It's on section "Pooling", on page 7 of the pdf (400 in the original), second column, last
parargraph starting with "We reanalyze Table 1, ...".
In that paragraph, the authors state that they estimate the probability of
admission of a female applicant "by multiplying the estimated probability of
admission of any applicant (regardless of sex) to that department by the number
of women applying to it", which I find reasonable.
However, immediately afterwards they compare that probability with the
probability of an applicant being admitted given that the applicant is female.
I quote from the end of the second column and the start of the third one:
"Thus, if the chances of getting into a department were one-half for all
applicants to it, and 100 women applied, we would expect 50 women to be admitted
if they were being treated just like the men".
In other words, bias depends on whether the proportion of applicants of one
sex, out of all applicants of _that_ sex, was higher than the expectation
formed for applicants of any sex. It's a little confusing and the language in the article is not very precise. But that's what's up.
