A 2 million-person, campaign-wide field experiment shows how digital advertising affects voter turnout

Aggarwal, Minali, Jennifer Allen, Alexander Coppock, Dan Frankowski, Solomon Messing, Kelly Zhang, James Barnes, Andrew Beasley, Harry Hantman, and Sylvan Zheng. 2023. Nature Human Behaviour 7(3): 332–341. doi:10.1038/s41562-022-01487-4

We present the results of a large, US$8.9 million campaign-wide field experiment, conducted among 2 million moderate- and low-information persuadable voters in five battleground states during the 2020 US presidential election. Treatment group participants were exposed to an 8-month-long advertising programme delivered via social media, designed to persuade people to vote against Donald Trump and for Joe Biden. We found no evidence that the programme increased or decreased turnout on average. We found evidence of differential turnout effects by modelled level of Trump support: the campaign increased voting among Biden leaners by 0.4 percentage points (s.e. = 0.2 pp) and decreased voting among Trump leaners by 0.3 percentage points (s.e. = 0.3 pp) for a difference in conditional average treatment effects of 0.7 points (t_1,035,571 = −2.09; P = 0.036; D̂IC = 0.7 points; 95% confidence interval = −0.014 to 0). An important but exploratory finding is that the strongest differential effects appear in early voting data, which may inform future work on early campaigning in a post-COVID electoral environment. Our results indicate that differential mobilization effects of even large digital advertising campaigns in presidential elections are likely to be modest.

PDF Appendix Journal Replication archive (Harvard) Replication archive (Yale) Pre-analysis plan 1 Pre-analysis plan 2 BibTeX

Fig. 3 from paper: 2020 turnout rates by one-point bins of Trump support score and condition. The results are shown for early voting (left), election day voting (middle) and voting regardless of mode (right). The results in black are for the treatment group and the results in grey are for the control group. The error bars represent 95% CIs. Linear predictions from the unadjusted models reported in the bottom left facet of Fig. 2 are overlaid on the binned means, with shaded 95% confidence regions. The vertical scales of all three facets cover a 15 percentage point range, but the ranges differ across facets to emphasize the relevant variation. n = 1,999,282.