Understanding the educational inequalities in suicide attempts and their mediators: a Mendelian randomisation study

Data sources

This MR study used publicly accessible summary-level data from GWASs. Since no additional individual-level data were used, ethical approval and informed consent were not required. The data selection process was governed by a set of predefined criteria. We prioritised summary statistics that (1) originated from the largest GWASs available, conducted on studies composed entirely or predominantly of individuals of European ancestry; (2) were derived from populations with balanced gender distributions, thereby excluding sex-specific phenotypes such as breast cancer; (3) were not adjusted for heritable covariates, such as body mass index, in order to circumvent collider bias; (4) represented continuous variables rather than dichotomised forms (eg, the continuum of blood pressure vs the binary classification of hypertension), to uphold the suitability of the MR assumptions and enhance the statistical power. In most GWAS data sets, genetic associations were adjusted for age, sex and principal components of genetic ancestry. Detailed statistical descriptors for each phenotype are presented in online supplemental table 1.

Educational attainment

In this study, educational attainment was the focal exposure. Summary statistics for this exposure were sourced from the latest GWAS meta-analysis conducted by the Social Science Genetic Association Consortium (N=3 037 499).11 The educational attainment phenotype was constructed by mapping the highest level of education attained by individuals at a minimum age of 30 years, congruent with the International Standard Classification of Education System (1997). This classification facilitated the conversion to an equivalent number of years of full-time education. The mean educational duration was 15.4 years with a standard deviation (SD) of 3.4 years.

Suicide attempts

This study examined suicide attempts, the most important predictor of completed suicide, as the primary outcome of interest. To minimise potential bias due to sample overlap, particularly given that measures of educational attainment and some proposed mediators were partially or entirely derived from the UK Biobank cohort, we utilised GWAS summary statistics for suicide attempts that did not include the UK Biobank data. Summary statistics were collated from two independent GWASs: the Lundbeck Foundation Integrative Psychiatric Research (iPSYCH) based in Denmark (6024 cases and 44 240 controls)12 and the FinnGen conducted in Finland (8978 cases and 368 299 controls).13 Within the iPSYCH data set, the case definition was contingent on clinical diagnoses of suicide attempts as codified by the 10th edition of the International Classification of Diseases (ICD-10) criteria (X60-X84) through the Danish Psychiatric Central Research Register. A broader inclusion criterion was adopted, wherein individuals presented with a primary mental disorder diagnosis (ICD-10: F chapter) in conjunction with secondary diagnoses indicative of self-harm, poisoning by pharmaceutical substances or other means or injuries localised to the hand, wrist or forearm (ICD-10: T36-T50, T52-T60, S51, S55, S59, S61, S65, S69). Conversely, FinnGen identified cases exclusively based on suicide attempt diagnoses (ICD-10: X60-X84) via national medical registries. The control groups in both studies were population-based participants with no recorded instances of suicide attempts.

Mediators

The identification of putative mediators was guided by four stringent criteria: (1) evidence from the extant literature or scientific rationale should suggest that the mediator contributes to the causal pathway between educational attainment and suicide; (2) the mediator should be modifiable through lifestyle modifications or clinical interventions; (3) the mediator should be sufficiently widespread to imply significant public health relevance in the context of suicide and (4) accessible and well-powered GWAS summary statistics meet the standards necessary for inclusion in the analysis. Following these criteria, we identified 42 candidate mediators that were subsequently stratified into four domains: 11 socioeconomic and behavioural factors, 11 biological indicators, 7 mental health disorders and 13 physical health diseases (online supplemental table 1). A comprehensive discussion of the selected mediators, including their respective data sources, is provided in online supplemental methods 1.

Statistical analysis

Statistical analyses were performed using the ‘TwoSampleMR’, ‘MendelianRandomization’, ‘MRPRESSO’ and ‘RadialMR’ packages within the R V.4.1.1 software.

Instrument selection

In adherence to conventional norms for instrument selection, our study used a genome-wide significance threshold (p<5×10⁻⁸) to extract single-nucleotide polymorphisms (SNPs) with a minor allele frequency >0.01. The extracted SNPs were subjected to a clumping procedure to ensure allelic independence by linkage disequilibrium (r²<0.001, with a clumping distance of 10 000 kb), using the European reference panel. To align the SNP effects between the exposure and outcome data sets and avoid strand mismatches, we executed a harmonisation process. The strength of the instruments was appraised using the F-statistic, where values >10 denoted a sufficiently strong instrument.

Primary analysis

The Wald ratio was computed for individual SNPs to estimate the unbiased effect of exposure on the outcome. Wald ratios for singular SNPs were combined using a random-effects inverse variance-weighted (IVW) approach, applicable to both univariable and multivariable MR analyses.14 Under the assumptions of instrument validity and balanced horizontal pleiotropy, the IVW method yielded the most precise and robust estimates. The random-effects model further accounted for the inherent heterogeneity of the instruments.

Mediation analysis

Mediation analysis progressed with the inclusion of candidate mediators that conformed to the pre-established selection criteria (figure 2). Within this mediation paradigm, the total effect of exposure (educational attainment) on the outcome (suicide attempts) was divided into direct (unmediated) and indirect (mediated) components (figure 1). We employed a two-step MR strategy (also known as the product of the coefficient approach) for the individual assessment of the mediators.15 The first step was to estimate the effect of educational attainment on the mediator via univariable MR. The second step was to evaluate the effect of the mediator on suicide attempts, adjusting for educational attainment, within a multivariable MR model. The indirect effect was deduced by multiplying the estimates from both analytical steps. To ascertain the collective effect of mediators, a multivariable difference-in-coefficients (MR) method was conducted, where the direct educational effect was separated by adjusting for all mediators under consideration. The mediated proportion was determined as the ratio of the indirect effect to the total effect. The 95% confidence intervals (CIs) were calculated using the propagation-of-error method.

Request permission

Sensitivity analysis

To assess the potential bias due to horizontal pleiotropy, a spectrum of alternative univariable MR methods was employed, including the weighted median estimator, weighted mode estimator and MR-Egger regression.14 For the same purpose, the median-based and MR-Egger approaches were extended to a multivariable MR analysis. These methods are valid under certain assumptions. Cochran’s Q statistic was used to evaluate instrument heterogeneity, and the MR-Egger intercept was calculated to detect the presence of horizontal pleiotropy. The pleiotropy residual sum and outlier (MR-PRESSO) outlier test, radial MR and leave-one-out analysis were used to identify and mitigate the influence of outlying SNPs.14

Given the substantial correlation among education, intelligence and cognitive performance, we investigated the independent effects of educational attainment on suicide attempts. Multivariable MR was implemented with adjustments for intelligence and cognitive performance metrics, individually and in combination. Summary statistics for intelligence were sourced from a GWAS meta-analysis of 14 cohorts (N=2 69 867).16 Summary statistics for cognitive performance were obtained from a meta-analysis of the UK Biobank and Cognitive Genomics Consortium (N=2 57 841).17

Third, Steiger filtering was applied to validate the directionality of the posited relationships.14 This technique eliminates instrumental SNPs that exhibit greater variance in outcomes relative to exposure. Additionally, reverse-direction MR analyses were conducted using instrumental variables as mediators to elucidate putative causal effects on educational attainment. To assess the causal influence of suicide attempts on mediators and educational attainment, we constructed instruments for suicide attempts using SNPs satisfying a less stringent significance threshold (p<1×10⁻⁶), with a similar procedure for clumping because of the absence of genome-wide significant SNPs associated with suicide attempts.12 A comparable relaxation of the significance threshold was applied to anxiety and other candidate mediators encountering a paucity of SNPs reaching genome-wide significance to boost the statistical power. Additionally, an MR Robust Adjusted Profile Score method was introduced to mitigate the potential weak instrumental bias arising from the utilisation of the relaxed significance threshold.

Finally, to recognise the substantial sample overlap between GWASs of educational attainment and specific mediators (such as smoking) attributable to UK Biobank contributions, we employed genetic associations ascertained from an earlier GWAS of educational attainment, one that did not include UK Biobank data, to facilitate validation.18

Interpretation of findings

This study’s reporting and interpretation align with the Strengthening the Reporting of Observational Studies in Epidemiology using MR guidelines.19 The results are reported as odds ratios (ORs) with 95% CIs for each SD change in continuous exposure (such as educational level) and per unit change in the log odds scale for dichotomous exposure (such as depression). Considering the traditional overreliance on p values for interpretive validity, our study incorporates a tripartite evaluative framework: (1) association robustness: assessment of the association’s statistical significance and its replicability; (2) MR assumption suitability: scrutiny for minimal evidence of horizontal pleiotropy and (3) directional consistency: uniformity of effect directions across different studies and analytical methods.

To explore the relationship between educational attainment and suicide attempts, associations that achieved a nominal significance level (p<0.05) via the IVW approach in both the iPSYCH and FinnGen studies and satisfied the latter two evaluative criteria were considered credible. For the 42 putative mediators, we heightened the stringency of the first evaluative criterion: associations were required to exceed the Bonferroni-corrected significance threshold (p<1.19×10⁻³) in at least one study and simultaneously attain nominal significance (p<0.05) in another.

Instrument statistics

In total, 602 independent SNPs were identified as instrumental variables for educational attainment, explaining 2.78% of the variance. The mean F-statistic was 83 (range, 28–576). Detailed statistics for the genetic instruments corresponding to the mediators are presented in online supplemental table 2.

Effect of educational attainment on suicide attempts

Employing univariable MR to assess the total effect, the IVW method results indicated that each SD decrease (equivalent to 3.4 years) in genetically predicted educational attainment was associated with a 1.93-fold (95% CI: 1.60 to 2.33; p<0.001) higher risk of suicide attempts in the iPSYCH study and 2.13-fold (OR: 2.13; 95% CI: 1.82 to 2.49; p<0.001) in the FinnGen study. The combined analysis of these studies showed an OR of 2.05 (95% CI: 1.81 to 2.31; p<0.001) for suicide attempts.

Mediation analysis

Following stringent selection processes, five mediators from the initial 42 candidates were incorporated into the final stage of the mediation analysis (figure 2). In univariable MR analysis, genetically predicted lower educational attainment was associated with increased smoking exposure (beta: 0.40; 95% CI: 0.37 to 0.43; p<0.001), a greater number of chronic pain sites endured (beta: 0.30; 95% CI: 0.27 to 0.33; p<0.001) and higher risks of insomnia (beta: 0.38; 95% CI: 0.32 to 0.44; p<0.001), depression (beta: 0.42; 95% CI: 0.37 to 0.47; p<0.001) and anxiety (beta: 1.04; 95% CI: 0.86 to 1.21; p<0.001) (figure 3). These five mediators demonstrated consistent causal effects on suicide attempts using data from the iPSYCH and FinnGen studies. The combined analysis yielded ORs for suicide attempts of 2.80 (95% CI: 2.31 to 3.40; p<0.001) for increased genetically predicted smoking exposure, 2.41 (95% CI: 1.57 to 3.69; p<0.001) for a higher number of chronic pain sites, 1.81 (95% CI: 1.64 to 1.99; p<0.001) for genetic liability to insomnia, 2.51 (95% CI: 2.22 to 2.84; p<0.001) for genetic liability to depression and 1.61 (95% CI: 1.01 to 2.56; p=0.044) for genetic liability to anxiety (figure 3).

Request permission

In the combined analysis of iPSYCH and FinnGen, both smoking (proportion mediated: 47%; 95% CI: 29% to 66%) and depression (proportion mediated: 49%; 95% CI: 36% to 61%) alone mediated nearly half of the effect of educational attainment on suicide attempts, followed by anxiety (proportion mediated: 35%; 95% CI: 12% to 59%), chronic pain (proportion mediated: 36%; 95% CI: 0% to 84%) and insomnia (proportion mediated: 26%; 95% CI: 18% to 34%) (figure 4). In multivariable MR, these five mediators together explained 68% (95% CI: 40% to 96%) of the total effect of educational attainment on suicide attempts.

Request permission

Sensitivity analysis

The primary IVW results were consistent with the alternative pleiotropy-robust methods, both in terms of the direction and magnitude of the effect, except for MR-Egger regression because of its lower efficiency and susceptibility to outliers (online supplemental tables 3 and 4). There was significant heterogeneity in most analyses, as suggested by Cochran’s Q statistic; however, the MR-Egger intercept did not indicate evidence of horizontal pleiotropy (online supplemental table 5). Leave-one-out analysis revealed no specific SNP that significantly affected the results (online supplemental table 5). After excluding outliers detected through the MR-PRESSO outlier test and radial MR, the results remained consistent, and heterogeneity was no longer observed (online supplemental table 6). After controlling for intelligence and cognitive performance via multivariable MR, the effect of lower genetically predicted education on the risk of suicide attempts remained largely unchanged (online supplemental table 7). Steiger filtering yielded similar results after controlling for possible reverse causation (online supplemental table 8). In reverse-direction MR, there was evidence of a possible effect of genetically predicted smoking on education (online supplemental table 9), and this association might be driven by directional pleiotropy (p for MR-Egger intercept <0.001). The results also indicated that genetic liability for depression, chronic pain and suicide attempts could lead to lower educational attainment, with no indications of pleiotropy. Comparable results were obtained using summary statistics from a smaller educational GWAS without the UK Biobank (online supplemental table 10).

Main findings

Using a methodologically robust MR framework to enhance causal inference, this study indicated that a decrement of 3.4 years in educational attainment is causally associated with an approximately two-fold increased risk of suicide attempts. Our study also elucidates that smoking, chronic pain and mental health disorders collectively mediate a substantial proportion, estimated at 70%, of the association between educational attainment and the risk of suicide attempts.

Previous MR analyses utilising the iPSYCH data set have consistently reported that for each SD decrease in educational attainment, the risk of suicide attempts nearly doubled .20 21 This study corroborates these findings by employing genetic instruments for educational attainment derived from the most comprehensive GWAS to date, which are approximately three times the size of those used in previous studies and fortified by additional data procured from FinnGen. In contrast to the pronounced impact of educational attainment, our study does not substantiate the putative role of other socioeconomic indices such as household income and occupational class. These results partially support the observation that a low educational level is a more robust predictor of suicidal behaviour and broader health outcomes than income level.22 23 Educational attainment, reflective of the material and intellectual resources accrued from one’s family of origin and typically acquired during the early stages of life, is moderated by accessibility to education and intellectual capacity and potentially exerts influence on an individual’s cognitive abilities, thus shaping opportunities to adopt health-enhancing behaviours. Significantly, our multivariable MR analysis showed that the link between genetically predicted educational attainment and suicide attempts was not mitigated on adjustment for intelligence and cognitive performance, indicating the presence of mechanisms that act independently of these intermediary phenotypes.

To date, few studies have attempted to dissect the potential mediating pathways between low educational attainment and elevated suicide risk. Taylor et al, utilising data from the Australian National Survey, found that the prevalence of suicide attempts diminished marginally (from 2.9% to 2.5% in men and from 4.3% to 3.6% in women) within the lowest category of educational attainment after adjusting for psychiatric disorders.24 Using data from a representative health survey in Korea, Ki et al deployed a structured equation model to appraise the mediating roles of problematic drinking and mental and physical disorders.25 Their findings indicated that the impact of educational attainment on suicide attempts was mediated through physical illnesses (proportion mediated: 26%) and problem drinking (proportion mediated: 2%) but not through mental disorders characterised by anxiety or depression. Both studies employed cross-sectional designs, which inherently limited causal deductions because of temporal ambiguity. Our MR investigation supported the possibility of a bidirectional causal relationship between educational attainment and chronic pain, depression and suicide attempts. In a previous MR study, Rosoff et al reported that the magnitude of the association between genetically predicted education and suicide attempts remained largely unchanged in a multivariable MR framework, even after adjusting for smoking, alcohol consumption and psychiatric disorders.20 Nonetheless, that study was limited by its low power and did not formally implement a mediation analysis to quantify the mediating effects. In contrast, the current study leveraged the largest GWAS data sets available to scrutinise the mediating effects of a broad array of risk factors by employing rigorous MR methodologies, including both two-step MR and multivariable MR, to facilitate a comprehensive mediation analysis. Our findings suggest that smoking, chronic pain, insomnia, depression and anxiety contribute to over a quarter of the risk of suicide attempts, which is attributable to lower educational attainment.

In the present analysis, approximately 30% of the educational effect on suicide attempts remains unexplained. While alcohol consumption is a significant risk factor implicated in suicide26 and has been posited as a potential explanatory mechanism underlying the educational influence, the current study yields no substantial evidence to support the causal relationship between educational attainment and the total amount of weekly drinks. This is in alignment with a preceding MR study.27 Given the complex and diverse correlations between educational attainment and various patterns of alcohol consumption as well as alcohol-related consequences,28 the mediation effect by alternate alcohol use behaviours, such as binge drinking or alcohol dependency, remains a viable hypothesis that necessitates further investigation. Moreover, educational attainment has been connected to a spectrum of clinical outcomes that may increase the risk of suicide.21 The current study suggests that, except for chronic pain, no clinical markers (including blood pressure, lipidemic profiles, glycaemic traits, thyroid hormones and inflammatory cytokines) or physical illnesses (neoplastic, cardiometabolic, chronic respiratory and neurological diseases) demonstrate significant causal relationships with suicide attempts. These null results contrast with those of numerous observational studies, which may have been limited by uncontrolled confounders such as poor socioeconomic conditions, long treatment histories and the presence of comorbid mental health conditions and pain.29

Limitations

This study has several limitations. First, as with all MR analyses, avoiding the influence of pleiotropy is a salient challenge. Although multiple sensitivity analyses yielded consistent results and the MR-Egger intercept indicated minimal pleiotropic bias, the possibility of underlying pleiotropy cannot be completely excluded, particularly given the predominantly uncertain biological functions of the instrumental SNPs involved. Furthermore, the application of MR in mediation analyses presupposes additional assumptions such as the absence of interaction between the exposure and mediator, the absence of time-varying effects of the exposure and mediator and the linear impact of the exposure or mediator on the outcome.15 By nature, these assumptions are untestable with summary-level data, potentially rendering mediation estimates susceptible to bias. Despite this, two-step MR continues to serve as a credible approach for assessing the causal null hypothesis in the context of mediation analysis. The study also has the limitation of statistical power, particularly for certain mediators such as physical activity, where genetic variants explain only a minimal proportion of the variance. The possibility that some mediators have small-to-moderate effects on suicide attempts cannot be excluded. Additionally, the identification of suicide attempts via hospital records may omit cases that did not seek or receive medical care. Furthermore, despite known sexual dimorphism in suicide, the study design did not permit sex-stratified analysis because of the unavailability of sex-specific GWAS data. Similarly, the effect of age could not be distinguished. Finally, the study was based on data from individuals of predominately European ancestry, which limits the generalisability of the findings across diverse ethnic groups. This is exemplified by a quasi-experimental study indicating that higher educational attainment confers protection against suicide attempts in non-Hispanic White populations but not in Black populations.30

Implications

Historical policies aimed at prolonging mandatory education have yielded health improvements, and such efforts must be perpetuated. However, direct interventions targeting education may pose challenges without the accompanying social and political reforms. The findings of this study suggest that strategies to mitigate disparities in suicide attributable to lower educational levels should prioritise the reduction of smoking prevalence and management of chronic pain and mental health disorders. It is imperative that future research endeavours aim to identify additional mediators and elucidate the interactions among them to effectively address and reduce social disparities.