Investigating which applicants Zenkai's hiring algorithm recommends โ and how the recommendations change when the outcome variable does. N = 2,500.
Zenkai built a LASSO-penalized logistic regression on five years of applicant data. The model assigns each applicant a probability score, and the firm advances the top 25% to a partner interview.
The EEOC standard for adverse impact: a selection rate for any group below 80% of the highest group's rate is regarded as evidence of disparate impact, triggering Title VII review.
The algorithm Zenkai's data science team built. Standardized coefficients shown for the variables that survived LASSO penalization.
Select 3 or 4 of the 9 candidate variables. The simulator returns the precomputed top-25% composition under the currently selected outcome.
Selected: 0 / 3โ4 variables.
The same applicant data, the same algorithm class, the same penalization. Only the outcome variable differs.
The pedagogical line: Two algorithms with identical predictive accuracy (AUC = 0.849 on each) and the same identity-blind feature set can produce dramatically different equity outcomes โ because the choice of outcome variable is the choice of what the algorithm is for.