No evidence for heterogeneity in smoking-depressive symptoms associations across different smoking strata

Prior research indicates a causal effect of smoking on depression liability, yet the shape of this association remains unexplored. Distinguishing linear from non-linear effects is useful, as each suggests different intervention recommendations. A recent development is non-linear Mendelian Randomization (MR), where analyses are performed in strata with different average exposure levels. However, non-linear MR methods have produced biased estimates under different scenarios. We explore the causal effect of pack years of smoking on depression by performing linear MR and doubly-ranked non-linear MR in a sample of UK Biobank participants. We use two genetic instruments (for cigarettes per day and lifetime smoking) and incorporate sensitivity analyses and positive and negative control outcomes to evaluate potential bias. We found significant associations of genetically predicted packyears with sex and age across both linear and non-linear analyses for some strata, potentially reflecting participation bias. Only for the lifetime smoking instrument, stratified analyses revealed significant causal effects on depression in ‘smoking-exposed strata’. However, no single analysis was ideal, as there were associations with negative control outcomes in some strata. We emphasize the necessity of multiple sensitivity checks and control analyses when interpreting non-linear MR results. A particular concern is results from smoothing methods (fractional polynomials) when estimates are much more precise in some strata than others, dominating the shape of the estimate. Overall, our findings provide some evidence supporting smoking as a risk factor for depression, but no strong evidence for deviation from linearity.