in the general population—to be subjected to violence and to perpetrate violence against others. We generally found the magnitude of the associations to be similar across both outcomes, indicating a 3- to 4-fold elevated risk when the patients were compared with siblings who did not have psychiatric disorders. In addition, we found that having a diagnosis of schizophrenia was not associated with subsequent subjection to violence after we accounted for comorbid substance use and personality disorders. In contrast, we found that the same condition was the strongest risk factor for the perpetration of violence. Our findings underscore the need to address comorbid substance use and personality disorders to develop scalable approaches that assess and manage the risk of subjection to and perpetration of violence in people with psychiatric disorders. Supplement. eMethods. Population Sample, Definitions of Measured Confounders, Validation of Diagnoses, Complementary Sensitivity Analyses, and Moderation Effects

Conclusions In this large longitudinal cohort study, we found that individuals diagnosed with psychiatric disorders in Sweden were more likely than 2 comparison groups without psychiatric disorders—siblings and individuals of similar age and gender in the general population—to be subjected to violence and to perpetrate violence against others. We generally found the magnitude of the associations to be similar across both outcomes, indicating a 3- to 4-fold elevated risk when the patients were compared with siblings who did not have psychiatric disorders. In addition, we found that having a diagnosis of schizophrenia was not associated with subsequent subjection to violence after we accounted for comorbid substance use and personality disorders. In contrast, we found that the same condition was the strongest risk factor for the perpetration of violence. Our findings underscore the need to address comorbid substance use and personality disorders to develop scalable approaches that assess and manage the risk of subjection to and perpetration of violence in people with psychiatric disorders. Supplement. eMethods. Population Sample, Definitions of Measured Confounders, Validation of Diagnoses, Complementary Sensitivity Analyses, and Moderation Effects eTable 1. STROBE Statement eTable 2. ICD Codes eFigure 1. Sex and Age-Adjusted Cumulative Incidence Rates of Subjection to Violence Across the First 10 Years After the Onset of Any Psychiatric Disorder in Patients Compared With Unaffected Siblings and General Population Control Groups

Supplement. eMethods. Population Sample, Definitions of Measured Confounders, Validation of Diagnoses, Complementary Sensitivity Analyses, and Moderation Effects eTable 1. STROBE Statement eTable 2. ICD Codes eFigure 1. Sex and Age-Adjusted Cumulative Incidence Rates of Subjection to Violence Across the First 10 Years After the Onset of Any Psychiatric Disorder in Patients Compared With Unaffected Siblings and General Population Control Groups eFigure 2. Sex and Age-Adjusted Cumulative Incidence Rates of Perpetration of Violence Across the First 10 Years After the Onset of Any Psychiatric Disorder in Patients Compared With Unaffected Siblings and General Population Control Groups eFigure 3. Sensitivity Tests: Alternative Exposure and Outcome Definitions and Matching Criteria eFigure 4. Sensitivity Test: Associations Between Specific Psychiatric Disorders and Subsequent Risks of Subjection to Violence and Perpetration of Violence With and Without Adjustments for All Other Psychiatric Disorders eReferences. Click here for additional data file.

Introduction Dementia is one of the most feared medical conditions worldwide; it represents a significant global challenge to health and social care.1,2 Recent evidence suggests that dementia rates have decreased in the last few decades in the United Kingdom and other parts of Western Europe.3,4,5 Similarly, in the United States, the Framingham Heart Study has shown that age-specific incidence rates of dementia have decreased by almost 20% within the last few decades, and the greatest declines were apparent in individuals with higher educational attainment relative to more basic educational attainment.6

pe.3,4,5 Similarly, in the United States, the Framingham Heart Study has shown that age-specific incidence rates of dementia have decreased by almost 20% within the last few decades, and the greatest declines were apparent in individuals with higher educational attainment relative to more basic educational attainment.6 Education may serve different roles in the development of dementia: it is a proxy for early-life experiences and (parental) socioeconomic status (SES); it is related to future employment prospects, income, and wealth; it determines occupational exposures and characteristics of adult life (eg, job complexity, work stress, environmental exposures); and it provides lifelong skills for optimal mental abilities and mastery. Education is also thought to be a marker of cognitive reserve, which appears to be protective against cognitive impairment and dementia risk, offering an increased neural network and compensatory mechanisms throughout the life course, even when individuals are facing neuronal death.7 Recent systematic reviews have highlighted that low educational level was associated with a higher risk of dementia incidence8 as well as with greater risk of dementia-related death.9 Some of this evidence highlights that the role of education varies according to period and sociocultural context. The variation in country-specific regulations on compulsory schooling and variations in measurement could account for the differences reported in the literature.

ith greater risk of dementia-related death.9 Some of this evidence highlights that the role of education varies according to period and sociocultural context. The variation in country-specific regulations on compulsory schooling and variations in measurement could account for the differences reported in the literature. Moreover, given that education is typically completed many decades before dementia onset, other individual and area-based components of SES, such as wealth, income, and area deprivation, may provide a more accurate indication of current socioeconomic resources. Also, at older ages, accumulated wealth represents a more robust measure of socioeconomic resources than income or occupation alone.10,11 There are relatively few studies to date that have used socioeconomic indicators other than education to investigate dementia risk. A recent analysis of the Health and Retirement Study compared various SES markers, including parental education (an early-life indicator) and education and income (adult and late-life indicators) associated with late-life memory performance and decline. These findings indicated that income was most strongly associated with decline, although education was the most influential determinant of baseline memory.12

g parental education (an early-life indicator) and education and income (adult and late-life indicators) associated with late-life memory performance and decline. These findings indicated that income was most strongly associated with decline, although education was the most influential determinant of baseline memory.12 Another aspect of socioeconomic position involves neighborhood characteristics and the area of deprivation level, which combines information from multiple domains such as income, employment, education, skills, training, health, disability, crime, and barriers to housing into a single measure. Previous results from the English Longitudinal Study of Ageing (ELSA) showed that the index of multiple deprivation (IMD), the official measure of deprivation in England, was associated with cognitive performance in older age independently of education and SES. These findings indicated that older women had lower cognitive scores if they lived in an area classified in the bottom 20% of IMD when compared with those in the top (least deprived) quintile.13 In contrast, Meyer et al14 showed that neighborhood SES had limited effects on executive function, independent of personal characteristics such as education and ethnicity. They also showed that individuals with dementia living in neighborhoods with higher SES experienced faster rates of decline before further statistical adjustment for education and ethnicity.14 These findings are consistent with the cognitive reserve hypothesis, which acknowledges a rapid cognitive deterioration for people with higher education once the pathological process associated with dementia has been initiated.7 However, findings from the Seoul Dementia Management Project15 showed there were no additive or synergistic effects between individual-level and district-level of SES, highlighting that the individual level contributed more to the development of cognitive impairment than the district-level SES.

tia has been initiated.7 However, findings from the Seoul Dementia Management Project15 showed there were no additive or synergistic effects between individual-level and district-level of SES, highlighting that the individual level contributed more to the development of cognitive impairment than the district-level SES. We aimed to describe dementia incidence in a nationally representative cohort of British older adults and to investigate the association with different socioeconomic markers, both via the individual characteristics (education and wealth) and group-level characteristics (IMD). A second objective was to examine the role of socioeconomic markers between 2 independent age cohorts (those born from 1902 to 1925 and from 1926 to 1943). Methods Data The English Longitudinal Study of Ageing (ELSA) is a large, multidisciplinary study representative of the English population both in terms of socioeconomic profile and geographic region.16 There have been 7 waves of data collection over a follow-up period of up to 12 years, providing detailed information on health, well-being, and socioeconomic circumstances. We used all the available data spanning 12 years across wave 1 (2002-2003) to wave 7 (2014-2015). Refreshment samples were recruited at waves 3, 4, 6, and 7. For the current analyses, we included only participants aged 65 years and older who were free of dementia at their baseline assessment at either wave 1 or through the refreshment sample of wave 4 (Figure 1 for sample selection).

02-2003) to wave 7 (2014-2015). Refreshment samples were recruited at waves 3, 4, 6, and 7. For the current analyses, we included only participants aged 65 years and older who were free of dementia at their baseline assessment at either wave 1 or through the refreshment sample of wave 4 (Figure 1 for sample selection). Figure 1. Flowchart of the Individuals Included in Analyses Numbers of excluded persons are nonmutually exclusive. ELSA indicates the English Longitudinal Study of Ageing; IMD, index of multiple deprivations. Ethical approval for each one of the ELSA waves was granted by the National Research Ethics Service (London Multicentre Research Ethics Committee). All participants provided informed consent. Study Variables Dementia Ascertainment Dementia occurrence was determined at each wave using an algorithm based on a combination of self-reported or informant-reported physician diagnosis of dementia or Alzheimer disease or a score above the threshold of 3.38 on the 16-question Informant Questionnaire on Cognitive Decline in the Elderly.17 This questionnaire is administered to an informant (eg, a family member or a caregiver), who can evaluate the changes in the everyday cognitive function. Each item is scored from 1 (much improved) to 5 (much worse). The validity of this scale was previously examined,18 and the threshold used has both high specificity (0.84) and sensitivity (0.82).19

red to an informant (eg, a family member or a caregiver), who can evaluate the changes in the everyday cognitive function. Each item is scored from 1 (much improved) to 5 (much worse). The validity of this scale was previously examined,18 and the threshold used has both high specificity (0.84) and sensitivity (0.82).19 Socioeconomic Indicators We measured SES at baseline, including individual characteristics (education and wealth) and area-based characteristics (IMD). Educational attainment was classified into 4 categories: (1) having a university degree or higher; (2) having completed A-levels or the equivalent, which is comparable with high school graduation; (3) having completed education below the A-level; and (4) lacking formal qualifications. Wealth was calculated by summing wealth from property, possessions, housing, investments, savings, artwork, and jewelry, and net of debt16; this was divided into quintiles. The index of multiple deprivation (IMD) is a composite measure which combines multiple area-level SES indicators into a single deprivation score.20 We used the 2004 IMD for England (in which 1 was least deprived and 5 was most deprived). The highest levels of wealth, education, and IMD were used as the reference group. Covariates Based on previous findings,21 we considered baseline age, sex, marital status (married vs unmarried or widowed), and baseline health (eg, history of stroke, coronary heart disease, hypertension, and diabetes mellitus as potential confounders). Being male, married, and having no health conditions were used as the reference groups.

evious findings,21 we considered baseline age, sex, marital status (married vs unmarried or widowed), and baseline health (eg, history of stroke, coronary heart disease, hypertension, and diabetes mellitus as potential confounders). Being male, married, and having no health conditions were used as the reference groups. Age Cohorts To investigate the change in incidence rates over the last decade, we derived 2 groups: age cohort I (who were born between 1902-1925) and age cohort II (who were born between 1926-1943). This derivation was generated using a median split of all birth years (Figure 1).

evious findings,21 we considered baseline age, sex, marital status (married vs unmarried or widowed), and baseline health (eg, history of stroke, coronary heart disease, hypertension, and diabetes mellitus as potential confounders). Being male, married, and having no health conditions were used as the reference groups. Age Cohorts To investigate the change in incidence rates over the last decade, we derived 2 groups: age cohort I (who were born between 1902-1925) and age cohort II (who were born between 1926-1943). This derivation was generated using a median split of all birth years (Figure 1). Statistical Analyses Incidence rates of dementia were computed by age and sex per 1000 person-years. We performed χ2 tests to ascertain if there were significant differences between SES groups. To summarize the relationship between SES characteristics and dementia incidence, Cox proportional hazards models with age as the underlying time variable were used to calculate hazard ratios (HRs) and accompanying 95% CIs.22 We present the results from 4 models: model 1 included unadjusted HRs; model 2 included sex and marital status; model 3 included model 2 with further adjustment for baseline health indicators (stroke, hypertension, diabetes, and cardiovascular disease), and model 4 included model 3 and further adjusted for the additional socioeconomic indicators. We used a forward stepwise approach and the Akaike Information Criterion to select the model of best fit. Given that the original IMD quintile classification was slightly underpowered, we conducted a sensitivity analysis with the IMD regrouped into a binary variable (with quintile 1 [Q1] set to 1 and Q2-Q5 set to 2).

. We used a forward stepwise approach and the Akaike Information Criterion to select the model of best fit. Given that the original IMD quintile classification was slightly underpowered, we conducted a sensitivity analysis with the IMD regrouped into a binary variable (with quintile 1 [Q1] set to 1 and Q2-Q5 set to 2). The survival time was calculated using participants’ baseline age at study entry until the age they were found to be experiencing dementia, the point of their death, or the end of the study period (the last wave before dropout, or wave 7, which ran in 2014-2015). The Schoenfeld residual test was used to test the proportional hazards assumption of the models.23 For individuals who did not report an exact diagnosis date or for those whose dementia was ascertained with Informant Questionnaire on Cognitive Decline in the Elderly, we considered the midpoint between the wave where dementia was first ascertained and the previous wave where it was not. Mortality data were used for participants who had provided written consent for linkage to official records from the National Health Service central register; the records available the time of these analyses continued until February 2013. All analyses were weighted using the baseline cross-sectional weights derived in ELSA to ensure the sample is representative of the English population.24

ent for linkage to official records from the National Health Service central register; the records available the time of these analyses continued until February 2013. All analyses were weighted using the baseline cross-sectional weights derived in ELSA to ensure the sample is representative of the English population.24 Given that death is often considered a competing risk for dementia incidence, we conducted supplementary analyses using a modification of the Fine and Gray Subdistribution Hazards model25 to account for the competing risk of death, as described elsewhere26 (eFigure 1 in the Supplement). All analyses were conducted in Stata SE, Version 14 (StataCorp). Statistical significance was considered to be at or below the .05 level. Additional details are noted in the eAppendix in the Supplement.

n Hazards model25 to account for the competing risk of death, as described elsewhere26 (eFigure 1 in the Supplement). All analyses were conducted in Stata SE, Version 14 (StataCorp). Statistical significance was considered to be at or below the .05 level. Additional details are noted in the eAppendix in the Supplement. Results The sample included in these analyses was composed of 6220 individuals, accounting for 43 218 person-years (median follow-up duration, 7 years; range, 1-12 years). Of these, 463 (7.4%) were classified with dementia during the surveillance period, and 1971 (31.7%) died. The baseline median age was 73.2 years (interquartile range, 68.1-78.3 years), while the median age at the time of dementia ascertainment was 82.7 (interquartile range, 78.2-87.8 years). The sample included 6220 people, of whom 3410 (54.8%) were female and 2810 (45.8%) male, 3682 (59.2%) married, and 3288 (52.5%) without formal educational qualifications. Only 1049 of 6220 participants (16.9%) attended university. More men were educated to university degree level than women, while more women had no formal educational qualifications (χ23, 338.28; P ≤ .001). The baseline median wealth for the overall sample was £15 100 (approximately $21,470; interquartile range [IQR], £2700-£62 546 [$3839-$88 935.30]); for the lowest quintile, the median wealth as £120 (approximately $170.63; IQR, £0-£700 [$0-$995.34]), increasing to £180 000 ($255 936.94; IQR, £117 000-£309 100 [$166 375.78-$439 544.91]) in the highest quintile. Except for stroke, which showed no clear SES gradient, all other health conditions (cardiovascular disease, diabetes, and hypertension) were inversely associated with each one of the SES markers (results presented in eTable 1 of the Supplement).

IQR, £117 000-£309 100 [$166 375.78-$439 544.91]) in the highest quintile. Except for stroke, which showed no clear SES gradient, all other health conditions (cardiovascular disease, diabetes, and hypertension) were inversely associated with each one of the SES markers (results presented in eTable 1 of the Supplement). Age-adjusted and sex-adjusted incidence rates for the full ELSA sample and each specific age cohort are presented in Table 1 and Figure 2. The overall incidence rate (IR) was 11.32 per 1000 person-years (95% CI, 10.34-12.41 per 1000 person-years). As anticipated, there was a significant increase in dementia IRs with age from an incidence of 4.38 (95% CI, 3.49-5.57) in people aged 65 to 69 years to 24.69 (95% CI, 21.20-28.91) for those 80 years or older. The comparison between the 2 distinct age-periods cohorts shows a 30% reduction in the IRs of dementia for the overlapping age group of 75 to 79 years who were born between 1902 and 1925 (IR, 20.29; 95% CI, 16.45-25.28) and those born later between 1926 and 1943 (IR, 13.59; 95% CI, 10.33-18.20) (Table 1). There were no significant sex differences in the IRs of dementia (eTable 1 in the Supplement). Table 1. Dementia Incidence Rates Per 1000 Person-Years by Age Cohort Characteristic Total Cohort (n = 6220) Age Cohort I (n = 1808) Age Cohort II (n = 4412) No. (Cases of Dementia/ Censored) Incidence Rate (95% CI) No. (Cases of Dementia/ Censored) Incidence Rate (95% CI) No.

Age-adjusted and sex-adjusted incidence rates for the full ELSA sample and each specific age cohort are presented in Table 1 and Figure 2. The overall incidence rate (IR) was 11.32 per 1000 person-years (95% CI, 10.34-12.41 per 1000 person-years). As anticipated, there was a significant increase in dementia IRs with age from an incidence of 4.38 (95% CI, 3.49-5.57) in people aged 65 to 69 years to 24.69 (95% CI, 21.20-28.91) for those 80 years or older. The comparison between the 2 distinct age-periods cohorts shows a 30% reduction in the IRs of dementia for the overlapping age group of 75 to 79 years who were born between 1902 and 1925 (IR, 20.29; 95% CI, 16.45-25.28) and those born later between 1926 and 1943 (IR, 13.59; 95% CI, 10.33-18.20) (Table 1). There were no significant sex differences in the IRs of dementia (eTable 1 in the Supplement). Table 1. Dementia Incidence Rates Per 1000 Person-Years by Age Cohort Characteristic Total Cohort (n = 6220) Age Cohort I (n = 1808) Age Cohort II (n = 4412) No. (Cases of Dementia/ Censored) Incidence Rate (95% CI) No. (Cases of Dementia/ Censored) Incidence Rate (95% CI) No. (Cases of Dementia/ Censored) Incidence Rate (95% CI) Total 463/5757 11.32 (10.34-12.41) 239/1569 22.99 (20.31-26.11) 224/4188 7.06 (6.29-8.07) Age group,y 65-69 71/2008 4.38 (3.49-5.57) NA NA 71/2208 4.38 (3.49-5.57) 70-74 105/1705 8.30 (6.88-10.08) NA NA 105/1705 8.30 (6.88-10.09) 75-79 127/963 17.14 (14.49-20.41) 79/518 20.29 (16.45-25.28) 48/475 13.59 (10.33-18.20) ≥80 160/1051 24.69 (21.20-28.91) 160/1501 24.69 (21.20-28.91) NA NA Sex Male 187/2623 10.27 (8.92-11.89) 82/655 20.53 (16.64-25.60) 105/1968 7.24 (6.01-8.81) Female 276/3134 12.09 (10.76-13.63) 157/914 24.39 (20.95-28.55) 119/2220 6.92 (5.80-8.32) Marital status Married 254/3428 9.94 (8.81-11.26) 96/640 21.77 (17.93-26.68) 158/2788 7.35 (6.31-8.61) Single/divorced 209/2329 13.35 (11.68-15.33) 143/929 23.80 (20.36-28.07) 66/1400 6.48 (5.12-8.34) Education Higher education 73/976 9.85 (7.86-12.50) 37/178 26.22 (19.22-36.58) 36/798 5.72 (4.17-8.08) A-level 103/1444 9.17 (7.60-11.18) 48/325 19.78 (15.14-26.34) 55/1119 6.06 (4.69-7.99) >A-level 20/316 9.71 (6.31-15.70) 10/82 23.11 (12.56-46.79) 10/234 5.70 (3.14-11.46) No qualification 267/3021 13.08 (11.62-14.77) 144/984 23.46 (20.02-27.67) 123/2037 8.32 (7.00-9.97) Wealtha Q1 (Highest) 67/1062 7.92 (6.26-10.16) 33/213 19.28 (13.87-27.51) 34/848 4.88 (3.52-6.98) Q2 82/1096 10.11 (8.16-12.67) 36/230 22.16 (16.18-31.11) 46/866 6.61 (4.99-8.95) Q3 91/1154 11.03 (9.02-13.64) 49/289 23.33 (17.84-31.03) 42/865 6.66 (4.95-9.16) Q4 102/1139 12.54 (10.38-15.28) 44/322 21.48 (16.12-29.19) 58/817 9.19 (7.18-11.95) Q5 (Lowest) 121/1306 15.05 (12.62-18.10) 77/515 26.07 (21.02-32.70) 44/791 8.34 (6.27-11.35) Index of multiple deprivationb Q1 (Least deprived) 86/1291 8.62 (7.48-11.24) 49/333 19.27 (14.76-25.59) 37/958 4.82 (3.52-6.77) Q2 116/1221 12.47 (9.50-14.06) 57/325 25.38 (19.71-33.1)8 59/896 7.92 (6.18-10.30) Q3 97/1224 11.56 (10.20-15.10) 54/337 25.59 (19.73-33.71) 43/887 6.42 (4.80-8.76) Q4 90/1109 11.99 (9.21-13.88) 43/314 21.51 (16.20-29.14) 47/795 8.35 (6.34-11.23) Q5 (Most deprived) 74/913 12.64 (10.70-16.86) 36/260 23.30 (19.97-32.79) 38/652 8.70 (6.37-12.19) Stroke No 407/5170 10.87 (9.87-11.99) 213/1426 22.41 (19.66-25.65) 194/3744 6.68 (5.82-7.71) Y

7 25.59 (19.73-33.71) 43/887 6.42 (4.80-8.76) Q4 90/1109 11.99 (9.21-13.88) 43/314 21.51 (16.20-29.14) 47/795 8.35 (6.34-11.23) Q5 (Most deprived) 74/913 12.64 (10.70-16.86) 36/260 23.30 (19.97-32.79) 38/652 8.70 (6.37-12.19) Stroke No 407/5170 10.87 (9.87-11.99) 213/1426 22.41 (19.66-25.65) 194/3744 6.68 (5.82-7.71) Y es 56/587 16.47 (12.78-21.56) 26/143 29.47 (20.40-43.92) 30/444 11.53 (8.16-16.82) Hypertension No 240/3239 11.27 (9.93-12.85) 137/854 24.68 (20.97-29.23) 103/2385 5.64 (4.66-6.88) Yes 223/2518 12.44 (10.86-14.27) 102/715 20.16 (17.40-25.55) 121/1803 8.95 (7.53-10.73) Diabetes No 416/5215 11.12 (10.11-12.26) 223/1423 23.32 (20.51-26.60) 193/3792 6.64 (5.78-7.67) Yes 47/542 13.52 (10.28-18.13) 16/146 19.06 (11.95-32.22) 31/396 11.66 (8.34-16.80) Cardiovascular disease No 373/5081 10.17 (9.15-11.34) 193/1360 22.07 (19.24-25.43) 180/3721 6.65 (5.61-7.52) Yes 90/676 16.21 (13.67-19.36) 46/209 28.00 (21.11-37.82) 44/467 11.03 (8.28-15.01) Abbreviations: NA, not available; Q, quintile. a In wealth rankings, Q1 indicates highest wealth category; Q2, the second highest; Q3, the third highest; Q4, the fourth highest; and Q5, the lowest. b In the index of multiple deprivation, Q1 indicates least deprived; Q2, the second least deprived; Q3, the third least deprived; Q4, the fourth least deprived; and Q5, most deprived. Figure 2. Dementia Incidence Rates Per 1000 Person-Years in Men and Women Presented by Age-Groups in the English Longitudinal Study of Ageing Error bars indicate 95% CIs.

b In the index of multiple deprivation, Q1 indicates least deprived; Q2, the second least deprived; Q3, the third least deprived; Q4, the fourth least deprived; and Q5, most deprived. Figure 2. Dementia Incidence Rates Per 1000 Person-Years in Men and Women Presented by Age-Groups in the English Longitudinal Study of Ageing Error bars indicate 95% CIs. Individual and Area-Based Socioeconomic Markers The multivariable analyses are summarized in Table 2. Education was not significantly associated with dementia incidence, but wealth was a strong indicator. Per model 4, the hazards of developing dementia were higher for those in the lowest 2 quintiles of wealth (Q4: HR, 1.39; 95% CI, 1.00-1.95; and Q5: HR,  1.50; 95% CI, 1.05-2.13; P for trend = .04), compared with those in the highest quintile (Q1), independently of covariates, education, and area-level socioeconomic characteristics (Table 2 and Figure 3). Table 2. Hazard Ratios From Univariate and Multivariate Cox Regression Models by Age Cohort Characteristic Hazard Ratios (95% CI) per Model No.

Individual and Area-Based Socioeconomic Markers The multivariable analyses are summarized in Table 2. Education was not significantly associated with dementia incidence, but wealth was a strong indicator. Per model 4, the hazards of developing dementia were higher for those in the lowest 2 quintiles of wealth (Q4: HR, 1.39; 95% CI, 1.00-1.95; and Q5: HR,  1.50; 95% CI, 1.05-2.13; P for trend = .04), compared with those in the highest quintile (Q1), independently of covariates, education, and area-level socioeconomic characteristics (Table 2 and Figure 3). Table 2. Hazard Ratios From Univariate and Multivariate Cox Regression Models by Age Cohort Characteristic Hazard Ratios (95% CI) per Model No. (Cases/Censored) Person-Years Model 1a P Value for trend Model 2b P Value for trend Model 3c P Value for trend Model 4d P Value for trend ELSA Overall (N = 6220; 43 219 person-years) Education University degree 73/976 7974 1 [Reference] .27 1 [Reference] .23 1 [Reference] .32 1 [Reference] .99 A-level 103/1444 11 593 0.89 (0.67-1.21) 0.90 (0.66-1.23) 0.91 (0.67-1.23) 0.84 (0.62-1.14) <A-level 20/316 2270 0.87 (0.53-1.42) 0.86 (0.53-1.41) 0.83 (0.50-1.36) 0.74 (0.44-1.22) No qualification 267/3021 21 382 1.07 (0.82-1.39) 1.09 (0.83-1.43) 1.07 (0.81-1.40) 0.92 (0.70-1.23) Wealthe Q1 (Highest) 67/1062 8807 1 [Reference] .01 1 [Reference] .002 1 [Reference] .01 1 [Reference] .04 Q2 82/1096 8605 1.27 (0.92-1.76) 1.31 (0.94-1.80) 1.29 (0.94-1.79) 1.31 (0.94-1.82) Q3 91/1154 8670 1.29 (0.94-1.77) 1.33 (0.97-1.84) 1.29 (0.94-1.78) 1.28 (0.91-1.79) Q4 102/1139 8605 1.43 (1.05-1.95) 1.50 (1.09-2.05) 1.43 (1.05-1.96) 1.39 (1.00-1.95) Q5 (Lowest) 121/1306 8531 1.49 (1.10-2.01) 1.62 (1.19-2.21) 1.56 (1.14-2.13) 1.50 (1.05-2.13) Index of multiple deprivationf Q1 (Least deprived) 86/1291 10 235 1 [Reference] .04 1 [Reference] .02 1 [Reference] .04 1 [Reference] .35 Q2 116/1221 9734 1.44 (1.08-1.90) 1.45 (1.09-1.92) 1.47 (1.11-1.95) 1.41 (1.06-1.87) Q3 97/1224 9177 1.38 (1.00-1.79) 1.36 (1.02-1.82) 1.35 (1.01-1.81) 1.27 (0.94-1.72) Q4 90/1109 7827 1.39 (1.03-1.87) 1.42 (1.05-1.91) 1.37 (1.02-1.85) 1.25 (0.91-1.73) Q5 (Most deprived) 74/913 6246 1.45 (1.06-1.99) 1.51 (1.10-2.07) 1.47 (1.07-2.10) 1.28 (0.90-1.82) Age Cohort I (N = 1808; 10 484 person-years) Education University degree 37/178 1446 1 [Reference] .70 1 .66 1 [Reference] .64 1 [Reference] .30 A level 48/325 2388 0.74 (0.48-1.13) 0.72 (0.46-1.11) 0.74 (0.48-1.14) 0.70 (0.45-1.08) <A-level 10/82 494 0.85 (0.43-1.70) 0.89 (0.45-1.77) 0.87 (0.43-1.74 0.78 (0.39-1.55 No qualification 144/984 6156 0.83 (0.57-1.20) 0.81 (0.55-1.18) 0.81 (0.56-1.19) 0.72 (0.49-1.07) Wealthe Q1 (Highest) 33/213 1630 1 [Reference] .23 1 [Reference] .18 1 [Reference] .30 1 [Reference] .21 Q2 36/230 1683 1.14 (0.71-1.82) 1.15 (0.72-1.84) 1.18 (0.74-1.90) 1.19 (0.74-1.90) Q3 49/289 2056 1.21 (0.77-1.89) 1.22 (0.78-1.91) 1.25 (0.78-1.98) 1.25 (0.7

1.20) 0.81 (0.55-1.18) 0.81 (0.56-1.19) 0.72 (0.49-1.07) Wealthe Q1 (Highest) 33/213 1630 1 [Reference] .23 1 [Reference] .18 1 [Reference] .30 1 [Reference] .21 Q2 36/230 1683 1.14 (0.71-1.82) 1.15 (0.72-1.84) 1.18 (0.74-1.90) 1.19 (0.74-1.90) Q3 49/289 2056 1.21 (0.77-1.89) 1.22 (0.78-1.91) 1.25 (0.78-1.98) 1.25 (0.7 8-1.98) Q4 44/322 2057 1.12 (0.71-1.76) 1.13 (0.72-1.78) 1.12 (0.69-1.81) 1.12 (0.69-1.81) Q5 (Lowest) 77/515 3058 1.32 (0.87-1.98) 1.37 (0.90-2.07) 1.35 (0.85-2.14) 1.35 (0.85-2.14) Index of multiple deprivationf Q1 (Least deprived) 49/333 2452 1 [Reference] .55 1 [Reference] .55 1 [Reference] .60 1 [Reference] .94 Q2 57/325 2285 1.30 (0.88-1.91) 1.29 (0.89-1.91) 1.31 (0.89-1.93) 1.28 (0.86-1.90) Q3 54/337 2197 1.34 (0.91-1.97) 1.34 (0.91-1.97) 1.34 (0.91-1.98) 1.30 (0.86-1.96) Q4 43/314 1964 1.15 (0.76-1.73) 1.15 (0.76-1.73) 1.13 (0.75-1.70) 1.07 (0.68-1.69) Q5 (Most deprived) 36/260 1586 1.21 (0.78-1.86) 1.21 (0.78-1.86) 1.20 (0.78-1.85) 1.11 (0.68-1.81) Age Cohort II (N = 4412; 32 735 person-years) Education University degree 36/798 6527 1 [Reference] .03 1 [Reference] .02 1 [Reference] .04 1 [Reference] .002 A level 55/1119 9206 1.08 (0.71-1.65) 1.12 (0.73-1.74) 1.12 (0.73-1.73) 1.02 (0.65-1.59) <A-level 10/234 1776 0.87 (0.43-1.77) 0.83 (0.41-1.68 0.77 (0.37-1.59 0.68 (0.33-1.41 No qualification 123/2037 15 226 1.43 (1.01-2.04) 1.49 (1.01-2.19) 1.43 (0.97-2.11) 1.21 (0.81-1.79) Wealthe Q1 (Highest) 34/848 7178 1 [Reference] .01 1 [Reference] .001 1 [Reference] .05 1 [Reference] .05 Q2 46/866 6921 1.42 (0.90-2.22) 1.47 (0.93-2.31) 1.44 (0.92-2.27) 1.43 (0.89-2.29) Q3 42/865 6613 1.37 (0.86-2.17) 1.46 (0.92-2.31) 1.39 (0.88-2.21) 1.34 (0.81-2.19) Q4 58/817 6548 1.81 (1.18-2.77) 1.96 (1.28-3.01) 1.83 (1.19-2.82) 1.65 (1.02-2.70) Q5 (Lowest) 44/791 5475 1.73 (1.10-2.72) 2.02 (1.25-3.25) 1.82 (1.12-2.96) 1.68 (1.05-2.86) Index of multiple deprivationf Q1 (Least deprived) 37/958 7781 1 [Reference] .01 1 [Reference] .005 1 [Reference] .01 1 [Reference] .18 Q2 59/896 7448 1.66 (1.10-2.51) 1.67 (1.10-2.53) 1.73 (1.14-2.63) 1.62 (1.06-2.46) Q3 43/887 6978 1.37 (0.88-2.12) 1.39 (0.89-2.16) 1.39 (0.89-2.18) 1.27 (0.81-1.98) Q4 47/795 5860 1.77 (1.15-2.73) 1.82 (1.18-2.82) 1.80 (1.16-2.79) 1.55 (0.98-2.45) Q5 (Most deprived) 38/652 4658 1.87 (1.18-2.97) 1.98 (1.25-3.15) 1.89 (1.19-2.99) 1.50 (0.91-2.49) Abbreviations: Q, quintile; SES, socioeconomic indicators.

43/887 6978 1.37 (0.88-2.12) 1.39 (0.89-2.16) 1.39 (0.89-2.18) 1.27 (0.81-1.98) Q4 47/795 5860 1.77 (1.15-2.73) 1.82 (1.18-2.82) 1.80 (1.16-2.79) 1.55 (0.98-2.45) Q5 (Most deprived) 38/652 4658 1.87 (1.18-2.97) 1.98 (1.25-3.15) 1.89 (1.19-2.99) 1.50 (0.91-2.49) Abbreviations: Q, quintile; SES, socioeconomic indicators. a Model 1 used SES indicators analyzed individually, unadjusted. b Model 2 used 1 SES indicator at the time, adjusted for sex and marital status. c Model 3 used 1 SES indicator at the time, adjusted for sex, marital status, stroke, hypertension, diabetes, and cardiovascular disease. d Model 4 with all 3 SES indicators entered simultaneously, adjusted for sex, marital status, stroke, hypertension, diabetes, and cardiovascular disease. e In wealth rankings, Q1 indicates highest wealth category; Q2, the second highest; Q3, the third highest; Q4, the fourth highest; and Q5, the lowest. f In the index of multiple deprivation, Q1 indicates least deprived; Q2, the second least deprived; Q3, the third least deprived; Q4, the fourth least deprived; and Q5, most deprived. Figure 3. Smoothed Hazard Estimates by Age per 1000 Person-Years by Wealth Quintiles in the English Longitudinal Study of Ageing Wealth quintile 1 indicates the highest level of wealth; quintile 5, the lowest.

f In the index of multiple deprivation, Q1 indicates least deprived; Q2, the second least deprived; Q3, the third least deprived; Q4, the fourth least deprived; and Q5, most deprived. Figure 3. Smoothed Hazard Estimates by Age per 1000 Person-Years by Wealth Quintiles in the English Longitudinal Study of Ageing Wealth quintile 1 indicates the highest level of wealth; quintile 5, the lowest. Area-based characteristics measured with IMD were also associated with dementia incidence. In contrast with individuals in the least-deprived areas (IMD Q1), the remaining 4 quintiles showed an increase in the hazard risk of developing dementia in model 1 (Q2: HR, 1.44; 95% CI, 1.08-1.90; to Q5: HR, 1.45; 95% CI, 1.06-1.99; P for trend = .04). However, only the association with the second-highest quintile (Q2: HR, 1.41; 95% CI, 1.06-1.87) maintained statistical significance in the fully adjusted model, independent of the other individual markers of SES. Results from the first sensitivity analysis showed that those in the lowest 4 quintiles of IMD combined had increased risks of developing dementia (HR, 1.32; 95% CI, 1.03-1.69; model 4) compared with those living in the least deprived area (eTable 2 in the Supplement).

Area-based characteristics measured with IMD were also associated with dementia incidence. In contrast with individuals in the least-deprived areas (IMD Q1), the remaining 4 quintiles showed an increase in the hazard risk of developing dementia in model 1 (Q2: HR, 1.44; 95% CI, 1.08-1.90; to Q5: HR, 1.45; 95% CI, 1.06-1.99; P for trend = .04). However, only the association with the second-highest quintile (Q2: HR, 1.41; 95% CI, 1.06-1.87) maintained statistical significance in the fully adjusted model, independent of the other individual markers of SES. Results from the first sensitivity analysis showed that those in the lowest 4 quintiles of IMD combined had increased risks of developing dementia (HR, 1.32; 95% CI, 1.03-1.69; model 4) compared with those living in the least deprived area (eTable 2 in the Supplement). Individual and Area-Based Socioeconomic Markers Within Age Cohorts An investigation of age cohort showed that education was significantly associated with dementia for participants born between 1926 and 1943 (age cohort II), but not for those born earlier in the century (age cohort I). In age cohort II, there was a greater hazard risk of dementia for those with no education than those educated at university levels (HR, 1.43; 95% CI, 1.01-2.04; model 1). However, this association was no longer significant once health conditions had been entered, per model 3.

born earlier in the century (age cohort I). In age cohort II, there was a greater hazard risk of dementia for those with no education than those educated at university levels (HR, 1.43; 95% CI, 1.01-2.04; model 1). However, this association was no longer significant once health conditions had been entered, per model 3. Wealth also seemed to have a stronger association with dementia incidence within age cohort II, although this was not statistically significant. The association of IMD with subsequent dementia was comparable in age cohort II and the full sample, while differences between IMD quintiles were not present for age cohort I in models 1, 2, and 3, before adjusting for other SES markers. Our additional analyses considering the competing risk of death showed a similar pattern of decline in dementia incidence over time (eFigure 2 in the Supplement) and a stronger association between dementia incidence and all the SES markers including education, but with no age-cohort effects (eTable 3 in the Supplement).

Our additional analyses considering the competing risk of death showed a similar pattern of decline in dementia incidence over time (eFigure 2 in the Supplement) and a stronger association between dementia incidence and all the SES markers including education, but with no age-cohort effects (eTable 3 in the Supplement). Discussion In a representative sample of the English population aged 65 years and older, we found a positive association between lower wealth and dementia incidence that was independent of education, area-level deprivation, and covariates. This suggests a higher risk for individuals with fewer financial resources. The association was more consistent for participants born after 1926 compared with those born earlier in the 20th century. Additionally, there was evidence for reduced incidence among participants born more recently. However, the 2 age cohorts overlap only for the group aged 75 to 79 years. Differences between age cohorts in the incidence of early-onset vs later-onset dementias may also be present.

h those born earlier in the 20th century. Additionally, there was evidence for reduced incidence among participants born more recently. However, the 2 age cohorts overlap only for the group aged 75 to 79 years. Differences between age cohorts in the incidence of early-onset vs later-onset dementias may also be present. There are several possible explanations for the strong association of wealth with subsequent health outcomes. Wealth is an indicator of socioeconomic resources, and it could represent a gateway to more mentally stimulating environments independent of the level of educational attainment. Previous ELSA findings have shown that increased wealth facilitates greater digital literacy, which is in turn associated with a reduced risk of dementia.27 Furthermore, increased financial status could provide broader access to cultural resources and behaviors (eg, reading, theaters, social clubs) or increased social networks, which could ultimately contribute to higher cognitive reserve.7,28 The integrated psychosocial resource model proposed by Matthews and Gallo29 argues for the accumulation of psychosocial and physical protective factors. However, in our analyses, the relationship between wealth and dementia remained statistically significant even after controlling for health-related conditions associated with dementia.

ychosocial resource model proposed by Matthews and Gallo29 argues for the accumulation of psychosocial and physical protective factors. However, in our analyses, the relationship between wealth and dementia remained statistically significant even after controlling for health-related conditions associated with dementia. There is also evidence that persistent SES disadvantage is associated with impaired physiological functioning,30 increased risk of depression,31 vascular disease, and stroke.32 Other factors, such as reduced exercise, poor diet,33 and inflammatory vascular risk factors,34 may also play a part in the association between low SES (as defined by wealth) and increased dementia risk. Our data showed a differential SES distribution for the health conditions modeled as covariates in these analyses, except for stroke, which showed no clear SES gradient. Further work on the ELSA data could explore these mechanisms in more detail to be able to disentangle the mediating role of psychological, cardiovascular, and metabolic functions on the association between SES markers and dementia.

as covariates in these analyses, except for stroke, which showed no clear SES gradient. Further work on the ELSA data could explore these mechanisms in more detail to be able to disentangle the mediating role of psychological, cardiovascular, and metabolic functions on the association between SES markers and dementia. The lack of a contextual, area-based SES effect on dementia incidence is also notable. Previous ELSA findings have documented a link between neighborhood deprivation and cognitive functioning, independent of individual markers of SES, showing that individuals living in the most deprived area of England had significantly lower cognitive scores compared with those living in the most affluent regions.13 Our study found an inconsistent association between the area deprivation (IMD) and dementia incidence, with higher rates for individuals in the second quintile of IMD compared with the top quintile (who were least deprived). The reasons for this are not clear. Associations were observed for the lower IMD quintiles in first stages of covariates adjustment, but these were no longer significant when individual-level SES indicators were considered. This suggests that much of the effect of area deprivation is explained by the individual characteristics of the people living in those areas, rather than the features of the areas themselves.

irst stages of covariates adjustment, but these were no longer significant when individual-level SES indicators were considered. This suggests that much of the effect of area deprivation is explained by the individual characteristics of the people living in those areas, rather than the features of the areas themselves. In this cohort, education was not a robust predictor of dementia incidence. Given that this association was no longer significant after age and sex were taken into account, it is possible that this might be a specific cohort effect in the English population born and educated in the period surrounding the World War II. Support for this speculation comes from an extensive population cohort collaboration (the Epidemiological Clinicopathological Studies in Europe), which showed no apparent protective effect of education on the clinical presentation of dementia (eg, accumulation of pathology, pathological severity, and level of compensatory mechanisms for cognitive impairment).35 Their findings showed that individuals with higher education had heavier brains, suggesting greater cognitive reserve, but they were not necessarily able to compensate for the accumulation of vascular and neurodegenerative pathologies. However, the role of education might be sensitive to sociocultural context. Similar to our findings, other investigations from the Rotterdam Study,36 the Rochester Epidemiology Project,37 and the Baltimore Longitudinal Study of Aging38 reported a lack of association between dementia incidence and education.

athologies. However, the role of education might be sensitive to sociocultural context. Similar to our findings, other investigations from the Rotterdam Study,36 the Rochester Epidemiology Project,37 and the Baltimore Longitudinal Study of Aging38 reported a lack of association between dementia incidence and education. In contrast, findings from the Health and Retirement Study39 indicated that higher education was associated with a lower risk of dementia prevalence between 2000 and 2012, and in the Kungsholmen study,40 education remained significantly associated with dementia following adjustment for occupational class. Moreover, in the Canadian Study of Health and Aging,41 fewer years of education were associated with an increased risk of late-onset Alzheimer disease incidence, while subsequent results from a 10-year follow-up (1991-2001) within the same study showed that high complexity of work with people or things was associated with a reduced risk of most dementia types (Alzheimer and vascular dementia).42 These findings indicate a protective effect of the occupational demands on the brain achieved through a lifetime occupational exposure. It is therefore possible that individuals born before the World War II may not necessarily have been able to access higher education (because of military service, financial restrictions, and limited university place availability) but may have gained access to intellectually challenging jobs and growth opportunities after the war.

fore possible that individuals born before the World War II may not necessarily have been able to access higher education (because of military service, financial restrictions, and limited university place availability) but may have gained access to intellectually challenging jobs and growth opportunities after the war. Strengths To our knowledge, this is the first longitudinal study to examine multiple facets of SES characteristics at individual and group levels simultaneously in association with dementia incidence within an age-cohort context. Through the extensive monitoring of biennial interviews and a long-term follow-up, we were able to use an integrative approach to study the association between various socioeconomic factors and dementia incidence. Furthermore, we benefited from a more detailed assessment of wealth than what is available in most studies to date, because this measure was computed on the basis of accurate information on multiple individual components rather than broad categorization of assets.

n various socioeconomic factors and dementia incidence. Furthermore, we benefited from a more detailed assessment of wealth than what is available in most studies to date, because this measure was computed on the basis of accurate information on multiple individual components rather than broad categorization of assets. Limitations This study also has limitations. Given that the ascertainment of dementia diagnosis is still challenging in the UK health services and elsewhere, it is likely that the presented dementia IRs are underestimated. Other common issues such as nonresponse and subsequent attrition are familiar to most longitudinal surveys.43 Moreover, because of a relatively small sample of dementia cases, we did not explore the IRs of dementia by specific typology (eg, Alzheimer disease, vascular, mixed). Although ELSA is a demographically representative cohort, the race/ethnicity is 97% white16 and we were therefore unable to investigate the effects that race/ethnicity might have on the outcome of dementia. Furthermore, we did not investigate the difference in dementia incidence by geographical regions, given the high collinearity with IMD. Lastly, as in any observational study, we cannot exclude the risk of confounding by other factors. Avenues for future exploration include examining the mediating role of cardiovascular disease, lifestyle factors, medical care and other risk factors that could influence the association between SES and dementia.

ith IMD. Lastly, as in any observational study, we cannot exclude the risk of confounding by other factors. Avenues for future exploration include examining the mediating role of cardiovascular disease, lifestyle factors, medical care and other risk factors that could influence the association between SES and dementia. Conclusions In a nationally representative sample of English people 65 years and older, the hazard risk of dementia incidence was associated with socioeconomic indicators, notably wealth. Socioeconomic inequalities were more marked in individuals born in later years (from 1926 onwards) than in those born earlier (between 1900 and 1925). Public health strategies for dementia prevention should target socioeconomic gaps to reduce health disparities and protect those who are particularly disadvantaged in addition to addressing vascular risk factors such as hypertension, diabetes mellitus, smoking, and heart disease. Supplement. eAppendix. Supplementary materials description. eFigure 1. Flowchart of the individuals included in the competing risk hazard model analyses. eFigure 2. Stacked cumulative incidence plots for dementia and death as competing risks within each age cohort. eTable 1. Dementia incidence rates per 1,000 person-years and 95% confidence intervals (CI), according to baseline characteristics in the overall ELSA sample (N=6220; 463 dementia cases). eTable 2. Hazard Ratios (HR) and 95% confidence intervals (CI) from univariate and multivariate Cox regression models by age.

eFigure 2. Stacked cumulative incidence plots for dementia and death as competing risks within each age cohort. eTable 1. Dementia incidence rates per 1,000 person-years and 95% confidence intervals (CI), according to baseline characteristics in the overall ELSA sample (N=6220; 463 dementia cases). eTable 2. Hazard Ratios (HR) and 95% confidence intervals (CI) from univariate and multivariate Cox regression models by age. eTable 3. Hazard Ratios (HR) and 95% confidence intervals (CI) from univariate and multivariate Fine and Gray Subdistribution Hazards model adjusting for the competing risk of death. Click here for additional data file.