Elections experts Ken Goldstein, Walter Mebane, and Vincent Hutchings analyzed the results and key lessons of the 2018 Midterm Elections at a round table discussion hosted by the Center for Political Studies on November 13, 2018. A recording of the event is available below.
Ken Goldstein, Professor of Politics at the University of San Francisco
Ken Goldstein began his presentation by noting that there are still races that do not have a clear winner a week after the election, including the Senate and Governor’s races in Florida and the Governor’s race in Georgia.
Leading up to the Midterm Elections, some observers anticipated big wins for the Democratic Party. Goldstein observed that while there was a general lack of exuberance on the part of Democrats on election night, further reflection reveals that there were meaningful shifts in this election. Although the “blue wave” of Democratic wins didn’t materialize, the number of congressional seats changing away from the President’s party was of similar magnitude to past midterm elections.
Goldstein drew attention to the behavior of independent voters. Exit poll data show that independents favored Republican candidates for the House of Representatives in the past two midterm elections, as well as the 2016 General Election. In contrast, independent voters were more likely to vote for Democratic House candidates in 2018 by a margin of 12 percentage points.
Were the polls leading up to the election predictive of the actual outcome? Goldstein said they were fairly accurate, but reminded the audience that many congressional seats were not in play in this election. There are few high-quality state-level polls, which makes forecasting less accurate. More probability-based surveys that weight responses for education and race of the respondent would improve the accuracy of predictions.
Looking at the big picture trends, Goldstein observed that there was a substantial increase in the number of women running for office and winning, as well as large increases in non-white voters. He shared a map showing what the results of the presidential election would look if votes followed the same partisan break-down as the 2018 midterms. However, Goldstein cautioned that presidential campaigns are very different from congressional campaigns, and that a presidential candidate running a nation-wide campaign will face challenges in changing districts, especially in the Midwest.
Walter Mebane, Professor of Political Science and Statistics at the University of Michigan
Next, Walter Mebane presented analyses he has conducted using election forensics. Mebane coined the term “election forensics” to describe a set of statistical methods he developed to determine whether the results of an election accurately reflect the intentions of the electors.
Using Twitter data from the 2016 General Election Mebane analyzed reports of election incidents, including wait times and problems with voting. During the 2016 General Election people used Twitter to report different kinds of election incidents depending on their partisan affiliation. These incidents tended to be reported in replies to people with similar partisan affiliations.
Mebane discovered that there are partisan differences in the types of incidents that Twitter users shared during the 2016 General Election. For example, Republicans were less likely to report a long line to vote, but more likely to report registration problems. A significant conclusion from this finding is that such observational biases and communication silos suggest partisans tended to form different impressions of how the 2016 election went, supported by the divergent reported experiences. These patterns will probably continue in 2018, according to Mebane.
Vincent Hutchings, Professor of Political Science at the University of Michigan
Vincent Hutchings analyzed the shifting demographics of the American electorate. Hutchings presented data showing that Democratic voters have become more racially diverse in the past 20 years, while Republican voters have remained predominately white. Similarly, the Congress elected in 2018 is the most diverse in the history of the United States, but the increase in diversity has been primarily among Democrats elected to Congress.
Reviewing voting data by race, gender, age, marital status, and education, Hutchings notes that each demographic group voted for Democrats at a higher rate than they did in the 2014 Midterm Elections. However, the magnitude of change was different for each group.
Some elections experts wondered whether women would vote for Democrats at higher rates in 2018 in response to the #MeToo movement, the contentious confirmation of Justice Kavanaugh, and controversial remarks about women made by President Trump. Hutchings showed that, among white voters, men and women both shifted toward Democratic candidates, but the gender gap didn’t change. Married men and married women both moved toward the Democratic Party House candidates at roughly equal rates in 2018 compared to 2014. No matter how Hutchings examined gender, he found no evidence that white women behaved differently than comparable men, relative to their preferences four years ago.
Similarly, Hutchings observed meaningful trends related to generation and education. Voters under 30 years old voted for Democrats at a higher rate than voters under 30 in 2014. Democrats also increased gains among college-educated white voters compared to the 2014 Midterm Elections. Hutchings concluded that, while media may focus on gender differences between Democratic and Republican voters, more important differences are emerging along generational and educational lines, and these are trends to watch
“Gerrymandering”— when legislative maps are drawn to the advantage of one party over the other during redistricting—received its name in 1812, when Massachusetts Governor Elbridge Gerry signed off on a misshapen district that was said to resemble a salamander, which a newspaper dubbed a “gerrymander.”
But although the idea of gerrymandering has been around for a while, proving that a state’s legislature has deliberately skewed district lines to benefit one political party remains challenging.
The problem is that the mere presence of partisan bias in a district map tells us very little about the intentions of those drawing the districts. Factors such as racial segregation, housing and labor markets, and transportation infrastructure can lead to areas where one party’s supporters are more geographically clustered than those of the other party. When this happens, the party with a more concentrated support base achieves a smaller seat share because it racks up large numbers of “surplus” votes in the districts it wins, while falling just short of the winning threshold in many of the districts it loses.
Further, there are many benign reasons that legislatures may seek to redistrict voters—for example, to keep communities of interest together and facilitate the representation of minorities—that may have the unintended consequence of adding a partisan spin to the map.
The research of political scientists Jowei Chen and Jonathan Rodden is helping to differentiate cases of deliberate partisan gerrymandering from other redistricting efforts. Chen, Faculty Associate at the University of Michigan’s Center for Political Studies, and Rodden, Professor of Political Science at Stanford University, have devised a computer algorithm that ignores all partisan and racial considerations when drawing districts, and instead creates thousands of alternative district maps based on traditional districting goals, such as equalizing population, maximizing geographic compactness, and preserving county and municipal boundaries. These simulated maps are then compared against the district map that has been called into question to assess whether partisan goals motivated the legislature to deviate from traditional districting criteria.
We first wrote about Chen and Rodden’s work back in December 2016, detailing a 2015 paper in the Election Law Journal, which used the controversial 2012 Florida Congressional map to show how their approach can demonstrate and unconstitutional partisan gerrymander. Now, this work is back in the spotlight: Chen’s latest research has been cited in several cases of alleged gerrymandering that are currently working through the courts in Pennsylvania, North Carolina, Wisconsin and Maryland.
In January, Chen’s testimony as an expert witness was cited when the Pennsylvania Supreme Court threw out the state’s U.S. House of Representatives district map. In its opinion, the court said the Pennsylvania map unconstitutionally put partisan interests above other line-drawing criteria, such as eliminating municipal and county divisions.
The Pennsylvania districts in question were drawn by the Republican-controlled General Assembly in 2011. Immediately, the shape of the districts was an indicator that at least one traditional criterion of districting—compactness—had been overlooked.
Though few states define exactly what compactness means, it is generally taken to mean that all the voters within a district should live near one another, and that the boundaries of the district should be create a regular shape, rather than the sprawling polygon with donut holes or tentacles that characterized the Pennsylvania district map.
In particular, District 7—said to resemble Goofy kicking Donald Duck—had been called into question. “It is difficult to imagine how a district as roschachian and sprawling, which is contiguous in two locations only by virtue of a medical facility and a seafood/steakhouse, respectively, might plausibly be referred to as compact,” the court wrote.
Although there are more registered Democrats than Republicans in Pennsylvania, Democrats hold only five of the state’s 18 congressional districts. In the 2016 election, Democrats won each of their five House seats with an average of 75 percent of the vote while Republicans’ margin of victory was an average of 62 percent across their 13 districts. This is an indicator of “packing,” a gerrymandering practice that concentrates like-minded voters into as few districts as possible to deny them representation across districts.
Chen’s expert report assessed the district map and carried out simulations to generate alternative districting plans that strictly followed non-partisan, traditional districting criteria, and then measured the extent to which the current district map deviates from these simulated plans.
To measure the partisanship of the computer-simulated plans, Chen overlaid actual Pennsylvania election results from the past ten years onto the simulated districts, and calculated the number of districts that would have been won by Democrats and Republicans under each plan (see Figure 1).
The districting simulation process used precisely the same Census geographies and population data that the General Assembly used in creating congressional districts. In this way, the simulations were able to account for any geographical clustering of voters; if the population patterns of Pennsylvania voters naturally favor one party over the other, the simulated plans would capture that inherent bias.
Generally, the simulations created seven to ten Republican districts; not one of the 500 simulated districting plans created 13 Republican districts, as exists under the Republican-drawn district map. Thus, the map represented an extreme statistical outlier, a strong indication that the enacted plan was drawn with an overriding partisan intent to favor that political party. This led Chen to conclude “with overwhelmingly high statistical certainty that the enacted plan created a pro-Republican partisan outcome that would never have been possible under a districting process adhering to non-partisan traditional criteria.”
This table compares the simulated plans to the 2011 Pennsylvania district map with respect to these various districting criteria.
Following its ruling, on February 20 the Pennsylvania Supreme Court released a new congressional district map that has been described in a Washington Post analysis as “much more compact”. In response, the state’s Republican leadership announced plans to challenge the new map in court.
If Rick Snyder weren’t the Governor of Michigan, Donald Trump would probably have 16 fewer electoral votes. I say this not because I think Governor Snyder did anything improper, but because Michigan law provides a small electoral benefit to the Governor’s party in all statewide elections; candidates from that party are listed first on the ballot.
Yesterday, Donald Trump was declared the winner in Michigan by a mere 10,704 votes, out of nearly 5 million presidential votes cast. Although this is not the smallest state margin in recent history – President Bush won Florida and the election by 537 votes and Al Franken won his senate seat in Minnesota by 225 (after the result flipped in a recount) – it represented a margin of 0.22%. The best estimate of the effect of being listed first on the ballot in a presidential election is an improvement of the first-listed individual’s vote share of 0.31%. Thus, we would expect Hillary Clinton to have won Michigan by 0.4% if she were listed first and about 0.09% if neither candidate were consistently listed in the first position.
It may seem surprising to suggest that anyone’s presidential vote would hinge on the order of candidates’ names, but the evidence is strong. In a paper I published with colleagues in Public Opinion Quarterly in 2014, we looked at name order effects across 76 contests in California – one of the few states that rotates the order of candidates on the ballot – to estimate the size of this benefit. We later replicated the results in a study of North Dakota. Both times, we found that first-listed candidates received a benefit and that the effect was present, though smaller, at the presidential level.
There are many reasons that voters might choose the first name, even if they started their ballots without a predetermined presidential candidate. Some individuals might have been truly ambivalent and selected the first name they had heard of (in this case, “Trump”), others may have instead checked the first straight party box – listed in a similar order – without intending to select our new President-elect. Regardless of the cognitive mechanisms involved, the end result is clear – the “will” of the voters can be diverted by seemingly innocuous features of ballot design.
How broadly is this first-position benefit a problem? Across the country, only seven states vary the order of candidate names across precincts. Another nine choose a single random order for listing candidates in each contest, but use that same order across the entire state. And the rest generally use some combination of alphabetic ordering or a listing based on who won in prior elections at the state level. Michigan’s system – prioritizing the candidate who last won the Governor’s office – is among the most common methods.
Given the control that Republicans currently hold over governorships, this bias likely helps Republicans maintain their dominance over many state legislatures. And the effects of being listed first only grow as you move down the ballot. In our study of California, we found the average benefit for a Governor was also 0.31 percent*, Senators gained 0.37 percentage points of additional votes, and candidates for other statewide offices gained an average of 0.63 percentage points.
Source: Prevalence and moderators of the candidate name-order effect evidence from statewide general elections in California Pasek J., Schneider D., Krosnick J.A., Tahk A., Ophir E., Milligan C. (2014) Public Opinion Quarterly, 78 (2) , pp. 416-439.
For a better answer, we might look to the strategy adopted by Michigan’s neighbor to the South. Ohio produces a unique ballot for each precinct where the ordering of candidates’ names is rotated. Although it is too late to prevent this effect from altering the 2016 election, it may have less of an impact if everyone were not filling out ballots with the same candidates listed first.
*this would likely be larger if California did not elect its Governors in non-presidential years.
So-called “inflation truthers” have made recent news waves with claims that inflation is actually much higher than reported. Mainstream financialnewsorganizations have debunked the inflation truthers charges with the simple math of averages. But what if the truthers are just looking in the wrong place? That is, is there systematic bias not in reported inflation but projected inflation?
Grafström and Gandrud began with the widely accepted notion that under more liberal governments, the United States Federal Reserve tends to predict higher inflation. But why? Democratic administrations tend to try to lower unemployment, which causes higher inflation. Under more conservative governments, on the other hand, the Federal Reserve predicts lower inflation. Yet there exists little empirical support for these ideas. Instead, most work on inflation comes from the field of economics, with a focus on comparing Federal predictions with money market predictions.
To test these commonly held ideas, Grafström and Gandrud looked at the Federal Reserve’s predictions across time. The authors took Presidential party and actual monetary and fiscal policies into account. They found that, regardless of actual monetary and fiscal policies, under more liberal presidents, the Federal Reserve over-estimates inflation while under more conservative presidents, the Federal Reserve under-estimates inflation.
In the graph below, perfect predictions would create an error of 0. Points above the line correspond to over-estimation and points below the line correspond to under-estimation. As we can see, when a Democrat is president, estimate errors tend to be above the line, while the average of Republican errors falls below the line.
Errors in Inflation Forecasts Across Time by Presidential Party
Grafström and Gandrud also wondered if control of Congress plays a role. To test this, they considered the joint influence of presidential party and the majority party in Congress. As the graph below shows, presidential party drives the trend. Interestingly, a Republican controlled Congress makes the original results stronger. That is, with a Democratic president and Republican congress, there is greater over-estimation of inflation. Likewise, with a Republican president and Republican congress, there is greater under-estimation of inflation. The graph below illustrates these findings (0 would again represent a match between predicted and actual inflation)
Errors in Inflation Forecasts Across Time by Presidential and Congress Majority Parties
Given the clear links between presidential partisanship and inflation forecasts, the authors worry that this likely translates into biased monetary and fiscal policies. That is, over-estimated inflation under Democratic presidents may lead to more restrictive monetary and fiscal policies. On the other hand, under-estimated inflation under Republican presidents may lead to more expansive monetary and fiscal policies. In both cases, the policy changes would be based on forecasts biased by flawed but accepted rules of thumb about inflation under Democrat vs. Republican presidents.
Have you noticed how the products you look at online seem to follow you from site to site and the coupons you receive in the mail sometimes seem a little too targeted? This happens because a set of companies are gathering information about Americans and merging them together into vast marketing databases. In addition to creating awkwardly personal advertisements, these data might be useful for researchers who want to know about the kinds of people who are and are not responding to public opinion surveys.
But before marketing data are incorporated into social science analyses, it is important to know how accurate the information actually is. Indeed, there are many concerns about consumer data. It could be out of date, incomplete, linked to the wrong person, or simply false for a variety of reasons. If we don’t know when marketing data are accurate, it is going to be difficult to figure out how these data can be used.
This is where the work of Josh Pasek, Center for Political Studies (CPS) Faculty Associate and Assistant Professor of Communication, comes in. Pasek, along with S. Mo Jang, Curtiss L. Cobb, J. Michael Dennis, and Charles DiSogra, have a forthcoming paper in Public Opinion Quarterly about the utility of marketing data. With Gfk Custom Research, 25,000 random addresses were selected, with about 10% of those joining the study. The marketing data available on these individuals was then matched against data collected as part of the study.
Interestingly, many variables showed large discrepancies between the two sources. Incomes mismatched by more than $10,000 for 43% of participants, while education level differed in at least two measures for 25%. Even the number of people living at the address differed by two or more in 35% of cases. Pasek and colleagues also investigate missing data with three different analyses. Ultimately, they find that the amount of data missing from consumer data is vast.
But at the same time, the consumer data performed better than chance in predicting actual data for all variables. This may make them useful for marketing purposes, but Pasek cautions that social scientific applications could be problematic. As Pasek says, “The bottom line is that these data are not consistently accurate. Although they may be great for targeting people who are more likely to buy a particular brand of shoes, our results suggest that marketing databases don’t have the precision for many research purposes.”
This post is part of a series celebrating the 65th anniversary of the American National Election Studies (ANES). The posts will seek to highlight some of the many ways in which the ANES has benefited scholarship, the public, and the advancement of science.
In The American Voter, Campbell, Converse, Miller, and Stokes became part of this behavioral revolution as they considered audience traits in the context of politics. The main argument of the book holds that most American voters cast their ballots on the basis of party identification. Specifically, voter decisions pass through a funnel. At the opening of the funnel is party identification. With this lens, voters process issue agenda. They then narrow down to evaluate candidate traits. Finally, at the small end of the funnel is vote choice. This understanding of voters encompasses the “Michigan Model.”
In time, the Michigan Model was revised. The original Michigan Model held party identification as king. This thesis maps onto the strong post-World War II Democratic party, strengthened by Roosevelt. In the next few decades, party identification weakened. More recently, party identification reemerged stronger than ever due to a variety of factors, including changing campaign strategy and polarization.
So while these new generations of scholars find different balances between party identification and other factors influencing vote choice, The American Voter provided a bar against which this change could be measured.
The American Voter also enabled the tools of measurement with ANES. The American Voter utilized early waves of what would become the American National Election Studies (ANES), which Miller himself facilitated. The ANES developed into a multi-wave, decade-spanning project offering continuous data on the American electorate since 1948.
Cited over 6,500 times to date, the book remains a seminal text in political science.
