Suicide risk with selective serotonin reuptake inhibitors and new-generation serotonergic-noradrenergic antidepressants in adults: a systematic review and meta-analysis of observational studies

Search strategy and selection criteria

We conducted our systematic review according to the PRISMA guideline.³⁴ The review was registered with PROSPERO on January 21, 2020, but as of April 29, the submission was still being processed. As we were immediately informed that the evaluation process would take several months, the study protocol was also registered with the Open Science Framework on February 3, 2020 (https://osf.io/eaqwn/). We searched the databases MEDLINE, PsycINFO, Web of Science, PsycARTICLES and SCOPUS for all original studies in English published from 1990 to January 2020, using the terms “antidepressants”, “selective serotonin reuptake inhibitor”, “serotonin norepinephrine reuptake inhibitor”, and all individual new-generation antidepressants in combination with “suicide”, “suicide attempt”, or “self-harm”. The detailed search strategy and the Pubmed search term are shown in the appendix.

We additionally searched MEDLINE, PsycINFO and the German database PSYNDEX for non-English records, but found no additional studies that met our inclusion criteria. In order to find unpublished studies, we searched observational studies registered with ClinicalTrials.gov. One study met our inclusion criteria, but it was already included through our database search. Screening of titles and abstracts and subsequently assessments of full-texts were conducted independently by two investigators (SA and SK) during February and March 2020. Any discrepancies were resolved by consensus and, if necessary, through arbitration by the lead investigator (MPH).

The eligibility criteria were case-control and cohort studies in adults with depression or any other psychiatric and non-psychiatric treatment indication. We also included studies reporting risk estimates for any unspecified antidepressant class when it was possible to infer from the data or external sources (e.g. prescription rates for different antidepressant classes in the underlying population) that at least 75% of all prescriptions in the study sample were SSRI or SNA. We included only studies that report a risk estimate for an exposed group in relation to an unexposed group. Therefore, we excluded studies that compared the suicide risk with SSRI or SNA to other drugs (e.g. tricyclics) and studies that examined variations in suicide risk across different treatment phases within exposed subjects (e.g. suicide risk during acute therapy vs. continuation or maintenance therapy). In order to minimize confounding by indication, we required exposed and unexposed groups to be broadly comparable in terms of clinical and socioeconomic characteristics. We thus excluded studies that compared the suicide risk in antidepressant users with increased baseline suicide risk (e.g. psychiatric inpatients) to an unexposed group with low baseline risk (e.g. healthy people from the general population) if differences in baseline risk where statistically not controlled for (e.g. via propensity score matching or covariate adjustment). If the unexposed group consisted of patients with a baseline suicide risk comparable to the exposed group (e.g. when both groups comprised psychiatric inpatients), then we also included studies with unadjusted risk estimates.

Outcomes

Our primary outcomes were the risk of suicide and suicide attempt in people exposed to new-generation antidepressants relative to an unexposed group according to the statistically best-adjusted analysis reported in the primary study. Suicide and suicide attempt were analysed separately and, to allow for comparability with the Barbui study,²⁸ also as a composite outcome. We further conducted separate analyses for depression and any treatment indication unspecified, as well as for SSRI, SNA, and any new-generation antidepressant (comprising SSRI, SNA and any class unspecified). If studies reported not an overall risk estimate for the entire observation period, but instead separate estimates for mutually exclusive treatment periods (e.g. acute therapy, continuation and maintenance therapy), we always chose the risk estimate for the earliest treatment period. This decision was specified a priori in the protocol to minimise survivorship bias, as a consistent body of evidence indicates that the suicide risk is highest at the beginning of an illness episode when patients are acutely distressed.^35-38

Data analysis

To allow for comparability, study quality was rated with the same 10-point scale applied by Barbui et al.²⁸ This scale assesses study quality based on six domains (population framework; study design; description of demographic data; description of clinical data; description of outcome data; and covariate adjustment) and ranges from 0 (minimal quality) to 10 (maximal quality) points.^39,40 The scale and its rating system are shown in the appendix. The quality of evidence for each outcome was evaluated according to the GRADE system.^41,42 Study fCOI was coded present when a lead-author (i.e. first or last author) had received payments from the pharmaceutical industry (e.g. as speaker, consultant or adviser); when the professorship of a lead-author was funded by the pharmaceutical industry; or when the study was sponsored by the pharmaceutical industry. For it we not only assessed the fCOI statement made in the primary study, but also searched the internet and declarations in other publications by the same authors.

We used an inverse variance random-effects model and computed relative risk estimates (RE) with 95% confidence intervals based on the odds ratios or hazard ratios extracted from the primary studies. We used the Sidik-Jonkman and Hartung-Knapp adjustments recommended in the Cochrane Handbook.⁴³ Publication bias was inferred from funnel-plot asymmetry and tested with the trim-and-fill method.⁴⁴ Heterogeneity (inconsistency of between-study results) was estimated with I²-statistics. All analyses were conducted with the R packages metafor and metagen. The R-code and raw data are available online, https://osf.io/eaqwn/.

Patient involvement

Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research.

Search results and inclusion of primary studies

The literature search identified 9542 records, which were reduced to 6676 after removal of duplicates. After screening of titles and abstracts, further 6602 records were excluded and 74 full-text were assessed for eligibility. A final number of 32 studies were retained. Of these, 5 focused on unique conditions other than depression, other affective disorders or anxiety disorders. These studies were not included in the quantitative synthesis (meta-analysis) but the results are reported qualitatively in the appendix. Altogether 27 studies, of which 19 on depression (plus other affective disorders and anxiety disorders) and 8 on any treatment indication unspecified were included in the meta-analysis. Included in this dataset are also the 6 adult studies incorporated in the Barbui meta-analysis.²⁸ The study selection process is shown in Figure 1 and the characteristics of the primary studies included in the meta-analysis are summarized in the appendix. Briefly, 8 studies were conducted in North America (7 in the USA and 1 in Canada), 17 studies in Europe and 2 elsewhere (1 in Australia and 1 in New Zealand). 9 studies had fCOI (among them all 6 studies included the Barbui study) and 18 had no fCOI. Study quality ratings ranged from 4 to 9 points (median=8, interquartile range=2). Detailed quality ratings for all studies are shown in the appendix.

• Download figure
• Open in new tab

Figure 1:

PRISMA flow chart

Meta-analyses for depression

The meta-analytic results for the association between exposure to new-generation antidepressants and suicide risk in depression studies are shown in Figure 2. The summary effect estimate for suicide with any new-generation antidepressant was RE=1.29, 0.90-1.86. For SSRI the effect estimate was RE=1.09, 0.51-2.33 and for SNA it was RE=1.31, 0.62-2.76. The funnel-plot was asymmetrical and the trim-and-fill method suggested 5 missing studies. Imputation of missing studies revealed a significantly increased risk, RE=1.64, 1.14-2.36. Funnel-plots for all outcomes and Egger’s test for funnel-plot asymmetry are shown in the appendix. Study fCOI was a significant moderator (Q=6.35, df=1, p=0.012); studies with fCOI (RE=1.00, 0.75-1.34) reported considerably lower risk estimates than studies without fCOI (RE=2.29, 0.99-5.28). Forest-plots for fCOI subgroups are shown in the appendix.

• Download figure
• Open in new tab

Figure 2:

Forest-plots for suicide risk in patients with depression (including other affective disorders and anxiety disorders). SSRI: selective serotonin reuptake inhibitor; SNA: new-generation serotonergic-noradrenergic antidepressant (includes serotonin norepinephrine reuptake inhibitor, e.g. duloxetine and venlafaxine, and atypical antidepressants, e.g. mirtazapine and bupropion); S: Suicide; SA: Suicide attempts; RE: risk estimate

For suicide attempt the summary effect estimate for any new-generation antidepressant was RE=1.31, 1.03-1.68 (for SSRI, RE=0.99, 0.57-1.72; for SNA, RE=1.32, 0.79-2.23). Trim-and-fill method estimated that 8 studies were missing and with imputation of these studies the result was RE=1.61, 1.24-2.10. Study fCOI had a significant effect (Q=12.86, df=1, p<0.001); studies with fCOI (RE=0.91, 0.70-1.17) reported considerably lower risk estimates than studies without fCOI (RE=1.84, 1.31-2.58). The summary effect estimate for suicide and suicide attempt combined (composite outcome) with any new-generation antidepressant was RE=1.29, 1.06-1.57 (for SSRI, RE=1.03, 0.70-1.51; for SNA, RE=1.28, 0.86-1.90). Trim-and-fill method suggested that 12 studies were missing and the result with imputation of missing studies was RE=1.59, 1.28-1.98. Study fCOI was again a significant moderator (Q=21.87, df=1, p<0.001); studies with fCOI showed no clear association (RE=0.91, 0.77-1.09), whereas studies without fCOI reported significantly increased suicide risk (RE=1.94, 1.46-2.59). Of 12 studies that reported risk estimates for SSRI, 8 studies (67%) had fCOI and 4 studies (33%) had no fCOI. By contrast, among the 8 studies reporting risk estimates for any class unspecified, none (0%) had fCOI.

Meta-analyses for all treatment indications

The meta-analytic results for all treatment indications are shown in Figure 3. The risk estimate for suicide with any new-generation antidepressant was RE=1.44, 1.15-1.80 (for SSRI, RE=1.33, 0.86-2.04; for SNA, RE=1.28, 0.74-2.21). Trim-and-fill method estimated that 5 studies were missing and the corresponding result after imputation of missing studies was RE=1.62, 1.28-2.06. Study fCOI had a significant effect (Q=11.36, df=1, p<0.001). Studies with fCOI reported no increased risk (RE=1.08, 0.88-1.33), but studies without fCOI did (RE=1.98, 1.43-2.76).

• Download figure
• Open in new tab

Figure 3:

Forest-plots for suicide risk in patients with any treatment indication (including depression and any unspecified psychiatric and non-psychiatric condition). SSRI: selective serotonin reuptake inhibitor; SNA: new-generation serotonergic-noradrenergic antidepressant (includes serotonin norepinephrine reuptake inhibitor, e.g. duloxetine and venlafaxine, and atypical antidepressants, e.g. mirtazapine and bupropion); S: Suicide; SA: Suicide attempts; RE: risk estimate

The risk of suicide attempt with any new-generation antidepressant was RE=1.47, 1.17-1.85 (for SSRI, RE=1.08, 0.67-1.75; for SNA, RE=1.34, 0.81-2.21). Trim-and-fill method suggested that 7 studies were missing and the result after imputation of missing studies was RE=1.79, 1.40-2.28. Study fCOI was a significant moderator (Q=21.97, df=1, p<0.001); no clear effect was shown in studies with fCOI (RE=0.91, 0.70-1.17), but studies without fCOI revealed significantly increased risk (RE=2.05, 1.57-2.67).

The risk of suicide and suicide attempt combined with any new-generation antidepressant was RE=1.45, 1.23-1.70 (for SSRI, RE=1.19, 0.88-1.60; for SNA, RE=1.28, 0.90-1.80). Trim-and-fill method estimated that 14 studies were missing; the result after imputation of missing studies was RE=1.75, 1.47-2.08. Likewise, there was a significant effect for fCOI (Q=37.17, df=1, p<0.001); studies with fCOI reported no increased risk (RE=0.96, 0.82-1.12), but studies without fCOI showed significantly increased risk (RE=2.02, 1.66-2.46).

The quality of evidence according to GRADE was rated very low for all outcomes due to substantial inconsistency of between-study results (I²≥85%, see Figures 2 and 3).

Sensitivity and subgroup analysis

Three effect estimates were found to be potential outliers. A sensitivity analysis with these potential outlier effects excluded revealed no meaningful differences compared to the results based on the full dataset (see supplementary material). The subgroup analyses for the composite outcome in association with any new-generation antidepressant are reported in Table 1. No significant subgroup differences were found for main treatment indication (depression vs. any unspecified), age (old adults vs. young adults), study design (cohort vs. case-control), study quality (high vs. low), and covariate adjustment (yes vs. no). There was a significant subgroup-difference for antidepressant class. Risk estimates for SSRI (RE=1.19, 0.88-1.60) and SNA (RE=1.28, 0.90-1.80) did not differ meaningfully, but studies that examined any class unspecified reported higher risk estimates (RE=1.87, 1.55-2.25). North American studies (all but one from the USA) reported significantly reduced suicide risk with new-generation antidepressants (RE=0.82, 0.68-0.99), whereas European studies showed significantly increased risk (RE=1.82, 1.51-2.20).

The fCOI results for all outcomes are reported in detail above and the subgroup forest-plot is shown in Figure 4. In studies without fCOI visual inspection of the funnel-plot indicated no asymmetry. This finding was statistically supported by Egger’s test, t=-0.8456, df=31, p=0.404. By contrast, in studies with fCOI the funnel-plot was asymmetrical, both visually and statistically, t=-2.1368, df=26, p=0.042. Both funnel-plots are shown in the appendix. Study fCOI and study location were related; while 6 of 8 (75%) North American studies had fCOI, only 3 of 17 (18%) European studies had so. Study fCOI was also related to drug class examined. While 9 of 16 studies (56%) on SSRI or SNA had fCOI, 0 of 11 studies (0%) on any class unspecified had so. The significant subgroup differences for drug class and study location are thus at least in part attributable to study fCOI.

• Download figure
• Open in new tab

Figure 4:

Forest-plot for suicide risk by financial conflicts of interest (fCOI) subgroups. SSRI: selective serotonin reuptake inhibitor; SNA: new-generation serotonergic-noradrenergic antidepressant (includes serotonin norepinephrine reuptake inhibitor, e.g. duloxetine and venlafaxine, and atypical antidepressants, e.g. mirtazapine and bupropion); S: Suicide; SA: Suicide attempts; RE: risk estimate

Limitations

Population-based register-linkage studies allow studying very large and representative samples and to control for various socioeconomic and clinical covariates, which increases generalisability of findings and which minimizes confounding by indication. Yet these studies may have important limitations, such as incomplete assessment of exposure periods. For instance, in the cohort study by Tiihonen et al.,⁵⁶ the register used to assess antidepressant exposure contained only the purchase of medication from pharmacies in outpatient care, but not prescriptions from hospital care. This means that suicides occurring in hospital or shortly after discharge were necessarily ascribed to the unexposed period in many first-episode patients, although it is likely that in these cases antidepressants were prescribed by a hospital doctor. As the suicide risk is highest in the first weeks after hospital admission and immediately after discharge,⁵⁷ this incomplete assessment of exposure may result in an underestimation of the suicide risk in the exposed group. On the other hand, prescription registries record only whether drugs had been dispensed, but not whether the drugs were actually taken as prescribed. It is known that in pharmacotherapy for chronic conditions patients show rather low adherence rates.⁵⁸ If some people redeemed a prescription but did not use the drugs for any reason, this may result in an overestimation of the suicide risk in the exposed group.

Another issue is confounding by indication. Although most studies tried to minimize this bias by controlling for important covariates such as previous suicide attempts and depression severity, unlike RCT it cannot be excluded that patients with more severe forms of psychopathology were more likely to receive antidepressants. On the other hand, most SSRI and SNA are prescribed by GPs,^23,24 often for mild and subclinical depression.^59,60 Therefore, antidepressant use may not necessarily indicate more severe psychopathology in people with depression. For instance, Olfson et al.²⁴ showed that US patients with serious psychological distress (a strong risk factor for suicide) are no more often treated with antidepressants than patients with less serious or no distress. This resonates with our meta-analytic finding that adjusted risk estimates were not significantly lower than unadjusted estimates. Remarkably, the two studies that reported the strongest suicide-protective effects were both unadjusted.^61,62 Depending on the sample studied, confounding by indication could thus result in both over- and underestimation of risk estimates. In accordance, covariate adjustment was not able to explain the high inconsistency in between-study results. To better explain heterogeneity of risk estimates, more detailed methodological aspects presumably need to be considered (e.g. concomitant psychosocial support).