As a result of this joint learning, splits in the random forest are more likely to occur along informative genetic features that are orthogonal (that is, not correlated) to population structure.