THE STATE PRESIDENTIAL APPROVAL/STATE ECONOMY MODEL

Like many forecasts, we included presidential approval. Historically, even state-level forecasts have used national-level approval ratings (Hummel and Rothschild Reference Hummel and Rothschild2014; Jerôme and Jerôme-Speziari Reference Jerôme and Jerôme-Speziari2016; Klarner Reference Klarner2012). A key contribution of our approach is to estimate the percentage approving of the president in each state. Footnote ² Building on our earlier work (Enns and Koch Reference Enns and Koch2013; Enns, Lagodny, and Schuldt Reference Enns, Lagodny and Schuldt2017), we used multilevel regression with poststratification (MRP), a statistical technique for estimating state-level public opinion from national surveys (Gelman and Little Reference Gelman and Little1997; Lax and Phillips Reference Lax and Phillips2009; Pacheco Reference Pacheco2014). Our estimates used 70 surveys with almost 90,000 respondents from June and July of election years.Footnote ³ After estimating the percentage in each state who approve of the president, we followed Hummel and Rothschild’s (Reference Hummel and Rothschild2014) strategy (for national-level approval) and subtracted a constant (so that when our approval variable equals zero, it is roughly equivalent to having no incumbent advantage) and multiplied the approval rating by −1 when the incumbent was a Republican (because our outcome of interest is the Democratic vote share). The online appendix provides additional discussion of all variables.

Presidential election outcomes also reflect economic conditions. We used the Federal Reserve Bank of Philadelphia’s monthly index of coincident economic indicators to measure economic conditions in each state. These data begin in January 1979; therefore, 1980 is the first election included in the analysis. This index uses four separate economic components (i.e., nonfarm payroll employment, average hours worked in manufacturing, unemployment rate, and wage/salary disbursements) to measure current economic conditions in each state.Footnote ⁴ Although leading economic indicators might be preferable to coincident indicators for election forecasts (Erikson and Wlezien Reference Erikson and Wlezien2016), state leading indicators were not available after February 2020 because the Philadelphia Fed suspended release of these data due to measurement complications from the COVID-19 pandemic. Similar to Erikson and Wlezien (Reference Erikson and Wlezien2016), we calculated the cumulative percentage change in coincident indicators through June of the election year, weighting months closer to the election more heavily.

The model also includes each state’s deviation from the national vote in the past election (Campbell, Ali, and Jalalzai Reference Campbell, Ali and Jalalzai2006; Hummel and Rothschild Reference Hummel and Rothschild2014)Footnote ⁵ ; home state of the presidential and vice presidential candidates; percentage of the vote in each state that went to influential third-party candidates in the previous election; and a binary indicator for southern states, capturing their Republican lean during the analysis period (Enns and Lagodny Reference Enns and Lagodny2020).

Table 1 presents the estimated relationships between these variables and the percentage of Democratic two-party vote share in each state (and Washington, DC) from 1980 to 2016. The relationships are in the expected direction, they are estimated with substantial precision, and the model fit is impressive.

Table 1

Predicting State Presidential Vote (% Democratic of Two-Party Vote), 1980–2016

Notes: *=p<0.05. All variables measured at the state level. Standard error in parentheses.

ACCURACY OF OUR BEFORE-THE-FACT FORECASTS

We report “before-the-fact” forecasts, relying only on model estimates from previous elections and data through July of the election being forecasted. These forecasts correctly predict the winner in 88% of all states from 1984 to 2016. We had the most difficulty with 1992, possibly due to Ross Perot’s third-party success and forecasting based on only three previous elections. Since 2000, we correctly predicted 94% of all states. Figure 1 presents our predicted vote for each state (y-axis) and the actual vote (x-axis). Most values align closely with the 45-degree line, which highlights the accuracy of our before-the-fact forecasts.

Figure 1

Before-the-Fact Forecasts and Actual Democratic Vote Share, 1984–2016

To see how our forecasts compare with others, we first considered the state presidential vote forecasts in the 2012 PS forecasting symposium (Berry and Bickers Reference Berry and Bickers2012; Jerôme and Jerôme-Speziari Reference Jerôme and Jerôme-Speziari2012; Klarner Reference Klarner2012) and Hummel and Rothschild’s 2012 state forecast (Hummel and Rothschild Reference Hummel and Rothschild2014). By combining model estimates from 1980 to 2008 with data available through July 2012, we could evaluate what our model would have forecasted around the time these researchers made their forecasts. The top section of table 2 shows that our forecast performs quite well, with the lowest absolute mean error and second-lowest absolute median error. Our 2016 forecast also compares favorably to the 2016 PS state-level vote forecast (see the middle section of table 2).Footnote ⁶ We also forecasted the national two-party vote by weighting the forecasted vote share for each state by that state’s population. The bottom section of table 2 shows that our national forecasts (based on our state forecasts) performed about as well or better than other national forecasts that also were conducted approximately 100 days before the elections (Abramowitz Reference Abramowitz2016; Erikson and Wlezien Reference Erikson and Wlezien2016; Lewis-Beck and Tien Reference Lewis-Beck and Tien2016). Our before-the-fact forecast correctly predicted 8/9 of the last popular-vote winners (missing 1992) and 6/9 of the Electoral College winners (missing 1992, 2000, and 2016). Our mean/median annual Electoral College error (46/42) is less than the Electoral College votes of the swing states of Florida (29) and Pennsylvania (20) together.

Table 2

Comparison of Forecasts

Notes: ^†Lewis-Beck and Tien (Reference Lewis-Beck and Tien2016, table 1) reported out-of-sample (i.e., Jackknife) predictions (based on all data except the year being predicted). Therefore, we did the same to allow direct comparison with their model from 1980 to 2016. All other forecasts are “before-the-fact.”

2020: 6-IN-10 CHANCE BIDEN WINS, 4-IN-10 CHANCE TRUMP IS REELECTED

Based on data from 104 days before the election, our 2020 model predicted that Biden will be the next president. We estimated that Biden will win 54.5% of the two-party vote share and 290 Electoral College votes; however, there is substantial uncertainty around these forecasts. Figure 2 reports this uncertainty by plotting the expected 2020 outcome based on 70,000 simulations. These simulations incorporate uncertainty in our model’s parameter estimates, uncertainty based on model error in previous elections, and uncertainty based on measuring current economic conditions (see online appendix 2 for full simulation details). Slightly less than 60% of the simulations predicted a Biden win, meaning that he has about a 6-in-10 chance of winning, leaving Trump’s reelection chances at about four in 10. Although six in 10 simulations predicted a Biden win, almost one in five simulations (19%) predicted a razor-slim Trump victory, with Trump receiving exactly 270 Electoral College votes.

Figure 2

2020 Electoral College Forecast Based on 70,000 Simulations

Figure 3 reports our predictions for each state. Dark-blue distributions are forecasted to go Democratic and light-red distributions are forecasted to go Republican. The height of the distribution indicates Electoral College importance. Although Arizona and Wisconsin lean toward Biden (hence, the dark blue) and Iowa and Florida lean toward Trump (light red), the roughly symmetrical distributions around the vertical 50% line indicate that these states could go either way. At 100 days out, Biden had a slight edge. He should focus on Arizona, Wisconsin, Iowa, and Florida to maintain that edge.

Figure 3

Simulated 2020 Forecast by State

Notes: Dark blue = expected Democratic win, light red = expected Republican win; distributions reflect simulated outcomes (i.e., distributions centered around 50% could go either way); distribution height indicates Electoral College importance; and the number next to the state = Electoral College votes.