For many decades, preventing death from suicide has been an important focus in mental health research. Hence, researchers have aimed to find causal risk factors, which could inspire improvements of suicide intervention strategies. In this issue of the British Journal of Psychiatry, Harrison et al^{Reference Harrison, Munafo, Davey Smith and Wootton1} examined whether tobacco smoking is a causal risk factor for potentially fatal behaviours such as suicidal ideations or attempts. Their aim was to triangulate the available evidence by testing for observational associations and causality. The latter was done with genetically informed approaches, including Mendelian randomisation and single nucleotide polymorphism (SNP) logistic regression models. Mendelian randomisation is a technique where SNPs are used as instrumental variables to investigate causal directionality of associations, using observed data. There are many important assumptions to consider when conducting Mendelian randomisation, which are explained in the article. Triangulating evidence beyond observational results by using genetic variants decreases the likelihood of residual confounding, reverse causality and thus improving causal inference.

Harrison et al observed cross-sectional associations of all self-reported smoking behaviours with suicide attempts and ideation in 45 825–109 688 UK Biobank participants.^{Reference Harrison, Munafo, Davey Smith and Wootton1} Analyses were adjusted for age, gender and socioeconomic status. The strongest association was observed between smoking initiation and suicide attempt (odds ratio, 2.07; 95% CI 1.91–2.26; P < 0.001). Various methods of Mendelian randomisation were conducted to further explore the causal direction of this phenotypic cross-sectional relationship. They performed Mendelian randomisation with summary-level data-sets from published genome-wide association studies (using five complementing methods), individual-level data from UK Biobank and single SNP analyses (using an SNP in the CHRNA5-A3-B4 gene cluster, known to be strongly associated with heaviness of smoking). Mendelian randomisation approaches using summary-level results revealed some evidence for a causal effect of smoking initiation on suicide attempts. Subsequently, individual-level Mendelian randomisation or single SNP regression models did not yield any evidence for a causal association.

Although there was support for a phenotypic relationship between smoking behaviours and suicidality, genetically informed methods yielded no strong support for a causal effect of smoking on suicide attempt or suicidal ideation. Furthermore, >50% of the SNPs in the genetic instrument explained more variance in the outcome (i.e. suicide attempts) than in smoking initiation (see Supplementary Table 3 in Harrison et al^{Reference Harrison, Munafo, Davey Smith and Wootton1}), which increases the likelihood of reverse causation. Follow-up analyses suggested that risk-taking behaviour, which is well known to affect both substance use and (auto-)aggressive behaviours,^{Reference Fazel and Runeson2} might underlie the association between smoking and suicidality. This bidirectional relationship between smoking and risk-taking challenges the value of using this genetic instrument as a proxy for smoking initiation. Several previous Mendelian randomisation studies on smoking behaviours have included similar genetic instruments as the current study.^{Reference Vermeulen, Wootton, Treur, Sallis, Jones and Zammit3} It would be valuable to place more emphasis on variable definition and the validity of a genetic instrument in future Mendelian randomisation studies before drawing firm conclusions on causality. There is a great need for reporting guidelines for Mendelian randomisation studies to improve the quality of the evidence.^{Reference Burgess, Davey Smith and Davies4}

Studying causal risk factors for suicide and self-harm is notoriously complex, and treatment trials aimed at modifiable risk factors have been underpowered to unambiguously demonstrate a reduction of the incidence of suicide and self-harm.^{Reference Fazel and Runeson2} The present study by Harrison et al is no exception: tobacco smoking had no causal effect on suicidal ideation and self-harm.^{Reference Harrison, Munafo, Davey Smith and Wootton1} Instead of focussing on modifiable risk factors, perhaps suicide prevention research should shift from causal frameworks to prediction. Unfortunately, although suicide prediction models have overall very good accuracy, their predictive validity of correctly classifying a suicide death is near zero, i.e. extremely poor.^{Reference Belsher, Smolenski, Pruitt, Bush, Beech and Workman5} This can largely be explained by the low base rate of suicide death in the general population. Potentially, more common phenotypes across the suicide/self-harm continuum, such as suicidal ideation and (intentional or non-intentional) self-harm, have greater potential to assess in prediction studies. It is, however, important to acknowledge here that there is a clear lack of effective evidence-based interventions to treat patients at risk of suicide and repeated self-harm,^{Reference Fazel and Runeson2} raising important ethical questions warranting further exploration.

Although reducing tobacco smoking across the population will probably not result in a reduction of self-harm or suicide, tobacco should nonetheless still be treated and discouraged, to stop the tobacco epidemic causing 8 million deaths a year worldwide. Individuals who survive self-harm are at elevated risk for many poor outcomes later in life, not only death from suicide, including serious psychopathology and repeated self-harm. Future efforts should focus on how to identify vulnerable individuals and offer them the proper mental healthcare. Developing evidence-based interventions following self-harm should be a top priority.