Declared Support and Clientelism

Transcription

1 Declared Support and Clientelism Simeon Nichter University of California, San Diego Department of Political Science Salvatore Nunnari Bocconi University Department of Economics June 6, 2017 Abstract Recent studies of clientelism provide many insights about how elites use rewards to influence vote choices and turnout. This article shifts attention towards citizens and their choices beyond the ballot box. Under what conditions does clientelism influence citizens decisions to express political preferences publicly? When voters can obtain future benefits by declaring support for victorious candidates, their choices to display campaign posters, wear political paraphernalia or attend rallies may reflect more than just political preferences. We argue that various factors, such as political competition and candidates monitoring ability, heighten citizens propensity to declare support in response to clientelist inducements. Building on insights from our interviews and surveys, a theoretical framework reveals how and why such factors can distort patterns of political expression observed during electoral campaigns. We conduct two experiments in Brazil and the U.S., which predominantly corroborate predictions about declared support and clientelism. The authors thank Andy Baker, Jordan Gans-Morse, Pietro Ortoleva for comments, as well as Inbok Rhee, Mariana Carvalho-Barbosa, Anselm Rink, and Henrique Barbosa for outstanding research assistance. In addition, we appreciate feedback from participants at Stanford University s Conference on Contemporary Challenges to Inclusion and Representation in Latin America and at the 2015 Annual Meeting of the American Political Science Association. Simeon Nichter acknowledges support from the Hellman Foundation, the National Science Foundation, and the Harvard Academy for International and Area Studies.

2 1 Introduction In many parts of the world, citizens receive material benefits in contingent exchange for providing political support. A recent cross-country survey of 1,400 experts found that such patterns of clientelism exist to some degree in over 90 percent of nations, with clientelist efforts by elites reaching moderate or major levels in nearly three-fourths of countries (Kitschelt, 2013). This phenomenon is widely recognized to have a broad range of consequences for democracy and development. Clientelism often exacerbates political inequalities by allowing those with resources to buy votes from impoverished citizens, and undermines representation when vote choices no longer reflect recipients political preferences (Schaffer and Schedler, 2007; Stokes et al., 2013). Moreover, clientelism is frequently linked to numerous maladies that can stifle development, such as the under-provision of public goods, increased rent seeking and expanded public deficits (Hicken, 2011, 302 4; Keefer, 2007). The literature on clientelism tends to focus on the role of elites and often depicts citizens as passive recipients. In past decades, research focused on highly asymmetric bonds between patrons and clients, with exchanges rigidly controlled by elites (e.g., Cornelius, 1977; Powell, 1970). Within many patron-client relationships, voters had limited autonomy to make choices of their own volition, due to various factors such as restrictive land-tenure arrangements (Scott, 1972, 93; Hall, 1974). This traditional focus on elites and the depiction of citizens as passive recipients often continues in contemporary research on clientelism, especially in formal and quantitative studies. Particularly emblematic of this tendency, many analysts concentrate on the supply-side logic by which politicians and their representatives target citizens when distributing campaign handouts. For instance, Stokes (2005) contends that elites reward weakly opposed voters for vote-switching (i.e., vote buying ), whereas Nichter (2008) argues that they reward nonvoting supporters for showing up at the polls (i.e., turnout buying ). This elite-targeting focus is extended in various studies motivated by Stokes et al. (2013), who argue that party leaders efforts to target weakly opposed voters are 1

3 hindered by brokers who channel rewards to supporters. Other recent examples that predominantly focus on elites include studies of how machines combine multiple targeting strategies (Diaz-Cayeros, Estévez and Magaloni, 2016; Gans-Morse, Mazzuca and Nichter, 2014), and work on how politicians target reciprocal voters with handouts (Finan and Schechter, 2012). While such studies have greatly enhanced our understanding of elite strategies, they shed far less light on the choices of citizens in contexts with clientelism. Such choices deserve greater attention, especially because some qualitative work on the topic has revealed substantial citizen autonomy in various contexts. Indeed, scholars have long observed variation in the degree to which elites dominate contingent exchanges (Scott, 1972; Piattoni, 2001), and have documented a decline over the years in elite control of exchanges (e.g., Archer, 1990; Gay, 2006). Various factors may heighten citizens ability to make autonomous decisions, such as enhanced educational attainment, improved access to mass media and information technology, and growing urbanization. In some contexts, the transition from monopolistic to competitive clientelism has also amplified the role of voter choices; in particular, citizens may opt for alternative sources of benefits when a single machine no longer dominates all exchanges. Recent qualitative research continues to call attention to increasing voter autonomy in clientelist exchanges (e.g., Auyero, 2000; Hilgers, 2012; Taylor-Robinson, 2010), but the mechanisms underlying citizens choices remain largely unexplored. The present study advances this literature by elaborating and testing a theoretical framework focused on citizens choices in contexts with clientelism. In particular, we investigate the following question: Under what conditions does clientelism influence citizens decisions to express political preferences publicly? Unlike the dominant paradigm, which tends to relegate citizens to a passive role in exchanges, we argue that many citizens do not respond mechanically to elite offers. Voter responsiveness to clientelist inducements cannot be taken for granted, nor can it be deduced by simply comparing reward sizes to the strength of citizens political preferences. Instead, the present study explores how and why citizens will 2

4 often choose different actions in contexts with clientelism, depending on numerous contextual characteristics. In addition to this focus on citizen choices, another key contribution is that we investigate how clientelism can induce political expression beyond the ballot box. Unlike some qualitative studies, formal and quantitative work on clientelism tends to focus narrowly on voting. By contrast, we examine why citizens publicly express support for political candidates, through actions such as displaying campaign posters, wearing political paraphernalia and attending rallies. Many studies of American politics consider such activities to be important forms of democratic participation, which enable citizens to express their political preferences and potentially influence the selection of leaders (e.g., Verba and Nie, 1972; Huckfeldt and Sprague, 1995). But in much of the world, clientelism presents another understudied motivation. When voters can obtain future benefits by declaring support for victorious candidates, their decisions to participate publicly often reflect more than just political preferences. We argue that various factors affect citizens propensity to declare support in response to clientelist inducements. For instance, citizens may deem it especially advantageous to declare support for a clientelist candidate who distributes large rewards, is likely to win the election, and can easily observe declarations. But citizens may also balk at declaring for that candidate if doing so is costly: it might be challenging to obtain campaign materials or attend rallies, citizens might prefer another candidate ideologically, or they might live in neighborhoods where declaring for that candidate involves social costs. And in some contexts, citizens might even face punishments if they declare support for a candidate who loses the election. Considering such factors, we investigate how clientelism affects citizens choices about declared support. Evidence suggests that the logic of declared support warrants careful investigation. As explored in the next section, many of our interviewees in Northeast Brazil explained that declaring support for a victorious candidate would improve access to benefits after an election. We investigated such perceptions in a survey experiment, by randomly exposing online 3

5 participants across Brazil to different vignettes. Corroborating our qualitative findings, subjects perceived that declared supporters of the victorious mayor would have an easier time receiving health care, water cisterns and employment than other citizens. In line with such perceptions, two surveys we fielded in Brazil suggest that declared supporters are indeed more likely to receive benefits from elected politicians. Moreover, observational work suggests declared support is by no means limited to Brazil. Our recent book chapter on Mexico s 2012 presidential election suggests that declarers were significantly more likely to receive clientelist benefits than non-declarers, even when including various controls and municipal fixed effects. 1 Qualitative research in Argentina suggests that citizens who attend rallies, a potential form of declaration, are more likely to receive handouts (Auyero, 2000, 163; Szwarcberg, 2015, 66 7). In Ghana, fieldwork reveals that many citizens perceive that publicly expressing support will improve their ability to receive future benefits from elected politicians (Michelitch, 2013, ). And in Lebanon, Melani Cammett finds that citizens are more likely to receive benefits if they demonstrate their partisan commitment through various actions such as displaying posters and voting (2011, 75, 84 7; 2014, 128 9). While such studies do not elaborate or test the logic by which citizens declare support, they provide evidence of the phenomenon around the world and suggest that both analytical tasks would be important contributions. The present study addresses both of these challenges, as it is the first to develop and test a theoretical model of declared support. This model provides numerous predictions about voters declaration choices in contexts with clientelism, which we test experimentally. For instance, formalizing the intuition discussed above, it suggests that citizens are most likely to declare support when a clientelist politician: (a) offers larger material rewards, (b) can monitor declarations effectively, (c) is heavily favored to win the election, (d) is preferred on programmatic or ideological grounds; and (e) can be publicly supported without incurring social costs. Just as important, the model predicts how these factors will affect 1 [Citation omitted for peer review.] 4

6 citizens choices about whether to remain undeclared or to declare support for an opposition candidate who offers no clientelist benefits. Furthermore, it examines how several other important features of polities can influence political expression beyond the ballot box, such as whether citizens are punished for publicly supporting a defeated candidate, and whether clientelism is a monopolistic or competitive phenomenon. Overall, formal analyses yield a rich set of hypotheses about citizen choices and declared support in contexts with clientelism. Both online and laboratory experiments are employed to test these predictions about how inducements affect citizens decisions to express political preferences publicly. The present study thereby offers another important contribution, as it is the first to examine the choices of citizens in clientelism using experimental methods. A key advantage of this approach is that it isolates effects by changing only one aspect of the decision environment at a time; by contrast, testing predictions observationally would require disentangling various reasons why citizens declare. Our primary experiment involved 1,259 online participants from 1,061 municipalities across Brazil. To investigate mechanisms, this experiment exposed subjects to 10 distinct treatments, each testing theoretical predictions about multiple declaration actions. As discussed below, experimental results are consistent with 16 of the model s 21 unconditional predictions (76 percent), even when focusing exclusively on within-subject variation. Secondarily, we also conducted a smaller experiment with 144 university students in a supervised computer lab in the United States. This experiment offered greater control of the testing environment but weaker external validity than the Brazilian experiment. Even in this substantially different context, results are consistent with 15 of 18 unconditional predictions (83 percent) tested experimentally. 2 Taken together, our mixed-methods approach emphasizes and elucidates the often overlooked role of citizen choices in clientelism. Building on insights from fieldwork, a formal model unpacks the voter calculus of publicly expressing political support when contingent 2 The experiment in the U.S. included fewer treatments than in Brazil, and thus tested fewer predictions. 5

7 benefits are distributed. Experiments predominantly confirm predictions, which suggest how and why citizens consider numerous factors when choosing whether to declare support in response to inducements. 2 Motivating Evidence We first explore patterns of declared support in Brazil, in order to provide motivation for further analyses and to guide modeling assumptions. During electoral campaigns, many Brazilians undertake various actions to convey support for candidates. For instance, the 2007 LAPOP AmericasBarometer survey indicates that 22.1 percent of Brazilian respondents displayed a candidate s banner or poster at home or work, or placed a candidate s sticker on their cars. 3 Of course, such declarations need not involve clientelism. Among numerous motivations, citizens may declare support simply to express their political preferences or to help their preferred candidates with free advertising. But, as discussed below, qualitative and quantitative research points towards an intriguing link between declarations and clientelism. This section builds on 130 interviews of citizens and elites we conducted in the states of Bahia and Pernambuco, as well as two extensive surveys we fielded more broadly in Brazil. The link between clientelism and declared support was a common theme during these interviews. Many citizens perceived that they had obtained preferential access to various post-election benefits because they had declared for the victorious candidate during a municipal campaign. The example of medicine was repeatedly mentioned: politicians reportedly favor their declared supporters with benefits from the public health system, and also use their own funds and salaries to purchase medicine for them at private pharmacies. In the words of a mason, if in need of a costly medical procedure, he would turn to the politicians I voted for, who would help because I declare my vote before voting. 4 According to some 3 The survey included a single question about declarations (mentioning all these actions), but no questions about clientelism. 4 Author s interview, municipality in Bahia with 80,000 citizens (11/21/2008). 6

8 interviewees, declared supporters more easily procure health services such as ambulances during emergencies and transport for surgeries in distant capitals. A councilor paraphrased a famous quote when explaining how declared supporters sometimes receive favored access to healthcare: For friends of the king, everything, and for others, [n]othing... Not everyone thinks so, but this, practically, is customary. 5 As another example, many interviewees also expressed their belief that declared supporters are favored with employment. A homemaker explained that it is easier for someone who declared because a politician says: You were on my side. I am going to give you work. 6 Citizens explained that such benefits of declaring support are contingent on the candidate winning the election, in part due to the resources available to politicians in office. As a homemaker in another municipality insisted, if I declare my vote for a candidate who loses, he definitely won t help me... he didn t win, how will he help me? 7 Local politicians are able to monitor, albeit imperfectly, citizens declaration actions because campaigns tend to have extensive direct contact with voters before municipal elections. For example, in an survey we conducted of 3,716 citizens across rural Northeast Brazil, respondents received an average of 4.6 home visits by candidates representatives, and two-thirds of survey takers and nearly three-fourths of declarers believed that others would remember who declared support by placing campaign flags on their homes. 8 We explore below how the magnitude of clientelist benefits, as well politicians monitoring ability, is expected to influence declarations. Fieldwork identified numerous other factors, beyond clientelist inducements, that also influence citizens declaration decisions. As one might expect, many interviewees emphasized their preferences towards candidates when discussing why they declared support, mentioning policy proposals, candidate qualities, personal affinities and other considerations. Some citizens expressly noted how much they enjoyed displaying political paraphernalia and attending 5 Author s interview, municipality in Bahia with 60,000 citizens (11/4/2008). 6 Author s interview, municipality in Bahia with 10,000 citizens (10/22/2008). 7 Author s interview, municipality in Bahia with 15,000 citizens (01/13/2009). 8 This three-year panel survey involved a random sample of citizens in 40 municipalities across nine states in Northeast Brazil. [Author s survey; identity omitted for peer review.] 7

9 rallies of preferred candidates, while others indicated that such actions could influence others and thereby help elect their preferred candidates. On the other hand, other interviewees provided various reasons for remaining undeclared. Some citizens mentioned declaration costs, such as a computer technician who said he did not attend rallies because my time is sacred, and a temporary worker who complained it was too challenging to remove a candidate s bumper stickers or to repaint his home s wall. 9 Social repercussions also play a role in preventing some declarations; for instance, a maid explained she did not declare because the other candidate s supporters would complain a lot... they fight, they get angry. 10 Various interviewees pointed to perceived punishments as another reason to remain undeclared: in addition to rewarding declared supporters, elected politicians can also disfavor citizens who declared for defeated candidates. Such respondents frequently used the word marcação ( labeling or marking ) to describe how politicians identify citizens who declared against them and provide them fewer post-election benefits (i.e., even fewer than undeclared citizens). Uncertainty about the election outcome was also mentioned as another consideration; for example, a municipal guard explained why he did not declare support: Who is going to win? Who is going to lose? I don t know. 11 Building on these insights, our formal and experimental analyses below investigate the logic by which such factors affect declaration decisions. Do perceptions of a link between declarations and the receipt of benefits extend beyond interviewees in Northeast Brazil? To explore this question, we developed an online survey experiment conducted across Brazil. 12 We randomly exposed 1,995 participants in over a thousand municipalities to one of several vignettes, depicting citizens who had or had not 9 Author s interviews, municipalities in Bahia with 80,000 and 10,000 citizens, respectively (11/20/2008 and 10/22/2008). 10 Author s interview, municipality in Bahia with 100,000 citizens (12/22/2008). 11 Author s interview, municipality in Bahia with 10,000 citizens (10/4/2008). 12 As discussed in Section 4.1, we recruited participants using advertisements on Facebook, and subjects were fairly representative of Brazil s overall population in terms of age, gender, region and urban/rural mix. The survey experiment has a larger sample size, as it includes additional participants recruited similarly. 8

10 Figure 1: Declared Support and Perceived Difficulty of Obtaining Benefits Declared for Winner Declare Winner No Declaration Undeclared Medicine Declared for Winner Declare Winner No Declaration Undeclared Declared for Winner Declare Winner No Declaration Undeclared percent 0% 20% 40% 60% 80% 100% Water Employment Very Easy Easy Hard Very Hard Note: This survey experiment randomly assigned participants to view a vignette depicting a citizen who had or had not declared support during a fictitious mayoral campaign. Participants were then asked how difficult it would be for that citizen to obtain a medical treatment, water cistern, and employment after the election. Declared for Winner indicates subjects whose vignette depicted a declared supporter of the election winner. No Declaration indicates subjects whose vignette depicted an undeclared citizen. Source: Authors survey with 1,995 respondents recruited across Brazil on Facebook. Sample is fairly representative of Brazil s population in terms of age, gender, region and urban/rural mix. Section 4.1 provides further details. 9

11 declared support during a fictitious mayoral campaign. As shown in Figure 1, subjects who viewed a vignette depicting a declared supporter of the election winner indicated it would be easier for that citizen to obtain post-election benefits, than did subjects who viewed a vignette depicting an undeclared citizen. First, 43.1 percent of participants exposed to the declaredsupporter vignette perceived it would be easy or very easy for the citizen to obtain a medical treatment, compared to only 35.5 percent of those exposed to the undeclared-citizen vignette. Second, 39.7 percent of subjects randomly assigned to the declared-supporter vignette believed it would be easy or very easy for the citizen to receive a water cistern, compared to just 28.7 percent of those assigned to the undeclared-citizen vignette. And third, 38.6 percent of respondents viewing the declared-supporter vignette believed it would be easy or very easy for the citizen to obtain employment, compared to only 18.2 percent of those viewing the undeclared-citizen vignette. These differences, which are all statistically significant (at the.01 level), suggest that the survey participants across Brazil perceived a link between declared support and clientelism. 13 Beyond perceptions, are declarations actually linked to the benefits that citizens receive after an election? Although not the focus of this study, such evidence would provide additional motivation for why the logic of declared support warrants close investigation. In line with perceptions, two surveys we fielded in Brazil suggest that declared supporters are indeed more likely to receive benefits from elected politicians. In the online survey discussed above, we asked respondents if they had publicly declared support for a candidate during the 2012 municipal campaign, and about post-election private benefits they received during the subsequent term in office ( ). Approximately 27.6 percent of citizens who declared support for a victorious candidate received such benefits from elected politicians during their time in office, compared to only 11.1 percent of non-declarers. We also asked about assis- 13 Punishments may be used less frequently in Brazil. Other subjects were exposed to a vignette depicting a declared supporter of the election loser. Across all three benefits, these subjects thought he would experience greater difficulty than did subjects exposed to the declared-supporter vignette. But only for cisterns did they believe he would experience greater difficulty than did subjects exposed to the undeclared-citizen vignette. 10

12 tance from the municipal government during the same period, given that politicians may help clients obtain municipal benefits rather than use their own resources. Over 34.1 percent of declared supporters of elected politicians received private benefits from municipal offices, versus only 15.2 percent of non-declarers. Similar patterns were observed in the face-to-face panel survey we conducted across rural Northeast Brazil, though magnitudes are smaller in part because questions pertained to benefits distributed during a shorter time after the election (only one year). 14 Approximately 11.8 percent of respondents who declared support for a victorious candidate during the 2012 municipal campaign received private benefits from elected politicians the next year, versus just 4.0 percent of undeclared citizens. In addition, 8.4 percent of declared supporters received benefits from the municipal government in 2013, compared to just 4.0 percent of non-declarers. While all differences mentioned in this paragraph are statistically significant (at the 1 percent level), one might reasonably be concerned that they do not control for numerous variables that may be associated with both declaration and the receipt of benefits (e.g., income or partisanship). Elsewhere, we employ extensive regression analyses of both datasets to show that such patterns of favoritism towards declared supporters are robust to the inclusion of numerous control variables and municipal fixed effects. 15 In sum, qualitative and quantitative evidence from Brazil points to a substantial link between declared support and post-election benefits, both in terms of citizens perceptions and the benefits they report receiving. Building on this motivating evidence, we now develop and test a formal model of declared support. 14 In addition, this survey employed vignettes to examine perceptions. But instead of a survey experiment, it showed all respondents each vignette. For all three benefits, this observational approach revealed perceptions of favoritism towards declared supporters and punishments against declared opposers. 15 [Citation to author s book manuscript omitted for peer review.] 11

13 3 Formal Analysis 3.1 Setup of Model To investigate citizens choices about declared support, the present study develops a theoretical model with numerous predictions that are tested experimentally. Citizens are modeled as strategic individuals who decide whether and for whom to declare support not only on the basis of political preferences, but also on the basis of inducements and contextual characteristics. In the formal analysis, each citizen weighs whether to declare support for one of two candidates (A or B), or to remain undeclared. A citizen is assumed to decide whether and for whom to declare on the basis of her overall expected utility. In the base model, this utility depends on five factors: (1) her political preferences with respect to the election winner, (2) any reward she receives for declaring for clientelist candidate A, (3) the cost of declaring, (4) expressive utility from declaring, and (5) any impact of her declaration on the election outcome. For exposition, we first focus on the case of monopolistic clientelism involving rewards, and then examine extensions for the cases of competitive clientelism and the use of punishments. In the base model, the timing of the game is as follows: 1. Citizen i decides whether to declare for clientelist candidate A at cost c A 0, to declare for candidate B at cost c B 0, or to remain undeclared. 2. Candidate A observes citizens declarations with probability γ [0, 1]. 3. The election winner is decided (potentially influenced by declarations). 4. If clientelist candidate A wins, she distributes rewards r A 0 to all citizens observed to declare support for her. If B wins, no rewards are distributed. With respect to the factors that affect citizens expected utility, we make several assumptions. First, citizens have heterogeneous ideological preferences, ranging from a strong preference for A to a strong preference for B: x i (, ) is citizen i s ideological gain (if 12

14 positive) or loss (if negative) from A s election victory. 16 Second, clientelist candidate A distributes rewards (r A ) after the election, to citizens observed (with probability γ) to declare support for her during the campaign. Third, declaring involves candidate-specific costs (c A and c B ). These costs include material costs such as obtaining and placing a banner on one s house, or traveling to a candidate s rally. In addition, they include any social costs, such as being ostracized if one declares for A in a neighborhood mostly populated by B s supporters. Fourth, citizens may receive expressive utility from act itself of declaring in accordance with their preferences, regardless of who wins the election. We employ a dampening factor, δ [0, 1], to capture the degree to which declaring provides such expressive utility. Just as x i is citizen i s ideological gain or loss from A s election victory, δx i is her ideological gain or loss from declaring for A. And inversely, δx i is her ideological gain or loss from declaring for B. 17 Fifth, the model allows for the possibility that a citizen s declaration affects the election outcome. Candidate A s ex ante probability of winning the election is given by q [0, 1]. We assume that a declaration increases that candidate s probability of victory by α [0, min{q, 1 q}]. If a citizen declares for A (B), then A s (B s) probability of victory is increased by α and given there are two candidates, B s (A s) probability of victory declines by α This setup normalizes the ideological gain from B winning to 0; x i is the amount by which citizen i is better off or worse off when A wins relative to when B wins (due to ideological considerations). 17 All predictions are robust to the exclusion of expressive utility from the model. 18 The model is decision-theoretic, in that the behavior of citizens depends on exogenous parameters and not actions taken by other citizens. A consequence is that we assume a citizen s declaration affects the election outcome by a defined parameter, rather than by how it affects the behavior of other citizens. Nevertheless, our results do not depend on any specific assumptions about the magnitude of this effect. For experimental purposes, our decision-theoretic approach has important advantages over a game-theoretic model. 13

15 i Declared for A i Declared for B i Undeclared Candidate A Wins x i + δx i + γr A c A x i δx i c B x i Candidate B Wins δx i c A δx i c B 0 Table 1: Citizen i s Utility by Declaration Action and Election Outcome 3.2 Expected Utility Given these assumptions, Table 1 summarizes citizen i s utility contingent on her declaration action and the election outcome. We now investigate the expected utility from each declaration action, which enables us to predict how changes in parameter values affect citizens choices. The expected utility of citizen i when declaring support for candidate A is: EU i (A) = (q + α)(γr A + x i ) + δx i c A (1) There are three components of utility that the citizen receives by declaring a clientelist effect, an instrumental effect, and an expressive effect. The clientelist effect is the citizen s expected reward from declaring for A, which depends on three factors: the size of each reward that A distributes to declared supporters (r A ), the probability A observes declarations (γ), and A s probability of victory given the citizen s declaration (q + α). The instrumental effect is the citizen s expected ideological gain or loss from the election outcome, which depends on her preferences with regards to A (x i ) as well as A s probability of victory given the citizen s declaration (q + α). The expressive effect is the utility gained (lost) from the act of declaring support in accordance (discordance) with one s own ideological beliefs (x i, discounted by δ). In Equation (1), the first term includes both clientelist and instrumental effects. More specifically, it represents the incremental utility accrued from A s victory for both clientelist (γr A ) and instrumental (x i ) reasons weighted by the probability A wins given the citizen s declaration (q + α). The second term (δx i ) captures the expressive effect, and the fourth term (c A ) captures declaration costs. Next, the expected utility of citizen i when declaring for candidate B is: 14

16 EU i (B) = (q α)x i δx i c B (2) Although similar, Equations (1) and (2) exhibit several differences. First, there are instrumental and expressive effects if the citizen declares for B, but no clientelist effect because the base model assumes that only A provides rewards. An extension below examines the case of competitive clientelism in which both A and B provide rewards. Second, α is now negative, because the citizen reduces A s probability of victory when declaring for B. Third, the sign of the expressive utility term (δx i ) is now negative, because the act of declaring for B provides a utility gain (loss) to supporters of B (A). Finally, the expected utility of citizen i from remaining undeclared is: EU i ( ) = qx i (3) As shown, there are no clientelist or expressive effects if the citizen remains undeclared. However, there is an instrumental effect of remaining undeclared, which depends on her preferences with regards to A (x i ) and the ex ante probability that A wins the election (q). 3.3 Decision Rule Given these expected utilities, a straightforward approach would be to predict citizens declaration decisions deterministically. That is, one could make predictions by deriving which declaration action provides the highest expected utility to citizens according to their political preferences and other parameter values. Although deriving predictions in this manner is simple (as shown in Online Appendix A), it involves the unrealistic assumption that citizens never make mistakes during decision making. Consider the case in which a given citizen s expected utilities for two different actions for example, between declaring support for A and remaining undeclared are so similar that he is nearly indifferent between the two actions. In this scenario, minute errors in judgment about rewards, declaration costs, electoral odds, or other factors could lead to a suboptimal decision. Acknowledging this reality, many prominent studies that test formal predictions with field or laboratory experiments eschew 15

17 this type of deterministic approach and instead use probabilistic choice models (e.g., Luce, 1959; Harless and Camerer, 1994; Hey and Orme, 1994; Camerer and Ho, 1994). 19 We follow such precedent by similarly assuming that citizens can err during decision making. In particular, we allow a small degree of bounded rationality and assume that instead of optimizing, citizens choose according to a stochastic choice rule: they choose with positive probability all available actions, but are more likely to choose better alternatives. That is, in their randomization, they place more weight on actions that give them a higher payoff, and place lower weight on actions that give them a lower payoff. As is standard in the literature, we employ a Logit stochastic choice rule Predictions Formal analyses reveal how parameter changes such as increasing rewards or heightening candidate A s probability of victory affect citizens declaration choices. As discussed above, the model assumes that a citizen chooses each available action with a positive probability, with the probability distribution over actions determined by the relative expected utility from each action. When changing a parameter increases the payoff a citizen receives from one action relative to the others, she becomes more likely to choose that action. In the vast majority of cases, we can make unambiguous statements about the effects of parameter changes, which hold irrespective of contextual characteristics. Such statements are possible when changing a parameter affects the expected utility from one action, but does not affect or affects in the opposite direction the expected utility from the other two actions. 19 As other examples, probabilistic choice models have been employed to explain turnout decisions (Goeree and Holt, 2005; Levine and Palfrey, 2007; Feddersen, Gailmard and Sandroni, 2009), electoral platforms (Aragones and Palfrey, 2004), and international conflict (Signorino, 1999). 20 This choice rule is discussed in Appendix A. More generally, our predictions are robust to the use of any choice rule in which the probability of declaring for A is strictly increasing in expected utility from that action and strictly decreasing in the expected utility from the other two possible actions. 16

18 For example, when clientelist candidate A increases rewards (i.e., r A increases) and other variables are held constant, all citizens expected utility from declaring for A increases, but there is no effect on citizens expected utilities from declaring for B or remaining undeclared. Consequently, for citizens whose optimal action is declaring for A, making a mistake by declaring for B or remaining undeclared becomes costlier (in terms of foregone utility). And furthermore, for citizens whose optimal action is not declaring for A, making a mistake by declaring for A becomes less costly. Thus, with the stochastic choice rule, increasing r A increases the likelihood any citizen declares for A and decreases the likelihood that same citizen declares for B or remains undeclared. By contrast, in a few specified instances we cannot make unambiguous statements about the effects of a single parameter change, because its impact depends on the level of other parameters. In those cases, predictions are provided in Section 4 below, based on parameter values designated when implementing the experiment. Overall, the model provides the following predictions, which are thoroughly derived in Appendix A and tested experimentally below: H1 Reward Size: As A provides larger rewards, declarations for A increase ( ) declarations for B decrease πi (B) r A < 0, and non-declarations decrease ( ) πi (A) r A > 0, ( ) πi ( ) r A < 0. H2 ( Lopsided Election: ) As A s probability of winning increases, declarations for A increase πi (A) > 0. Effects on non-declarations and declaring for B depend on parameters. q H3 Social Cost: As the social ( cost of declaring ) for clientelist candidate( A increases, ) declarations for A decrease πi (A) c A < 0, declarations for B increase πi (B) c A > 0, and ( ) non-declarations increase πi ( ) c A > 0. ( ) H4 Monitoring: As A s monitoring ability increases, declarations for A increase πi (A) > 0, γ ( ) ( ) declarations for B decrease πi (B) < 0, and non-declarations decrease πi ( ) < 0. γ γ H5 Expressive Utility: As the utility of declaring in accordance with preferences increases, ( declarations for A increase among A s supporters, ) but decrease among B s supporters πi (A) > 0 if x δ i > 0; π i(a) < 0 if x δ i < 0. Declarations for B increase among B s sup- 17

19 ( porters, but decrease among A s supporters πi (B) δ Declarations by indifferent citizens are unaffected. ) > 0 if x i < 0; π i(b) < 0 if x δ i > 0. H6 Election Influence: As the election influence of declaring increases, declarations for A increase among A s supporters ( and indifferent citizens. Declarations ) for B also decrease among these citizens πi (A) > 0 and π i(b) < 0 if x α α i 0. Effects on declarations by B s supporters depend on parameter values. 3.5 Extension with Competitive Clientelism Given that competitive clientelism is observed in some contexts, a first extension adapts the base model so that both candidates can distribute rewards. As before, if candidate A wins, she distributes rewards r A 0 to all citizens observed to declare support for A. But now, if B wins, that candidate also distributes rewards r B 0 to all citizens observed to declare support for B. In this extension, the expected utilities of citizen i from declaring for A and from remaining undeclared are unchanged from the base model. However, the expected utility of citizen i from declaring for candidate B becomes: EU i (B) = (q α)x i + (1 q + α)γr B δx i c B (4) Whereas Equation 2 only included instrumental and expressive effects, Equation 5 also includes a clientelist effect (in the second term, which is new). Applying the steps elaborated above, formal analysis yields the following predictions (derived in Appendix A): H7 Competitive ( Clientelism: ) As candidate B provides( larger rewards, ) declarations for A decrease πi (A) r B < 0, declarations for B increase πi (B) r B > 0, and non-declarations ( ) decrease πi ( ) r B < Extension with Punishments A second extension adapts the base model to allow candidate A to punish citizens who declared against her. As in the initial setup, if candidate A wins, she distributes rewards r A 0 to all citizens observed to declare support for A. But now, if candidate A wins, 18

20 she also imposes punishments p A 0 on all citizens observed to declare support for B. In this extension, the expected utilities of citizen i from declaring for A and from remaining undeclared are unchanged from the base model. However, the expected utility of citizen i from declaring for candidate B becomes: EU i (B) = (q α)(x i γp A ) δx i c B (5) Similar to the competitive clientelism extension, Equation 6 adds a clientelist effect to Equation 2, but it involves negative instead of positive inducements. Formal analysis provides the following predictions (derived in Appendix A): H8 Punishments: As the ( clientelist) candidate A imposes greater( punishments, ) declarations for A increase πi (A) p A > 0, declarations for B decrease πi (B) p A < 0, and nondeclarations increase πi ( ) ( ) p A > 0. 4 Test of Predictions in Brazil 4.1 Experimental Design The formal analyses developed above provide intriguing hypotheses, but to what extent do they offer meaningful predictions about human behavior? If citizens are exposed to the model s conditions, their declaration choices might be entirely unaffected by rewards, political competition and other factors or they may change in unpredicted ways. Such findings would cast serious doubt on the assumptions of our model. By contrast, if citizens tend to respond as predicted, it would heighten confidence in our theoretical insights about how and why clientelism influences political expression beyond the ballot box. To investigate whether the model provides meaningful predictions about human behavior, we developed experiments fielded in both Brazil and the United States. Controlled experiments play an important role in rigorously testing predictions from formal models in the social sciences (Falk and Heckman, 2009; McDermott, 2002). Unlike observational studies, they allow the 19

21 researcher to manipulate variables systematically while holding all other conditions constant, thereby helping to test mechanisms and causal effects. This high degree of control over the decision making environment enables us to test, with significant precision, hypotheses elaborated in the prior section. Given that external validity is always an important concern with experiments, we focused our testing efforts on Brazil a country where we had already observed patterns of declared support (see Section 2). We also sought to obtain a large subject pool from across the nation, and thus recruited 1,259 participants from 1,061 municipalities. To recruit participants, we broadcast advertisements on Facebook in October-December 2016, following an established strategy employed in Brazil (Samuels and Zucco, 2013, 2014; Boas, 2014). Facebook s impressive reach in the nation makes it a particularly useful tool for recruiting subjects. Brazil is Facebook s third-largest market globally, with 123 million registered Facebook users in 2017, compared to an overall population of 207 million. 21 Of course, Facebook is by no means perfectly representative of Brazil s population, and certain types of users may be more inclined to click on advertisements. As such, we advertised more extensively to specified demographic subgroups, particularly women and the elderly. As shown in Appendix B, our experimental sample mirrors Brazil s overall population fairly closely with respect to gender, age, region and urban/rural mix. Given the paucity of research on clientelism outside of Brazil s largest cities, we displayed advertisements in both urban and rural areas of municipalities with populations up to 250,000 citizens. This inclusion criterion captures 98.3 percent of all municipalities, with 59.7 percent of the nation s population. Our sample proved to be quite familiar with clientelism: beyond findings discussed in Section 2, 87.3 percent of participants thought that clientelist benefits were distributed frequently or very frequently by candidates in their municipalities, and 14.4 percent reported that they had themselves received such handouts in Brazil s Technology Start-Ups Lure Capital Back After Recession, Financial Times, 5/15/2017; IBGE,

22 Many Brazilians do not use electronic payments, so subjects were recruited and incentivized with a lottery, in which we awarded a total of four iphones. After clicking on our advertisement, Facebook users were redirected to a separate web page, consented to participation, and commenced the experiment. Following Berinsky, Margolis and Sances (2014), we inserted two screener questions at different points of the survey, enabling us to control for participant attentiveness in multivariate analyses. The experiment elicited participants willingness to declare support for fictitious candidates, employing incentives to manipulate clientelist inducements and preferences about candidates. As incentives, participants earned additional lottery tickets, thereby increasing their chance of winning an iphone. Subjects could expend a small number of tickets to declare support for one of two candidates (A or B) by displaying a corresponding flag on her fictitious home, potentially affecting the election outcome. We induced political preferences for the two fictitious candidates using a standard methodology in experimental economics: a reward mechanism in which different electoral outcomes generate different monetary values (Induced Value Theory, Smith 1976). In particular, we induced a stronger ideological affiliation with a candidate by increasing the number of iphone lottery tickets a citizen received from that candidate s victory (regardless of whether she declared support). Subjects were assigned randomly to one of seven partisan types in which each type was induced to have distinct preferences, ranging from a strong preference for candidate A to a strong preference for candidate B. 22 We introduced clientelist rewards by increasing the number of lottery tickets received if a citizen declared for clientelist candidate A and A won the election; competitive clientelism is also explored below. As shown in the appendix, before subjects chose whether and for whom to declare, they viewed a simple vignette communicating information associated with 22 The sample includes 186 subjects assigned to be strong A supporters, 193 moderate A supporters, 193 weak A supporters, 175 indifferent citizens, 171 weak B supporters, 162 moderate B supporters, and 179 strong B supporters (with x i = 30, 15, 5, 0, 5, 15, and 30, respectively). 21

23 each choice, including rewards, declaration costs, candidates probability of victory, and the probability that declarations are observed. Once a subject submitted her choice (declare support for A; declare support for B; or no declaration), the election winner was determined by the computer using the odds resulting from the citizen s declaration decision. The identity of the election winner and the resulting clientelistic rewards (if any) determined the subject s earnings for each election. To test model predictions, the experiment employed ten distinct treatments, each involving a fictitious election with different contextual characteristics. Using a within-subject design, each participant was exposed to all ten treatments. This design has important advantages over a between-subject design. First, it enables us to control for unobserved individual characteristics that may affect citizens choices, such as risk aversion and cognitive abilities, through the use of fixed effects in regression analyses. And second, it increases the statistical power of our analyses because each participant contributes an observation for each of the ten treatments. However, observations are not independent in a within-subject design, and it is possible that participants behave differently in later treatments. Potential reasons are carry over effects stemming from fatigue or practice, or demand effects stemming from subjects attempts to satisfy what they perceive are the experimenter s expectations. We address such concerns by randomizing the order in which subjects observe treatments, and by controlling for within-subject error correlation in regression analyses. In the Baseline Clientelism treatment, subjects were presented with a close election between a clientelist candidate who delivers rewards to declared supporters once elected, and a non-clientelist candidate who delivers no such rewards. We use behavior in this baseline as a benchmark to evaluate the effect of the treatment variables. A fundamental attribute of the experimental setup is that other treatments modify only one factor at a time, thereby 22

24 leaving all other elements of the decision environment constant. 23 Such comparisons enable us to draw inferences about the causal link between contextual factors and citizen behavior. In the No Clientelism treatment, we remove clientelist rewards for declaration, but maintain all other elements identical to the Baseline treatment. Comparing these two treatments enables us to elicit the causal impact of clientelism on the prevalence of declared support, testing Hypothesis H1. Next, the Lopsided Election treatment tests Hypothesis H2 by increasing the clientelist candidate s probability of victory from 50 to 80 percent, before declarations of support. To examine Hypothesis H3, the Social Cost treatment investigates the scenario in which declaring for the clientelist candidate involves greater costs than declaring for the opposition. Testing Hypothesis H4, the Low Monitoring treatment adapts the Baseline Clientelism treatment to consider the case in which the clientelist candidate has a lower ability to observe declarations. In the Expressive Utility treatment, we examine Hypothesis H5 by introducing a benefit from declaring in accordance with one s preferences and a cost from declaring against one s preferences (regardless of the election outcome). In the No Election Influence treatment, we study the case where declaring for a candidate has no effect on that candidate s probability of winning the election, testing Hypothesis H6. With the Competitive Clientelism treatment, we examine Hypothesis H7 by considering the scenario in which both candidates distribute rewards to their own declared supporters. Finally, we use the Punishment & Reward and Punishment Only treatments to investigate effects when the clientelist candidate, if elected, punishes citizens who declared for the opposition, testing Hypothesis H8. The Punishment & Reward treatment leaves in place the baseline s rewards, while the Punishment Only treatment eliminates them. To summarize the experimental design, Table 2 shows the parameters used in these ten treatments. 23

25 Table 2: Brazil Parameters for Experimental Design Treatment r A q c A γ r B p A δ α c B Baseline Clientelism No Clientelism Lopsided Election Social Cost Low Monitoring Competitive Clientelism Punishment Only Clientelism and Punishment Expressive Utility No Election Influence Note: r A and r B are rewards offered by candidates A and B, respectively; q is A s ex-ante probability of winning the election; c A and c B are costs of declaring for each candidate, γ is the probability declarations are observed; p A is the punishment imposed by A; δ reflects the degree to which declaring provides expressive utility; and α is the impact of declarations on the electoral odds. Red text indicates parameters that differ from Baseline Clientelism. 4.2 Experimental Results In our Brazilian experiment, across all treatments and partisan types, subjects most frequently chose to declare for the clientelist candidate. That is, 44.2 percent of 12,590 total observations involved declarations for candidate A. In addition, 32.4 percent were declarations for opposition candidate B, and 23.4 percent were decisions to remain undeclared. Each of 1,259 participants who completed the experiment made ten declaration decisions; all findings below are robust when including subjects who did not complete the experiment. To provide a descriptive overview, Figure 3 shows the distribution of declaration decisions, by partisan type, for the ten experimental treatments. Unsurprisingly given the structure of incentives, participants induced to prefer candidate A were most likely to declare for A (i.e., the solid lines slope downward). In addition, participants induced to prefer candidate B were most likely to declare for B (i.e., the dashed lines slope upward). More interestingly, 23 As discussed below, in treatments involving punishments, benchmarks other than Baseline Clientelism are employed so that only one environmental factor is modified at a time. 24

26 Figure 2: Brazil Declaration Choices of Participants, by Treatment Baseline Clientelism No Clientelism Lopsided Election Social Cost Low Monitoring Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Partisan Type Partisan Type Partisan Type Partisan Type Partisan Type Competitive Clientelism Punishment Only Clientelism & Punish Expressive Utility No Election Influence Partisan Type Partisan Type Partisan Type Partisan Type Partisan Type Declare for A Declare for B No Declaration Note: For each treatment, figures reflect the share of participants in Brazilian experiment (N = 1,259) who declared for A, declared for B, and did not declare. Shares are shown for each partisan type labeled on horizontal axes: (1) strong A supporter, (2) moderate A supporter, (3) weak A supporter, (4) indifferent citizen, (5) weak B supporter, (6) moderate B supporter, and (7) strong B supporter. Participants were randomly assigned to partisan types, and induced to hold such preferences for the fictitious candidates. declaration choices varied substantially across treatments in line with formal predictions. For instance, when compared to the Baseline Clientelism treatment, fewer subjects in each partisan type declared for A when no rewards were distributed (No Clientelism), while more declared for A when that clientelist candidate was heavily favored to win the election (Lopsided Election). In the Competitive Clientelism treatment, which involved the scenario of both candidates offering contingent rewards, more subjects across the political spectrum declared for opposition candidate B than with Baseline Clientelism. As examined below, 25

27 more rigorous analyses reveal a wide range of significant differences that are consistent with predictions of our theoretical model. In order to test predictions, we first examine how the overall proportion of subjects undertaking each declaration action varies across treatments. These proportions are shown in Table 3, with declarations for clientelist candidate A in Panel A, declarations for opposition candidate B in Panel B, and non-declarations in Panel C. To foreshadow results, findings in this table are significant and consistent with 16 of the model s 21 unconditional predictions (76 percent), including all predictions about declarations for clientelist candidate A. For two other treatments in which the model yields different predictions across partisan types, we provide a more nuanced discussion below of proportions among subgroups of participants. After employing differences in proportions to assess the model s predictions, we then demonstrate that findings remain robust when conducting multivariate regressions and examining within-subject variation across treatments. First, we examine how citizens declarations change in response to clientelist inducements. The No Clientelism treatment reveals the level of sincere declarations; that is, the level of declared support in the absence of rewards, given the structure of incentives in the experiment. Any differences between No Clientelism and our benchmark treatment, Baseline Clientelism, reflect induced declarations. Recall that when clientelist candidate A eliminates rewards, Hypothesis H1 predicts that declarations for A decrease, declarations for B increase, and non-declarations increase. Experimental findings in Table 3 corroborate all three predictions and reveal how inducements shape public expressions of political support. In Panel A, declarations for A decrease from 46.0 percent in Baseline Clientelism to 37.1 percent in No Clientelism. In Panel B, declarations for B increase from 30.5 to 33.8 percent. And in Panel C, non-declarations increase from 23.5 to 29.1 percent. These differences are significant at the.01,.05, and.01 levels, respectively. We next turn to political competition, a contextual characteristic that can shape the nature of clientelist exchanges (Kitschelt and Wilkinson, 2007, 28; Corstange, 2017, 2). How 26

28 Table 3: Brazil Declaration Choices by Treatment Panel A: Declaration for Clientelist Candidate A Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Decreases Lopsided Election Increases Social Cost Decreases Low Monitoring Decreases Competitive Clientelism Decreases Punishment Only Increases Clientelism and Punishment Increases Panel B: Declaration for Opposition Candidate B Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Increases Lopsided Election Decreases Social Cost Increases Low Monitoring Increases Competitive Clientelism Increases Punishment Only Decreases Clientelism and Punishment Decreases Panel C: No Declaration Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Increases Lopsided Election Decreases Social Cost Increases Low Monitoring Increases Competitive Clientelism Decreases Punishment Only Increases Clientelism and Punishment Decreases Notes: 1,259 observations. Predictions aggregate across all partisan types. p-values refer to one-tailed difference in proportions Z test. All results are robust to using two-tailed tests, with the exception of Low Monitoring in Panel A (p-value: 0.138) and Lopsided Election in Panel C (p-value: 0.125). To ensure analyses isolate the effect of a single parameter change: (1) Punishment Only is compared to No Clientelism, and (2) Clientelism and Punishment is compared to Punishment Only. Results are robust to including subjects who completed only some treatments. 27

29 might this factor influence citizens choices about whether to express political preferences publicly? To explore this question, we test how declarations change if clientelist candidate A is heavily favored to win the election. Hypothesis H2 suggests that declarations for A should increase; furthermore, given the parameter values employed in the experiment, the model predicts that declarations for B and non-declarations both decline. A comparison of the Lopsided Election and Baseline Clientelism treatments in Table 2 confirms all three predictions. Declarations for A increase from 46.0 percent in Baseline Clientelism to 51.6 percent in Lopsided Election. In addition, declarations for B decrease from 30.5 to 27.4 percent, while non-declarations decrease from 23.5 to 21.0 percent. These differences are significant at the.01,.05 and.10 levels, respectively. Even in contexts without clientelism, the social context of a neighborhood can discourage or encourage various forms of political expression (Huckfeldt, 1979, 581), and citizens decisions to display yard signs and bumper stickers in particular can be influenced by neighbors (Huckfeldt and Sprague, 1992, 78; Makse and Sokhey, 2014, 211). What if citizens who declare for clientelist candidate A suffer social costs imposed by their neighbors? In this case, Hypothesis H3 predicts that declarations for A decrease, while declarations for B and non-declarations increase. Comparing the Social Cost and Baseline Clientelism treatments provides empirical support for two of these three predictions. In Table 2, declarations for A decrease from 46.0 percent in Baseline Clientelism to 40.2 percent in Social Cost. By contrast, declarations for B increase from 30.5 to 35.3 percent. Both of these differences concord with expectations and are significant at the.01 level. For non-declarations, however, the difference between the two treatments is statistically indistinguishable (though it has the predicted sign). Many studies of clientelism suggest that elites monitor citizens in order to minimize opportunistic defection. Given that such efforts involve considerable organizational infrastructure and resources (Stokes, 2005; Kitschelt and Wilkinson, 2007), the capacity to monitor citizens may vary considerably. How does a politician s monitoring capability affect voters 28

30 decisions to express political preferences publicly? According to Hypothesis H4, when clientelist candidate A s ability to monitor declarations declines, the model predicts: declarations for A decrease, declarations for B increase, and non-declarations increase. Although differences in proportions between the Low Monitoring and Baseline Clientelism treatments have the predicted sign in all three panels, only in Panel A is this difference statistically significant. More specifically, declarations for A decrease from 46.0 percent in Baseline Clientelism to 43.1 percent in Low Monitoring (p =.069). In numerous countries, clientelism is not a monopolistic phenomenon, but rather involves multiple machines providing rewards in the same localities. For instance, Kitschelt s (2013) cross-national survey reveals that competitive clientelism exists in nations such as Hungary, Ghana, Indonesia, Nigeria and Taiwan. How would such competitive clientelism affect citizens choices to express preferences publicly? To investigate this question, we consider the case in which both candidates A and B provide contingent rewards. Hypothesis H7 predicts that when both candidates distribute rewards, declarations for A decrease, declarations for B increase, and non-declarations decrease. Experimental results corroborate all three predictions regarding this competitive scenario. Declarations for A decrease from 46.0 percent in Baseline Clientelism to 41.2 percent in Competitive Clientelism. Also as expected, declarations for B increase from 30.5 to 38.7 percent and non-declarations decrease from 23.5 to 20.1 percent. These differences are significant at the.01,.01 and.05 levels, respectively. Although studies of clientelism tend to focus on positive inducements, elites in some contexts employ punishments to influence voter behavior. For example, Mares and Young (2016, 268) emphasize that negative inducements play an important role in clientelism in Africa, noting that a substantially greater share of Afrobarometer respondents report fear of punishments than offers of rewards. How do citizens respond in terms of declared support when punishments are used? We first investigate a Punishment Only treatment, in which A exclusively punishes citizens who declared for B. Comparing results to the No Clientelism treatment isolates the causal effect of introducing punishments; that is, any 29

31 differences between the two treatments reflects declarations suppressed by negative inducements. 24 Hypothesis H8 predicts that when A introduces punishments, declarations for A increase, declarations for B decrease, and non-declarations increase. Findings are consistent with the first two of these predictions. Declarations for A increase from 37.1 percent in No Clientelism to 42.4 percent in Punishment Only, a difference significant at the.01 level. Also as predicted, declarations for B decrease from 33.8 to 31.1 percent (p =.068). However, contrary to expectations, non-declarations decrease from 29.1 to 26.5 percent (p =.077). These findings reveal how punishments can suppress public expressions of political support, but it is also important to recognize that elites often mix their use of positive and negative inducements (Mares and Young, 2016, 268). How are voters declaration choices affected if candidate A combines punishments with rewards? To isolate the effect, we compare the aforementioned Punishment Only treatment with a Clientelism and Punishment treatment, in which A rewards its own declared supporters and punishes those who declared for B. Building on our discussion of Hypothesis H1 above, A s addition of rewards is expected to increase declarations for A, decrease declarations for B, and decrease non-declarations. Results dovetail with two of these predictions. Declarations for A increase from 42.4 percent in Punishment Only to 47.2 percent in Clientelism and Punishment. Also as expected, nondeclarations significantly decrease from 26.5 to 21.6 percent between these two treatments. Both differences are significant at the.01 level. By contrast, there is no perceptible difference with respect to declarations for B. Studies of voting behavior often point to expressive utility as a reason why citizens vote even if they are unlikely to influence the election outcome (e.g., Downs, 1957, Riker and Ordeshook, 1968, Fiorina, 1976). We examine how these two factors expressive utility and election influence affect citizens decisions to express political preferences beyond 24 As shown in Table 2, this comparison isolates the effect because it changes only one parameter value. By contrast, comparing Punishment Only and Baseline Clientelism would not isolate the causal effect as it changes two parameter values simultaneously (removing rewards and introducing punishments). 30

32 the ballot box in contexts with clientelism. Unlike factors in Table 3, the model suggests that these two factors impact on declarations depends on citizens political preferences. Hypothesis H5 predicts that as citizens obtain greater utility from the act of declaring in accordance with their preferences, declarations for candidate A (B) increase among A s (B s) supporters, but decrease among B s (A s) supporters. As shown in Online Appendix B, the treatment that introduces expressive utility generally corroborates these predictions, though splitting the sample reduces statistical precision. Among supporters of A, declarations for A increase from 53.5 percent in Baseline Clientelism to 58.4 percent in Expressive Utility, while declarations for B fall from 27.4 to 26.2 percent (p =.048 and.107, respectively). Results also conform with expectations for supporters of B: declarations for B increase from 33.8 percent in Baseline Clientelism to 45.1 percent in Expressive Utility, while declarations for A fall from 36.9 to 33.2 percent (p =.000 and.107, respectively). Furthermore, experimental results corroborate another formal prediction about expressive utility: declaration actions by indifferent citizens are unaffected. 25 While research in the United States suggests that many survey respondents believe that lawn signs can influence votes (Makse and Sokhey, 2014, 210), what effects are expected in contexts where such forms of declaration have no influence whatsoever on the election outcome? As mentioned, Hypothesis H6 provides heterogeneous predictions. Among A s supporters and indifferent citizens, as the influence of declaring on the election falls, declarations for candidate A (B) should decrease (increase). And among B s supporters, as this influence falls, declarations for candidate A (B) should increase (decrease). 26 As shown in Online Appendix B, the experiment yields perceptible differences with predicted signs for each effect, but most are statistically indistinguishable from zero which is unsurprising given the loss of precision when examining subsamples. Among A s supporters and indiffer- 25 However, this expected null effect likely stems from the smaller sample size; subjects were randomly assigned to seven partisan types, of which only one is indifferent. 26 Predictions in this sentence are based on parameter values in the experimental design. Recall that Hypothesis H6 s predictions about declarations by B s supporters depend on parameter values. 31

33 ent citizens, declarations for A decrease from 52.2 percent in Baseline Clientelism to 49.5 percent in No Election Influence, while declarations for B increase from 28.2 to 30.1 percent (p =.150 and.213). Also in line with predictions, when examining B s supporters, declarations for A increase from 36.9 percent in Baseline Clientelism to 40.4 percent in No Election Influence, and declarations for B decrease from 33.8 to 28.5 percent (p =.124 and.034). Thus far, we have estimated treatment effects by comparing differences in proportions. Such findings, which were predominantly consistent with formal predictions, do not rely on parametric assumptions. To show robustness, we next conduct multivariate regressions that control for key variables and examine within-subject variation across treatments. This step involves pooling observations across treatments and adopting a basic parametric structure. More specifically, we employ logistic regressions and assume that declaration decisions are a function of each treatment as well as political preferences, survey round, and screener performance. 27 With respect to these covariates, political preferences (x i ) for one of the fictitious candidates were induced with payoffs (see Section 4.1). Survey round is included to control for the possibility that experience within the experiment affects declaration decisions. The screener performance variable, which indicates how many screener questions the subject answered correctly (0 to 2), controls for respondents level of attentiveness. These screener questions follow the work of Berinsky et al. (2014) and were inserted at different points of the survey. Some specifications include subject fixed effects to investigate within-subject variation across treatments. This step controls for any characteristics that do not vary across treatments for a given participant, such as age, education and gender. 28 For 20 of the model s 21 unconditional predictions, findings from multivariate regressions are similar to differences of proportions with respect to the accuracy of predictions and statistical significance. Table 4 focuses on rewards; coefficients for each treatment indicate marginal effects and are shown in comparison to Baseline Clientelism (the excluded 27 Findings are also robust when using multinomial logit. 28 Specifications with fixed effects omit the political preferences and screener performance variables, which do not vary within a given participant. 32

34 Table 4: Brazil Estimates of Average Treatment Effects, Rewards (Logit) Declare for A Declare for B No Declaration Treatment (1) (2) (3) (4) (5) (6) No Clientelism *** *** 0.032** 0.049** 0.056*** 0.104*** (0.017) (0.021) (0.015) (0.021) (0.015) (0.024) Lopsided Election 0.058*** 0.080*** ** ** * ** (0.016) (0.021) (0.015) (0.021) (0.013) (0.023) Social Cost *** *** 0.047*** 0.072*** (0.016) (0.021) (0.016) (0.022) (0.014) (0.023) Low Monitoring * (0.017) (0.021) (0.015) (0.021) (0.014) (0.024) Competitive Clientelism *** *** 0.080*** 0.121*** *** *** (0.016) (0.021) (0.016) (0.022) (0.013) (0.023) Round 0.007*** 0.012*** *** *** (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) Partisan Type 0.003*** *** *** (0.000) (0.000) (0.000) Screener *** (0.010) (0.010) (0.009) Subject Fixed Effects No Yes No Yes No Yes Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. Baseline Clientelism is the excluded treatment category, so that coefficients report differences from that baseline. Robust standard errors are reported, clustered by subject in columns 1, 3, and 5. 33

35 treatment category). Similar to the findings above, predictions for all three declaration actions are confirmed for No Clientelism, Lopsided Election, and Competitive Clientelism. In addition, the Social Cost treatment again confirms two of three predictions. By contrast, Low Monitoring is the only treatment in which results are not robust to multivariate analyses. Whereas one prediction conforms to expectations with differences of proportions, this only holds in specifications with fixed effects (at the.10 level). Next, Table 5 examines punishments. It employs the identical methodology as Table 4, but compares with different treatments to ensure that only one aspect of the decision environment is changed at a time. As with differences in proportions, multivariate regressions reveal that Punishment Only and Clientelism and Punishment treatments are again consistent with two of three predictions. Moreover, Appendix B shows that multivariate analyses are also similar to unadjusted differences when considering the two factors with heterogeneous predictions (Expressive Utility and No Election Influence). More broadly, results from the Brazil experiment predominantly conform to theoretical expectations, whether analyses employ differences of proportions or multivariate regressions. 5 Test of Predictions in U.S. 5.1 Experimental Design To show robustness of predictions, we also conducted a laboratory experiment in the United States. At the outset, it should be emphasized that the Brazil experiment serves as our primary test: it is superior with respect to external validity, it involves far more participants (1,259 vs. 144 in the U.S.), it includes more treatments (ten vs. eight in the U.S.), and it offers a more refined test of predictions by inducing more partisan types (seven vs. five in the U.S.). Nevertheless, the US experiment also offers several advantages, and thereby further heightens confidence in our theoretical insights about how and why clientelism influ- 34

36 Table 5: Brazil Estimates of Average Treatment Effects, Punishments (Logit) Declare for A Declare for B No Declaration Treatment (1) (2) (3) (4) (5) (6) Punishment Only 0.053*** * * (0.016) (0.015) (0.014) Clientelism & Punishment 0.047** *** (0.020) (0.018) (0.017) Round 0.007** 0.009*** *** *** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Partisan Type 0.002*** 0.004*** *** *** *** (0.001) (0.001) (0.001) (0.000) (0.001) (0.000) Screener ** ** 0.034*** 0.029*** (0.013) (0.011) (0.013) (0.011) (0.012) (0.010) Subject Fixed Effects No No No No No No Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. To isolate causal effects, Punishment Only specifications employ No Clientelism as the excluded category, and Clientelism & Punishment specifications employ Punishment Only as the excluded category. Robust standard errors are reported, clustered by subject. Online Appendix B shows that findings are robust with subject fixed effects. 35

37 ences political expression beyond the ballot box. First, its implementation in a supervised computer laboratory, as opposed to online, offers greater testing control. For example, the US experiment offered fewer potential distractions when respondents received instructions and made declaration decisions. Second, the physical presence of subjects in a laboratory facilitated the use of an alternative incentive scheme: the payment of cash at the end of the experiment, instead of the iphone lottery employed in Brazil. To the extent that cash provides stronger incentives for participants, its use should facilitate the detection of treatment effects. And third, given that the subject pool in the U.S. is less likely to be familiar with clientelism, participants may potentially be less affected by social desirability bias. The fact that results in both experiments predominantly accord with predictions despite substantially different testing conditions lends credence to our theoretical framework. However, the magnitude of findings should not be compared across settings, not least because the US experiment involved fewer partisan types, different parameter values, and an alternative incentive scheme. The US experiment was conducted at the Columbia Experimental Laboratory for the Social Sciences. All participants were undergraduate or graduate students at Columbia University, recruited from a database of volunteer subjects. We held six sessions, with 24 participants independently answering questions in each session, totaling 144 subjects. No subject participated in more than one session. Instructions were first read aloud; then, subjects reviewed the same instructions again on their computer screens before commencing the experiment. In the experiment, subjects made declaration decisions in 16 fictitious elections. More specifically, participants completed two rounds of eight randomly ordered treatments, employing parameter values described in Online Appendix C. 29 The US experiment was conducted in person, so as mentioned we were able to use cash as incentives. Each participant was paid his or her total earnings across all elections in cash at the end of the session. Av- 29 Based on feedback after the US experiment, we included two more treatments in the Brazil experiment. The US setting provided approximately 40 minutes with participants in the laboratory, enabling two rounds. 36

38 erage earnings per participant, excluding the $5 show-up fee, were $23.16 (with a standard deviation of $6.47). 5.2 Experimental Results Similar to the Brazil experiment, we first test predictions by examining how the overall proportion of subjects undertaking each declaration action varies across treatments. These proportions are shown in Table 6, with declarations for clientelist candidate A in Panel A, declarations for opposition candidate B in Panel B, and non-declarations in Panel C. As shown, for 15 of 18 unconditional predictions (83 percent) tested in the US experiment, findings are significant and consistent with predictions. This section discusses these findings, as well as another treatment with heterogeneous predictions across partisan types. Note that two treatments discussed in the context of Brazil, Social Cost and Expressive Utility, were untested in the US experiment. 30 All findings in the US experiment are robust to using multivariate regressions, as conducted in the Brazil experiment. 31 When examining the effect of clientelist rewards on declarations, all three findings in Table 6 are consistent with predictions in Hypothesis H1. In Panel A, declarations for A decrease from 76.7 percent in Baseline Clientelism to 18.1 percent in No Clientelism. Moreover, declarations for B increase from 17.7 to 29.2 percent in Panel B, and non-declarations increase from 5.6 to 52.8 percent in Panel C. All three differences are statistically significant (at the.01 level). With regards to clientelist candidate A being heavily favored to win the election, the US experiment corroborates one of three predictions. In line with Hypothesis H2, declarations for B decreased from 17.7 percent in Baseline Clientelism to 13.5 percent in Lopsided Election (significant at the.10 level). Declarations for A increased as predicted but the difference 30 We conducted the US experiment first, and later examined these two factors in Brazil. 31 Online Appendix C shows these regression analyses. 37

39 Table 6: United States Declaration Choices by Treatment Panel A: Declaration for Clientelist Candidate A Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Decreases Lopsided Election Increases Low Monitoring Decreases Competitive Clientelism Decreases Punishment Only Increases Clientelism and Punishment Increases Panel B: Declaration for Opposition Candidate B Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Increases Lopsided Election Decreases Low Monitoring Increases Competitive Clientelism Increases Punishment Only Decreases Clientelism and Punishment Decreases Panel C: No Declaration Treatment Prediction Proportion Difference p-value As Predicted? Baseline Clientelism Baseline Baseline Baseline Baseline No Clientelism Increases Lopsided Election Decreases Low Monitoring Increases Competitive Clientelism Decreases Punishment Only Increases Clientelism and Punishment Decreases Notes: 288 observations. Predictions aggregate across all partisan types. p-values refer to one-tailed difference in proportions Z test. All results are robust to using two-tailed tests, except Lopsided Election in Panel B (p-value: 0.169). To ensure analyses isolate the effect of a single parameter change: (1) Punishment Only is compared to No Clientelism, and (2) Clientelism and Punishment is compared to Punishment Only. 38

40 is insignificant (p =.182). For non-declarations, contrary to expectations, participants responded almost identically when exposed to the two treatments. When clientelist candidate A has less capacity to monitor declarations, all three US findings significantly concord with predictions. As predicted in Hypothesis H4, declarations for A fall from 76.7 percent in Baseline Clientelism to 47.9 percent in Low Monitoring. Also as expected, declarations for B increase from 17.7 to 24.3 percent, and non-declarations increase from 5.6 to 27.8 percent. These findings are significant at the.01,.05, and.01 levels, respectively. In the scenario where both candidates provide rewards, all three US findings dovetail with formal predictions. As suggested by Hypothesis H7, declarations for A decrease from 76.7 percent in Baseline Clientelism to 49.0 percent in Competitive Clientelism. Moreover, declarations for B increase from 17.7 to 49.0 percent, and non-declarations decrease from 5.6 to 2.1 percent. These findings are significant at the.01,.01, and.05 level, respectively. When clientelist candidate A uses punishments instead of rewards, results from the US experiment confirm two of three predictions. As in Brazil, to isolate the effect of punishments, we compare a treatment in which A uses neither rewards nor punishments, to a treatment in which A punishes citizens who declared for B. 32 In accordance with Hypothesis H8, declarations for B decrease from 29.2 percent in No Clientelism to 12.2 percent in Punishment Only. Also as predicted, non-declarations increase from 52.8 to 69.1 percent. While both of these findings are significant (at the.01 level), the two treatments have an unexpectedly imperceptible difference in declarations for A. In addition, the US experiment corroborates all three formal predictions about the scenario in which candidate A combines these punishments with rewards (Clientelism and Punishment). As in Brazil, we isolate effects by comparing this treatment to Punishment Only. As predicted, declarations for A increase from 18.8 percent in Punishment Only to 78.5 per- 32 As discussed above, comparing Punishments Only and Baseline Clientelism is inappropriate because it changes two parameter values simultaneously. 39

41 cent in Clientelism and Punishment. Also conforming with expectations, declarations for B decline from 12.2 percent to 6.9 percent, and non-declarations decreased from 69.1 to 14.6 percent. These findings are significant at the.01,.05, and.05 levels, respectively. Finally, results from the US experiment are consistent with three of four predictions regarding the case in which declarations have no influence on the election outcome. 33 Recall that predictions for this factor depend on citizens political preferences. As expected, declarations for A decrease from 93.7 percent in Baseline Clientelism to 64.4 percent in No Election Influence among A s supporters and indifferent citizens, but increase from 50.9 to 77.2 percent among B s supporters. Also conforming with expectations, declarations for B decrease from 41.2 to 6.1 percent among B s supporters. These three findings are all statistically significant (at the.01 level). By contrast, results for a fourth prediction are insignificant: whereas A s supporters and indifferent citizens are expected to increase their declarations for B, they decrease in the experiment from 2.3 to 1.1 percent (p =.21). As in the Brazilian experiment, results from the US experiment are similar when conducting multivariate regressions instead of examining differences of proportions. More specifically, the online appendix shows robustness when pooling observations across treatments and conducting logistic regressions that include controls and/or fixed effects. Overall, experiments in both Brazil and the U.S. predominantly corroborate formal predictions about declared support. 6 Discussion The present study has emphasized the role of citizen choices in clientelism, thereby redirecting the elite focus adopted by most prominent research on the topic. The phenomenon of declared support not only counters the usual depiction of citizens as passive recipients, but also underscores how their scope of choices in clientelism extends well beyond the ballot box. 33 These findings are shown in Online Appendix C. 40

42 When voters can obtain future benefits by declaring support for victorious candidates, their decisions to display campaign posters, wear political paraphernalia or attend rallies often reflect more than just political preferences. Formal analysis elaborates how and why clientelism can influence citizens choices to express political support publicly. And furthermore, it suggests how various factors affect citizens propensity to declare support in response to clientelist inducements. For example, citizens are more likely to declare support for a clientelist candidate who can distribute large rewards and monitor declarations effectively, but are less likely to do so if that candidate is electorally weak or cannot be publicly supported without social repercussions. Moreover, several other features of polities can influence political expression, such as whether citizens are punished for publicly supporting a defeated candidate, and whether there exists competition between clientelist candidates. Predictions from our theoretical model are predominantly corroborated by two experiments implemented in substantially different contexts. The primary experiment conducted with 1,259 online participants across Brazil yields findings consistent with 16 of the model s 21 unconditional predictions (76 percent). Results in another smaller experiment, conducted with 144 subjects in a supervised lab in the U.S., are consistent with 15 of the 18 unconditional predictions (83 percent) that we tested. The model thus offers meaningful predictions about human behavior in both experiments, suggesting that it provides useful theoretical insights about how and why clientelism influences political expression beyond the ballot box. These analyses not only elaborate and test the logic of declared support, but also lay the groundwork for further investigation into the role of citizens in clientelism. One important step is to examine patterns of declared support in other countries. Our fieldwork and surveys suggest a strong link between declarations and clientelism in Brazil, and this study s introduction discussed more limited evidence from Argentina, Ghana, Lebanon and Mexico. In order to test the external validity of our findings thoroughly, it would be fruitful to collect data about declared support in settings with diverse characteristics. Furthermore, while declared support is an important phenomenon, it is by no means the only action that 41

43 citizens can take to influence their receipt of contingent benefits. Important directions for future research are exploring the various modalities by which citizens can shape clientelist exchanges, and examining the conditions under which citizens are more or less motivated to undertake such actions. These topics also warrant close attention in contexts where citizens face substantial constraints to autonomous decision making, such as under some authoritarian regimes. Overall, a broader analytical lens that considers the choices of citizens and not just those of elites holds substantial promise to deepen our understanding of contingent exchanges. Given the various consequences of clientelism for both democracy and development, improving our knowledge about this phenomenon would be a significant contribution. 42

44 References Aragones, Enriqueta and Thomas R. Palfrey The Effect of Candidate Quality on Electoral Equilibrium: An Experimental Study. American Political Science Review 98(01): Archer, Ronald P The Transition from Traditional to Broker Clientelism in Colombia: Political Stability and Social Unrest. Kellogg Institute for International Studies, Working Paper #140. Auyero, Javier Poor People s Politics: Peronist Survival Networks and the Legacy of Evita. Durham, NC: Duke University Press. Berinsky, Adam J., Michele F. Margolis and Michael W. Sances Separating the Shirkers from the Workers? Making Sure Respondents Pay Attention on Self-Administered Surveys. American Journal of Political Science 58(3): Boas, Taylor C Pastor Paulo vs. Doctor Carlos: Professional Titles as Voting Heuristics in Brazil. Journal of Politics in Latin America 6(2): Camerer, Colin F. and Teck-Hua Ho Violations of the Betweenness Axiom and Nonlinearity in Probability. Journal of Risk and Uncertainty 8(2): Cammett, Melani C Partisan Activism and Access to Welfare in Lebanon. Studies in Comparative International Development 46(1): Cammett, Melani C Compassionate Communalism: Welfare and Sectarianism in Lebanon. Ithaca, NY: Cornell University Press. Cornelius, Wayne A Leaders, Followers, and Official Patrons in Urban Mexico. In Friends, Followers, and Factions: A Reader in Political Clientelism, ed. Carl H. Landé Steffen W. Schmidt, Laura Guasti and James C. Scott. Berkeley, CA: University of California Press. 43

45 Corstange, Daniel Clientelism in Competitive and Uncompetitive Elections. Comparative Political Studies, OnlineFirst. Diaz-Cayeros, Alberto, Federico Estévez and Beatriz Magaloni Strategies of Vote- Buying: Democracy, Clientelism, and Poverty Relief in Mexico. New York, NY: Cambridge University Press. Downs, Anthony An Economic Theory of Political Action in a Democracy. Journal of Political Economy 65(2): Falk, Armin and James J. Heckman Lab Experiments are a Major Source of Knowledge in the Social Sciences. Science 326(5952): Feddersen, Timothy, Sean Gailmard and Alvaro Sandroni Moral Bias in Large Elections: Theory and Experimental Evidence. American Political Science Review 103(2): Finan, Frederico and Laura Schechter Vote-Buying and Reciprocity. Econometrica 80(2): Fiorina, Morris P The Voting Decision: Instrumental and Expressive Aspects. The Journal of Politics 38(2): Gans-Morse, Jordan, Sebastian Mazzuca and Simeon Nichter Varieties of Clientelism: Machine Politics during Elections. American Journal of Political Science 58(2): Gay, Robert The Even More Difficult Transition from Clientelism to Citizenship: Lessons from Brazil. In Out of the Shadows: Political Action and the Informal Economy in Latin America, ed. Patricia Fernandez-Kelly and Jon Shefner. University Park, PA: Penn State University Press. Goeree, Jacob and Charles Holt An Explanation of Anomalous Behavior in Models of Political Participation. American Political Science Review 99(02):

46 Hall, Anthony Patron-Client Relations. The Journal of Peasant Studies 1(4): Harless, David W. and Colin F. Camerer The Predictive Utility of Generalized Expected Utility Theories. Econometrica 62(6): Hey, John D. and Chris Orme Investigating Generalizations of Expected Utility Theory Using Experimental Data. Econometrica 62(6): Hicken, Allen Clientelism. Annual Review of Political Science 14: Hilgers, Tina Clientelism in Everyday Latin American Politics. New York, NY: Palgrave Macmillan. Huckfeldt, R. Robert Political Participation and the Neighborhood Social Context. American Journal of Political Science 23(3): Huckfeldt, R. Robert and John Sprague Political Parties and Electoral Mobilization: Political Structure, Social Structure, and the Party Canvass. American Political Science Review 86(1): Huckfeldt, R. Robert and John Sprague Citizens, Politics and Social Communication: Information and Influence in an Election Campaign. New York, NY: Cambridge University Press. Keefer, Philip Clientelism, Credibility, and the Policy Choices of Young Democracies. American Journal of Political Science 51(4): Kitschelt, Herbert Democratic Accountability and Linkages Project. Durham, NC: Duke University. Kitschelt, Herbert and Steven Wilkinson Patrons, Clients, and Policies: Patterns of Democratic Accountability and Political Competition. New York, NY: Cambridge University Press. 45

47 Levine, David K and Thomas R Palfrey The Paradox of Voter Participation? A Laboratory Study. American Political Science Review 101(01): Luce, Duncan R Individual Choice Behavior: A Theoretical Analysis. New York, NY: Wiley. Makse, Todd and Anand E Sokhey The Displaying of Yard Signs as a Form of Political Participation. Political Behavior 36(1): Mares, Isabela and Lauren Young Buying, Expropriating, and Stealing Votes. Annual Review of Political Science 19: McDermott, Rose Experimental Methods in Political Science. Annual Review of Political Science 5(1): McFadden, Daniel Conditional Logit Analysis of Qualitative Choice Behavior. In Frontiers in Econometrics. New York, NY: Academic Press, pp Michelitch, Kristin G Electoral Competition and Interpartisan Economic Discrimination. Ph.D. Dissertation, New York University. Nichter, Simeon Vote Buying or Turnout Buying? Machine Politics and the Secret Ballot. American Political Science Review 102(1): Piattoni, Simona Clientelism, Interests, and Democratic Representation: The European Experience in Historical and Comparative Perspective. New York, NY: Cambridge University Press. Powell, John Duncan Peasant Society and Clientelist Politics. American Political Science Review 64(2): Riker, William H and Peter C Ordeshook A Theory of the Calculus of Voting. American Political Science Review 62(01):

48 Samuels, David and Cesar Zucco The Power of Partisanship in Brazil: Evidence from Survey Experiments. American Journal of Political Science 58(1): Samuels, David J and Cesar Zucco Using Facebook as a Subject Recruitment Tool for Survey-Experimental Research. Unpublished Manuscript. Schaffer, Frederic Charles and Andreas Schedler What Is Vote Buying? In Elections for Sale: The Causes and Consequences of Vote Buying, ed. Frederic Charles Schaffer. Lynne Reinner Publishers, pp Scott, James C Patron-Client Politics and Political Change in Southeast Asia. American Political Science Review 66(1): Signorino, Curtis S Strategic Interaction and the Statistical Analysis of International Conflict. American Political Science Review 93(02): Smith, Vernon L Experimental Economics: Induced Value Theory. The American Economic Review 66(2): Stokes, Susan C Perverse Accountability: A Formal Model of Machine Politics with Evidence from Argentina. American Political Science Review 99(3): Stokes, Susan, Thad Dunning, Marcelo Nazareno and Valeria Brusco Brokers, Voters and Clientelism. New York, NY: Cambridge University Press. Szwarcberg, Mariela Mobilizing Poor Voters: Machine Politics, Clientelism, and Social Networks in Argentina. New York, NY: Cambridge University Press. Taylor-Robinson, Michelle M Do the Poor Count? Democratic Institutions and Accountability in a Context of Poverty. University Park, PA: Penn State University Press. Verba, Sidney and Norman H Nie Participation in America. New York: Harper & Row. 47

49 Appendix A - Formal Analysis Logit Stochastic Choice Rule As described in Section 3.3, we assume that citizens choose according to a stochastic choice rule: they choose with positive probability all available actions, but are more likely to choose better alternatives. That is, in their randomization, they place more weight on actions that give them a higher payoff, and place lower weight on actions that give them a lower payoff. As is standard in the literature, we employ a Logit stochastic choice rule. In particular, the probability that citizen i chooses declaration action j = {A, B, } is: π i (j) = exp(λeu i (j)) exp(λeu i (A)) + exp(λeu i (B)) + exp(λeu i ( )) (6) Where EU i (A), EU i (B), and EU i ( ) are as in equations (1), (2), and (3) and λ [0, ) is the Logit response parameter. With λ = 0, citizens choices are insensitive to payoffs and all actions are chosen with the same probability. As λ approaches infinity, the likelihood of choosing the best action increases. For interior levels of λ, the probability of choosing each action is positive and increasing in the relative utility that the action delivers. Note that the Logit stochastic choice rule can be derived endogenously from a random utility model in which the unobserved utilities (or error terms) are independently and identically distributed extreme values (Luce, 1959; McFadden, 1974). It should be noted that although the Logit specification we employ is the most common stochastic choice rule used among studies that test predictions experimentally, our predictions hold more generally. More specifically, they are robust to using any choice rule in which the probability of declaring for A is strictly increasing in expected utility from that action and strictly decreasing in the expected utility from the other two possible actions. 48

50 Comparative Statics with Stochastic Choice As discussed in the text, when changing a parameter affects the expected utility from one action and it does not affect the expected utility from the other two actions or it affects the expected utility from the other two actions in the opposite direction, we can make an unambiguous statement on how this parameter affects the distribution of probability over actions. When this is not the case, the impact of a parameter on the probability distribution over actions depends also on the other parameters (and on the degree of noise, λ). We first consider the base model, showing comparative statics for Hypotheses 1-6. H1 Reward Size: As A provides larger rewards, declarations for A increase ( ) declarations for B decrease πi (B) r A < 0, and non-declarations decrease ( ) πi (A) r A > 0, ( ) πi ( ) r A < 0. Consider the effect of a marginal change in the rewards offered by A. EU i (A) r A = (q + α)γ EU i (B) r A = 0 EU i ( ) r A = 0 Increasing r A (keeping all other variables constant) increases EU i (A) for all citizens but does not affect EU i (B) or EU i ( ). As a consequence, increasing r A increases π i (A), decreases π i (B) and decreases π( ), as defined in equation (4) that is, it increases the likelihood any citizen declares for A at the expense of the likelihood that the same citizen declares for B or remains undeclared. H2 Lopsided Election: As A s probability of winning increases, declarations for A increase ( ) πi (A) > 0. Effects on non-declarations and declaring for B depend on parameters. q Consider the effect of a marginal change in the support for candidate A. EU i (A) q = x i + γr A EU i (B) q = x i EU i ( ) q = x i 49

51 Increasing q by a single unit, increases EU i (A) by x i + γr A units, EU i (B) by x i units and EU i ( ) by x i units. This means that declaring for A has become relatively more attractive than both other actions for any citizen. Whether a citizen is more or less likely to declare for B or remain undeclared also depends on how the relative utility between these two actions has changed which, in general, is a function of the other variables and of λ. H3 Social Cost: As the social cost of declaring for clientelist candidate A increases, declarations for A decrease πi (A) ( ) ( ) c A < 0, declarations for B increase πi (B) c A > 0, and ( ) non-declarations increase πi ( ) c A > 0. Consider the effect of a marginal change in the (social) cost of declaring for A. EU i (A) c A = 1 EU i (B) c A = 0 EU i ( ) c A = 0 Increasing c A by a single unit decreases EU i (A) for all citizens but does not affect EU i (B) or EU i ( ). As a consequence, increasing c A decreases π i (A), increases π i (B) and increases π i ( ), as defined in Equation (6) that is, it decreases the likelihood any citizen declares for A, while increasing the likelihood that the same citizen declares for B or remains undeclared. H4 Monitoring: As A s monitoring ability increases, declarations for A increase ( ) declarations for B decrease πi (B) < 0, and non-declarations decrease γ ( πi (A) ) γ > 0 ),. ( πi ( ) γ < 0 Consider the effect of a marginal change in A s ability to monitor declarations. EU i (A) γ = (q + α)r A EU i (B) γ = 0 EU i ( ) γ = 0 Increasing γ of a single unit, increases EU i (A) of (q +α)r A units, and does not affect EU i (B) and EU i ( ). This means that increasing γ makes declaration for A relatively more attractive with respect to both other actions for any citizen. 50

52 H5 Expressive Utility: As the utility of declaring in accordance with preferences increases, declarations for A increase among A s supporters, but decrease among B s supporters ( ) πi (A) > 0 if x δ i > 0; π i(a) < 0 if x δ i < 0. Declarations for B increase among B s supporters, but decrease among A s supporters πi (B) ( ) > 0 if x δ i < 0; π i(b) < 0 if x δ i > 0. Declarations by indifferent citizens are unaffected. Consider the effect of a marginal change in expressive utility. EU i (A) δ = x i EU i (B) δ = x i EU i ( ) δ = 0 Increasing δ, increases the expected utility any citizen derives from supporting her favorite candidate, decreases the expected utility she derives from supporting the other candidate, and does not affect the expected utility from remaining undeclared. As such, increasing δ increases the likelihood any citizen with x i > 0 (x i < 0) declares for A (B) and decreases the likelihood the same citizen declares for B (A). The impact of δ on the probability of remaining undeclared for citizens with x i 0 cannot be determined unambiguously. The degree of expressive utility does not affect the behavior of neutral citizens (that is, citizens with x i = 0). H6 Election Influence: As the election influence of declaring increases, declarations for A increase among A s supporters and indifferent citizens. Declarations for B also decrease ( ) among these citizens πi (A) > 0 and π i(b) < 0 if x α α i 0. Effects on declarations by B s supporters depend on parameter values. Consider the effect of a marginal change in the election influence of declaring. EU i (A) α = x i + γr A EU i (B) α = x i EU i ( ) α = 0 Increasing α by a single unit increases EU i (A) by x i + γr A units, EU i (B) by x i units and does not affect EU i ( ). For citizens with x i > 0, declaring for A has become relatively more 51

53 attractive than both other actions and declaring for B has become relatively less attractive than both other actions. Therefore, increasing α increases the likelihood these citizens declare for A and decreases the likelihood they declare for B. For citizens with x i < 0, whether they are more or less likely to declare for either candidate depends also on the other variables. Next, we examine the extension with competitive clientelism. H7 Competitive Clientelism: As candidate B provides larger rewards, declarations for A ( ) ( ) decrease πi (A) r B < 0, declarations for B increase πi (B) r B > 0, and non-declarations ( ) decrease πi ( ) r B < 0. Consider the effect of a marginal change in the size of rewards offered by candidate B. EU i (A) r B = (1 q + α)γ EU i (B) r B = 0 EU i ( ) r B = 0 Increasing r B, increases EU i (B) in equation (5) for all citizens but does not affect EU i (A) or EU i ( ). As a consequence, increasing r B increases the likelihood any citizen declares for B to the expense of the likelihood the same citizen declares for A or remains undeclared. Next, we examine the extension with punishments. H8 Punishments: As the clientelist candidate A imposes greater punishments, declarations for A increase πi (A) ( ) ( ) p A > 0, declarations for B decrease πi (B) p A < 0, and nondeclarations increase πi ( ) ( ) p A > 0. Consider the effect of a marginal change in the punishment imposed by candidate A. EU i (A) p A = 0 EU i (B) p A = (q α)γ EU i ( ) p A = 0 Increasing p A, decreases EU i (B) in equation (6) for all citizens but does not affect EU i (A) or EU i ( ). As a consequence, increasing r B decreases the likelihood any citizen declares for A to the advantage of the likelihood the same citizen declares for A or remains undeclared. 52

54 Appendix B - Brazil Experiment Table 7: Brazil Characteristics of Online Sample vs. Nation Overall Online Sample Brazil Overall Gender Female 46.2% 49.0% Male 53.8% 51.0% Age % 31.0% % 22.3% % 18.5% % 13.6% % 8.1% % 6.4% Region Center-West 6.2% 7.4% North 4.9% 8.3% Northeast 30.6% 27.8% South 20.0% 14.4% Southeast 38.4% 42.1% Urban Rural 19.2% 15.6% Urban 80.8% 84.4% Notes: Characteristics of online sample are self-reported by participants in the declared support experiment. These participants were recruited through Facebook advertisements, as described in Section 4.1. Characteristics of Brazil overall reflect 2010 data from Brazil s census bureau (Instituto Brasileiro de Geografia e Estatística). 53

55 ONLINE APPENDICES - NOT FOR PUBLICATION Online Appendix A - Model Using Deterministic Choice We also show the logic of declared support is similar when employing a simpler deterministic model, in which citizens choose actions with certainty. Predictions of the deterministic model are equivalent to those of the stochastic choice model, in terms of the aggregate distribution of actions, when considering a continuum of citizens whose ideal points are distributed over a sufficiently large spectrum (e.g., x i [, ]). However, implementing an experiment involves using a narrower spectrum, in which some predictions diverge. Whenever the deterministic model predicts an effect, the stochastic choice model makes the identical prediction. But in some cases, the deterministic model predicts no effect as subjects ideal points are insufficiently close to cutoffs and it unrealistically assumes citizens never make mistakes whereas the stochastic choice model allows mistakes and thus predicts an effect. To simplify exposition, our deterministic analysis focuses on the base model presented in Section 3.1 and assumes that declaration costs are sufficiently large relative to clientelist benefits such that there exist citizens who remain undeclared. This assumption is realistic, given that during real-world campaigns, not every citizen publicly expresses support for a candidate. 34 Citizens with sufficiently intense ideological preferences will always declare support for their preferred candidate. For such citizens, the increased probability of their favorite candidate winning and/or the expressive utility from declaration are worth the cost of declaring and dominate any clientelistic considerations. On the other hand, clientelist and/or declaration costs weigh more heavily on the decisions of citizens with weaker ideological preferences, who prefer to remain undeclared or even declare for the other candidate. This intuition is captured in Figure 3, which shows the ideological space along which citizens can be arranged according to their ideological preferences. Moving along the spectrum of 34 Given the restrictions on the other parameters, a sufficient assumption to ensure the existence of non-declarers is c A + c B > γ(q + α)r A. 54

56 ideological preferences, the incentive to declare support for candidate A increases as x i (the ideological gain from A winning) rises. Declare for B Remain Undeclared Declare for A x* B x* A x i Figure 3: Optimal Behavior as a Function of Ideological Preferences Citizens on the left, with smaller values of x i, are supporters of B, whereas citizens on the right, with higher values of x i, are supporters of A. The right cutpoint (x A ) represents a citizen whose ideological preferences make her indifferent between declaring for A and remaining undeclared. The left cutpoint (x B ) represents a citizen whose ideological preferences make her indifferent between declaring for B and remaining undeclared. The assumption that there exist citizens who remain undeclared enables us to focus on the case in which x B < x A. All citizens to the left of x B (i.e., who prefer B more strongly than x B ) declare support for B. By contrast, all citizens to the right of x A (i.e., who prefer A more strongly than x A ) declare support for A. Citizens between x A and x B remain undeclared. Depending on the contextual characteristics that is, on the values of the model parameters both cutpoints may represent supporters of the same party (i.e., x A > x B > 0 or 0 > x A > x B ). When this is the case, clientelist considerations dominate instrumental and expressive considerations for some citizens, who declare for the candidate they dislike. To derive these two cutpoints, we observe that citizen i prefers declaring for B over remaining undeclared when EU i (B) > EU i ( ), and prefers declaring for A over remaining undeclared when EU i (A) > EU i ( ). Substituting equations and solving yields: x B = c B α + δ x A = c A (q + α)γr A α + δ (7) Our objective is to derive comparative static results for the effect of increasing each variable on the fraction of citizens who declare for A, declare for B, or remain undeclared 55

57 (keeping every other variable constant). Recall that we assume there exist some citizens who remain undeclared (that is, x A > x B ). Thus, increasing x B means increasing the fraction of citizens who declare for B rather than remaining undeclared; decreasing x B means decreasing the fraction of citizens who declare for B rather than remaining undeclared. On the other hand, increasing x A means decreasing the fraction of citizens who declare for A rather than remaining undeclared; decreasing x A means increasing the fraction of citizens who declare for A rather than remaining undeclared. This means that to determine the marginal effect of each variable on the fraction of citizens who declare for B (A), it is enough to determine the sign of the partial derivative of that variable on the cutpoint x B (x A ). The citizens who remain undeclared are those whose ideology falls in between the two cutpoints. Thus, the fraction of undeclared citizens is proportional to the distance between the two cutpoints, or to x A x B, which is always positive because we focus on the case where x A > x B. We have: x A x B = c A + c B γ(q + α)r A α + δ (8) To determine the marginal effect of each variable on the fraction of citizens who remain undeclared, we consider the partial derivatives of each variable on (x A x B ). H1 Reward Size: As clientelist candidate A provides larger rewards, declarations for A ( ) ( ) x increase A x r A < 0, declarations for B are unaffected B r A = 0, and non-declarations ( ) (x decrease A x B ) r A < 0. Proof: x B does not depend on r A. The derivative of x A with respect to r A is: x A γ(α + q) = r A α + δ which is always negative. Regarding non-declarations, increasing r A decreases the numerator of equation (8) and does not affect its denominator. 56

58 H2 Machine Support: As clientelist candidate A s probability of winning increases, declarations for A increase A x ( ) ( ) x < 0, declarations for B are unaffected B = 0, and q q ( ) (x non-declarations decrease A x B ) < 0. q Proof: x B does not depend on q. The derivative of x A with respect to q is: x A q = γr A α + δ which is always negative. Regarding non-declarations, increasing q decreases the numerator of equation (8) and does not affect its denominator. H3 Social Cost: As the social cost of declaring for clientelist candidate A increases, declarations for A decrease A x ( ) ( ) x c A > 0, declarations for B are unaffected B c A = 0, and ( ) (x non-declarations increase A x B ) > 0. q Proof: x B does not depend on c A. The derivative of x A with respect to c A is: x A c A = 1 α + δ which is always positive. Regarding non-declarations, increasing c A increases the numerator of equation (8) and does not affect its denominator. H4 Monitoring: As the clientelist candidate A s ability to monitor how citizens declare ( x A increases, declarations for A increase ), < 0 declarations for B are unaffected γ ( ( ) x B ), = 0 (x and non-declarations decrease A x B ) < 0. γ γ Proof: x B does not depend on γ. The derivative of x A with respect to γ is: x A γ = (α + q)r A α + δ which is always negative. Regarding non-declarations, increasing γ decreases the numerator of equation (8) and does not affect its denominator. 57

59 H5 Expressive Utility: As the utility of declaring in accordance with preferences increases, declarations for A increase if x A > 0 and decrease if x A < 0; declarations for B increase ( ) ( ) x B (x > 0, and non-declarations decrease A x B ) < 0. δ δ Proof: The derivative of x B with respect to δ is: x B δ = c B (α + δ) 2 which is always positive. The derivative of x A with respect to δ is: x A δ = c A (q + α)γr A (α + δ) 2 This is positive if and only if (q+α)γr A > c A (that is, if and only if x A < 0). Regarding non-declarations, increasing δ increases the denominator of equation (8) and does not affect its numerator. H6 Election Influence: As the election influence of declaring increases, declarations for A increase if γ(q δ)r A < c A and decrease if γ(q δ)r A > c A ; declarations for B increase ( ( ) x B ), > 0 (x and non-declarations decrease A x B ) < 0. α α Proof: The derivative of x B with respect to α is: x B α = c B (α + δ) 2 which is always positive. The derivative of x A with respect to α is: x A α = γ(q δ)r A c A (α + δ) 2 which is positive if and only if γ(q δ)r A > c A. Regarding non-declarations, increasing α reduces the numerator of equation (8) and increases its denominator. 58

60 Online Appendix B - Brazil Experiment Table 8: Brazil Observed Choices vs. Predictions, Heterogeneous Treatment Effects Expressive Utility Treatment Panel A: Declaration for Clientelist Candidate A Treatment Prediction Proportion Difference p-value As Predicted? Supporters of A Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility Increases Indifferent Citizens Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility No Effect Supporters of B Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility Decreases Panel B: Declaration for Opposition Candidate B Treatment Prediction Proportion Difference p-value As Predicted? Supporters of A Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility Decreases Indifferent Citizens Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility No Effect Supporters of B Baseline Clientelism Baseline Baseline Baseline Baseline Expressive Utility Increases Notes: Number of observations is 572 for A Supporters, 175 for Indifferent Citizens, and 512 for B Supporters; p-values refer to one-tailed difference in proportions Z test. All results are robust to using two-tailed tests. The model does not make unambiguous predictions for No Declaration, so no panel is shown for that dependent variable. 59

61 Table 9: Brazil Observed Choices vs. Predictions, Heterogeneous Treatment Effects No Election Influence Treatment Panel A: Declaration for Clientelist Candidate A Treatment Prediction Proportion Difference p-value As Predicted? A Supporters & Neutrals Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Decreases B Supporters Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Increases Panel B: Declaration for Opposition Candidate B Treatment Prediction Proportion Difference p-value As Predicted? A Supporters & Neutrals Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Increases B Supporters Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Decreases Notes: Number of observations is 572 for A Supporters, 175 for Indifferent Citizens, and 512 for B Supporters; p-values refer to one-tailed difference in proportions Z test. All results are robust to using two-tailed tests. The model does not make unambiguous predictions for No Declaration, so no panel is shown for that dependent variable. 60

62 Table 10: Brazil Estimates of Heterogeneous Treatment Effects (Logit) Without Subject Fixed Effects Panel A: Declaration for Clientelist Candidate A (1) (2) (3) (4) Expressive Utility 0.049** (0.022) (0.026) No Election Influence (0.025) (0.025) Round *** (0.005) (0.005) (0.005) (0.005) Partisan Type 0.004** 0.005*** *** (0.002) (0.002) (0.002) (0.002) Screener 0.056*** ** (0.020) (0.019) (0.019) (0.020) Sample A Supporters B Supporters A Supporters B Supporters Prediction Increases Decreases Decreases Increases Subjects Fixed Effects No No No No Observations Panel B: Declaration for Opposition Candidate B (1) (2) (3) (4) Expressive Utility *** (0.021) (0.026) No Election Influence ** (0.022) (0.025) Round * (0.005) (0.005) (0.005) (0.005) Partisan Type ** * ** * (0.002) (0.002) (0.001) (0.002) Screener *** *** (0.017) (0.019) (0.017) (0.019) Sample A Supporters B Supporters A Supporters B Supporters Prediction Decreases Increases Increases Decreases Subject Fixed Effects No No No No Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. Baseline Clientelism is the excluded treatment category, so that coefficients report differences from that baseline. Robust standard errors are reported, clustered by subject. 61

63 Table 11: Brazil Estimates of Heterogeneous Treatment Effects (Logit) With Subject Fixed Effects Panel A: Declaration for Clientelist Candidate A (1) (2) (3) (4) Expressive Utility 0.178*** ** (0.053) (0.053) No Election Influence * (0.050) (0.055) Round 0.048*** *** (0.012) (0.011) (0.011) (0.013) Sample A Supporters B Supporters A Supporters B Supporters Prediction Increases Decreases Decreases Increases Subject Fixed Effects Yes Yes Yes Yes Observations Panel B: Declaration for Opposition Candidate B (1) (2) (3) (4) Expressive Utility *** (0.059) (0.050) No Election Influence *** (0.056) (0.055) Round *** ** (0.013) (0.012) (0.013) (0.012) Sample A Supporters B Supporters A Supporters B Supporters Prediction Decreases Increases Increases Decreases Subject Fixed Effects Yes Yes Yes Yes Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. Baseline Clientelism is the excluded treatment category, so that coefficients report differences from that baseline. Robust standard errors are reported. 62

64 Table 12: Brazil Estimates of Average Treatment Effects, Punishments (Logit) (Shows Robustness of Table 5 to Inclusion of Subject Fixed Effects) Declare for A Declare for B No Declaration Treatment (1) (2) (3) (4) (5) (6) Punishment Only 0.152*** *** ** (0.034) (0.036) (0.039) Clientelism & Punishment 0.133*** *** (0.033) (0.035) (0.039) Round 0.013* 0.036*** *** *** (0.008) (0.007) (0.008) (0.008) (0.009) (0.010) Subjects Fixed Effects Yes Yes Yes Yes Yes Yes Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. To isolate causal effects, Punishment Only specifications employ No Clientelism as the excluded category, and Clientelism & Punishment specifications employ Punishment Only as the excluded category. Robust standard errors are reported, clustered by subject. 63

65 Online Appendix C - US Experiment Table 13: United States Parameters for Experimental Design Treatment r A q γ r B p A α c Baseline Clientelism No Clientelism Lopsided Election Low Monitoring Competitive Clientelism Punishment Only Clientelism and Punishment No Election Influence Note: r A and r B are rewards offered by candidates A and B, respectively; q is A s ex-ante probability of winning the election; c is the cost of declaring for either candidate, γ is the probability declarations are observed; p A is the punishment imposed by A; and α is the impact of declarations on the electoral odds. Red text indicates parameters that differ from Baseline Clientelism. 64

66 Figure 4: United States Declaration Choices of Participants, by Treatment Baseline Clientelism No Clientelism Lopsided Election Low Monitoring Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Share of Choices Partisan Type Partisan Type Partisan Type Partisan Type Competitive Clientelism Punishment Only Clientelism & Punish No Election Influence Partisan Type Partisan Type Partisan Type Partisan Type Declare for A Declare for B No Declaration Note: For each treatment, figures reflect the share of participants in US experiment (N = 144) who declared for A, declared for B, and did not declare. Shares are shown for each partisan type labeled on horizontal axes: (1) strong A supporter, (2) weak A supporter, (3) indifferent citizen, (4) weak B supporter, and (5) strong B supporter. Participants were randomly assigned to these partisan types, and induced to hold such preferences for the fictitious candidates. 65

67 Table 14: US Observed Choices vs. Predictions, Heterogeneous Treatment Effects No Election Influence Treatment Panel A: Declaration for Clientelist Candidate A Treatment Prediction Proportion Difference p-value As Predicted? A Supporters & Neutrals Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Decreases B Supporters Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Increases Panel B: Declaration for Opposition Candidate B Treatment Prediction Proportion Difference p-value As Predicted? A Supporters & Neutrals Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Increases B Supporters Baseline Clientelism Baseline Baseline Baseline Baseline No Election Influence Decreases Notes: Number of observations is 114 for A Supporters, 60 for Indifferent Citizens, and 114 for B Supporters; p-values refer to one-tailed difference in proportions Z test. All results are robust to using two-tailed tests. The model does not make unambiguous predictions for No Declaration, so no panel is shown for that dependent variable. 66

68 Table 15: US Estimates of Average Treatment Effects, Rewards (Logit) Declare for A Declare for B No Declaration (1) (2) (3) (4) (5) (6) No Clientelism *** *** 0.115*** 0.268*** 0.472*** 0.735*** (0.024) (0.023) (0.020) (0.036) (0.032) (0.034) Lopsided Election 0.031* 0.044* ** *** (0.018) (0.024) (0.017) (0.036) (0.017) (0.028) Low Monitoring *** *** 0.066*** 0.154*** 0.222*** 0.346*** (0.027) (0.030) (0.020) (0.037) (0.029) (0.033) Competitive Clientelism *** *** 0.312*** 0.732*** ** ** (0.027) (0.029) (0.026) (0.035) (0.016) (0.023) Partisan Type 0.024*** *** (0.001) (0.001) (0.001) Screener (0.012) (0.012) (0.013) Subject Fixed Effects No Yes No Yes No Yes Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. Baseline Clientelism is the excluded treatment category, so that coefficients report differences from that baseline. Robust standard errors are reported, clustered by subject in columns 1, 3, and 5. 67

69 Table 16: US Estimates of Average Treatment Effects, Punishments (Logit) Panel A: Punishment Only Declare for A Declare for B No Declaration (1) (2) (3) (4) (5) (6) Punishment Only *** *** 0.163*** 0.566*** (0.021) (0.115) (0.019) (0.028) (0.030) (0.064) Partisan Type 0.020*** *** (0.001) (0.001) (0.003) Screener (0.019) (0.015) (0.032) Subject Fixed Effects No Yes No Yes No Yes Observations Panel B: Clientelism & Punishment Declare for A Declare for B No Declaration (1) (2) (3) (4) (5) (6) Clientelism & Punishment 0.597*** 0.977*** ** *** *** *** (0.028) (0.011) (0.021) (0.110) (0.034) (0.020) Partisan Type 0.018*** *** *** (0.001) (0.002) (0.002) Screener (0.019) (0.016) (0.021) Subject Fixed Effects No Yes No Yes No Yes Observations Note: : p < 0.10, : p < 0.05, : p < Coefficients report marginal effects from logistic regressions. Each observation corresponds to a decision in the experiment. Dependent variables are indicators coded 1 if the decision was as listed in the respective column headers; 0 otherwise. Independent variables are indicators coded 1 if the decision for a given observation corresponds to the treatment listed in the respective rows; 0 otherwise. To isolate causal effects, Punishment Only specifications employ No Clientelism as the excluded category, and Clientelism & Punishment specifications employ Punishment Only as the excluded category. Robust standard errors are reported, clustered by subject in columns 1, 3 and 5. 68

70 Online Appendix D - Brazil Experiment Screenshot Examples Instructions (Page 1 of 2) TRANSLATION: Thank you for participating! You already have 50 TICKETS for the IPhone 5S lottery. You will now play 10 games to earn more tickets. The more tickets you have, the more chances you will have to win an Iphone. Every game has an election. Two candidates run for mayor the yellow candidate and the green candidate. In each game, you will have the option to place a yellow or green flag on your house. If you put up a flag, you increase the chances of that candidate winning the election. You can also choose to place no flag on your house. 69

71 Instructions (Page 2 of 2) TRANSLATION: Read the instructions carefully. The tickets you earn for each choice can change from one question to another. In some games, the candidate who wins may reward you if you placed his flag on your house, or he may penalize you if you placed his opponent s flag. After each game, the computer chooses the winner. [IMAGE: Yellow Ball Chosen Yellow Candidate Wins. Green Ball Chosen Green Candidate Wins.] The number of IPhone tickets you will earn depends on who wins the election and your decision about the flag. Remember that the candidates and the flags are not real! Click when you are ready to play. 70

72 Weak Supporter of Candidate A (Partisan Type 3) No Clientelism Treatment, Options Page Client TRANSLATION: PLEASE CHOOSE ONE OF THE FOLLOWING OPTIONS: [NO FLAG] If you place NO flag: 5 yellow balls and 5 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 44 tickets; If the green ball is chosen, the green candidate wins and you earn 34 tickets. [YELLOW FLAG] If you place a YELLOW flag: 6 yellow balls and 4 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 40 tickets; If the green ball is chosen, the green candidate wins and you earn 30 tickets. [GREEN FLAG] If you place a GREEN flag: 4 yellow balls and 6 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 40 tickets; If the green ball is chosen, the green candidate wins and you earn 30 tickets. 71

73 Weak Supporter of Candidate A (Partisan Type 3) No Clientelism Treatment, Outcome Page Yellow Flag Chosen, Yellow Candidate Wins No Client TRANSLATION: YOU EARNED 40 TICKETS FOR THE IPHONE LOTTERY! [IM- AGE: Yellow flag selected, Yellow ball chosen.] You earned 40 tickets! You placed a yellow flag. The computer chose a yellow ball, so the yellow candidate won. [BUTTON: CLICK TO EARN MORE TICKETS!]. 72

74 Weak Supporter of Candidate A (Partisan Type 3) Baseline Clientelism Treatment, Options Page line TRANSLATION: PLEASE CHOOSE ONE OF THE FOLLOWING OPTIONS: [NO FLAG] If you place NO flag: 5 yellow balls and 5 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 44 tickets; If the green ball is chosen, the green candidate wins and you earn 34 tickets. [YELLOW FLAG] If you place a YELLOW flag: 6 yellow balls and 4 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins. He rewards you for placing a yellow flag, so you earn 50 tickets; If the green ball is chosen, the green candidate wins and you earn 30 tickets. [GREEN FLAG] If you place a GREEN flag: 4 yellow balls and 6 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 40 tickets; If the green ball is chosen, the green candidate wins and you earn 30 tickets. 73

75 Weak Supporter of Candidate A (Partisan Type 3) Baseline Clientelism Treatment, Outcome Page No Flag Chosen, Green Candidate Wins TRANSLATION: YOU EARNED 34 TICKETS FOR THE IPHONE LOTTERY! [IM- AGE: No flag selected, Green ball chosen.] You earned 34 tickets! You did not place a flag. The computer chose a green ball, so the green candidate won. [BUTTON: CLICK TO EARN MORE TICKETS!]. 74

76 Weak Supporter of Candidate B (Partisan Type 5) No Clientelism Treatment, Options Page lient TRANSLATION: PLEASE CHOOSE ONE OF THE FOLLOWING OPTIONS: [NO FLAG] If you place NO flag: 5 yellow balls and 5 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 34 tickets; If the green ball is chosen, the green candidate wins and you earn 44 tickets. [YELLOW FLAG] If you place a YELLOW flag: 6 yellow balls and 4 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 30 tickets; If the green ball is chosen, the green candidate wins and you earn 40 tickets. [GREEN FLAG] If you place a GREEN flag: 4 yellow balls and 6 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 30 tickets; If the green ball is chosen, the green candidate wins and you earn 40 tickets. 75

77 Weak Supporter of Candidate B (Partisan Type 5) No Clientelism Treatment, Outcome Page Green Flag Chosen, Yellow Candidate Wins TRANSLATION: YOU EARNED 30 TICKETS FOR THE IPHONE LOTTERY! [IM- AGE: Green flag selected, Yellow ball chosen.] You earned 30 tickets! You placed a green flag. The computer chose a yellow ball, so the yellow candidate won. [BUTTON: CLICK TO EARN MORE TICKETS!]. 76

78 Weak Supporter of Candidate B (Partisan Type 5) Baseline Clientelism Treatment, Options Page nt TRANSLATION: PLEASE CHOOSE ONE OF THE FOLLOWING OPTIONS: [NO FLAG] If you place NO flag: 5 yellow balls and 5 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 34 tickets; If the green ball is chosen, the green candidate wins and you earn 44 tickets. [YELLOW FLAG] If you place a YELLOW flag: 6 yellow balls and 4 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins. He rewards you for placing a yellow flag, so you earn 40 tickets; If the green ball is chosen, the green candidate wins and you earn 40 tickets. [GREEN FLAG] If you place a GREEN flag: 4 yellow balls and 6 green balls in the jar; If the yellow ball is chosen, the yellow candidate wins and you earn 30 tickets; If the green ball is chosen, the green candidate wins and you earn 40 tickets. 77

79 Weak Supporter of Candidate B (Partisan Type 5) Baseline Clientelism Treatment, Outcome Page Yellow Flag Chosen, Yellow Candidate Wins TRANSLATION: YOU EARNED 40 TICKETS FOR THE IPHONE LOTTERY! [IM- AGE: Green flag selected, Yellow ball chosen.] You earned 40 tickets! You placed a yellow flag. The computer chose a yellow ball, so the yellow candidate won. He rewards you for placing a yellow flag, so you earn 40 tickets. [BUTTON: CLICK TO EARN MORE TICK- ETS!]. 78

80 Online Appendix E - US Experiment Screenshot Examples US Experiment: Instructions (Page 1 of 2) 79

81 US Experiment: Instructions (Page 2 of 2) 80

82 US Experiment: Indifferent Citizen Baseline Clientelism Treatment, Options Page 81

83 US Experiment: Indifferent Citizen Baseline Clientelism Treatment, Outcome Page Blue Flag Chosen, Blue Candidate Wins 82

84 US Experiment: Moderate Supporter of Candidate A Baseline Clientelism Treatment, Options Page 83

85 US Experiment: Moderate Supporter of Candidate A Baseline Clientelism Treatment, Outcome Page Blue Flag Chosen, Blue Candidate Wins 84