gender-pay-gap
70.6%, 19.1% or 15.0%. All three are correct. That is the problem.
See code →The problem
In gender pay gap debates, each side cites its own number and accuses the other of lying. The curious thing is that both are right: the three most cited figures are all correct, but they answer different questions. The analysis does not take sides — it makes explicit why all three are simultaneously true and how choosing between them is not a technical decision but a political one.
The data
IPUMS-CPS (Current Population Survey): nationally representative US sample with gross annual wages, weekly hours worked, industry sector, occupation (3-digit SOC) and educational level. The source is the same one the BLS uses for its official statistics.
Approach
- Gross gap (70.6%): direct comparison of mean annual wages with no adjustment — the figure usually cited by advocacy
- Adjusted for sector and occupation (19.1%): controls for the type of work each group chooses — the figure usually cited in liberal economic debate
- Further adjusted for hours worked (15.0%): also controls for working hours — the figure preferred by those arguing the gap is choice, not discrimination
- Omitted variable bias analysis: why "adjusting for occupation" does not neutralise bias if occupational segregation is itself an effect of discrimination
- Multiple regression to decompose the gap into explained and residual parts under each set of controls
Tools
jupyter
pandasResult
There is no single correct figure — there is a figure for each question. The analysis makes visible what is sacrificed at each adjustment step and why "adjusted" is not synonymous with "fair": adjusting for occupation presupposes that occupational choice is free, which is precisely what is in dispute.
Next steps
The current analysis is descriptive and uses OLS regression. The most interesting extension is causal: applying instrumental variables or a regression discontinuity design to separate the effect of direct discrimination from the effect of occupational segregation, which conventional adjustment methods cannot do.