I had some free time this week between conference travel and figured I'd throw something together to really drive home how stupid/bad/lazy that NYT analysis was. I can't find the original post, but mar_k posted this counter from FT which I think does a good job of highlighting how dumb the underlying idea that millennials are somehow drifting inevitably towards conservativism. My initial thought after reading those NYT tweets was that voter makeup, even within just one generation, is probably not static and qualitatively different from one election to the other. As a cohort ages, some members become wealthier, some poorer, some are immiserated and even disenfranchised, so to presume that the monolithic 'millennial' voter is homogenous across time is weird. To be fair, they don't really make that claim, but the not-so-subtle implied mechanism is something in the vein of the If a man is not a socialist by the time he is 20, he has no heart. If he is not a conservative by the time he is 40, he has no brain-quote; so these cohort compositional changes are really important. Maybe generations do become more conservative over time, but maybe this is entirely just the result of attrition? If so, that's a hell of an endogeneity issue.

Anyway, to address this (and because it's freely available) I downloaded Ohio's voter registrations rolls, which includes the complete universe of registered OH voters, their place of residence, party affiliation, voting history, etc. While not perfect, these data are way more useful than voter turnouts, since I can construct a panel and observe individual-voter behavior across time, rather than depend on a clumsy, repeated cross-sectional aggregated count (as an aside- why the fuck are NYT """"data scientists"""" using voter-turnouts? Just about anyone can download voter rolls; it's free in most states, and I'm confident that the Times can spring to spend a few thousand bucks to get their staff access to the data that isn't).

Some limitations to note: First, I'm using a 10% sample (n=792,907, my poor laptop was struggling enough with this, but I can send my R code to anyone that wanted to replicate my results for the whole sample), though the law of large numbers is probably able to safeguard us from any serious bias here; next, this is just Ohio, so not necessarily representative of national-trends- but since we're interested in looking at within-generational changes, these results should still be useful; lastly, because Ohio has open primaries, I can only observe party affiliation based on participation in the primaries when they select which ballot they want to vote on. Obviously, primary-voting-dorks are not the same as general election voters, but again, we're looking at within-cohort changes over time; so unless said dorks are somehow more affected by the immutable drift towards conservatism as they age, these estimates can still capture generalizable trends. At a minimum these data should do a better job at describing the desired underlying changes than voter turnouts.

To perform the analysis I construct a panel from the voter rolls so that I have one observation for every primary election that each voter has participated in, as well as their party affiliation for each election (another note on this: I only have voter history data back to 2000, and the panel is unbalanced since not everyone has been alive and eligible to vote in the same number of elections, nor actually do so in every election that they can- so again, not perfect as I can't completely control for broader attrition effects). I compute the age of each voter for all primaries that they participate in, as well as assign them to one of four generations: Boomer, Gen X, Millennial, Gen Z. To derive the figures in panel A of the top figure, I estimate a linear probability model which regresses the dichotomous party affiliation variable (is Republican) on a series of Generation X Age at Election interaction terms (that is, I regress up to four different variables for each age; e.g. 20-and-a-Boomer, 20-and-a-GenXer, 20-and-a-Millennial, etc.). I cluster standard errors at the unit of treatment, so year of birth. Panel A therefore depicts the conditional relative likelihood for each age to be a voting Republican, by generation. Without any additional input, it looks like there might be some bones to the conservatism-comes-with-age argument: Across the three depicted generations (Zoomers are dropped due to low sample size), voters appear to be more likely to vote Republican as they get older.

HOWEVER, these results are functionally identical to those performed with voter turnout, and do not describe changes in voting patterns from aging. Instead, in panel B, I include unique fixed effects for every voter, as well as for each primary election. The argument here is that we're only interested identifying the effects of aging on party affiliation propensities, so I control for all time-invariant properties (say, your race, ethnicity, educational attainment level of parents, location of birth, etc.- all affect voting behavior, but have nothing to do with aging) with the voter fixed effects. The primary election fixed effect is meant to control for any broader turnout effects at that specific time which could be confounding the aging effect (for instance, whether a particular primary is for an incumbent, the degree of voter registration laws, as well as just general political trends not related to age). As you can see, with the inclusion of these controls, age appears to have essentially no significant effect on party affiliation (the 95% CI stretches well across the 0 axis for Millennials and- while omitted in the figure- does for Boomers as well). In fact, the only significant aging effect is observed for the oldest Gen-Xers.

To more concretely describe this supposed aging effect, I also estimate a linear probability model for party affiliation in 2023 based on historical voting practices. In panel A of this figure, I estimate the likelihood of being affiliated with the democrats by year of birth while controlling for age when first registered to vote, and zip code of residence. I do this for the entire pooled population, as well as for the subsample of individuals who have ever voted Republican. Panel B is the opposite: I estimate the likelihood of Republican party affiliation by year of birth with the same controls, and also estimate this on the subsample of voters who have ever been Democrats. Again, these results refute the inevitability of an aging conservatism: Millennials demonstrate slightly lower (higher) propensities to change party affiliation from Democrats (Republicans) when compared to older generations. Unsurprisingly, older Gen-Xers appear to demonstrate the highest propensities to switch affiliation (cuz they’re fucking rats).

Okay- that’s enough for this wall of text. Holler if yinzes see anything horribly awry or want me to fiddle around with the data more. Vermont, Florida and Washington all seem to have decent-enough data as well, so maybe if time-permits I’ll boot-up my workstation and do with with a few million more observations.