Read my Lips! Experimental Evidence on the Effects of Electoral Competition on Shirking and Trust

Transcription

1 Vol 5 No 7 Read my Lips! Experimental Evidence on the Effects of Electoral Competition on Shirking and Trust Gari Walkowitz (University of Cologne) Arne R. Weiss (University of Cologne)

2 Read my Lips! Experimental Evidence on the Eects of Electoral Competition on Shirking and Trust Gari Walkowitz Arne R. Weiss November 30, 2014 We experimentally test whether electoral competition reduces shirking behavior by oce-holders and increases citizens' trust. We hypothesize that competition increases campaign promises by oce-holders, who feel committed to what they promise. Using a novel repeated multi-person investment-game with periodic elections, we indeed nd that elected oce-holders shirk less (i.e., they back-transfer more to citizens relative to investments) as compared to randomly appointed oce-holders. Surprisingly, this eect cannot be explained through competition inating the level of electoral promises. Nevertheless, promises do matter; in fact, they carry greater weight for the behavior of elected oce-holders than for their randomly appointed counterparts. Elections also have a positive short-term eect on citizens' trust by cutting o both low and excessively high promises. JEL Classication: D72, D02, D03, C71, C91 Keywords: Elections, Promises, Shirking, Trust Game We thank Dennis Dittrich, Simon Gächter, Hong Geng, Heike Hennig-Schmidt, Michel Maréchal, Bettina Rockenbach, Reinhard Selten, Stephan Tontrup, and Irenaeus Wol for valuable comments and suggestions during all stages of the project. We also thank participants of research seminars in Bonn, Cologne, ESA Lyon, and Erfurt for helpful discussions. Department of Corporate Development and Business Ethics, University of Cologne, Albertus Magnus-Platz, Cologne, Germany. Phone: , gari.walkowitz@unikoeln.de. Experimental and Behavioral Economics, Department of Economics, University of Cologne, Albertus-Magnus- Platz, Cologne, Germany. Phone: , arne.weiss@uni-koeln.de. 1

3 It is useless to hold a person to anything he says while he is in love, drunk, or running for oceshirley MacLaine Broken vows are like broken mirrors. They leave those who held to them bleeding and staring at fractured images of themselves.richard Paul Evans 1 Introduction Shirley MacLain might be wrong after allat least when it comes to politics. Despite the stereotype, no doubt fueled by a long list of read my lips!-promises and outright lies, politicians are actually quite reliable promise-keepers. 1 In a meta-study that spans 21 cases from dierent countries over the past 40 years, Pétry and Collette (2009) nd that 67% of electoral promises were fullled. This gives some vindication to what Manin et al. (1999) call the mandate view of elections (starting with Hotelling, 1929): Candidates oer a menu of electoral platforms, from which voters choose what they like best. The winning candidate will then implement her proposed policies. This view was criticized by rational choice scholars (Ferejohn, 1986; Austen-Smith and Banks, 1989) as naivelyignoring that, once elected, politicians may well have better things to do than keeping their word of serving the nation, e.g., ying to World Cup matches instead of going through tedious reports or stealing public funds instead of investing in education. The rst line of defense of the mandate view is that voters do not like to be cheated and would therefore punish incumbents who broke their promises. 2 However, this sounds more straightforward than it is since punishment (i.e., voting against the preferred candidate) is a dominated strategy for rational voters. 3 Theoretical and empirical advances in behavioral economics and social psychology on guilt aversion, cost of lying and preferences for consistency (Cialdini, 1993; Eyster, 2002; Charness and Dufwenberg, 2006; Vanberg, 2008; Battigalli and Dufwenberg, 2009; Lundquist et al., 2009; Kartik, 2009; Falk and Zimmermann, 2013; Guadagno and Cialdini, 2010) suggest further mechanisms that link action to words, even in a static setting of a single election. 4 Such theories do not only provide a micro foundation of the mandate view. They furthermore suggest that voters 1 Read my lips! No new taxes (George H. W. Bush); Nobody has the intention of building a wall. (Walter Ulbricht); If you like your health care plan, you'll be able to keep your health care plan, period. (Barack Obama); I only make promises that I will be able to keep. Everything I say will be done. (Francois Hollande)... 2 Harrington Jr. (1993) goes as far as putting this argument into political science folklore. 3 Additional sanctioning strategies become credible (which would not be subgame-perfect with standard preferences) based on theories of strong reciprocity and betrayal-aversion (e.g., Falk and Fischbacher, 2006; Bohnet et al., 2008): Voters might then be willing to retaliate promise-breaking, i.e., vote against their preferred candidate. 4 The psychological costs of lying have been explained, for instance, through self-image concerns (see Mazar et al., 2008; Fischbacher and Föllmi-Heusi, 2013 andmore poeticallyrichard Paul Evans). 2

4 might benet from campaign-based electoral competition beyond selecting among policies, screening candidates (e.g., Carrillo and Mariotti, 2001; Cooter, 2002; Besley, 2005) and providing incentives through the shadow of future elections: Electoral competition might force politicians to promise and deliver more than they would in the absence of electoral competition. In this paper, we study this conjecture in a controlled laboratory experiment. Our aim is to better understand the behavioral mechanisms that shape the relationship between electoral promises and promise-keeping in oce, with the presence of strong incentives not to do so. Moreover, we also investigate how the above relationship aects citizens' trust in repeated economic exchanges between oce-holders and citizens. The Link between Words and Action There are several papers that try to provide a micro-foundation of the link between campaign promises and political action through the channel of re-election pressures. In Harrington Jr. (1993), heterogeneity in beliefs among both voters and candidates about the economic eectiveness of policies can lead candidates to reveal their true policy intentions. Voters update their beliefs after observing economic outcomes during the rst term. Candidates therefore prefer to implement policies they believe to be eective and to be elected by a median voter whose initial beliefs are in line with the candidate's beliefs. This can keep campaign promises meaningful despite term limits and politicians being purely oce-motivated. However, the crucial assumption in his model is the alignment of interests between candidates and voters. By contrast, we are interested in situations of conicting interests between the voters and the elected ocials. In the model by Aragonès et al. (2007), candidates may stick to their promises for fear of voter punishment. This punishment is prospective as an implicit part of voters' strategy to believe a candidate's promises as long as the candidate has a good reputation, i.e., has never reneged in the past. This allows for subgame-perfect equilibria (in a game with innite horizon) in which promises are believed and kept. Campaign promises are therefore a coordinating device for voters and do not matter per se. Voters ultimately only care about the policies implemented. In terms of policy outcomes, any subgameperfect equilibrium in the game with promises is therefore also a subgame-perfect equilibrium in the same game without promises. Two recent papers provide a more direct link between words and actions by assuming some form of cost of lying among candidates. In Callander and Wilkie (2007), even a small fraction of honest candidates, assumed to have a non-zero cost of lying, induces meaning and structure in electoral campaigns. Despite their relative disadvantage in the ability to lie, honest candidates win a substantial share of elections. In the model of Kartik and 3

5 McAfee (2007), candidates either have or do not have character. Candidates with character are inexible and will do exactly what they promised, whereas candidates with no character will say whatever it takes to get elected. Voters care about policies and character. Similar in spirit, both models come to the conclusion that cost of lying or character can lead to imperfect pandering to the center and meaningful campaigns. In contrast to Harrington Jr. (1993) and Aragonès et al. (2007), these results hold even for a single election, i.e., without any re-election incentives. The causal logic in both models runs from policy intentions to promises. This becomes most obvious in the model of Callander and Wilkie (2007). The choice of platforms at the campaign stage is constrained by cost of lying, whereas behavior is not constrained by campaign promises: Following Callander and Wilkie (2007), elected candidates will implement their policy intentions regardless of what they promised. However, it is actually the reverse path that has received the most attention and empirical support from the literature on guilt-aversion and cost of lying, as well as from studies on the positive eects of exchanges of promises on subsequent levels of cooperative behavior (e.g., Ostrom et al. 1992; Kerr and Kaufman-Gilliland 1994; Sally 1995; Bicchieri and Lev-On 2007). 5 If candidates indeed choose promises and policies strategically, the same politician might act more in line with voters' interests if she is elected based on campaign promises than if she assumed oce without an election, irrespective of any looming future elections. As an example, assume that citizens' investments increase welfare, but the government may decide to tax the returns of the investments. The oce-holders cannot precommit to respect private property rights but can promise to do so. If the oce-holders feel bound to their promises, electoral competition may lead candidates to promise and enact lower tax-rates than they would have in the absence of elections. This in turn might foster citizens' trust in the oce-holders, thereby increasing investments. We will test these conjectures using an experimental approach. Studying the Role of Promises with an Experiment Separating the motivation to comply with promises from reputational concerns is very dicult with eld data because oce-holders often aim to be re-elected. Furthermore, elections also serve as a selection device (Cooter, 2002) such that elected oce-holders might be dierent types than oce-holders that came into oce without elections, making it dicult to draw causal inference from electoral competition to shirking behavior and trust. 6 The laboratory allows us to draw 5 See the hypotheses section 3 for more details. 6 The (semi-)separating equilibria in the game-theoretic models of Callander and Wilkie (2007) and Kartik and McAfee (2007) also predict that elected oce-holders are, in expectations, dierent types than oce-holders that 4

6 causal inferences from electoral competition to oce-holders' shirking behavior and citizens' trust in a strictly controlled environment. In our experiment, we can rule out reputational concerns, control for selection eects and exogenously determine whether oce-holders are subject to electoral competition. In addition, all participants act fully anonymously via the computer network (see details below). Two related dictator-game studies model and experimentally test the interaction between promises, electoral competition and shirking behavior. Based on a signaling model with guilt-averse candidates who have heterogeneous pro-social preferences, Geng et al. (2011) predict that electoral competition inates oce-holders' promises on giving behavior and leads to higher transfers. The authors nd partial experimental support for their model: Elected oce-holders act less selshly than randomly elected oce-holders if electoral campaigns are based on promises. Promises are not signicantly higher in their election treatment but there is a stronger correlation between promises and behavior for elected than for randomly appointed oce-holders. This is contrary to results from another experimental study by Corazzini et al. (2014), who nd electoral competition to raise promises. However, the theoretical prediction, as well as empirical result that electoral competition based on campaign promises leads to reduced shirking behavior, is shared between these two studies. 7 While both studies are important contributions to a better understanding of the role of campaign promises, they leave open whether their results generalize beyond the simple setting of a one-shot dictator game and whether electoral competition also inuences voter trust. In both studies, there is very little reason for candidates to promise anything in the respective control treatment with random appointment of oce-holders because recipients have no say in these treatments. Indeed, to promise nothing is also a theoretical prediction in the control treatment of Corazzini et al. (2014). Yet, even in autocratic regimes citizens are important actors who co-determine welfare through their decisions. As agents with their own economic interests, their decisions crucially hinge on their trust put in the government. Following the example above, governments, whether democratic or not, may promise not to conscate assets after successful investments. Will electoral competition increase these promises further? By employing a multi-person investment game, we can answer this question and test for the eects of electoral competition on trust. Mapping the example above, in our design oce-holders decide how to divide money whereas citizenswho can elect the oce-holderdetermine, through their investment decisions, the size of the disposable budget. The investment game is repeated over ve came into oce without elections. 7 Corazzini et al. (2014) use a signaling model with a cost of lying that depends on the approval rate. Candidates either have a high or low sensitivity to the psychological cost of lying. 5

7 election periods, in which reputation building across periods is ruled out. Citizens can invest by sending money to an oce-holder who was either elected by the citizens or exogenously determined. Investments are doubled by the experimenter. Before the election, oce-holder candidates make promises on their intended back-transfers. Once elected into oce, an oce-holder can, however, freely deviate from her electoral promises. As a major dierence to previous studies, we also measure candidates' shirking attitudes beforehand. This allows us to test whether candidates can successfully signal their types during the electoral campaign. We will investigate if 1) elected oce-holders transfer back higher amounts relative to investments (our measure of shirking) than randomly appointed oce-holders and whether 2) citizens transfer higher amounts (a measure of trust) when they can elect the oce-holders. We nd strong eects of elections on oce-holders' shirking behavior and limited eects on citizens' trust. Oce-holders use their power in a less self-oriented way when elected instead of randomly appointed. This nding is robust to controlling for dierences in players' individual shirking attitudes and the size of the budget. Surprisingly, and extending upon previous ndings, this eect cannot be explained by the hypothesized mechanism of electoral competition raising promises. Yet, promises do matter. In fact, the driving force behind the lower level of self-oriented behavior is the stronger correlation between oce-holders' behavior and promises under electoral competition as compared to when they are exogeneously appointed. This context-dependency of the eects of promises is a puzzle for current theories of cost of lying or guilt aversion. A tentative explanation is that an election based on promises concludes an implicit contract between the candidate and the electorate. Elections also have a short-term eect on trust. They aect trust by leading to the right kind of promises, i.e., under elections promises are compressed such that low promises as well as excessively high promises are cut o. The paper is organized as follows: Section 2 introduces the experimental design. Hypotheses are derived in section 3. Results are presented in section 4. Section 5 concludes. 2 Experimental Implementation In the following, we rst illustrate our experimental design; then we describe the procedures we have applied. 6

8 2.1 Experimental Design In our experimental treatment (VOT), participants are randomly selected into anonymous groups of eight players, each consisting of ve citizens and three oce-holder candidates. Citizens and candidates are selected into their roles based on their randomly assigned roles during the ex-ante measurement of their individual shirking and trust attitude (see below). The citizens elect one of the three candidates as an oce-holder. After election, the oce-holder plays with each of the ve citizens a one-on-one investment game similar to Berg et al. (1995), i.e., the oce-holder acts as a responder (or trustee) to each citizen's investment. All players initially receive an endowment of 10 points (1 Point = 1 RMB) 8. Subsequently, all citizens individually choose an investment level a i {0, 1,..., 10}. Investments are doubled by the experimenter and transferred to the elected oce-holder. The oce-holder then separately decides for every investment a i the amount b i {0, 1,..., 2 a i } to be back-transferred to each citizen. The oce holder can therefore freely decide how much of the received amount to back-transfer. The payo of a citizen is given by π Si = 10 a i + b i. Accordingly, an oce-holder's payo is calculated as π R = a i b i. The two non-elected candidates, who do not become an oce-holder, pocket their endowment. The oce-holder is always able to receive at least the same payo as the citizen. The oceholder eectively plays a dictator game with each of the group's citizens, the pie being the received amount 2 a i. The return-on-investment (RoI = b i /a i 1) implies a measure of shirking towards the citizen. The level of shirking corresponds to the degree that the oce-holder exploits her power to her own advantage. As a lower ROI means a higher prot for the oce-holder at the expense of the citizen's prot, it implies a higher level of shirking. We introduce voting into this basic investment framework by implementing ve election periods. Each election period starts with an election stage, followed by ve rounds of the multi-person investment game. Hence, in total, the game consists of 25 decision rounds across ve election periods. At the election stage, all candidates have to announce a non-binding back-transfer strategy, which is shown to all citizens in the group. This back-transfer strategy is a feasible vector of backtransfers they promise to execute in case of their election for any possible investment a i = 0, 1,..., 10 received from an individual citizen. Candidates are assigned a unique identication code so that each candidate can be unambiguously linked to a promised backtransfer strategy. This code is valid for one period (i.e., ve out of 25 decision rounds). By contrast, citizens have no identication codes; oce-holders can only determine their respective back-transfers based on the investments received. After seeing the candidates' back-transfer promises, the ve citizens are asked to elect 8 RMB stands for Renminbi, the national currency of China. 7

9 one oce-holder out of the three candidates by majority rule. Finally, the ve citizens receive information about the outcome of the election; the corresponding identication code of the winning candidate is shown to the citizens. 9 The non-elected candidates do not participate in the investment game that follows and do not receive any information about it. In order to analyze the long-term eects of voting we use a repeated setup, which allows oceholders, candidates, and citizens to learn. Roles and groups are kept constant. Every ve rounds a new oce-holder is chosen. Prior to any oce-holder election, candidates receive new identication codes. Furthermore, the position of candidates' promises on the computer screen is shued for every oce-holder election. Hence, oce-holders can build a reputation within periods, but not across. The backtransfer strategy of the elected oce-holders is also shown to the non-elected candidates. This was done in order to set a focal point on the winning back-transfer strategy, which allows learning among candidates on how to win elections. The control treatment (RAN) is identical to VOT, with the only dierence that the election is replaced by a random draw such that each candidate faces a probability of 1/3 of being selected in each period. 10 A further important feature of our experimental design is the ex-ante measurement of individual dierences in players' baseline levels of shirking and trust. We conducted a bilateral one-shot investment game (henceforth OS) before the experimental treatments RAN and VOT were launched. This game had identical parameters to the multi-person investment game in RAN and VOT, with two main dierences: First, every later citizen (in the treatments RAN and VOT) was matched with one later candidate. Second, each later candidate decided by the strategy vector method (Selten, 1967), i.e., they had to decide for every possible amount sent how much to transfer back. Subjects only received feedback on the resulting payos from the OS at the very end of the experimental session. Subjects were not given any information about which experiment would follow afterward but were truthfully told that no information would be transferred between the OS and the later experiment. Each later candidate's strategy prole elicited in OS can be seen as an attitude towards shirking. 11 More precisely, the strategies provide a baseline level of ROI, averaged over the entire range of possible investments and also for every possible investment amount separately. Similarly, the OS investment decision serves as a measure of individual attitudes towards trust. Measuring candidates' shirking attitudes allows us to test whether our design is indeed successful in preventing 9 The exact voting outcome, i.e., how many votes each candidate received, is not displayed in order to focus on the procedure of voting, rather than on the role of approval rates. 10 Most importantly, candidates go through the same procedure of stating promises. 11 Note that the concept of attitudes, as used in psychology, are only assumed to be stable over time (but not over contexts), whereas preferences, as used in rational choice models, are usually assumed to be stable over both contexts and time. Other papers that used the strategy method to elicit individual types of behavior include, e.g., Fischbacher and Gächter (2010), Fischbacher et al. (2001). 8

10 Treatment Groups Citizens Candidates Group members RAN 9 5 x 9 = 45 3 x 9 = 27 8 x 9 = 72 VOT 9 5 x 9 = 45 3 x 9 = 27 8 x 9 = 72 Total Table 1: Number of subjects per treatments included in our experiment the potential selective function of elections. Furthermore, we can more precisely measure the eect of voting on oce-holders' shirking behavior by controlling for any individual dierences in shirking attitudes. 2.2 Experimental Procedure We ran the experiments in Chengdu, the capital and largest city of Sichuan province, People's Republic of China. Most Chinese people have very little experience with voting mechanisms as elections only take place in some rural areas (Pastor and Tan, 2000). China also scores low on the Voice and Accountability indicator of the World Bank Governance indicators (Kaufmann et al., 2008). The lack of experience with voting procedures is an advantage for testing our hypotheses. Chinese subjects should have less pre-understanding of how an elected oce-holder should or does behave than subjects in countries with more frequent elections. In this way, experimental results may be less confounded by subjects' personal experiences with voting procedures and the behavior of particular elected oce-holders. Instructions on the screen and on paper presented the experiment in rather neutral terms. However we emphasized the hierarchical relation between the oce-holder and the citizens: citizens were called citizens and the oce-holders president. Candidates were called candidates. The decision by citizens is called investing and the oce-holders' back-transfer decision was called transferring back. 12 In OS, investors were called Type A and responders Type B. In total, 144 subjects participated in the computerized laboratory experiment. Subjects were recruited from the student subject-pool of the Herbert A. Simon & Reinhard Selten Behavioral Decision Research Lab, Southwest Jiaotong University, Chengdu, P.R. of China. We collected 18 independent observations in total (see Table 1 for an overview on treatments and participants). The experiment was programmed in ASP.NET. Separate cubicles with curtains made sure that inter- 12 A translated version of the instructions can be found in the Appendix. 9

11 action between the subjects was completely anonymous. All procedures were common knowledge, including that the experiment started only after all subjects had correctly answered a set of test questions about the experiment, including how earnings were calculated. Subjects' earnings were set at local standards; subjects earned an average of 91 RMB (about 12 USD at the time of the experiment), paid out in cash individually at the end of the experiment. 3 Research Hypotheses It is straightforward to derive the subgame-perfect Nash-equilibrium of our game when moneymaximizing rationality is common knowledge. Since there is a known end to the game, the logic of backward-induction rules out the existence of cooperative equilibria that may otherwise be present in innite play. Therefore, investments are predicted to be zero throughout the game in both treatments, and consequently voting is predicted to not have any inuence on shirking. This prediction, however, is at odds with a vast amount of experimental literature on one-shot and repeated trust games (see e.g., Cardenas and Carpenter, 2008; Johnson and Mislin, 2011; Bornhorst et al., 2010). Once we relax the assumption of common knowledge of rationality (e.g., Kreps and Wilson, 1982) or allow for arguments in the utility function other than own income, a more complex picture emerges, but the picture looks dierent depending on the theory painting it. Outcome-based theories of pro-social preferences (Fehr and Schmidt, 1999; Bolton and Ockenfels, 2000) may explain cooperation in nite play but would predict no treatment eects since promises are no more than cheap talk. By contrast, theories of guilt-aversion, cost of lying and taste for consistency depart from the cheap talk assumption. Based on these theories, a voting mechanism may inuence behavior if voting aects the formation of promises. If promises for in-oce behavior are more than cheap talk, higher promised ROI may imply higher actual ROI: Hypothesis 1: Higher promised ROI lead to higher ROI. This hypothesis follows naturally from theories in which lying (Ellingsen and Johannesson, 2004; Kartik, 2009) or behavioral inconsistency (Eyster, 2002; Falk and Zimmermann, 2013) directly enters negatively into the utility function: Higher promised ROI may drive up actual ROI if, in utility terms, the cost from deviating from what was promised is higher than the sacrice of sticking to the promise. H1 may also be motivated by models of guilt-aversion (e.g., Battigalli and Dufwenberg, 2009; Charness and Dufwenberg, 2006), in which promises matter indirectly through their impact 10

12 on second-order beliefs. These models would predict a link between promises and behavior if promise-makers believe that their promises matter to the promise-receivers; when beliefs are formed rationally, this needs to hold in equilibrium. This is formally modeled by Geng et al. (2011) for dictator giving-behavior when candidates are guilt-averse but dier in their pro-social preferences. In their model, competition under the voting mechanism may raise promises and transfers compared to a random appointment of the oce-holder. Competition raises promises because in equilibrium these serve as a credible signal of the candidates' in-oce behavior. The model by Corazzini et al. (2014), in which candidates dier in their cost of lying, yields the same prediction. Given that H1 holds, citizens will anticipate the link between promises and behavior, which leads us to predict: Hypothesis 2: Higher promised ROI improve electoral prospects: Candidates receive more votes if they promise higher ROI compared to rival candidates. Based on the predicted voting behavior, candidates have an incentive to raise their promises beyond what they would promise in the absence of competition, i.e., when selection into oce is random. We therefore predict: Hypothesis 3: Promised ROI are higher in VOT than in RAN. If promised ROI are indeed higher in VOT than in RAN (H3) and higher promised ROI indeed lead to higher actual ROI (H1), then we can, for our principal research question, predict: Hypothesis 4: For given levels of investment, elected oce-holders in VOT will transfer back higher amounts than randomly-determined oce-holders in RAN. Given that H4 holds, can we expect investments to be higher in VOT than in RAN? Not necessarily: If citizens are risk-neutral and prot-maximizing, investments follow a step-function with zero investments below an expected return on investment of zero and full investments above. Under this assumption, investments would only dier between the treatments if the expected return on investment is below zero in RAN and above zero in VOT. However, if there are risk-averse citizens who demand a premium for taking the risk of investing, which is suggested by previous experiments (e.g., Holt and Laury, 2005; Dohmen et al., 2011), then investments might indeed be responsive to dierent levels of return on investment. Therefore, we predict: Hypothesis 5: Investments are higher in VOT than in RAN. 11

13 We now turn to our data in order to see whether our hypotheses receive empirical support. 4 Results 4.1 Does Voting Decrease Shirking? We start by looking at the behavior of elected oce-holders across treatments (see panel 1 from Figure 1) at the level of the statistically independent groups. We nd average ROI over all 25 rounds to be weakly signicantly higher in VOT (34.88%) compared to RAN (16.63%) (p =.084, two-sided FisherPitman Permutation Test for two independent samples, henceforth FPPI). 13 By controlling for players' individual dierences in shirking attitudes, as elicited in OS, we are able to measure the treatment eect more precisely. As it turns out, the estimated treatment eect increases considerably. Average ROI Average difference to OS ROI Round Round RAN VOT RAN VOT Figure 1: Treatment comparison of mean returns on investment over time. Panel 1 shows uncontrolled mean levels; panel 2 displays mean levels as deviations from the OS strategy Elected oce-holders in VOT deviate on average more from their OS ROI-level compared to their counterparts in RAN (62.0% vs. 28.3%, p =.018, FPPI, also see panel 2 in Figure 1, which shows oce-holders' average deviation from their OS strategy over time). The marked increase in 13 All statistical tests in this paper are two-sided. 12

14 the treatment eect is due to elected oce-holders in VOT (31.3%) having a somewhat higher baseline level of shirking than randomly determined oce-holders in RAN (16.4%), albeit not signicantly so (p =.409, FPPI). 14 This also shows that the detected treatment dierences are not driven by citizens successfully selecting, relative to the random mechanism, the candidates with lower shirking attitudes. If at all, voters performed worse in this regard than the random device: Treatment-dierences in OS- ROI are somewhat smaller in the sample of all candidates (22.8% in VOT versus 20.5% in RAN, p =.884, FPPI) than in the sub-sample of (s)elected oce-holders, as shown above. Taken together, introducing a voting procedure clearly decreases oce-holders' shirking, in line with our principal hypothesis H4. What drives higher ROI? Based on the hypothesized mechanism, we expect higher ROI in VOT to be driven by higher promised ROI. Therefore, we rst investigate whether elected oce-holders in VOT promise to give back more compared to RAN, as predicted by H3. We nd that oce-holders promised on average higher ROI in VOT (66.1%) than in RAN (52.1%), in line with H3. However, because this dierence is not statistically signicant (p =.244, FPPI), we cannot reject, at conventional levels of condence, that this treatment dierence only arose by chance. At rst glance, this is surprising because candidates who promise higher average ROI relative to their rival candidates enjoy an electoral advantage (see Model 1 of Table 2). In this regression, a voter's choice for one of the three candidates is explained by each of the candidates' level of promises relative to the mean level of promises in the group. We therefore use McFadden's alternative specic conditional logit model (see Cameron and Trivedi, 2010). While Model 1 gives support to H2, Model 2 shows that this hypothesis needs to be qualied: In the range of hyperfair promises (the dummy variable hyper-fair is equal to 1 for promised ROI larger than 5/6) 15 increasing promises further from the group's mean no longer improves electoral prospects; they might even become bleaker. 16 According to the information criteria AIC and BIC, Model 2 is 10 Back transfer OS i i= Average baseline ROI are calculated as 1, with i being the investment level. i=1 i 15 Note that an ROI of 5/6 equalized, on average, payos between the oce-holder and the citizens (as the oce-holder plays a trust game with each of the 5 citizens and furthermore receives an endowment of 10). 16 Since a very small number of clusters (below 10, according to the simulations by Bertrand et al., 2004) can lead to a considerable over-rejection of the null hypothesis, we run a robustness check through the wild cluster-bootstrap procedure, which is recommended by Cameron et al. (2008) to provide asymptotic renement. Implementation in Stata is done with the program cgmwildboot for linear regression models, written by Judson Caskey. The dependend variable is therefore the number of votes received by each candidate. The results are qualitatively the 13

15 Votes for candidates (1) (2) Dierence from mean promise.022*** (.009).038*** (.012) Dierence from mean promise Hyper-fair promise -.098*** (.027 ) N AIC BIC Table 2: Coecients from a conditional logistic regression analysis. The dependent variable is Votes for candidates. Hyper-fair promises are coded as 1. Robust standard errors clustered at the group-level in parentheses. *** p <.01, ** p <.05, * p <.1 superior to Model 1. A likely explanation is that hyper-fair promises were deemed less credible by the voters than promises in a lower range. This in turn might have prevented candidates in VOTdespite the competitive pressurefrom increasing their promises signicantly beyond the level in RAN. In order to dig deeper and to answer whether promises actually mattered for behavior, we try to explain individual ROIs by means of regression analysis. This also allows us to control for oce-holders' shirking-attitudes as well as the interaction between investments and back-transfers. Considering Models 1 to 3 in Table 3, the regression analyses paint a similar picture as the previous non-parametric statistics with respect to the main eect: The coecient of the treatment dummy VOT (which is coded 1 for VOT and 0 for RAN) is about.2 in all specications (and always signicantly dierent from zero). This is very close to the simple dierence in means between the treatments reported above. Hence, neither shirking attitudes nor promised ROI can explain why ROI are about 20% higher in VOT than in RAN. 17 The signicant treatment eect in Model 3 is not particularly surprising in light of the insignicant dierences between promised ROI in VOT and RAN. Nevertheless, it clearly shows that inated promises are not the driving force behind the main treatment eect. Note, however, that promised ROI have a strong inuence on behavior, as was hypothesized in H1. In Model 4, we additionally include investments as a regressor in order to see whether part of the treatment eect is a reward of more trust. As we will analyze in greater detail in section same and obtainable upon request from the authors. 17 The strong predictive power of ROI in OS and ROI during the main experiments adds plausibility to treating the former as a measure of shirking attitudes. 14

16 ROI (1) (2) (3) (4) (5) (6) VOT.184**.223***.207***.189*** (.08) (.05) (.05) (.11) (.04) OS-ROI.260***.255***.236***.245***.245*** (.04) (.04) (.03) (.03) (.03) Promised ROI.225**.210***.140**.140** (.09) (.07) (.07) (.06) Investment.028***.027***.027*** (.01) (.01) (.01) VOT promised ROI *** (.19) (.07) Constant.195***.214*** ** (.06) (.04) (.05) (.04) (.06) (.04) N 1,737 1,737 1,737 1,737 1,737 1,737 R AIC BIC Table 3: Regression analysis of oce-holders' ROI. Robust standard errors clustered at the grouplevel in parentheses. *** p <.01, ** p <.05, * p <.1 4.2, investments are somewhat higher in VOT than in RAN. However, while trust is indeed rewarded overproportionally, Model 4 shows that this explains very little of the treatment eect as the coecient stays almost the same (.19). To dig deeper, Model 5 goes one step further and allows the eects of promises on behavior to dier between the treatments. In Model 5, promised ROI represents the additional inuence of promises in RAN, whereas the interaction dummy VOT promised ROI shows the additional inuence in VOT. Since Model 5 includes a slope dummy for the partial eects of promises, the dummy VOT is now the estimated dierence between the treatments when promised ROI is equal to zero and all other explanatory variables are held constant. This dierence is clearly not dierent from zero. Therefore, we drop the dummy VOT in Model 6. Note that the point estimates of Models 5 and 6 are almost identical, yet the precision of the interaction dummy VOT promised ROI in Model 6 is greatly improved (because of reduced multicollinearity). Furthermore, according to AIC and BIC, Model 6 is the most informative. Model 6 shows that promises carry signicantly greater weight for behavior in VOT than in RAN: There is a steeper relationship between actual and promised ROI in VOT compared to RAN. As the most important conclusion of Model 5, the treatment dummy turns insignicant once we control for the dierential eects of promises on behavior. Taken together, Models 5 and 6 show that the treatment eect eectively runs through promises: At low levels 15

17 of promises (such as promising a zero return on investment), oce-holder behavior does not dier between the treatments; at high promised ROI, oce-holders give back signicantly more in VOT than in RAN. Long-term Eects on Shirking To study whether the treatment eect on oce-holders' behavior fades over time, we look at differences in the very last period. In these rounds (21-25), oce-holders in VOT still deviate weakly signicantly more from their shirking attitudes, as measured in OS, compared to oce-holders in RAN (p =.085, FPPI). However, in a repetition of the regression analysis of Table 3, a signicant treatment eect can no longer be detected (see Table 4). RoI (1) (2) (3) (4) (5) (6) VOT (.13) (.12) (.11) (.11) (.28) OS-ROI.309***.327***.322***.346***.341*** (.10) (.10) (.10) (.10) (.10) Promised ROI.338**.330**.229*.273** (.12) (.13) (.11) (.12) Investment (.01) (.01) (.01) VOT promised ROI * (.36) (.14) Constant.266***.236*** (.07) (.08) (.11) (.11) (.10) (.10) N R AIC BIC Table 4: Regression analysis of oce-holder's ROI for rounds 21 to 25. Robust standard errors clustered at the group-level in parentheses. *** p <.01, ** p <.05, * p <.1 Despite a point estimate in a similar range as in Table 3 (once we control for shirking attitudes, in Models 2 to 4), it fails to be signicant at conventional levels. The same holds true for the role of promises. The point estimate for the additional eect of promises on behavior in VOT compared to RAN is even higher in the last period than overall (.478 in Model 5 vs..292 in Model 5 from Table 3) and yet is not signicant (p=.204). However, once we drop the treatment dummy VOT (which is, as in Model 5, not signicant) in Model 6, the interaction between promises and the treatment 16

18 does become weakly signicant (p=.064). Again, following AIC and BIC, Model 6 represents the best t of our data. The regression analysis might suer from lower power due to the reduced number of observations (309 compared to 1,737 in the full model). As an interesting side note, shirking attitudes are not only a highly signicant predictor even in the last ve rounds but have the strongest inuence on actual behavior (based on standardized coecients). 18 To conclude, based on the point estimates, a similar picture emerges for the last period as for the entire experiment: The voting procedure seems to lead to less shirking oce-holder behavior as well as to a stronger inuence of promises on behavior. However, possibly because of the low power of a restricted mode with few observations, we cannot condently conclude that this is more than mere chance. 4.2 Does Voting foster Citizens' Trust? In order to look at the eects of the voting procedure on citizens' trust, it is insightful to rst look at behavior in the rst round. Here, citizens have not yet been exposed to any payo-relevant decisions by the oce-holder; hence, any dierences in rst round investments would be clear evidence for a treatment eect independent of oce-holders' back-transfer behavior. Investments start o somewhat higher in VOT (6.69) than in RAN (5.47), though the treatment dierence fails to reach signicance in a conservative non-parametric test at the level of independent groups (p=.149, FPPI). Controlling for players' trust-attitudes, we do, however, nd a signicant treatment eect: Citizens deviate signicantly more from their OS-investment level in VOT (3.44) than in RAN (1.31) (p=.003, FPPI; also see regression Model 2 in Table 5). Interestingly, Model 3 shows no signicant positive linear relationship between promises and investments. A closer look at the data reveals that there is a positive linear relationship but a negative quadratic relationship (see Model 4). Consequently, investments rise with promises at low levels of promised ROI but decrease when promises increase further. Citizens obviously distrust high promises. This mirrors the earlier nding that the electoral advantage of higher promises disappears at near-maximum promises. As the most important conclusion of Model 4, the treatment dummy turns insignicant once we allow for a non-linear relationship between promises and investments. Where then does the treatment eect in Model 2 come from? As the interaction dummy between the voting treatment and promises is not signicant (and therefore omitted from Table 5), we can conclude that the type of promises given in VOT are more conducive to triggering investments than the type of promises in RAN. The former tend to be higher (.59 vs..44), though not signicantly 18 Results are available from the authors. 17

19 Investment (1) (2) (3) (4) VOT ** 1.369**.362 (.77) (.66) (.62) (.59) OS-transfer.455***.464***.427*** (.14) (.14) (.14) Average promised ROI ** (1.32) (3.63) Average promised ROI ** (3.21) Constant 5.467*** 3.577*** 2.683*** 2.470*** (.69) (.74) (.91) (.84) N R AIC BIC Table 5: Citizen behavior in the rst round. Robust standard errors clustered at the group-level in parentheses. *** p <.01, ** p <.05, * p <.1 so (p=.340, FPPI). More importantly, promises in VOT seem to be concentrated in the better range. In the random treatment, we nd promised ROI (slightly) below zero (-.09) as well as at thenot very crediblemaximum of 1. By contrast, only promises in the upper, yet interior, range are selected in the voting treatment (between.45 and.81 in VOT). In line also with the voting analysis (see Table 2), voters seem to have distrusted the candidates who promised low and those candidates who promised excessively high ROI. 18

20 Average investment Average difference to OS investment Round Round RAN VOT RAN VOT Figure 2: Treatment comparison of average investments over time. Panel 1 shows uncontrolled means; panel 2 displays the mean deviation from trust attitudes. Investment (1) (2) (3) (4) (5) (6) VOT (.85) (.85) (.80) (.84) (.56) (.55) OS-transfer.247***.254***.247***.193***.202*** (.05) (.06) (.06) (.06) (.06) Average promised ROI *.873 (1.23) (2.85) (2.17) Average promised ROI (2.73) (2.11) ROI last round 3.018*** 3.033*** (.34) (.35) Constant 5.169*** 4.143*** 3.741*** 3.334*** 4.799*** 4.492*** (.64) (.74) (1.23) (1.07) (.77) (.56) N 2,250 2,250 2,250 2,250 1,721 1,721 R AIC BIC Table 6: Citizens behavior over all rounds. Robust standard errors clustered at the group-level in parentheses. *** p <.01, ** p <.05, * p <.1 19

21 Averaging over the entire course of the experiment, investments are about 20% higher in VOT (6.23) than in RAN (5.17) (see Figure 2), which, however, is not a signicant dierence (p=.242, FPPI, see also Model 1 in Table 6). Controlling for trust attitudes in the regression analysis of Table 6 leads a dierent picture compared to a non-parametric analysis. The latter shows that citizens deviate from their OS-investment level weakly signicantly more in VOT (by 2.98) than in RAN (1.01) (p=.057, FPPI), whereas Model 2 does not show a signicant treatment dummy (p=.153). From the second round onward, citizens learn about the back-transfer behavior of the oce-holders. Therefore, Model 5 includes lagged ROI (from the previous round) as an explanatory variable. Once we control for protability experiences, promises lose the already weak explanatory power they have in Model 4. This model shows a weakly signicant positive linear and an insignicant quadratic relationship between promised ROI and investments (without controlling for squared promises, in Model 3, the linear relationship is not signicant). Model 6 (which is the most informative one according to AIC and BIC) therefore drops promised ROI as a linear and quadratic explanatory variable. It reveals that there is clearly no treatment eect on trust that is independent from oceholders' behavior and, interestingly, that trust attitudes still matter despite the wealth of feedback citizens receive about oce-holders' behavior. Looking at the evolution of average treatment dierences in trust, we nd investments to be higher for about the rst half of the experiment (see left panel in Figure 2). Treatment dierences are more stable if we control for trust attitudes (see the right panel in 2). Nevertheless, the treatment eect on investments found in the rst round peters out over time, and trust becomes mainly determined by oce-holders' behavior. Consequently, we only nd weak support for hypothesis H Is Voting Protable for Citizens? We have shown that investments and the protability of investing are higher in VOT than in RAN. As investments are on average protable (i.e., ROI > 0), this translates into citizens earning slightly more in VOT (12.59 points) compared to RAN (11.25) (p=.060, FPPI). Interestingly, however, oce-holders in RAN (29.58) do not make signicantly higher prots than oce-holders in VOT (28.2) (p=.439, FPPI), despite the fact that ROI are higher in VOT than in RAN. 19 The reason for this lies in the behavior of the citizens. Whereas oce-holders in VOT sacrice in any given round part of their own prots for the benet of their constituency, this is, in dynamic terms, almost fully 19 Recall that oce-holders interact with 5 citizens. Consequently, a dierence of about 1.3 in prots is relatively smaller for oce-holders than for citizens. 20

22 compensated by the somewhat higher investments in VOT, which increase overall earnings. 20 The distribution of income consequently diers only slightly between treatments. In RAN, citizens earn on average 18.3 points less than oce-holders. In VOT, this dierence (15.6) is smaller than in RAN at a borderline level of signicance (p =.103, FPPI). Thus, our data deliver some evidence that our voting treatment leads to higher and more balanced payos for citizens without signicantly decreasing oce-holders' income. 5 Conclusion In this paper, we studied the role of promises in periodic elections for oce-holders' shirking behavior and citizens' trust. We hypothesized that electoral competition would inate promises and, furthermore, lead to higher actual returns-on-investment. We also expected citizens to invest more when oce-holders are elected as compared to being randomly selected. We nd a positive eect of the election procedure on in-oce behavior. Elected oce-holders shirk less than their randomly appointed counterparts, as they provide ROI that are almost 20 percentage-points higher. This eect is stable over time. However, the hypothesized mechanism only gets mixed support: Even though citizens reward candidates who promised high returns on investments, promises are not signicantly higher in the voting treatment. In line with this observation, the treatment eect cannot be explained by controlling for the level of promised ROI. Yet promises do matter. In fact, they carry greater weight in the voting treatment than in the random treatment. Controlling for this interaction does explain the treatment eect, that is, oce-holders' behavior only diers when they promise high returns on investment. It seems that the election procedure creates (or strengthens) an implicit contract between oce-holders and citizens: In the absence of elections, canditates' promises are without consequence for the selection into oce. In the voting treatment, by contrast, candidates are clearly elected because of the promises they made. Oce-holders may therefore feel more committed to their promise. The election procedure also aects trust. The estimated treatment dierence is an increase of 20 percent due to the voting treatment. However, this dierence is only signicant in the short-term and diminishes over the course of the experiment. The eect in the rst round, before citizens are exposed to oce-holders' shirking, can be explained by promised returns on investment being in a range that is more conducive to inducing investments. Citizens likely choose those candidates who 20 However, we cannot conclude condently that overall earnings, as a measure of eciency, are indeed higher in VOT. As they increase proportionally with investments, a signicant dierence in eciency levels cannot be detected (82.85% in VOT and 78.04%, p =.242, FPPI). 21