74% of the people these newspapers name carry no textual sign of their gender.
Measured on 1,062,339 people in 220,221 articles sampled from 1900 to 1963. Before anyone asks how the share of women changed, that number decides what the share can mean. This atlas tracks both, and tests whether the answer survives a change of method.
See what survives the robustness grid- 55.3% → 68.3%women's share of gender-signalled people, 1900–21 to 1948–63trend +3.11 pp per decade, 95% CI +1.41 to +4.81
- +1.30 to +3.37pp per decade: the trend in women's share of authority roles, depending on which of 5 methods measures “authority”same people, same years; compare methods
- 899of 900 analysis specifications show a rising women's shareslope range −0.27 to +5.84 pp per decade; the grid
- 0.38–0.84F1 for “authority” across methods, against reference labelsreference labels were produced by an AI annotator, not humans; details
The question
How did the representation of women and men in American newspapers change between 1900 and 1963, and how much of any answer is an artefact of the NLP method used to measure it?
The second half of the question is the point. A trend line from one pipeline says little on its own. Here every headline estimate is re-measured with different rules for who counts as a person, what counts as evidence of gender, and what counts as a role. Estimates are shown with the corpus's changing composition held fixed and with it left free.
What holds, and what does not
Each hypothesis was written down before any measurement was run. A status of supported means the pre-registered criterion holds in the primary specification and the sign survives every cell of the robustness grid; tentative means only the first; unsupported means neither. Statuses are computed by scripts/claims.py.
The share of gender-signalled people who are signalled as women increases, 1900-1963.
tentative · H1 · evidence
Women's share of authority roles is below their share of all people in every period, and the gap narrows.
unsupported · H2 · evidence
Adjusting for topic and newspaper changes the 1900-1963 trend by at least 25% of the raw slope.
supported · H3 · evidence
Role-measurement methods produce materially different trend effect sizes (ratio > 2 or sign disagreement for at least one headline role).
supported · H4 · evidence
The lexical method (A) yields a larger absolute trend than the context-aware methods (B, D).
tentative · H5 · evidence
The UNKNOWN share changes over time, and the women's-share trend is sensitive to how UNKNOWN is handled.
tentative · H6 · evidence
Among attributed quotations, women's share of speakers is lower than their share of gender-signalled people.
supported · H7 · evidence
Table view
| Year | primary: honorific + pronoun | honorific only | + gendered nouns |
|---|---|---|---|
| 1900 | 43.9% (41.6%–46.2%) · n=10,179 | 56.6% · n=7,190 | 43.0% · n=10,680 |
| 1903 | 47.2% (44.4%–50.2%) · n=9,710 | 61.8% · n=6,695 | 46.2% · n=10,245 |
| 1906 | 51.2% (49.2%–53.1%) · n=10,076 | 67.2% · n=7,067 | 50.0% · n=10,628 |
| 1909 | 51.9% (49.7%–53.9%) · n=10,407 | 68.3% · n=7,410 | 50.5% · n=10,995 |
| 1912 | 57.3% (55.4%–59.2%) · n=10,932 | 73.3% · n=8,023 | 55.9% · n=11,525 |
| 1915 | 61.7% (59.7%–63.7%) · n=10,758 | 76.9% · n=8,064 | 60.2% · n=11,341 |
| 1918 | 62.0% (60.0%–63.9%) · n=9,877 | 77.0% · n=7,509 | 60.0% · n=10,490 |
| 1921 | 64.4% (62.9%–65.9%) · n=12,646 | 77.5% · n=9,944 | 62.4% · n=13,382 |
| 1924 | 63.4% (61.6%–65.3%) · n=11,238 | 80.0% · n=8,366 | 61.7% · n=11,853 |
| 1927 | 67.8% (66.0%–69.5%) · n=12,593 | 85.1% · n=9,431 | 66.0% · n=13,240 |
| 1930 | 69.6% (68.1%–71.1%) · n=12,738 | 85.2% · n=9,791 | 67.2% · n=13,551 |
| 1933 | 67.5% (65.6%–69.2%) · n=13,605 | 83.3% · n=10,379 | 65.5% · n=14,473 |
| 1936 | 68.1% (66.4%–69.8%) · n=13,908 | 83.7% · n=10,675 | 65.9% · n=14,901 |
| 1939 | 70.4% (68.9%–72.0%) · n=15,002 | 83.4% · n=12,187 | 68.5% · n=15,879 |
| 1942 | 67.3% (65.6%–68.9%) · n=11,785 | 83.5% · n=9,035 | 64.7% · n=12,679 |
| 1945 | 65.2% (63.6%–66.8%) · n=11,206 | 81.8% · n=8,479 | 62.2% · n=12,430 |
| 1948 | 66.4% (64.9%–68.0%) · n=14,117 | 77.0% · n=11,641 | 64.3% · n=15,179 |
| 1951 | 67.3% (65.5%–68.8%) · n=14,259 | 79.5% · n=11,632 | 65.4% · n=15,303 |
| 1954 | 68.5% (67.1%–70.0%) · n=13,714 | 80.6% · n=11,231 | 66.5% · n=14,656 |
| 1957 | 71.5% (70.2%–72.7%) · n=15,212 | 81.7% · n=12,846 | 69.3% · n=16,264 |
| 1960 | 70.2% (68.5%–71.4%) · n=15,445 | 80.4% · n=13,033 | 67.8% · n=16,785 |
| 1963 | 65.0% (63.3%–66.5%) · n=12,309 | 78.0% · n=9,767 | 62.4% · n=13,636 |
AmericanStories (Dell et al. 2023, CC-BY-4.0), built on Chronicling America (Library of Congress). 10,000 sampled articles per year, 22 years.