Labor Day weekend, or thereabouts, is typically when the American Political Science Association (APSA) holds its annual convention – an event in which hundreds of geeky men and women (the men wearing blue jackets, khaki pants and red ties, the women an assortment of power suits) invade some urban locale to present their latest research, drink lattes and run around everywhere with their identification badges always showing. This year it was held in Washington, DC – an appropriate setting because several of my colleagues took the opportunity to present their midterm election forecast models. Long-time readers will recall that a similar set of forecast models, all issued around Labor Day, proved quite prescient in predicting the outcome of the 2008 presidential election. If I recall correctly, only one prominent forecaster got the result of that race wrong, and most of the models came quite close to predicting the actual proportion of the popular vote Obama won.
I have long warned, however, that forecasting midterms elections, at least in the aggregate, is trickier; rather than predicting one outcome, forecasters have to deal with 435 House races and 33 (or 34) Senate races that do not all respond in identical fashion to the same set of factors. To be sure, not all of these House and Senate races are competitive. But many of these models roll these races together to form one observation for each midterm, so the forecasts are often based on a very small sample size. (Not all modelers do this, however.) So we should expect more variation in the midterm forecast models, and that’s precisely what has occurred. Before presenting the results of these models, I want to provide a brief overview of the underlying assumptions on which political scientists base their projections. I’ll then discuss the results from five of the models presented at APSA, focusing on the House. As we shall see, political scientists don’t necessarily agree on the likely outcome of the 2010 midterm races.
The history of midterm congressional election forecasting dates back at least to 1974, when Edward Tufte used regression analysis to predict the share of the national vote received by House candidates from the president’s party. Tufte assumed that midterms were in part a referendum on the president’s performance, so his model included a variable measuring the president’s popularity, as well as a second variable assessing the growth rate in real disposable personal income. These assumptions – if not the specific variables – have made their way into almost all subsequent forecast models. But the more recent forecast models, while building on Tufte’s precedent, have also tweaked his model in important ways. To begin, rather than focus on the popular vote, more recent models have tried to predict the actual number of seats won – a much trickier proposition. A second permutation has been to include a measure of “seat exposure” in the forecast model. The idea is that a party that occupies more seats is vulnerable to greater losses, whereas the minority party is likely to lose fewer seats. A variation on this is to include a measure for how long the majority party has been in control. Again, the idea is that the longer in power, the more vulnerable the party is to defeat.
The final development in forecasting models was to include some measure of voters’ attitudes toward the two parties as captured, for example, by the results of the generic ballot poll that I’ve written about extensively. So today most forecast models include some variable measuring the state of the economy, a second assessing the president’s performance, a third that tries to control for the number of seats in play for each party and, in some models, a fourth variable assessing voters’ attitudes toward the two parties. How these variables are operationalized, however, can vary from model to model, and not every variable is necessarily included in each model.
We see, then, that not all forecast models are alike. Which model one chooses is based in large part on what one believes determines the midterm results. Historically, as most of you know, more often than not the president’s party loses seats in the midterm – this is the so-called midterm loss phenomenon about which I’ve written extensively in previous posts. Broadly speaking, that phenomenon is usually attributed to some combination of the following factors:
1. As a referendum on the president;
2. As an effort to balance (or split) partisan control of government, often by voting against the president’s party
3. As a reflection of the changing composition of the midterm voting pool compared to turnout in presidential election years (the surge-and-decline) thesis.
My point is that political scientists have some latitude when putting together these models. Moreover, the models are constantly being updated, as each new election allows for additional out-of-sample forecasting and further tweaking of the models’ parameters.
With this background information, let’s turn to the five models unveiled at the APSA convention. What do they predict? (Keep in mind that Democrats currently hold 257 seats in the House, and that 218 is the dividing line between majority and minority status).
|Researcher||Mean Dem Seat Loss in House||Dem’s Lose House Majority?|
|Alfred Cuzan||-37 (-30 seats if “exceptional” years in which the president’s party actually gained seats are included in the model)||NO|
|Alan Abramowitz||-49 (assuming 5% Repub. generic ballot advantage)||YES|
|Michael Lewis-Beck, Charles Tien||-22||NO|
|Joseph Bafumi, Robert Erikson and Chris Wlezien||-51||YES|
Two of the models have the Democrats retaining control of the House, and three do not. How can this be? The explanation is rooted in the different assumptions built into each model. Cuzan’s Midterm Election Model, for example, analyzes the number of seats lost by the incumbent party as a function of the total number of seats held by that party as well as changes in GDP and in inflation. The Beck-Tien model, which shows the smallest Democrat loss, is based on the change in disposable income over the first six months of the election year, the President’s approval rating in June of that year, and a dummy variable indicating whether it is a midterm or presidential election year. (They use the same model to predict congressional outcomes in presidential elections years).
In contrast to Cuzan and Lewis-Beck/Tien, the Bafumi, Campbell and Abramowitz models predict much bigger Democrat losses. What is the difference with their models? One difference, as Pollster’s Mark Blumenthal points out here is that the models that incorporate some measure of the generic ballot question (“Do you plan to vote for the Republican or Democrat candidate?”) predict a much greater Democrat seat loss than do those that do not use this variable. Both the Abramowitz and Bafumi forecast model include results from the generic polling survey. In fact, the Bafumi model only uses the generic polling and a measure of the president’s party’s seats to predict the popular House vote, from which they then derive projections for House seats.
But this can’t be the only explanation. Note that Campbell’s model does not incorporate the generic ballot question results. Instead, he uses data, based on Charlie Cook’s assessments, that measures how many seats are considered vulnerable as well as presidential approval figures to arrive at a 51 Democrat-seat loss.
I’ll have much more to say about these forecast models, and others, before presenting my own predictions. But there are two important issues to keep in mind. First, I’ve presented the average predicted seat loss generated by these models – but not the uncertainty of those projections. The best forecasters are careful to provide some measure of the uncertainty underlying their projections, as based on previous results. That confidence interval can be more important than the projected results. Second, any model that is based in part on the generic ballots results is subject to more uncertainty because those survey results can vary during the remaining five weeks. (I’ll have more to say about this in another post.) Thus Abramowitz shows results ranging from a Democrat loss of 23 seats if Democrats are favored by 10% on the generic ballot over Republicans (highly unlikely given current polling) to a loss of 57 seats assuming a 10% Republican advantage on the generic ballot (more likely).
In parsing the differences among these models, we ought not to lose sight of the most important take-home message: none of them predict that Democrats will gain seats come November. If political scientists are correct, there will be no reprise of FDR’s first midterm results, or even George W. Bush’s in 2002. The only question is the size of the Democrats’ loss in the House. Will it fall closer to the traditional 24 or so seats characterizing most midterms during the post-World War II era, or is this likely to be one of those “wave” elections in which the party in power is swept out of office? I’ll give you my answer – but first I want to discuss the Senate projections. Here Democrats are likely to do better.