Large-scale genome-wide association studies (GWASs) have the power to detect very small effects. Steiger filtering is a method for selecting valid genetic instrumental variables (IVs) from very large GWASs with statistical power to identify genetic variants that are related to the trait of interest at a genome-wide significance threshold because it is downstream of many other traits that the genetic variant influences. For example, a sufficiently large GWAS of a biomarker (biomarker A) may identify numerous genetic variants that influence another biomarker (biomarker B) because biomarker A is influenced by biomarker B.
If those genetic variants are included in an MR study as an instrumental variable (IV) estimating the effect of biomarker A on biomarker B without knowledge of their biology (as is typical from GWAS results), the result could (erroneously) suggest that biomarker A causes biomarker B (when the opposite is true). These reverse causal IVs will bias the MR estimate. In these situations, Steiger filtering uses a statistical test to identify the stronger of bidirectional effects (e.g., biomarker A on biomarker B or biomarker B on biomarker A) to select valid IVs from very large GWASs that can be used in MR. Steiger filtering assumes that a valid IV should explain more variance in the exposure than the outcome and removes those genetic variants that do not satisfy this criterion. If biomarker A has much more measurement error than biomarker B, then the variance explained by the SNP on biomarker A might be smaller than on biomarker B even though it primarily influences biomarker A. Some patterns of confounding can also lead to this problem.
References
- Hemani G, Tilling K, Davey Smith G. Orienting the causal relationship between imprecisely measured traits using GWAS summary data. PLOS Genetics 2017; 13: e1007081.
- Hemani G, Bowden J, Haycock PC, et al. Automating Mendelian randomization through machine learning to construct a putative causal map of the human phenome. bioRxiv 2017: 173682.