Research Design and Data

The data consists of five experimental studies, which I conducted between 2020 and 2022. The studies include four conjoint experiments with ten country-year samples and one vignette experiment with seven country-year samples. Although pre-registered, this was not for the purpose of testing whether young people are less committed to democracy. This makes the comprehensive scope of the data even more important as we need to rule out that idiosyncratic findings of individual studies were cherry-picked for the purpose of this letter (Wuttke, Gavras, and Schoen Reference Wuttke, Gavras and Schoen2022). Using all of these studies to test the proposition that young people punish undemocratic behaviour less may be more ethically defensible than conducting new, pre-registered studies as that would imply collecting more data from respondents than is strictly necessary to test the proposition. Table 1 summarizes the studies in terms of included countries, year of fielding, design, number of tasks, number of respondents, and number of candidate observations.

Table 1.

Included studies

All studies manipulated undemocratic candidate behaviour in election scenarios, measured respondent age, and had voting intentions as the outcome. The four conjoint experiments follow a paired-profile logic where undemocratic behaviour was assigned randomly for the two candidate profiles independently. Studies 1 and 2 included ten paired-profile scenario tasks, whereas Studies 4 and 5 had fifteen and eighteen tasks, respectively. The vignette experiment was a single-profile, single-task experiment in which respondents were asked to evaluate a political leader from their party who did or did not violate democratic principles without variation in any other candidate attribute. This differs from the conjoint experiments, which all assign several candidate attributes beyond undemocratic behaviour.

All experiments assigned one democratic or undemocratic behaviour to each candidate profile. Therefore, I employ a binary indicator of democratic/undemocratic behaviour that is consistent across the studies. The undemocratic behaviours include violations of the democratic cornerstones of free and fair elections, civil liberties, and/or checks and balances. These violations typically consist of incremental violations such as gerrymandering, altering the media landscape in one's favour, or ignoring court rulings (Graham and Svolik Reference Graham and Svolik2020). All undemocratic behaviours and information about which study they were used in are included, along with illustrations of treatment scenarios and ethical considerations, in Appendix A.

Regarding the outcome, the measures of voting intentions vary slightly between the studies. Studies 1 and 3 employed a five-point scale, asking respondents how likely they would be to vote for each candidate; Study 2 used a similar seven-point voting likelihood scale; and Studies 2 and 5 employed a forced-choice outcome, asking respondents to select between the candidate profiles in an election scenario. I rescale all outcomes to 0–1 such that 0 signals not selecting the candidate (forced-choice) or a low likelihood of voting for the candidate (scales). Correspondingly, 1 signals selecting the candidate (forced-choice) or a high likelihood of voting for the candidate (scales).

Studies 1, 3, 4, and 5 were fielded with Lucid, whereas Study 2 was fielded with Amazon Mechanical Turk (MTurk). These platforms have different advantages and challenges. MTurk respondents largely consist of professional survey takers, and the platform suffers somewhat from fraudulent respondent behaviour. On the other hand, Lucid recruits respondents from a diverse set of sources – even mobile games (Ternovski and Orr Reference Ternovski and Orr2022). Therefore, respondents on Lucid are less professional and more similar to respondents of nationally representative surveys than MTurkers (Coppock and McClellan Reference Coppock and McClellan2019; Ternovski and Orr Reference Ternovski and Orr2022). Indeed, the samples for Studies 1, 3, 4, and 5 meet nationally representative quotas on gender and age.

However, using Lucid entails other challenges – in particular, inattentiveness, which has increased with exploding demands for data from online surveys during and after the pandemic (Ternovski and Orr Reference Ternovski and Orr2022). As expected, Study 2 suffers more from professional survey-taking and lack of representativeness, whereas the remaining studies are challenged by inattentiveness. Although this provides a good point of departure for checking whether the results hold across different types of respondents, I draw on additional data sources – for example, YouGov, pre-pandemic Lucid, and the Russian Election Study – from other studies in Appendix D (see also the Robustness Checks and Auxiliary Analyses section).

Estimation and Modelling

I estimate the average effects of the binary undemocratic behaviour indicator and the interaction between this indicator and respondent age using OLS regression with respondent-clustered standard errors (Hainmueller, Hopkins, and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014). I do this for each country-year sample, for each study, and for the pooled sample. When examining the pooled sample, I estimate country-year and study fixed effects as well as random effects to avoid relying strictly on assumptions about whether the interaction between undemocratic behaviour and age varies across contexts (Incerti Reference Incerti2020; Schwarz and Coppock Reference Schwarz and Coppock2022).

I rely on a simple measure of age in years rather than a generational measure, as the data does not allow disentangling life-cycle effects from generational effects because it has no temporal component (Wuttke, Gavras, and Schoen Reference Wuttke, Gavras and Schoen2022). Moreover, I estimate the interaction with age linearly. This linear approximation allows me to rely on one single estimand and show comparable estimates across studies. I complement the linear model with a flexible kernel estimator provided by Hainmueller, Mummolo, and Xu (Reference Hainmueller, Mummolo and Xu2019) when illustrating the pooled results to check the linearity assumption. The proposition that young people are less committed to democracy gains support if the negative effects of undemocratic behaviour increase with age, generally corroborated by negative interaction terms between undemocratic behaviour and age.

Young People Punish Undemocratic Behaviour Less

Figure 1 shows the average effects of undemocratic behaviour on voting intentions (upper row) and the linear interaction between undemocratic behaviour and age in years (lower row). The left column shows the results for each of the seventeen country-year samples, whereas the right column shows the results for each of the five studies. The upper row shows that undemocratic behaviour produces significant negative effects in all country-year samples and studies. The effects vary between 1.5 and 25.5 percentage points. The pooled random and fixed effects estimates yield effect sizes between what corresponds to a 3.5–5.5 percentage points loss in vote share for undemocratic candidates.

Figure 1.

Average effects of undemocratic behaviour on voting intentions and interaction between undemocratic behaviour and age across country-year samples and studies.

Turning to the direct test of the proposition that young people punish undemocratic behaviour less, the lower row of Fig. 1 largely shows that the negative effect of undemocratic behaviour increases with age, thus supporting the proposition. All studies yield negative interaction coefficients, whereas all but two country-year samples produce negative coefficients. Four out of five study-based interaction coefficients are statistically significant, whereas seven out of seventeen country-year-based coefficients are significant. All four pooled estimates are significantly negative, with p < 0.001. These rather unambiguous results for the study-level and pooled sample estimates are important, as individual country-year samples may not be sufficiently well-powered.

Relatedly, Study 3 simultaneously provides the largest and most statistically uncertain coefficients. The statistical uncertainty is because the vignette experiment is less well-powered than the conjoint experiments, which provide multiple observations per respondent. The study-level coefficient for Study 3 is, in fact, numerically very large and significant statistically. Thus, the finding that young people punish undemocratic behaviour less does not seem to be exclusive to the vignette or conjoint experiments. Moreover, there is no tendency for the findings to be particular to any specific country or region. We find old democracies among the countries with the largest age differences (Italy and some US samples), with medium-sized age differences (UK samples), and with the smallest age differences (other US samples). Similarly, younger democracies such as South Africa and Poland are found among those with the strongest differences, whereas South Korea and the Czech Republic yield the smallest differences.

To assess how substantive the age differences are, Fig. 2 shows the pooled random effects estimates graphically. Illustrations of the fixed effects estimates, which are very similar to the random effects estimates, are included in Appendix B. I complement the linear model with a flexible kernel estimator and include underlying age histograms to document common support in the interaction (Hainmueller, Mummolo, and Xu Reference Hainmueller, Mummolo and Xu2019).Footnote ¹ I illustrate the results for respondents below eighty-five years as there is a lack of common support for respondents aged eighty-five and above, who constitute 0.2 per cent of the data. Depending on which model we focus on, the estimates suggest that the oldest people punish undemocratic behaviour by an effect corresponding to an 8–12 percentage point loss in vote share. By contrast, the youngest people punish undemocratic behaviour by a minuscule 0.5–1 percentage point. Therefore, the differences in effects between young and older people are substantial. Moreover, the kernel estimate shows that the interaction is quite linear.

Figure 2.

Illustration of interaction between undemocratic behaviour (UB) and age on voting intentions. Random effects.

In sum, I find comprehensive and rather consistent support for the proposition that young people punish undemocratic behaviour less than older people. The adverse effects of undemocratic behaviour on voting intentions are smallest among the youngest, strongest among the oldest, and generally increase substantially with age.

Robustness Checks and Auxiliary Analyses

In Appendix C, I provide a series of robustness checks and auxiliary analyses that deal with the issue of respondent attentiveness in different ways, as we may suspect that the effects of undemocratic behaviour are attenuated for young people because they are less attentive to surveys (Ternovski and Orr Reference Ternovski and Orr2022). First, I show that the results are robust to controlling for attentiveness. Second, I show that the results are fairly similar for inattentive and attentive respondents. Third, I investigate whether age differences are similar to other factors, such that effects increase with age regardless of what we estimate the effects of. I estimate the effects of other controversial candidate attributes, such as anti-redistributive and anti-minority/nativist behaviour and the effects of partisanship and policy agreement between respondents and candidates. I manipulated or measured these factors in Studies 1, 2, and 4. Partisanship was also measured in Study 5.

Unlike the effects of undemocratic behaviour, age differences in the effects of policy agreement are heterogeneous across studies. These effects increase with age in Studies 1 and 4 but decrease with age in Study 2. The negative effects of anti-minority or nativist behaviour – which includes restrictions on immigrants, abortion, and gay marriage – are clearly largest among young people and decrease with age. The effects of anti-redistributive behaviour are largely homogeneous across age but – if anything – also tend to decrease with age. Finally, the effects of partisanship increase with age, but – given the remainder of the evidence speaking against differences being due to attentiveness – this may simply be because young people are less attached to parties than older people (Dalton and Weldon Reference Dalton and Weldon2007, 185). In sum, the results are robust to accounting for survey attentiveness, whereas age differences on effects of several other factors follow patterns that are distinct from the differences related to sanctioning of undemocratic behaviour. Thus, survey attentiveness does not seem to attenuate the effects of undemocratic behaviour among young people.

In Appendix D, I conduct an additional analysis, where I draw on available data from eight related published studies (Aarslew Reference Aarslew2023; Carey et al. Reference Carey2022; Graham and Svolik Reference Graham and Svolik2020; Krishnarajan Reference Krishnarajan2023; Lewandowsky and Jankowski Reference Lewandowsky and Jankowski2023; Mares and Visconti Reference Mares and Visconti2020; Reuter and Szakonyi Reference Reuter and Szakonyi2021; Saikkonen and Christensen Reference Saikkonen and Christensen2022), which, among other sources, include data from YouGov, pre-pandemic Lucid, and the Russian Election Study. This additional analysis shows that the results are robust to using survey platforms other than Lucid and MTurk, as the interaction between undemocratic behaviour and age – compared with the original pooled effects shown in Fig. 1 – are even stronger when pooled across these other studies. Finally, in Appendix E I show that the findings are quite externally robust to possible differences between the included samples and populations of interest using the the approach proposed by Devaux and Egami (Reference Devaux and Egami2022).