Antidepressant Use and Risk of Manic Episodes in Children and Adolescents With Unipolar Depression

Key Points Question Do pediatric patients diagnosed with unipolar depression who receive antidepressant treatment have an increased incidence of mania/hypomania compared with patients not treated with antidepressants? Findings In this cohort study including 43 677 children and adolescents diagnosed with unipolar depression, during a 12-week follow-up, patients treated with antidepressants did not have an increased incidence of mania/hypomania compared with patients who did not initiate antidepressant treatment. Hospitalizations, parental bipolar disorder, and use of antipsychotics and antiepileptics were the most important predictors of mania/hypomania by 12 weeks. Meaning This study found no evidence of treatment-emergent mania/hypomania in children and adolescents with unipolar depression.


Per-protocol analysis
To conduct per-protocol analysis where individuals were censored if they switch from on-treatment to off-treatment, or vice versa, we first calculated the estimated treatment periods for individuals who initiated antidepressant medication.A continuous treatment period with an antidepressant was defined based on the assumption that two dispensed prescriptions falling within 120 days of each other belong to the same treatment period.Timeframe of 120 days was selected because oral psychiatric medications are typically not dispensed for more than 90 days at a time in Sweden (so-called 90-day rule).We added 30 days to the 90 days to account for potential treatment non-adherence.Treatment periods were defined independently of the study follow-up.At the last or single dispensation in a treatment period, treatment end date was estimated by adding the study population median number of days between consecutive dispenses to the date of dispensation (40 days in our cohort).This definition is based on prior work [1][2].Further, we estimated treatment periods for the time-varying psychotropic treatment covariates in the same way as when defining treatment periods for antidepressant medications (i.e., using the 120-day assumption and adding the median number of days).
In per-protocol analyses, individuals initially assigned to the control group were censored on the date they initiated antidepressant medication.Individuals assigned to the treatment group were censored if their estimated treatment period ended before the end of follow-up (at 12/52 weeks).Data were split by week.For each week of follow-up, we assumed that individuals were randomly censored conditional on their baseline and time-varying covariates.The time-varying covariates were indicators for the use of benzodiazepines, antipsychotics, and any other (nonantidepressant) psychotropic drugs over the follow-up (see eTable 1 for variable definitions).All time-varying covariates were updated weekly.We estimated time-varying treatment adherence weights, which accounted for both baseline (covariates specified in the manuscript) and time-varying confounders.We weighted each individual at each week of follow-up by the product of the baseline IPT weights (the same weight also used in the ITT analysis) and the time-varying censoring weights (specified earlier in this section).All weights were stabilized, and truncated at the 99 th percentile.We used a discrete-time hazards model (a pooled logistic model) to estimate cumulative incidences and risk ratios for 12 and 52 weeks of follow-up, and calculated 95% confidence intervals for the estimates using non-parametric bootstraps.
We conducted two additional robustness checks for the per-protocol analysis.In the first one, we defined treatment periods assuming that two dispensed prescriptions falling within 60 days of each other belonged to the same treatment period.Treatment end date was estimated in the same manner as in the main approach.In the second robustness check, we estimated treatment periods using the total quantity of medication purchased instead of relying on elapsed time between dispensations.The Prescribed Drug Register includes the size of tablet package (i.e.number of tablets per package) and the number of packages an individual purchased.The duration of treatment period was calculated from the total number of tablets, assuming that the daily dosage was consumed once per day (i.e., 1 pill per day).

Prediction of mania: coding of predictor variables
We investigated which patient characteristics were associated with mania/hypomania in the 12-week of follow-up.For this analysis, we removed individuals with "unknown" as the response category in any of the variables (n=2,197).Some of the predictor variables did not have sufficient number of observations to estimate a separate coefficient.We therefore recoded or combined a subset of the predictor variables to increase statistical power.Family education was recoded as 0=primary school or high school, 1=university level; family income percentile as 0=80 th percentile or below, 1=over 80 th ; number of prior hospitalizations as 0=no hospitalizations, 1=1 or more hospitalizations.Diagnosis for alcohol or drug use disorder were combined into one variable, where having either one of these diagnoses or having both diagnoses were coded as 1, and having neither was coded was 0. In the same manner, we combined 1) conduct disorder and personality disorder diagnosis, 2) poisoning by alcohol and poisoning

Sensitivity analysis: cloning-censoring-weighting approach
We conducted cloning-censoring-weighting analysis detailed in Maringe et al. [3].We used a grace period of 77 days so that the grace period would not be shorter than the primary follow-up of 12 weeks.In this design, follow-up starts at depression diagnosis.At this time, two copies (clones) of each individual are created: one clone is allocated to the treatment group (initiate antidepressant treatment within 77 days of diagnosis), the other clone is allocated to the control group (no initiation of antidepressant treatment within 77 days of diagnosis), forcing the study arms to be identical at baseline.If an individual initiated antidepressants during the grace period, the clone in the control group was censored.If an individual did not initiate antidepressants during the grace period, a clone of the individual remained in each arm for 77 days, after which the clone in the treatment group was censored.Cloning accounts for confounding at baseline, but the artificial censoring is usually informative (i.e., treatment decision is based on characteristics also associated with the outcome), and can therefore introduce selection bias.We addressed this using inverse-probability-of-censoring weighting, with the purpose of up-weighting patients remaining in the risk set so that they represent censored patients, and maintain the comparability of the study arms throughout the grace period.To estimate the weights, we fitted a Cox regression model which was then used to predict the individual probabilities of remaining uncensored at each time of event.The model included covariates specified in the manuscript, and was estimated separately for each arm.The weights are the inverse of these probabilities.We estimated ITT and perprotocol effects.The 120-day gap approach described in the earlier section of eMethods was used to estimate the length of medication treatment periods in the per-protocol analysis.Weighted Kaplan-Meier was used to estimate cumulative incidences, and weighted Cox regression to estimate hazard ratios, in the 12 and 52-week follow-up.0.19 (0.12-0.26) 0.22 (0.16-0.28) 0.03 (-0.06-0.12)1.16 (0.75-1.79) 1 pill per day treatment periods, PP j 0.19 (0.13-0.25) 0.19 (0.14-0.24) 0.00 (-0.07-0.07)1.00 (0.70-1.45) Note: PP=per-protocol.Estimates are from the intention-to-treat (ITT) approach unless stated otherwise.
a Complete cases model excluded individuals with missing information in any covariate, cohort n=41,480.
b Individuals who dispensed antipsychotic or antiepileptic medications within 4 months before follow-up start were excluded, cohort n=42,346.
c The grace period was extended to 120 days after diagnosis in defining the treatment group.
d We considered outcomes to have occurred only in individuals who had at least two diagnoses, or one diagnosis and one medication dispensation.
e Outcome was extended to include dispensations for olanzapine (with indication for "mood stabilizing" purpose).
f Follow-up length was extended to from 12 to 18 weeks.
g Individuals were censored if they were hospitalized for any mental health reason (excluding the outcome).
h Cloning-censoring-weighting approach described in detail in Supplemental eMethods.Follow-up began from the index diagnosis.
i When defining antidepressant treatment periods, we assumed that two dispensations falling within 60 days of each other belonged to the same treatment period.See Supplemental eMethods for details of this analysis.
j Antidepressant treatment period lengths were estimated from the total quantity of pills dispensed, and assuming that patients take 1 pill per day.

eFigure 2 .
The Estimated Cumulative Incidence of Mania/Hypomania Over 52 Weeks of Follow-up.Green line represents the treatment group and red line represents the control group.Intention-to-treat estimates.Panel A: weighted data.Panel B: non-weighted data.

Prescription Time-distribution Matching to Assign Start of Follow-up eTable 3. Frequencies of Different Antidepressant Medications in the Treatment Group
Note: N/A = For these medications, there were too few patients to provide exact numbers without compromising identifiability.eTable 4.

Weighted Descriptive Statistics of the Cohort
© 2023 Virtanen S et al.JAMA Psychiatry.

eTable 5. Sex-stratified Analyses: Cumulative Incidence and the Relative Risk of Mania/Hypomania in the Treatment and Control Groups at 12 and 52 Weeks of Follow-up
Note: Estimates are from the intention-to-treat analysis eTable 6.

Sensitivity Analyses: Non-weighted Cumulative Incidence and the Relative Risk of Mania/Hypomania in the Treatment and Control Groups at 12 Weeks of Follow-up
.