Visible home styles in Congress

Transcription

1 Visible home styles in Congress L. Jason Anastasopoulos Dhruvil Badani Crystal Lee Shiry Ginosar Jake Ryland Williams July 17, 2017 Abstract While members of Congress routinely communicate with constituents using images, there is little systematic knowledge about how images are used as a means of strategic communication due to computational limitations. New developments in computer vision, however, are bringing the systematic study of images within reach. We develop a framework for understanding visual political communication by extending Fenno s analysis of home style to images and then apply this framework to study racial representation in 19,000 images collected from MCs Facebook profiles using a machine learning technique known as convolutional neural networks. We demonstrate that Democrats and Republicans in the House of Representatives strategically use the race of individuals that they pose with for political ends. When compared with their district demographics, Democrats tend to over-represent African-Americans in photos that they post on social media while Republican House members tend to under-represent African-Americans in their photos. Keywords: Congress, home style, convolutional neural network, deep learning. (Corresponding Author); Microsoft Visiting Professor, Center for Information Technology Policy, School of Engineering & Woodrow Wilson School of Public and International Affairs, Princeton University, Princeton, NJ, USA; ljanastas@princeton.edu; Tel:

2 Introduction Before becoming president in 1963 after John F. Kennedy s assassination, Lyndon Johnson spent 12 years as a Democratic representative of Texas 10th Congressional district and another 12 years representing Texas as a Senator. During his time in Congress, he voted against practically every piece of pro-civil rights legislation (Caro 1990, 2002) yet became known as one of America s foremost civil rights champions shortly after signing the landmark Civil Rights Act of 1964 into law (Kotz 2005). Figure 1 Johnson meets with Martin Luther King and other civil rights leaders in the oval office. Source: Lyndon Baines Johnson Library and Museum. Photo by Yoichi Okamoto. Johnson s transformation from civil rights opponent to civil rights leader in such as short period of time was due in no small part to his skillful abilities to manipulate his public image through photographs (Altschuler 1986; 2

3 Figure 2 President Johnson meets with civil rights activists John Lewis and James Farmer.Source: Photo by Yoichi Okamoto. Zarefsky 2004; Duganne 2013). Johnson made it a point to invite photographers into the oval office and elsewhere to take pictures of meetings with African-American civil rights leaders and activists such as Martin Luther King (Figure 1) and others (Figure 2). He was the first president to appoint an official White House photographer, Yoichi Okamoto, now known as the godfather of White House photography, who served under him from 1963 to 1969, snapping now iconic shots of Johnson conducting business in the White House (Estrin 2013). Okamoto s photos are considered masterful not only because of their aesthetic qualities such as lighting, angles, etc., but also because they are believed to have perfected a visual persuasion technique known as juxtaposition, which simply involves placing an individual or object next to another 3

4 object, person or group in a photograph with the goal of creating associations between political figures with something that an object, person or group represents. By consistently posing Johnson with civil rights leaders in his shots, Okamoto sought to link Johnson with the civil rights movements in the minds of viewers. While photographs of political figures are replete with intentional and unintentional juxtapositions, there is little empirical evidence demonstrating how, or if, this technique and others used in photography affect how political figures are perceived by the public. Developing an empirical basis for measuring the political uses and effects of visual persuasion techniques such as juxtaposition is crucial to understanding key aspects of the photo laden modern day political landscape. Non verbal cues contained in political images and photographs have been found to affect voting behavior and other politically relevant attitudes and have demonstrated that they often have potent effects on how politicians are perceived (Rosenberg et al. 1986; Lau and Redlawsk 2001; Campbell, Green, and Layman 2011; Lenz and Lawson 2011; Arceneaux 2012). Facial features of candidates, body language, clothing and the context that they tend to appear have all been found to influence how political figures are perceived by voters (Hayes, Lawless, and Baitinger 2014). With the explosion of social media as a means of constituency communication and improvements in camera technology, text posts combined with photographs and images containing thousands of non verbal cues have be- 4

5 come a routine means of daily communication between members of Congress, their constituents and the public at large through a number of online platforms such as Twitter, Facebook and Instagram and MCs own.gov web pages hosted on U.S. government servers (Esterling, Lazer, and Neblo 2012; Barberá 2015; Bond and Messing 2015; Accetti and Wolkenstein 2017; Ryan 2017a). In the United States, juxtapositions of political figures with members of constituency groups are particularly relevant to understanding present day home styles of members of the U.S. House of Representatives and Senate (Fenno 1978) because they can help us understand how members of Congress utilize visual media to convey key elements of MCs home styles. In this study we develop a framework for understanding political communication through visual media such as photos and images by extending Fenno s analysis of home style (Fenno 1978) and introduce the concept of a visible home style and study the visible home styles (Fenno 1978) of House and Senate members by comparing racial demographics of the photographs that they post on Facebook with the racial demographics of their respective constituency groups 1. To accomplish this, we analyzed approximately 19,000 photographs using a deep learning technique known as convolutional neural networks (CNNs) to identify the race of individuals that House members pose 1 In addition to this we demonstrate that image features identified in our framework causally affect how politicians are perceived by the public on dimensions related to home style using an image manipulation experiment. 5

6 with in the photographs that they post on their official Facebook profiles 2. 1 A Framework for Political Image Analysis A brief exploration of the literature on the political uses of images suggests that what distinguishes a political image from other images is that political images are usually created with the intention of persuading viewers to side with a political candidate, party or cause (Hutchings, Walton Jr, and Benjamin 2010; Dilliplane, Goldman, and Mutz 2013; Prior 2013; Baker 2015). Starting with politicians use of images, the focus of our empirical study below, we premise our framework on the Mayhewian electoral connection assumption that the behavior of democratically elected politicians can be largely explained by their desire to get re-elected (Mayhew 1974; Giger and Klüver 2016; Lopes da Fonseca 2017; Ryan 2017b). Under this premise, images are used by politicians to increase their likelihood of re-election or otherwise advance their careers. The primary goal of political image analysis, then, is to understand how politicians use photographs to fulfill their reelection goals or advance their careers. 2 The photographs were sub sampled from a database that we collected of 192,000 photographs from the Facebook profiles of 230 members of the House of Representatives and 52 members of the Senate. Photographs collected were based entirely on availability from Facebook at the time that the photographs were collected. 6

7 Figure 3 Photos from Speaker Paul Ryan s Facebook profile showing him with a flag in the background (left) and speaking to military personnel (right). Source: On this front, Fenno s analysis of legislator home style (Fenno 1978) provides guidance. According to Fenno, home style is an idiosyncratic set of behaviors legislators adopt as a means of fostering trust among their constituents for the purpose of increasing their chances re-election. Legislator home style is comprised of three main elements: (1) how she allocates resources; (2) how she presents herself to others (presentation of self) and (3) how she explains what she is doing when outside of her district (Washington activities). As we describe in more detail below, photographs provide an excellent means of communicating presentation of self and Washington activities to constituents. For Republican House members, this might involve posting photographs on social media posing with American flags in the background or addressing 7

8 Figure 4 Pictures from minority leader Nancy Pelosi s Facebook profile showing her posing next to union members on strike in her district (left) and speaking at a gun control rally (right). Source: military personnel (Figure 3). For Democratic House members, this might involve posing with union members in their districts or giving speeches on gun control as in Figure 4. With this in mind, we first identify and discuss three essential features of photos which can shape how politicians are perceived: objects, people and poses. 8

9 Figure 5 Former Texas Governor Rick Perry posing with rifles in a gun shop. Source Objects Objects and symbols in photos contain political meaning when juxtaposed with a political figure or when they are displayed by themselves. Objects tend to contain abstract information about the personal qualities of politicians or their policy positions. For example, the Democratic donkey and Republican elephant are recognizable by many as representing American political parties but also stand for the ideas and policy positions taken by each party. The American flag, an object and a symbol, represents the geopolitical construct of the United States but depending on context can represent American values and ideals, American patriotism, American military power or a combination of all three. Objects such as guns and military equipment can be used to represent specific policy stances such as support or opposition to gun control, support for veterans, military interventions and so on. Thus, 9

10 a politician might convey opposition to gun control and/or support for the Second Amendment by posing in a gun shop as former Republican Texas Governor Rick Perry does in Figure People Figure 6 New Jersey governor Chris Christie embraces President Obama during his visit to the state after Hurricane Sandy in Source Other people in photographs where a politician is the subject are politically relevant to the extent that they can convey information about the politician to their constituents. This can be accomplished through posing with a well-known person (celebrity, political figure, etc.) or unknown people. When posing with unknown individuals, visible features (race, gender, 10

11 veteran status, age etc.) of these individuals can shape how the politician is perceived. For example, a white politician who consistently poses with people from other racial groups may be trying to signal that they identify or empathize with members of that group or that their policy positions will benefit members of that group. In our image experiment below (see Appendix A), we demonstrate that the race of individuals that politicians pose with affects how they are perceived by politicians in terms of partisanship/party identification and identification and empathy with minority groups. Regarding famous individuals, the information conveyed depends largely upon how the famous individual is perceived by the viewer. For a recent example, Figure 6 contains a photo of Republican New Jersey governor Chris Christie embracing President Obama after a visit to assess some of the damage caused by Hurricane Sandy in While right-leaning media sources such as Breitbart excoriated Christie, mainstream sources praised Christie s bi-partisanship 3. 3 The New York Times, for example, covered the meeting between Obama and Christie with a story whose headline read No Partisan Fire at the Shore: An Obama-Christie Reunion. Source: The New York Times, r=0 11

12 1.3 Poses Poses, which include actions and facial expressions, are among the most well studied aspects of visual persuasion. In a recent study by (Joo et al. 2014), for example, the authors find that certain facial expressions and actions convey both positive and negative politically relevant information on a number of dimensions. Facial displays such as smiles and frowns can convey satisfaction or sadness with a certain political event or outcome. Gestures such as finger pointing, hand waving, hand shaking, hugging and so on can demonstrate different types of engagement with constituents and other political leaders. 2 Visible Home Styles Visible home styles refer to how the politically relevant features of visual media discussed above (objects, people and poses). Figure 7 Image attributes which reflect elements of photographic home style. Blue = Reflects home style element; White = Does not reflect home style element. Figure 7 breaks down each aspect of home style and identifies which fea- 12

13 tures of images correspond to each aspect. Our theory of visible home style suggests that each image feature reflects clusters of home style aspects. Objects, for example, can represent identification with constituents and Washington activities, but are unlikely to represent empathy and qualification. Other people in photographs, the focus of our empirical study below, can potentially convey empathy or identification with constituent groups. For example, photographs in which members of Congress pose with veterans or members of minority groups may communicate to constituents and others that the MC cares about members of that group, empathizes with them or at the very least spends time with them. Demonstrations of identification and empathy through images convey the message to constituents that I am one of you and I understand you and think like you do. (Fenno 1978; Fiorina and Rohde 1991; Harden 2013; Parker and Goodman 2013; Desmarais, La Raja, and Kowal 2015; Miller, Saunders, and Farhart 2016) Photos which show MCs participating in activities that their constituents enjoy such as hunting, attending sporting events, performing manual labor etc. are one means of accomplishing this. Another means is through posing with members of constituency groups that are likely to vote for the MC in the next election. If this were true, we would expect Republicans to strategically post pictures with members of groups that tend to vote Republican, such as veterans as we see with House Speaker Paul Ryan in Figure 8. Similarly, Democrats would be more likely to do so by posting pictures 13

14 Figure 8 House Speaker Paul Ryan Poses with a Vietnam Veteran. Source: of themselves with members of minority groups such as African-Americans and Hispanics, members of labor unions and so forth. Rep. Louise Slaughter (D-NY 25) provides an good example of this in a photo collage posted on her Facebook profile page in Figure 9 where she meets with with African- American, Hispanic and white members of her district. As we demonstrate with an empirical study below, we find strong evidence of an electoral connection with visible home styles. While Democratic House members tended to over represent African Americans in the photos that they post on Facebook, Republican House members, especially those in in Southern states, tended to under represent African Americans in the photos that they post. We argue that the former suggests that Democratic 14

15 Figure 9 A photo collage posted on Rep. Louise Slaughter s (D-NY 25) Facebook profile where she is posing with African- American, Hispanic and white constituents at a festival in her district. Source: House members use photos to elicit identification and empathy among minority constituents 4 while the latter suggests that Republican House members, especially those in districts which have historically high levels of racial resentment, do the same among white constituents by excluding photos with African Americans. 4 See Appendix A for a description and analysis of an experiment which provides causal evidence for this claim. 15

16 Stage Definition (1) Acquisition Posing w/ other individuals at events and having someone else (a staffer, other people at events etc.) take a photo. (2) Selection Of the photos taken, selecting those to post for the purpose of furthering the politician s career goals. Table 1 Stages of photo selection for display on MCs Facebook profile photos. At each stage, electoral/strategic considerations can come into play. 3 Exploring visual home styles with convolutional neural networks Here we seek to explore partisan variation in the use of racial characteristics of individuals that members of Congress pose with in the photographs that they post. Specifically, we are interested in understanding how images may be used strategically to elicit identification and empathy among constituents, something that we have discovered causal evidence for in an experiment discussed in Appendix A. To accomplish this, we first discuss how images could be used by MCs to achieve these ends and conduct a deep learning analysis of MCs Facebook images using convolutional neural networks as a means providing evidence for the strategic use of images by Democrats and Republicans in the House of Representatives. 3.1 The strategic use of images by members of Congress Unfortunately, there is little guidance as to what constitutes the strategic use of images in a given context on the internet so we define this here in the 16

17 context of politicians posting photos with other individuals. In defining the strategic use of images, we must consider the steps required of the politician to post a picture posing with other individuals. These include: (1) acquisition posing with other individuals and having someone else (a staffer, other people at events etc.) take the shot and; (2) selection selecting which photographs to post for the purpose of furthering the politician s career goals. At each step, electoral/strategic motivations may be involved. For example, in an effort to appeal to minority constituents, Democrats may choose to pose more frequently with members of specific minority groups that they are targeting for the next election and may also choose to post a higher proportion of photographs with members of those minority groups. Republicans, on the other hand, may choose to pose more frequently with veteran constituents that they are targeting, as we have seen with Paul Ryan s photo above. We focus on the selection stage as a means of understanding the strategic use of images among Democrats and Republicans in the House of Representatives because determining strategic behavior during the acquisition phase requires on the ground knowledge that is not readily available. We first discuss our method of detecting strategic posting and the hypotheses that we test using the image data that we have acquired. As we have discovered in our experiment above, group characteristics such as race tend to elicit identification and empathy among our survey takers. Because of this, we have decided to focus on the race of individuals that politicians pose with and explore whether members of Congress use the race 17

18 of the people that that pose with in a strategic manner. Model Expectations Strategic Implication Null E(θ R,D ) E(θ R,I ) None. Strategic E(θ R,D ) < E(θ R,I ) Appeal to constituents of racial group R. Table 2 Expectations for the % of each race in the MCs district (E(θ R,D )) vs. the % of each race in the Facebook image set (E(θ R,I )). Under random posting, we would expect that the % of members of each racial group would be roughly the same in the MCs district and in thier Facebook profile image set. Under strategic posting, we would expect that the MC would purposefully over represent a racial/ethnic group that they are targeting. In order to determine whether images are used strategically, we first define a null hypothesis that we can test the image data the we collected against. Imagine a scenario in which a member of Congress randomly took photographs with individuals in her district (acquisition) and also randomly posted them (selection). When we collect information about the race of the individuals that these members of Congress posed with, we would expect to find that the % white, black, hispanic etc. that are estimated from the photographs, represented by E(θ R,I ), would be roughly equivalent to the population % of whites, blacks and hispanics in the district, represented by E(θ R,D ), simply as a consequence of the Central Limit Theorem. We define this scenario as the Null model which is what we would expect in an absence of strategy at either the posting or the acquisition stage or both. Deviations from the null model, then provide evidence of strategy. For example, if a member of Congress represents a district in which roughly 25% of the population is white, but 90% of the individuals in their photographs 18

19 are white, we would argue that this provides evidence of strategic posting. Of course, it is not necessarily true that the strategy is causally linked to race. The hypothetical member of Congress mentioned above, for example, might tend to post images of corporate campaign donors who tend to overwhelmingly be white in an effort to collect more donations from certain groups. Either way, deviations from the null model will present us with some evidence that strategic decisions are being made on that individuals social media site. Given these definitions and our reliance on the electoral connection model of MCs behavior, we test three hypotheses using a deep learning technique known as convolutional neural networks (CNNs) to conduct an empirical analysis of visual home styles of members of the House and Senate. Specifically, we are interested in exploring the following hypotheses related to racial representation and visual home styles: H1: Because Republican members of the House of Representatives generally do not rely on votes from minority constituents, they will not make efforts to signal identification and empathy with members of minority groups by posing with them in photos and/or posting these photos on social media. H2: Conversely, because Democrats tend to rely more heavily on votes from minority groups (Barreto, Segura, and Woods 2004; Cameron, Epstein, and O halloran 1996; DeNardo 1980), they will make efforts 19

20 to signal identification and empathy with minority groups by posing with them in photos and/or posting them on social media. H3: As a consequence of H1 and H2, we would expect African Americans to be overrepresented (% black in photos greater than the % black in the district) in the photographs of white Republican House members and under represented (% black in photos greater than the % black in the district or state) in the photographs white Democratic House members 5 To test these hypotheses, we collected approximately 192,000 Facebook profile photographs from 230 members of the 114th House of Representatives and 52 members of the Senate 6. We then use a deep learning method known as convolutional neural networks (CNNs) to identify the race of individuals that members of Congress pose with from a sub sample of approximately 5 We focus on African Americans for a number of reasons. First, African Americans comprise a very important voting bloc, especially for Democratic House members. Second, they are the minority group with the highest classification accuracy according to our trained convolutional neural network classifier. 6 These were the members of Congress available to us at the time that we acquired the photographs. See Appendix B for details about photo acquisition and analysis. For each member of the House and Senate, we acquired the entire set of photos from their Facebook profiles. 20

21 19,000 photographs. Finally, we calculate the racial demographics within MCs Facebook profiles and compare them to district demographics from American Community Survey 5-year estimates. Our data collection process is described in Appendix A. 3.2 Methods Before describing how we identified the race of individuals that members of Congress pose with from their photos, we provide some background on image analysis and convolutional neural networks Image as Data Figure 10 Images of Nancy Pelosi as represented on a machine by a matrix of pixel intensities. 21

22 When we view an image on a computer screen, we are seeing a collection of pixels stacked in a certain order. Grayscale (black and white) images are represented on a machine as a single matrix of pixel intensity values which represent the brightness of the image. The most common format for pixel intensity values is the byte image which is stored as an 8-bit integer that takes on a range of integer values between 0 and 255. Color images are also stored as pixel intensity values, but instead of containing pixels values across a light/dark dimension, color images contain pixel intensity values three dimensions, or channels: red, green and blue. Pixel intensity values for both grayscale and color images are thus represented as a one or three matrices of pixel intensity values in the range of 0 to 255, respectively. Because images, both color and grayscale, are represented on machines as either one or three matrices, each image in a machine is typically represented as a series of tensors or arrays of multidimensional arrays. A convolutional neural network is supervised machine learning technique, or put more simply, a model, that uses pixel location and intensity values as data for the purpose of assigning class probabilities to labeled images. Deep Learning and Convolutional Neural Networks Convolutional neural networks (CNNs) are a type of artificial neural network employed in image analysis with high rates of success for image classification tasks. Like all artificial neural networks, they are comprised of layers which are functions estimated from data that are then pieced together 22

23 to form more complex non-linear functions of the kind required to perform complicated tasks such as identifying features from pixel intensity data alone. In functional form, a neural network is simply the composition of many functions f() with associated weights w, layers L, and data θ. f(θ) = f L (...f 2 (f 1 (θ, w 1 ), w 2 )...), w L ) (1) Weights for f(θ) are estimated such that they minimize the difference between predicted and actual class labels as provided by the data by minimizing the loss function λ(z i, f(θ i, w)) where z i are the true, labeled values and f(θ i, w)) are the predicted values as produced by the model. The empirical loss is then the average loss over each of the examples i: Λ(w) = 1 n λ(z i, f(θ i, w)) (2) n i=1 The loss function is minimized through an optimization technique known as back-propagation which continuously updates the loss function through a series of forward and backward passes through the model. During the forward pass, the data is first passed through the model in Equation 1 with a randomly initialized set of weights and predicted values are generated. Gradients for the weights in each of the layers are then computed and adjusted accordingly using a technique known as stochastic gradient descent: f w l = w l f(θ) (3) 23

24 What differentiates convolutional neural networks from artificial neural networks in general is that all of the functions from equation 1 are operators that are translation invariant. In other words, the convolution aspect of convolutional neural networks slide the set of functions across portions of pixel data in the image similar to the way in which basic filters, such as a Gaussian blur, work to transform images. Figure 11 Depiction of a convolutional neural network race classifier. The classifier utilizes pixel intensity data to estimate a complex non-linear function which can accurately predict high level features in images such as race. Here, we trained a convolutional neural network race classifier for Facebook images based on the VGG architecture (depicted in Figure 11 below), a deep classification network that was pre trained on the ImageNet Large Scale Visual Recognition (ISLVRC) dataset (Krizhevsky, Sutskever, and Hinton 2012). ISLVRC is a database containing a standard set of images used for bench-marking deep learning image classification models. We fine-tuned the 16-layer VGG model on 17,500 PubFig (Kumar et al. 2009) training images using the race annotations provided with the dataset as ground truth. Additionally, we included in the training set 44,000 portraits of American 24

25 high school seniors from a large Yearbook dataset that were classified with high confidence. Figure 12 Faces in yearbook and Facebook images detected using Viola- Jones algorithm with Haarcascades, cropped and converted to grayscale as a pre-processing step. All training images were cropped in a manner similar to Figure reffig:cropped by first detecting faces using Haarcascades (Gorbenko and Popov 2012; Lienhart and Maydt 2002), a modified version of the Viola-Jones algorithm (Viola and Jones 2001) to include only the face to match the Facebook test images. Additionally all training and Facebook test images were converted to grayscale. We split the training set into 61,000 training and 17,000 validation images, and used these to fine-tune the classifier for the labels White, African-American, Asian and Hispanic over 100,000 iterations (backpropagation passes). Since training the classifier on the PubFig and Yearbook datasets and testing on Facebook portraits constitutes a domain shift that 25

26 Predicted Class Ground Truth White Black Asian Hispanic Accuracy White 28,119 1, Black 620 3, Asian , Hispanic Table 3 Confusion matrix for the convolutional neural network image classifier. Outlined numbers are the number of correctly identified individuals of each race. hurt the classification performance, we improved the classification accuracy via a process of bootstrapping by manually verifying the high-confidence race classifications and adding these into the training set. Afterward, we further fine-tuned our classifier by training 20,000 iterations on the augmented training set. The final average cross-validated accuracies were 90% for Whites, 85% for African-American, 75% for Asian and 65% for Hispanic as can been seen in the confusion matrix in Table 3. 4 Results After classifying the race of all individuals in House and Senate members Facebook profiles, we test the hypotheses discussed above by estimating Facebook profile demographics for each member of Congress and compare these demographics to district and state demographics from American Community Survey (ACS) 5-year estimates. Images containing the actual members of Congress themselves were excluded from the analysis. We begin our analysis 26

27 Figure 13 Proportion African-American, Hispanic and Asian in Facebook photos of white House of Representative Members. by exploring the raw estimated proportion of African-Americans, Hispanic and Asians in the Facebook profiles of white Republican and Democratic House and Senate members. Figures 13 and 14 are box plots of the raw estimated proportion African- American, Hispanic and Asians in the Facebook profiles of white representatives in the House and Senate, respectively. The clear and consistent pattern that emerges from these figures is that white Democrats in both chambers appear to include a significantly higher proportion of African-Americans and 27

28 Figure 14 Proportion African-American, Hispanic and Asian in Facebook photos of white Senate Members. Asians in their images, but similar proportions of Hispanic voters. This provides some evidence for our hypothesis that race is an essential component of Democrat s, but not Republican s visible home styles, especially since white representatives in both chambers exhibit roughly the same patterns. The question that remains, however, is whether these differences are due to chance or whether they reflect the strategic use of images as defined above. As we discussed above, evidence of strategic behavior in this context is implied by over representation of minority group members which constitute a 28

29 Value Meaning Strategic Implication R i = 0 Representation Suggests null model. R i > 0 Over representation Suggests strategic catering to African American constituents. R i < 0 Under representation May suggest strategic catering to other minority groups. Table 4 Values of R i and their strategic implications. Among Democratic MCs we would expect R i > 0 if they tend to use photos to boost their re election chances by appealing to core minority constituencies while among Republican MCs we would expect R i = 0 since they have little incentive to appeal to African Americans. significant portion of the MCs base of support. Thus, if Democrats tend to use images strategically, we would expect to find that they have a significantly higher % of African Americans in their photos than they do in their districts while Republicans would post roughly the same % of African Americans in their photos as they have in their districts (see the Null Model in Table 2 for more details). Below, we explore this hypothesis by comparing the % of African Americans identified in MCs Facebook images with their district (House) or state (Senate) % African American. Specifically, we create a dependent variable which measures representation of African Americans in MCs Facebook photos as compared with their district demographics: R i = % Black in MC i s Facebook Photos % Black in MC i s District (4) We begin our analysis by exploring the distribution of R i among white 29

30 Figure 15 Distribution of black representation R i in Facebook photos among white Democratic and Republican House Members (% Black in Facebook Photos - % Black in district). White Democratic House members clearly over represent African Americans more frequently than Republican House members suggesting a strategic use of photos. A null model distribution should be symmetrical around zero. Democrats and Republicans in the House of Representatives. From Figure 15 it appears as if Democrats tend to over represent African Americans in their Facebook photos (R i > 0) while Republicans do not. To explore this further, we estimated two models. The first model is a logistic regression of a dichotomous transformed version of R i such that if R i > 0, R i = 1 or 0 otherwise: 30

31 logit(e[r > 0 Democrat, X]) = α + βdemocrat + Xγ + ɛ (5) Parameter estimates and predicted probabilities from Equation 5 give us a sense of the conditional probability of over representation of African Americans among Democratic and Republican House members. Figure 16 Predicted probability of black over representation in Facebook photos among Democratic and Republican House members from the logistic regression model: logit(e[r > 0 Democrat, X]) = α + βdemocrat + Xγ + ɛ Predicted conditional probability estimates of over representation produced from Model (3) in Table 5 are plotted in Figure 17. The average pre- 31

32 Figure 17 Distributions of the predicted probability of black over representation in Facebook photos from the logistic regression model: logit(e[r > 0 Democrat, X]) = α + βdemocrat + Xγ + ɛ dicted conditional probability that Democrats will over represent African Americans in their Facebook photos 7 is 57.1% (52.1%, 62.2%) and for Republicans it is 34.7% (30.8%, 38.7%). Clearly, Democratic MCs, even after conditioning on race, region and other district demographics, are far more likely to over represent African Americans in their photos than are Republicans. Indeed, a closer look at the predicted conditional probability distributions among Democrats and Republicans (Figure 17) reveals more interesting information. 7 95% confidence intervals in parentheses 32

33 DV: 1 = %Black in Photos > %Black in District, 0 = Otherwise (1) (2) (3) Democrat 0.918*** 0.803** 0.849* (0.272) (0.305) (0.406) White MC ** (0.380) (0.696) South *** (0.410) % White in District *** (1.659) % Hispanic in District *** (1.380) N R p 0.05 p 0.01 p Table 5 Logistic regression of Over-representation on MCs party and other covariates. Coefficents reported are raw coefficient values estimated from the equation: logit(e[r > 0 Democrat, X]) = α + βdemocrat + Xγ + ɛ. Odds ratio calculations (exp(democrat)) suggest that Democratic MCs are more than twice as likely (exp(0.849)) to over-represent African Americans in their photos. While predicted probabilities of over representation among Democrats is unimodal with a center clearly higher than Republicans, Republicans predicted probabilities are bimodal with one group of Republicans centering around 50% and other group of Republicans centering around 11%. This observation led us to an exploration of African American under representation in Facebook photos among Democrats and Republicans. In addition to over representation of African Americans among Democrats, Republicans, especially those in districts which are very conservative or those districts in which constituents may harbor racial resentment (Feldman and Huddy 2005) may 33

34 Figure 18 % of MCs in which African Americans are under represented in Facebook photos in Southern and Non-Southern districts. There are large, statistically significant differences in under representation among Republicans in Southern and Non-Southern districts with 88.7% of Southern Republicans under representing African Americans. strategically under represent African Americans. Given the South s history of racial tensions (Acharya, Blackwell, and Sen 2016), we might expect Southern members of Congress, especially those in more conservative/republican districts to under represent African American s in their photos. Figure 18 contains estimates of the % of MCs that under represent African American s in their photos by party and region. While both Democrats and Republicans in the South tend to under represent African Americans, it is clear that Southern Republicans have the highest rates of black under representation 34

35 of all groups at 88.7% of Southern Republicans under representing African Americans in their photos. DV: Black under representation in photos. (1) (2) (3) Republican 0.918*** (0.272) (0.331) (0.407) Republican x South * 1.885* (0.677) (0.833) South (0.456) (0.665) District % Black *** (2.317) White MC ** (0.882) District % Hispanic * (1.303) N p 0.05 p 0.01 p Table 6 Logistic regression of Under representation on MCs Republican party affiliation and Southern district. Coefficents reported are raw coefficient values estimated from the equation: logit(e[r < 0 Republican, South, X]) = α+β 1 Republican+β 2 Republican x South+β 3 South+Xγ+ɛ. Odds ratio calculations (exp(republican)) suggest that Republican MCs in the South are over 6 times more likely (exp(1.885) = 6.6) to under-represent African Americans in their photos. To confirm this result conditional on relevant district and member covariates, we estimated the following logistic regression model in which the dependent variable was coded as 1 if the MC under represented African Americans in their photos, Republican is party affiliation of the MC and South is whether the MC s district is in a Southern state: 35

36 logit(e[r < 0 Republican, South, X]) =α + β 1 Republican + β 2 Republican x South + β 3 South + Xγ + ɛ (6) Table 6 model (3) contains raw coefficient estimates from Equation 6. Conditional on member and district covariates, we find that Southern Republicans are about 6.6 times more likely 8 than all other House members to under represent African Americans in their photos. There are several explanations for this finding. One possibility is that Southern Republicans simply don t pose with African Americans by chance. This explanation seems rather unlikely for several reasons. First, African Americans comprise a significantly higher % of most districts in Southern states than non-southern states, yet under representation is far less common among Republicans in non Southern states (46.2% of districts v. 88.7% of districts in Southern states). Second, we saw from results in Table 6 after conditioning on % black in the MCs district, the interaction term β 2 on Republican x South actually increased. Another, more likely explanation for this phenomenon is that Republican MCs in Southern states, which have historically have high rates of racial resentment among whites, may seek to minimize any possibility of antagonizing constituents who harbor racial resentment or feelings of 8 exp(β 2 ) = exp(1.885) =

37 racial animosity. While a more detailed discussion of this claim is beyond the score of this paper, we hope that students of Congress explore this claim in more depth in the future. 5 Discussion The use of images by political figures to manipulate public opinion and sentiment is by no means a new phenomenon. As discussed above, shrewd political figures such as Lyndon Johnson recognized the potential that images had to shape how they were perceived by the public and accordingly appointed the first White House photographer to do just that during his term in office. At the same time, however, photographs taken by brave journalists and chroniclers of the Vietnam War laid bare the horrors and devastation of a war which eventually led to Johnson s steep decline in popularity and his eventual decision to not run for a second term. While the use of images to achieve political ends is not new, our ability to systematically study and understand how, when and why they are used has only recently been made possible by the prolific use of images by politicians via the internet and recent developments in computer vision which allow for fast and accurate identification of complex features from labeled image data. Here, we take advantage of both developments in an effort to provide a broad framework for political image analysis using these techniques and simultaneously demonstrate how they can be used to understand the mod- 37

38 ern relevance of home style (Fenno 1978). By breaking down images into their simplest political elements: objects, people and poses, our framework provides a basis from which scholars can explore which aspects of images are used by politicians and others for the purpose of influencing opinion. In addition to this, we make contributions to an understanding of home style in the 21st century. We first develop a theory of visible home style by linking our broader image analysis framework to Fenno s main conceptual elements of home style. We then demonstrate that the characteristics of individuals that politicians pose with affect how they are perceived across a number of politically relevant dimensions related to home style which include qualification, identification and empathy. Using convolutional neural networks and an empirical analysis of photos posted by members of the House and Senate on their Facebook profiles, we provide evidence that electoral pressures shape visible home styles. Specifically, we show that Democratic members of Congress over represent African Americans in their Facebook profile photos in part, we argue, as a means of eliciting identification and empathy with African American constituents, a claim that is backed up with causal evidence from an experiment that we conducted that can be found in Appendix A. On the other hand, virtually all Souther Republican House members in our sample employed the opposite strategy by under representing African Americans in their Facebook photos, potentially in an effort to appease white constituents who may foster higher levels of racial resentment. 38

39 References Accetti, Carlo Invernizzi, and Fabio Wolkenstein The crisis of party democracy, cognitive mobilization, and the case for making parties more deliberative. American Political Science Review 111 (1): Acharya, Avidit, Matthew Blackwell, and Maya Sen The political legacy of American slavery. The Journal of Politics 78 (3): Ahn, TK, Robert Huckfeldt, Alexander K Mayer, and John Barry Ryan Expertise and bias in political communication networks. American Journal of Political Science 57 (2): Altschuler, Bruce E Lyndon Johnson and the public polls. Public Opinion Quarterly 50 (3): Arceneaux, Kevin Cognitive biases and the strength of political arguments. American Journal of Political Science 56 (2): Baker, Andy Race, paternalism, and foreign aid: Evidence from US public opinion. American Political Science Review 109 (1): Barberá, Pablo Birds of the Same Feather Tweet Together. Bayesian Ideal Point Estimation Using Twitter Data. Political Analysis 23 (1): Barreto, Matt A, Gary M Segura, and Nathan D Woods The mobi- 39

40 lizing effect of majority minority districts on Latino turnout. American Political Science Review 98 (01): Bishin, Benjamin G, Thomas J Hayes, Matthew B Incantalupo, and Charles Anthony Smith Opinion Backlash and Public Attitudes: Are Political Advances in Gay Rights Counterproductive? American Journal of Political Science 60 (3): Bond, Robert, and Solomon Messing Quantifying social medias political space: Estimating ideology from publicly revealed preferences on Facebook. American Political Science Review 109 (01): Bonica, Adam Ideology and interests in the political marketplace. American Journal of Political Science 57 (2): Broockman, David E Black politicians are more intrinsically motivated to advance Blacks interests: A field experiment manipulating political incentives. American Journal of Political Science 57 (3): Cameron, Charles, David Epstein, and Sharyn O halloran Do majority-minority districts maximize substantive black representation in Congress? American Political Science Review 90 (04): Campbell, David E, John C Green, and Geoffrey C Layman The party faithful: Partisan images, candidate religion, and the electoral impact of party identification. American Journal of Political Science 55 (1):

41 Caro, Robert A Means of Ascent: The Years of Lyndon Johnson. New York: Alfred A. Caro, Robert A Master of the Senate. Vol. 3 Alfred a Knopf Incorporated. DeNardo, James Turnout and the Vote: The Joke s on the Democrats. American Political Science Review 74 (02): Desmarais, Bruce A, Raymond J La Raja, and Michael S Kowal The fates of challengers in US House elections: The role of extended party networks in supporting candidates and shaping electoral outcomes. American Journal of Political Science 59 (1): Dilliplane, Susanna, Seth K Goldman, and Diana C Mutz Televised exposure to politics: New measures for a fragmented media environment. American Journal of Political Science 57 (1): Duganne, Erina The Photographic Legacy of Lyndon Baines Johnson. Photography and Culture 6 (3): Esterling, Kevin M, David MJ Lazer, and Michael A Neblo Connecting to Constituents: The Diffusion of Representation Practices among Congressional Websites. Political Research Quarterly: Estrin, James Photographing the White House From the Inside.. 41

42 Feldman, Stanley, and Leonie Huddy Racial resentment and white opposition to race-conscious programs: Principles or prejudice? American Journal of Political Science 49 (1): Fenno Home style: House members in their districts. HarperCollins. Fiorina, Morris P, and David W Rohde Home style and Washington work: studies of congressional politics. University of Michigan Press. Gerber, Elisabeth R, Adam Douglas Henry, and Mark Lubell Political homophily and collaboration in regional planning networks. American Journal of Political Science 57 (3): Giger, Nathalie, and Heike Klüver Voting against your constituents? How lobbying affects representation. American Journal of Political Science 60 (1): Gorbenko, Anna, and Vladimir Popov On face detection from compressed video streams. Applied Mathematical Sciences 6 (96): Harden, Jeffrey J Multidimensional responsiveness: the determinants of legislators representational priorities. Legislative Studies Quarterly 38 (2): Hayes, Danny, Jennifer L Lawless, and Gail Baitinger Who cares what they wear? Media, gender, and the influence of candidate appearance. Social Science Quarterly 95 (5):

43 Hutchings, Vincent L, Hanes Walton Jr, and Andrea Benjamin The impact of explicit racial cues on gender differences in support for confederate symbols and partisanship. The Journal of Politics 72 (4): Joo, Jungseock, Weixin Li, Francis F Steen, and Song-Chun Zhu Visual persuasion: Inferring communicative intents of images. In 2014 IEEE Conference on Computer Vision and Pattern Recognition. IEEE pp VisualPersuasion_CVPR14.pdf. Kotz, Nick Judgment Days: Lyndon Baines Johnson, Martin Luther King, Jr., and the Laws That Changed America. Houghton Mifflin Harcourt. Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E Hinton Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems. pp imagenet-classification-with-deep-convolutional-nn.pdf. Kumar, Neeraj, Alexander C. Berg, Peter N. Belhumeur, and Shree K. Nayar Attribute and Simile Classifiers for Face Verification. In In IEEE International Conference on Computer Vision (ICCV. com/papers/kbbn09iccv.pdf. Lau, Richard R, and David P Redlawsk Advantages and disadvan- 43

44 tages of cognitive heuristics in political decision making. American Journal of Political Science: Lenz, Gabriel S, and Chappell Lawson Looking the part: Television leads less informed citizens to vote based on candidates appearance. American Journal of Political Science 55 (3): Lienhart, Rainer, and Jochen Maydt An extended set of haar-like features for rapid object detection. In Image Processing Proceedings International Conference on. Vol. 1 IEEE pp. I cf c5ccd e b6c1.pdf. Lopes da Fonseca, Mariana Identifying the Source of Incumbency Advantage through a Constitutional Reform. American Journal of Political Science. Marschall, Melissa J, and Amanda Rutherford Voting Rights for Whom? Examining the Effects of the Voting Rights Act on Latino Political Incorporation. American Journal of Political Science 60 (3): Mayhew, David R Congress: The electoral connection. Yale University Press. Miller, Joanne M, Kyle L Saunders, and Christina E Farhart Conspiracy endorsement as motivated reasoning: The moderating roles of political 44

45 knowledge and trust. American Journal of Political Science 60 (4): Parker, David CW, and Craig Goodman Our States Never Had Better Friends: Resource Allocation, Home Styles, and Dual Representation in the Senate. Political Research Quarterly 66 (2): Prior, Markus Visual political knowledge: A different road to competence? The Journal of Politics 76 (1): Rosenberg, Shawn W, Lisa Bohan, Patrick McCafferty, and Kevin Harris The image and the vote: The effect of candidate presentation on voter preference. American Journal of Political Science: Ryan, Timothy J. 2017a. How Do Indifferent Voters Decide? The Political Importance of Implicit Attitudes. American Journal of Political Science. Ryan, Timothy J. 2017b. No compromise: Political consequences of moralized attitudes. American Journal of Political Science 61 (2): Viola, Paul, and Michael Jones Rapid object detection using a boosted cascade of simple features. In Computer Vision and Pattern Recognition, CVPR Proceedings of the 2001 IEEE Computer Society Conference on. Vol. 1 IEEE pp. I gatech.edu/paper_of_week/viola01rapid.pdf. Zarefsky, David Presidential rhetoric and the power of definition. Presidential Studies Quarterly 34 (3):

46 Appendix Appendix A: The people you pose with experiment Figure 19 Questions asked to respondents and their relationship to each element of home style. The empirical portion of our paper below explores the connection between racial representation and visual home styles through a systematic examination of the race of individuals that members of Congress pose with and the images they choose to post on their social media sites. Before moving on to this analysis, however, we first seek to demonstrate that group characteristics such as the race and gender of people that members of Congress pose with causally affect politically relevant outcomes related to home style. According to literature exploring how the group characteristics of politicians themselves affect how they are perceived by individuals who are members of the same groups, we would expect that posing with members of groups that have po- 46

47 litical relevance (i.e. minority group members, women, veterans, etc.) would elicit greater support for the politician among members of those groups (Ahn et al. 2013; Bonica 2013; Broockman 2013; Gerber, Henry, and Lubell 2013; Bishin et al. 2016; Marschall and Rutherford 2016) To accomplish this, we designed an image manipulation experiment in which the treatment randomized was a series of images containing a member of Congress posing by himself (control) or next to people of different genders and races which was presented to survey respondents 9. Outcomes collected were respondent s best guess about the MCs qualities that directly relate to presentations of self relevant to qualification, identification and empathy as shown in Figure 19. No information other than the photograph and a sentence explaining that the person in the photograph is a politician in the United States was provided to respondents. Figure 20 contains each of the potential image treatments that respondents were randomly exposed to. The political subject of the experiment that was chosen was Republican Rep. Lou Barletta who represents Pennsylvania s 11th District. Rep. Barletta was chosen for several reasons. First, he is a relatively obscure political figure 10, thus making it unlikely that respondents will base their judgments about him based on anything outside of the images. Second, he happens to have many pictures on his Facebook profile where he is posing with people of different races and genders with the same 9 See Appendix C for details about subject recruitment and survey design. 10 Less than 1% of survey respondents recognized Barletta. 47

48 Figure 20 Experimental treatments shown to subjects taken from Facebook photos. background, allowing us to isolate the effects of the race and gender of the person standing next to him on the outcomes shown in Figure 19. We find treatment effects for race and gender across different categories of political outcomes. Posing next to African-Americans made respondents more likely to believe that Barletta was a Democrat and was more liberal. 62% of respondents guessed that he was a Democrat when he was posed next to African-Americans compared with 36% to 44% across all other categories of photos. Posing next to African-Americans also led non-white respondents to be more likely to agree that Barletta shared their values as can be seen in Figure 22. Interestingly, we also find effects of gender on perceptions of qualification and competence. When Barletta poses next to a woman, respondents are more likely to agree that he is more trustworthy (Figure 23) and knowledge- 48