1. Data and empirical strategy

We use all ANES data covering elections from 1956 to 2018.Footnote ¹ Our outcome of interest is self-reported turnout in election s _n, predicted by a dummy that denotes whether election s _n is of same type (Presidential or Midterm) as election s ₁—each respondents’ first eligible election. We obtain this information by subtracting individuals’ age from the year of the survey, which provides us with their year of birth. Knowing the year of birth allows us to predict with high accuracy individuals’ first eligible election.Footnote ² We use no information as to whether respondents actually voted in their first eligible election. Rather, we stick to the reduced form, employing type-of-election eligibility as a predictor of turnout in the year of the survey. Moreover, we omit from the analysis the elections used to classify respondents in terms of their treatment status, i.e. individuals’ first pair of Presidential and Congressional elections. Additional information on the data can be found in the online Appendix.

Identification relies on two assumptions: that potential turnout rates in same-type elections would be the same for both groups had there been no difference in their type of first-eligible election (ignorability); and that type-of-first-election eligibility does not affect future turnout in same-type elections through other ways except via its effect on the type of election one encounters when coming of age (exclusion).Footnote ³ The first assumption, ignorability, would be violated if individuals (or their parents) could in some way select themselves (or their children) into the type of first-eligible election, a rather unlikely scenario.

Exclusion would be violated if type-of-election eligibility alone affected turnout patterns without this effect going through the experience of the election itself. This could be the case if for example Presidential-eligibles ended up being for some reason older or more educated than Midterm-eligibles. Figure 1 illustrates the logic upon which the design is based. The alternation of black and grey stripes in the left-hand panel suggests the absence of any monotone association between year-of-birth and type-of-election eligibility. It is this alteration we aim to leverage for identification purposes. We believe that eventually belonging in one of the two groups is as-good-as-random. The right-hand side of the graph, which compares the age distribution between the two groups, does suggest a high degree of overlap. That said, we do find some imbalances in the age and level of education of respondents, which, as Table I shows, are statistically significant. To address this concern, we employ various specifications, including all or some of these variables as covariates. We also include a dummy for whether the current election is Presidential or Midterm. Finally, since survey respondents vary in age, they also vary in terms of how distant the outcome election, s _n, is from the treatment election, s ₁. For some people their outcome election is the third or fourth Presidential or Midterm election in which they are of age to vote, while for others it might be their 10th eligible election. To account for such heterogeneity, we include election-counter fixed effects.

Figure 1. Visualizing the identification strategy: comparing groups according to the type of first eligible election.

Table 1. Difference in means between Presidential and Midterm eligibles

Another way in which exclusion could be violated is if Presidential- or Midterm-year eligibility affects future turnout in Presidential or Midterm elections respectively via other routes than the first-eligible election itself. For example, if high-school curriculum also changes in years of Presidential or Midterm elections, it could be that effects are not due to the election itself but rather due to cohort-specific changes in school curriculum.Footnote ⁴ We cannot test this assumption, but we try to gauge whether it holds by looking into several post-treatment outcomes that are not election-specific. Our treatment is based on the idea that similar contextual stimuli as those experienced in first eligible election should moderate behavioral responses. By extension, we expect that attitudinal outcomes not distinguishing between types of elections remain unaffected by our treatment. We use all such outcomes available in the cumulative ANES survey, as shown in Table 2. The means of the two groups are close to identical across all outcomes. Presidential- and Midterm-eligibles do not differ in terms of their attitudes toward government spending; whether they approve or not the President; whether they think that parties are only interested in casting votes; or whether parties differ. They are also indistinguishable when it comes to regional residence as well as residential mobility. Finally, when looking also at the supplementary items included in the ANES (available for only a third of the total number of respondents) we find that the two groups do not differ in terms of internal or external efficacy, trust in government, opinions about whether voters have a say in politics; whether voting matters; or whether politics is too complicated. Among these 17 items, only one indicates a difference which albeit very small is significant at the 5 percent level. Figure 2 shows that when arraying the p-values obtained from this exercise against a uniform distribution, we find only small variations from the 45 degree line, which corresponds to the level of balance expected under true randomization. Without being in any way exhaustive, the evidence enhances our confidence that the effects reported here are due to type-of-election eligibility.

Figure 2. Balance on observables between Presidential and Midterm eligibles.

Note: The graph displays p-values corresponding to a difference-of-means test between Presidential and Midterm eligibles. The dashed diagonal line denotes the 45-degree line, which is where the p-values would lie, if they came from a uniform distribution. The blue dashed lines denote the 5 percent cutoff point and the red dashed lines denote the 10 percent level.

Table 2. Differences between midterm and presidential eligibles on political outcomes

Thus, our baseline model takes the following form:

(1)$$Vote_{i, s> 1} = \sum_{s = 2}^{15} \alpha_s E_{s} + \lambda_1 Same_{is} + \delta P_s + X\beta_2 + u_i$$

where Same _is is coded as one if the election in question is of same type as first-eligible election and zero otherwise. E _s indexes election-counter dummies, where $s = \{ 2,\; \, \dots ,\; \, 15\}$. Every s represents two elections, one Midterm and one Presidential, which together comprise a pair. Thus, s = 2 means that i is interviewed after the second election she was eligible to participate, either Presidential or Midterm, and so on, up until s = 15. On average, our respondents encounter their first eligible election when they are 20 years old. Thus, going up to s = 15 means that we trace a 60-year-long life trajectory, from 20 to approximately 80 years of age. P _s is a dummy for Presidential elections that controls for differences in turnout between these elections and midterm ones. X is a vector of control variables. In our preferred specification, we include all covariates included in Table 1.Footnote ⁵ In the online Appendix we use two alternative specifications. The first includes only education and age, the only two covariates we deem necessary for identification.Footnote ⁶ The second is intended to assess the robustness of these estimates by including contemporaneous sociodemographics as proxies for socioeconomic background: income, region of residence by the time of the ANES interview, and a trichotomy distinguishing respondents between those residing in central cities, suburban cities, or rural towns (Urbanism).Footnote ⁷ Conditional on these observables, we assume as-good-as-random assignment of type of first eligible election. If this is the case, λ ₁ recovers the effect of same-type election on probability of voting in election s _n. Since eligibility does not mean voting, equation (1) yields the intent-to-treat (ITT) effect of voting in the same type of election as the election in which one was first eligible to vote.

Equation (1) does not distinguish between types of elections. This is done in equation (2), which interacts P _s with Same _is:Footnote ⁸

(2)$$Vote_{i, s> 1} = \sum_{s = 2}^{15} \alpha_s E_{s} + \lambda_1 Same_{is} + \delta P_{s} + \lambda_2 Same_{i} P_{s} + v_i$$

Now, λ ₁ yields the effect of same-type election for Midterm elections and λ ₁ + λ ₂ denote the same effect for Presidential elections. Finally, we decompose the same-type election effects across individuals’ life trajectory. Equation (3) does this, by interacting Same _i with each pair of elections in respondent's life span, up until election pair s = 15:

(3)$$Vote_i = \sum_{s = 2}^{15} \alpha_s E_{s} + \lambda_1 Same_{is} + \sum_{s = 2}^{15} \delta_s Same_{i} E_s + \delta P_{s} + X\beta_2 + u_i$$

For each pair of election, 2 ≤ s ≤ 15, λ ₁ + δ _s yields the effect of same-type-eligibility on the probability of voting in election s. For all elections s > 15, the effect is summarized by λ ₁. Equation (4), below, further distinguishes between each type of election:

(4)$$\eqalign{\matrix{& V_i = \displaystyle\sum_{s = 2}^{15} \alpha_s E_{s} + \lambda_1 Same_{is} + \delta P_{s} + \displaystyle\sum_{s = 2}^{15} \beta_s E_{s} Same_{i} \cr & \quad\quad\quad\quad\quad\quad+ \displaystyle\sum_{s = 2}^{15} \theta_s E_{s} P_{s} + \lambda_2 Same_{i} P_s + \displaystyle\sum_{s = 2}^{15} \zeta_s E_{s} P_{s} Same_{i} + X\beta_2 + u_i }}$$

Now, for each 2 ≤ s ≤ 15, λ ₁ + β _s yields the same-type-of-election effect for all Midterm elections while λ ₁ + β _s + λ ₂ + ζ _s yields the same-type-of-election effect on turnout in Presidential elections.

3. Policy implications

If the type of first eligible election matters, it should also help us understand the effects of policies designed to reduce voting age. Reducing the age threshold of voting eligibility has become increasingly popular as a way of increasing youth political knowledge (Rosenqvist, Reference Rosenqvist2020) and locking young individuals into voting habits (Franklin, Reference Franklin2020). If such policies change the type of first-eligible elections, they may also have hitherto neglected side-effects on the probability of voting for subsequent elections of different types. To shed light on these effects, we leverage the laws of 1972 reducing the voting age from 21 to 18 years old that affected most US States.

The 1972 law made voters born in years such as 1963, 1967, or 1971 first eligible to vote in midterm elections at age 19. In the absence of the reform, they would have been first eligible to vote in a Presidential election at age 21. On the other hand, people born a year before remain first eligible to vote in Midterm elections regardless of the reform. Comparing people aged 19 and 20 at the time of congressional elections before and after the reform, we can therefore test whether making one group to experience congressional elections first affects future turnout. Doing so also presents the advantage of accounting for potential cohort-specific fixed effects, which was impossible with the previous research design.

To test the effect of switching from a presidential to congressional election as a first eligible voting experience, we use the same data as in the previous analyses and estimate the following equation:

(5)$$\eqalign{\matrix{& Voted{\scriptstyle i,\; s> 1} = \gamma_{0}Change_{i} + \gamma_{1}Presidential_{i}Change_{i} + \delta Presidential_{i}\cr & \qquad\qquad\qquad\quad\quad\quad+ \left[\displaystyle\sum_{s = 2}^{15} \alpha_{0s}E_{s}\right] + X_{i}\beta_{0} + Presidential_{i}\left(\left[\displaystyle\sum_{s = 2}^{15} \alpha_{1s}E_{s}\right] + X_{i}\beta_{1}\right) + v_{is} }}$$

In the equation, Change equals one if the individual is 19 years old in the year of congressional elections and is born after 1955—thereby experiencing a switch in their first eligible election from Presidential to Midterm. Alternatively, Change equals zero if the voter turns 19 in the year of congressional elections and is born before 1955 or turns 20 in the year of such elections and therefore always remains first eligible for presidential elections. We also include our Presidential dummy to control for turnout differences between these two elections as well as an interaction between Presidential and Change. This means that γ ₀ represents the effect of switching to midterm elections on subsequent midterm election turnout while γ ₀ + γ ₁ represents the effect of switching to midterm on subsequent presidential elections’ turnout. Finally, we include the same covariates and election-counter fixed effects as before in a fully interacted model to allow for these to depend on the type of election voters experience. To verify the robustness of our results, we also run the same regressions with state code fixed effects and their corresponding interactions with the presidential dummy. Results do not change, albeit becoming less precise.

Table 3 shows the findings.Footnote ⁹ Regardless of model specification, switching one's first election from Presidential to Midterm increases the likelihood of voting in subsequent Midterm elections by three to four percentage points and decreases the probability of voting in subsequent Presidential elections by between four and five percentage points.

Table 3. Effect of changing one's first election type from Presidential to Midterm on voting for subsequent elections.

Robust standard errors in parentheses

***p < 0.01, **p < 0.05, *p < 0.1

Table C1 in the online Appendix repeats the same exercise focusing on the group who experience a change from Midterm to Presidential elections after the reform. As expected the coefficients are of opposite sign to the ones shown in Table 3.

To make sure that we are not rather measuring the effect of one group now being eligible to vote in an election a year earlier than the other, Table C2 in the online Appendix compares 18 and 19 year olds after the reform who are eligible for the same first election assigning treatment to 19 year olds (models 1 and 2). We find no impact both for congressional and presidential elections. We repeat the exercise comparing 21 and 22 years who became first eligible for the same election pre-1972 (models 3 and 4) and, again, find no effect.