Estimating the risk of suicide associated with mental disorders: A systematic review and meta-regression analysis

1The Global Burden of Disease Study (GBD) estimated suicide to be the 15th leading cause of years of life lost (YLLs) in 2019, up from the 18th in 2010. Suicide is a major public health concern globally with nearly 800,000 deaths due to suicide worldwide. It also affects people across the life course – children, adolescents, as well as older adults. Of all suicide deaths, 58% occur between the ages of 15–49 years which indicates the large magnitude of potentially productive years of life lost. In older age, major life events and complex transitions may be precipitants for suicide. According to GBD 2019 results, death rates from suicide are much higher in older age groups than in younger ones (24·53 per 100,000 for ages greater than 70 and 14·25 per 100,000 between 50 and 69 years vs 11·19 per 100,000 between 15 and 49 years). There are also differences in sex-specific suicide rates, with the suicide death rate among males being consistently higher than that of females over time (13·5 per 100,000 in males vs 6·1 in females in 2019). Globally, India and China continue to report the highest number of deaths by suicide along with distinctly dominant suicide methods such as poisoning, hanging, and self-immolation. This suggests a continued need to assess underlying risk factors and potential target areas for suicide prevention interventions in low and middle-income countries.

2The risk of suicide attributable to mental disorders (MDs) is well established in the literature. Existing reviews indicate that suicide risk is particularly high at the time of diagnosis for affective disorders, and schizophrenia. Up to 80% of suicide deaths have been attributed to a mental or substance use disorder in high income countries compared to approximately 70% in low income countries. However, the World Health Organization (WHO) classification followed by the GBD study captures suicide related deaths under ‘Injuries’ as the direct (rather than underlying) cause of death, while MDs are grouped separately within ‘Non-Communicable diseases. To estimate the proportion of suicide related deaths due to mental and substance use disorders (MSDs), previous work by Ferrari and collaborators applied comparative risk assessment (CRA) methodology to GBD 2010 data. They obtained RR estimates from a review of the literature which were then pooled and combined with GBD 2010 prevalence estimates to generate population-attributable fractions (PAFs). In their review of this literature, Ferrari and collaborators indicated that 62·2% of the burden estimated for suicide in an earlier iteration of the GBD study could be attributed to MSDs. MSDs moved from the 5th to the 3rd leading class of disease burden once suicide YLLs were reassigned to MSDs.

3Given these findings, it is important to periodically re-assess the effect of MDs on premature mortality risk by updating the evidence base for the relationship between MDs and suicide. The work by Ferrari and Collaborators was conducted using GBD 2010 data and systematic literature reviews conducted in 2014. Analyses from this previous work largely comprised pooled RRs obtained using a simple meta-analytic approach. This approach did not account for sources of variation that may contribute to differences in reported suicide risk. These are important to consider in order to understand the impact of differences in study methods and subgroups on reported magnitude of suicide risk which, to date, remain largely unexplored. Existing reviews on this topic typically define broad inclusion criteria including vulnerable or high-risk subgroups. These may introduce measurement error and subsequently result in potentially inflated risk estimates. The analytic approach in these reviews also typically involve pooled analyses that do not fully adjust for sources of measurement error (e.g. study design, sampling strategy, etc.) or assess the impact of other ecological variables (e.g. region or individual characteristics). This current systematic literature review aims to account for these limitations by providing updated suicide risk estimates adjusted in a meta-regression for key demographic and methodological covariates that may contribute to variation in study-reported estimates. We use the earlier review by as the basis and template for this work. We expand the search strings to incorporate self-harm and self-injury to maximize the scope of potentially eligible studies. Relevant data from 2010 to 2019 updates this systematic review, this the most comprehensive review of this risk-outcome pair to the best of our knowledge. This review will also contribute to measurement of mortality that can be indirectly attributed to mental disorders.

4The MDs included in this paper were those identified previously as risk factors for suicide. They were defined according to the Diagnostic and Statistical Manual of Mental Disorders (DSM) or the International Classification of Diseases (ICD) diagnostic criteria. The mental disorders included were major depressive disorder (MDD), dysthymia, bipolar disorder, anxiety disorders, and schizophrenia. Suicide was defined as cases meeting criteria for ICD cause of death codes for intentional self-inflicted poisoning or injury.

5We expanded upon the 2014 systematic review by Ferrari and colleagues which covered the literature between 1966 and 2010 (See appendix). The search protocol adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and are described in the appendix. Three major scholarly databases, namely Medline, Embase, and PsychINFO were searched between January 2010 and June 2019 Studies were eligible for inclusion based on the following criteria:Considered MDs (as defined by DSM or ICD) as a risk factor associated with death by suicideReported a relative risk (RR) with 95% uncertainty, or provided sufficient information for these to be calculatedWere individual-level case-control or cohort studies defining a clear temporal association between the exposure (presence of mental disorder(s)) and outcome (suicide). Prospective cohort studies were considered the gold standard for inclusion. Studies using a cross-sectional design were excluded as this approach does not allow for a clear temporal association to be assessed.Study sample was representative of the general population of a given location (e.g.: community-based studies). Therefore, studies based on treated samples, linked psychiatric treatment registries, health insurance or claims data, or vulnerable populations (prisoners, pregnant women, veterans, homeless, etc.) were excluded.

6Considered MDs (as defined by DSM or ICD) as a risk factor associated with death by suicide

7Reported a relative risk (RR) with 95% uncertainty, or provided sufficient information for these to be calculated

8Were individual-level case-control or cohort studies defining a clear temporal association between the exposure (presence of mental disorder(s)) and outcome (suicide). Prospective cohort studies were considered the gold standard for inclusion. Studies using a cross-sectional design were excluded as this approach does not allow for a clear temporal association to be assessed.

9Study sample was representative of the general population of a given location (e.g.: community-based studies). Therefore, studies based on treated samples, linked psychiatric treatment registries, health insurance or claims data, or vulnerable populations (prisoners, pregnant women, veterans, homeless, etc.) were excluded.

10In addition to the electronic database search, experts in the field were consulted and recently published reviews on this topic were reviewed for any additional data sources. All data from earlier reviews were first re-assessed for eligibility and accepted only if they met criteria for this review. Relevant information on effect sizes, study methodology and sample characteristics were extracted and compiled for each study. References were managed and duplicates were removed via EndNote citation management software. MM conducted the data searches and data extraction. AF and DS re-assessed study eligibility as needed and reconciled any questions around study inclusion.

11A multi-level meta-regression approach was used to obtain pooled relative risk estimates (RRs) and 95% uncertainty intervals for each disorder. We chose a multi-level meta-regression approach as it is best suited for data that are inherently nested or clustered within larger groups, as was the case here. Most studies in our dataset reported multiple effect sizes by disorder, age and/or sex and therefore were not independent observations. The multi-level approach with random effect terms for studies and disorder groups allowed us to account for any within and between study dependencies in the data at these levels. An unstructured variance-covariance matrix was used for random effects variances. The details and applications of this method are described in detail elsewhere. Exploratory forest plots for disorder-specific RRs and variance-covariance matrices of models with different correlation thresholds are reported in the appendix.

12We made use of a series of covariates to quantify and adjust for (where appropriate) sources of heterogeneity in the data. A detailed explanation of covariates is provided in Table 1. The inclusion of covariates in the model occurred sequentially, starting with demographic variables known to be primary sources of variation (age, sex, disorder type, location); methodological variables (follow-up time, sampling type, response rate, study design, estimate adjustment); and the Sociodemographic Index (SDI) and the Health Access Quality Index (HAQI). Changes to effect-sizes due to the inclusion of new variables were reviewed at each step.

13Due to considerable variation in the data collection and analytical methodology used between studies, several analytical choices were made in order to maximize data use. If studies reported odds ratios for suicide, these were converted to RRs using methods described in the appendix. Data from categories such as ‘mental disorders’, ‘mood and psychotic disorders’ or ‘mood and alcohol use disorders’ were not used as these very broad categories were less informative than more specific disorder categories. We considered the use of a location-level covariate using either GBD super-region classificationsor World Bank income level groups. However, only 6 out of 20 studies were from low or middle-income settings, which may have limited statistical power to detect a significant effect in our analysis. For this reason, a location covariate was not used in the final model. Covariates in our final model included age, sex, disorder type, estimate adjustment, use of psychological autopsy method, and study design.

14In order to test if specific estimates had greater influence on pooled results compared to others, exploratory Baujat plots were generated. These plots display the squared residuals on the X-axis and influence on the model on the Y-axis. Any estimates that had high residuals or high influence were determined to be outliers. Studies with outliers were re-assessed and those determined to have limitations in quality and/or measurement methods were excluded. Sensitivity analyses were conducted by assessing if overall results changed significantly upon removal of outliers as observed from the Baujat plots (Appendix). The distribution of variance across the levels included in our model was calculated using methods described in Cheung et al. (2014). This provided the proportion of the overall variance attributed to each level. All meta-regression analyses were conducted using the metafor package in R v3·5·2.

15A total of 20 studies and 69 individual RRs were included in the final analysis (See Fig. 1). These studies covered 10 countries and 4 GBD regions. Anxiety disorders and MDD were the most commonly assessed MDs. (See Table 2 for a summary of selected studies). There was wide variation in reported sample sizes (84–68378) and follow-up time (1 year–64 years).

16Table 3 presents the meta-regression model coefficients and Table 4 presents the predicted RRs adjusted for select covariates. Disorder type, sex, age, study design, estimate adjustment, and use of psychological autopsy (PA) were tested in the model for suicide risk (Table 3). The amount of the overall variance attributed to random effects was 73.14%. We found no significant associations between suicide risk and response rate, follow-up time, sampling type, high vs low/middle-income locations, SDI, or HAQI. These variables were therefore not included in the final model. Our sensitivity analyses also showed no major change in significance and direction of estimates upon removal of outliers (4 observations – see appendix).

17Age was associated with an increase in suicide risk (Table 3). Among methodological covariates, the use of retrospective study design was associated with lower suicide risk compared to prospective study design. There was no significant effect of estimate adjustment or use of the psychological autopsy method.

18Suicide risk was significantly associated with the presence of MDD, anxiety disorders, bipolar disorder, dysthymia, and schizophrenia (Table 3). Predicted RRs were the highest for MDD, bipolar disorder, and schizophrenia (Table 4). RRs for females were consistently lower than RRs for males across all disorders (although this association was not significant).

19RRs adjusted for sex, age, disorder, study design, estimate adjustment, and psych-autopsy method.

20Our results confirm that MDs are important risk factors for suicide. This association remained after adjusting for methodological factors. The earlier review provided pooled estimates of suicide risk for each disorder obtained via a random effects meta-analytic approach. This updated review improves upon this approach by using a multi-level meta-regression model with predictive covariates that can better utilize (in its nested approach) all the data available across studies and more robustly explain the heterogeneity in the data reported between studies to separate measurement error from variability that is ‘real’ and due to ecological variables.

21Our findings showed that MDD, dysthymia, anxiety disorders, bipolar disorder, and schizophrenia were significantly associated with suicide risk. The selection of disorders attaining statistical significance, are consistent with meta-analytic results from other studies exploring suicide mortality due to MDs (Baxter et al., n.d.;;; ). An earlier review on anxiety as a risk factor for suicide did not find anxiety disorders to be an important predictor of death by suicide. This may have been due to the inclusion of studies that used symptom scales to assess anxiety disorders – which are best used for screening or symptomatic evaluation rather than provide a clinical diagnosis. Secondly, the inclusion of any individual anxiety diagnosis may have attenuated the pooled risk of suicide. Reviews may also focus only on specific disorders. Although these provide an important profile of disorder-specific suicide risk, they exclude psychiatric comorbidity - an important contributor to and perhaps catalyst for suicide risk. Reviews may also draw from national treatment registers for data from samples with mental disorders. National treatment registers are a useful source of data on mental disorders often drawn from large samples and allow for a more efficient study of rare outcomes such as death by suicide. However, it is important to consider that this mode of data collection may also be inherently biased as it does not capture mental disorder cases among those who may not have made contacts with any health service. It is likely that register-based diagnoses may reflect a higher severity of mental disorders than what may be present in the general population. Therefore, the use of registry-based data is expected to be associated with an increase in or overestimation of pooled estimates of suicide risk. Some reviews also examine suicidal behavior (such as suicide attempts and/or self-injury with an intent to die). A recent review by Gili and colleagues examine the evidence for mental and substance use disorders as risk factors for suicidal behavior (suicidal acts and attempts) among young people. Although this review provides important insights into the risk for suicide attempts that frequently precede suicide deaths in younger populations, the analyses use ORs instead of RRs and use broad groups of disorders which may limit interpretation for disorder-specific risks especially for depressive disorders which are known to account for a large proportion of suicide risk. It is important to highlight that the disorder-specific pooled RRs obtained here are considerably lower than those found in the earlier review by. The use of demographic, methodological, and location-level covariates may have contributed to this difference. We adjusted for study-level random effects using a hierarchical approach that is also likely to yield lower adjusted RRs for suicide compared to a standard meta-analytic approach that has typically been used to generate pooled effect sizes in the past. We also used stricter inclusion criteria and data prepping procedures than what has been used in the existing literature. Furthermore, meta-analytical studies will often use odds ratios as an approximation for RRs if the prevalence of suicide is known to be very low. This may well overestimate RRs and lead to higher effect sizes we usually see in other studies. This study attempts to take suicide risk measurement a step further by including covariates that account for measurement error. Although uncertainty in the model remains, we believe that the estimates are more robust as a result of this.

22In our model, older age was associated with higher RRs. This is consistent with findings in the literature where older age is likely to be associated with an increased risk of suicide (. Among methodological variables, the use of a retrospective study design was associated with lowered suicide risk. It is possible that a prospective study design is better at establishing a temporal association between mental disorders and suicide which may not be possible in a retrospective design. Additionally, studies using a retrospective design may not be able to capture the impact of confounders on mental disorders and suicide as accurately as those using a prospective design.

23We found no significant associations between the use of psychological autopsy and suicide risk. This may be due to the low number of non-psych autopsy studies that were included in the analyses. Overestimation or underestimation of suicide risk in psych-autopsy studies may be due to biases in how psychopathology is measured, specifically the choice of informants, choice of comparison group, and lowered event recall - especially if the assessment is conducted much later than the date of suicide. Although our study did not empirically assess these components beyond a dichotomous measure of psychological autopsy use, it is reasonable to assume that they may contribute to the heterogeneity in reported suicide risk beyond what we found in our analyses.

24Adjusted RRs were not significantly different from unadjusted RRs in our study. However, it was important to include this covariate in our model since RRs that controlled for known potential confounders were likely be more accurate in estimating true suicide risk compared to crude RRs. A possible reason for not detecting a significant effect may be the low number of studies adjusted RRs. There is considerable variation in the number and type of confounders controlled for between studies. Some studies may control for a single confounder (such as age, or educational attainment). Other studies may control for a larger number and range of confounders such as psychiatric comorbidity, family history of MDs, socioeconomic status, physical illnesses, and others. Although our current analyses did not empirically assess the impact of this variation on suicide risk, it is plausible that studies that control for as many potential confounders as possible likely report estimates of a higher quality and reliability compared to studies that do not.

25We were unable to empirically assess location-level differences in RRs due to the lack of sufficient studies from LMICs. Out of 20 studies, 14 were from high-income locations - mostly Western Europe and North America. Given the lack of income level-specific estimates across disorders, the uncertainty across the results produced is greater and interpretability of the resulting sex- and region-specific RR are lowered. The sparsity of data from LMICs may be due to variation and constraints in country-level reporting of suicide deaths. Suicide continues to be a criminal offence in many countries in Africa, Asia, and South America, which potentially leads to systematic underreporting of suicide cases. The global data coverage for MDs is also low for LMICs. Better quality data from LMICs along with more sex-specific data on MDs and suicide may allow us to make more substantive conclusions about regional and sex-specific patterns in RRs for suicide.

26This systematic review provides a useful insight into the recent available evidence on the risk of suicide among those with versus without MDs and the heterogeneity that may exist in the data based on study-specific attributes. Our findings of MDs being significant predictors of suicide risk are largely consistent with findings from other individual-level and meta-analytic studies. Based on the GBD study, MDs are the some of the leading causes of non-fatal disability (, 2013). However, very few MDs are attributed as direct causes of suicide despite overwhelming evidence in the literature. Findings from this study may better inform the burden from premature mortality directly attributed to MDs in GBD.

27These findings also support the need for adequate mental healthcare as an important suicide prevention measure. Research using World Mental Health surveys shows that a large treatment gap exists for those exhibiting suicidal ideation and behaviors. Only 39% of those with suicidality receive any form of treatment worldwide. This proportion is even lower in low-income countries where only 17% of those with suicidality receive any form of treatment. Although there exists evidence in support of select interventions for suicide prevention, improving prevention efforts at earlier stages of care may increase the likelihood of fulfilling this unmet need for care. These findings also emphasize the continued need for primary interventions for suicide prevention that focus on adequate screening for and detection of MDs in different contexts such as community, primary care, and school-based settings. However, precision in MD detection is dependent on the screening tool and classification thresholds used - particularly in varying cultural contexts. Arango and colleagues define secondary and tertiary prevention efforts in addition to primary prevention efforts in their review. Secondary prevention efforts that focus on reducing the incidence of MDs and tertiary prevention efforts that treat those with established MDs are important as they address the role of MDs in suicide prevention. A systematic review on suicide prevention strategies found that in addition to direct suicide prevention strategies such as restricting access to lethal means, pharmacological and psychological treatment of MDs are also important in suicide prevention.

28Several limitations constrained our analyses. First, the geographic representativeness of our analysis is limited. Although our inclusion and exclusion criteria ensured the selection of high quality studies, these were mostly from high-income countries in North America and Western Europe. Therefore, studies from low and middle-income countries are under-represented in this analysis indicating the need for more published data from these regions. Second, we were unable to provide estimates for additional psychiatric disorders such as eating disorders due to the lack of studies that fulfilled our selection criteria. Third, we acknowledge that the uncertainty intervals for predicted RRs were at times wide and overlapping between disorders due to lack of data for some disorders and by income-level. Fourth, analyzing data for all available MDs in a single meta-regression model may either attenuate or amplify the magnitude of disorder-specific RRs. Fifth, the standard meta-regression framework allows for uncertainty intervals of predicted RRs to capture error attributed to fixed effects only and not between-study heterogeneity. Sixth, the scope of this analysis was confined to measuring the risk of suicide deaths only. While suicidal ideation and attempts are important precursors to death by suicide, our study did not explore these outcomes as they may have different risk profiles compared to completed suicides. Lastly, although we tried to quantify as many methodological sources of variation as possible, we were unable to account for all sources of measurement error. We excluded methodological covariates such as sampling type, follow-up duration, and response rate as there was insufficient power to detect an effect.

29MDs are associated with an increased risk for suicide. This synthesized analysis shows that the magnitude of study-reported risk may vary depending on study methodology and disorders examined. Therefore, study quality and choice of methods are important to consider in public reports on aggregated evidence. More research is needed to better quantify evidence for MD burden and suicide risk with higher precision in low and middle-income settings. These efforts will assist policy makers in framing evidence-based suicide prevention strategies and make mental health care an important part of their agendas.

30This work was supported by the 10.13039/100000865Bill and Melinda Gates Foundation, Queensland 10.13039/501100003921Department of Health, and the 10.13039/501100000925National Health and Medical Research Council Early Career Fellowship Grant APP1121516.

31The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

32Modhurima Moitra: Conceptualization, Methodology, Software, Formal analysis, Data curation, Writing – original draft, Visualization. Damian Santomauro: Conceptualization, Methodology, Software, Validation, Formal analysis, Writing – review & editing, Supervision. Louisa Degenhardt: Resources, Data curation, Writing – review & editing. Pamela Y. Collins: Writing – review & editing, Supervision. Harvey Whiteford: Writing – review & editing. Theo Vos: Writing – review & editing. Alize Ferrari: Conceptualization, Methodology, Validation, Writing – review & editing, Supervision.

33We declare no competing interests.