Association between suicide attempts and selective serotonin reuptake inhibitors: systematic review of randomised controlled trials

Literature search strategy

We conducted a systematic literature search to identify all randomised controlled trials of SSRIs indexed on Medline between 1967 and June 2003. The search strategy combined the text terms “SSRI”, “serotonin uptake inhibitors”, “fluoxetine”, “Prozac”, “sertraline”, “Zoloft”, “paroxetine”, “Paxil”, “fluvoxamine”, “Luvox”, “Citalopram”, and “Celexa” with the Dickersin filter for randomised controlled trials.12 In addition, we searched the Cochrane Collaboration's register of controlled trials (November 2004) for trials produced by the Cochrane depression, anxiety, and neurosis group with the same strategy. We also reviewed the bibliographies of three systematic reviews13^–15 and identified trials to identify relevant reports. Three authors (SD, BH, and DF) independently reviewed all citations retrieved from the electronic search to identify potentially relevant trials. Each citation was reviewed by at least two individuals. When a unanimous decision could not be reached, a third reviewer was consulted to resolve the difference.

Identification of articles and abstraction of data

To be eligible for inclusion, studies had to be randomised controlled trials comparing an SSRI with either placebo or an active non-SSRI control. We included clinical trials that evaluated SSRIs for any clinical condition. We excluded abstracts, crossover trials, and all trials whose follow up was less than one week. Crossover trials were excluded because of the difficulty in appropriately attributing an outcome to treatment and the poor reporting of the relation between adverse events and treatment.

We developed a standardised data abstraction form that included the condition treated, modes of treatment compared, duration of treatment, the number of patients randomly assigned to each treatment group, the number of patients reported to have completed treatment, patients' demographics, and funding sources. Because our primary aim was to evaluate a rare, serious event and not effectiveness of treatment, we did not quantify the quality of individual study reports by using a formal quality scale. We limited eligibility to trials that were truly randomised and examined individual sources of clinical and methodological heterogeneity including clinical indication, trial duration, sex, age, sample size, and dropouts.

Outcomes

The primary outcome, suicide attempts, included both fatal and non-fatal acts of suicide. We documented rates of fatal and nonfatal suicide attempts separately. Fatal suicide attempts were self inflicted acts resulting in death, as reported in the primary studies. We made conservative assumptions to deal with the published reporting of non-fatal suicide attempts. The authors had literally to use the term “suicide.” The one exception was the use of the term “overdose.” If the authors explicitly reported that there were no adverse or serious adverse events, we recorded that there were no fatal or non-fatal suicide attempts. If no suicide attempts were mentioned but the authors accounted for all adverse events and reasons for discontinuation, we recorded zero suicide attempts. Subjects for which the authors did not indicate a reason for withdrawal or discontinuation we did not count as suicide attempts.

Methodological considerations

We documented how adverse events were reported, dropout rates, sample size, and the number of trials that did not report adverse events. To deal with poor reporting of adverse events, we included a “not reported” category. This category comprised trials that did not mention adverse events or reasons for discontinuation of therapy, provided an incomplete listing of all adverse events, or did not explicitly state that no serious adverse events had been observed. We also documented the proportion of trials that chose to report adverse events beyond percentage thresholds (for example, 5%) or occurring in more than a defined number of patients. We determined the proportion of studies with dropout rates exceeding 15% and 25% and reported the size of trials as the proportion of trials with a total number of patients less than 50, between 50 and 100, and exceeding 100.

Analysis

As an initial description of the risk of suicide overall and in major comparisons, we calculated the absolute risk per 1000 patients treated by dividing the number of events (suicide attempts) by individuals exposed to therapies and multiplying by 1000. To account for exposure time, we calculated the number of episodes of suicide attempts per 1000 person years of exposure by assuming a constant risk over the first year and using a weighted average of exposures.

To evaluate the association between suicide attempts and the use of SSRIs, we undertook three separate meta-analyses: SSRIs compared with placebo, with tricyclic antidepressants, and with other active forms of treatment excluding placebo and tricyclic antidepressants. Within each comparison, we tested the association between suicide attempts and the use of SSRIs by calculating odds ratios using fixed effects models. We used Peto's methods to calculate odds ratios and 95% confidence intervals.16^–18 An odds ratio greater than 1 implies greater risk in the SSRI group, and an odds ratio less than 1 implies greater risk in the non-SSRI group. We conducted separate meta-analyses for the number of fatal and non-fatal suicide attempts. We did not incorporate trials categorised as “not reported” into the analyses.

A priori subgroups of interest were based on age, the duration of the study follow up, proportion of women, and primary diagnosis of participants in the trials (major depression, depression, and other conditions). We examined the reported partial or total funding source (funded by, compared with not funded by, the pharmaceutical industry). We also conducted a cumulative meta-analysis to evaluate the temporal sequence of evidence of effect.

Possible explanations for our findings

In this meta-analysis, the increase in the number of suicide attempts was not associated with a comparable increase in the risks of fatal suicide attempts. Several explanations are plausible. We observed non-significant divergent risks of suicide among different clinical conditions. Estimates for patients with major depression favoured a decrease in suicides with SSRIs, whereas patients with depression and other clinical indications may have as much as an eightfold increase in the rates of suicide, thus resulting in an overall null effect. In all instances, the number of events was too small to generate sufficiently narrow confidence intervals. If the mechanism of action thought responsible for inducing suicidality is true, the agitation and akathasia known to occur with this class of agents may have affected non-depressed and depressed patients differently, inducing more distress in patients with less severe clinical conditions than in those with severe depression. This may account for the greater number of suicide attempts in patients without severe depression. Another explanation could be that these trials do not reflect true practice and that treating more severely depressed patients with a higher inherent risk of suicide in a controlled environment may produce a more favourable ratio of risks to benefits. These explanations would reconcile the observed benefits attributed to SSRIs with the proposed risks associated with the induced agitation that accompanies initiation of treatment, missed doses, decreases of dosage, or discontinuation of treatment. One implication from our findings is that patients with mild illness who are being treated without supervision in the community may require closer monitoring by general practitioners, family, friends, or work colleagues.

A review of published and unpublished sources documented increased rates of suicide in patients with depression when records from the FDA were considered.19 Our review noted suicide attempts at a rate of 3.9 episodes per 1000 patients, whereas suicide attempts documented by Healey approximated 15.3 episodes per 1000 patients treated with SSRIs, a 3.9-fold difference in rates. The difference in rates implies that a substantial proportion of suicide attempts have gone unreported.8 20

Limitations

As additional evidence of difficulties in reporting, we were unable to find documentation confirming or refuting suicide attempts in 51 205 of the 87 650 patients. We conducted a survey of a random sample of 35 (10%) trials that did not report suicide attempts. Of the 17 responders, two reported suicide attempts, seven reported no suicide attempts, and eight confirmed that these data were not collected. Of the two responses that reported suicide attempts, one reported a non-fatal suicide attempt in the SSRI group and none in the placebo group, and the other reported a non-fatal suicide attempt in the SSRI group and two non-fatal attempts in the group taking tricyclic antidepressants. In our random sample of non-responders, 22.2% of trials (n = 2) therefore reported a suicide attempt compared with 18.6% of trials (n = 64) in our entire sample.

Only one trial (0.14%) mentioned a potential association between suicidality or any aspect of self inflicted injuries and SSRIs in their background or discussion sections,21 despite repeated concerns about this adverse effect expressed in scientific discourse.4^–6 One hundred and four of the 702 trials reported adverse events that occurred in excess of a prespecified threshold of either 3%, 5%, or 10% of patients or above a certain number of patients (for example, three patients). As a consequence, rare but lethal complications may have gone unreported or under-reported.

In addition to under-reporting, we documented other important limitations. Of 493 trials that reported dropout rates, 28.7% (n = 18 217) of the 63 478 patients dropped out. In most study areas, patients who are lost to follow up tend to be less compliant with treatment, do not derive comparable benefits, and have a greater frequency of adverse outcomes compared with other patients in trials.22 High rates of losses to follow up may therefore have hindered the ability to detect risks of suicide.

Trial size and duration of follow up are obstacles to detecting associations between SSRIs and rare adverse events. Clinical trials have focused largely on symptoms rather than long term outcomes, such as resolution of depression, prevention of relapse, and long term quality of life. As a consequence, 62.3% of trials (n = 437) enrolled fewer than 100 patients, which made rare events difficult to document in individual trials. In a comparable manner, the mean duration of treatment and follow up in published trials was 10.8 weeks with fewer than 6300 (7.0%) patients in 49 (6.6%) trials followed for more than six months. It is therefore impossible to infer rates of long term risks and benefits of treatment, especially in relation to other therapies. At best, we can make assumptions only on the basis of rates of events from short periods of follow up documented in systematic reviews.

Several study manoeuvres were introduced into clinical trials that may alter response to treatment and rates of suicide and suicide attempts. In 29 trials representing 4243 patients, investigators limited trial entry to those patients who were known to respond to and tolerate SSRIs. Restricting eligibility in this manner would effectively diminish adverse events during the conduct of the trial. In addition, some trials enrolled patients receiving SSRIs into a placebo arm without an adequate washout period, thereby potentially attributing adverse events associated with the discontinuation of treatment to the placebo or attributing adverse events to placebo in patients who were successfully treated by SSRIs.

Conclusions

Despite the limitations of the 702 primary reports we synthesised information by using conservative outcome definitions and documented an association between suicide attempts and the use of SSRIs. A more accurate estimation of the risks of suicide would be garnered from investigators fully and accurately disclosing all events. Our review also showed major limitations in the published medical literature. Doctors rely on published reports for their treatment decisions, making open and complete reporting scientifically and ethically essential.