Analyzing Wisconsin’s Record Midterm Turnout

Friday, May 10, 2019

2018 was a high water mark for midterm voter turnout in Wisconsin — almost 60 percent turned out. While that’s still lower than most presidential elections, it’s the highest midterm since at least 1948.

Turnout over time

There’s a fair amount of variation from county to county:

There’s a fairly big range between the highest turnout counties and the lowest ones. Perhaps unsurprisingly, the change since the last midterm is also quite large.

Looking at counties by counties doesn’t tell the whole story, however. We can use counties as a way of looking at factors like race and partisanship that aren’t geographic in and of themselves but are correlated with it.


In 2016, turnout among black voters dropped by one-fifth, according to a Center for American Progress study that used county-level data. Did that pattern repeat in 2018?

It didn’t — if anything, the opposite happened.

Plot of turnout change between midterms versus the percentage of people of color in each county

\(R^2 = 0.0674\); trend line equation: \(y=0.0012x + 0.1389\)

It looks like there’s a stronger effect among counties with a lower population of people of color. Let’s do the same analysis just for those counties:

Plot of turnout change between midterms versus the percentage of people of color in each county

\(R^2 = 0.1525\); trend line equation: \(y=0.0040x + 0.1146\)

Even if we deliberately exclude outliers that don’t conform to the trend, there’s not much of an effect (the line is basically flat) nor is this trendline a particularly good fit even after removing the outliers.

So it’s clear that this time around, areas with more people of color did not see a reduction in turnout. However, areas with more people of color tend to have lower turnout:

Plot of turnout by racial percentage, done as small multiples of year.

Breaking it down by race, the picture is actually a little more complicated. While turnout is persistently lower in counties with a greater percentage of Native Americans and persistently higher in counties with a greater percentage of Whites, it varies among other groups:

Plot of turnout by racial percentage, done as small multiples of race and year.

In case it’s not obvious, each race is scaled differently. For example, the multiracial column’s scale goes from 0 to over 3 percent, but the Native American column’s scale goes from 0 to above 75 percent.

A note on how race is calculated

Race is tricky to define for calculation purposes for much the same reason regions of the U.S. are, albeit not as contentious. In short, it’s socially constructed, and it’s not always constructed in ways that are convenient for statistical analysis.

The Census Bureau treats whether a person is Hispanic as separate from their race or races. However, most Americans think of being Hispanic as a race, including most Hispanic Americans themselves. The Obama administration considered a combined question that also increased the number of races, but the Trump administration opted to keep the system used in 2000 and 2010.

Here I define a person of color to be any person who’s Black, Native American, Hispanic, or Asian American, even if they’re also White. This is to reflect the reality that people who are both White and another race are often not perceived as White (at least not as solely White).


Looking at partisanship is more predictive than looking at race, although it still doesn’t capture much of the story.

Turnout change in 2018 compared to democratic margin

\(R^2 = 0.2032\); trend line equation: \(y=0.1222x + 0.1639\)

Turnout change in 2016 compared to democratic margin

\(R^2 = 0.2064\); trend line equation: \(y=-0.0758x + -0.0239\)

I experimented with adding additional variables like age and race to try to explain more of the variance in turnout, without much luck.

One suggestive finding is that 2018 turnout isn’t closely correlated to 2016 turnout, suggesting that there’s quite a bit of variation even once you account for things like demographics, which are stable over the long term.

Voter ID

After voter ID laws took effect in Wisconsin in 2016 — long after they were first enacted — the question immediately became ‘Did the laws impact the results in Wisconsin?’

There’s no question voter ID laws make it harder for certain groups to vote, as reporting by the Wisconsin Center for Investigative Journalism and my own reporting for Rewire News has concluded.

What’s less clear is the size of the effect. Currently, the balance of evidence appears to be that it lowers turnout by roughly two percentage points:

  • A 2017 UW-Madison study (PDF) specifically looked at Wisconsin’s law. It surveyed 2,400 residents of Milwaukee and Dane counties who didn’t vote in 2016. It found that without voter ID, turnout in 2016 would have been between 0.9 and 1.8 percentage points higher in those counties.
  • A 2017 study reviewed research, including the GAO one below, and found a similar size effect, while noting in some cases it might not be statistically indistinguishable from zero.
  • A 2014 Government Accountability Office report (PDF) compared Kansas and Tennessee, which enacted voter ID laws, turnout to their neighbors’ turnout. It found a 1.9-2.2 percentage decline in Tennessee and a 2.2-3.2 percent decline in Kansas.
  • Back in 2012, Nate Silver found that most studies predict turnout would be lowered by around 2-3 percentage points — in line with the later studies above.

In contrast, a 2019 study found that strict voter ID laws don’t have an effect. To be precise, it found a small effect (less than 1 percentage point) that was statistically indistinguishable from no effect. My inclination is to say that you’d expect the occasional null result when studying a small effect. Of course, if future studies were to show no statistical difference, I’d re-evaluate.

What counts as a significant change in turnout depends on your perspective. If your goal is to get as many people as possible to vote, lowering it by two percentage points isn’t great, but your biggest concern is probably changing factors that affect turnout more. If your concern is social justice, a small overall effect doesn’t matter as much as the disproportionate effect on, say, black voters, which is much clearer. Finally, if your concern is whether close elections are affected, two percentage points can be crucial.

Surprisingly, most studies don’t measure the partisan impact. In Silver’s 2012 post, he estimated a one percent increase in turnout gave Democrats a 0.5 percent advantage. If that applied to Wisconsin in 2016, that means voter ID could have plausibly cost Hillary Clinton Wisconsin — something she’s in fact argued.

At first glace, my analysis of the 2018 midterms contradicts these studies showing a negative effect — to recap, the first midterm election after the laws took effect indicated an increase in turnout, including in Dane and Milwaukee counties.

However, it’s possible to harmonize the 2018 turnout outcome with the other research and analysis. In fact, I can think of at least two possible explanations:

  • It’s possible that the people affected by voter suppression don’t (usually) vote in midterms. If that’s the case, it’s no surprise there’s no decline since the last midterm election, although the fact it was record-breaking is surprising in itself. Take this with a grain of salt, however: I couldn’t find any research looking at this hypothesis and the Annual Review of Political Science paper I mention above found a study (p. 10.15 ) that found the opposite.

  • There’s a fair amount of year-to-year variance. If you look at historic voter turnout, swings of 4-5 percentage points are common. In other words, it wouldn’t be surprising if a 2 percent change was entirely masked by election-to-election variation.

I think the second explanation is more likely, but until there are more elections’ results to analyze or surveys of individual voters asking them about their past voting behavior and ID, I don’t think we can say for sure.

One major limitation of my analysis is that it’s at the county level, so it isn’t very granular. Taking Milwaukee county as an example, we’re essentially lumping together parts of the county as demographically different as the suburbs, the south side, and several colleges and universities. In the future, I’d like to redo this analysis at the ward level.

Further Reading