Correlates of Voter Turnout

The first national-level turnout studies explain variations in voter turnout by focusing on a selected sample of lower house elections in OECD countries. For example, Jackman (1987) analyzes 19 democracies, Jackman and Miller (1995) analyze 23, and Blais and Carty (1990) and Powell (1986) include 20. Normally, these studies focus on a series of factors such as “socio-economic environment, the constitutional setting, and the party system,” (Blais & Dobrzynska, 1998, p. 241).

Recent years have seen a proliferation of potential theoretical factors shaping turnout as well as an expansion of coverage. For instance, voter turnout is depressed with the size of a political community (Remmer, 2010), economic globalization (Steiner, 2010), corruption (Stockemer, LaMontagne, & Scruggs, 2013), and ethnic diversity (Martínez i Coma & Nai, 2017). Terrorism, in turn, increases the electorates’ attention on national politics and, consequently, turnout increases (Robbins, Hunter, & Murray, 2013). Similarly, while previous studies focused mostly on economically developed democracies, Blais and Dobrzynska (1998) were the first to have a truly global collection of democratic elections. More recent analyses focus on other regions including Latin America (Fornos, Timothy Power, & Garand, 2004), post-Cold War Eastern Europe (Kostadinova, 2003), Africa (Kuenzi & Lambright, 2007), and Muslim-majority countries (Stockemer & Khazaei, 2014).

These works share a number of similarities including the use of three general types of independent variables: socioeconomic, institutional, and political (Geys, 2006; Blais & Dobrzyinska, 1998). Socioeconomic variables include factors like economic growth and urbanization. Institutional variables capture the institutional arrangements under which the election take place (e.g., electoral formula). Finally, political variables account for specific electoral outcomes (e.g., the margin of victory).³

The literature’s limitations appear when comparing their results. For example, one of the most cited studies by Blais and Dobrzyinska (1998) find that turnout is significantly affected by economic development, literacy rates, population size and density, compulsory voting laws, minimum voting ages, the electoral system, the number of political parties seeking seats, and the election’s competitiveness. Endersby and Krieckhaus (2008) reach similar empirical conclusions but suggest that context is important. Along the same lines, Martínez i Coma (2016) confirms some of Blais and Dobrzyinska’s (1998) findings but not others—notably those relating to electoral systems and economic development.

Furthermore, and more important for the purpose of this paper, a consensus does not yet exist on the robustness of these variables and, consequently, on what variables should be considered for a core model of cross-national aggregate voter turnout. For example, of the eight socio-economic factors used in the three articles mentioned above, only one variable is considered in all three—population.⁴ Put simply, comparative studies provide mixed evidence for the robustness of particular factors affecting voter turnout. In summary, after over 50 years of research the literature still has not coalesced around a core model of turnout; different sets of variables are used in different analysis.⁵ The literature, “draw[s] on relatively small samples, differing sets of observations, divergent operationalizations of turnout, and a wide array of theoretical models, the literature has mainly converged around relatively commonsensical observations” (Remmer, 2010, p. 277).

One popular means of evaluating a literature’s lessons is a meta-analysis, which basically assesses whether (or not) an independent variable affects a dependent variable.⁶ A meta-analysis, or an “analysis of analyses” (Glass, 1976, p. 3), normally employs two procedures: “vote-counting” and “combined tests.” The former counts the number of times a given coefficient is significant and in the expected direction. In such cases, it is counted as a ‘success’; otherwise, it is considered an anomaly. The higher the success rate, the more likely it is that we are to be observing a real association between the variables. A combined test is “based on the summation of the actual test statistics provided in each study” (Geys, 2006, p. 640).⁷

Such meta-analyses have been done in the voter turnout literature. For example, Geys (2006, p. 641) identifies twelve significant socio-economic, political, and institutional factors affecting turnout, while highlighting that none are “omnipresent in the literature.”⁸ This is “partly due to the lack of a firm theoretical model at the basis of variable selection,” and he argues for the “construction of a ‘core’ model of turnout” (Geys, 2006, pp. 641, 653). Along similar lines, Stockemer’s (2017, p. 712) meta-analysis of 135 studies from 2004 to 2013 identifies over 100 potentially salient variables, thoroughly analyzes 10 of them, and concludes “no variable is omnipresent or appears in most studies. Rather, different variables are used in various contexts.” The divergence pointed out by Stockemer (2017) explains the different results from his and Geys’ (2006) meta-analysis. They assess the “success” or robustness of 18 variables, 5 of those common in both works. Even for those five common variables,⁹ when comparing the studies’ success rates, there are three common results: compulsory voting and population size impact turnout while income inequality does not. They diverge on the impact of election closeness and PR system because Geys (2006) finds they affect turnout, while Stockemer (2017) does not. Unsurprisingly then, Stockemer (2017, p. 712) acknowledges that the “literature is far from establishing a core turnout model.”

Meta-analysis’s main limitation is not procedural but conceptual. In the end, even the most exhaustive meta-analyses like those above are circumscribed by their samples. This implies that we are unsure about the robustness of such results, given that most researchers’ robustness checks are ad hoc. “They identify a set of competing explanations and see if their empirical results hold once they control for some variables that might be consistent with those explanations” (Hegre & Sambanis, 2006, p. 509). In other words, given all possible variable combinations, we cannot be sure that the selected model and the results presented are the ‘right’ ones. Usually, sensitivity (or robustness) checks estimate a series of regressions with alternative specifications. However, the number of possible alternative specifications is, of course, substantially higher. For example, without assuming any initial knowledge of variable selection, 5 variables lead to 32 possible model specifications, 6 variables have 64 combinations, 7 variables have 5040, 8 have 40,320, etcetera. The number of permutations increases exponentially as the number of variables increases; therefore, any particular study is likely showing a tiny proportion of the possible combinations. Hence, even the most demanding and detailed meta-analysis will only cover a tiny percentage of the multiple possible combinations that may affect voter turnout.

The standard regression framework on which meta-analyses rely has two other limitations. First, a particular variable’s statistical significance may be sensitive to the inclusion/exclusion of other variables. As Leamer (1983, p. 38) concludes, “an inference is not believable if it is fragile, if it can be reversed by minor changes in assumptions.” Second, even when theories point to particular mechanisms, they are not “refined enough to inform the choice of the empirical measure to be used to proxy for such factors/mechanisms” (Carmignani et al., 2014, p. 516). For example, should we use GDP per capita (as four studies considered in this paper do) as a proxy, the log of GDP per capita (two studies), or the log of GDP at purchaser’s price parity (two studies)? Should we use one measure or two? How would results change, for example, if instead of using Laakso and Taagepera’s (1979) effective number of parties, one decides to use the disaggregated number of parties or the number of parties weighted by their vote share?

An extreme bounds analysis estimates a set of regressions with the following functional form:where Y is the dependent variable (in our case, voter turnout), I is a vector of core variables included in all models, M is the variable of interest, Z is a set of controls, and u is the error term (Levine & Renelt, 1992). I is the set of variables always included—the “base” or “core” variables—because the literature suggests a well-established relationship with the dependent variable. What is the I vector of variables for electoral turnout? Unfortunately, as we show below, less than those that one, a priori, may think.

EBA repeatedly estimates the equation with a different set of Z controls in each regression. Since every regression produces a coefficient for all included variables, all the regressions create a distribution of such coefficients. In order to decide whether the coefficients are robust, researchers have relied on two main criteria. The first by Leamer (1983) suggests that a variable should be considered robustly related to the outcome variable if, and only if, the lower and upper extremes of a variable’s coefficients have the same sign.¹⁰ Specifically, the extreme upper (lower) bound is defined by the maximum (minimum) value of the variable of interest plus (minus) two standard deviations. If the variable of interest remains of the same sign at both upper and lower bounds, then such relation among the variables is said to be “robust.” When the variable of interest does not keep the same sign at both upper and lower bounds, then such relation among the variables is said to be “fragile.” In short, only after running all possible regressions including all variables and only if all estimates are in the same direction, are results considered robust. Sala-i-Martin (1997) finds Leamer’s standard to be overly restrictive in most cases because it is likely that if enough model specifications are analyzed, and assuming that the distribution of β has both some positive and negative support, it is likely that the signs of the coefficients will change at least once. In fact, following Leamer’s criteria if a single regression produces a coefficient of the opposite sign large enough to shift one of the bounds, then the variable is considered not robust. Sala-i-Martin (1997) proposes to look at the entire distribution of coefficients and conclude a variable is robustly related to the outcome variable when a large percentage—say 90–95%—of the coefficient’s distribution is either above or below zero.

These criteria, then, can lead to different substantive conclusions. In the extreme, if enough regressions are run and the distribution of the estimators have some “positive and some negative support, then one is bound to find one regression for which the estimated coefficient changes signs” (Sala-i-Martin, 1997, p. 179). Indeed, by following Leamer’s approach, we may conclude that the knowledge about many social phenomena is scarce and thereby make a consequential Type-II error. In contrast, as Plümper and Traunmüller (2020, p. 149) recently show, Leamer’s EBA has “an extremely low probability of producing false positives” while, Sala-i-Martin is “more likely to suffer from identifying false positives than the inferential rule it replaced” (Plümper & Traunmüller, 2020, p. 149). Previous works have, in the main, relied on Sala-i-Martin’s rather than Leamer’s approach.¹¹ However, both criteria are informative, so we follow Hafner-Burton (2005) and report results according to both criteria. As is clear below, there are substantive differences in what is considered robust predictors of turnout using these different criteria, and we compare our results using both criteria.

In order to identify the most common predictors of voter turnout, we rely on Geys (2006), Geys and Cancela (2016), and Stockemer (2017) meta-analyses of 83, 185, and 130 works, respectively. For our analysis, we included all English language, national-level, comparative, peer-reviewed journal articles focused on voter turnout published between 1980 and 2017. We therefore exclude case studies, studies focusing on local, regional, or provincial elections, book-length studies, and studies not in English. We focus on the national level because local dynamics are likely distinct from those at the national level. Furthermore, logistically it also makes sense to exclude works where the underlying data are not comparable to other cases: for example, exploring the effects of Norwegian school referendums (Kaniovski & Miller, 2006) on voter turnout in non-Norwegian countries is not possible. This decision implies that some factors, like campaign expenditures, cannot be examined given the almost total lack of available data outside the US. The 44 included studies are listed in the Appendix.¹⁵

As mentioned above, turnout predictors are usually organized into three groups: socio-economic, institutional, and political; and we follow this approach when organizing 127 independent variables derived from the 42 articles we examine.¹⁶ We find 41 socio-economic factors, 48 institutional factors, and 38 political factors. Such a large number of independent variables reinforce the diversity of empirical approaches in the literature and the need to clearly determine what robustly affects turnout (and what does not).

Tables 1, 2 and 3 summarizes each group of variables. The first column presents the number of times a variable is used in the literature; the second column includes the variable name or concept; the third column show the ways in which the variable has been measured (if available); the fourth column present the directional effect in turnout—sub-divided in four sub-columns, one accounting for each possible result. When the variable had a positive impact for turnout, it is labelled as ‘positive’, ‘negative’ when the contrary; ‘NS’ suggests a non-significant result, while ‘mixed’ captures those results when the results vary depending on the model.

Table 1 includes 48 institutional factors. The most frequent factor is compulsory voting, which is measured in three different ways. Twenty-eight studies find that compulsory voting has a positive and significant effect while five find it not significant. Two other variables—the number of political parties and proportional representation (PR)—are the next frequent (19 times each). This illustrates the literature’s differences in measurement; the former is operationalized in 10 different ways, the latter in two. We have created a straightforward measure of agreement among studies by dividing the most frequent result by the number of studies that use such measure and multiplying it by 100. The higher the percentage, the more established the finding.¹⁷ For compulsory voting, the degree of agreement is over 84%. Other variables are less established. For example, the agreement on the impact of the number of parties or the impact of proportional representation electoral systems is unclear, with about 53% agreement. Regarding the former, while 10 cases find that higher number of parties, lead to lower turnout, 8 do not find it significant. Likewise, 10 studies show that countries under proportional representation systems show higher turnout while 7 find it not significant. Table 2’s socio-economic variables include two sub-groups, socioeconomic characteristics and geographical dummies. First, the indicators gathering relevant socio-economic characteristics of a given society, such as size of population, GDP, GDP growth, and urbanization. Twenty-two studies include GDP as an independent variable. Eight find that GDP has a positive impact on turnout, while seven show a non-significant relationship, three report a negative influence, and four show mixed results. Results are not much better for the rest of such variables. Second, there are geographical dummies for specific countries or regions. Most notable about these variables is that including a variable for Switzerland and/or for the US almost always are negatively related to turnout. There are two patterns worth mentioning when discussing Table 3’s political variables. First, not many political variables appear in the articles we examined. This is surprising given the fundamentally political nature of turning out to vote. An exception is “closeness/competitiveness” that appears in 21 studies (almost half of our sample). Furthermore, the level of agreement for this variable is below 50%. Second, we only find a high level of agreement for the previous election turnout level (though such variable only appears in five studies).

In sum, three important findings arise from this initial literature review. First, out of 127 distinct variables, less than half (44%) appear more than once. Even the most frequently used indicator, compulsory voting, was included in less than 75% of the examined studies. Second, among the 55 variables that appear in more than 1 article, over half (57%) are measured in more than one way. Third, it seems more generally that turnout studies face a paradox—while voting is mainly a political act, the most common empirically tested arguments in the literature are of institutional or socioeconomic mechanisms. Only recently have some articles examined the impact of terrorist attacks (Robbins et al., 2013), corruption (Stockemer et al., 2013) or electoral dynamics (Martínez i Coma & Trinh, 2017) on turnout.

Table 4 condenses this information and also offers an overview of the distribution of the 70 variables for which we have data. These results strongly suggest that a standard model of turnout does not yet exist, and few factors—especially institutional and socioeconomic—have a consistently established effect on voter turnout.

Including all 127 predictors discussed above in our empirical models is not possible due to data availability—several variables are available for only a short span of time, specific region (i.e., Europe) or a clear set of countries (i.e., OECD members).¹⁸ The final line in Table 4 describes the distribution of the seventy variables we include, and Appendix Tables A2, A3 and A4 present these variables’ summary statistics, operationalization, and sources.

Given the absence of a commonly accepted model of voter turnout, in our selection of core variables we rely on several theoretical assumptions consistent with the literature and empirical regularities to include five variables in all of our models. From an institutional perspective, compulsory voting has been found to affect turnout. 84% of our 44 studies find that compulsory voting increases turnout, which is likely the literature’s most commonly accepted finding. The electoral system is also a recurrent variable of study under the (challenged) assumption that voter turnout is usually higher in proportional representation systems (Blais & Carty, 1990). We also include two socio-economic factors—per capita gross national income and population (both logged to control for outliers). The former accounts for the literature’s finding that economic development fosters turnout while the latter controls for the relationship between community size (an element assessed both by Geys, 2006 and Stockemer, 2017) and turnout. Finally, we include the level of turnout in the previous election because the literature suggests that turnout may be habit forming and has its own inertia (Geys, 2006).¹⁹ While we acknowledge that more core variables could be included, we believe our core turnout predictors are consistent with the literature we survey.

Given that several variables measure similar concepts and/or that the measurement of such variables may be related, multicollinearity may be a risk when including such a large number of predictors. Levine and Renelt (1992, p. 945)²⁰ rely on three common strategies to reduce this risk: (1) they limit the total number of explanatory variables to eight at most; (2) they limit the number of Z controls to seven; (3) and they further restrict their Z variables by excluding variables that a priori may measure the same phenomenon. Similarly, Hartwig and Sturm (2014) drop some variables when they are highly correlated. To avoid artificially inflating our estimates, we have followed their first and third strategies (we limit the total number of core and independent variables to eight and we exclude multiple measures of the same underlying phenomena).^21,22

The final estimation consideration is what type of model to run. The most commonly used method in the 44 reviewed studies was ordinary least squares (OLS) with robust standard errors clustered by country (15 articles). Regression models with fixed effects and random effects were used in several (6 articles and 7 articles respectively) of the 44 articles examined in the literature review above.²³ Nevertheless, there are reasons to expect with eight predictors that there is the real risk of unobserved variable bias and risk of unmeasured unit-effects. Furthermore, like most of the literature we are interested in the reasons for differences in turnout both between and within countries. There is an ongoing methodological debate about which is appropriate for time-series cross-section data such as ours that is outside the scope of this article.²⁴ Therefore, we estimate two series of models. The first series of OLS models with fixed effects focuses on time-varying explanations for turnout. Our I vector includes 5 variables, and the Z vector has 53 variables. The second series of models use random effects and clustered standard errors by country to include time-invariant variables. A random effects approach relies on the strong assumption that the included predictors are uncorrelated with country-specific intercepts (i.e., unobserved factors), but it does allow for time-invariant factors’ effects on turnout to be estimated. Our I vector for the random effects models includes 5 variables, and the Z vector has 65 variables. In both series of models, continuous independent variables are standardized in order to aid the comparability of coefficients and a number of continuous predictors were log-transformed to reduce the effects of extreme values.²⁵