Sunny Side Up

The Snap Election Forecasting Model of British Elections

Ross E. Burkhart

School of Public Service

Boise State University

            I make an election forecast of the 12^th December 2019 British general election based on a common structural model of elections, with a small innovation that accounts for snap elections. The forecast suggests that the Conservative government will lose seats, but with a larger number of parties than normal likely to win seats, the governing party or coalition that will emerge from the election is uncertain.

            This particular British election may very well be hazardous to forecast, given the controversial circumstances surrounding its calling, the high number of parties which will be contesting the constituencies with plausible chances of victory, and uncertainty over how the voters will weigh the primary issue, Brexit. (To call this election the "Brexit Election" is almost an understatement!) Structural forecasting models that rely on "heavy" traditional variables to make a forecast, such as economic and governmental performance, may well go awry given the special nature of this election. Yet forecasters cannot be choosy and remain credible. The election is upon the country in less than a month's time.

            For context, the structural models are pioneering British election forecasting forays. They have been joined recently by aggregate polling models from several polling firms and synthetic models combining the two approaches. The structural models employ economic and political variables measured with a good lead time before the election date, in order to generate a forecast well beforehand. Forecasting models are evaluated under three criteria: (1) sound theory behind the forecast, (2) non-trivial lead time for the forecast, and (3) model performance. Structural models that invoke common theories of voting (referendum voting à la V.O. Key, voter fatigue à la Alan Abramowitz) suggest an external validity to the forecasting enterprise. Combined with sufficient lead time, and good model performance, structural models play an integral forecasting role.

            All of these forecasting techniques presuppose a similar cognitive voter dynamic. The election dates are usually known months in advance, allowing voters the time to ready themselves for evaluating the parties and candidates. But what about snap elections? These occur in the Westminster parliamentary system for three reasons: (1) the government loses votes of confidence in the House of Commons, (2) the Prime Minister seeks to take advantage of her or his popularity to win more House seats, (3) the government is stymied prior to the end of Parliament in passing legislation due to deadlock in the House chamber, often caused by an attrition of seats held due to defections, resignations, and death.

            Snap elections characterize the current election campaign, and several other prior national-level general elections over the past 60 years. These elections are listed below, along with the date of dissolution of Parliament and the election date itself. (The dissolution of Parliament rapidly follows the declaration of the general election by the Prime Minister.)

TABLE 1: BRITISH SNAP ELECTIONS

From a practical standpoint, it makes sense that voters take stock of the condition of the country at the time of dissolution for a snap election, as while there may have been buildup to the dissolution, by definition there is some surprise to the declaration timing itself. Hence, voters should feel uncertainty in their evaluation, and uncertainty breeds a quick and dirty evaluation, right at the time of the election. Hence, the election forecast will have little lead time in the snap election environment, but conditions six months prior seem out-of-date.

            I re-create a structural British elections model forecast equation from the extant literature, with the innovation of utilizing data from the dissolution date for the snap elections. The modern structural modeling involves elections and data beginning in 1955, which is where I begin my dataset. The model is informed by Lewis-Beck, Nadeau, and Bélanger (2004), where the variables are as follows: percentage seats won by the incumbent party as the dependent variable (denoted as S below), and three independent variables: the public approval rate of the Prime Minister or P (either six months prior during non-snap elections or at the time of dissolution during snap elections), the jobless rate six months prior to the election or J, and the number of terms in office for the incumbent party assuming a fatigue effect for long-running spells in power for one party or T. The functional form of the model is thus: S = f (P, J, T)

            I estimate the model using a Prais-Winsten regression, Cochrane-Orcutt procedure (Stata statistical software) in order to account for autocorrelation. The model results are as follows:

S =   20.50 + .73*P + 1.61*J - 3.78*T

       (1.45)  (3.03)   (2.74)   (1.82)

R² (adj.) = .68   S.E.E. = 5.42    D-W = 1.84    N = 15

Where R² (adj.) = adjusted R-squared statistic, S.E.E. = standard error of estimate, D-W = Durbin-Watson statistic for autocorrelation after the Cochrane-Orcutt procedure, N = number of elections, figures in parentheses = absolute t-ratios.

            Model performance is good, if surprising. All independent variables reach conventional levels of statistical significance. The popularity and terms independent variables are signed in the expected directions: more popular prime ministers win more seats, and too lengthy a stay in government leads to seat declines. Strangely, the unemployment rate is positively connected to seat increase for the incumbent party. This may seem perverse, but perhaps is reflective of a generally high level of unemployment in Britain as the time period sampled went on, and may also suggest electoral forgiveness for high jobless rates if they trade off for lower levels of inflation and general income growth. (A technical point: there is no significant multicollinearity among the independent variables, with the highest Variance Inflation Factors (VIFs) being 1.5.)

            The forecast for 2019 is made simply by plugging in the relevant measurements for 2019, with the snap election polling at the time of dissolution revealing 40 percent approval of the Prime Minister, Boris Johnson.

S = 20.50 + (0.73*40) + (1.61*3.8) - (3.78*3)

S = 44.48

.4448 * 650 = 289 seats won by the government party

This forecasted result does not suggest a resounding victory for the Conservative government at the polls, rather a loss of nine seats after dissolution, but there is more than the usual uncertainty surrounding this particular snap election. (Witness the high standard error of estimate and relatively lower R-squared statistic.) While at once the Conservative government has gained in popularity in recent months, its lengthy stay in power (since 2010) may find the fatigue factor pushing voters toward the opposition. Low levels of unemployment actually give this variable the least impact on seat change. However, with several parties poised to win seats across the country, 289 seats may well be enough for the Conservatives to form the next government.