> It's a consequence that hiring practices should favour women who are equally qualified on paper, because they have likely overcome more barriers in practice and, all other things being equal, are better hires.
I want to argue against this from a purely utilitarian standpoint, where we care about getting better hire, and don't care about any ethical implications. I think your logic is just bad math.
You're right, and women have more barriers to entry into STEM. And if some barriers are already removed (as this study suggest), I can assume that others still exist. Let's make the numbers simpler, and assume that these barriers make it 2 times harder for a given woman to get into some experience level (I don't know real numbers anyway, and they don't matter).
However, when we observe the effect of these barriers, we simply see that there are 2 times more men on that level than women. That's the whole effect. Statistically, if a woman had 2 times less chance of getting through, and there are already 2 times less women than men, then we already see the full effect of this barrier: there's no evidence that there should be some hidden variable to explain this barrier.
However, if we would know that women's barrier makes it 10 times as harder, and still, there are only 2 times as many men as women, then we would need some other data to explain this; namely, that these women are actually 5 times as good as the men.
But we don't know that. The whole knowledge about these barriers that we have comes from the outcome: we see that there are 2 times as many men as women, and that's how we assume (correctly, I think) that there is a barrier. But if we start from this outcome data, we can't through some magic come back to it and add a hidden variable - it's just a logic loop, and a strange one: we see data, make conclusions from observations, and from these conclusions change our observation of this very data to see it as incomplete, without help of any data points outside this original observation. How could such logic be correct?