Signaling Through Emphatic Legislative Speech

The starting point for our theoretical account of emphatic legislative speech is that we conceive of speeches as signals. The signaling function of legislative speech is well established in research on political communication (see, for example, Grimmer Reference Grimmer2013; Highton and Rocca Reference Highton and Rocca2005; Hill and Hurley Reference Hill and Hurley2002), yet there is a notable gap in existing accounts of speech signaling. While there is little doubt that legislators are mindful of public preferences when speaking in public, it has often remained unacknowledged how unlikely it is that such signals will reach their intended audience.

The motivating idea for this contribution is that legislators are aware that the vast majority of speech signals will go unnoticed by the public, which is why legislators can and do make efforts to make their contributions more visible. Particularly in the current media environment, legislators can rely on a push strategy by posting their speeches on their websites or on social media. Legislators can also pursue a pull strategy by trying to create ‘broadcastable’ moments, hoping that their contributions get picked up by traditional or social media. In practice, push and pull strategies will go hand in hand in that legislators try to create good soundbites, which they then post on social media in the hopes that their messages might go viral.

In trying to create good soundbites, legislators can rely on a host of rhetorical strategies that have been elaborated since antiquity. While captivating oratory can never be separated from the substance of what is being said, it is never a mere textual matter either. Instead, there are a great number of non-verbal characteristics that distinguish a good speech from a bad one. This might cover physical aspects such as gesturing, pose, and body movement, as well as auditory features such as pace, pitch, and volume.

We expect that legislators deliberately rely on these techniques to improve the odds that their signal reaches its intended audience. Notably, we are not interested in whether a signal is actually perceived but whether legislators predictably vary their speech delivery, suggesting a deliberate use of these techniques for strategic ends. In this sense, we conceptualize speech delivery as a deliberate effort on the part of legislators rather than an unconscious indicator of legislators’ true beliefs. To be sure, conceptualizing speech delivery as deliberate should not be equated with insincere. It is easily conceivable that legislators often feel strongly about an issue, resulting in a forceful delivery. Yet, the notion of deliberate emphasis suggests that, for the most part, professional politicians can choose to hide their true feelings when giving a speech if they consider doing so politically advantageous. With regard to our theoretical interest, then, we assume that legislators can choose to send a signal by way of an emphatic delivery.

There is tentative evidence that emphatic and emotional appeals are in fact more likely to be perceived by the public. Given the difficulty of systematically quantifying the emphasis in political speech (Cochrane et al. Reference Cochrane, Rheault, Godbout, Whyte, Wong and Borwein2022), there is little direct evidence on this question. The most direct study linking speech emphasis and media visibility is presented by Dietrich et al. (Reference Dietrich, Schultz and Jaquith2018). Analyzing audio data from floor speeches in the US House, the authors show that more emotionally charged speeches are more likely to be broadcast and receive media coverage. Beyond the direct evidence, there are various studies on public engagement with political messages, which consistently show that greater emotional intensity predicts the success of political messages on social media (Brady et al. Reference Brady, Wills, Jost, Tucker and Van Bavel2017; Heiss et al. Reference Heiss, Schmuck and Matthes2019; Nave et al. Reference Nave, Shirman and Tenenboim-Weinblatt2018; Peeters et al. Reference Peeters, Opgenhaffen, Kreutz and van Aelst2023), as well as showing that legislators’ rhetorical skills and emotional appeals predict their visibility in traditional media (Amsalem et al. Reference Amsalem, Sheafer, Walgrave and Loewen2017; Lupacheva and Mölder Reference Lupacheva and Mölder2024; Maier and Nai Reference Maier and Nai2020; Sheafer Reference Sheafer2001 Reference Sheafer2008; Sheafer and Wolfsfeld Reference Sheafer and Wolfsfeld2004; Wolfsfeld and Sheafer Reference Wolfsfeld and Sheafer2006).

For a first test of our new measure of legislative speech emphasis, we rely on one of the most well-established context factors in research on legislative behavior – the effect of public preferences. We expect that legislators only choose an emphatic delivery when their preferences are aligned with the preferences of their electorate. Only under conditions of alignment should we expect legislators go out of their way to try to send a signal about where they stand politically.Footnote ¹

In formulating this expectation, we are able to add nuance to research on legislative signaling. Arguably the most politically consequential signal that legislators can send is their plenary vote. While there is ample evidence that public preferences shape legislators’ voting record, legislators often find themselves voting against their district, either because of strongly held personal beliefs or, more commonly, because of pressures from the party leadership, where the pressure to fall in line with the party should be especially strong in the current climate of political polarization. Consequently, while legislators may often find themselves voting against their district, there is little reason to expect legislators to advertise that fact to their electorate by delivering an emphatic speech.

Even though there is little reason to expect legislators to emphatically highlight a vote against the district majority, one might wonder whether legislators with unpopular positions would not be better off not taking the floor at all. While trying to fly under the radar is not an unreasonable strategy, it can also prove dangerous when an unpopular vote comes to light. Therefore, legislators who find themselves voting against their district may prefer to take the floor to explain their vote. At the same time, it would rarely be wise to call attention to an unpopular stance by creating a good soundbite. We can therefore summarize that legislators whose vote is aligned with the preferences of their district are more likely to give an emphatic appeal than legislators with an unpopular stance. As an empirical matter, this expectation also helps distinguish between a deliberate and an unconscious account of speech delivery. If public preferences systematically shape legislators’ speech delivery, this is unlikely to result from legislators’ true beliefs that accidentally shine through in their delivery.

Measuring Emphasis in Legislative Speech Using Automated Video Analysis

To study the emphasis in legislative speech, we analyze video recordings of key debates in the House of Representatives. The video recordings of the debates were compiled from the online archives of the House. The sample contains video recordings of all debates on the twenty-five pieces of legislation. We manually discard irrelevant sequences to ensure that we only analyze footage where the camera fully captures the speaker.Footnote ⁴ This results in seventy-seven hours of video footage comprising 2,341 speeches by 543 legislators. Table A2 in the Online Appendix lists the debates and the pieces of legislation.

We employ computer vision to measure the speech emphasis. Specifically, we generalize manual annotations based on a set of training videos to all videos in the sample. We start by drawing an additional sample of 245 speeches by 116 legislators on 37 bills from the 115th Congress as training/test data. We manually selected 245 speeches rather than choosing them at random to ensure sufficient variation in emphasis across our training and test data. This approach allowed us to incorporate a mix of high-, mid-, and low-emphasis speeches in both datasets. The resulting videos were split into 184 training and 61 test videos. Four trained coders were tasked with annotating the emphasis in the speeches using a seven-point Likert scale, ranging from − 3 (low emphasis) to + 3 (high emphasis), for every non-overlapping two-second segment.Footnote ⁵

Every video was annotated by two randomly selected coders to better judge the emphasis in the videos and to evaluate inter-rater agreement. As continuous annotation of video data is subject to different reaction times and mental processing speeds, annotations can move out of sync. Therefore, we align the annotation sequences by the two coders using the mean absolute error distance. This alignment shifts values by at most two seconds, that is, by one segment. Table 2 summarizes the key values of the manually annotated dataset. On average, the two annotators deviate by less than 0.5 scale points based on the mean absolute error across all two-second segments in the training and test data. Additionally, we report Lin’s concordance correlation coefficient and Pearson’s correlation coefficient as common measures for inter-rater agreement.Footnote ⁶

Table 2.

Summary metrics for the annotated data set and model evaluation

*Note: predictions drawn from a clipped standard normal distribution, 1,000 runs.

Predicting speech emphasis using a convolutional neural network

To estimate emphasis scores for speeches outside the manually annotated set, we use the training data to train a multimodal convolutional neural network using audio and video inputs. The goal of the network is to assign emphasis scores for each two-second segment. As context information is useful for predicting the current emphasis state of a speaker, we include the surrounding two-second segments for the prediction. Thus, the model takes an input of six seconds of audio and video data for each two-second segment prediction. Before feeding the data into the network, we perform a series of preprocessing steps which we describe in Online Appendix C.

To predict the speech emphasis, we employ a convolutional neural network.Footnote ⁷ CNNs typically comprise two stages: feature learning and prediction. In the first stage, feature learning, the convolutional base of the model learns hierarchies of modular patterns in the input data. These features are represented in feature vectors that constitute the output of the convolutional base. In the second stage, prediction, this feature vector is fed into a second neural network, which uses the features from the first stage to predict outcome values, in our case, emphasis scores. If trained on a large enough dataset, features learned in the convolutional base are sufficiently generic to be useful for a wide variety of classification tasks. Therefore, especially in cases with small training data, pre-trained networks are commonly used for feature extraction and have proven to be highly effective (Carreira and Zisserman Reference Carreira and Zisserman2017; Chollet and Allaire Reference Chollet and Allaire2018).

As our manual training data is limited, we use a pre-trained network to extract features for the video input. Specifically, we use a state-of-the art pseudo-3D-Resnet CNN (Qiu et al. Reference Qiu, Yao and Mei2017). This network is pre-trained on the Kinetics data set, which is commonly used for human action recognition (Kay et al. Reference Kay, Carreira, Simonyan, Zhang, Hillier, Vijayanarasimhan, Viola, Green, Back, Natsev, Suleyman and Zisserman2017). The network takes 299 × 299 pixel images as input and generates a 2,048-dimensional feature vector. For the audio input, we use a Soundnet-like subnetwork (Aytar et al. Reference Aytar, Vondrick and Torralba2016). This network takes the audio input and generates a 512-dimensional feature vector.

After passing the video and audio inputs through these two networks in the feature learning stage, we obtain two feature vectors that summarize the video and audio inputs. In the next step, we combine both vectors and pass them to two fully connected layers. The final layer produces values in the [ − 1, + 1] range. To match the output to the original emphasis scale, we multiply the predicted values by 3 to obtain scores ranging from − 3 to + 3. Figure 1 visualizes the network architecture.

Figure 1.

Convolutional neural network architecture.

To train the model, we use a mean absolute error loss function. This function minimizes the distance between values predicted by the model and the mean emphasis scores provided by the human annotators. To prevent the neural network from overfitting, we add dropout to the fully connected layers in the second stage (Srivastava et al. Reference Srivastava, Hinton, Krizhevsky, Sutskever and Salakhutdinov2014).Footnote ⁸ We use the Adam algorithm to optimize the model’s parameters (Kingma and Ba Reference Kingma and Ba2014).

Applying the model to footage of key vote debates

We apply the trained model to the video footage from all debates in our sample. The network predicts emphasis scores for each two-second sequence. For example, a one-minute speech contains thirty consecutive emphasis scores. To generate one emphasis score per speech, it is necessary to aggregate the individual scores. The simplest approach would be to calculate the average emphasis of each speech. Such an approach would ignore that speeches differ in length. This means that a multi-minute speech with thirty seconds of intense delivery would score lower than a one-minute speech with the same sequence. In line with the argument that legislators attempt to signal their issue positions by giving passionate speeches in the hopes of being amplified by traditional or social media, it is sensible to focus on shorter sequences within speeches. Legislators are aware that only short sequences of their speeches may be picked up and broadcast to the public. Therefore, it is sufficient to deliver a short but high-intensity appeal as part of a longer speech, such that short and long speeches with the same high-intensity sequence should score the same. Consequently, we select the thirty-second sequence with the highest within-speech average emphasis to score the speeches.Footnote ⁹ For the same reason, if a legislator delivered more than one speech on a bill, we select the speech with the highest emphasis score for the analysis. We explain this aggregation procedure in more detail in Online Appendix D.

Table 1 provides summary statistics for the resulting data. The number of legislators who delivered speeches on a bill ranges from 13 to 231. Mean emphasis scores range from − 0.11 (Kate’s Law) to 0.72 (End Don’t Ask Don’t Tell Act). Figure 2 provides additional information on the distribution of the emphasis scores across all debates, which range from − 1.5 to 2.2. Thus, the distribution does not reach the extremes of the emphasis scale, running from − 3 to + 3. This is unsurprising as we average over 30-second segments. The distribution can be characterized as approximately normal with a mean of 0.29 and a standard deviation of 0.70.

Figure 2.

Visualization of the emphasis scores and their distribution.

Note: Each line in the upper panel depicts the estimated speech emphasis over the course of the thirty-second sequences. The three highlighted sequences depict the emphasis scores of the speeches with the highest and lowest average emphasis scores (by Rosa L. DeLauro and John Conyers, Jr.) and the speech with the highest within-speech variance (by John Lewis). The video frames give an impression of how increased levels of gesturing and facial expression are linked to higher estimated emphasis scores. The density curve on the right depicts the distribution of the average emphasis scores as used in the analysis. The two boxplots represent the distributions of average emphasis scores by Democrats (D) and Republicans (R).

Model evaluation

We now turn to the evaluation of the neural network. We present the results of two validation exercises. First, we apply the trained model to the held-out test set of sixty-one videos and compare the model predictions with the human annotations throughout all two-second segments within those speeches. In addition, we apply our aggregation algorithm to all of these speeches using both the model predictions and the human annotations and compare the results. Table 2 shows the results of this comparison, along with the results based on random guessing (drawing values from a clipped standard normal distribution) and zero guessing (predicting an emphasis score of 0). As evaluation criteria, we compute mean absolute errors (MAE), Lin’s concordance correlation coefficient (CCC), and Pearson’s correlation coefficient (PCC).

Unsurprisingly, the correlations are essentially zero under random guessing. For zero guessing, the correlation is defined as zero as a constant cannot correlate with a variable. For both random and zero guessing, we observe MAE values close to one standard deviation of the underlying label distribution. The neural network achieves considerably lower MAE values. Based on the MAE metric, the machine prediction is 0.552 scale points off from the human annotation when considering two-second segments. This figure is close to the human interrater MAE. For thirty-second aggregates, the MAE decreases further to 0.460, suggesting that aggregation helps average out random noise in the data. Unlike the guessed values, the model predictions show high correlations for both the CCC and the PCC metric. As before, the correlation between the machine prediction and the human annotators is in the same range as the correlation between the human annotators. We thus conclude that the neural network reliably predicts the speech emphasis.

Our second validation is based on coders’ ratings of thirty- rather than two-second segments. We generated 150 speech pairs based on stratified samples from all thirty-second sequences we used in the analysis.Footnote ¹⁰ For each pair, we asked two coders to indicate which of the speakers displayed a higher level of emphasis, and compared their ratings with the model ratings. Coder and model ratings are considered aligned when the speech identified as more emphatic by the coder receives a higher emphasis score from the model. If the model assigns nearly identical scores to both speeches, we would expect coder ratings to align with the model predictions in around 50 per cent of the cases. As the difference in model emphasis scores increases, we expect coder ratings to be more likely to align with the model predictions.

Naturally, the two coders do not always agree on which of two speeches is more emphatic. Overall, they disagree on twenty-three of the 150 speech pairs. Panel A of Figure 3 plots the probability of agreement between the coders against the difference of predicted emphasis scores by the model. It shows that the probability of the two coders agreeing on the emphasis ordering of a speech pair increases as the speeches are rated more distinct by the model.

Panel B focuses on speech pairs where both coders agree and compares their ratings to the model predictions. As expected, coder and model ratings almost always align when the model predicts a substantial difference in emphasis between the two speeches. However, when the model predicts smaller differences in emphasis, coder ratings are more likely to diverge from the model predictions.

Panel C addresses the fact that the model is trained on speeches from the 115th legislative term, while the analysis includes speeches from the 111th to 115th term. These terms include speakers who were not in the training data, and speaking styles may have evolved, making it more difficult for the model to assess speeches from before the 115th term. To assess whether this is indeed the case, Panel C differentiates between speech pairs that include at least one speech from the 115th legislative term, and that are based on speeches from earlier terms. Indeed, the probability of alignment between coder and model ratings decreases slightly for pairs based on speeches prior to the 115th House. To address this, we include a dummy variable in our analyses, indicating whether a speech is from the 115th term or before.

Figure 3.

Model evaluation based on pairwise comparisons.

Note: Predicted probabilities and confidence intervals are based on bivariate logistic models, regressing agreement on the difference in predicted emphasis between two speeches. Panel A shows the predicted probability that the two coders agree on which speaker displays greater emphasis. Agreement increases as the model predicts less similar emphasis levels between the two speeches. Panel B is based on pairs where both coders agree on the more emphatic speech, displaying the predicted probability that their ratings align with the model predictions. Agreement between coders and the model increases as the model identifies greater differences in emphasis between the speeches. Panel C distinguishes between pairs that include at least one speech from the 115th legislative term and those that do not. Disagreement between the model and coder ratings is more likely for pairs without speeches from the 115th legislative term.

While the model demonstrates satisfactory performance in the validation exercises, it is not without error. This raises the question of when and why the model makes incorrect predictions. To explore this, we conducted additional quantitative and qualitative analyses of the variation in prediction errors across speeches in our test dataset. We summarize the key insights from this analysis below and provide further details in Appendix E.

First, the model rarely predicts values above + 2 or below − 2, suggesting it may be subject to some attenuation bias. Second, our assessment of the speeches with the highest average prediction error suggests that, compared to human annotators, the model is less sensitive to hand movements occurring in front of the body of the speaker and to gestures made with a closed rather than an open hand. In contrast, the model is more responsive to large, clearly visible hand movements. This observation is plausible as open hands are better visible than closed hands, and gestures stand out more clearly against the background when positioned beside the body of the speaker. Third, we observed that the model tends to overestimate the emphasis of speakers who naturally speak with a strong voice. This is understandable, as a naturally strong voice can resemble an emphatic one. The tendency may correlate with the gender of the speaker, which we account for by controlling for gender in the analysis.

Estimating District-Level Bill Preferences

The key independent variable is the extent to which legislators’ votes align with constituency preferences. We draw on survey data from multiple waves of the CCES to estimate district-level preferences. Each survey contains multiple questions on specific bills. Respondents are provided with the title and a short summary of the bill and are asked how they would have voted.Footnote ¹¹ Matching each bill in our sample to a CCES item enables us to estimate district preferences towards the bills.Footnote ¹²

Despite the large sample size of the CCES, it is not designed to be representative of congressional district populations. Thus, simply disaggregating survey answers to estimate district-level preferences would likely yield biased estimates. To overcome this challenge, we rely on the widely used multilevel regression and post-stratification (MrP) for estimating district preferences (Gelman and Little Reference Gelman and Little1997; Lax and Phillips Reference Lax and Phillips2009; Warshaw and Rodden Reference Warshaw and Rodden2012). To assess the reliability of the estimates, we complement the MrP approach with Bayesian regression trees and post-stratification (BARP) (Bisbee Reference Bisbee2019). We employ MrP and BARP to estimate district-level preferences for the twenty-five bills in our sample. The estimates range from zero to one, where high values indicate high levels of support.Footnote ¹³

In the next step, we match these estimates to the representatives’ voting records on the twenty-five bills. Substantively, we are interested in the extent to which legislators’ votes align with the preferences of their electorate. To compute an alignment score, we code roll call votes as one for legislators who voted in favor of a bill and zero for those who voted against it. To assess the extent to which legislators’ votes align with the preferences of their electorate, we calculate the absolute difference between legislators’ votes (YES=1, NO=0) and the preferences of their districts and subtract this from the maximum distance, 1. We label this variable Vote–District Alignment. Values close to 1 indicate high levels of agreement between legislators and their districts, values close to 0 indicate low levels of agreement. Formally:

(1) $$ {\rm VOTE-DISTRICT\ ALIGNMENT} = 1 - |{\rm YES\ VOTE} - {\rm DISTRICT\ PREFERENCE}| $$

We present the distributions of the Vote–District Alignment variable by bill and vote choice in Figure 4. The bright density curves depict the distributions of the Vote–District Alignment for legislators who voted yes, the dark densities depict the distributions for legislators who voted no. For most bills, MrP and BARP result in similar distributions of the alignment between legislator vote and district preference. In most instances, the distribution for legislators who vote yes diverge from those who vote no. Consider the State Children’s Health Insurance Act (HR 2, 111th) as an example. Although there is variation between the districts, almost all districts were fairly supportive of the bill. Thus, the Vote–District Alignment scores for legislators who voted for the bill are significantly higher than the values for legislators who voted against the bill. This also means that we observe numerous instances where legislators’ votes were misaligned with their districts. Arguably, this finding can be traced back in no small part to the high levels of polarization and the resulting pressure to vote along party lines, such that legislators often find themselves between a rock and a hard place, where they either have to vote against the preferences of their constituents or risk upsetting the party leadership.

Figure 4.

Distributions of alignment between legislator votes and district preferences.

Individual Versus Macro-Level Explanation

Having shown evidence for a link between district opinion and speech emphasis among legislators who rise in opposition, we now proceed to test whether the proposed individual-level explanation is driving the findings. It is conceivable that legislators whose preferences are more consistently aligned with those of their constituents might be more likely to signal this alignment to their constituents by giving more emphatic soundbites overall. We are thus interested in differentiating between such an explanation at the legislator level from an explanation where legislators are more emphatic when their position is aligned with their district in a particular debate.

We adjust the statistical model to study the isolated effects of both explanations. To that end, we partition the total variation of vote–district alignment into two parts: variation within districts and variation between districts. Within vote–district alignment is defined as the deviation of the vote-alignment on a specific bill from legislators’ average vote-alignment, which reflects the ‘legislator-in-debate’ level explanation. Between vote–district alignment is defined as legislators’ average vote-alignment, which reflects the generic legislator level explanation.

We specify a within–between random effects (REWB) model (Bell et al. Reference Bell, Fairbrother and Jones2019; Bell and Jones Reference Bell and Jones2015) to estimate the effects of both variance components in the same model. The model is specified as follows:

(2) $$ y_{it} = \beta _{0} + \beta _{1W}(x_{it} - \bar x_{i}) + \beta _{2B} \bar x_{i} + \sum _{j=1}^{k} \gamma _{j} z_{ji} + (\upsilon _{i} + \upsilon _{t} + \epsilon _{it}) $$

where y _it represents legislator i’s emphasis in debate t on a specific bill. x _it is the alignment between legislator i’s vote after debate t and the preference of their district. $\bar {{x_i}} )$ is the average alignment between legislator i’s votes and the preference of their district. υ _i are random intercepts for legislator i and υ _t are random intercepts for debate t. z _ji represent the same k individual-level control variables as before.

The model has several properties worth noting. Importantly, the independent variable of interest, vote-alignment, enters the model in two forms: First, in its de-meaned form $({x_{it}} - \bar {{x_i}} )$ , that is, the deviation of the vote-alignment from legislators’ average vote-alignment. The coefficient β _1W represents the average within effect of vote-alignment, that is, the expected change in a legislator’s speech emphasis caused by variation of preferences within their district. Thus, positive values of β _1W would constitute evidence for the individual-level explanation, making β _1W the main coefficient of interest. The coefficient is equivalent to individual fixed effects and is independent of differences in legislators’ delivery styles. Second, the model incorporates legislators’ average vote-alignment ( $\bar {{x_i}} $ ) as a covariate. The coefficient β _2B represents the average between effect of vote-alignment. This effect captures differences in emphasis between legislators with varying average levels of vote-alignment, that is, legislators with more or less average district support for their floor votes. Thus, positive values of β _2B would constitute evidence for the macro-level explanation.

The results in Table 4 provide support for both the individual and the macro-level mechanism. Starting with the individual mechanism, models (1) and (2) show that there is a positive within effect of vote–district alignment on speech emphasis. This result indicates that legislators who rise in opposition vary their delivery depending on how closely their vote is aligned with the preference of their district. Specifically, legislators who rise in opposition deliver more emphatic speeches as their districts become increasingly hostile to the bill. Figure 5 helps to assess the size of this effect. Consider a legislator who rises in opposition in two debates on different bills. In the first debate, 50 per cent of the voters in their district are – like them – opposed to the bill. In the second debate, 75 per cent are opposed, making their vote more aligned with public opinion in their district.Footnote ¹⁷ Based on the estimates from model (1), we would expect the legislator to deliver a speech that scores 0.19 points higher on the emphasis scale during the second debate compared to a speech during the first debate. This is equivalent to 0.45 standard deviations of the de-meaned speech emphasis.

Table 4.

Within–between multilevel specifications with legislator and debate random effects. Parentheses report heteroskedasticity consistent wild bootstrap standard errors (Modugno and Giannerini, Reference Modugno and Giannerini2015; Loy, et al. Reference Loy, Steele and Korobova2023)

Note: The dependent variable is the level of emphasis of a legislative speech. Parentheses show wild bootstrap standard errors.

Figure 5.

First differences and 95 per cent confidence intervals illustrating the expected change of speech emphasis in response to increased alignment between a legislator’s No vote and public opinion in the district.

Note: Wild cluster bootstrap confidence intervals based on model (1) and model (2) in Table 4. The baseline value of the de-meaned district alignment is set to − 0.28 (minimum for the MrP estimates), the mean level of vote-alignment is set to the empirical mean (0.53 for MrP, 0.51 for BARP), Republican is set to zero, vote with party is set to 1, seniority is set to its mean (15.7), gender is set to zero, ideology is set to the empirical mean (0.44).

The results also provide evidence for the macro-level mechanism. Both models show a positive between effect of vote–district alignment on speech emphasis (1.28 and 1.16 depending on the public opinion estimate). This indicates that legislators whose opposing vote constantly shows high alignment with their electorate tend to deliver more emphatic speeches compared to legislators with less support for their votes. To assess the substantive meaning of this effect, consider two otherwise similar legislators whose voters differ in their attitudes towards key pieces of legislation, where legislator A enjoys high alignment between her votes and public opinion and B does not. Suppose that the difference in vote alignment between legislators A and B amounts to 25 percentage points on average.Footnote ¹⁸ Model (1) predicts that this difference has implications for how legislators A and B present themselves on the floor: On average, legislator A would deliver speeches that score 0.32 points higher on the emphasis scale compared to legislator B. This amounts to 0.55 standard deviations of the mean emphasis score.

Taken together, the results from the REWB model on legislators who rise in opposition provide evidence for both the individual-level and the macro-level mechanism. Regarding the individual-level mechanism, legislators deliver more emphatic speeches as their vote becomes more aligned with their district. At the same time, legislators whose votes constantly show high alignment with their districts deliver more emphatic speeches on average.

The results from models (3) and (4) echo the findings from the models in Table 3. They show that this finding does not generalize to legislators who rise in support of a bill. The magnitudes of the estimated within and between estimates are not distinguishable from zero at conventional levels of statistical significance.