Campaign Context and Preference Dynamics in U.S. Presidential Elections

Abstract

Previous scholarship finds that campaigns “matter” in that these enterprises have the capacity to influence voter preferences. Insights about how campaigns exert effects on preferences are less abundant, however. In this paper, we elaborate a theory of campaign effects that proposes campaigns matter, at least in part, because they function as a filter to mediate the impact of events. By amplifying and reinforcing the impact of relevant events, campaigns help voters process campaign developments, permitting citizens to form, crystallize or update candidate preferences. In quadrennial presidential contests, active campaigns are generally limited to battleground states, setting up natural experiments that allow scholars to investigate the claim that competitiveness influences campaign dynamics by generating vigorous campaigns that intensify the impact of events. We examine this hypothesis by comparing the dynamics of voter preferences for U.S. president in battleground and non-battleground states in the fall 2008 campaign using statewide survey data. The results confirm our hypothesis by showing that events exerted both short- and longer-term effects in battleground states, while any impact on voter sentiment in non-battleground states was short-lived.

Acknowledgement

The author is grateful to Robert Erikson, to the anonymous reviewers and to editors (current and former) for valuable comments and suggestions.

Notes

For previous cycles, see Panagopoulos and Wielhouwer (2008); Panagopoulos (2006); Bergan et al. (2005).

Admittedly, the preference estimates we employ are not based on random samples of voters in respective subgroups of states. This approach is not ideal given the potential for variability in the composition of states that comprise the daily poll of polls, but that does not necessarily compromise the reliability of these measures. We note that subsets of state polls vary for both battleground and non-battleground samples. As an additional robustness check, we note the overall, daily measures of Obama support for all states in our sample are significantly correlated with the daily averages of all available national samples as compiled by Real Clear Politics (r = 0.42, p < 0.01) as well as with Gallup's national, daily tracking poll estimates for this period (r = 0.47, p < 0.01) This suggests the estimates of voter sentiments we employ are, in the very least, useful approximations. Nevertheless, we adjust estimates for possible state effects as described below.

This adjustment scheme accounts for any bias potentially introduced by the fact that subsets of states comprising the battleground and non-battleground samples can vary by controlling for effects attributable to state-specific factors. A separate, potential concern is bias attributable to measurement error if there was simply a higher density of state-level polling in battleground states. We note this was not the case in our overall samples. The mean number of statewide polls included in our daily polls-of-polls was actually somewhat higher (nine) for the non-battleground subset, compared to the battleground pool (eight), but the difference was not statistically significant at conventional (p < 0.05) levels. Accordingly, we find no evidence that the concentration of polling differed substantially across the two key subsets of statewide poll samples.

We acknowledge that we combine surveys that differed in terms of sample types (likely voters, registered voters), polling organizations and survey methodologies to obtain the daily “polls of polls” estimates we employ. We note this approach was commonly adopted by poll aggregators in the 2008 election cycle.

We acknowledge there are alternative indicators of states' competitiveness, but these are not necessarily superior. Absent detailed information about the deployment of campaign resources including staffing, volunteers, advertising and mobilization efforts, alternative measures do not necessarily offer greater precision. Moreover, alternative indicators of battleground status generally overlap considerably (see Panagopoulos (2009a) for details).

As a robustness check, we replicate the analyses below using alternative designations for battleground states and find the results change only trivially. We detect no substantive changes. Data available from the author upon request.

See Huang and Shaw (2009) and Shaw (2006) for details.

Spending and appearance data are reported in Huang and Shaw (2009). We note further that overall state-by-state spending and appearances by the Obama and McCain campaigns are highly correlated; for spending, r = 0.94 (p < 0.01) and for appearances r = 0.78 (p < 0.01).

All correlations reported throughout are Pearson's r product-moment correlation coefficients for unadjusted data, unless otherwise indicated.

Lowess (locally weighted scatter plot smoothing) creates a new value for each time point based on the results of regressions using a designated number of surrounding data points. Predictions from these regressions are weighted based on their temporal distance from the point in question to generate the new value (see Erikson and Wlezien 1999 for additional details).

We focus our attention on the fall campaign. See Linn et al. (2009) for an analysis of overall campaign dynamics over a longer period in 2008.

Results that are substantively similar to those described below obtain when we estimated standard linear regression models with a lagged dependent variable on the right hand side. Data available upon request.

For example, we assert the economic crisis ensued following the Lehman Brothers collapse on September 15, and all polls included in our time series conducted after this date are denoted accordingly We note that alternative date designations as the start of the economic crisis do not alter the key findings.

Alternative designations for the duration of short-term effects ranging from one to five days were also considered and produced substantively similar results. Estimates are not presented but are available upon request.

It is likely this is an artifact of little variation in our indicator for the GOP convention given this spans nearly the complete duration of the period of the current study. Also, the adjusted preference estimates diminish the number of available observations slightly, so the impact of the GOP convention cannot be estimated for the non-battleground sample.

We acknowledge preferences in uncompetitive states exhibit somewhat less, but comparable, variance overall when adjusted for sampling variability. (Data available upon request; see also Panagopoulos (2009a)).

As a robustness check, we conduct a parallel exercise using survey data from the 2004 election cycle. We analyze post-Labor Day (September 6) presidential vote intentions using national, rolling, cross-sectional data gathered by the 2004 National Annenberg Election Survey (NAES), taking only preferences for Kerry or Bush into account. (Corresponding data for 2008 were not available for use; see Romer et al., 2006 for details about the 2004 NAES. Procedures adopted for this analysis followed Panagopoulos (2009a)). We divide respondents by battleground status based on Shaw's (2006) classification. Accordingly, battleground states in 2004 were FL, IA, ME, MO, MN, NH, NM, NV, OH, OR, PA, WA, WI, WV. In this analysis, we examine the long-term impact of four key events using the same coding scheme adopted for the 2004 analysis: three presidential debates (September 30, October 8 and October 13) and the vice presidential debate on October 5. The results of the Prais-Winsten regression, presented in Appendix 1, support the main contention we advance in this paper and are consistent with the 2008 findings we report above. Specifically, only the first presidential debate in 2004 appears to be significantly related to voter preference trajectories, but the effect is isolated to respondents in battleground states. Additional details are available upon request.