4 ms·

> Finally, 25% of their data is with physicians with names that preclude gender assignment by their algorithm. If we assume that to be a mix of men and women, t

> Finally, 25% of their data is with physicians with names that preclude gender assignment by their algorithm. If we assume that to be a mix of men and women, that group should perform somewhere between the male and female physicians. Why exclude this data from the analysis, rather than analyze it to confirm the trend? I think leaving that out makes a lot of sense actually. You can't assume that the mix is 50-50 just because you don't know the data. Anecdotally, I know more women who use a male or gender-agnostic version of their name professionally than I do men. It may be more common in fields other than medicine, but for example we only know "JK Rowling" as "JK" because her publishers were worried that having an author with a feminine name would make it harder to sell the book to boys.

carbocation 8y ago
> You can't assume that the mix is 50-50 I don't assume that it's a 50/50 split. I assume it's somewhere between 0% female to 100% female. And it's a big sample. If a mix of (men and women) doesn't perform somewhere between the all-male and all-female groups, that's a big problem with their hypothesis. They're saying something equivalent to Group A scores mean 90 and Group B scores mean 95. If you then take a very large set of people that is drawn from Group A and Group B with an unknown proportion and they score mean 85 or 100, that would be fairly devastating to the claim. I'm not saying that this should be weighted into the main analysis. I'm saying they should test their hypothesis with this.