The Economy and the Dynamics of the 2008 Presidential Campaign: Evidence from the National Annenberg Election Study

Abstract

This paper considers the impact of economic perceptions on vote intention in the 2008 presidential campaign with data from the two components of the National Annenberg Election Survey. It addresses a controversy over whether the collapse of Lehman Brothers and its aftermath altered the terms of competition, and produced the late‐campaign widening of Barack Obama's lead and his comfortable victory. Detailed attention to the chronology of the campaign, made possible by the structure of NAES data collection, indicates that the timing of key shifts is inconsistent with a simple economic interpretation of vote‐intention dynamics. Multivariate analyses indicate that the economy‐vote link weakened at the critical point.

Acknowledgments

This is a greatly revised version of paper presented to the Geurin‐Pettus conference on the 2008 US Election at Southern Methodist University‐in‐Taos. Elements of this paper and related work have been presented to audiences at Princeton, Stanford, the 2009 annual meetings of the Midwest Political Science Association and the American Political Science Association, and to the NAES graduate seminar at the Annenberg School for Communication. We thank these audiences and workshop participants for many helpful suggestions. Special thanks are due to Harold Stanley, Richard Niemi, and Herb Weisberg for pointed comments on the first draft. All responsibility for the contents of this draft remains with the authors.

Notes

1. For most respondents vote intention was captured by a question about intention framed as if the election were held on the day of interview. As early voting opened, respondents were asked if they had voted already and if so for whom. For early voters actual vote is substituted for intention. Respondents with no vote intention or a third‐party one are excluded from all analyses and figures. Details on voting questions can be found in the Appendix.

2. For a comparison of the last week's NAES telephone findings to the actual outcome, see Figure 5 in Johnston and Thorson (2009a).

3. Campbell (2008a, 142ff) proposes mechanisms that might explain the sequence.

4. Claims about 2000 and 2004 are based on NAES data as reported in Johnston et al. (2004) and Johnston and Thorson (2009a).

5. Awkwardly for predictions to the 2008 result, retrospective judgments seem less powerful when there is no incumbent in the race (Nadeau & Lewis‐Beck, 2001).

6. Indeed, on the basis of spring‐vs‐fall comparisons inside the ANES, Holbrook is generally skeptical of dynamics in economic perception.

7. But see Sniderman and Stiglitz (2008), who show that Barack Obama's race played a two‐fold role, repelling some whites but attracting others. Although Sniderman and Stiglitz decline to comment on net effect, their analysis suggests that the two effects roughly cancelled each other and were unlikely to produce an effect of the magnitude supposed by Lewis‐Beck and Tien.

8. Their pooling is executed by Stimson's (2004) algorithm. They do not remark on the fact that the pooling produces late‐campaign estimates of the horserace and especially of John McCain's share strikingly at odds with other aggregators (and implicitly with the NAES tracking), not to mention with the actual result.

9. Details on question wording appear in the Appendix.

10. Inspection of the daily plot for the days before 8 September suggests that there were no net shifts before that date.

11. For the sake of graphical clarity, we do not report daily numbers. The extra plots would introduce clutter and force us to increase the vertical scale to a degree that would obscure differences along the smoothed plots.

12. See, for instance, Pollster.com ⟨http://www.pollster.com/polls/us/08-us-pres-ge-mvo.php⟩, which has McCain peaking on 6 or 7 September and Obama regaining the lead on 13 September; and RealClearPolitics, ⟨http://www.realclearpolitics.com/epolls/2008/president/us/general_election_mccain_vs_obama-225.html⟩, which has McCain peaking on 8 September, either 8 September or 13 September as the turning point, and the lines crossing on 17 September. Reconstruction of Nate Silver's 538 series is difficult at this remove, but trolling its Poll Reports archive indicates that the 538's Electoral College model returned to forecasting an Obama victory on 18 September. Silver characterizes his algorithm as conservative. A tangible indication of this is that his system did not assign the lead to John McCain until 10 September, by which date the other trackers had McCain past his peak.

13. This is consistent with the point made emphatically by Evans and Andersen (2006).

14. Strikingly, the mean rating among Republicans was much higher in early September than in late August. The simplest interpretation that the euphoria generated by the party's convention extended to the economy. Republicans account for all of the apparent gains in optimism noted in Figure 3.

15. The estimation presented uses a random‐effects setup. The alternative would be a fixed‐effects setup, which is tantamount to assigning a dummy variable to each respondent. Fixed‐effects estimations are unbiased but can be inefficient. Random‐effects estimations are vulnerable to bias but deploy the information efficiently. A Hausman test of differences in coefficients between setups suggests that there are none of note, hence our use of the more efficient design.

16. Not included are any factors more proximate to vote intention, in particular respondents' ratings of candidates. When candidate ratings enter the estimation, they overwhelm the direct effect of all other variables, even (to an extent) party identification. Even variables that are themselves powerful factors in candidate ratings have their effects masked. We return to a discussion of candidates in the conclusion, however.

17. The mechanics of daily estimation are another reason why this paper uses linear probability estimations. For many days, logit or probit estimations of this paper's very simple model yield perfect predictions, such that coefficients are both massive and massively imprecise. Because it is fated to miss the target, the linear probability model yields consistently more plausible estimates.

18. Significantly different, that is, for collections of consecutive days. On several individual days, coefficients tended toward zero, as a reflection of error in small samples.

19. We do not report confidence intervals in Figure 5 for reasons of graphical clarity. Some sense of the order of magnitude can be gained from average standard errors for daily estimations. These are: for web panel regression, 0.05; for web wave 3–wave 4 lagged estimations, 0.10; and for telephone pre‐post, 0.15. By these criteria, the dynamic impact of the economy is never significantly different from zero in the telephone sample, sometimes significant in the web lagged estimation, and always significant in the web panel estimation. More to the point, shifts in the smoothed values for estimations look to be significant, at least for web estimations. For instance, the pre‐meltdown shift in the web lagged estimation traverses about 1.5 standard errors. The corresponding shift in the panel regression traverses about 1.0 standard errors. Bear in mind that the standard errors in this discussion are based on noisy daily data. The loess smoother, in contrast, pools information over a minimum of 10 days and a maximum of 20 days.