The paper
Derek E. Holliday, Yphtach Lelkes, and Sean J. Westwood (2024). The July 2024 Trump assassination attempt was followed by lower in-group support for partisan violence and increased group unity. Proceedings of the National Academy of Sciences 121(49), e2414689121.
doi:10.1073/pnas.2414689121 (opens in a new tab) · Replication data: osf.io/uphyd (opens in a new tab)
Main text
p. 1, Abstract
Published
We exploit the natural variation induced by the assassination attempt and large daily survey coverage (preattempt: 3,572; postattempt: 703; and 690 in a panel) [...]
Correction
We exploit the natural variation induced by the assassination attempt and large daily survey coverage (preattempt: 2,308; postattempt: 506; and 690 in a panel) [...]
Basis These are the Democrats and Republicans analyzed in Fig. 1; footnote ‡ limits the analysis to partisans. See the correction to p. 2 for the source of 3,572 and 703.
p. 2, Data
Published
Our analysis focuses on a month: 26 d before the attempt and the 4 d after, with a total sample of 3,572 preattempt and 703 postattempt respondents (out of 104,123 total sample respondents).
Correction
Our analysis focuses on a month: 26 d before the attempt and the 4 d after, with a total sample of 2,308 preattempt partisan respondents (1,282 Democrats, 1,026 Republicans) and 506 postattempt partisan respondents (273 Democrats, 233 Republicans), out of 104,123 total sample respondents.
Basis These are the samples analyzed in Fig. 1. The published 3,572 counts every interview, independents included, from midnight Eastern time on June 16 to midnight on July 16; it contains the 597 postattempt interviews and the 171 interviews from July 13 that the analysis excludes. The published 703 counts every interview dated July 14 to 16 in the archive's UTC timestamps, but the analysis window ends at midnight Pacific time on July 16 and keeps 52 of the 158 interviews dated July 16.
p. 2, Increased In-Group Attachment and Decreased Support for Partisan Violence
Published
or in perceived out-party support for partisan murder (Republicans: −0.4, P = 0.86; Democrats: −0.04, P = 0.99 on a 0 to 100 scale).
Correction
or in perceived out-party support for partisan murder (Republicans: 0.42, P = 0.86; Democrats: −0.04, P = 0.99 on a 0 to 100 scale).
Basis Republicans' mean perceived out-party support for murder was 40.5 before the attempt and 40.9 after it. Every other number in the sentence is after minus before. The change is not significant, so the sentence's conclusion stands.
p. 2, Fig. 1B, y-axis label
Published
Mean Count of Political Violnce Supported
Correction
Mean Count of Political Violence Supported
p. 2, Robustness
Published
As a robustness test, we constructed every possible window from our full dataset of 104,123 interviews (from 9/16/2022 to 7/12/2024). Of the 637 tests, the largest mean difference in support for partisan murder among Republicans is nearly 50 times smaller than the difference observed between the pre- and postassassination attempt.
Correction
As a robustness test, we constructed every possible window from our full dataset of 104,123 interviews (from 9/16/2022 to 7/12/2024), each a 27-day preperiod followed by a 5-day postperiod. Across the 632 windows with postperiod interviews, the weighted change in support for partisan murder among Republicans ranges from −3.38 to 6.83 percentage points. A decrease at least as large as the 1.97-point decrease observed between the pre- and postassassination attempt periods occurs in 20 windows (3.2%), and a change of at least that size in either direction occurs in 58 windows (9.2%).
Basis The published comparison divided the observed change in percentage points (1.97) by the largest placebo change in the archived code (figure1_placebo.R), which is stored as a proportion (0.0394); 1.97/0.0394 is about 50. In matching units the largest placebo change in that code is 3.94 points, 2.0 times the observed change. The changes above are weighted, as in Fig. 1; the archived code compares unweighted means. It generates 636 windows, 632 of which contain postperiod interviews; the text prints 637.
p. 2, Robustness
Published
We have balance on covariates in both the pre- and posttreatment samples (age, gender, race, partisanship 7-point, and political interest).
Correction
We compared the pre- and posttreatment samples within each party on age, gender, race, partisanship 7-point, and political interest. Democrats are balanced on all five covariates (each P ≥ 0.12; joint test P = 0.38). Republicans interviewed after the attempt are more often white (83.6% vs. 75.6%, P = 0.005) and more interested in politics (mean 1.64 vs. 1.78 on a 1 to 4 scale on which 1 is most interested, P = 0.022; joint test P = 0.039). Adjusting for the five covariates, the Republican estimates are −1.96 percentage points for support for partisan murder (P = 0.002), −0.09 for the number of violent acts supported (P = 0.003), 2.50 for in-party feeling (P = 0.069), and −1.13 for out-party feeling (P = 0.547).
Weighted covariate means before and after the attempt within each party (Democrats 1,282 before, 273 after; Republicans 1,026 before, 233 after). P values from the weighted t-test used for Fig. 1; the joint test is a weighted logistic regression of the postattempt indicator on all five covariates. Political interest excludes "don't know". Covariate Democrats before Democrats after Democrats P Republicans before Republicans after Republicans P Age (years) 47.10 47.51 0.725 52.58 50.94 0.205 Female 53.0% 58.0% 0.128 47.2% 46.4% 0.825 White 56.0% 52.8% 0.334 75.6% 83.6% 0.005 Partisanship (7-point) 1.63 1.69 0.294 6.23 6.26 0.617 Political interest (1 = most, 4 = least) 1.84 1.83 0.919 1.78 1.64 0.022 Joint test, all five covariates 0.380 0.039 Basis Every estimate in the paper compares the two periods within a party. Means are weighted, and each covariate is tested with the weighted t-test used for Fig. 1. The adjusted estimates add the five covariates to the regression in the SI (Extended Methods), with the survey weights.
p. 3, Discussion
Published
The assassination attempt occurred just before the Republican National Convention, although our post-data collection was completed the day before Trump’s speech. While the event and the convention together may represent a bundled treatment, our placebo results suggest that it is unlikely the differences we observed occurred by chance.
Correction
The assassination attempt occurred just before the Republican National Convention, which opened on July 15. Our postattempt interviews ended on the evening of July 15 (Eastern time), during the convention’s first day, and the archived data end on July 16, two days before Trump’s speech. The event and the convention together may represent a bundled treatment. In our placebo windows, a decrease in Republican support for partisan murder as large as the one we observed occurs in 3.2% of windows, so the observed decrease lies in the tail of the placebo distribution but is not unique to the period after the attempt.
Basis The convention ran from July 15 to July 18, and Trump spoke on July 18. Interview times are the archived timestamps. The placebo share is from the corrected robustness test on p. 2.
Replication archive
Replication archive, recode.R (analysis window)
Published
filter(starttime_parsed >= as.POSIXct("2024-06-16") & starttime_parsed <= as.POSIXct("2024-07-16"))Correction
filter(starttime_parsed >= as.POSIXct("2024-06-16", tz = "America/Los_Angeles") & starttime_parsed <= as.POSIXct("2024-07-16", tz = "America/Los_Angeles")). Without a time zone, the window edges fall at midnight in the time zone of the computer running the code, while the pre- and postattempt periods use UTC dates. The published estimates correspond to Pacific-time edges, and the table gives every Fig. 1 estimate under Pacific, Eastern and UTC edges.Fig. 1 estimates (after minus before) with the analysis window's edges set to midnight in each time zone. Analyzed partisans before/after: Pacific 2,308/506, Eastern 2,332/506, UTC 2,347/457. Outcome Group Pacific edges (published) Eastern edges UTC edges Support for partisan murder (pp) Democrats −0.83 (P = 0.333) −0.80 (P = 0.346) −0.60 (P = 0.520) Support for partisan murder (pp) Republicans −1.97 (P = 0.001) −1.95 (P = 0.001) −1.92 (P = 0.001) Support for partisan murder (pp) MAGA Republicans −2.99 (P = 0.006) −2.95 (P = 0.007) −2.90 (P = 0.009) Violent acts supported (0-4) Democrats −0.05 (P = 0.165) −0.05 (P = 0.167) −0.04 (P = 0.257) Violent acts supported (0-4) Republicans −0.08 (P = 0.002) −0.08 (P = 0.002) −0.08 (P = 0.004) Violent acts supported (0-4) MAGA Republicans −0.12 (P = 0.021) −0.12 (P = 0.022) −0.11 (P = 0.032) In-party feeling Democrats −0.28 (P = 0.854) −0.20 (P = 0.896) −0.66 (P = 0.675) In-party feeling Republicans 2.89 (P = 0.039) 2.88 (P = 0.039) 3.07 (P = 0.037) In-party feeling MAGA Republicans 2.88 (P = 0.071) 2.92 (P = 0.067) 3.20 (P = 0.049) Out-party feeling Democrats −0.98 (P = 0.579) −0.97 (P = 0.580) −1.19 (P = 0.517) Out-party feeling Republicans −2.34 (P = 0.186) −2.33 (P = 0.187) −2.69 (P = 0.142) Out-party feeling MAGA Republicans −0.57 (P = 0.817) −0.62 (P = 0.800) −0.73 (P = 0.773) Basis On a computer set to Eastern time the window gains 28 preattempt interviews (24 of them partisans): the Democratic in-party estimate becomes −0.20 (published −0.28), and the Democratic murder test gives P = 0.346 (published 0.333). With UTC edges, 49 postattempt partisans drop out and the MAGA in-party test gives P = 0.049 (published 0.071).