The paper
Marc S. Jacob, Yphtach Lelkes, and Sean J. Westwood (2026). Entrepreneurs of conflict: A descriptive analysis of when and how political elites use divisive rhetoric. PNAS Nexus 5(3), pgag038.
doi:10.1093/pnasnexus/pgag038 (opens in a new tab) · Replication data: osf.io/v5nr8 (opens in a new tab)
Main text
p. 2, Introduction
Published
Other accounts suggest that legislators skillfully blend policy advocacy with partisan attacks or that negativity is primarily a feature of election seasons (12, 21, 22).
Correction
Other accounts suggest that legislators skillfully blend policy advocacy with partisan attacks or that negativity is primarily a feature of election seasons (12, 21).
Basis Ref. 22, Green, Palmquist and Schickler's Partisan Hearts and Minds: Political Parties and the Social Identities of Voters, studies party identification among voters. It does not examine legislators' rhetoric or legislative activity. The same citation is removed on p. 9.
p. 3, Personal attacks and representation
Published
The persistence of vitriolic rhetoric outside traditional campaign periods suggests that American politics has entered what Sydney (50) termed the “permanent campaign”
Correction
The persistence of vitriolic rhetoric outside traditional campaign periods suggests that American politics has entered what Blumenthal (50) termed the “permanent campaign”
Basis Ref. 50 is Sidney Blumenthal, The Permanent Campaign (1980). See the correction to the reference list.
p. 4, Data sources
Published
All data collection was exhaustive (covering the full period of the 118th Congress).
Correction
All data collection was exhaustive (covering the full period of the 118th Congress), with one exception: the cable-news mention counts run from September 12, 2022 to September 5, 2024. Of the mentions used in the analysis, 19.4% predate the start of the 118th Congress on January 3, 2023, and the final 120 days of the Congress are not covered.
Basis Dates and counts are from the archived mention file (data/media.csv). The archived code (code/preparation/media.R) sums each legislator's mentions over the whole file without a date filter.
p. 4, Human validation procedure, steps 1 and 4
Published
Both coders independently classified a random sample of 500 statements from our corpus, using the same prompt as the LLM. [...] For personal attacks model performance was: 97% accuracy, 98% precision, and 92% recall. For critical debate, the model achieved accuracy of 81%, precision of 84%, and recall of 92%.
Correction
Both coders independently classified a sample of 500 statements from our corpus, stratified on GPT-4o's labels, using the same prompt as the LLM. [...] In this stratified sample, model performance for personal attacks was: 97% accuracy, 98% precision, and 92% recall. For critical debate, the model achieved accuracy of 81%, precision of 84%, and recall of 92%. Because the strata are GPT-4o's own labels, precision carries over to the full corpus, while recall and accuracy describe the stratified sample. Weighted to the corpus, where GPT-4o flags 1.3% of statements as personal attacks and 28.9% as critical debate, recall is 28% for personal attacks and 51% for critical debate, and accuracy is 97% and 72%. The implied true rates are 4.6% of statements for personal attacks and 47% for critical debate, so the measured rate of personal attacks understates the true rate by a factor of about 3.5.
Basis The validation statements, strata and labels are not archived, so the confusion matrices are recovered from the published percentages. With 500 statements, 105 integer matrices round to the personal-attack figures. All of them contain 127–186 gold-standard attacks (25–37% of the sample), and in all of them GPT-4o flags 24–35% of the sample, against 1.3% of statements in the corpus (p. 5). The 356 matrices that fit the critical-debate figures are 71–79% critical debate, with a false-positive rate of at least 43%, above every weekly critical-debate rate (maximum 42.1%). The corpus weights are the weekly means on p. 5. Values use the central matrix (140 true positives, 3 false positives, 12 false negatives and 345 true negatives for personal attacks). Across all consistent matrices, personal-attack recall spans 23–33% and the true rate 3.9–5.5%. With sampling error in the 12 missed attacks (Wilson 95% interval), recall spans 18–40% and the true rate 3.2–7.0%. For critical debate, recall spans 44–57%, accuracy 64–77% and the true rate 43–56% across matrices. See also the corrected SI Tables S1 and S2.
p. 7, Electoral performance
Published
Once we control for pre-election district competitiveness—operationalized using the Cook Partisan Voting Index (PVI) (66)—the coefficient on personal attacks becomes statistically indistinguishable from zero.
Correction
Once we control for pre-election district competitiveness—operationalized using the Cook Partisan Voting Index (PVI) (66)—the coefficient on personal attacks becomes statistically indistinguishable from zero. This model is not reported. SI Table S7, which controls for years in Congress, party, gender, and ideological extremity but not PVI, also shows no significant association (Model 1: −0.239, SE 0.250).
Basis SI Table S7 has no Competitiveness row, and its archived code (code/tables/tab_d3.R) does not include PVI. The lower panel of Fig. 4C, whose estimate is positive with a 95% interval above zero, controls only for time in office, party, and chamber (see the correction to the Fig. 4 caption).
p. 8, Fig. 3 caption
Published
Attacks are also three times more common on social media than on other sources (A).
Correction
Attacks are also more common on social media than on other sources: about 1.6 times as common as in press releases, and at least three times as common as in newsletters and floor speeches (A).
Basis The source means on p. 5 are 1.8% of X posts, 1.1% of press releases, 0.6% of newsletters, and 0.5% of floor speech segments. X exceeds press releases by a factor of 1.6, newsletters by 3.0, and floor speech by 3.6.
p. 8, Table 1
Published
Table 1. The most attacked people and groups. [...] Total 17,553
Correction
Corrected Table 1. The listed rows sum to 17,404.
Corrected Table 1: the most attacked people and groups. The total is the sum of the listed rows. The Joe Biden, Kamala Harris and Donald Trump rows combine several target strings; the Democrats, Republicans, Mayorkas, Tim Walz, Elon Musk and Fauci rows each count one exact string. Target N Joe Biden 8,546 Kamala Harris 2,958 Donald Trump 2,930 Democrats 779 Republicans 614 The Judiciary 537 Mayorkas 411 Tim Walz 321 Elon Musk 154 Fauci 154 Total 17,404 Basis The printed total exceeds the sum of the rows by 149, the number of attacks on “Merrick Garland” in the archived target file (data/attacks.csv), a target below the table's threshold of 150. The rows also follow different counting rules. The Democrats, Republicans, Mayorkas, Tim Walz, Elon Musk and Fauci rows each equal the count of one exact target string. The Joe Biden, Kamala Harris and Donald Trump rows combine strings: the exact strings “Joe Biden”, “Kamala Harris” and “Donald Trump” account for 4,785, 2,042 and 1,692 attacks. Of the 8,604 targets that contain “Biden”, the strings other than “Hunter Biden” (344) sum to 8,260, and the strings other than “Biden family” (194) sum to 8,410. Both are below the published 8,546, so the Joe Biden row includes attacks on his son and on his family. The archive's data/targets.csv lists other counts (Democrats 3,890, Republicans 3,527) and does not correspond to Table 1.
p. 8, Legislative efficacy
Published
they receive fewer assignments to prestigious standing committees, with the exception of senators, for whom no difference emerges (see SI Section 5 for full results).
Correction
they receive fewer assignments to prestigious standing committees, with the exception of senators, for whom no difference emerges (see SI Section 5 for full results; the Senate Republican estimate rests on 8 senators and 30 committee assignments).
Basis The committee models control for the Cook PVI, which exists only for House districts. The archived code joins PVI to legislators by state and district. Senators have no district, and the at-large House seats parse to a missing district, so only senators from Alaska, Delaware, North Dakota, South Dakota, Vermont, and Wyoming receive a value. The Senate Republican model in SI Table S32 therefore keeps 8 of the 49 Republican senators. The Senate Democratic model omits PVI and keeps all 48 Democratic senators. See the correction to SI Section 5.
p. 9, Fig. 4 caption
Published
The models include time in office, party, chamber, gender, district competitiveness (PVI), and ideological extremity (absolute value of the GGUM score) as controls.
Correction
The models in (B)–(D) include time in office, party, and chamber as controls. The model in (A) also includes gender, district competitiveness (PVI), and ideological extremity (absolute value of the GGUM score); because PVI is defined only for House districts, it includes only the 12 senators from the six states with a single at-large House seat.
Basis Control sets as in the archived code/figures/fig_4.R. In the archived data, PVI is joined to legislators by state and district, and only the senators from Alaska, Delaware, North Dakota, South Dakota, Vermont, and Wyoming receive a value (see the correction to p. 8). The other 88 senators drop out of any model that controls for PVI.
p. 9, Legislative efficacy
Published
That is, personal attacks do not translate into greater substantive legislative activity or policy victories (22).
Correction
That is, personal attacks do not translate into greater substantive legislative activity or policy victories.
Basis Ref. 22 studies voters' party identification (see the correction to p. 2).
p. 10, Fig. 5 caption
Published
In the bottom, we show standardized regression estimates with 90% and 95% CI and tenure, party, and chamber controls.
Correction
In the bottom, we show standardized regression estimates with 90% and 95% CI, controlling for tenure, party, chamber, gender, district competitiveness (PVI), and ideological extremity. Because PVI is defined only for House districts, the only senators in these models are those from the six states with a single at-large House seat.
Basis Control set as in the archived code/figures/fig_5.R. See the correction to the Fig. 4 caption for the PVI join.
p. 10, Changes in net worth
Published
This null finding holds in multivariate models controlling for tenure, baseline wealth, party, and chamber (see lower panel of Fig. 5).
Correction
This null finding holds in multivariate models controlling for tenure, party, chamber, gender, district competitiveness, and ideological extremity (see lower panel of Fig. 5 and SI Tables S25 and S26).
Basis No archived net-worth model controls for baseline wealth (code/figures/fig_5.R; code/tables/tab_d21.R and tab_d22.R).
p. 11, Conclusion
Published
Media outlets tend to emphasize insults and conflict over substantive discussion, creating what Iyengar and Kinder (73) term a “spotlight” incentive structure that appears to encourage conflict entrepreneurship.
Correction
Media outlets tend to emphasize insults and conflict over substantive discussion, creating a “spotlight” incentive structure that appears to encourage conflict entrepreneurship.
Basis Iyengar and Kinder (73) study the agenda-setting and priming effects of television news on viewers. The incentive structure for legislators is this article's own interpretation.
p. 13, Methods, District out-group affect
Published
Respondents rated Democrats and Republicans on separate 0–100 feeling thermometers; we define out-group affect as the thermometer rating of the respondent’s nonpreferred party, rescaled so higher values indicate greater dislike.
Correction
Respondents rated Democrats and Republicans on separate 0–100 feeling thermometers; we define out-group affect as the thermometer rating of the respondent’s nonpreferred party, rescaled so higher values indicate greater dislike. The district estimates used in SI Tables S27 and S28 are not rescaled: in those tables, higher values indicate warmer feelings toward the other party.
Basis The archived district estimates (data/out_cd.csv), which the archived code merges and labels “District Affective Polarization”, range from 17.0 to 37.3 across the 435 districts (mean 23.5), the scale of out-party thermometer ratings. They correlate at −0.77 with the archived in-party minus out-party gap (data/affpol_cd.csv, mean 50.6), which the code does not use.
p. 13, footnote d
Published
We further subdivided districts into competitive (PVI within ±5 points), semicompetitive (PVI between ±5 and ±15 points), and safe (PVI beyond ±15 points) categories and found no systematic relationship between conflict intensity and electoral performance at any competitiveness level (see Fig. S5).
Correction
We further subdivided districts into terciles of absolute PVI: competitive (PVI within ±8.67 points), semicompetitive (PVI between ±8.68 and ±17.04 points), and safe (PVI of ±17.06 points or more). The coefficient on conflict intensity is negative at every competitiveness level and statistically significant in safe districts (SI Table S7, Models 4–6: −0.419, −0.593, and −0.708; see Fig. S5).
Basis The archived code (code/preparation/elections.R) assigns terciles of absolute PVI across the 435 House districts in the archived Cook PVI file; the bounds above are the tercile limits in that file. Rebuilding the samples from the archived files reproduces the group sizes in SI Table S7 (140, 139, and 143). The safe-district coefficient is −0.708 (SE 0.356).
pp. 13–15, References
Published
12 Zeitzoff T. Nasty politics: the logic of insults, threats, and incitement. Oxford University Press, 2023. https://academic-oup-com.stanford.idm.oclc.org/book/46138. 14 Trussler M, Soroka S. 2014a. Consumer demand for cynical and negative news frames. Int J Press Polit. 19(3):360–379. 30 Trussler M, Soroka S. 2014b. Consumer demand for cynical and negative news frames. Int J Press Polit. 19(3):360–379. 32 Green J, Jennings W. 2017. Party reputations and policy priorities: how issue ownership shapes executive and legislative agendas. Br J Polit Sci. 29(2):443–466. 37 Ansolabehere S, Iyengar S, Simon A, Valentino N. Going negative: how political advertisements shrink and polarize the electorate. Free Press, New York, 1995. 50 Sydney B. The permanent campaign: inside the world of elite political operatives. Beacon Press, Boston, 1980. 74 Fenno RF. House members in their districts. Little, Brown, 1978.
Correction
12 Zeitzoff T. Nasty politics: the logic of insults, threats, and incitement. Oxford University Press, 2023.
14 Trussler M, Soroka S. 2014. Consumer demand for cynical and negative news frames. Int J Press Polit. 19(3):360–379.
30 The same article as ref. 14; citations of ref. 30 refer to ref. 14.
32 Green J, Jennings W. 2017. Party reputations and policy priorities: how issue ownership shapes executive and legislative agendas. Br J Polit Sci. 49(2):443–466.
37 Ansolabehere S, Iyengar S. Going negative: how political advertisements shrink and polarize the electorate. Free Press, New York, 1995. The same book as ref. 46.
50 Blumenthal S. The permanent campaign: inside the world of elite political operatives. Beacon Press, Boston, 1980.
74 Fenno RF. Home style: House members in their districts. Little, Brown, 1978.
Basis Going Negative (1995) is by Ansolabehere and Iyengar, the authors of ref. 46. Green and Jennings is in volume 49 of the British Journal of Political Science (doi:10.1017/S0007123416000636). Ref. 12 printed a university library proxy address.
Supplementary information
SI p. 2, Tables S1 and S2
Published
Table S1: Classification performance for personal attacks. [...] Table S2: Classification performance for critical debate.
Tables S1 and S2, “GPT 4o” rows: Accuracy 97%, Precision 98%, Recall 92% (personal attacks); Accuracy 81%, Precision 84%, Recall 92% (critical debate).
Correction
Corrected Tables S1 and S2. The validation sample was stratified on GPT-4o's labels, so the published rows describe that sample. Weighted to the corpus, GPT-4o's recall is 28% for personal attacks and 51% for critical debate, and its precision is unchanged.
Corrected Table S1: classification performance for personal attacks. The validation sample of 500 statements was stratified on GPT-4o's labels; the first two rows describe that sample. The last row weights GPT-4o's performance to the corpus, where it flags 1.3% of statements (central confusion matrix: TP 140, FP 3, FN 12, TN 345). Accuracy Precision Recall Trained annotators (validation sample) 92% 56% 82% GPT 4o (validation sample) 97% 98% 92% GPT 4o (weighted to the corpus) 97% 98% 28% Corrected Table S2: classification performance for critical debate. Rows as in Table S1; the corpus weight is GPT-4o's critical-debate flag rate of 28.9% (central confusion matrix: TP 342, FP 65, FN 30, TN 63). Accuracy Precision Recall Trained annotators (validation sample) 80% 83% 91% GPT 4o (validation sample) 81% 84% 92% GPT 4o (weighted to the corpus) 72% 84% 51% Basis See the correction to p. 4 for the method and the range across all confusion matrices consistent with the published percentages. The trained-annotator rows are reproduced as published; they also describe the stratified sample.
SI Section 3, p. 8
Published
Among Republicans, Representatives Marjorie Taylor Greene, Clay Higgins, and Darrell Issa are the most extreme conflict entrepreneurs.
Correction
Among Republicans, Representatives Lance Gooden, Marjorie Taylor Greene, and Darrell Issa are the most extreme conflict entrepreneurs.
Basis Figure S1 ranks Gooden, Greene, and Issa first through third among Republicans; Higgins is twelfth.
SI Section 4, p. 13
Published
The models include time in office, party, chamber, gender, district competitiveness (PVI), and ideological extremity (absolute value of the GGUM score) as controls.
Correction
Unless a table shows otherwise, the models include time in office, party, chamber, gender, district competitiveness (PVI), and ideological extremity (absolute value of the GGUM score) as controls. Tables S7 and S8 omit chamber and PVI, Table S23 omits gender, PVI, and ideological extremity, and Tables S27 and S28 cover the House only. PVI is defined only for House districts. Models that control for it include only the senators from Alaska, Delaware, North Dakota, South Dakota, Vermont, and Wyoming, the six states with a single at-large House seat, and the other 88 senators drop out. In Tables S5, S6, S9–S22, and S24–S26, the “Senate (vs. House)” coefficient therefore compares House members with at most 12 senators.
Basis The archived code (code/preparation/elections.R and merging_data.R) joins PVI to legislators by state and district. Senators have no district, and the at-large codes (“AK-AL” and so on) parse to a missing district, so the two senators of each at-large state receive that state's at-large PVI, and no other senator and none of the 6 at-large House members receives one. Rebuilding the samples from the archived files reproduces the published sizes: 440 in Table S5 (213 Democrats and 227 Republicans in Table S6), of whom 11 are senators, against 530 in Table S23, which omits PVI. Estimates for the full Senate need the legislators' rhetoric scores, which are not in the public archive.
SI p. 15, Table S6
Published
Table S6: OLS: Regressing ideal points on personal attacks.
Table S6, Models 4 and 5: “Ideological Extremity” −1.000* (0.000) and 1.005* (0.005); R² 1.000 and 0.999.
Correction
Table S6 is withdrawn. Its outcome is each legislator's signed GGUM ideal point, and its control for ideological extremity is the absolute value of the same ideal point. Within each party the two are equal up to sign for nearly every legislator: all 213 Democrats have negative ideal points, and 226 of the 227 Republicans have positive ones. The control therefore reproduces the outcome, and the personal-attack coefficients in Models 4 and 5 are zero by construction. Models 1–3 likewise regress the ideal point on its own absolute value.
Basis Code: code/tables/tab_d2.R. Re-estimating Models 4 and 5 from the archived ideal points (data/ideology.rds) and controls without the personal-attack share gives extremity coefficients of −1.000 and 1.004 (published: −1.000 and 1.005) and R² of 1.000 and 0.999, as published. A model with extremity as the outcome needs the legislators' personal-attack shares, which are not in the public archive.
SI pp. 16–17, Tables S7 and S8, Models 1–3
Published
Tables S7 and S8, Models 1–3: “Num. obs.” 429.
Correction
Models 1–3 of Tables S7 and S8 include one senator, Robert Menendez of New Jersey, who is assigned the 2022 vote share of Representative Rob Menendez (NJ-8, 72.9%). The House sample has 428 members.
Basis The archived code joins 2022 House returns to legislators by full name (code/preparation/elections.R and merging_data.R), and both members are “Robert Menendez” in the archived legislator file. Rebuilding the sample from the archived files reproduces the published 429 observations and the subgroup sizes of 140, 139, and 143; the senator has no district and is in none of the subgroups. The vote-share models are described as House-only (p. 12). Estimates without the senator need the legislators' rhetoric scores, which are not in the public archive.
SI pp. 20–21, captions of Tables S11 and S12
Published
Table S11: OLS: Regressing out-of-state donations on the relative personal attack score. [...] Table S12: OLS: Regressing out-of-state donations on the relative personal attack score.
Correction
Table S11: OLS: Regressing out-of-state donations (log) on personal attacks. [...] Table S12: OLS: Regressing out-of-state donations (log) on critical debate.
Basis The archived code (code/tables/tab_d7.R and tab_d8.R) regresses logged out-of-state donations on the personal-attack share and on the critical-debate share, the predictors the tables label “Personal attacks (%)” and “Critical debate (%)”.
SI pp. 24–25, captions of Tables S15 and S16
Published
Table S15: OLS: Regressing number of in-state donors (log) on personal attacks. [...] Table S16: OLS: Regressing number of in-state donors (log) on critical debate.
Correction
Table S15: OLS: Regressing number of out-of-state donors (log) on personal attacks. [...] Table S16: OLS: Regressing number of out-of-state donors (log) on critical debate.
Basis Tables S13 and S14 cover in-state donors. The archived code for Tables S15 and S16 (code/tables/tab_d11.R and tab_d12.R) models out-of-state donors, as the SI text on p. 13 states.
SI p. 30, Figure S5 caption
Published
Figure S5: The relationship between personal attacks and election returns by district competitiveness (Cook PVI dataset).
Correction
Figure S5: The relationship between personal attacks and election returns by district competitiveness (terciles of absolute Cook PVI; see the correction to footnote d). Members whose 2022 vote share exceeded 90% are not shown.
Basis The archived code (code/figures/fig_d1.R) drops vote shares above 90%. In the archived data this removes 17 of the 422 House members who have a vote share and a competitiveness group.
SI pp. 35–36, Tables S25 and S26
Published
Tables S25 and S26, row label “Years Serving Congress”.
Correction
Row label “Year first serving in Congress”. Expressed as years of service, the coefficients in this row change sign; their standard errors are unchanged.
Basis The archived code (code/tables/tab_d21.R and tab_d22.R) enters the calendar year in which the member first served, where the other tables use years in Congress.
SI p. 37, Table S27 caption and Section 4.6
Published
Table S27: OLS: Regressing district affective polarization on personal attacks (House only). Robust standard errors in parenthesis. [...] Table S27 shows linear models regressing bill (co-)sponsorship in personal attacks. S28 shows a similar model with critical debate as a predictor.
Correction
Table S27: OLS: Regressing district out-party warmth (mean thermometer rating of the other party; higher values indicate warmer feelings) on personal attacks (House only). Robust standard errors in parenthesis. [...] Table S27 shows linear models regressing district out-party warmth on personal attacks. S28 shows a similar model with critical debate as a predictor.
Basis The outcome in the archived code (code/tables/tab_d24.R and tab_d25.R) is the district mean in data/out_cd.csv, the unrescaled out-party thermometer (see the correction to the Methods, p. 13). On the dislike scale defined in the Methods, every coefficient in Tables S27 and S28 changes sign, with the same standard errors and fit. The section text described bill (co-)sponsorship, the outcome of Tables S23 and S24.
SI p. 38, Table S28 caption
Published
Table S28: OLS: Regressing district affective polarization on critical debate (House only). Robust standard errors in parenthesis.
Correction
Table S28: OLS: Regressing district out-party warmth (mean thermometer rating of the other party; higher values indicate warmer feelings) on critical debate (House only). Robust standard errors in parenthesis.
Basis See the correction to SI p. 37.
SI Section 5, p. 40
Published
Note that most members of Congress are assigned to several committees, which is why we consider each assignment separately. We cluster standard errors at the legislator level.
Correction
Note that most members of Congress are assigned to several committees, which is why we consider each assignment separately. We cluster standard errors at the legislator level. All models control for years in Congress, gender, and ideological extremity; the pooled models also control for party and chamber; and all models except the Senate Democratic models and the Senate Republican model in Table S30 also control for district competitiveness (PVI). Because PVI is defined only for House districts, the pooled models include only the 12 senators from the six states with a single at-large House seat, and the Senate Republican models in Tables S31–S33 include 8 senators (30 assignments), against 49 senators (206 assignments) in Table S30.
Basis Control sets as in the archived code/tables/tab_e2.R–tab_e5.R. Rebuilding the samples from the archived legislator, PVI and committee files reproduces the published Num. obs. and N Clusters: 938 assignments and 431 legislators in the pooled models, 30 and 8 for Senate Republicans in Tables S31–S33, 206 and 49 for Senate Republicans in Table S30, and 219 and 48 for Senate Democrats. See the correction to SI Section 4 for the PVI join.