Nuanced use of PCs with GRMs

,

We had a helpful discussion on the selection of covariates for GWAS analysis on Day4. While GRMs capture genetic relatedness, including PCs is also recommended to account for overall genotype similarity across populations.

I am working on a problem where I use the logistic SAIGE model and include PCs to account for population stratification. However, the statistical significance of my results changes depending on how many PCs I include.

Several suggestions came up during the discussion:

  1. Use a GRM plus PC1–PC4 if the data come from a single country, and include more PCs if the data come from multiple countries.

  2. Check lambda.

  3. Generate an elbow plot for the PCs and identify the PCs that explain the greatest amount of variance.

I tried looking for publications that support these approaches. I was wondering whether anyone has come across articles that discuss the nuanced use of these parameters as covariates, especially related to point 1.