Pharmacological and Somatic Treatment Effects on Suicide in Adults: A Systematic Review and Meta-Analysis

A systematic search was performed on MEDLINE using the following search criteria without language restriction before December 15, 2019: ([Mood stabilizers OR Depakote OR valproate OR valproic acid OR lamotrigine OR lamictal OR carbamazepine OR tegretol] OR [Antipsychotics or neuroleptics or quetiapine or Seroquel or olanzapine or Zyprexa or zydis or Latuda or lurasidone or aripiprazole or abilify or rexulta or brexpiprazole or ziprasidone or Geodon or risperidone or Risperdal or asenapine or saphris] OR [Lithium] OR [clozapine] OR [ect OR electroconvulsive therapy] OR [TMS or transcranial magnetic stimulation] OR [vagus nerve stimulation OR VNS]) AND suicide. Two authors (Anubhav Kidambi and Victor J. Avila-Quintero) independently identified relevant studies among the results of the systematic search. References of included studies were also screened for relevant studies. Resources from Yale University were used to translate studies published in languages other than English. This study was con- ducted according to the MOOSE guidelines (Stroup et al., 2000; see Supporting Information).

We only included studies that

Randomized trials, observational cohort studies, and case-control studies were included. Given the broad scope of this meta-analysis, we did not include psychotherapeutic treatments in our review (D'Anci et al., 2019; Fox et al., 2020). Furthermore, given the extensive literature on the topic, we did not include standard oral antidepressants (such as selective serotonin reuptake inhibitors) in this review (Carpenter et al., 2011; Gibbons et al., 2012; Jick et al., 2004; Perlis et al., 2007).

Two authors (Samuel T. Wilkinson and Daniel T. Diaz) independently extracted the data using a standardized form: study design, intervention of interest, comparison group, odds ratio (OR)/hazard ratio (HR)/relative risk, number of suicide deaths, and denominator in each comparison group. Where studies had some but not all necessary data for meta-analysis, efforts were made to contact corresponding au- thors. In extracting data, we prioritized adjusted ORs or HRs calculated by the authors of the original papers. Where this information was not available, we calculated unadjusted ORs using raw data. For observational studies where cohort groups often differed with respect to person-time of exposure, we favored (where available) the use of person-years in the denominator. For events with rare occurrences such as suicide, ORs and relative risk yield approximately equal results (Rothman et al., 2012). In reporting total number of subjects exposed to a given intervention, total numbers of subjects were used; in instances where the published literature only reported person-years, these were used but are separated from counts as noted below. Quality of studies was assessed using the Newcastle–Ottawa scale. In specific, the Newcastle–Ottawa scale scores studies on quality based on selection of cohorts, comparability of cohorts, and outcome. Four possible points are given for selection of cohorts (maximum points awarded for a cohort that is truly representative of patients in the community with mental illness, if the nonexposed cohort is drawn from the same community as the exposed cohort, if the exposure is ascertained through a secure record, and there is clear demonstration that the outcome was not present at the start of the study. Two possible points are given for comparability if the study controls for basic potential confounders (socioeconomic status, age, race, and gender) as well as content-specific confounders. Three possible points are given for outcome if the assessment of outcome is done independently or by reference to secured records (i.e., medical records), if the follow-up was long enough for the outcome to occur, and if there is adequacy of follow-up cohorts (Wells et al., 2000).

Statistical analyses were performed using STATA 15.1 and Comprehensive Meta-Analysis 3.0 (CMA; Biostat). The primary outcome for each comparison was suicide death, which was assessed as described in each individual study (either through prospective observation or through records database matching [i.e., the National Death Index or similar death registries]). The time over which suicide was assessed varied by study; as a rule, we used the follow-up time given in each individual study.

Our primary outcome measure for this meta-analysis was effect sizes of pharmacological and somatic interventions with respect to suicide death. We report the random-effects model as our primary analysis because it is likely to be more conservative and more appropriate for the data as we expected significant heterogeneity between studies given differences in design, approach, doses of drug used, and other features.

The Q-statistic was used to evaluate heterogeneity between studies. This metric provides a statistical test indicating whether different effect sizes between studies are attributable to subject level sampling error alone or other sources of variability. We estimated the degree of heterogeneity using the I-squared statistic, which approximates the proportion of total variance attributable to between-study variance. We also assessed publication bias graphically (by use of funnel plot, which plots effect size vs. standard error for each study) and statistically by use of the Egger's test (Egger et al., 1997).

Where possible, we stratified meta-analyses by clinical diagnosis and comparison group. This involved comparison of the intervention of interest to another active control (almost always, a psychotropic medication) or to placebo/no intervention. The placebo/no intervention groups were comprised of randomized trials and reports of cohort studies that did not specify what treatments patients were receiving who were not receiving the intervention of interest.

Meta-analysis included 36 studies involving 58,244 subjects treated with Lithium and 87,965 controls (in addition to 103,487 person-years of lithium and 160,729.2 person-years of control follow-up). Studies included 20 cohort studies, 3 case-control studies, and 13 randomized trials. Seventeen studies included subjects exclusively with bipolar disorder and 19 studies included subjects with other diagnoses.

Figure 1 displays the effects of pharmacological treatment with lithium on suicide risk. Among studies conducted with individuals with bipolar disorder, lithium was associated with a significant reduction in the odds of suicide compared to an active control (OR = .58, 95% confidence interval [CI] = 0.40–0.85, z = −2.78, p = .005, k = 12; heterogeneity: Q = 14.5, I² = 24.1%; Figure 1) and when compared to placebo or no specific intervention (OR = .46, 95% CI = 0.25–0.82, z = −2.61, p = .009, k = 9; heterogeneity: Q = 18.9, I² = 57.1%). Among studies conducted with individuals with mixed psychiatric disorders, lithium was similarly associated with a reduction in odds of suicide compared to placebo or no intervention (OR = .27, 95% CI = 0.17–0.45, z = −5.21, p < .001, k = 12; heterogeneity: Q = 9.67, I² = 0.0%) but not compared to an active control (OR = .89, 95% CI = 0.64–1.23, z = −0.73, p = .468, k = 7; heterogeneity: Q = 6.77, I² = 11.4%).

There was evidence of publication bias when lithium was compared against no active control among patients with bipolar disorder using the Egger's test (z = −3.30, p = .013) and was also suggested by funnel plot asymmetry. A reduction in suicide odds was no longer significant when the trim and fill method was used to adjust for funnel plot asymmetry (adjusted OR = 0.63, 95% CI = 0.35–1.15, p = .12). There was no evidence of publication bias for the other lithium analyses. Funnel plots examining all lithium outcomes are provided in Figure S2.

In sensitivity analyses where only randomized trials were included, estimates were generally similar to the overall analysis though without statistical significance (bipolar disorder, vs. active control OR = 0.74, 95% CI = 0.21–2.69, z = −0.45, p = .65, k = 5; insufficient data for bipolar disorder without active control; mixed disorders, vs. active control, OR = 0.34, 95% CI = 0.07 to 1.55, z = −1.40, p = .16, k = 4; mixed disorders, vs. placebo, OR = 0.65, 95% CI = 0.10–4.23, z = −0.45, p = .65).

Meta-analysis included seven studies involving 5424 subjects treated with clozapine and 56,548 controls (in addition to 100,204 person-years of clozapine and 36,473 person-years of control follow-up). Included studies were comprised of four cohorts, two case-control studies, and one randomized trial. All studies were conducted in patients with psychotic disorders. Most studies (four out of seven) reported that the comparison group was taking a non-clozapine antipsychotic; one study specified that the comparison group was not taking an antipsychotic. Figure 2 displays the effects of pharmacological treatment with clozapine on the risk of suicide. Clozapine was associated with a reduction in the odds of suicide (OR = 0.46, 95% CI = 0.27–0.81, z = −2.70, p = .007, k = 7; Figure 2). There was a large amount of heterogeneity among studies for this comparison (Q = 18.3, I² = 67.3%). There was no evidence of publication bias according to a funnel plot (Figure S3) or the Egger's test (z = 1.36, p = .233). There were insufficient randomized trials to conduct sensitivity analyses.

Meta-analysis included 11 studies involving 1100 subjects treated with ECT and 6077 controls (in addition to 8508 person-years of ECT and 29,221 person-years of control follow-up). Included studies were comprised of seven cohorts, three case-control studies, and one randomized trial. No ECT was the comparison group in all studies and in most cases involved some form of psychotropic medicine therapy. Figure 3 depicts the association of ECT with risk of suicide death. The association between ECT and lower odds of suicide did not achieve statistical significance (OR = 0.77, 95% CI = 0.58–1.00, z = −1.93, p = .053, k = 11; Figure 3). There was a small amount of heterogeneity between studies (Q = 11.8, I² = 15.6%). Neither the funnel plot (Figure S4) nor Egger's test (z = −1.24, p = .245) suggested evidence of publication bias. Notably, many of these studies were older and of lower quality (Table S3). There were insufficient randomized trials to conduct sensitivity analyses.

Meta-analysis included 12 studies involving 46,569 subjects treated with antipsychotics and 10,454 controls (in addition to 16,907 person-years of antipsychotics and 11,349 person-years of control follow-up). Included studies were comprised of five cohorts, two case-control studies, and five randomized trials. Given protective effects against suicide previously demonstrated with lithium, we stratified the analysis based on comparison group (lithium vs. non-lithium comparator). Additionally, studies were stratified by the psychiatric diagnosis of the study sample.

Among patients with bipolar disorder, the association between antipsychotic use and a reduction in odds of suicide when compared to non-lithium therapies or no therapies did not achieve statistical significance (OR = .73, 95% CI = 0.51–1.05, z = −1.70, p = .090, k = 6; het-erogeneity: Q = 4.2, I² = 0.0%; Figure 4). Among patients with psychotic disorders, the association between antipsychotic use and a reduction in risk of suicide when compared to non-lithium therapies or no therapies did not achieve statistical significance (OR = 0.39, 95% CI = 0.14–1.07, z = −1.82, p = .069, k = 6; heterogeneity: Q = 8.91, I² = 43.9%). Neither the funnel plots nor Egger's tests (z = 1.80, p = .15 for bipolar disorder; z = 0.36, p = .74 for psychotic disorder) suggested publication bias for these comparisons. When compared to lithium, no statistically significant association was seen (OR = 2.68, 95% CI = 0.75–9.55, z = 1.52, p = .128; heterogeneity: Q = 2.44, I² = 18.0%).

There were insufficient randomized trials to conduct sensitivity analyses of antipsychotics in patients with bipolar disorder. Sensitivity analyses using only randomized trials in patients with psychotic disorder, antipsychotic versus non-lithium therapies, or no therapies showed (OR = 0.69, 95% CI = 0.21–2.24, z = −0.62, p = .54, k = 3).

Meta-analysis included 16 studies involving 54,135 subjects treated with antiepileptic mood stabilizers and 61,875 controls (in addition to 93,629 person-years of antiepileptic mood stabilizers and 87,104 person-years of control follow-up). Included studies were comprised of 11 cohorts, 1 case-control study, and 4 randomized trials. Given the protective effects against suicide previously demonstrated with lithium (Cipriani et al., 2013), we stratified the analysis based on comparison group (lithium vs. non-lithium comparator). When compared to non-lithium therapies or no therapies, there was no significant association with risk of suicide (OR = 0.91, 95% CI = 0.36–2.27, z = −0.21, p = .84, k = 8; heterogeneity: Q = 31.2, I² = 77.6%; Figure 5). Neither the funnel plot nor Egger's test suggested publication bias (z = 1.24, p = .261). When compared to lithium, antiepileptic mood stabilizers were associated with a higher risk of suicide death (OR = 1.35, 95% CI = 1.01–1.81, z = 2.04, p = .042, k = 13; heterogeneity: Q = 16.8, I² = 28.6%). Both inspection of the funnel plot and Egger's test (z = 2.28, p = .043) suggested publication bias for this comparison. When the Trim and Fill method was used to adjust for potential funnel plot asymmetry, the adjusted odds ratio was 1.23 (adjusted OR = 1.23, 95% CI = 0.91–1.68, p = .18, Q = 23.0) for antiepileptic mood stabilizers compared to lithium.

There were insufficient data to conduct sensitivity analyses in randomized trials.

There was an insufficient number of studies (k = 2) to meta-analyze the effects of VNS on suicide-related outcomes (Aaronson et al., 2017; Olin et al., 2012). Our search revealed no studies assessing the effect of TMS, Magnetic Seizure Therapy, or transcranial direct current stimulation on risk of suicide.

Lithium is indicated for use in bipolar disorder (acute treatment for mixed, manic, or depressive phases as well as for maintenance therapy) and as adjunctive therapy for unipolar depression. Findings suggest that lithium is effective in protecting against suicide in bipolar disorder. In some clinical contexts (compared to no intervention), lithium may also be protective in unipolar depression and psychotic disorders. The finding that lithium is not better than an active comparator in unipolar depression and psychotic disorders suggests that interventions specifically indicated for these disorders (i.e., antipsychotics for psychotic disorders) may more effectively treat the underlying disorder and consequently reduce the risk of suicide.

Despite evidence of its protective effects, the use of lithium has waned in recent decades (Kessing et al., 2016; Rhee, Olfson, Nierenberg, et al., 2020; Rhee, Olfson, Sint, et al., 2020). A number of factors have likely contributed to this trend, including the availability of newer generation antipsychotics that are also indicated for bipolar disorder and have active marketing support from pharmaceutical companies. Also, lithium has tolerability concerns, a need for frequent plasma monitoring, and risks of long-term renal impairment that may limit its use in select populations. However, the decline in lithium use in concert with the rapid increase in the use of antipsychotics for bipolar disorder highlights the importance of investigation into the comparative protective effects of these treatments.

Antipsychotic clozapine was associated with a reduction in risk of suicide death. Because of the risk of agranulocytosis associated with clozapine, this medicine is subject to a strict Risk Evaluation and Mitigation Strategy (REMS), that requires regular complete blood count monitoring. Along with an unfavorable metabolic safety profile, strict REMS requirements have limited its use in practice. Clozapine is generally not used until other antipsychotics have failed to produce adequate symptom relief. Notably, clozapine is approved by the Food and Drug Administration for reducing the risk of recurrent suicidal behavior in patients with schizophrenia or schizoaffective disorder.

Understanding treatment approaches that reduce the risk of suicide may shed light on the mechanistic basis of suicide risk. Advances in elucidating the mechanism of lithium suggest that it may serve to upregulate neuroprotective factors (i.e., brain-derived neurotrophic factor and BCL-2) and downregulate apoptotic factors (Malhi et al., 2013). Furthermore, lithium may also work through the modulation of dopaminergic, glutamatergic, and GABAergic pathways (Malhi et al., 2013). Lithium's beneficial effects on impulsivity and aggression may mediate this reduction in suicide risk from a psychological perspective (Cipriani et al., 2013). The mechanisms by which clozapine reduces suicide risk is less clear, but given its wide array of effects may include simultaneous modulation of multiple neurotransmitters (Meltzer et al., 2000) (monoamine-based) and intracellular (Leveque et al., 2000) (i.e., modulation of N-methyl-d-aspartate receptor expression) systems.

While practice standards suggest ECT is to be used for patients with mood disorders at risk of suicide, the evidence of the effects of ECT on suicide death have been inconsistent. Given the nature of ECT, blinded studies are essentially no longer possible due to ethical and practical concerns. Furthermore, because ECT is usually reserved for the most severely ill patients, observational studies that adequately account for confounding by indication are inherently difficult to conduct. While ECT has consistently been associated with lower risks of all-cause mortality (Jorgensen et al., 2020; Rhee et al., 2021), further carefully conducted research is needed to examine the extent to which ECT may provide protection against suicide risk.

One critical area of future investigation involves improving access to therapies that have been shown to reduce the risk of suicidal behaviors. For instance, lithium, clozapine, and ECT are all underused in community settings (Baldessarini et al., 2007; Kelly et al., 2018; Rhee, Olfson, Nierenberg, et al., 2020; Rhee, Olfson, Sint, et al., 2020; Wilkinson et al., 2018), limiting their potential public health impact on suicide prevention.

While several treatments have demonstrated evidence of reducing the risk of suicide or suicidal behaviors, improved therapies are needed given the side effect profiles of many of these medicines. A key challenge involves the development of treatments that are targeted to treat patients with suicidal ideation. Over the last several years, clinical trials of antidepressants and antipsychotics have had more restrictive eligibility criteria (Zimmerman et al., 2015), to the point that a large majority of clinical trials excluded patients with even low risk of suicide. As a result, several of the most commonly used treatments have little evidence of the efficacy of protective effects against suicide in patients who are at the highest risk. Reasons for excluding patients at some risk of suicide from clinical trials involve liability, the ethical quandary of allocating someone at risk of suicide to a placebo condition, higher study costs due to more safeguards, and the theoretical concern that treatments may worsen suicidal ideation (Pearson et al., 2001). In 2018, the FDA issued im- portant guidance on including patients with suicidal ideation in clinical trials (Food and Drug Administration, 2018); this guidance advises that patients at risk of suicide can be included in clinical trials for the development of interventions that target those most at risk. One key principle from this guidance is that a standard of care option should be included in lieu of standalone placebo in the comparator arm in trials where patients at significant risk of suicide are included.

Several limitations of this meta-analysis require comment. First, the inclusion of non-randomized studies can bias results if comparator groups are different in ways that are not measured or controlled for. For instance, the severity of illness (i.e., in mood disorders) is difficult to account for in non-randomized studies and may introduce confounding. However, all studies included passed peer review and the majority were of high quality with moderate to good comparability between groups (Tables S1–S5). While the Newcastle–Ottawa Scale can have fair to high interrater reliability (Oremus et al., 2012), its use for reducing bias in meta-analyses is recommended by the Cochrane Collaboration (Higgins et al., 2019). Second, while not a perfect assessment, there was significant heterogeneity in three of the meta-analyses performed: lithium in bipolar disorder versus no active control; clozapine; and antiepileptic mood stabilizers versus no antiepileptic mood stabilizers. (For the other seven analyses, heterogeneity was not significant.) Third, among some studies included, the comparator groups were not consistently defined. However, significant findings were also found where comparator groups were consistently defined (i.e., lithium vs. antiepileptic mood stabilizers). Fourth, we did not conduct subgroup analyses when significant heterogeneity was found because there were too few studies for each outcome. Finally, a lack of inclusion of studies involving children and adolescents limits the generalizability of these findings to broader populations.

Lithium and clozapine have consistent data supporting protective effects against suicide in certain psychiatric disorders.