In many scientific disciplines the proportion of women at doctoral level is close to parity or above. The proportion at full professorial level is not, and the drop between those two points is where the loss occurs.
This is a different problem from getting girls into physics, it happens to people who have already made every commitment the field could ask for, and it receives a fraction of the attention.
Where the loss is concentrated
The steepest attrition is in the postdoctoral years and at the transition to an independent position.
The structural features of that period are worth stating plainly, because they explain a great deal without requiring any assumption about differential ability or commitment.
It is long. Several years of fixed-term contracts is normal in many fields, sometimes considerably more.
It is geographically mobile. Moving between institutions, often internationally, is treated as evidence of seriousness. This interacts badly with partnership and with caring responsibilities, and the interaction is asymmetric in practice.
It is badly paid relative to the qualification. Postdoctoral salaries lag equivalent private-sector positions by a wide margin.
It coincides almost exactly with the years in which people who are going to have children generally have them.
And it culminates in an evaluation — the hiring or tenure decision — that is unusually unstructured and unusually dependent on reputation and sponsorship.
The two-body problem, quantified
Academics partner with other academics at high rates. This produces a well-documented coordination problem: two people requiring positions in the same location, in a market with few positions and no geographical flexibility.
Survey data on dual-career academic couples has consistently found that where one career is deprioritised to resolve the conflict, it is more often the woman's. The asymmetry is not enormous but it is consistent across surveys and institutions.
This is a structural problem with a structural solution — partner hiring programmes exist and are used — but the solution is unevenly available and heavily dependent on the seniority of the person being recruited.
Evaluation, letters and citations
Academic hiring depends heavily on reference letters and on publication record, and both have been studied for systematic differences.
Analyses of recommendation letters have repeatedly found differences in language: letters for women contain more communal and effort-related terms and fewer standout adjectives referring to ability. Studies of how such letters are read find that these differences affect evaluations.
Analyses of citation patterns have found gaps in citation rates that persist after controlling for a range of factors, with self-citation and network effects among the proposed mechanisms. Findings differ by field and effect sizes vary.
Studies of authorship credit have found that women's contributions to collaborative work are attributed to them less reliably, particularly in fields where author order carries meaning.
None of these individually is large. In a market where a small number of candidates are selected from a large qualified pool, small consistent differences at each stage are exactly what produces large differences in outcome.
What has been shown to help
Structured evaluation. Search committees using specified criteria agreed before candidates are reviewed produce different outcomes than committees discussing candidates freely. Several institutions have documented shifts after making this change.
Extending the tenure clock for caring responsibilities has been implemented widely, and its evaluation is instructive: at least one well-known analysis found that gender-neutral clock extensions benefited men more than women, because the additional time was used differently. This is a useful caution that a neutral policy in an asymmetric environment does not produce neutral outcomes.
Cluster hiring — recruiting several people at once rather than singly — has been reported to produce more diverse outcomes, plausibly because it reduces the weight of any single comparison.
Reducing the number of moves required. Longer postdoctoral contracts and clearer routes to independence reduce the number of points at which a coordination problem can end a career.
The industry comparison
Many people who leave academic science do not leave science. They move to industry, government or clinical roles, and a substantial share report better pay, more stability and more predictable geography.
This complicates the loss framing. A researcher who moves to a pharmaceutical company is not lost to science.
What is lost is her presence in the institutions that train the next cohort, allocate research funding and decide what gets studied. Those functions are concentrated in academia, and the composition of the people performing them is not a neutral matter.
The point that generalises
Recruitment is popular because it is visible and because its subjects are young enough to be photographed. Retention is unpopular because it requires an institution to examine its own promotion and evaluation processes.
The arithmetic is not sympathetic to that preference. A field losing a large share of its trained women after a decade of investment in each of them is not solving anything by admitting more at the bottom.