Measuring the Effects of Voter Identification Laws
Nate Silver, NY Times – Almost every day, I get e-mails and Twitter messages asking me about the effect of voter identification laws on turnout. Most of these messages, I presume, are from Democrats. They worry that more onerous laws, like those in Pennsylvania, could make it more difficult for Democratic-leaning voting groups like African-Americans and young voters to participate in this November’s election.
These concerns are perfectly logical — although it is also possible to exaggerate the effects that these laws might have. Academic studies suggest that they very probably reduce turnout, but not by more than a couple of percentage points. And although Democratic voters may be more affected by the laws, some Republican voters will be disenfranchised by them, too.
As I mentioned, there are quite a number of academic studies that seek to evaluate the effect of identification laws on voter turnout; John Sides has compiled a list of them here; or you can do some searching for yourself on Google Scholar.
On the surface, these studies seem to disagree with one another about whether or not there is any effect on turnout from harsher voter identification laws. But if you read them in more detail, you’ll find that much of the disagreement is semantic rather than substantive.
There is something of a consensus in the literature, in fact, about the rough magnitude of the effects. The stricter laws, like those that require photo identification, seem to decrease turnout by about 2 percent as a share of the registered voter population.
Whether this effect is deemed to be “statistically significant” or not varies from study to study. It depends on what particular type of statistical test the researcher has applied, and how much data he or she is looking at.
Statistical significance, however, is a funny concept. It has mostly to do with the volume of data that you have, and the sampling error that this introduces. Effects that may be of little practical significance can be statistically significant if you have tons and tons of data. Conversely, findings that have some substantive, real-world impact may not be deemed statistically significant, if the data is sparse or noisy.
My view is that something which might reduce turnout by 2 percent in a key state is meaningful in a practical sense — at least if you looking at the election in a detail-oriented way, as we often do.
Statistical significance tests start by specifying a null hypothesis. In the case of these studies, the null hypothesis is that voter identification laws do not impact turnout. Then it’s a question of whether the data is robust enough to persuade you otherwise. Some studies say that it is, and others say it isn’t.
However, the null hypothesis is not very logical in this case. Why should we give the benefit of the doubt to notion that voter ID laws will not affect turnout? The mechanism for how these laws work is very simple, after all. Some people show up at the polling place and find that they are not able to cast a ballot (or must vote by provisional ballot) when they otherwise would have voted. It would be stunning if these laws didn’t have some downward effect on the number of legal votes counted. (If you’re using Bayesian statistics, the hypothesis that voter ID laws do impact turnout would be your prior belief.)