Are Young Women More Liberal Than Expected?

This is a technical report on a side project, not part of my book. It revisits a question I wrote about in early 2024, with two more years of data and a better model. Comments and corrections are welcome.

Background¶

In January 2024, an article in the Financial Times, “A new global gender divide is emerging”, reported that in the United States, women aged 18 to 30 had become 30 percentage points more liberal than men of the same age, and that the gap had opened up in six years. The figure was based on Gallup data, backfilled with the General Social Survey (GSS), and adjusted by subtracting the trend in the overall population.

I tried to replicate that result, and wrote four posts about it:

• In “Is the Ideology Gap Growing?” I found that in the GSS, young women are more likely than young men to say they are liberal, by 5–10 percentage points, but the only evidence that the gap was growing came from the 2022 data point for young men, which is probably an error: the sampling weights move it by almost 30 percentage points.

• In “Political Alignment, Affiliation, and Attitudes” I extended the analysis to party affiliation, where the gap among young people was almost 20 points in the 1990s and 5–10 points recently, and to 15 policy questions, where the gap was small and not growing.

• In “The Political Gender Gap is Not Growing” I found the same in the American National Election Studies (ANES): the gap in political alignment appeared in the 1980s and has been nearly constant since.

• In “The Gender Gap in Political Beliefs Is Small” I found that on 30 policy questions in the ANES, men and women give similar answers, so the political gap is mostly in what people say about themselves.

Since then, both surveys have released 2024 data. And the question has shifted a little. The claim I hear now is not just that there is a gender gap among young people, but something like this: young men are about as conservative as we would expect, based on long-term trends, and young women are more liberal than we would expect.

That’s a claim about a baseline, so to test it we need to say what “expected” means, and we need a model that can estimate the relevant quantities with honest uncertainty. That’s what this report does.

The question¶

Here’s the version of “expected” I’ll use.

For each sex, I compute how far people aged 18 to 29 sit from their own sex overall: for example, the percentage of young women who say they are liberal, minus the percentage of all women who say they are liberal. I’ll call that the young offset. It is close in spirit to the FT adjustment, which subtracts the trend in the general population, except that I subtract the trend for each sex separately.

If young women are “more liberal than expected”, their offset should have grown relative to young men’s. So the quantity I’ll focus on is a difference in differences: the change in young women’s offset since about 2010, minus the change in young men’s offset over the same period. Young men are the natural comparison because they lived through the same years at the same age. If the two offsets moved together before 2010 and young women’s moved further after, that is evidence young women are more liberal than the earlier pattern predicts.

I also considered two other baselines. One is to forecast each group from its trend before some cutoff year, and compare the forecast with what happened after; but smooth trends extrapolate with intervals that widen quickly, so this test can rarely detect anything over a decade. The other is a cohort model, comparing the newest birth cohorts with the trend across earlier cohorts; I built one, and it can’t detect a departure confined to the newest cohorts (see the appendix).

The model¶

The survey questions are seven-point scales. The GSS asks:

We hear a lot of talk these days about liberals and conservatives. I’m going to show you a seven-point scale on which the political views that people might hold are arranged from extremely liberal–point 1–to extremely conservative–point 7. Where would you place yourself on this scale?

The party identification question, in both surveys, also produces seven categories, from strong Democrat to strong Republican, with independents who lean toward a party in between.

In my earlier posts, I collapsed these scales into three groups and plotted percentages. That’s easy to understand, but it has two problems. First, it throws away information: a move from “slightly liberal” to “liberal” doesn’t show up. Second, the percentages for men and women, or young and old, are separate quantities with nothing that puts them on a common scale.

So for this report, I used an ordinal model. It assumes that each group of respondents has a position on a hidden (latent) scale from liberal to conservative, and that the survey responses come from that position: a group further to the right is more likely to choose the conservative categories. The boundaries between categories, called cutpoints, are the same for everyone: women and men, all ages, all years. That’s what puts every group on one scale, so it makes sense to say that one group is further left than another, or has moved further left over time.

The groups are women and men in three age bands, 18–29, 30–49, and 50 and older. For each group and survey year, the model estimates a position, built from a few smooth curves:

• a trend shared by everyone;

• a gap between women and men;

• an offset for each age band, and a separate offset for each age band and sex.

Each curve is a random walk on the slope, which lets it bend but penalizes wiggles, so years with small samples borrow strength from their neighbors. The model also includes a term that lets the spread of positions change over time, which matters when the share of moderates or independents rises or falls, and terms for year-to-year noise beyond what sampling explains. The likelihood uses the sampling weights, with sample sizes reduced to the effective sample size each year.

From the fitted model I compute the quantities in the previous section, as percentages of each group in each category. The shaded areas in the figures are 84% intervals: if the intervals for two estimates don’t overlap, the difference between them is approximately significant at the 5% level.

Before trusting the model, I checked it two ways. First, I generated fake data from known curves and checked that the model recovers them; it does, including a recent change in young women’s offset. Second, I compared the model’s predictions with the data; for ideology and party, in both surveys, 84–97% of observed values since 2010 fall inside the predicted intervals, close to the expected 84%. The appendix has the details, including several versions I tried and dropped.

Results: ideology¶

The following figure shows the young offset in the liberal share — the percentage of 18–29-year-olds who say they are liberal, minus the percentage of all respondents of the same sex — for women and men. The left panel is the GSS, and the right panel is the ANES.

In the GSS, the young offset was 3 points for both sexes in 1990 and 7–8 points in 2010. In 2024, it was 11.5 points for young women and 7.2 for young men. In the ANES, it was 4–5 points for both in 1990; by 2008 it was 11.6 points for women and 9.0 for men; in 2024 it was 12.2 and 8.5.

So in both surveys, young people moved further to the left of their own sex between 1990 and 2010, and young women and men moved together, more or less. In the GSS, they diverged after 2010. In the ANES, the difference opened between 1990 and 2008 and hasn’t grown since.

Here’s the difference in differences, with 84% intervals:

In the GSS, young women’s offset grew 3.4 points more than young men’s, and the interval excludes zero. But as I found in 2024, the 2022 data point for young men looks wrong, so I ran the model again without the 2022 wave. Without it, the difference in differences is 1.7 points, and the interval includes zero. In the ANES, the difference is 1.2 points, with an interval from −2.7 to 4.0.

The next figure shows the gap between women and men in the liberal share, by age.

In the GSS, the gap among 18–29-year-olds was 4.1 points in 2010 and 8.7 points in 2024; among older groups it was 3–4 points in both years. In the ANES, the gap among 18–29-year-olds was 8.8 points in 2008 and 12.3 points in 2024; it also grew among 30–49-year-olds (6.2 to 8.7) and people 50 and older (4.2 to 7.3). Without the 2022 wave, the GSS gap among young people in 2024 is 5.5 points.

For comparison with the FT figure, which shows the percentage who say they are liberal minus the percentage who say they are conservative, the same “net” gap among 18–29-year-olds in 2024 is 15.9 points in the GSS (10.0 without 2022) and 22.7 points in the ANES.

Those numbers are not adjusted for the overall trend, so they are not directly comparable with the FT’s 30 points, but they are in the same direction and smaller.

Interpretation¶

Two things are well supported. In both surveys, young women are more likely than young men to say they are liberal, by 9–12 points in 2024. And the gap among young people is bigger than among older people.

The claim that young women are more liberal than expected, meaning their offset has grown faster than young men’s, has weaker support. The difference in differences is positive in all three estimates, between 1 and 3.5 points of liberal share. But only the GSS estimate that includes the 2022 wave excludes zero, and that wave is the one I flagged as suspicious in 2024. Without it, the GSS and the ANES agree: a small widening that the data can’t distinguish from none.

The ANES has one more caution. In recent waves, 15–25% of ANES respondents say they “haven’t thought much about” where they are on the liberal-conservative scale, and young women say so more often than young men (24% and 16% in 2020–2024). Those respondents are dropped from the analysis, so the young women who remain are a more selected group than the young men. When I count them as moderates instead, the difference between young women’s and young men’s offsets disappears.

As for young men, they are where the comparison says they should be, in one sense: their offset in the GSS was 7.0 points in 2010 and 7.2 in 2024, so it hasn’t changed. But it was rising before 2010, so against their own earlier trend, young men stopped getting more liberal relative to men overall. Without the 2022 wave, even that stall mostly goes away (7.3 to 8.5). In both surveys, young men are more likely than men overall to say they are liberal, by 7–9 points.

Results: party identification¶

For party identification, I collapsed the seven categories into three: Democrat (including independents who lean Democratic), Independent, and Republican (including those who lean Republican). With all seven categories, the model can’t fit the data, because the share of pure independents and the share of strong partisans grew at the same time in the GSS, while weak partisans and leaners declined; a model with one position per group can’t produce that pattern. With three categories, and a spread that varies by age band — young people are more likely to be independents than the model would otherwise predict — the model fits well.

The next figure shows the young offset in the Democratic share, for women and men.

In both surveys, young people were less likely than their elders to identify as Democrats in the 1980s, by 4–7 points in 1990, and the offset has risen since. In the GSS, the young offset in 2024 is 1.5 points for women and −1.6 for men; in the ANES, it is 4.6 and 4.3. Young women and men have moved together for most of the period in both surveys.

None of these intervals excludes zero, and the ANES estimate goes the other way.

The next figure shows the gap between women and men in the Democratic share, by age.

In the GSS, the gap among 18–29-year-olds was 5.1 points in 2010 and 11.4 points in 2024, compared with 7.5 points among older groups; without the 2022 wave, the young gap in 2024 is 7.3 points. In the ANES, the gap in 2024 is 6–7 points in every age group; among young people it was 8.2 points in 2008 and 6.8 in 2024.

I also checked the ANES pilot studies, which fill in some of the years between the presidential surveys (2018, 2019, 2022), but they are opt-in online panels with small samples of young people, and they differ from the main surveys in years where both are available. They don’t add usable information about young people (see the appendix).

Interpretation¶

In party identification, there is no evidence that young women are more Democratic than expected. Among people 30 and older, women are more likely than men to identify as Democrats, by 6–8 points in both surveys, and young people of both sexes have moved toward the Democrats relative to their elders since 1990. But young women and young men moved together. The one hint of a widening gap among young people, in the GSS, depends on the 2022 wave, and the ANES shows nothing like it.

This is consistent with what I found in 2024: in the ANES, and in the GSS without the 2022 wave, the gap among young people in party affiliation is 5–10 points, and not growing.

Conclusions¶

Here’s what I conclude.

Ideology. Young women are more likely than young men to say they are liberal, by 9–12 percentage points in 2024, and that gap is bigger than among older people. The claim that young women are more liberal than expected — that they have moved left of their own sex faster than young men have — is supported only weakly. The difference in differences is between 1 and 3.5 points of liberal share in three estimates, and the only one whose interval excludes zero depends on a GSS data point that is probably an error. In the ANES, the result also depends on how we handle respondents who haven’t thought about the question.

Young men. Young men are more liberal than men overall, as they have been since the 1970s. In the GSS, their offset hasn’t changed since 2010, so in that sense they are about as liberal — or as conservative — as expected; but that’s partly because the 2022 data point pulls them to the right. There is no evidence in either survey that young men have become more conservative relative to men overall.

Party identification. There is no evidence that young women are more Democratic than expected, or that young men are more Republican than expected. Young women and young men have moved together.

Compared with 2024. My conclusions are mostly the same as in my 2024 posts. With a model that puts everyone on one scale and two more years of data, the GSS shows a small widening of the ideology gap among young people, measured against each sex’s own average, but not one I would bet on, and not one the ANES confirms. Nothing in either survey is close to a 30-point gap that opened in six years.

Appendix: data and model checks¶

Data¶

• GSS: polviews, 1974–2024, and partyid, 1972–2024 (with “other party” dropped), weighted with wtssps. GSS waves are mapped to a two-year grid (2021 counts as 2020). “Don’t know” is 2–3% of respondents since 2019 and is dropped.

• ANES: the Cumulative Data File, VCF0803 (liberal-conservative self-placement, 1972–2024) and VCF0301 (party identification, 1952–2024), weighted with VCF0009z. Since 2004, the ANES time series is every four years.

• “Haven’t thought.” In the ANES, 15–25% of respondents in recent waves say they haven’t thought much about where they are on the liberal-conservative scale, more often women than men, and more often young women than young men. Main fits drop them; a second fit counts them as moderate.

• ANES 2012. About 3,800 respondents interviewed online in 2012 weren’t offered the “haven’t thought” option (none chose it, compared with 20–26% face to face), so their answers come from a different format. They are dropped for ideology and kept for party identification, where “don’t know” is under 1%.

• GSS 2022. In 2024 I found that the sampling weights move the 2022 estimate for young men by almost 30 points. The main fits keep the 2022 wave; the results sections also report fits without it.

• ANES pilots. The 2016, 2018, 2019, 2020 and 2022 pilot studies ask party identification the standard way, but they are opt-in online panels; only 2022 asks the liberal-conservative question with the time-series wording. Among 18–29-year-olds (94–257 per sex), the pilots show 17–32% pure independents, compared with 11–18% in the model fitted to the time series, including in 2016 and 2020, when both are available. The difference between young women and young men in the Democratic share changes sign from one pilot to the next. I used them as a check only.

The model¶

For group g (sex × age band) in year t, with sex sign s (+1 for women, −1 for men) and age band a, the latent position is

η = μ(t) + s · gap(t) / 2 + age(a, t) + s · agegap(a, t) / 2 + shocks,

and the probability of answering in category k or below is the logistic function of (cₖ − η) / spread(t, a), with cutpoints cₖ shared by all groups and years. The curves are second-order random walks; the age terms sum to zero across age bands. Counts are weighted and reduced to the effective sample size of each group and year, with a multinomial likelihood. The spread varies over time for ideology, and over time and by age band for party identification. Code: notebooks/ordinal/ordinal.py in the repository for this project.

Checks¶

• Fit. For each group, year and collapsed category, I checked whether the observed percentage falls inside the model’s 84% predictive interval. Since 2010, the coverage is 84% for GSS ideology, 97% for ANES ideology, and 97% for party identification in both surveys. The weakest fit is the share of moderates in the GSS: women are more likely than men to choose “moderate”, by more than one shared scale can explain, mostly at the expense of “slightly conservative”. The liberal and conservative shares, which the results rely on, fit well.

• Recovery. I simulated data with known curves, including a change in young women’s offset in the most recent years, using the same sample sizes as the real data. The model recovered the young offset within its 84% interval for 94% of years in the GSS design and 82% in the ANES design.

• Things I tried and dropped. A spread that differs between women and men (small, and confounded with the gap when there are only three categories). A seven-category model for party identification (can’t fit the simultaneous growth of independents and strong partisans). A period-cohort version, with one curve over survey years and one over birth years: in a simulation, it recovered only a quarter of a change confined to the newest cohorts, with or without heavier-tailed curves, because those cohorts appear in only a few surveys, at young ages; and without an age term, it attributes changes over the life course to period and cohort.