Earlier this week, I took on the subject of mandatory voting and how it could affect American elections. The post attracted some conversation in comments, on Twitter, and in e-mail to me. So I wanted to elaborate in this little Q&A format. Are you ready for 1,700 more words on mandatory voting? No doubt.
Figuring out the impact of mandatory voting depends on ascertaining the political attitudes of nonvoters. How can we do that, especially since nonvoters obviously don’t vote?
If the goal is to compare the views of all voters and all nonvoters in a particular election, here are ways to compare the views of voters and nonvoters that are prominent in the scholarly literature:
- Using surveys, compare the views of those who say that they voted and those who say that they did not vote. Here is one example.
- Using exit polls of voters, estimate how demographic characteristics are related to the vote via a statistical model. Then use the results of that model to impute the views of nonvoters in other surveys. Here is one example from my research. (There are other ways to go about this imputation, but the basic idea is the same.)
- Using a survey merged with actual turnout data from voter files, compare the views of confirmed voters and nonvoters. Here is one example.
What are the advantages and disadvantages of each?
The first method is simple and easy to implement in lots of surveys. However, it relies on self-reported turnout after elections, which is usually exaggerated. In one study that compared this method to the third method — the one that used verified turnout from the voter file — the authors found that the first method tended to exaggerate the differences between voters and nonvoters. But not every study necessarily finds that pattern.
Nevertheless, the third method still finds at least some differences between voters and nonvoters (see also here). Moreover, scholars who have used the first method often argue that there are not always substantively significant differences between voters and nonvoters anyway. So there is at least some common ground here.
The second method depends on the assumption that the relationship between demographic characteristics and the vote is the same among voters and nonvoters. In one sense, this is a reasonable assumption: campaigns often heighten the links between demographic characteristics and the vote. So, if nonvoters were required to vote, you could assume that the campaign would have this effect on them. Nevertheless, this is an assumption, and perhaps it would not hold true in every election.
The third method has the advantage of relying on verified turnout. However, not all survey respondents can be found in the voter file. It is probably safe to assume that those respondents were not registered voters and thus did not vote. But there may be at least a few cases where the problem was simply that the information that the survey gathered about the respondent simply didn’t match the information recorded when the respondent registered to vote.
It may also be the case that voter file data is not 100 percent accurate. The states and localities that supply these data can make mistakes. It is not clear, however, that these mistakes are numerous enough to change the results in large-scale analyses like the ones described here.
What do these methods find about nonvoters?
The results are pretty similar, actually. All of these methods tend to show that nonvoters are more likely than voters to report voting for Democratic candidates and to take liberal positions on at least some issues. As I said in the first post, the differences tend to be in the single digits of percentage points.
Have these differences grown over time?
You might expect that nonvoters and voters will have diverged because there has been growth in Democratic-leaning groups that do not vote at high rates, relatively speaking (e.g., Latinos or millennials). Much of the published literature has not extended beyond 2006.
Here are two data points from more recent elections. First, using voter files, Anthony Fowler compared the party registration of voters and nonvoters in 2010. He found that nonvoters were 10 points more likely to be registered Democrats.
Second, using the 2012 Cooperative Congressional Election Survey merged with voter file data, I compared the self-reported party identification of verified voters and nonvoters — with the assumption, as stated above, that survey respondents who could not be matched to the voter file were in fact not registered to vote. (I also eliminated a handful of people who stated that they were noncitizens.)
Here, I found that the percent who identified as a Democrat or said that they leaned toward the Democratic Party was actually the very similar among voters (49 percent) and nonvoters (48). However, the percent who identified as Republican was larger among voters (41) than nonvoters (32).
In our previous work on this subject, Jack Citrin, Eric Schickler and I computed the “partisan differential” — the difference in the Democratic candidate’s percent of the major-party vote among both voters and nonvoters. The analogous quantity using these party identification statistics from 2012 — that is, the Democrats’ percentage of major-party identifiers — is (49 percent/49 percent+41 percent) minus (48 percent/48 percent +32 percent), or about six points.
Fowler’s 10-point figure is also calculated in this fashion. Both statistics — the 10-point figure and the 6-point figure — are larger, on average, than the typical partisan differentials that we found in our study of Senate and state-level presidential elections from 1990-2006. We found that the typical partisan differential in the states was roughly two to four points.
There was, however, variation across years and across states. So the overall figures from 2010 and 2012 likely conceal some differences across states. And the methodology we employed for 1990-2006 (method No. 2 above) is different than the method used here (No. 3 above). We should be careful about the comparisons.
Nevertheless, these results suggest that, on average, the gap between voters and nonvoters may be larger in 2010 and 2012 than in at least some earlier elections.
Are the differences between voters and nonvoters large enough to change the outcomes of elections, assuming that there was 100 percent turnout?
Previous literature that has studied presidential and senatorial elections has found relatively few elections where the outcome would have been different. A key reason for this is that many elections are not competitive enough, and nonvoters are not distinctive enough, for nonvoters to elect a different winner.
There are exceptions, however. In my first post, I noted that Citrin, Schickler and I found that full turnout would have turned Gore’s narrow Electoral College defeat into a victory. We found the same in 2004 for Kerry. Either would have been consequential, to say the least. But such cases are the exception in the presidential and Senate races that have been studied to date.
As Zoltan Hajnal discusses, we are likely to find larger effects of full turnout in local elections — where the differences between voters and nonvoters may be larger, on average, and where turnout is routinely very low.
The open question is whether the somewhat larger differences between voters and nonvoters observed in 2010 and 2012 would have affected those elections, under the assumption of full turnout. I do not know of any detailed simulations for those years, and I haven’t conducted any myself.
What about the case of Australia?
Australia is the very rare case where we can observe the consequences of the implementation of compulsory voting. Anthony Fowler has found that it increased the vote for the Labor Party by seven to 10 points and also increased spending on pensions.
What does this case tell us about the United States?
The value of the Australian case is that we can get a reasonably clean before-and-after snapshot of the consequences of compulsory voting. But since Australia implemented this in the mid-20th century, one can reasonably can ask whether the results generalize to the United States in 2015. On the one hand, Australia experienced some of the same biases in turnout then as the United States does now — such as higher turnout among the relatively wealthy compared to the poor. On the other hand, Australia differs from the United States in other respects, such as its party system and voting rules.
Does this mean that making turnout higher or lower would have no consequences for elections?
No, it does not. My post was focused on the literature that tries to examine the impact of mandatory voting. That is different than the question about higher or lower turnout. Higher or lower turnout could have implications for both who wins elections and for subsequent policy, depending on the which kinds of voters comprise the “newly mobilized.”
For example, Fowler found that the type of voter who might not vote because of, say, a rainy day is more Democratic. And, although increased turnout by, say, racial minorities or the poor might not elect a different candidate in every case, it may change the calculus of elected leaders and produce policies closer to what those groups want. For example, here are two studies that find that the turnout of the poor is correlated with the generosity of social welfare benefits.
This does not mean, of course, that higher turnout means that Democrats or liberals will always win everywhere. For example, this study by Thomas Hansford and Brad Gomez exploited the fact that rain does affect turnout to see how higher turnout affects election outcomes. (Rain serves as the “instrumental variable” in the analysis.) They found that higher turnout helps Democrats more often that it helps Republicans, but the impact of higher turnout depends on how Democratic an area is and whether the incumbent is a Democrat or Republican. (There are questions too about how to interpret the results when rain is used as an instrumental variable. No research design is beyond reproach, as discussed above.)
Would any electoral reform that increases turnout therefore end up benefiting Democrats?
No. It’s really important not make this extrapolation based either on the simple descriptive differences between voters and nonvoters or based on any findings from the literature on mandatory voting. The voters who are mobilized by electoral reforms (or mobilization efforts by campaigns) are typically not representative of nonvoters. For this reason, electoral reform does not necessarily change the overall partisan complexion of the electorate (see here or here or here, for example).
I don’t really understand how you can even simulate a radical change like compulsory voting. How can we possibly know what would happen?
This is a perfectly reasonable questions. As I noted in my first post, politicians and parties would undoubtedly change their strategies if compulsory voting were law. It is difficult to predict the impact of those strategies. If we assume that both parties adapt in a rational fashion, then their adaptations may not give either side a clear advantage. But it is hard to know now.
I thank Anthony Fowler and Eitan Hersh for helpful conversations.