Consensus Points to 50-Seat G.O.P. Gain in House, but May Understate Uncertainty

FiveThirtyEight’s projection for the U.S. House shows little change from last week. Republicans are given a 73 percent chance of taking over the House, up incrementally from 72 percent last week. During an average simulation run, Republicans finished with 227 seats, up from 226 last week; this would suggest a net gain of 48 seats from the 179 they hold currently.

However, there is considerable uncertainty in the forecast because of the unusually large number of House seats now in play. A gain of as large as 70-80 seats is not completely out of the question if everything broke right for Republicans. Conversely, if Democrats managed to see a material rebound in their national standing over the final two weeks of the campaign, they could lose as few as 20-30 seats, as relatively few individual districts are certain pickups for Republicans.

In past weeks, we have written about the divergence between the various indicators that the model uses — for instance, the generic ballot, as compared against polls of individual districts. Increasingly, however, these metrics are falling into alignment.

Some generic ballot polls have shown incremental improvement in Democrats standing — although they still trail by roughly 6 points among likely voters on the generic ballot, according to our model’s estimate. According to one commonly used formula, a Republican lead of 6 points on the generic ballot would translate into a gain of about 50 seats.

Meanwhile, more and more polls of individual districts are being released (in excess of 150 districts have now received some type of polling this year). While it is somewhat difficult to determine whether these local polls show movement toward one party or the other over all, their relative abundance allows us to come to some general conclusions. For instance, 50 of the 73 districts that are closest to the national average in terms of presidential voting (those with a Partisan Voting Index of between R+3 and D+3) have been polled. The Republican candidate leads by an average of between 3 and 4 points in our adjusted polling average in these districts. Although many of the leads are small and the Democrat is ahead in our polling average in 18 of the 50, this nevertheless suggests that the “typical” House district slightly favors Republicans on the basis of local polling, which means that they are more likely than not to take a majority of the House over all.

Expert forecasts like those used by Cook Political are also employed in our model. These groups had been somewhat cautious before — but they have continued to move their forecasts in the direction of Republicans, and most of them are now suggesting a Republican gain closer to 45 or 50 seats than 35 or 40. These expert forecasts are now in strong alignment with the local polls. For instance, all 40 of the districts currently rated by a tossup by Cook Political have been polled. On average, our adjusted polling average of local polls in those districts show the two candidates essentially tied, with Republicans at 44 percent of the vote and Democrats at 43.

Finally, a poll of competitive districts conducted by National Public Radio — although it is not used directly in our model — is also consistent with an “over-under” line of a 50-seat Republican gain, according to an analysis by Glen Bolger of Public Opinion Strategies, who helped conduct the survey.

So, a broad array of indicators will get you somewhere in the vicinity of a 50-seat Republican gain — as does our model. Some wind up a little north of it and others a little south of it, but they paint a reasonably consistent picture.

This consensus, however, is somewhat misleading — because it does not adequately capture the fact that there is considerable uncertainty on either side of that 50 seat estimate, more so than in past elections. This is for two principal reasons.

First, as we mentioned, about twice as many seats are in play as in other recent elections. Thus, fairly subtle shifts in the political environment between now and Nov. 2 could have relatively profound implications for the seat count.

Second, there is considerable disagreement among pollsters on the magnitude of the enthusiasm gap. If Gallup’s likely voter model, which implies extremely lopsided turnout in favor Republicans, were to be correct, G.O.P. gains would be well in excess of 50 seats. Other turnout models, however, imply more like a 4- or 5-point enthusiasm gap, which would be more consistent with patterns in a typical midterm election. With an enthusiasm gap of that magnitude, Democrats would probably lose the House only narrowly and would have decent chances of holding onto it.

It is also important to remember that there are some factors, like the fact that many pollsters do not include cellphones in their sample, that could potentially result in the polls underestimating the position of Democrats. Our model assumes there is a chance that the overall “consensus” of polling could be off, which could affect the results in a great number of individual districts. This is one reason that it thinks such disparate outcomes as a 70-seat Republican gain or a mere 30-seat gain are not exceptionally unlikely. But such systematic bias in the polling could run in either direction.

Some trading markets imply that Republican chances of a House takeover are already in the neighborhood of 85 percent; in my view, this is too confident a position. But likewise, one should not dismiss the possibility of very large Republican gains.

*-*

There has also been one technical change to the model; readers who are not interested in arcane statistical details may wish to stop here.

In reviewing the performance on the model over past election cycles, I noticed that its error in individual districts was somewhat leptokurtic or “fat-tailed.” What this means in plain English is that it the model was more likely to make the occasional very large error (for instance, missing the margin by 20 points or more), but less likely to make medium-size errors, than a normal distribution would imply. We have now built in a correction for this, the upshot of which is that candidates who appear to be very significant underdogs are now given slightly better chances, e.g., a candidate who was considered a 500-to-1 long shot before might now be considered a 100-to-1 long shot. The adjustment has very little effect, however, on the overall forecast, or on races that had already appeared to be highly competitive.