[x] Group lasso is not doing that much better than lasso in R^2. Some improvement ideas:
[x] Try different phenotype than height?
[x] We may need to pool all population to get enough ancestry variation. The point is that we're probably not going to do that much better on white_british, but pooling can improve performance on the other populations.
[x] Check ancestry information for white british population that there's sufficient counts.
[x] Include population information as unpenalized features.
[x] Try Mexico biobank
[x] Try subsetting UK biobank to those born outside UK