Table 1.

Summary of the Two Approaches to Mendelian Randomization

	Individual‐level data MR	Summary‐level data MR
Requirements	Measured exposure, outcome in the same population. Genotype dosage for all instruments in the same population (or a polygenic risk score)	SNP‐exposure association results and SNP‐outcome association results from separate populations, including: Effect allele Other allele Effect allele frequency Beta for the per allele effect on the exposure or outcome (unit increase or log odds) Standard error for the beta
Possible analysis methods	To test for causal effect: linear/logistic regression of SNP genotype or polygenic risk score on the outcome To quantify causal effect: Single SNP: Wald ratio estimate βoutcome/βexposure Single/multiple SNPs/polygenic risk score: Two‐stage least‐squares regression	To test for causal effect: determine SNP‐outcome effect using summary statistics from a published GWAS To quantify causal effect: Single SNP: Wald ratio estimate Multiple SNPs: an inverse‐variance weighted meta‐analysis of the Wald ratio estimate for each SNP
Testing the relevance assumption	First‐stage F‐statistic	Mean F‐statistic for SNP‐exposure association
Testing the independence assumption	Associations between the instrument(s) and potential confounders can be directly tested for all known/measured confounders	N/A
Testing the exclusion‐restriction assumption	Sargan test for heterogeneity in individual SNP results	Cochran's Q statistic as a measure of heterogeneity in Wald ratio estimates MR‐Egger intercept as a measure of the average effect of the SNP on the outcome when there is no effect of the SNP on the exposure
Pleiotropy‐robust methods	MVMR MR‐GENIUS controls for some directional pleiotropy( 106 ) sisVIVE and adaptive LASSO for outlier removal( 107 , 108 )	Several methods, including MR‐Egger, weighted median, weighted mode, MR‐CAUSE, MR‐PRESSO, MVMR, reviewed in Sanderson et al.(9) Can be broadly categorized as outlier adjustment, outlier removal, or estimate adjustment methods
Benefits	More flexibility in models (eg, can test for non‐linear effects) and covariates Ability to perform subgroup analyses (eg, sex‐stratified)	Larger sample sizes increase power Greater range of sensitivity analyses to determine pleiotropy‐robust estimates of causal effect
Limitations	Sample limited to those with measured exposure, outcome, and genotype, often restricting sample size Fewer methods to interrogate pleiotropy Weak instrument bias is toward the observational (confounded) estimate, potentially resulting in type 1 error	Assumes the two study populations are drawn from the same underlying population in terms of ethnicity, sex distribution, etc. Weak instrument bias toward the null, resulting in type 2 error Overlap in individuals between samples can result in bias toward the observational estimate in the presence of weak instruments (type 1 error) Unable to control which covariates are adjusted for Unable to perform subgroup analyses unless summary statistics are available for specific subgroups for both exposure and outcome