Is it fine to adjust first 10-PCs always? Why only first 10-PCs?
Has anyone come up with a more precise way of using PC to control for population stratification other than using for 10 components - always felt very arbitrary
this depends on the population heterogeneity of your sample, so not always 10 PCs
larger samples usually require more
depending on sample and analysis type 10PCs is too many. If you have a homogenous population then 4 can be plenty. Over correction will show up as lambda >1
try running a phenotypical regression analysis first and see how many PC are significant
You could plot the Eigenvalues and look for an elbow or also look at 80% variance explained
You can also control for popstrat more precisely using more fine-grained variables beyond PCs — for example, the deCODE Icelandic dataset controls for the county a person was born in using dummy codes
What would be the problem when controlling for too many PCs?
Most of the time, you don’t know what the PCs represent. So it’s possible that you may dilute your signal if the PC related to e.g. case control status and a genuine SNP.