Electoral rewards for political grandstanding

Significance In many representative democracies, legislators sometimes focus on making political points rather than making policy. Previous studies assumed or argued that they do so in an expectation of gaining more votes in the following election, but this claim has not been studied systematically. By looking at data on legislators’ statements, this paper demonstrates that US House representatives who made statements conveying political messages more intensely in any given two-year term tended to gain higher vote shares in the following election. In contrast, organized donors did not react to legislators’ political speeches but rewarded their legislative achievements instead. This raises concerns as legislators may represent organized interests when making policy while pleasing voters only by giving impressive, political speeches.

to the public is shown in Figure S2. This is a tweet with a video clip in which Alexandria Ocasio-Cortez, a Democratic House 30 Representative of New York, condemns Republicans for their anti-environmental move at a House Financial Services Committee 31 hearing on environmental justice. ‡ This video clip was originally excerpted from the YouTube channel of the committee and 32 tweeted by a social media user @nowthisnews. Then, the video clip quickly went viral as soon as she retweeted it on her own 33 wall. § The view count of this video clip is 9.4 million, and it was retweeted more than 5,800 times and liked by more than 34 20,000 users. 35 In addition, conducted a more comprehensive analysis of congressmen's tweets. Among the 12,080 tweets that congressmen 36 posted during the 115th Congress using their official Twitter accounts and contain the word "hearing", I randomly selected 100 37 tweets for the analysis. I find 88 of them are about congressional committee hearings and 48 of those about hearings contain 38 either the video clip for a hearing session or members' statements made during the hearings they mentioned. 39 To better assess how many video clips containing congressional hearing sessions congressmen post for public view, I collected 40 YouTube data. Using the official YouTube channels of the members of the 115th Congress provided by GovTrack.us, I extracted 41 information of the video clips that a random sample of 96 members posted during the 115th Congress. ¶ Around 354 (8%) out

89
In terms of the workers' quality, it was closely analyzed and monitored to achieve high quality work following the practice 90 introduced in Carlson and Montgomery (2). Specifically, only 500 to 1000 comparisons were posted online at a time, which 91 constitutes a batch. After retrieving each batch of work, poorly performing workers, defined as the ones providing answers 92 deviant from other workers, were identified by running a random utility model which estimates both the SentimentIt score as 93 well as the worker quality parameter for each worker. Then, the poor workers whose quality parameter estimates are below 1 94 are banned from further participating in the project. Also, note that in the process of computing the SentimentIt score, the 95 workers' decisions are weighted by their quality. (See Carlson and Montgomery (2) for more details.)

96
The inter-coder reliability is less relevant in this project because the paragraphs to be compared in each pairwise comparison 97 task are randomly matched. Thus, coders worked on different pairs rather than the same set making it challenging to compute 98 the conventional inter-coder reliability. However, I could identify 298 duplicate pairs out of over 30,000 comparisons, and 99 there were 149 unique pairs. Thus, each of the 149 were presented only twice in total. On 91 out 149 pairs (61%), the coders 100 agreed.
(2) conducted the same analysis using 50 movie reviews. In their project, workers were asked to choose a review which 101 has more positive tone. They found 1049 duplicates with 423 unique comparisons done more than once in their data. The  Type  Most Frequent Word Stems  Grandstanding  countri, problem, realli, congress, said, tri, concern, american, good, system, help, secur, lot, even,  fact, put, first, job, money, administr, provid, bill, law, point, last, mean, busi, forward, may,  understand, service, someth, hope, percent, believ, care, fund, cost, sure, health, part, unit, continu,  act, depart, presid, public, happen, area, number, give, done, got, kind, interest, made, certain,  appreci, inform, chang, wit, respons, million, support, report, legisl, agenc, whether, differ, hous,  protect, two, seem, process, great, everi, plan, polici, abl, tell, increas, move, day, budget, find policy issue and ask of witnesses' expectation on policy consequences, etc. Based on these analyses, the Grandstanding Score seems to successfully 119 capture key characteristics of grandstanding statements and distinguishes them from non-grandstanding statements.

120
The Five Most Grandstanding Statements. [1] " And even the environmental movement should realize that the worst polluters in the world were 121 the Socialist and Communist governments. People take better care of their own private property than is taken of property that is in public 122 ownership. I haven't been to this area, but I bet these people take good care of their property. It is so sad to think that a lot of these people, 123 or most of them, are people who came from Cuba where land was confiscated and taken away from them by a Communist dictator, and now 124 they come to the United States, a country that is supposed to be a free country, yet these people are coming all the way from south Florida to 125 Washington, DC, to try to defend their little piece of the American dream. The unbelievable thing is, this is happening all over this country. These 126 liberal, left-wing, Socialist, big-government types think they can run everybody's life better than they can themselves. So they don't mind coming 127 in and doing a really cruel thing and taking away this property. And I just think that we need to speak out against it and that it needs to stop, or 128 at least hopefully we can slow it down a little bit. Thank you."

129
[2] " Let me just say that this was a half-a-billion-dollar mistake and I would ask that we point out that there was an action taken by your 130 body under a justification of a legal definition that I think is a threat to both Democrats and Republicans that the word "is" is, and that for 131 somebody to sit there and ignore the law and redefine the word "is" I think the American people are outraged that a half-a-billion-dollar issue was 132 raised while legal jargon was ignoring the fact that the law is in there. And I don't think Democrat or Republican wants to have to add in every 133 law that it will never happen. I yield back to the gentleman." [3] " And were never willing to admit just how oppressive they really were. I think it Muslims around the world, many if not most who could be friends and be open to these kind of ideas of accepting people and not oppressing 136 somebody simply because they worship God in a different way. We need to call to task the Saudis and the Pakistanis and these other people who 137 have supposedly been on our side and quit trying to treat them with kid gloves because it ain't going to work. These regimes are basically gangster 138 regimes in terms of the way they treat their people and it shouldn't be tolerated and the United States has done that. Shame on us. Washington Post and Fox News. 14 articles were relevant to this hearing: 7 from each outlet. Then, I manually extracted members' statements 174 that were quoted in these news articles and identified members' statements that include these quotes from the hearing transcript data set. I found 175 12 statements in total. Then, I conducted a two-sample t-test that compares the average Grandstanding Score of these statements (68.37) and the 176 average Grandstanding Score of the rest of the statements that members made in this hearing (43.96). The gap between the two is statistically 177 significant (p-value: 0.000019). This analysis suggests that members' statements scoring high in the Grandstanding Score tend to be perceived as 178 newsworthy by journalists as well, which provides additional substantive validation of the Grandstanding Score.

180
A. Main Results. This section provides full regression results that generated Table 1 and Figure 3 in the main text. In particular, Table S3 is a full 181 version of Table 1 in the main text; Figure S4 presents two regression results based on which Figure 3 was drawn.  Note: * p<0.1; * * p<0.05; * * * p<0.01 Congress and committee fixed effects and hearing and legislator random effects are included. Standard errors are in parentheses.
have been captured by the variables included in the main analysis, this section replicates the analysis controlling for the unit-specific linear time 184 trend. For this, I take two approaches: using fixed effects and random effects. First, interactions between the member fixed effects and the time 185 variable are added to the list of controls in the full model. To capture time, the congress's number is used as a numeric variable. In this model, 579 186 more parameters are estimated. Second, an alternative way to control for the unit-specific time trend without sacrificing the degrees of freedom is 187 to allow a random slope for the time variable to vary by member. § § 188 Table S5 present results for these two regression models. In the first model addressing the unit-specific time trend using fixed effects, the 189 coefficient on the Grandstanding Score is not statistically significant. However, in the second model, the Grandstanding Score has a positive and 190 statistically significant coefficient. Thus, these tests provide mixed results. Why does the result in the first model addressing all the member-specific variations using fixed effects differ from the results of other model 192 § § In this model, only the random slope is introduced without a random intercept for each member because member fixed effects are already included in the model. However, I also tried adding a random intercept for each member, but the coefficients and standard errors for fixed effect variables including the Grandstanding Score do not change.

191
specifications? There can be several potential reasons for this.
First, almost 80% of the members in my data served only one, two, three, or four terms during the period that the data covers. Thus, controlling 194 for a unit-specific linear time trend might have soaked up too much variation in the vote share of these short-lived members leaving little room to 195 be explained by other substantively important covariates. In Model 2 presented in the main text, those who served only a single term are dropped as they do not allow analysis of within-member coefficients on covariates may grow larger in general. Then, the model is more likely to induce false negatives.

218
In summary, the set of analyses presented in this section generated mixed results. However, given the statistical and theoretical reasonings 219 described above and the analyses reported in the main text showing that reverse causality is unlikely, the overall findings imply that members' 220 grandstanding efforts have a correlational effect on their vote share. increase the incumbent's vote share is to mobilize supporters who did not vote previously to turn out to the extent that its increase is greater than 226 any increase in the turnout rate of non-supporters. † † † However, the turnout of both groups may decrease in midterm elections and increase in 227 presidential elections. Technically, then, as long as in-partisans are relatively better mobilized than independents or out-partisans compared to the 228 previous election after accounting for the dip in the midterm election, the incumbent's vote share may increase. Indeed, partisan messages are 229 known to increase voter turnout especially among base supporters (6). Therefore, political messages that members send out are likely to better 230 mobilize in-partisan voters and increase the incumbent's vote share in the following election.

231
The second way that an incumbent's reelection chances may increase is through persuading independents or opposite partisans to switch 232 their votes in favor of the incumbent. Theoretically speaking, if turnout rates of both supporters and non-supporters are fixed, the only way 233 that an incumbent's vote share can increase is to win votes from those who did not support the incumbent in the previous election. Politicians' 234 position-taking messages are found to have a persuasion effect by shifting public opinion (7). Studies note that voters tend to be persuaded by a 235 like-minded candidate's messages (8) but dissuaded by an unlike-minded candidate's messages, and the size of the dissuasion effect is larger than 236 that of the persuasion effect (9). However, the persuasion effect among strong in-partisans and the dissuasion effect among strong out-partisans are 237 unlikely to affect their vote choices and will not contribute to vote shifts. Instead, vote shifts can be induced by a persuasion effect on independents 238 or lean out-partisans and a backfire effect on lean in-partisans. As long as the former trumps the latter, the incumbent's grandstanding is likely to 239 have a positive persuasion effect and increase their vote share in the following election.

240
However, note that these two mechanisms-mobilization and persuasion effects-may work simultaneously. This section tests whether any of 241 these mechanisms are at play and, if so, to what extent. I test the following two hypotheses:

242
Mobilization Hypothesis: As a member grandstands more in any given congress, in-partisan supporters in their district are more likely than 243 independents or out-partisans to be mobilized to turn out in the following election. The data include multiple variables that allow testing various mechanisms. Theoretically, of my particular interest are respondents' validated 256 turnout § § § and whether they said they voted for the incumbent. Given the errors and biases in self-reported vote choices documented in previous 257 literature (11) and especially low response rate for the vote choice question, however, to complement the analysis, I also test the effect of members' 258 grandstanding on respondents' partisan affiliation with the incumbent party. All these three dependent variables are binary indicators. ¶ ¶ ¶ The 259 unit of analysis in this section is an individual respondent, and I look at the over-time changes in each of the respondents.

260
The main explanatory variable is the incumbent's Grandstanding Score, and I control for the following potential confounders. At the 261 respondent-level, I control for age, the level of education, family income, the level of political interest and binary indicators for registered voters 262 and unemployment. 17 At the member-level, I include the Legislative Effectiveness Score (LES), seniority and its squared value, challenger quality, 263 campaign spending, 18 previous vote share, and binary indicators for party leaders and minority members. 19 Then, in the model, respondent fixed 264 effects and congress fixed effects are included. As the dependent variables are binary, I use a fixed-effects logit model (a.k.a. conditional logistic 265 regression) using "clogit" function in "survival" R package (13).
266 Table S7 presents regression results. The first model tests the mobilization hypothesis which proposes that there can be a mobilization effect 267 especially among in-partisans rather than among independents and out-partisans. The Grandstanding Score is interacted with partisan groups 268 with independents as a reference group. The result shows that there is neither a mobilization effect nor a demobilization effect in any group.

269
The second model tests the persuasion hypothesis, that is, whether incumbents' grandstanding tends to motivate non-supporters more likely 270 to vote for the incumbent. However, it finds no evidence of vote switches in any group. Even without the interaction with partisan groups, the 271 politicians' grandstanding tendency does not have any independent effects on voter turnout or vote choices. (See Table S8 for the results.)

272
These results are inconsistent with the member-level findings presented in the main text. The potential reason for this inconsistency might

281
The third model in Table S7 examines the effect of grandstanding on voters' affiliation with the incumbent party. Interestingly, I find that as (15.29%) is higher than the non-response rate among those who did not (13.36%) although the difference between the two groups is not statistically 299 significant based on the two-sided z-test of two-sample proportions. However, also note that the difference between the two groups is over 4%  Table S9. The results show that out-partisans tend to switch their votes to the incumbent as the incumbent grandstands more. The 310 coefficient on the interaction term between the Grandstanding Score and the indicator for out-partisans is positive and highly significant, and 311 the marginal effect of the Grandstanding Score for this group is positive. The analysis further shows no effect of the incumbent's grandstanding 312 on turnout and partisanship. Although the results are slightly different from those shown in Table S7, they corroborate the evidence for the 313 persuasion effect.

314
Then, why would the incumbents' grandstanding induce vote switches among their constituents who did not support the incumbent in the 315 previous election? There can be multiple ways how message contents make some voters switch their votes in favor of the incumbent. For example, 316 the incumbent's policy stance can be adopted by message receivers (7); the incumbent's presentation of self may increase their valence or likability 317 that voters perceive (14, 15); the same may be at work at the party-level rather than at the individual member-level, from the perspective of 318 message politics (16), such that voters may evaluate the incumbent party more highly or favorably as the messages contain partisan messages.

319
However, testing each of these psychological mechanisms is beyond the scope of this analysis.

320
In summary, the findings presented in this section suggests that the increase in the vote share of the grandstanding incumbents can be explained 321 by voters' vote switches to the incumbent party rather than the mobilization effect. This is consistent with Spenkuch and Toniatti's finding that 322 TV campaign advertisements have a persuasion effect but do not affect voters' turnout.(17) 323 § § § I only use validated voters because self-reported turnout is often found untruthful in previous studies (e.g., (10)), and I confirmed that the current dataset suffers from this issue as well. ¶ ¶ ¶ Except for the turnout record, the other three dependent variables were originally ordered variables, but I converted them to binary variables for the following reason. There are neither many ready-made statistical packages that estimates a fixed-effects ordered logit model consistently nor a consensus on the validity of a few new approaches to estimate the estimators. One option recently made available is the "feologit" command in Stata (12), but none of my model specifications were fit due to non-convergence. 17 The coding rules are presented in the next section.
18 First of all, I control for the effect of campaign spending in three different forms depending on the dependent variable. As in the main analysis where the dependent variable was members' vote share, I control for the spending ratio between the incumbent and challenger when the dependent variables are survey respondents' vote choices and partisanship choices because these choices are made in zero-sum competition. However, voters' turnout is likely to increase as the absolute amount of campaign spending from all candidate's increases. Thus, in the turnout model, I control for the spending of both candidates and spending of the incumbent, respectively. 19 The interaction between the indicators for president party members and midterm elections is not included. First, because the president did not change during the period covered in this analysis, due to the absence of variations within respondents, the coefficient is not estimated for this variable. Second, the midterm indicator is same as the fixed effect for the 112th Congress. 20 However, it makes sense that less people would refuse to respond to the vote choice question because only those who turned out to vote can say who they voted for. Still, 4.4% out of those who said they voted did not report their vote choices.  Note: * p<0.1; * * p<0.05; * * * p<0.01 Note: Robust standard errors are reported in parentheses. The coefficients for the following variables were not estimated and thus not reported in the table: Registered voter, vote share (t−1) , challenger quality, campaign spending variables, and indicators for the 112th and 113th Congresses. works in voters' mind but was intended to provide the first analysis using the best possible available observational data. Given the nature of the 325 data and the research design used in this analysis, the findings presented here are necessarily correlational. Future research involving a randomized 326 experiment may be able to further confirm the causal path and extend the investigation of the micro-level mechanisms.

327
In-partisan: 1 for affiliation with the incumbent party; 0 otherwise. 332 Out-partisan: 1 for affiliation with the opposition party; 0 otherwise.

334
Education: 1 for no high school education; 2 for high school graduate; 3 for some college education; 4 for two years of college education; 5 for four 335 years of college education; and 6 for postgraduate education.  Income: Family income in an 18-point scale with 1 for less than $10,000 and 18 for $25,000 or more.