The forthcoming article”Is Voting Habit Forming? New Evidence from Experiments and Regression Discontinuities” by Alexander Coppock and Donald Green is summarized by the authors here:
Voting begets voting. Not just in the sense that current voters are more likely to vote in the future than current non-voters, but in the sense that voting in one election is itself responsible for raising the probability that an individual will vote in future elections. The size of this effect varies substantially by place, time, and individual voter, but we estimate the average effect to be approximately 10 percentage points. Put another way, consider two identical individuals both of whom have 30% chance of voting in 2012. If one of them were randomly induced to vote in 2008, our estimates suggest she would have 40% chance of voting in 2016.
Estimating the impact of voting on voting is difficult because voters and non-voters differ in many ways, most of which is unobservable. In order to understand the effects of voting itself, we need to create or exploit a situation in which individuals are similar except for the fact that some vote and others don’t. We take advantage of two such situations: large-scale Get-Out-The-Vote (GOTV) experiments and the voting-age eligibility cutoff. Both strategies have been used before, but our article applies them to many new circumstances, allowing us to document how much the effect varies across contexts and individuals.
Large-scale GOTV experiments aid the study of voter habit by randomly encouraging some people to vote but not others. For example, in the 2006 social pressure experiment, some voters were sent mailers prior to the August 2006 election that displayed their past vote history and indicated that researchers would be watching to see if they turned out or not (the so-called “Self” mailing). Those in the control group turned out at a rate of 31%, those who got the Self mailing turned out at 36% – the treatment increased turnout by 5 percentage points.
Interestingly, the treatment versus control difference in voting behavior persisted for at least the next 6 years: in every August election between 2006 and 2012, the treatment group voted at a statistically significantly higher rate. The figure below shows how the voting rates of the two randomly formed groups evolved over time.
In order to estimate the causal effect of the act of voting itself, we have to invoke an assumption (the exclusion restriction) that states that the reason the treatment group continues to vote at a higher rate was because they were more likely to vote in the 2006 election – and not any features of the initial mailer itself. Under the exclusion restriction, we can estimate the effect of voting on future voting among “compliers” – those who vote if and only if they are encouraged by the mailers to do so. For example, in August 2010, the difference in turnout between the treatment and control groups was a statistically significant 0.8 percentage points. Our estimate of the effect of voting on voting, then, is 0.008 / .05 = 0.16, or an estimated treatment effect of 16 percentage points.
This pattern holds not only for this experiment but also for similar experiments conducted in 2007 and 2009.
The Eligibility Cutoffs
The voting age eligibility cutoff at 18 years of age affords another opportunity for us to study groups that are otherwise similar except for voting. In this setup, those who are “just-eligible” form our “treatment group” and those who are “just-ineligible” our “control group”. Our unit of analysis is the birthdate cohort and our dependent variable is the total number of votes cast by that birthdate cohort. (This choice gets us around the problem that voter files do not include people who aren’t registered to vote.).
The graph below has three panels. In the top panel, we plot the raw 2008 vote totals, by birthdate, of those who were born within a year on either side of the 2008 eligibility cutoff. The Florida Secretary of State records approximately 250 votes per birthdate for those just-eligible to vote, and (appropriately) 0 votes per birthdate among the ineligible.
The middle panel shows the 2012 vote totals – the estimated difference at the cutoff is approximately 25 votes. There are two interesting features of the raw data that this plot makes clear: first, you can see the separation between weekdays and weekend days: fewer babies are born on the weekend. Second, you can see the seasonality in birth trends: more babies are born in the fall.
The bottom panel accounts for both sources of variation by subtracting off the vote totals of those born one year prior. (We were careful to match days of the week, so this lag is not exactly 365 days.) When we account for the one-year lag, a large source of variation disappears – we no longer see either the day-of-the-week pattern or the seasonality trend. The estimate at the cutoff is still approximately 25 votes per birthday cohort.
Again under an exclusion restriction, our estimate of the effect of voting on future voting is the ratio of the two estimates: 25/250 = 0.10 or 10 percentage points. We chose to show the Florida estimates here because the estimate for Florida is very close to our average estimate.
Variation in the estimates
The 10 percentage point average masks quite a bit of variation in the voter habit effect. In our paper, we document how the effect varies depending on the type of election, the closeness of the election, and the state. It’s also important to reiterate that both the experimental estimates and the eligibility estimates only apply to “compliers,” those who are moved by the initial encouragement to vote.
Encouragements to vote (either from GOTV messages or eligibility at 18) have long-enduring consequences: Individuals who get these encouragements continue for many years to vote at higher rates than their otherwise comparable counterparts. This robust empirical pattern suggests that electoral participation can become a habit acquired through the act of voting itself.