Think of it like this. Linear models in statistics break out the prediction for a given case in terms of main effects and interactions. All women get hit by the main effect of being a woman; it just so happens that some women are buoyed by other main effects like wealth, whiteness, etc.

Now, it may turn out that you might want to evaluate an aggregate measure of disadvantage per person, so you can sort people by level-of-disadvantage and apply that first. But that method may be brittle; if you choose the relative weighting of various forms of disadvantage wrong, then you're applying your benefit in a way that deviates from the "true" perfect-information order anyway.

One way of resolving this is to say "we're not going to attempt to apply this benefit in a perfectly Rawlsian order; we're just going to pick a main effect to attack and, while doing less good than a perfect-information solution, will still almost certainly do more good than harm.".