Can the training objective hide the conditional signal we later want to test? A classifier trained for marginal prediction can exploit confounders and compress away useful within-group differences.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results