Probably Overthinking It

In previous articles (here, here, and here) I’ve looked at evidence of a gender gap in political alignment (liberal or conservative), party affiliation (Democrat or Republican), and policy preferences.

Using data from the GSS, I found that women are more likely to say they are liberal, and more likely to say they are Democrats, by 5-10 percentage points. But in their responses to 15 policy questions that most distinguish conservatives and liberals, men and women give similar answers.

In other words, the political gap is mostly in what people say about themselves, not in what they believe about specific policy questions.

Now let’s see if we get similar results with ANES data. As with the GSS, I looked for questions where liberals and conservatives give different answers. From those, I selected questions about specific policies, plus four questions related to moral foundations, with preference for questions asked over a long period of time. Here are the 16 topics that met these criteria:

For each question, I identified one or more responses that were more likely to be given by conservatives, which is what I’m calling “conservative responses”.

Not every respondent was asked every question, so I used a Bayesian method based on item response theory to fill missing values. You can get the details of the method here.

As in the GSS data, the average number of conservative responses has gone down over time.

Men give more conservative responses than women, on average, but the differences is only half a question, and the gap is not getting bigger.

Among people younger than 30, the gap is closer to 1 question, on average. And it is not growing.

In summary:

• In the ANES, there is no evidence of a growing gender gap in political alignment, party affiliation, or policy preferences.
• In both the GSS and the ANES the gap in policy preferences is small and not growing.

The details of this analysis are in this Jupyter notebook.

What about economics?

Many of the questions in the previous section are about social issues. On economic issues some of the patterns are different. Here are 15 questions I selected that are mostly about federal spending.

Unlike the social issues, which trend liberal over time, responses to these questions are almost unchanged.

In the general population, the gender gap is about 0.5 questions and not growing.

Among young adults, the gender gap is smaller, and not growing.

On a total of 30 questions where conservatives and liberal disagree, men and women provide similar responses.

I am happy to announce the third edition of Think Python, which will be published by O’Reilly Media later this year.

You can read the online version of the book here. I’ve posted the Preface and the first four chapters — more on the way soon!

You can read the Early Release and pre-order from O’Reilly, or pre-order the third edition on Amazon.

Here is an excerpt from the Preface that explains…

What’s new in the third edition?

The biggest changes in this edition were driven by two new technologies — Jupyter notebooks and virtual assistants.

Each chapter of this book is a Jupyter notebook, which is a document that contains both ordinary text and code. For me, that makes it easier to write the code, test it, and keep it consistent with the text. For readers, it means you can run the code, modify it, and work on the exercises, all in one place.

The other big change is that I’ve added advice for working with virtual assistants like ChatGPT and using them to accelerate your learning. When the previous edition of this book was published in 2016, the predecessors of these tools were far less useful and most people were unaware of them. Now they are a standard tool for software engineering, and I think they will be a transformational tool for learning to program — and learning a lot of other things, too.

The other changes in the book were motivated by my regrets about the second edition.

The first is that I did not emphasize software testing. That was already a regrettable omission in 2016, but with the advent of virtual assistants, automated testing has become even more important. So this edition presents Python’s most widely-used testing tools, doctest and unittest, and includes several exercises where you can practice working with them.

My other regret is that the exercises in the second edition were uneven — some were more interesting than others and some were too hard. Moving to Jupyter notebooks helped me develop and test a more engaging and effective sequence of exercises.

In this revision, the sequence of topics is almost the same, but I rearranged a few of the chapters and compressed two short chapters into one. Also, I expanded the coverage of strings to include regular expressions.

A few chapters use turtle graphics. In previous editions, I used Python’s turtle module, but unfortunately it doesn’t work in Jupyter notebooks. So I replaced it with a new turtle module that should be easier to use. Here’s what it looks like in the notebooks.

Finally, I rewrote a substantial fraction of the text, clarifying places that needed it and cutting back in places where I was not as concise as I could be.

I am very proud of this new edition — I hope you like it!

In a previous article, I used data from the General Social Survey (GSS) to see if there is a growing gender gap among young people in political alignment, party affiliation, or political attitudes. So far, the answer is no.

• Young women are more likely than men to say they are liberal by 5-10 percentage points. But there is little or no evidence that the gap is growing.
• Young women are more likely to say they are Democrats. In the 1990s, the gap was almost 20 percentage points. Now it is only 5-10 percentage points. So there’s no evidence this gap is growing — if anything, it is shrinking.
• To 15 questions related to policies and attitudes, young men give slightly more conservative responses than women, on average, but the gap is small and consistent over time — there is no evidence it is growing.

Ryan Burge has done a similar analysis with data from the Cooperative Election Study (CSE). Looking at stated political alignment, he finds that young women are more likely to say they are liberal by 5-10 percentage points. But there is no evidence that the gap is growing.

That leaves one other long-running survey to consider, the American National Election Studies (ANES). I have been meaning to explore this dataset for a long time, so this project is a perfect excuse.

This Jupyter notebook shows my analysis of alignment and party affiliation. I’ll get to beliefs and attitudes next week.

Alignment

This figure shows the percent who say they are liberal minus the percent who say they are conservative, for men and women ages 18-29.

It looks like the gender gap in political alignment appeared in the 1980s, but it has been nearly constant since then.

Affiliation

This figure shows the percent who say they are Democrats minus the percent who say they are Republicans, for men and women ages 18-29.

The gender gap in party affiliation has been mostly constant since the 1970s. It might have been a little wider in the 1990s, and might be shrinking now.

So what’s up with Gallup?

The results from GSS, CES, and ANES are consistent: there is no evidence of a growing gender gap in alignment, affiliation, or attitudes. So why does the Gallup data tell a different story?

Here’s the figure from the Financial Times article again, zooming in on just the US data.

First, I think this figure is misleading. As explained in this tweet, the data here have been adjusted by subtracting off the trend in the general population. As a result, the figure gives the impression that young men now are more likely to identify as conservative than in the past, and that’s not true. They are more likely to identify as liberal, but this trend is moving slightly slower than in the general population.

But misleading or not, this way of showing the data doesn’t change the headline result, which is that the gender gap in this dataset has grown substantially, from about 10 percentage points in 2010 to about 30 percentage points now.

On Twitter, the author of the FT article points out that one difference is that the sample size is bigger for the Gallup data than the datasets I looked at — and that’s true. Sample size explains why the variability from year to year is smaller in the Gallup data, but it does not explain why we see a big trend in the Gallup data that does not exist at in the other datasets.

As a next step, I would ideally like to access the Gallup data so I can replicate the analysis in the FT article and explore reasons for the discrepancy. If anyone with access to the Gallup data can and will share it with me, let me know.

Barring that, we are left with two criteria to consider: plausibility and preponderance of evidence.

Plausibility: The size of the changes in the Gallup data are at least surprising if not implausible. A change of 20 percentage points in 10 years is unlikely, especially in an analysis like this where we follow an age group over time — so the composition of the group changes slowly.

Preponderance of evidence: At this point see a trend in one analysis of one dataset, and no sign of that result in several analyses of three other similar datasets.

Until we see better evidence to support the surprising claim, it seems most likely that the gender gap among young people is not growing, and is currently no larger than it has been in the past.

Is there a growing gender gap in the U.S?

Alignment

A recent article in the Financial Times suggests that among young people there is a growing gender gap in political alignment on a spectrum from liberal to conservative.

In last week’s post, I tried to replicate this result using data from the General Social Survey. I generated the following figure, which shows the percentage of liberals minus the percentage of conservatives from 1988 to 2021, among people 18 to 29 years old. The analysis is in this Jupyter notebook.

Women are more likely to say they are liberal by 5-10 percentage points. But there is little or no evidence that the gap is growing.

Party Affiliation

This figure shows the percentage of Democrats minus the percentage of Republicans from 1988 to 2021. The analysis is in this Jupyter notebook.

Women are more likely than men to say they are Democrats. In the 1990s, the gap was almost 20 percentage points. Now it is only 5-10 percentage points. So there’s no evidence this gap is growing — if anything, it is shrinking.

Attitudes and beliefs

To quantify political attitudes, I will take advantage of a method I used in Chapter 12 of Probably Overthinking It. In the General Social Survey, I chose 15 questions where there is the biggest difference in the responses of people who identify as liberal or conservative. Then I estimated the number of conservative responses from each respondent.

The following figure shows the average number of conservative responses for young men and women since 1974. The analysis is in this Jupyter notebook.

Men give slightly more conservative responses than women, on average, but the gap is small and consistent over time — there is no evidence it is growing.

In summary, GSS data provides no support for the claim that there is a growing gender gap in political alignment, affiliation, or attitudes.

The video from my PyData Global 2023 talk, Extremes, outliers, and GOATS, is available now:

The slides are here.

There are two Jupyter notebooks that contain the analysis I presented:

• Chapter 4 of Probably Overthinking It
• An exploration of Gaussian and lognormal models for anthropomorphic measurements

Here’s the abstract:

The fastest runners are much faster than we expect from a Gaussian distribution, and the best chess players are much better. In almost every field of human endeavor, there are outliers who stand out even among the most talented people in the world. Where do they come from?

In this talk, I present as possible explanations two data-generating processes that yield lognormal distributions, and show that these models describe many real-world scenarios in natural and social sciences, engineering, and business. And I suggest methods — using SciPy tools — for identifying these distributions, estimating their parameters, and generating predictions.

This talk is based on Chapter 4 of Probably Overthinking It. If you liked the talk, you’ll love the book 🙂

Thanks to the organizers of PyData Global and NumFOCUS!