Abstract
Gender gaps in voter turnout are usually studied using opinion surveys rather than official census data. This is because administrative censuses usually do not disaggregate turnout according to voters’ sex. Without this official information, much of the research on gender gaps in electoral turnout relies on survey respondents’ self-reported behavior, either before or after an election. The decision to use survey data implies facing several potential drawbacks. Among them are the turnout overstatement bias and the attrition or nonresponse bias, both affecting the estimation of factors explaining turnout and any related statistical analysis. Furthermore, these biases may be correlated with covariates such as gender: men, more than women, may systematically overstate their electoral participation. We analyze turnout gender gaps in Chile, comparing national surveys with official administrative data, which in Chile are publicly available. Crucially, the latter includes the official record of sex, age, and the electoral behavior—whether the individual voted or not—for about 14 million registered individuals. Based on a series of statistical models, we find that analysis based on survey data is likely to rule out gender gaps in electoral participation. Carrying out the same exercises, but with official data, leads to the opposite conclusion, namely, that there is a sizable gender gap favoring women.
Introduction
The “gender gap” in electoral turnout has been widely discussed by Verba et al. (1978), Lehman Schlozman et al. (1995), Verba et al. (1997), Norris (2002), Desposato and Norrander (2009), among others. In general, these authors find women’s electoral turnout to be lower than men’s, a situation that responds to their low levels of “political commitment,” as expressed in disinterest in participating in electoral campaigns or in discussions of a political nature, a notion endorsed by Van Deth (2000) and Burns et al. (2001).
However, according to Carreras (2018), the panorama has changed in recent decades. Several democracies exhibit either a negligible gender gap, or even higher levels of electoral turnout among women than men, although this coexists with lower levels of political commitment among women. Carreras (2018) suggests that this paradox—that women vote more than men despite having lower levels of political commitment—responds to the greater “civic duty” felt by women. Added to this is an increase in the female labor force in the past decades, which may also contribute to explaining their greater attendance at the polls (Cebula and Alexander, 2017).
Importantly, much of the discussion and evidence on turnout gender gaps in the literature is based on opinion survey results. 1 The absence in most countries of administrative censuses that separately record the electoral turnout of men and women, makes it difficult to validate these results or, at least, to contrast them with official data. Interestingly, an exception to this is work by Dahlgaard et al. (2019), who report that when based on survey data, women turn out as much as men; that is, there is no gender gap; whereas when based on official Danish records they conclude that “using self-reported turnout instead of validated turnout would lead to the incorrect conclusion that men are about as likely to vote as women.”
In this article, we carry out a similar exercise, comparing estimated gender gaps obtained from surveys with those obtained from the official administrative census. We highlight three findings. First, we find that in opinion surveys men overreport their electoral turnout more than women. This is implemented by computing turnout rates according to self-reports on both past and future voting behavior. These figures are then compared with the respective population turnout rates, as an approximation. This is because we cannot individualize voters in the official data in order to compute the actual voting behavior of those who participated in the survey. Second, we find that gender is not a statistically significant predictor when statistical models are run on survey data. This is at odds with administrative census records according to which a sizable gender gap favoring women is found. Third, we find that the gender gap favors women especially within the youngest age group (comprising those aged between 18 and 40 years), but favors men among the elderly (those aged above 72 years).
Data and Method
The Chilean Electoral Service (Servel) releases official information on electoral turnout ever since 2012, containing information about all eligible individuals registered in the official electoral record. Specifically, the data include information about every individual’s age, sex, party membership, 2 area of residence (municipality), and electoral turnout. Each individual is coded as having “voted” or not (“did not vote”). The database amounts to nearly 14 million entries for the most recent presidential election held in 2017.
These data are compared with the opinion survey results. From the existing literature, there is consensus that surveys contain biases, one of which is the interviewees’ tendency to overreport their electoral turnout when referring to either a past or upcoming election. 3 That is, individuals are prone to declare they have voted or intend to vote, when not. On average, overreporting electoral turnout amounts to approximately 10% (DeBell et al., 2018). Depending on the specific effect to be measured (education, income, sex, among others), inferences from such results may be biased.
In Table 1, we present a simple exercise to evaluate the extent of this bias. We calculate the percentage of individuals who declared they were registered in the electoral registers based on the pre-electoral survey carried out by the Center for Public Studies (CEP) between 1993 and 2009 (presidential and concurrent parliamentary elections). During this period, and until just before the 2012 municipal election, registration was voluntary while voting of those who were registered, was mandatory. We use the following question in the survey: “Are you registered to vote?.” For the 2013 and 2017 (both consisting of a presidential and a concurrent parliamentary election)—when registration was automatic and voting voluntary—we compute the percentage of respondents who declared they were sure to go out to vote in the upcoming election. The question in this case was the following: “And in your case, will you go to vote in the next presidential election?” (2013 and 2017, respectively). The possible replies were, “Yes, I will definitely go to vote”; “Probably yes”; “Probably not”; “No, definitely I will not go to vote.” We consider only the first alternative as a predisposition to vote. 4
Registration and Turnout in Surveys and Official Data.
Source: Own compilation with data from www.cepchile.cl, www.servel.cl, and www.ine.cl.
The first CEP survey following the 2013 election was taken in July 2014, while the first survey following the 2017 election was taken in October–November 2018.
In addition, we compute the actual figures from the official electoral records (1993–2009) and turnout (2013–2017), according to official Servel data. We compare these figures with those from survey-based computations. As shown in the table, there is indeed a significant overreporting behavior of electoral participation. On average, this overreporting is higher among men than women, at least between 1993 and 2009. For example, in 2009, 72.1% of men in the survey reported that they were registered, while the official registration figure for men was 64%. In contrast, 73.4% of women in the survey reported being registered, while the official figure was 68.1%. Thus, men overreported their expected turnout by 8.1 points and women by 5.3 points. Figures for 2013 change significantly. The overreporting rate in men was 4.5 points while among women it was −0.3, indicative of women’s underreporting. In the 2017 election, finally, results are even more surprising: men’s overreporting is 3.9 points, while among women is −7.1, which implies an increase in women’s underreporting compared with 2013.
We also implement a similar exercise comparing actual turnout with self-reported turnout in the first post-election CEP survey (i.e. retrospective voting self-report). In this case, the question in the survey is whether the individual had voted or not in the latest election or not. For 2013, overreporting was alike between men (14.8%) and women (14.2%), but in 2017 the difference there is a wider gap. Men’s overreport rate is 19.2%, while for women it is 11.6%.
One important qualification is in order. Although we have referred to and focused our discussion on overstatement bias, one cannot rule out the existence of a nonresponse or attrition bias, which may be taking place in conjunction with the overreporting bias. 5 Generally, those who vote are more likely to participate and are prone to provide answers in surveys. The figures in the table, thus, might be also explained by the self-selection of those who vote, inflating thereby turnout rates. Unfortunately, one cannot disentangle both effects with the information at hand. Without validation carried out with precise identification of those individuals who participated in the respective survey, one cannot compute the actual figures. We are aware of this limitation and our conclusions are cautious in relation to this point. Notwithstanding, as an extenuating consideration to the qualification above and our results, in a recent related work in which the data allow disentangling both biases, the authors conclude that sex is not correlated to the probability of taking the survey (see Dahlgaard et al., 2019).
Results
Polling station tallies are an alternative to opinion surveys when studying gender gaps. Awkwardly, between 1988 and 2009, in Chile there were polling stations for men and polling stations for women. This allowed direct and simple computation of voting gender gaps among registered individuals. However, these figures were uninteresting, because almost every eligible citizen had registered for the 1989 foundational election that was to oust General Augusto Pinochet from power. Combined with the obligatory voting rule for those registered, in subsequent elections the close to universal registered population slightly changed throughout due to a meager generational replacement effect. In 2012, an electoral reform abolished gender-segregated polling stations, making this type of analysis no longer possible. Yet, almost at the same time, the Servel released a database disaggregated down to the individual level, containing information about each registered individual’s sex, age, party membership, and, not least, their attendance to vote.
In Tables 2 and 3, we present the results of the estimation of turnout models using the ordinary least squares (Table 2) and logistic (Table 3) estimators. We compare the results for different specifications. In the models that use official data (Servel, “AC” in the table), the dependent variable is the actual turnout rate (percentage) for the last presidential and parliamentary election held in 2017. 6 In the case of the CEP survey (“S” in the table), we use the question from the survey conducted between October and November of 2018: “Did you vote in the last presidential and parliamentary election in November 2017?” 7 In both cases, the dependent variable is a dummy, with a value of 1 if the individual participated or reports having participated, respectively; and of “0” in the opposite case. 8 In all the specifications we use the dichotomous variable “FEMALE,” with takes the value 1 if the respondent is “female,” and “0” otherwise. As controls, we use the individual’s age, AGE, and its square, AGE2. These variables are interacted with the variable of interest “FEMALE” (AGE × FEMALE and AGE2 × FEMALE).
Turnout Gender Gaps in CEP Survey versus Official Census Data—Estimation Using the Ordinary Least Squares (OLS) a Estimator.
Source: Own compilation with data from www.servel.cl and www.cepchile.cl.
S: survey; C: administrative census.
These results are robust—in the survey—when observations are weighted and errors are clustered at the municipal level. These results are available upon request.
Corresponds to a significance of 10%; **corresponds to a significance of 5%; and ***corresponds to a significance of 1%.
Turnout Gender Gaps: CEP survey versus Official Census Data—Estimation Using the Logistic Estimator (LOGIT).
Source: Own compilation with data from www.servel.cl and www.cepchile.cl.
S: survey; C: administrative census.
Corresponds to a significance of 10%; **corresponds to a significance of 5%; and ***corresponds to a significance of 1%.
The results are consistent for both estimators. In the estimation using data from the CEP survey (“S” in the table), the coefficient for “FEMALE” is not statistically significant and even has a negative sign in one of the three specifications, which is contrary to what our analysis with official data conveys. In the third specification (column 5 in the tables) the linear predicted turnout difference (FEMALE vs male) is negative for most of the age groups in the sample (see below). These results contrast with regressions on data from the official Servel records. In the first two specifications (columns 2 and 4, respectively), FEMALE is positive and statistically significant. As for the third specification (column 6 in each table, respectively), the linear predicted gender gap is positive for all age groups except among the eldest age groups. Interestingly, as opposed to the gender effect, the estimation of age effects and the interaction terms have the same sign, are closer to the respective census estimation, and are statistically significant (see columns 3 and 5), with the sole exception of AGE × FEMALE in column 5. To show these results more comprehensibly, from the OLS specifications estimated in columns (5) and (6), respectively, we obtain a graph comparing the gender effect electoral turnout for every age group. In the vertical axis in Graph 1 the turnout gender gap estimates are displayed. In the horizontal axis is age. Positive values indicate a larger female turnout and negative values indicate that female turnout is less than male. We find from official data that the overall effect of being a “woman” is positive for women aged between 18 and 71 years; for women aged 72 years or more, participation is lower than male. Naturally, confidence intervals from population samples are almost imperceptible, and thus we exclude them from the graph.
The Marginal “Female” Effect on Electoral Turnout (Census and Survey Data).
These results contrast with the same analysis carried out on survey data. The difference between men and women, is statistically not different from zero for nearly all age groups. For ages from 34 to 64 years, however, the predicted values imply that being a woman does have a positive effect on turnout, although this effect is considerably less compared with the results obtained from the census data. These results are replicated for estimation using the logistic estimator, which are shown in Table 3.
Similarly, we compare the linear prediction of turnout according to gender and age, obtained from the estimation of the LOGIT models in column 5 and 6 of Table 3. The results are displayed in Graphs 2 and 3, respectively. Graph 2 shows the evolution of turnout rates, for women and men, across age groups spanning from 18 to 88 years, using census data. The figure depicts a gap favoring women, especially among those aged 18–63 years. In contrast, there is a similar trend in the estimated probability of voting using survey data, but only for middle-aged groups. Among the youngest and eldest age groups, however, the estimated male electoral participation slightly exceeds that of women. Consequently, and as we have argued here, the results obtained with census data differ significantly from those obtained with survey data.
Predicted Turnout Probability for Gender and Age Groups: Administrative Census Data.
Predicted Turnout Probability for Gender and Age Groups: CEP Survey Data.
Although these results indicate a large difference between the survey and census data, leading to very different conclusions, we also note that the estimated confidence intervals for the predicted values in the survey are quite wide, especially among the age groups in which the gender gap is most noticeable. The survey is “noisy.” In line with this, even in the absence of overreporting and nonresponse effects, one may still discard a gender gap out of the plausible low power of the hypothesis tests, as reflected by the estimated confidence intervals. 9
One avenue for further research is to study the estimation of weighted regressions, where weights are obtained from validated data. As suggested by Dahlgaard et al. (2019), this procedure should improve estimation of those covariates correlated with the overreporting and nonresponse bias in surveys. The strategy consists of giving more weight to those observations for which actual turnout is higher in the population sample in those age and gender groups, for instance, to which the observations belong in the survey sample. The logic is straightforward: the larger the actual turnout, the less scope there is for either nonresponse or overreporting bias. 10
Conclusion
We have studied turnout gender gaps in a developing country such as Chile. In doing so, we compare the results conveyed by opinion survey data, with those obtained from census data. This is a methodological approach which can be implemented in very few countries. In many democracies official electoral data are not publicly available. Furthermore, when so, it does not distinguish female voting behavior from male. In this sense, Chile is an exceptional case, and the evidence presented here is novel.
The results draw attention to the possible risks borne by statistical analyses based on opinion surveys. While these appear to rule out turnout gender gaps, evidence from the administrative census moves in the opposite direction. Census data reveals the existence of a sizable gender gap in electoral turnout that favors women, though moderated by age. This concurs with evidence from several developed countries where women turn out more than men. However, as the evidence we have presented here suggests, if a researcher were to use data sources similar to those used in most studies analyzing turnout gender gaps—that is, opinion surveys—she or he would arrive at quite a different result. Our findings are coherent with the results of Dahlgaard et al. (2019). A novelty in our work is that it brings new evidence from a developing country, whereas Dahlgaard et al. (2019) do so with data from an advanced industrialized democracy. Our results, as theirs, are robust.
This research note, therefore, aims to sound a warning bell for voter turnout studies based on opinion survey data. One matter of concern is the presence of overreporting voting behavior and nonresponse bias, both of which may be correlated with gender, as our findings suggest. Another is its impact on the design of public policies aimed at increasing turnout. For instance, public initiatives might lead to the implementation of campaigns with the intention of stimulating electoral turnout, while ignoring gender gaps precisely because that is what opinion surveys would tend to recommend. Instead, in the Chilean case, when using data from the administrative census, public policies in favor of participation will be better targeted, becoming instruments that, at the end of the day, will strengthen democracy.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research is sponsored by The National Fund for Scientific and Technological Development, FONDECYT, Project No. 11200297.
