Towards Data Science

A Computational Analysis of Women’s Suffrage in Historical Newspapers

7.5内容质量

TL;DR · AI 摘要

分析显示1915-1917年纽约报纸对女性选举权报道激增,但加州数据存在严重偏差,OCR技术与查询定义影响历史趋势测量准确性。

核心要点

  • 纽约1917年报纸中'women’s suffrage'出现频率是加州的2.3倍
  • OCR错误导致15%的关键词识别失败,需人工校验关键数据
  • 加州报纸来源仅占全州出版物的6.8%,存在严重采样偏差

结构提纲

按章节快速跳转。

  1. 明确分析1890-1920年加州与纽约报纸中女性选举权报道的时空差异

  2. 揭示Chronicling America数据库仅覆盖5%的同期报纸出版物

  3. 采用严格六词查询与宽泛查询双重标准衡量关注度

  4. 发现纽约1917年报道量激增与加州持续低水平的显著差异

  5. 指出OCR误差率与报纸存档不完整性的双重数据质量威胁

思维导图

用一张图看清主题之间的关系。

查看大纲文本(无障碍 / 无 JS 友好)
  • 历史报纸数据分析
    • 数据局限性
      • 采样偏差
      • OCR误差
    • 分析方法
      • 严格查询
      • 宽泛查询
    • 时空差异
      • 纽约激增
      • 加州平稳

金句 / Highlights

值得收藏与分享的关键句。

#数据科学#历史分析#OCR#大数据
打开原文

A Computational Analysis of Women’s Suffrage in Historical Newspapers | Towards Data Science

Data Science

A Computational Analysis of Women’s Suffrage in Historical Newspapers

Using 600,546 digitized newspaper pages to measure how suffrage coverage changed from 1890 to 1920

Renish Joshi

November 4, 2026

11 min read

Prefer us on Google

Share

Instructing Yeowomen , photographed between ca. 1915 and 1920. (Library of Congress, public domain)

The Question

In what years was there the highest coverage of suffrage for women in California and New York between 1890 and 1920, and were the trends similar in the two states?

Both California and New York constitute useful grounds for historical comparison. In California, women's suffrage was introduced in 1911. In New York, a suffrage amendment failed in 1915 but passed in 1917. The Nineteenth Amendment was transmitted to the states in 1919 and became ratified in 1920.

Though these dates serve as useful context, they are not used to guide the analysis process. While this task was originally seen as just a historical counting exercise, it turned into a wider problem of measurement.

This problem can be seen to be relevant not only to the study of historical newspaper archives. Practitioners dealing with datasets have to contend with cases when their size, composition, language, and quality change over time. This may mean that an identified trend is due either to the change in the phenomenon under observation or to a change in the process of recording it.

The dataset is not the population

The data come from the Library of Congress’ Chronicling America database, retrieved through the Tim-Pinecone/LOC-Chronicling-America dataset on Hugging Face and licensed under CC0 1.0 .

Every line stands for an OCR’d page of a newspaper and includes data such as title, city, year, and other information. The Parquet files were retrieved remotely through DuckDB.

The dataset includes 600,546 pages: 290,397 from California and 310,149 from New York.

However, the archive cannot be considered to be a stable sample of the press.

Five New York titles have been included throughout the time span, and in 30 out of 31 years, one newspaper accounted for more than half of the pages. There are 21 titles from California, but in no year have more than six titles been present.

The bottom line is clear:

The data describe the archive, not the whole newspaper press.

Defining “suffrage attention”

Unit of analysis: page.

A page is defined as having at least one search term within it regardless of how many times it appears.

The strict query uses six phrases:

  • woman suffrage
  • women’s suffrage
  • female suffrage
  • votes for women
  • suffragist
  • suffragette A broader comparison query searches for any occurrence of suffrage .

There were 14,971 matching pages in the strict query search against 28,283 matching pages in the broad query search for the entire sample period.

“Attention,” then, is the share of archived pages that include one of the above phrases. This measure does not capture the level of public backing. In addition to support, opposition, satire, ads, and news, any form of attention is captured in this measure.

Why raw counts can mislead

A raw count mixes up two different variables: (1) the number of pages that the archive has for that year, and (2) how often suffrage is found in these pages. An increase in the number of pages in the archive can bring up the number of matches even when the rate stays the same.

It can be seen in the current data as well. There is a Spearman correlation between the raw suffrage-page count and the archive size for California and New York of 0.59 and 0.63, respectively. Once normalized, these correlations decrease to 0.17 and 0.38, respectively.

It is very easy to see in New York. There are the most matching pages in 1913 (1,227), while the normalized rate peaks in 1909.

There is the same problem in California. There are only 96 matching pages in 1914, but the archive itself contains only 2,248 California pages that year. When normalized, 1914 turns out to have one of the highest rates of attention.

A more general lesson is that when dataset size changes over time, raw counts contain some information about the dataset size as well.

Normalizing attention

The key indicator utilized is:

Suffrage Rate = (matched pages ÷ archived pages) × 10,000

An index of 50 would imply one matched page in every 200 archived pages.

While missing values were distinguished from zeroes, a zero means that pages were available but there were no matches, while missing means that there was no evidence at all for that particular state/year.

After normalization, both states show a similar trend, which includes a rise in the 1890s, low rates in the beginning of the 1900s and then a stronger surge starting roughly in 1908.

In California, the peak comes in 1913 at 735.9.

In New York, the peak comes in 1909 at 853.6, followed by 1913 at almost the same level at 851.9.

The greatest increase from one year to another in New York occurs from 1907 to 1908, with the rate rising from 72.2 to 406.1.

However, it should be noted that these peak values require verification.

Does the search definition change the story?

Keyword queries are modeling choices.

A narrow query can miss relevant language, while a broad query can add noise.

The strict and broad series track each other closely: their annual rates correlate at 0.94 in California and 0.98 in New York .

The broad query finds roughly three times as many pages.

But California’s peak changes.

The strict query peaks in 1913 .

The broad query peaks in 1896 .

The general pattern is therefore robust, while the exact peak is not.

That distinction matters: a broad conclusion can remain stable even when a headline statistic depends on the model.

When one newspaper dominates the state

Normalizing deals with the problem of archive size inequality but does not solve the problem of archive content inequality.

The main rate is equal in its weight for each page. Hence, a newspaper that accounts for 80% of the pages will provide about 80% of the signal.

In order to see the difference between these two measures, another measure was used in which all newspapers have equal weight.

The overall correlation between these two measures in the time series is very high: 0.96 for California and 0.997 for New York.

But it becomes important at certain peaks.

New York’s 1909 peak

In 1909, the proportion of pages published by The Sun in New York was 84%, and the suffrage rate is about 907 out of 10,000. Therefore, the series based on page weight peaks in 1909. With the same weight for each newspaper, the peak occurs in 1913. Neither of the two methods of weighting is wrong, but they solve different problems. When all pages have the same weight, the year 1909 is the peak. If all newspapers have the same weight, then the peak is 1913.

California’s language problem

California provides another case study.

From 1915 through 1919, the main California newspaper in the database was L’Italia, an Italian-language newspaper out of San Francisco.

Italian-, Spanish-, and German-language newspapers accounted for over a quarter of the pages printed in California in each year of a seven-year stretch, rising as high as 79% in 1917.

The search terms used were English.

The average rate of suffrage for those years with non-English content was 60.3 per 10,000 among all California newspapers.

The rate rises to 130.9 when the calculation is limited to only English newspapers.

While the fall after 1913 can still be observed, part of the magnitude of the drop is due to the change in language in the underlying corpus of the query.

A constant query does not imply a constant measure if the underlying corpus varies.

Testing for change points

Visual trends could potentially affect storytelling in a backward-looking manner.

For examining potential changes in structures in a systematic way, the analysis used change-point detection by means of the PELT method in the ruptures package.

Four penalties were evaluated, and only a change point detected by at least three penalties within one year of each other was considered stable.

Historical events were introduced only after change-point detection.

No stable change point was detected as stable for either state.

The only detection took place around 1908 in New York, but only in the case of the lowest penalty.

It does not mean that there was no change in the series; on the contrary, the increase of New York starts in 1907.

What it means is that the change detected is not stable in relation to reasonable values of the penalty parameter.

Lowering the penalty in order to detect more change points would defeat the purpose of sensitivity analysis.

Null results can also carry information, especially when produced by a predetermined procedure.

What TF-IDF found

To examine the discourse related to suffrage, the states were divided into four time periods:

  • 1890-1900
  • 1901-1910
  • 1911-1917
  • 1918-1920

Up to 750 matched pages per state-time period were included, resulting in an NLP corpus of 5,352 pages. After excluding English stop words and the original search terms, a TF-IDF model including unigrams and bigrams was created.

This process identified another issue with archival text – many salient terms would have more to do with archival features rather than with historical language.

The early New York papers contained many OCR artifacts, such as “tbe”, “tbo”, and “ofthe”. The California time periods had terms like “los angeles”, “el centro”, and “imperial valley”, which corresponded to the newspaper that was dominant in the sample. Late California papers featured Italian function words like “della”, “che”, and “sono”.

Thus, time-based NLP identified OCR quality, newspaper name, and language composition before identifying historical terms. However, a clearer historical trend emerged in the 1918-1920 New York time period with terms like “president wilson”, “war”, “league nations”, “amendment”, and “ratification” becoming prominent. While TF-IDF was helpful, it partially served as an indication of archival features.

What survived every check?

Several findings remained after changing the denominator, query definition, weighting scheme, and language assumptions.

  • Both states show a strong wave of suffrage attention from roughly 1908 to 1915. That pattern survives normalization, title balancing, the broad query, and the language check.
  • The exact peak year is less stable than the overall trend. New York peaks in 1909 under page weighting but 1913 under raw counts and title balancing. California peaks in 1913 under the strict query but 1896 under the broad query.
  • Archive composition materially changes the result. One newspaper contributes 84% of New York’s pages in 1909, while non-English newspapers contribute as much as 79% of California’s archive in 1917.
  • A visible trend is not automatically a stable statistical regime shift. No PELT breakpoint remained stable across the tested penalties.

What this analysis cannot tell us

Just because there is a discussion in the newspaper does not mean that public opinion has been affected. A periodical can discuss suffrage for a variety of reasons, such as support, opposition, ridicule, or simply because it is considered newsworthy.

Moreover, the analysis cannot determine cause and effect. The observed correlation between the times when peaks occurred during political events does not prove whether politics had an impact on the press or whether the press impacted politics.

This archive is not a representative sample of all of the press. In New York, there is only a small number of Manhattan newspapers discussed, while the composition of the newspapers in California varies over time.

OCR errors can distort the text. In a manual check of 40 matched pages, 37 turned out to be relevant, while two were false hits and one had no mention of women’s suffrage at all. Seven excerpts showed considerable damage done by OCR.

Therefore, the analysis studies what is possible to learn from the digitized archive.

What the archive teaches about data analysis

The initial question was one of history:

At what point did newspaper coverage of women’s suffrage hit its peak?

A more practical question of methodology emerged:

How consistent is the same pattern when the dataset itself is part of the problem?

While the broad 1908–1915 wave continued, the specific peaks were no longer consistent.

The California dip became less pronounced after accounting for language composition.

The New York peak of 1909 lost some strength following correction for concentration of newspapers.

Additionally, the NLP analysis often found shifts in the archive that preceded language changes.

The larger implications are clear:

  • Figure out the denominator before interpreting the numerator.
  • Distinguish missing data from zeroes.
  • Observe the composition of sources changing.
  • Analyze both the number and the language of documents.
  • Treat the search terms as model choices.
  • Check OCR output manually.
  • Expect NLP to show characteristics of the dataset itself.
  • Test if the statistical result holds with other plausible configurations.

Archival text data are not simply waiting to be counted but rather are dynamic samples. The aim is not to perfect all the samples but to find which results hold up after their imperfections are made clear.

Reproducibility

The analysis uses the Library of Congress Chronicling America corpus through the Tim-Pinecone/LOC-Chronicling-America Hugging Face dataset.

The version used in the analysis corresponds to Hugging Face commit:

781d9ba6cb605320764bec3fdcecad6c0c263096

All 62 state-year queries were collected between October 24 and October 25, 2026 UTC.

The pipeline was written in Python and uses DuckDB for remote Parquet queries, pandas and NumPy for processing, ruptures for change-point detection, scikit-learn-compatible TF-IDF processing for text analysis, and Matplotlib for figures.

NLP sampling is deterministic using md5(_id) ordering, while manual QA sampling uses a fixed random seed.

Running:

python run_all.py

rebuilds the collected aggregates, analysis, NLP outputs, figures, and reports and performs the project’s validation checks.

The final run completed all validation checks successfully.

Image provenance: All analytical figures and visualizations in this article were created from the results of this analysis.

Featured image: Bain News Service, Instructing Yeowomen , ca. 1915–1920. George Grantham Bain Collection, Library of Congress Prints and Photographs Division. No known restrictions on publication.

References

Library of Congress. Chronicling America: Historic American Newspapers. National Digital Newspaper Program, Library of Congress and National Endowment for the Humanities.

Condello, Tim. Chronicling America: Nationwide Partitioned Apache Parquet Newspaper Dataset. Hugging Face, 2026.

California Secretary of State. History of Women’s Suffrage in California.

National Park Service. The Fight for Women’s Suffrage: It Happened in New York, Finally.

National Archives and Records Administration. 19th Amendment to the U.S. Constitution: Women’s Right to Vote.

Library of Congress. Chronicling America — Technical Information: Optical Character Recognition.

DuckDB. HTTP Parquet Import / HTTP(S) Support Documentation.

Truong, C., Oudre, L., and Vayatis, N. ruptures: Change Point Detection in Python. ruptures documentation.

Killick, R., Fearnhead, P., and Eckley, I. A. “Optimal Detection of Changepoints With a Linear Computational Cost.” Journal of the American Statistical Association , 107(500), 2012, pp. 1590–1598.

Scikit-learn. Feature Extraction: TF–IDF Term Weighting.