Consequences of Lowering the Voting Age to 16: Lessons from Comparative Research

Abstract

Two quasi-mechanical forces push in different directions when we consider consequences of lowering the voting age to 16. On the one hand, lowering the voting age would provide votes to young adults still in school and living in their parental homes. These circumstances should (theory tells us) boost the turnout of those individuals not only at their first election but throughout their ensuing lifetimes. Considering that the previous such reform (lowering the voting age to 18) had the opposite consequences (as this chapter explains), finding a way to undo the deleterious consequences of that reform has a high priority in the minds of many, and Votes at 16 might just do the trick. On the other hand, are sixteen-year-olds mature enough to understand the consequences of their party choices? Or might they rather simply vote more-or-less at random, adding to the volatility of election outcomes that has already been growing apace in recent years? This chapter uses empirical evidence from historic cases in an attempt to evaluate these possibly countervailing effects.

Appendix

Aggregate data for this chapter are taken from the IDEA voter turnout database, augmented by information from Wikipedia regarding the seat shares of the top two parties following each election. Election coverage was as follows: Argentina 1983–2017 (19 elections); Austria 1975–2017 (13); Bolivia 1979–2014 (10); Brazil 1978–2018 (11); Chile 1989–2017 (8); Colombia 1974–2018 (13); Ecuador 1984–2017 (13); Nicaragua 1984–2016 (7); Paraguay 1973–2018 (11); Uruguay 1984–2014 (7); and Venezuela 1973–2015 (10).

Survey data for South-American countries are provided by the LatinoBarometer (LB) database for the same countries as listed above; the only countries in the database with seven or more elections since their 1970–1980 emergence from authoritarian rule and for which LB surveys were conducted in any election years. Surveys for Brazil in 2002 and for Ecuador in 2009 are dropped because of anomalous numbers of respondents with ages that placed them as having been eligible to vote at 16 (86% and 1%, respectively). The remaining numbers of surveys conducted in election years were as follows for each country: Argentina 10; Bolivia 4; Brazil 3; Chile 5; Colombia 4; Ecuador 5; Nicaragua 5; Paraguay 4; Peru 5; Uruguay 2; and Venezuela 5.

Survey data for Austrian elections in 2008. 2013 and 2017 is taken from the Austrian National Election Study (AUTNES) website. For 2008 only a postelection survey is available (conducted in 2009). In 2013 and 2017 there were additional preelection surveys. Because of the small N available in only one country and our special interest in the elections of 2013 and 2017 (the only elections for which we have individual-level data relevant to the study of volatility) I pool the pre and postelection data where available.

The main text refers to error correction models (ECMs) which are presented here. Table 2.5 contains two models: Model M for the full dataset and Model N for the same data truncated to contain only the final 7 electoral contests held in each country. The outcome (dependent variable) is the difference in turnout (“∆Turnout”, using the Greek letter Delta for “difference”) found by comparing each election with the previous one in temporal sequence. The first input (independent variable), shown in Row 2, is a version of the same outcome lagged by one time-point. In the context of an ECM, its coefficient is known as the “error correction parameter”.

Table 2.5 Error correction effects on aggregate-level turnout in South-American countries and Austria of changes in vote-eligibility, controlling for other predictors of turnout

Because it is negative, the error correction parameter suggests what is sometimes known as “regression towards the mean”—the tendency of any deviation from long-run equilibrium to be “corrected” (or “decay”) over the passage of time. An error correction parameter of –0.85 in Model N suggests that 85% of any short-term effect decays within a single time period (approximately four years in our data, the average period between elections). The remaining inputs all come in pairs, each differenced coefficient being paired with a lagged coefficient for the same input. Differenced coefficients show the short-term effect of the variable concerned (the effect that will decay at the rate established by the error correction parameter) while the corresponding lagged coefficient shows the long-term effect—an effect that contributes to changes in the equilibrium level from which future short-term deviations will occur.

Model M shows no significant effect of eligibility to Vote at 16, neither short-term (Row 7) nor long-term (Row 8). These same failures were seen Model A of Table 2.1 in the main text, which used the same unbalanced data as are used in Model M. But Model N does show a long-term effect of 6.59 (Row 8), though this is only barely significant at conventional levels. That long-term effect is not significantly greater than the value found in Table 2.1, Model B, of 6.08% for the effect on turnout resulting from a switch from eligibility at 18 to eligibility at 16; but, more importantly, tells us that the effect in question is of a long-term nature. Evidently, the truncated dataset removes a composition artifact that masked this finding when time-series of different lengths were analyzed together.¹

Figure 2.4 replicates Fig. 2.2 in the main text but using only data for Austria. In that country we have only three elections with age-16-eligible voters, thus the maximum age that any of these voters have had the chance to reach in 2017 is 26. So the distinction between the traces for age-16-eligibles and older voters extends only over 9 years. But during this period there is no sign of the diminution in this distinction that we saw in the main text’s Fig. 2.2 for South-American countries. Indeed, the trace for respondents eligible to Vote at 16 appears to rise with increasing age. Confidence intervals are quite wide, however, and comparing the high margin of this interval for the youngest members of the cohort with the low margin for its oldest members we see that the upward slope is in fact not statistically significant and might even be downward. Statistically speaking, there is no difference between this graph and the one shown in Fig. 2.2 for South-American countries.

Fig. 2.4

Turnout distinctions for Austria between citizens eligible to Vote at 16 and others, with increasing age

In Table 2.6 we replicate the volatility analysis using error correction models. In the unbalanced dataset (Model O) we find an effect (5.84 in Row 3), rather larger than was found in Table 2.3 of the main text (2.07 at most). This is as it should be since Table 2.3 averaged over all future elections an effect of which 89% would have dissipated by the time of the next election following the one at which 16-year olds first voted (according to the error correction parameter in Row 2 of Table 2.6). (See the text associated with Table 2.5 for an explanation of the effects shown in an error correction model) This effect is far smaller and not remotely significant in Model P, which employs the balanced dataset, seemingly due to estimation problems.²

Table 2.6 Error correction effects on aggregate-level volatility in South-American countries and Austria of changes in vote-eligibility

Notes

• 1. Statisticians express concerns regarding possibly spurious effects in ECMs for variables that trend over time (Kennedy, 2008, pp. 307–313)—certainly the case for turnout. But turnout shows a lot of variability, and its overall trend is not statistically significant in our data except for Colombia (in the full dataset) and Paraguay (in the balanced dataset). No other country displays what is known in the econometric jargon as a “unit root” for turnout and our measure of Votes at 16 shows no such unit root for any country. Electoral clarity does show a unit root for five other countries (three in the balanced data), but the effects of the lower voting age in the full dataset rise to match those in the balanced data if the countries with unit roots on either variable are all omitted. For the balanced data, if Paraguay alone is omitted and electoral clarity is dropped from the model then the long-term effect of lowering the voting age rises to 7.1%, still significant at 0.10.

• 2. Again we need to be concerned with the possibility of unit roots in our data. In addition to those mentioned in Note 1, four countries show a unit root for volatility in both datasets: Chile, Nicaragua, Paraguay, and Venezuela. But if the countries with unit roots for any relevant variable are all omitted from the volatility analysis the results are actually strengthened, with short-term effects of age-16-eligibility rising to 8.1 in the unbalanced data and to 6.9 in the balanced data. Both coefficients prove significant at p < 0.05. Since none of the problematic countries are ones in which the voting age was lowered to 16, but were simply included as controls, one could take the finding of unit roots in four of those control countries as disqualifying them for this role due to non-comparability with countries in which the voting age was lowered. Dropping the data for these four countries would still leave four South American control countries to match the four South American test countries, reifying the findings described in this note.