Hostname: page-component-6898cfcb55-w5hgh Total loading time: 0 Render date: 2026-11-11T18:59:56.416Z Has data issue: false hasContentIssue false
  • English
  • Français

Post-Election Reflections on Our Pre-Election Predictions

Published online by Cambridge University Press:  06 April 2005

Christopher Wlezien and
Robert S. Erikson
Christopher Wlezien
Affiliation:
Oxford University
Robert S. Erikson
Affiliation:
Columbia University
Rights & Permissions [Opens in a new window]

Extract

Prior to the 2004 presidential election we provided forecasts of the final vote relying ontwo different models. The first model follows what we did in 1996 and 2000 and relies on twovariables:

  1. the cumulated weighted growth in leading economic indicators (LEI) through the 13th quarter of the sitting president's term—March of the election year—which provides advance information about income growth leading up to Election Day while also summarizing prior income growth; and

  2. the current value of presidential approval, which absorbs some economic effects but also reflects non-economic aspects of incumbent performance.

Information

Type
Symposium
Information
PS: Political Science & Politics , Volume 38 , Issue 1 , January 2005, pp. 25 - 26
Copyright
© 2005 by the American Political Science Association

Prior to the 2004 presidential election we provided forecasts of the final vote relying on two different models. The first model follows what we did in 1996 and 2000 and relies on two variables:

  1. the cumulated weighted growth in leading economic indicators (LEI) through the 13th quarter of the sitting president's term—March of the election year—which provides advance information about income growth leading up to Election Day while also summarizing prior income growth; and
  2. the current value of presidential approval, which absorbs some economic effects but also reflects non-economic aspects of incumbent performance.

The second model adds a third variable: the incumbent party candidate's share in the most recent trial-heat polls, which allows us to capture comparative evaluations of the candidates to the moment. For more details, see Wlezien and Erikson (2004).

We initially provided forecasts in June and then again in August, the latter just before the Republican convention. The forecasts are of George W. Bush's share of the two-party vote, ignoring Independent Ralph Nader and other candidates. In June, we forecasted 52.8% and 52.2% using the respective models; in August, we forecasted 52.9% and 51.7%. The last number is what we reported in the October issue of PS. This is because the model on which it is based, including trial-heat polls, dominates the other using data from the late summer. That is, presidential approval adds nothing to what we know in the trial-heat polls at that point in time. Earlier in the year, presidential approval is a much better predictor.1

At the very beginning of the year, polls tell us absolutely nothing about the Election Day vote. When we regress the actual vote on early poll results, the adjusted R-squared is less than zero. (Also see Campbell 2000; Wlezien and Erikson 2002). This holds true using polls from January and much of February. After the New Hampshire primary, and especially as we move into March, the polls begin to offer real information but approval does too, independent of the polls. By late-summer, as we have noted, the latter no longer is true. Leading economic indicators on the other hand do add information until the very end of the campaign (Wlezien and Erikson 2004).

Given the final result of 51.2%, our second model obviously did well in August, missing by only 0.5%. Even the first model, excluding the trial heats, performed well, overpredicting the Bush share by a mere 1.7%. The corresponding forecasts from June were only slightly (0.2%) less accurate on average. What does this tell us? It's not that we have the magic model, as there isn't one. Let us consider the models' performance, beginning with the economy. The value of LEI growth through the first quarter of 2004 was slightly above average—exactly seventh best of the last 14 election years. It implied a healthy level of income growth during the remainder of the election year. Interesting, this did not follow. The actual rate of income growth in 2004 was disappointing, and cumulative growth for the term through the 15th quarter (the end of September) was only a bit more than half of what we would have predicted using our measure of LEI growth. The value in 2004 ranks as 11th best out of the last 14 election years, surpassing only 1980, 1960, and 1992, years when the in-party got the boot.2

Estimated cumulative income growth through quarter 15 of 2004 is 0.38 as compared to an expected 0.75. The mean level over the 14 elections between 1952 and 2004 if 0.62 and standard deviation is 0.36. For specifics on measurement, see Wlezien and Erikson (1996; 2004).

This is not good company to keep as an incumbent president running for reelection, though note that the next best year and closest comparable year is 1996, when Bill Clinton won in a landslide. Thus, income growth in 2004, while poor, did not by itself provide a sufficient basis for change.3

We can predict the vote using quarter 15 cumulative income growth and final presidential approval (see Wlezien and Erikson 1996). Given the late-October approval ratings (50% on average) from the Gallup poll and the realized value of income growth, we still would predict a victory for Bush with 50.5% of the vote share.

Of course, other things were at work, e.g., terrorism and the war in Iraq. Taken together, the other issues evidently didn't hurt Bush on Election Day given the campaign; indeed, they appeared to help. Throughout the election year, his approval rating hovered around 50%, at the cusp of the reelection danger zone. In other words, Bush enjoyed a residual level of popularity that exceeded what the mediocre income growth numbers warranted.4

Based on quarter 15 cumulative income growth in 2004, we would predict a final approval rating of 47%, three points below the actual value (see note 3). Note that the adjusted R-squared for the equation relating income growth and approval is low (0.20) and the standard error very high (11.4 percentage points), which tells us that predictions of approval based solely on income growth are not highly reliable. It also implies that non–economic factors generally matter more.

This afforded him the slightest of advantages heading into the fall general election campaign. Had his popularity sunk further toward that of his father in 1992, our model would have targeted him for defeat.

Although we could foresee a certain advantage for Bush earlier in the year, the likely outcome did not become clear in the polls themselves until Labor Day. Figure 1 shows Bush's percentage share of the two-party vote, ignoring Nader and other candidates, in the trial-heat polls conducted from January to November, 2004.5

For this exercise, we follow what we've done in previous election years using the many polls available at pollingreport.com. Where multiple results, reflecting different sampling universes, were reported for the same polling organizations and dates, data for the universe that best approximates the actual voting electorate is used, e.g., a likely voter sample instead of a registered sample. Although this follows our practice in previous years, relying on likely voter samples is not without its problems (see Erikson et al. 2004). Where there is overlap in polls, usually in tracking polls, conducted by the same survey houses for the same reporting organizations, only separate samples are used here. For example, where a survey house operates a tracking poll and reports 3-day moving averages, poll results for every third day are selected. This leaves 273 separate national polls. Since most polls are conducted over multiple days, each poll is dated by the middle day of the period the survey is in the field. The procedure allows readings for 151 days in 2004, 50 of which are during the 56-day post-Labor Day campaign. No adjustment to the polls is made for differences in question wording, sampling universe, survey house, or the various likely voter screens and weighting procedures. Some of these differences are of real consequence, however, especially the practices of different survey houses (see Erikson and Wlezien 1999). Poll results vary from day to day in part because the polls reported on different days are conducted by different houses.

The observations are the daily poll-of-polls and represent the Bush share for all respondents aggregated by the mid-date of the reported polling period. The data in Figure 1 exhibit pattern but also considerable noise, largely the result of survey error. The smoothed series offers more clarity.6

The series was generated using the “lowess” (locally weighted scatterplot smoothing) procedure. This procedure creates a new value for each time point based on the results of regressions using a designated number of surrounding data points (see Jacoby 1997). To generate the new value, predictions from these regressions are weighted based on their temporal distance from the particular point in question. Lowess tends to follow the data quite well, though the degree to which it does depends on the bandwidth, or share of time points in the full series, one uses to generate the smoothed values. For this exercise, a bandwidth of 0.1 was used to allow a reasonably close fit.

Bush began the year well ahead and quickly lost the lead as Senator John Kerry became the apparent Democrat nominee. His share stabilized at about 50% until the Republican convention, which produced a surge in support that persisted until the onset of the debates, approximately 30 days before the election, when his share dropped sharply. During the final month, Bush consistently polled about 51%, only a fraction less than what he received at the ballot box and we predicted in August.

Trial-Heat Polls Aggregated by Date, 2004

It was a narrow victory for President Bush. Although there was not an overwhelming basis in the public's mind for Bush's reelection in 2004, there also was not a clear basis for change. It thus was Bush's advantage, one that he realized on Election Day. The advantage was foreseeable in advance of the election. It also may make the outcome more understandable after the fact. The final verdict awaits the careful sifting of evidence, and this has only just begun.7

Consider the example set by Johnston, Hagen, and Jamieson (2004) in their analysis of the 2000 presidential election campaign.

A version of this article was presented at the Annual Meeting of the Northeastern Political Science Association, Boston, November 13, 2004. We are grateful to Liam Spender and Jason Kelly for collecting and organizing the trial heat poll data used in this article and to Richard Eichenberg, Steve Fisher, Michael Hagen, Mark Kayser, and Mark Wrighton for comments.

References

Campbell James E. 2000 The American Campaign: U.S. Presidential Campaigns and the National Vote. College Station: Texas A&M University Press.Google Scholar
Erikson Robert S. Costas Panagopoulos, and Christopher Wlezien. 2004 “Likely (and Unlikely) Voters and the Assessment of Campaign Dynamics.” Public Opinion Quarterly 68: 588– 601.Google Scholar
Erikson Robert S., and Christopher Wlezien. 1999 “Presidential Poll as a Time Series: The Case of 1996.” Public Opinion Quarterly 63: 163– 177.Google Scholar
Jacoby William. 1997 Statistical Graphics for Univariate and Bivariate Data. Thousand Oaks, CA: Sage Publications.Google Scholar
Johnston Richard, Michael G. Hagen, and Kathleen Hall Jamieson. 2004 The 2000 Presidential Election and the Foundations of Party Politics. Cambridge: Cambridge University Press.CrossRefGoogle Scholar
Wlezien Christopher, and Robert S. Erikson. 2004 “The Fundamentals, the Polls, and the Presidential Vote.” PS: Political Science and Politics 37 (October): 747– 751.Google Scholar
Wlezien Christopher, and Robert S. Erikson. 2002 “The Timeline of Presidential Election Campaigns.” Journal of Politics 64: 969– 993.Google Scholar
Wlezien Christopher, and Robert S. Erikson. 1996 “Temporal Horizons and Presidential Election Forecasts.” American Politics Quarterly 24: 492– 505.Google Scholar