The determinants of trump’s defeat: what if the COVID-19 pandemic did not matter?

ABSTRACT

While the COVID-19 pandemic significantly impacted the lives of many people worldwide, it is debated whether it led to the defeat of incumbent Donald Trump in the 2020 American presidential election. We argue that the COVID-19 pandemic had a significantly negative impact on Trump’s support due to his conflicting and populist rhetoric, which culminated in contradictory behavior at a time when Americans sought a consistent leader to “rally round the flag.” We use both waves of the ANES 2020 Survey to determine what support for Trump would have looked like if the COVID-19 pandemic had lower influence in citizens’ electoral decision-making compared to real world conditions using regressions and a counterfactual strategy. Our findings suggest that Trump’s electoral defeat depended on multiple factors, with aggregate-level analyzes suggesting that Trump would have received more support had the management of the health crisis mattered less in voting decisions.

Acknowledgements

We thank David Dreyer at Lenoir-Rhyne University as well as the participants at the 2021 American Political Science Association Annual Meeting for their comments and feedback on an early version of this manuscript. A special thanks also to the anonymous referees. Their feedback throughout the review process has been incredibly valuable to the quality of our work.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability disclaimer

Data sharing is not applicable to this article as no new data were created or analyzed in this study. The 2020 ANES data can be downloaded from https://electionstudies.org.

Notes

1 As a robustness check, we also test the effect of an alternative 7-point scale measuring respondents’ ideological positions on the liberal-conservative continuum. Results are highly similar and shown in the Appendix.

2 As a counterpoint, however, it could be argued that in a highly polarized environment, national economic evaluations are to some extent endogenous to partisan, or candidate, support. Although research has shown that voters tend to rely less on partisan cues for their economic evaluations when economic conditions worsen (Dickerson and Ondercin 2017), as was the case in the US during the pandemic crisis, this is definitely a point to be more carefully explored in future research. In this study, we expect the controls for Democratic vs. Republican self-placement, supplemented by additional checks in the Appendix for ideological positions on the liberal vs. conservative scale, to at least partially account for this potential mechanism.

3 This pattern aligns closely with recent scholarship arguing the emergence of a new educational divide across Western democracies (Attewell 2022; Hooghe and Marks 2018).

4 The same models have been also replicated based on the logit function. Results do not show any substantive differences from those already shown in Table 1 in terms of direction, strength, and significance of the effects (p < 0.05). The models are available in the Appendix, where we also provide a replication of our full model based on state-level clustered standard errors in place of fixed effects. In this case, we again found results are highly similar to those in the main text.

5 For a general overview of this approach, please refer to Van der Brug, Van der Eijk, and Franklin (2007, 137–169).

6 The average probability shows the highest predictive accuracy also in the case of vote intentions, with 94.41% of cases correctly classified vis-à-vis the 90.79% and the 65.18% of the median probability and default 0.5 cutoff respectively. Differences across thresholds are less outstanding when in the logistic versions of the models (see Appendix), although in this case the average probability also stands out (steadily beyond 95% of predictive accuracy), establishing itself as a top-performing criteria for our analysis. For more details on the comparison between average predicted probability and other threshold criteria for binary classifications, please refer to Freeman and Moisen (2008).

7 In both cases, these slight differences are statistically insignificant (95% CI).

8 As a robustness check, all models and the corresponding counterfactual vote shares shown here have also been replicated using logit estimations. The results, available in Appendix, highly resemble the findings discussed in the main text.