Forecasting House Seats from Generic Congressional Polls: A Post-Mortem

Based on information available in July, we predicted that the Republicans would receive 52.9% of the total House vote and end up holding 229 seats, gaining control from the Democrats in the process (Bafumi, Erikson, and Wlezien Reference Bafumi, Erikson and Wlezien2010b). Our national vote forecast proved to be nearly correct, undershooting the actual Republican share (53.8%) by slightly less than one percentage point.Footnote ¹ Our seat forecast was a little less accurate. Although we did foresee the House changing hands, we did not predict such a large Republican windfall in seats—we forecast a “mere” swing of 50 seats, which was short of the actual outcome by about 13 seats.Footnote ² The Republican seat total of 242, however, was well within the 95% confidence interval (199 to 259).Footnote ³

We used a two-step method to forecast House seats, informed by equations based on past elections. First, we simulated the midterm vote division based on the generic polls and the fact that the president is a Democrat. Second, with the simulated estimates of the partisan tide in place, we simulated the outcomes of the 435 individual contests. Our seats model incorporated the predicted swing of the national vote to forecast the vote in all districts, using separate equations for open seats and races with incumbent candidates (see Bafumi, Erikson, and Wlezien Reference Bafumi, Erikson and Wlezien2010b).

The first step worked well. The national vote reflects poll results at the time of the forecast plus the voters' tendency to gravitate away from the party of the president (Bafumi, Erikson, and Wlezien Reference Bafumi, Erikson and Wlezien2010a). Although the generic ballot polls at midyear showed a roughly even partisan division, our simulations projected that the Republicans would widen their once-small lead, which is exactly what happened. That our mean simulation underestimated the Republican vote share by nearly one percentage point by itself accounts for an underestimate of about five or six seats going Republican. In other words, had we forecast the actual vote share with perfect precision, our seat forecast would have been within seven or eight seats of the outcome, rather than off by 13. Our forecast thus lost a little in the translation from the vote to seats. The equation for predicting open seats worked very well, as we appear to have mispredicted only two winners—one in each direction, producing no net effect. The equation for incumbent seats did not work as well, as it slightly underestimated (by seven to eight seats) the degree of incumbent loss. With this handful of exceptions, the 2010 election played out fairly close to our forecast based on patterns in previous midterm elections.