2018 House Predictions

2018 House Predictions:

A Democratic Majority

House seat predictions
The heavy line shows the mean prediction with shaded areas showing the 95 percent confidence interval based on the model’s standard deviation. (Click for higher quality.)

Likelihood of Democrats earning a given number of seats (drawn from 40,000 simulations of the 2018 House predictions model).

2018 midterm election predictions
Likelihood of Democrats winning a given number of House seats in 2018. (Click for higher quality.)

How the 2018 House predictions work

The model

The model regresses the number of Democratic House seats won in a given election on these three variables across 41 House elections:

  1. Forecasted congressional popular vote (drawn from adjusted generic congressional polls)
  2. The previous number of seats won by the Democratic Party
  3. Whether it’s 1948 (an outlying year best explained by adding a dummy variable; see Cuzan and Bundrick)

Intuitively simple, it explains 95% of the Democratic seat variation across the elections with a RSME of 6.921.

2018 house seat predictions
Regression seat predictions versus actual seats won.  The point in the top-left quadrant is a cycle in which the model predicted the Democrats would be the minority party but they actually won more than 218 seats; the bottom-right quadrant shows the opposite.  In both cases the actual result fell within the margin of error.

Forecast methodology

The following steps outline how I generate the 2018 House predictions:

  1. Forecast numbers from each new generic Congress poll
  2. Run the model at least 40,000 times, with each iteration using a randomly selected point from a normal distribution centered around the mean forecasted popular congressional vote result
  3. Generate the mean seat graph
  4. Populate the histogram showing seat likelihood

Handling uncertainty

Since one variable — the popular House vote — cannot be known ahead of time, the model must address the inherent uncertainty in the popular vote forecast (which can be best interpreted as a range of likely outcomes).  To do so, each simulation uses a different popular vote number drawn from a normal distribution centered at the mean forecasted and with a standard deviation of the popular vote forecast’s RSME.

Those outputs, at least 40,000 of them generated with each new generic Congress poll (again, see the tracker), populate the seat distribution graph whereas the first shows the mean Democratic seat output as well as the margin of error from the model itself — it does not include results expected to occur around 5 percent of the time.

Potential shortcomings

The model does not look at individual races.  Doing so proves harder given the local nature of many House races, the inconsistency of time-series data (redistricting renders unusable some invaluable variables), and other difficult to forecast variables (eg, straight-ticket voting and levels of geographic sorting that polarize districts and leave fewer and fewer uncompetitive).  It’s entirely possible that while Democrats currently have a large forecasted popular vote lead, a leading indicator of House seat gains, their actual benefits may be less than expected because geographic sorting has created too many safe seats unaffected even by a large swing.