Elif Ferreira· Compensation Consultant, Fairlead Shipping
Asked in Pay Equity & Transparency ·
I am the analyst for a US software company headquartered in Massachusetts, about 1,100 employees. Leadership has asked for gender by race and ethnicity, which is reasonable, but once I cross six ethnicity categories with gender and then look within job level, three quarters of the cells have fewer than ten people and many have one or two.
Running a regression with interaction terms gives me coefficients with confidence intervals wider than the effects. Reporting simple averages for cells of three people feels both meaningless and a privacy problem.
What do people do in practice? I do not want to tell leadership it cannot be done.
10 helpful · 6 insightful · 2 agree · 3 replies · 399 views · 3 following
Accepted answer· by Oscar Novak
Two tools for two questions, at about 1,400 people. For the statistical question, we keep the company-wide model with gender, ethnicity and the interaction, and report only the handful of groups large enough to estimate, with intervals shown. We say explicitly which groups could not be assessed. For everyone else we do a cohort review rather than statistics. Every employee in a small group gets an individual comparison against the predicted pay for their role, level and tenure, and anyone more than 5% below is reviewed by a person. It caught 14 cases last year that no coefficient would ever have shown.
Read in the conversation ↓
Benefits Manager, Meridian Health Systems
We set a minimum cell size of 10 for anything reported outside the analysis team and 30 for anything we describe as a finding. Below that the cell is pooled upward, first across adjacent levels, then into a broader group. Leadership accepted it once we showed them a cell where one resignation flipped the sign of the gap.
Compensation Analyst, Kestrel Aerospace
Accepted answer
Two tools for two questions, at about 1,400 people.
For the statistical question, we keep the company-wide model with gender, ethnicity and the interaction, and report only the handful of groups large enough to estimate, with intervals shown. We say explicitly which groups could not be assessed.
For everyone else we do a cohort review rather than statistics. Every employee in a small group gets an individual comparison against the predicted pay for their role, level and tenure, and anyone more than 5% below is reviewed by a person. It caught 14 cases last year that no coefficient would ever have shown.
Elif AlvarezCRP
Director, Total Rewards, Kestrel Aerospace
