How Many People Vote Together?
Before testing our hypotheses, we examine the extent to which individuals vote together with their household members in Denmark (see Table 1) and in comparison with other countries where survey data are available (Table 2). The data in Table 1 uses the Danish register data which contain the timing of the vote from time stamps on the voter list. The individuals in the samples are divided depending on whether they did not vote, voted by post (in Denmark postal voting is a form of early voting usually cast at the city hall or citizen service centers), voted at the polling station alone or voted at the polling station with another household member.
Between 29 and 35% of all eligible individuals voted at the polling station on Election Day with someone from their household in each of the three elections. If we look only at voters, 41–51% voted within a minute of someone else in their household. This is, to our knowledge, the first evidence on the level of voting together from large-scale administrative data and it shows that voting together is a very common phenomenon. The 41–51% share of voters voting with others is remarkable when taking into account that some voters live alone and therefore by definition cannot vote with others in their household, while others vote by post and therefore cannot vote together on Election Day. If we restrict the sample to only individuals living in 2 + sized households, the share of voters voting with others increase to 56% (2015 general election), 60% (2013 municipal elections) and 69% (2014 European Parliament election).
Looking across the three types of elections suggests that, despite large variations in salience between the three elections, differences are relatively modest. However, one interesting pattern appears: while the absolute percentage of individuals voting together is higher when turnout is high (29, 33 and 35% across the three elections ordered by overall turnout), the relative proportion voting together is higher in low salience elections (51, 46 and 41% respectively). This indicates that individuals voting together may be more resilient to factors that reduce turnout, perhaps because voting together decreases the costs and increase the peripheral benefits of voting. Together, these findings provide prima facie evidence of the companion effect.
It is relevant to ask if the findings from Denmark also hold elsewhere. In order to do so we have fielded identical survey items in the Danish Election Study and the British Election Study across different types of elections in the two countries, aiming to estimate the level of voting together in Denmark and the UK.Footnote 6 We also utilize previously reported findings from Italy, Canada, Scotland and Wales (Fieldhouse and Cutts 2012).Footnote 7 Table 2 replicates the findings for voting together from Table 1 based on election studies from the six different locations. As the percent voting together of all electors is likely to be inflated by the fact that there is substantially under-reporting of non-voting in the surveys, we focus on the distribution of voting modes among voters.
Overall across all these countries between roughly one-third and two-thirds of those reporting having voted indicated that they voted together with someone at the polling station. There is some variation across countries with Denmark having twice the proportion of voters voting together compared to the UK (slightly less than 60% compared to 31%). A large part of this difference seems to be driven by the use of postal voting in the UK (postal voting is rare—approximately 5%—in Denmark and was therefore not given as a response option in the surveys). About 43% of British voters voting at a polling station voted with another person compared to approximately 58% in the Danish surveys. The difference between Britain and Denmark highlights a potentially negative aspect of postal voting. Postal voting may diminish some of the social aspects of voting by making voting a more individualized activity which could be more vulnerable to decline (Burden et al. 2014). Figures for Scotland, Canada and Wales resemble the UK while the numbers for Italy are close to the Danish ones. The findings show that across different countries and across different types of elections, voting together is a widespread phenomenon and is not specific to a single country or type of election.
The percentage of polling station voters who vote with others in Denmark is about 58% in the survey data compared to 45–49% in the register data from the same elections. The difference is most likely to be due to the fact that the register data only identifies individuals voting together who share a household. Slightly more than 10% of those voting together in the survey data report voting only with someone outside the household. Furthermore, we cannot dismiss the possibility of over-reporting of voting together in the survey data. However, overall the survey data and register data is quite consistent.
Who Votes with Others?
Having established that voting together is a frequently occurring and general phenomenon we dig deeper into what types of individuals are most likely to vote with others. We expect that voting together is largely driven by opportunity, i.e. household size (no. of eligible individuals in the household) and marriage. To confirm this, for each election we calculate the share of each mode of voting (not voting, postal voting, voting alone and voting together), by the variables of interest. Figure 1 shows the rates of voting together by household size for each of the three Danish elections.
For all elections household size is strongly related to voting together. This is not surprising insofar as voting together (by definition) can only occur when the household size is greater than one. Looking at multi-person households, the descriptive relationship between size and voting together is modest. In absolute terms the share voting together declines slightly with household size in all elections—for instance, in the 2013 municipal elections 46% voted with others in two elector households while the corresponding number was 32% in large households (more than four electors). However, this is mainly because there are more non-voters in large households. When disregarding non-voters, the relative share of lone voters and individuals voting together is virtually constant across household size—for instance, in the 2013 municipal elections, the ratio of individuals voting together and voting alone is approximately 1.7:1 for household sizes of both two and greater-than-four. A possible explanation of the similar patterns in household sizes greater than two is that the opportunities of voting together in larger households may be offset by weaker ties among household members. Figure 2 shows an equivalent chart for marital status.
The charts confirm that married couples vote together more frequently than the rest of the population. The percentage of non-married individuals voting together is 15–22% across the elections while the corresponding numbers for married individuals is 45–51%.
To provide more insight into the differences between groups when controlling for a range of demographic predictors of turnout, we estimate multinomial logistic regression models for each election. The dependent variable is voting mode (voting alone, not voting, postal voting and voting together). The reference category is voting alone. We include a range of usual suspects as controls: age, age squared, age-cubed,Footnote 8 gender, educational level (5 categories), income, children in the household, residential stability and ethnicity (3 categories). We restrict the models to include household sizes greater than one since single-individual households by definition cannot vote together. In Fig. 3 we show the predicted probabilities (averaged over observed values) for voting together compared to the corresponding probabilities for voting alone. Note that confidence intervals are plotted but are not visible due to the large sample size. Numerical results can be found in Appendix Tables 3, 4, 5 of the appendix where we also show results split on household sizes.
The results are consistent across elections. Even when taking household size into account a married person has a higher likelihood of voting together compared to voting alone (see also the positive coefficients for married individuals in the multinomial logit models in Appendix Tables 3, 4, 5). Unsurprisingly, the differences between married and unmarried individuals are smaller than when household size and other factors are not taken into account, but the share voting together is still around 10% points higher for married couples.
The results for other variables are also interesting. In the European elections, highly educated voters were more likely to vote together in absolute terms, but relative to voting alone, the proportion is lower than other groups across all elections (see also the negative coefficients for the high education groups in Appendix Tables 3, 4, 5). This may be because those with lower levels of education are more likely to drop out of voting when they have nobody to vote with. In other words having a voting companion may be especially important when an individual does not otherwise have the resources to vote. The corollary of this is that highly educated citizens are relatively less likely to vote together (compared to alone) possibly because their resources or norms of voting make them less reliant on the social benefits of voting, and are more likely to vote even if that means voting alone. Another potential explanation is that, insofar as the highly educated on average work longer work hours (Deding and Filges 2009), they might find it more difficult to coordinate going to polling station with a family member. Non-Western immigrants are less likely to vote together than ethnic Danes, while older people are more likely than the young. In further analyses (Appendix Tables 3, 4, 5) we have tested the robustness of the results to splitting the models on household sizes instead of controlling for household size. The results are generally consistent across models. Appendix Table 6 replicates the findings with survey data from the UK and Denmark (see the appendix). Again, the replication with survey data provides similar results as the objective data and across the two different contexts UK and Denmark.
The Relationship Between Opportunity of Voting Together and the Likelihood of Voting
We have now documented that individuals indeed vote together and that the tendency of voting together is, at least, partly is driven by opportunity and social intimacy. This is interesting insofar as it informs us about how people vote. Whilst the descriptive patterns are indicative of a connection between the opportunity for voting together and actual turnout, the causal effect—whether voting together affects turnout—still remains unproven. In other words, does the availability of a voting partner cause an increased probability of voting? Recent studies have found that households are perhaps the most important unit for inter-personal mobilization (Nickerson 2008; Sinclair et al. 2012; Bhatti et al. 2017). This could, at least partly be due to the possibility of accompanying each other to the polls, as suggested by the companion effect.
We noted above that it is difficult to demonstrate whether voting together bears any causal significance on turnout as voting together can (by definition) only occur among voters. In other words how do we know if a non-voter would have voted had they had the option to vote in company? We can go some way towards measuring the opportunity to vote together with network survey data by examining the impact of inter-personal mobilization—that is whether a respondent is asked by a discussant to vote (Rosenstone and Hansen 1993). In the 2014 European Parliament wave of the British Election Study Internet Panel (Fieldhouse et al. 2015) this was asked in a discussant ego-network module alongside whether each discussant accompanied the respondent to vote. These data show a high degree of correspondence between being asked to vote by a discussion partner and voting together: 74% of discussants who asked a respondent to vote actually accompanied the discussant to the polling station. By contrast less than 1% of those voting together did so without having been asked. Nevertheless, another 17% of those asked also voted in company, but not with the discussant who invited them. This demonstrates an imperfect correspondence between conventional measures of inter-personal mobilization (being asked to vote) and voting together. Moreover, this still does not tell us whether each respondent would have voted had they never been asked or had the opportunity to vote together never arisen. This absence of a reliable counterfactual (only having data on voting together for voters) makes it difficult to assess the causal importance of the companion effect in cross-sectional data. To get a better understanding of this we examine whether individuals who gain the opportunity to vote with a companion have a higher propensity to vote than individuals who lose the opportunity. More specifically we conduct two analyses. First, we test whether acquiring a potential voting companion, from one election to the next, leads to increased turnout probability (H1). Second, we test whether losing a voting companion leads to a greater fall in turnout probability than losing a non-companion (H2).
To examine the consequences of acquiring a potential voting companion, we use data from individuals included in our register data about whom we also have information on turnout (though not timing) from the 2009 Danish municipal elections. This means that we can create an individual level two-wave panel of the same type of elections, with a sufficiently long time-lag for a substantial number of voters to have changed their living circumstances. Specifically, we examine whether individuals who previously lived alone in 2009 but lived at least one other elector in 2013 saw an increase their probability of voting. We also test the reverse of this—whether losing a potential voting partner (from a multi-elector household to a single elector household)Footnote 9 leads to a decrease in probability of voting. The sample for this analysis is all individuals in our data who were eligible in the 2009 and 2013 municipal elections (for 2009 we have access to data from 44 of the 98 Danish municipalities). A challenge for this analysis is the possibility that unobserved characteristics of citizens are correlated both with changes in household composition and changes in turnout behavior. By stratifying our analysis by previous turnout behavior we adopt the equivalent of a change score model which provides some protection against the effects of unobserved time-constant variables (Allison 1990; Berrington et al. 2006). We cannot eliminate this potential threat entirely, but we mitigate it as much as possible by matching on a range of pre-treatment characteristics using coarsened exact matching or CEM (Blackwell et al. 2009; Iacus et al. 2012). Subsequently, we conduct standard regression on the matched sample with appropriate weights to take into account differences in the relative number of treatment and control observations between strata.
In our models we split the sample depending on whether individuals voted or abstained at the outset in 2009 to allow for asymmetrical effects on previous voters and non-voters. This allows us to take into account that change over time could be dependent on the initial level of turnout. As the sample is stratified by turnout at the outset, the dependent variable is simply turnout in 2013 (0—abstained or 1—voted). The key independent variable is the change in the household type of the individual. We consider four different treatment statuses, one for losing a potential voting partner, two for no change (either no household partner in both periods or a partner in both periods), and another for gaining a partner. In our analyses we compare having no partner in both periods with gaining a partner, and having a partner in both periods with losing a partner. We match exactly on pre-treatment age (one category for each year of age), education (5 categories), civic status (married vs. non-married), income (6 categories), and residential stability (9 categories). The combination of these variables provides us with more than 40,000 potential strata. After the matching we conduct a standard logistic regression of change between 2009 and 2013 on a range of variables. Figure 4 depicts the results graphically, while Appendix Table 7 of the appendix show the results numerically (see Appendix Table 8 of the appendix for a robustness test without matching which yields similar conclusions).
The results in Fig. 4 and Appendix Table 7 indicate that a change in the availability of a potential voting partner is highly consequential for individual turnout. The results are especially consistent for those who abstained in 2009 (the bottom half of Fig. 4 and model 1–2 in Appendix Table 7). For those who were in a single elector household in 2009 gaining a potential partner resulted in an increase in turnout of about 10% points compared to those who did not gain a partner (bottom left of Fig. 4 and model 1 of Appendix Table 7). Among those who did have a potential voting partner in 2009, losing a potential partner resulted in a 5% point decrease in turnout (bottom right of Fig. 4 and model 2 of Appendix Table 7). For those who voted at the outset, losing a partner resulted in a 6% point drop in turnout (top right of Fig. 4 and model 4 of Appendix Table 7), but there is almost no effect of gaining one (top left of Fig. 4 and model 3 of Appendix Table 7). This might be because individuals voting at the outset were very likely to vote regardless of gaining a partner. Moreover, as well as inducing turnout, gaining a potential voting partner might disrupt previous voting patterns. For example, inevitably some subjects (including those that voted in 2009) gained a non-voting partner, which may have a demobilizing effect (Partheymüller and Schmitt-Beck 2012). In further analysis we tested this potential demobilizing effect by splitting the sample by whether those gaining a partner were joined by someone who was a voter or a non-voter in the previous election.Footnote 10 The analysis (reported in Appendix Fig. 6) shows that the effect of gaining a partner is positive for non-voters moving in with either a voter or a non-voter, although the positive effect is larger for those who gained a voting partner. Moreover, even prior-voters who gained a non-voting partner saw no discernible drop in turnout. Together these findings suggest that the positive impact of gaining the opportunity for voting together (the companion effect) outweighs any potential negative effects of anti-voting social norms. To sum-up, the results in Fig. 4 provide support for our first hypothesis: acquiring a potential voting companion leads to an increased turnout probability, whilst losing one has the opposite effect.
We noted above that, despite the panel design, it is possible that observed correlation between changes in turnout behavior and household status could be the result of a third factor driving both. An alternative way of approaching the question, which overcomes this, is to examine whether individuals who voted together in one election behave differently in subsequent elections to those who lived together but did not vote together (H2). More specifically, we can look at whether individuals who lose a voting partner are more adversely affected than individuals who split from a person they did not vote with. By focusing only on households that broke up we avoid the problem of unobserved variables that correlate with both household break up and changing turnout
To test this, we focus our analysis on households with two eligible-electors who voted in the 2013 election, which subsequently split up in the 6 months period between the 2013 and the 2014 elections. In other words, they did not live together at 2014 election, but did so in the 2013 election. In contrast to the previous analysis this has the advantage that, we are able to test directly the effect of the loss a voting companion, as opposed to any other household partner. We do not make any restrictions on their new household, i.e. they can be single or live with someone new, but we include an indicator of whether they lived with someone else in 2014. We look at the 2013 and the 2014 elections as 2013 is the first election in which time stamps are available (i.e. we do not have time stamps in 2009).
We estimate a logit model where the dependent variable is turnout in the 2014 European parliament election. We restrict the sample to those who voted in 2013 and had a partner who also voted to maximize the comparability of the ‘treatment’ and ‘control’ since, in both groups, the subject lived with an elector who voted in the previous election. The only difference between the groups is that in the companion ‘treatment’ group the pair attended the polling station together. Thus the key independent variable is whether the individual voted with the partner in 2013. Our expectation is that voting together in 2013 would have a negative effect on the change in turnout between the elections, as this would imply the loss of a voting partner as opposed to the loss of a partner who voted separately. In other words, if having a voting companion is important, then losing a voting companion should be more detrimental to turnout than losing a non-companion. This also allows us to separate the effect of merely living with a voter (which might be associated with increased normative influence or increased flows of information) from the effect of the opportunity to vote together. In other words, if the effects are just as large for the loss of a ‘non-companion’ co-habitee, this would suggest it is not the companion effect at play (and vice versa). The results are presented in Fig. 5. As in Fig. 4 we base the model on a matched sample created by CEM on pre-treatment variables and we apply appropriate weights in the regression. Note that the elections are only 6 months apart and therefore we are not able to control rigorously for time-varying variables which are mainly annual in the Danish registers. In Table 9 of the appendix we show the results numerically and in Appendix Table 10 we present an alternative model with no prior matching which yields similar results.
In line with the expectations we find a negative estimate of more than 8% points for individuals who split from a voting companion compared to someone who voted separately (see Fig. 5 or Appendix Table 9 ). In other words the negative effect of splitting is markedly higher for individuals who in the first election voted together than for individuals who split from a person they did not vote with—even though focus only on voters. We cannot completely exclude the possibility that these differences are partly driven by relevant unobserved differences between splitting couples that had previously voted together and alone. In other words there may be some unobserved factor that is correlated with both the transient component of turnout and whether a voting companion or non-companion was lost (e.g. splitting from a spouse compared to a flat-mate). However, given the protection offered by the panel design, the restriction of the sample, and the CEM, along with the substantial size of the effect, the results provide strong evidence in favor of the companion effect as an important mechanism driving turnout.