The FDA asked eight industry sponsors of 12 marketed antidepressant products (bupropion, citalopram, duloxetine, escitalopram, fluoxetine, fluoxetine/olanzapine (ultimately excluded from the analysis), fluvoxamine, mirtazapine, nefazodone, paroxetine, sertraline, and venlafaxine) for datasets from all double blind randomised placebo controlled trials of antidepressant in adults for any indication.
The variables included in these datasets provided detailed information about individual participants. Sponsors’ dataset submissions were received by the FDA between September 2005 and September 2006 as electronic files (in SAS transport file format).
Data were requested from completed, double blind randomised placebo controlled trials with at least 20 participants in each treatment arm. Trials limited to known drug responders, such as those using randomised withdrawal designs, were not included; such studies do not examine the effects of initiating treatment and would eliminate as non-responders those who had shown suicidality during drug treatment.
We asked sponsors to provide a list of all known trials, indicating which trials the sponsor planned to include and which they intended to exclude from the dataset and why. We then provided feedback to the sponsors on which trials should be included in the final dataset. Sponsors summarised the characteristics of the trials included in the datasets in the form of two tables: one providing the dose, duration, and number of participants per trial, and the other providing the trial exclusion criteria.
Other than dataset formats, instructions for identifying and classifying events possibly related to suicidality (suicidal thoughts and actions) and the event classification process, we did not specify who or how companies should retrieve or compile the information we requested. Each company has its own system for archiving and maintaining its records of clinical studies. They are in the best position to judge the optimal approach to completing these tasks. Each company designated a person to serve as the main contact with the FDA.
Adverse events in these trials were solicited by general inquiry and recorded in case report forms. Following the approach used in the paediatric study,4
sponsors were asked to search their electronic databases for adverse events reported during the double blind phase of treatment for terms related to suicidality. Because it was difficult to determine whether events represented a change in condition or resulted from a pre-existing condition, all events reported during the double blind phase were included. Events that occurred more than a day after the randomised treatment stopped were excluded.
The data request letter asked sponsors to search clinical trial databases for preferred terms, verbatim terms, and any comment fields for the following text strings: “accident-”, “attempt”, “burn”, “cut”, “drown”, “gas”, “gun”, “hang”, “hung”, “immolat”, “injur-”, “jump”, “monoxide”, “mutilat-”, “overdos-”, “self damag-”, “self harm”, “self inflict”, “self injur-”, “shoot”, “slash”, “suic-”, “poison”, “asphyxiation”, “suffocation”, “firearm”. All events identified by this search were considered as possibly related to suicidality, unless they were identified as “false positive” results: events that included any of these text strings but were not related to suicidality. For example, “epigastric pain” would be identified in the search for the text string “gas.” Sponsors submitted listings of the events they classified as “false positives,” which were reviewed by FDA staff.
The datasets included 406 clinical trials with 103
491 participants. Six trials were duplicated in the submissions. We excluded 28 other trials: 23 because at least one trial arm contained fewer than 20 participants, three because data at the patient level were not available, and two because the study drug was a combined antidepressant/antipsychotic. We also excluded participants assigned to a non-antidepressant active control drug (608), leaving a total of 372 trials with 99
231 participants (fig 1). Most of the studies were unpublished; those that had been published in some form seldom contained information concerning suicidality in the publication. None of the studies was included in the previous study of paediatric trials.
Fig 1 Flow diagram of exclusions of trials and participants (numbers shown in parentheses)
Determination of suicidality outcomes
Sponsors prepared case narratives for each event possibly related to suicidality (suicidal thoughts and actions). Details that might bias classification (such as treatment assignment) were removed. Because of the large number of participants, the sponsors and not the FDA adjudicated events. Adjudicators, who were blinded to treatment assignment, classified events using the approach of Posner et al.7
Events were classified into seven mutually exclusive categories: 1 completed suicide, 2 suicide attempt, 3 preparatory acts towards imminent suicidal behaviour, 4 suicidal ideation, 5 self injurious behaviour, intent unknown, 6 not enough information (fatal), and 7 not enough information (non-fatal). For participants with multiple events, we coded only the most severe event (for example, completed suicide would be coded rather than suicide attempt if both had occurred).
Three individuals with appropriate training and expertise independently rated events. Sponsors engaged outside contractors experienced in the method of Posner et al,7
either to conduct the rating process or to train the sponsor’s staff. If the three raters were not unanimous in their ratings, a discussion among the raters, led by a fourth rater, was conducted with the goal of achieving consensus. If consensus was not achieved, the event was rated as indeterminate (category 6 or 7).
The primary outcome was defined as suicidal ideation or worse (categories 1, 2, 3, or 4). This corresponded to the primary outcome (definitive suicidal behaviour/ideation) used in the FDA’s paediatric analysis of suicidality. The second outcome variable, considered in parallel with the primary outcome, was preparatory actions or worse (categories 1, 2, or 3), also called suicidal behaviour. We looked at ideation alone for the sake of comparison but did not consider it a clinically important outcome because the category excludes both those who had neither suicidal ideation nor behaviour and those who had both. Because of the previous finding of increased suicidality with antidepressant use in children and adolescents, we considered findings within the general outcomes of suicidal ideation or worse and suicidal behaviour or worse that would be consistent with this observation in a set order. The primary hypothesis for each outcome was that the outcome would be increased among all adults using antidepressants. The principal secondary hypothesis, contingent on rejection of the primary hypothesis, was that the subpopulation of adults most similar to children and adolescents—that is, young adults (defined as age <25)—would show an increased risk. If this second hypothesis was confirmed, we would then consider the possibility of an increased risk in the next youngest 10 year group (age 25-34), and so forth.
We used conditional logistic regression to calculate odds ratios and obtained risk differences with population averaged general estimating equations. These methods were chosen for computational speed and ease of inclusion of covariates. The insensitivity of the results to the method was supported by obtaining similar results for the principal analyses with other techniques: exact methods, Mantel-Haenszel, Bayesian, and unconditional and random effects logistic regression.8
All analyses were conditioned or stratified by study. To examine trial heterogeneity, we added treatment by trial interaction terms to the model. Heterogeneity of effect by drug and drug class was similarly modelled. We used a random effects logistic regression model to model age and age-treatment interaction as continuous variables in a post hoc analysis. Analyses were performed with Stata version 9.2 and SAS version 9.1.
Subgroup analyses were performed based on demographics, characteristics at the trial level, indication, and drug class. As we were particularly interested in age because of the association of suicidality with antidepressant use in the paediatric population, we performed analyses using age and the interaction of age with treatment as both categorical and continuous variables.
To obtain results that could be directly compared with two published meta-analyses of suicidality in clinical trials of selective serotonin reuptake inhibitors (SSRIs),5 6
we compared odds ratios for SSRI with placebo for completed suicide (outcome 1), non-fatal self harm (outcomes 2 and 5), and suicidal ideation alone (outcome 4), and, for SSRIs compared with tricyclics, for any suicide attempt (outcomes 1 and 2).
We classified treatment indications into one of five indication groups by FDA physicians: major depressive disorder, other depressive disorders, other psychiatric disorders, other behavioural disorders, and non-behavioural disorders. We divided the 18 antidepressant drugs used as either primary drugs or active controls in adult trials into five classes: selective serotonin reuptake inhibitors (SSRIs), serotonin-noradrenaline (norepinephrine) reuptake inhibitors (SNRIs), other modern antidepressants, tricyclic antidepressants, and other antidepressants.