As they did in 2010 and 2012 and may continue to do for a while, the Republicans won the House in the 2014 elections. This outcome didn’t surprise anyone. The current partisan balance of seats is 244 for the Republicans and 185 for the Democrats with six more outcomes yet to be finalized. With 218 seats needed for majority status, this is an obviously good outcome for the Republicans.
In a post last week I summarized the final House forecasts from a variety of sources, and together they ranged from 239 to 250 Republican seats. Our final Election Lab forecast was for the Republicans to win 243 seats with a 90 percent confidence interval of 236 to 251 seats. Our forecast proved reasonably accurate.
What about the individual House elections? There are 429 races for which the results are finalized, and our model correctly predicted 415 of them (97 percent). There does not appear to be a party bias for the incorrect predictions: six of these 14 races were won by Democrats and eight were won by Republicans.
As we and others have explained, evaluating a model solely based on the number of correct predictions can pose a variety of problems. Another way to assess the forecasts is to compare the actual outcomes to the predicted probabilities of winning. Using this metric, forecasts do better when they predict the winner and do so with confidence (and they do worse when they predict the loser and do so with confidence). The measure is called the Brier score, and lower scores mean a forecast has less error (and therefore is better). So forecasters hope for a score of 0, with the worst being 1.0.
For each of our forecasts for the 429 election outcomes we computed the Brier score. Overall, the average score was .024, which matches the best Senate score (made by Drew Linzer for Daily Kos) and is marginally better than our Senate score of .027. Our median score (0.0) makes clear that many House races were easy to forecast correctly with 100 percent confidence.
That said, if we only focus on close elections, such as those where the margin of victory was less than 10 points, the accuracy of the predictions should be worse because those elections are more difficult to forecast. That is the case. We correctly predicted 27 of the 40 close elections (68 percent) with a notably higher average Brier score of .22.
What was our single biggest forecasting error? In Florida’s second House district, the Republican incumbent Steve Southerland lost to Democratic challenger Gwen Graham, with Graham narrowly winning with 50.6 percent of the two-party vote. Our forecast was for Southerland to win with 54.3 percent of the vote, and the model assessed Southerland’s chances of winning at 96 percent.
What went wrong? Three things. First, the model gives significant weight to incumbency and the midterm penalty, and those factors favored Southerland.
Second, the model considers whether polling data should lead us to revise the forecast based on fundamentals, and in this district, there was not much polling (only 2 post Labor Day polls). The first in September did favor Graham, but the second one in October favored Southerland and our model averaging process gives significantly more weight to polls conducted closer to Election Day (perhaps a feature of the model that should be reconsidered for elections like this one with so few available polls.)
Third, the model does not take into account idiosyncratic factors like the Democrat challenger being the daughter of a popular former governor and senator from the state (Bob Graham) who would be able to draw Bill Clinton and Jimmy Buffet to rallies on her behalf.
To be sure, it does not appear that other forecasters saw Graham winning the race easily, but they certainly did not forecast it to be an easy Republican victory as we did. Indeed, the Cook Report and the Rothenberg Report gave the race their most competitive ranking, which better accords with the actual result.
You win many, you lose some.
[NOTE: This post originally said that there was only 1 poll after Labor Day in Florida’s second district. In fact, there were two.]