The paper
Sean J. Westwood, Erik Peterson, and Yphtach Lelkes (2019). Are there Still Limits on Partisan Prejudice?. Public Opinion Quarterly 83(3), 584–597.
doi:10.1093/poq/nfz034 (opens in a new tab) · No public replication archive · Other public data used: doi:10.7910/DVN/D5WS0Y (opens in a new tab)
Main text
p. 585
Published
Ninety percent of Hillary Clinton’s ads attacked Donald Trump’s character (Fowler, Ridout, and Franz 2016), and Trump regularly called for Clinton’s jailing.
Correction
About 90 percent of the ads attacking Donald Trump focused on his personal characteristics (Fowler, Ridout, and Franz 2016), and Trump regularly called for Clinton’s jailing.
Basis Fowler, Ridout and Franz (2016) report that fewer than 10 percent of ads attacking Trump focused on his policies and about 90 percent on Trump as an individual, and that over 60 percent of ads supporting Clinton were solely about candidate characteristics. They do not report the share of all Clinton ads that attacked Trump’s character.
p. 585, footnote 1
Published
Our 2014 data were collected six to seven months before the midterm elections.
Correction
Our baseline data were collected before the 2014 midterm elections; the Study 3.1 baseline was fielded in 2013 (Lelkes and Westwood 2017).
Basis Lelkes and Westwood (2017), the source of the four baseline studies, states that its Study 3.1 (N = 395), the baseline for Study 3.1 here, “was fielded in 2013” (footnote 15). The baseline labelled “2014” throughout the article and SI therefore includes data collected in 2013.
p. 587
Published
The 2017 data (n = 1,377) were collected in June, and the 2014 data (n = 2,045) were collected in several surveys conducted over the first half of the year.
Correction
The 2017 data (n = 1,390) were collected in June, and the baseline data (n = 2,035) were collected in several surveys conducted in 2013 and the first half of 2014.
Basis 1,390 is the sum of the four 2017 study samples reported on pp. 589–591 (265, 351, 372 and 402). 2,035 is the number of respondents with an affective polarization score in the four public baseline files. Together they give the N of 3,425 printed in Figure 1A; the printed 1,377 and 2,045 sum to 3,422. On the 2013 fieldwork, see the correction to footnote 1.
p. 587
Published
This difference is rescaled between zero and one.
Correction
This difference is rescaled between zero and one, but not with the same anchors in every survey: a zero difference between the two thermometers scores 0.468 in the Study 3.1 baseline and 0.500 in the Study 3.2 baseline, so the pooled baseline scores do not share one fixed scale.
Basis In the public baseline files the score is an exact linear function of the in-party minus out-party thermometer difference (thermometers scored 0 to 1): 0.468 + 0.532 × difference in Study 3.1 and 0.500 + 0.500 × difference in Study 3.2. The 2017 data are not public, so the anchors of the 2017 scores, and a comparison of the two years on one fixed scale, cannot be checked from public data.
p. 587, description of Figure 1A
Published
The top panel (A) shows the average level of affective polarization in the two nonelection surveys is nearly identical (mean difference = –0.002, 95 percent confidence interval [–0.014, 0.010]).
Correction
The top panel (A) shows the average level of affective polarization in the two nonelection surveys is nearly identical (2017 mean 0.685, 2014 mean 0.684; difference 0.001).
Basis The text reads the difference as 2017 minus 2014 (“less polarized in 2017”). Under that convention the means printed in Figure 1A give 0.001, and no rounding of them gives −0.002. The 2017 data are not public, so the interval cannot be recomputed from public data. The scores compared are not on one fixed scale (see the previous correction).
p. 590, Study 2 sample
Published
A majority of respondents (59.85 percent) had at least a college degree, 25.36 percent had some college, and 15.10 percent had a high school degree or less.
Correction
A majority of respondents (59.54 percent) had at least a college degree, 25.36 percent had some college, and 15.10 percent had a high school degree or less.
Basis The three printed shares sum to 100.31 percent. The corrected college share is the complement of the other two.
p. 591, Study 3.2 sample
Published
The replication sample (n = 402) was slightly more Democratic (54.98 percent) than Republican (47.31 percent). The sample included more white respondents (70.90 percent) than nonwhite respondents (20.10 percent)
Correction
The replication sample (n = 402) was slightly more Democratic (54.98 percent) than Republican (45.02 percent). The sample included more white respondents (70.90 percent) than nonwhite respondents (29.10 percent)
Basis The printed party shares sum to 102.29 percent, and 47.31 percent repeats the Republican share of the Study 3.1 sample. The printed race shares sum to 91.00 percent. The corrected values are the complements of the Democratic and white shares.
p. 594
Published
In study 1, those with high levels of affective polarization suppress criticism of copartisans (β = 0.747, t = −5.823, p < 0.001)
Correction
In study 1, those with high levels of affective polarization suppress criticism of copartisans (β = −0.747, t = −5.823, p < 0.001)
Basis A coefficient and its t statistic have the same sign. The negative sign also matches the sentence: more polarized respondents are less willing to publish criticism of copartisans.
Supplementary information
SI p. 1, contents, and headings of sections 1.2, 2.2, 3.3 and 4.2 (pp. 2, 3, 6 and 8)
Published
Regression models in tabular forum
Correction
Regression models in tabular form
SI p. 4, Tables A5 and A6
Published
Martial Status
Correction
Marital Status
SI p. 5, Figure A1 and caption
Published
Mean differences with 95% confidence intervals. Differences between 2014 and 2017 that are significant at the .05 level are in green.
Figure A1, “Co-Partisans” panel drawn in black, with the figure note quoted above.
Correction
Figure A1, co-partisan panel drawn in green, as is the opposing-partisan panel. Caption: “Year means with 95% confidence intervals. Differences between 2014 and 2017 that are significant at the .05 level are in green.”
Basis The figure plots the mean fine in each year. In the co-partisan panel the 2014 interval ends at about $2,130 and the 2017 interval begins at about $2,200 (measured on the published figure), so the intervals do not overlap and the 2014 to 2017 increase is significant at the .05 level. In the public 2014 data the co-partisan mean fine is $1,741 (95% CI $1,333 to $2,149, n = 192), the value plotted for 2014. The 2017 data are not public.
SI p. 7, Table A9
Published
Table A9: OLS Models for Study 3 [...] Fine Amount [...] Note: *p¡0.05
Correction
Table A9: OLS Models for Study 3.1 [...] Log of (Fine Amount + 1) [...] Note: *p<0.05
Basis The 2014 columns reproduce from the public Study 3.1 baseline data only with log(fine + 1) as the outcome: constant 4.18, opposing partisans 0.84, N = 367. In dollars the 2014 co-partisan mean is $1,741 and the opposing-partisan difference $1,311. SI Figure A2 labels the same outcome “Log of Fine Amount”.
SI p. 8, Table A11
Published
Table A11: OLS Models for Study 4 [...] Agree with Punishment
Correction
Table A11: OLS Models for Study 3.2 [...] Support for Investigation
Basis Table A11 is in the Study 3.2 section and its 2017 N (402) is the Study 3.2 sample. The Study 3.2 outcome is support for an investigation of the donations, measured on a seven-point scale recoded to 0–1 (p. 591).