In a population of 20 906 children and adolescents 10 to 18 years of age who were beginning antidepressant therapy in British Columbia between 1997 and 2005, we observed no clinically meaningful variation in the risk of suicidal acts according to antidepressant agent within the class of SSRIs or between antidepressant classes. We observed some numeric variations in the effect sizes between individual agents, but the 95% CIs were substantially overlapping. Because of the small number of patients receiving MAOIs, we did not attempt to compare the risk of suicidal acts between MAOIs and SSRIs. Sertraline might have demonstrated an increased risk of suicidality among children with a recorded diagnosis of ADHD, although this finding was very unstable because of the few events. The majority of events occurred in the first half-year after treatment initiation. Our results are consistent with the findings of several other observational studies that reported small or no differences in rates of suicides and suicide attempts between antidepressant classes,8,9
as well as meta-analyses of randomized, controlled trial data that found no heterogeneity among SSRIs.19–21
Studies that do not find statistically significant differences merit special attention to their statistical power to detect a clinically meaningful difference.22
Because of sample size, we had the greatest power to detect a difference in suicidal act rates between venlafaxine and the SSRI drug class. We could exclude a hazard reduction of >11% and an increase of >108% with a confidence level of 95% for venlafaxine users versus SSRI users. Our estimates of the comparative safety of fluvoxamine versus fluoxetine were the least precise. For the fluvoxamine versus fluoxetine comparison, we could exclude a hazard reduction of >54% and an increase of >143%.
During the 9-year study period, the suicide risk among all British Columbia residents 13 to 17 years of age, as reported by the British Columbia Coroner's Office and British Columbia Statistics, averaged ~0.052 suicide deaths per 1000 people.23–25
The rate we observed after initiation of antidepressant use was 5 times higher, which likely reflects the current depressed state and greater degree of psychiatric comorbidity in our population.
A major strength of our study is that the large, stable, study population allowed us to examine a variety of medications and important subgroups. We had an adequate sample size to restrict our study to new initiators of antidepressants with documented diagnoses of depression. An incident user design reduces the likelihood of missing early adverse events, allows for an evaluation of risks over time, ensures that the assessment of patient baseline characteristics is not influenced by effects of antidepressant treatment, and reduces the likelihood that current treatment assignment is influenced by past drug-related experiences, such as adverse affects and refractory symptoms. Our decision to censor data for subjects at treatment discontinuation and to use a proportional-hazards analysis inherently controls for differences in treatment persistence. To obtain clearly identified exposure groups, we compared monotherapies and censored patient follow-up data as soon as the patient switched drugs or augmented therapy. This analytic strategy makes treatment groups more comparable with respect to initial health state and avoids analytic difficulties associated with comparisons of patients who escalate or change treatment in response to treatment failure or adverse effects with patients who do not. This reduces the generalizability of our findings to patients receiving monotherapy; however, this reduction in generalizability is outweighed by the improvement in validity. The extent of generalizability would depend on the extent to which prescribing by British Columbia physicians and underlying suicide risk and comorbidity patterns among British Columbia children and adolescents are similar to those in other regions and health care settings.
Nonrandomized studies using health care utilization data are particularly scrutinized for their control of confounding and the potential for misclassifying diagnoses.26
Confounding would occur if certain antidepressants were more likely to be given to patients with a greater background risk of suicide. Therefore, we controlled for sociodemographic, clinical, and health care utilization factors likely to be independent predictors of suicidality, by using traditional multivariate and high-dimensional propensity score techniques. However, our ability to adjust fully for mental health status was limited by the measurement and reporting of mental health conditions as ICD-9 codes on insurance claims to the provincial government. Random misclassification of confounders in health care utilization databases leads to incomplete adjustment of confounding bias.27
For example, our finding that expanding the population to include non–treatment-naive subjects increased the apparent protective effect of tricyclic antidepressants, compared with SSRIs, suggests that there may be residual confounding by the presence of treatment-resistant depression or other suicide risk factors in this analysis.
In terms of the validity of our suicide attempt definition, intentional self-harm E-codes do not distinguish between suicidality and self-harm without suicidal intent, despite their widespread use in research and surveillance to identify suicide attempts. However, a study of the validity of deliberate self-harm E-codes to identify suicide attempts reported a positive predictive value of 86% for these codes, relative to the standard method of medical chart review.15
To the extent that suicide attempts are assigned the correct E-codes, these admissions should be captured in the British Columbia hospital discharge database, because this database had an E-coding completeness rate of >95% for the years evaluated. Although it is possible that some suicide attempts were recorded as unintentional injuries or injuries of unknown intent, a Canadian study found that E-codes and medical chart review yielded consistent estimates of the proportions of poisoning-related hospitalizations among subjects 18 to 24 years of age that were attributable to intentional self-harm, and a US study reported 95% agreement between E-codes and physician opinions regarding the intent of injuries that resulted in hospitalization. We did not have data on suicide attempts that were treated in an emergency department without inpatient admission. However, the majority of pediatric and adolescent patients with nonfatal self-harm seen in the emergency department are admitted to the hospital,11
and our inclusion of only more-severe attempts that required hospitalization is likely to increase the specificity of our definition, at some cost in sensitivity. High specificity is desirable, in that relative risk estimates are unbiased if outcomes are assessed with 100% specificity, even if sensitivity is far lower.28