# Lecture 12: historical and current gender pay gaps ## Historical earnings `gender_history.csv` uses the Census Bureau's [Table P-20](https://www2.census.gov/programs-surveys/cps/tables/time-series/historical-income-people/p20.xlsx), median annual earnings of workers aged 25 and over by sex and education. The original workbook is saved as `source/p20.xlsx`, downloaded October 8, 2026. It includes income years through 2025; the classroom extract retains 1991–2024 to end in the same year as the ACS investigation. There are 204 rows: 34 years, two sexes, and three education groups (all workers, bachelor's degree, master's degree). The lecture filters to the two degrees. One row describes a year/sex/education group. `pay` is published median annual earnings in thousands of that year's dollars. These include full- and part-time workers. Run `Rscript course_design/prepare_gender_history.R` from f26. Where the workbook reports two survey versions for a year (2010, 2013, 2017), preparation retains the first, newer-methodology row. No medians are calculated from the old group means or combined across degrees. The F25 mean-earnings source remains unchanged in F25. Source workbook SHA-256: `c7d52097e0e57f76489b88bd99dc4d0d1cd92e450e500037da5b28dc4337cc4b`. The female/male ratio is unchanged if both earnings amounts receive the same annual inflation adjustment. This series measures annual earnings; [Pew's headline estimate](https://www.pewresearch.org/short-reads/2025/03/04/gender-pay-gap-in-us-has-narrowed-slightly-over-2-decades/) uses median hourly earnings among workers aged 16 and over. ## Current ACS workers `gender_acs.csv` contains all 27,028 eligible person records from the national 2024 ACS one-year public-use file already downloaded for Lectures 10–11. It uses Lecture 11's filters: currently employed civilian wage/salary workers (ESR 1–2, COW 1–5), ages 25–50, computer/math occupations (SOCP starts with 15), bachelor's or master's degree (SCHL 21–22), at least 35 usual weekly hours and 50 weeks worked, positive wages and person weights. Run `uv run --project course_design/skill_returns/acs python course_design/prepare_gender_acs.py` from f26. The file retains every eligible person and uses `pay = WAGP * ADJINC / 1000000 / 1000`, in thousands of 2024 dollars. There is no trimming, sampling, or wage imputation. ```text Column Meaning sex SEX: 1 Male, 2 Female age AGEP: age in years degree SCHL: Bachelor's (21), Master's (22) pay Adjusted annual wage/salary income, $1,000 of 2024 dollars hours WKHP: usual weekly hours worked weeks WKWN: weeks worked in the previous 12 months sector Private (COW 1–2), Public (COW 3–5) occupation_code OCCP: Census occupation code weight PWGTP: person survey weight ``` [2024 ACS documentation](https://www.census.gov/programs-surveys/acs/microdata/documentation/2024.html) defines the released fields. SEX records male/female sex categories; this field does not record gender identity. Source archive checksums and extraction metadata are in `course_design/skill_returns/acs/data/acs_2024.json`. `gender_occupations.csv` supplies one label for each of the 16 occupation codes present. These unchanged codes and labels are taken from the saved Census 2018 ACS occupation list also used in Lecture 10's current-pay preparation. The 2024 SOCP patterns for these codes agree with that list's occupation families. Some codes describe several occupations together; `Other mathematical science occupations` should not be relabeled as data scientists alone. The optional exercise selects four named occupations with more than 100 sampled men and women each. ## Classroom analysis ACS descriptive earnings summaries use sample medians, with each respondent receiving equal weight. Person weights remain available. The main age investigation joins the occupation names and keeps software developers with bachelor's degrees: 7,653 workers. Pooled median pay at ages 25–29 is $111,677.50 for each sex. At ages 40–49 it is $121,830 for women and $159,394.25 for men. Keeping exactly 40-hour workers leaves the young medians equal and gives an older-group gap of $30,965.125. Age-group medians are computed from individual records. The age regression uses the same 7,653 individual records with ordinary lm(), which models mean pay. Its section explicitly switches to mean earnings and displays sample means alongside the fitted lines. It does not estimate a median regression. The interaction expresses two separate simple regressions; no general many-predictor regression or quantile regression is introduced. The earnings-distribution investigation returns to all 27,028 workers. quantile() computes within-sex 10th and 90th percentiles; median() supplies the 50th. Female p10/p50/p90 are $55,838.75/$101,525/$203,050; male values are $60,915/$121,830/$253,812.50. These use R's default quantile method and equal respondent weights. Students reshape the three percentile columns for plotting. The corresponding dollar gaps are about $5,076/$20,305/$50,763. The historical CPS groups and the current ACS computer/math sample describe different populations. They are analyzed separately. Age comparisons use different workers observed in 2024, not repeated observations of the same people. The hourly calculation divides annual pay by usual hours times weeks worked and is an approximation. Regression coefficient tests use ordinary lm() assumptions and do not estimate causal discrimination effects. Keep this methodological context in presenter notes and oral discussion.