Table 5.

Most common instrumental variable assumptions in mendelian randomisation and examples of possible assessments or sensitivity analyses

Assumptions	Examples of possible assessments
Relevance: genetic variants are associated with the exposure of interest	Report F statistic
Independence: genetic variants share no unmeasured cause with the outcome	Report the associations of plausible confounders with both the genetic variant(s) and the outcome; how population stratification has been taken into account (e.g. through principal component adjustment); and unmeasured confounding sensitivity metrics for the variant-outcome association59 68 69 75
Exclusion restriction: genetic variants do not affect the outcome except through their potential effect on the exposure of interest	Report results from MR Egger regression slope estimate as well as the intercept and its 95% confidence intervals; and results using negative control outcomes or negative control populations
Homogeneity (two stage least squares): there is a constant causal effect of the exposure of interest on the outcome	Report the instrumental variable effect estimate for different measurable subpopulations (eg, stratified by age, race or ethnic origin, sex, or socioeconomic status); for continuous outcomes, report on variance by level of instrument72
InSIDE (MR Egger): associations of genetic variants with the exposure variable must be independent of its direct effects on the outcome	Also report the effect estimates from other estimators that do not require this assumption (eg, median and modal based tests).