Probably Overthinking It

There’s a famous (but unattributable) quip that goes something like this:

If you are not a liberal at 25, you have no heart. If you are not a conservative at 35, you have no brain.

In Chapter 12 of Probably Overthinking It, I used data from the General Social Survey (GSS) to show that the quip is half right. Older people hold more conservative views. But as people get older, their views don’t become more conservative; on average, they become a little more liberal.

For the book, I used data through 2021 and methods that were good enough. Since then, the GSS has released data from 2022 and 2024, and while working on my next book, You Would Choose Now, I developed better methods for separating the effects of age, birth year, and current events. So it’s time for an update.

This is the first in a series of articles. In this one, we’ll see whether people get more conservative as they get older. The answer is still no.

How to measure conservatism

If someone says people get more conservative with age, what do they mean? I assume that people have in mind some bundle of beliefs, attitudes, and behaviors they consider conservative. But different people might imagine a different bundle.

For Probably Overthinking It, I searched the GSS for the questions that most distinguish liberals and conservatives, and then I curated a list of 15 that cover a variety of topics and were asked often over the course of the survey. That worked well enough, but curating involves judgment calls, which is where researcher bias can sneak in.

This time I replaced curation with a mostly automated selection process that starts with 6,943 variables in the GSS dataset and selects 53 questions that meet the requirements (described at the end of this article).

To combine responses to 53 questions into a single measure, I use a Bayesian IRT (item response theory) model to estimate a latent factor of conservatism for each of 64,632 respondents from 1972 to 2024 who were asked and answered at least five questions. This model makes it possible to compare people who answered different sets of questions, even in cases where there is no overlap. More details about the process are in this technical appendix.

Finally, for each respondent, I express their estimated conservatism as a percentage, which you can think of as the number of conservative responses they would give to a hypothetical pool of 100 questions. Now let’s see some results.

Older people are more conservative

If you think about the people you know, it’s likely that the older ones are more conservative, on average, than the younger ones. To see what that difference looks like in the data, I divided the respondents by the decade when they were interviewed and computed the average percentage of conservative responses as a function of age.

The following figure shows the results, with one line for each decade of the survey from the 1970s to the 2020s.

We can see two trends:

1. In every decade, older people give more conservative answers than younger people. In the 1970s, conservatism increased by about 3.9 percentage points per decade of age. Since then the slope has been a little smaller, about 2.9 points per decade in the 1980s and 1990s, and 2.1 to 2.4 points per decade since 2000. But in every decade, the oldest respondents are 15 to 25 points more conservative than the youngest.
2. Every line is lower than the one before. For example, at every age, people surveyed in the 2020s are more liberal than people the same age surveyed in the 2010s.

So the first half of the quip is right: older people are more conservative. But that doesn’t mean that people become more conservative as they get older.

People don’t get more conservative

To see why, we’ll group people by decade of birth, rather than decade of survey. The following figure shows percent conservative responses as a function of age again; each line represents a decade of birth, from the 1900s to the 2000s.

Within each group, the lines are mostly flat or declining. As people get older, they don’t become more conservative.

If this seems like a contradiction, remember Simpson’s paradox, which I wrote about in this previous post and in Chapter 11 of Probably Overthinking It: the trend in every group goes one way, and the overall trend goes the other way.

However, this figure has a subtle problem. Each line represents people born over a ten-year span, and the mix of birth years changes as we follow the line. At the beginning of each line, it includes only the oldest members of the cohort, and at the end it includes only the youngest. I was aware of this problem when I was working on Probably Overthinking It. Now I have a way to solve it.

Following each birth year

The solution is to estimate a trajectory for each birth year and then aggregate the results, rather than aggregating before estimating. But with less data for each birth year, we need a model to smooth out the estimates.

I developed a model that solves this problem, which I described in this blog post. This model is also the kernel of the analysis I’m doing for my new book-in-progress, You Would Choose Now.

It estimates a component for each year of birth and a component for each year of the survey. Together, they describe a trajectory for each birth year over the years it was observed. The following figure shows those trajectories.

Each line represents a single birth year. For example, the yellow-green line at the top left shows people born in 1900, who were 72 when the survey started in 1972. The purple line at the bottom right shows people born in 2006, who became eligible for the survey when they turned 18 in 2024.

Two patterns are clear:

• Each birth year starts out more liberal than the one before. People born in 1900 gave conservative answers to about two-thirds of the questions; people born after 2000 give conservative answers to about 40%.
• As each birth year ages, its line drifts down. Following the same people over time, their answers become more liberal, not more conservative.

We can decompose these results into a cohort component that shows how conservatism varies from one birth year to the next, controlling for changes over time, and a period component that shows how conservatism varies from year to year, controlling for generational replacement.

The following figure shows the cohort component: the average level of conservatism for each birth year, as if every birth year had been surveyed in every year of the survey.

The decline is steepest between people born in the 1930s and 1950s; it pauses for people born in the 1950s and 1960s, and it continues after that.

The following figure shows the period component: the average level of conservatism in each year of the survey, as if the population had the same mix of birth years throughout.

With the mix of birth years held constant, conservatism declines slowly. From 1972 to 2024, the total change within cohorts is −4.6 percentage points. If we fit a straight line, the rate of change is −0.82 percentage points per decade, with a 90% interval from −0.97 to −0.66. So the change is small, but we can be confident about the direction: following the same people over time, they become a little more liberal.

Finally, we can use these components to break down the total change. Over the course of the survey, average conservatism fell by 16.4 percentage points. Of that, about 11.8 points are due to generational replacement — older, more conservative people leaving the survey population and younger, more liberal people joining it — and about 4.6 points are due to changing minds. So more than two-thirds of the change is generational.

Every group moves left

The model in the previous section is based on the simplifying assumption that the cohort and period components don’t interact. In reality, they might; for example, if an event affects one cohort more than another. To see whether this assumption a reasonable, we can fit the same model separately for each decade of birth, and see what the trend looks like in each group.

The following figure shows the period component for each decade of birth, holding its mix of birth years fixed, so the shifting composition of each group can’t move the line.

Within the groups, the lines are flat or declining. People born in the 1950s moved a little toward conservative until about 1980, and toward liberal after that. The dotted line shows the overall average, which falls faster than most of the groups, because it includes the effect of generational replacement.

The following figure shows the same analysis with respondents grouped by social generation, using the Pew Research Center’s definitions.

The results are the same: no group of birth years, became more conservative over the course of the survey. All of them became more liberal or stayed about the same.

Choose your own adventure

All of these results depend on the set of questions I chose. So if you think I chose badly, you might not be convinced by my conclusion.

For example, you might think the questions I selected put too much weight on social issues and not enough on economic issues. You might not agree with my decision to exclude religious beliefs or confidence in institutions. Or you might think there should be more questions about crime.

In future articles, I’ll run the analysis with different bundles of variables. We’ll see that the results are quite robust — pretty much any bundle of questions that covers a mixture of topics leads to the same conclusions: people do not become more conservative with age.

But if you don’t want to wait, you can try it yourself. I have created an online calculator, called The Overtron, where you can select whatever bundle of questions you want and run the whole analysis in your browser. Try it out, and let me know what you find!

Question selection

Here are the details of the selection process I used. In general, I tried to keep my judgment out of it, but you’ll see that there are a couple of steps where I intervened.

1. Start with everything. The GSS cumulative file has 6,943 variables. I kept the ones that record responses to questions asked in at least 8 waves over a span of at least 15 years, with between 2 and 7 possible answers. That leaves 429.
2. Keep the opinions. From those, I kept the questions that ask for an opinion about a public issue or a social norm. I left out demographics (like age and income), behavior (like smoking or owning a gun), personal circumstances (like happiness and health), and, as I did for the book, religious belief and practice (we’ll revisit that one). I also left out questions about party and voting, which would make the measure circular. That eliminates 259 and leaves 170 questions.
3. Identify the conservative answers. For each question, I split the answers into two groups at the point that best separates people who identify as conservative from people who identify as liberal. The answers conservatives choose more often are the conservative side. For example, if the question is “Do you favor or oppose the death penalty for persons convicted of murder?”, the conservative answer is “Favor”.
4. Measure how well each question separates them. For each question, I estimated how strongly the conservative answer is associated with self-identified ideology, controlling for the year of the survey and the respondent’s year of birth. Steps 3 and 4 don’t eliminate any questions.
5. Coverage. For the final list, a question had to be asked in at least half of the survey years (16 of 32). That’s because the measure compares people surveyed decades apart, so we need long-running questions that overlap in time. If too many questions are short-lived, people surveyed in different years would be measured by different questions, and changes over time would mix real changes in beliefs with changes in the questions. This step eliminates 49 questions and leaves 121.
6. Stability. The difference between conservatives and liberals couldn’t reverse. If the same answer was given more often by conservatives in one decade and by liberals in another, the question is not a stable indicator of conservatism. This step eliminates 1 question and leaves 120.
7. No duplicates. Some questions were asked with two different wordings, on different versions of the survey, so I kept one of each pair. And when several questions cover the same topic, I kept the one that best separates conservatives and liberals, from step 4. This step eliminates 53 questions (12 alternate wordings and 41 others on the same topic) and leaves 67.
8. No confidence questions. I left out 13 questions about confidence in institutions — like the press, organized religion, and the military — because they evaluate institutions rather than take a position on an issue. I made this decision after seeing results that included them, so in a later article I’ll show what difference they make. One of the 13 was already eliminated in step 6, so this step eliminates 12 and leaves 55.
9. Check for drift. A question “drifts” if its connection to conservatism changes over time. This step ended up excluding only two questions, so it doesn’t have much impact on the results. That leaves the 53 questions I used.

The details of this process are in the technical appendix.