| Relevance assumption: instrumental variable is strongly associated with the exposure |
Measured genetic variants associated with exposure act as presumed instrumental variable Plausibility assessment: F-statistic, risk difference or using genetic variants found by large-scale GWAS to be associated with the exposure Weak instrumental variable bias
|
All unmeasured familial causes specific to the exposure act as instrumental variable; co-twin’s exposure variable used as a proxy Plausibility assessment: within-pair correlation in the exposure All familial causes specific to the exposure are stronger than a finite number of genetic variants associated with the exposure
|
| Independence assumption: instrumental variable is independent of any confounder of the relationship between exposure and outcome |
Validity assessment: biological knowledge, test between genetic variants and confounders, evidence from GWAS of confounders Approaches for invalidity: removing invalid genetic variants, adjusting for population stratification, adjusting for potential confounders
|
Xco-twin is theoretically unrelated to unmeasured confounders specific to an individual only; any relation with unmeasured confounders shared between twins is captured by the familial confounder, SXY Still works even if Xco-twin is also associated with Yself through SXY; inference on causation based on the changes in regression coefficients still apply
|
| Exclusion restriction assumption: the association between the instrumental variable and the outcome is mediated through the exposure variable only |
Validity assessment: biological knowledge, test between genetic variants and potential alternative mediators, evidence from GWAS of the outcome or potential alternative mediators Approaches for invalidity: removing invalid genetic variants, statistical methods such MR-Egger regression and Weighted Median Estimator
|
Xco-twin is theoretically not related to Yself through potential alternative pathways in which Xself is uninvolved Still works even if Xco-twin is also associated with Yself through a mediator shared between twins; inference on causation based on the changes in regression coefficients still apply
|
