The above image shows the forecasted Democrat and Republican 2018 House popular vote over time. Bold lines denote the mean prediction and the shaded areas above and below the line represent the margins of error. Individual dots naturally represent individual polls, each adjusted according to the methodology described below the table.
Generic Congress Polls
This table displays all logged generic Congress polls. It displays the pollster, date, aforementioned pollster grade, sample size, voter screen, current model weight, and the adjusted Democrat and Republican vote shares after applying Bafumi et al’s transformations. This will be updated with each new poll.
Forecasting the Popular House Vote: How it Works
The model builds on work done by political scientists Joseph Bafumi, Robert Erikson, and Christopher Wiezien who studied how (and how well) certain indicators predicted the House of Representatives midterm popular vote.
They found that the generic congressional poll offers observers and analysts the most powerful predictor of forthcoming election results. Naturally, the poll’s predictive power increases as the time to the election decreases, but the researchers noted that even 12 months before the election, slight adjustments to the generic poll accurately forecasts the future vote.
Numbers 1-3 outline Bafumi et al’s model. Steps 4-6 explain how this model builds on Bafumi et al’s work.
- Take the Democrat’s two-party poll share and express it as a deviation from 50 (ie, if a poll shows Democrats with 54 percent of the two-party responses, its deviation from 50 would be 4).
- Adjust this number according to voter screens. Polls of all adults or registered voters tend to overstate Democratic support viz a viz polls of likely voters. This step normalizes all polls to those of likely voters, benefitting Republicans.
- Put the resulting number into the model to forecast the Democratic popular vote’s deviation from 50. The model changes for different time periods in the run-up to election day, but, for purposes of this experiment, always take into account generic poll results and the current presidential party.
- Translate the deviation from 50 into forecasted Democratic and Republican two-party vote shares.
- Collect all polling data and further weight responses based on 538 pollster rating, poll age, voter screen, and sample size.
- Use the weighted average to forecast the November 2018 midterm popular vote shares for the two parties.
Early polls tend to overstate the support for the incumbent president’s party, which, historically, fares poorly during midterm elections. As the time to election day falls, the out-party (defined as the party not holding the presidency) tends to gain more support and post stronger generic Congress poll numbers. Therefore, the Bafumi, et al, model boosts the out-party’s forecasted popular vote results at all points leading up to the election, but with the out-party bonus being highest 240+ days before the midterm.
Bafumi et al’s analysis stretches 300 days prior to the midterm analysis. Though we’re currently 400+ days from the 2018 midterm, the forecast here provided uses Bafumi’s February model coefficients. If the out-party pattern extends back further, this model may well understate the forecasted Democratic vote share, though likely not by much.