Loss of predictive power of polygenic risk score when dataset contains missing variants

Question

I am trying to calculate polygenic risk scores (PRS) scores for a new dataset. This dataset does not have all the variants that the PRS score needs. The PRS score I am interested in has 40 variants, but in my dataset only 20 of those variants are present. What is the correct way of assessing the decreased predictive power of the PRS score due to the missing variants?

I assume it is not as easy as summing the effective weights of those variants I have and comparing to the full sum (all variants needed for the PRS).

Thanks a lot and sorry for newbie question!

Answer 1

In a sample with all 40, determine the accuracy of the PRS. Then in the same sample, remove the 20 you are missing in the other set, and then determine the accuracy with the missing variants. This will give you an idea of the level of attenuation. However, you may not have access to such a dataset. In that case, I would recommend imputing your data before calculating a polygenic score. (It's likely that the GWAS the PRS is from used imputation itself.)

Answer 2

Which software package around you using, PRSice? Generally missing values comprising half the total observations are extremely difficult to calculate using regression analysis (or many statistical analyses for that matter.

One way forward is to include (an) example(s) of all 40 variants into your data set and perform the analysis with and without the added data, then compare this against a data set where all 40 variants are observed and your samples are excluded. You might find the missing values have low probability of PRS and this would then explain your results.

Missing analysis in regression is a specialist area of statistics and is not for the faint hearted, so a work-around is the best way forward.