Weather, mood, and voting: An experimental analysis of the effect of weather beyond turnout Anna Bassi*^ Abstract Theoretical and empirical studies show that inclement weather on an election day reduces turnout, potentially swinging the results of the election. Psychology studies, however, show that weather affects individual mood, which - in turn -affects individual decision-making activity potentially beyond the simple decision to turn out on an election day. This paper evaluates the effect of weather, through its effect on mood, on the way in which voters who do turn out decide to cast their votes. The paper provides experimental evidence of the effect of weather on voting when candidates are perceived as being more or less risky. Findings show that, after controlling for policy preferences, partisanship, and other background variables, bad weather depresses individual mood and risk tolerance, i.e.. voters are more likely to vote for the candidate who is perceived to be less risky. This effect is present whether meteorological conditions are measured with objective or subjective measures. This draft: May 20, 2013. Word-count: 8467 *The University of North Carolina at Chapel Hill, Department of Political Science ^An earlier version of this paper has previously circulated as "The indian rain dance of the incumbent. The effect of weather beyond turnout." I am grateful to Daniel Butler, David Cesarini, Stanley Feldman, Andrew Healy, Luke Keele, Neil Malhotra, Michael MacKuen, and Agnieszka Tymula for helpful comments. I would like to thank the Behavioral Lab at UNC Kenan-Flagler for providing part of the financial support for this paper, and seminar's participants at the 2012 Midwest Political Science Association, 2013 Southern Political Science Association, and 6th Annual NYU-CESS Experimental Political Science Conference for providing valuable feedback. All errors remain my own. 1 Much has been written about the effect of bad weather on Presidential Election Day turnout and whether it benefits one party over another. The issue of weather and its impact on elections is one that receives constant media and political campaigner hype at every election. Understanding the possible effects of weather conditions on voting behavior is crucial: in today's contentious elections, even a small difference in numbers of votes over a large geographical area could tilt the vote one way or the other. For example, Ludlum (1984) finds that weather proved decisive in the presidential election of 1960, when John F. Kennedy defeated Richard Nixon by a razor-sharp margin of 11,500 votes in five states. Heavy rains, caused by a cold front in swing states like Illinois, deterred rural voters (mostly Republicans) from going to the polls, while not affecting the urban voters (mostly Democrats) in Chicago. The analysis of other cases (Ludlum, 1984) confirm this effect, providing support for an axiom of New York politics that holds that [...a rainy day favors a Democratic candidate since the upstate Republicans would not turn out in full in inclement weather, while the urban Democrats would not be put to undue inconvenience.] (Ludlum, 1984, p. 102) With Republicans continuing to hold a greater share of the vote in rural areas and Democrats continuing to dominate in urban areas, the New York politics' adage seems not to hold anymore. Recent studies (De Nardo, 1980; Knack, 1994; Gatrell and Bierly, 2002; Gomez et al, 2007, Keele and Morgan, 2013) show mixed results. Although the issue of weather and its effects on elections is one that invariably arises in every election year, most of the literature focuses on the effect of weather on turnout exclusively: these studies fail to investigate whether the effect of weather goes beyond turnout, affecting the way in which voters, who do turn out, decide how to cast their vote. However, some findings (Gomez et al., 2007) suggest that the ef- feet of weather on parties' vote share may be greater than the indirect effect through turnout (their findings show that for every inch of rain above average, turnout decreases only by approximately 0.9%, but that the Republican candidate receives approximately an extra 2.5% of the votes). Few studies have investigated this question. The effect of extreme weather (such as natural disasters) has been analyzed in a context of public opinion and elections by Healy and Malhotra (2010), who analyze the effect of tornados on electoral outcomes, finding that voters appear to reward or punish the incumbents according to their perceived performance in handling the disaster. Gerber (2013) studies the relationship between partisanship and climate policy of the local government. Cohen (2011) tackles the question of how general weather affects public opinion, finding a positive relation between presidential approval ratings and sunshine exposure. To have a better understanding of the effect of weather on voters' decisions, we need to both assess the effect of weather on voters' actions and beliefs and to understand the mechanism through which climate conditions exert such an effect. Psychology literature suggests that weather affects both conscious and subconscious mood and that mood affects human behavior. Whether, however, the effect of weather on mood is strong enough to drive human's behavior is less clear. Laboratory experiments provide the best tool for testing these behavioral questions because of the precise control that they afford and the possibility to analyze data that do not naturally occur (Woon, 2012b). Bassi et al. (2013) provides experimental evidence of the link between weather, mood, and risk aversion. Similarly, this paper investigates whether such an effect is present in a voting setting. Specifically, this paper identifies the existence of an effect of weather on individual voting decisions. Findings suggest that sunlight and good weather have a positive impact on the likelihood of voting for riskier candidates, while voters rely more heavily on less risky candidates in bad weather. This result holds for both objective and subjective measures of weather conditions. 2 Furthermore, this paper helps identify a specific pathway through which weather affects voting decisions. The paper provides an analysis of the mechanism at work by employing a psychological questionnaire called PANAS-X (Watson and Clark, 1994) used to measure respondents' moods. I find that "positive mood" feelings such as self-assurance and attentiveness display a statistically significant decrease in bad weather conditions, while sadness displays a statistically significant increase. Results also show that positive mood feelings and states that are sensible to weather conditions are also positively associated with the likelihood of voting for a riskier candidate. I interpret these findings as offering evidence of a causal mechanism at work: the impact of weather (through mood) on voting choice. Subjects are more willing to accept the level of risk associated with a risky candidate when they are in a better mood. 1 Weather, Mood, and Behavior The impact of sunlight and weather in general on human mood has been widely examined in the clinical psychology literature. For example, good mood has been associated with low levels of humidity (Sanders and Brizzolara, 1982); high levels of sunlight (Cunnigham, 1979; Parrot and Sabini, 1990; and Schwartz and Clore, 1983); high barometric pressure (Goldstein, 1972); and high temperature (Cunningham, 1979; Howarth and Hoffman, 1984). Furthermore, the effect of temperature and sunlight is especially strong in the spring, when people have been deprived of such weather (Keller et al., 2005). By the same token, mood has been proven to significantly affect how individuals make decisions. For example, research in experimental psychology has proven that mood affects the way agents make decisions in risky or uncertain environments by affecting their levels of risk aversion. A person's mood may affect the subjective judgment of the 3 likelihood of a future event (Wright and Bower, 1992): a happy person is "optimistic," i.e., she reports higher probabilities for positive events and lower probabilities for negative events. Conversely, a sad person is "pessimistic," perceiving lower (higher) probabilities to be attached with positive (negative) events. Experimental studies have documented a negative link between anxiety (and depression) and "sensation seeking" measures, which have been extensively documented to be reliable proxies for risk-taking behavior. More recently, Eisenberg et al. (1998) show in experimental studies that depressed individuals tend also to be more risk averse in a series of hypothetical everyday-life situations. In this framework, people in a good mood would be more willing to engage in activities and choices yielding a degree of risk. This conduct has been identified in the literature as "mood-risk tolerance" channel. Bassi et al. (2013) show that weather affects risk aversion through this "mood-risk channel." Mood on Election Day can swing voter choice of a marginal voter - the one who is still undecided - through this "mood-risk tolerance" channel. A voter who is virtually or almost indifferent between two candidates might lean toward the "riskier" candidate when she feels in an upbeat mood, or might resort to the "safer" candidate when she feels more depressed or pessimistic. If the "safer" choice is also equivalent to the status quo, this effect can also be interpreted as "status quo bias," "loss aversion," "endowment effect," or "regret avoidance." A related stream of literature analyzes the effect of emotions on risk assessments. Johnson and Tversky (1983) show a positive link between emotions and risk assessments, indicating that optimistic emotions lead to optimist risk assessments and vice versa. Hsee and Weber (1997) suggest that positive emotions lead to greater risk seeking because people are more optimistic about future outcomes, while negative emotions such as anxiety make agents more pessimistic about the future and thus more risk-averse. What makes the literature on mood and emotions complementary but not equivalent is the fact that, although mood and emotions are tightly linked, not all negative emotions lead to negative mood states and viceversa (DeSteno et al., 4 2000). Druckman and McDermott (2008) show that emotions need to be differentiated beyond their positive or negative general state, as different negative emotions exert opposite effects on individuals' risk attitudes. MacKuen et al. (2010), Huddy et al. (2007), and Feldman et al. (2013) stress that negative emotions such as anxiety need to be distinguished from other negative emotional responses such as anger. This paper explores the way in which weather influences how individuals vote between candidates framed as risky choices via a "mood-risk tolerance" channel. Mood may provide an important key in explaining the different effects of weather variables on voting behavior. The goal of this paper is to enrich, rather than negate, earlier findings on the effect of emotions on elections and public opinion by analyzing how the mood variables that are susceptible to change with weather conditions affect voting. I next describe how individuals' risk assessments map into voting decisions. A similar approach has been taken by Eckel, El-Gamal, and Wilson (2009), who investigate the link between risk preferences and emotions in a sample of hurricane Katrina evacuees right after the natural disaster and a year later, finding the first sample to be more risk-loving than the second. The authors ascribe the higher risk tolerance of the first sample to the prominence of negative emotions. 2 Uncertainty, Prospects, and Risk Attitudes As Berinsky and Lewis (2007) suggest, analyzing the question of whether risk attitudes affect voters' preferences regarding risky candidates requires discussion of the (1) specification of the voters' utility function over different candidates; and (2) specification of how the uncertainty about future outcomes enters in their utility calculations. Concerning the first question, the standard approach - the expected utility model -assumes that the individual utility function is concave, that is, the marginal utility 5 of an additional dollar diminishes when the total utility increases (Bernulli, 1954). However, Kahneman and Tversky (1979) argue that individuals display diminishing marginal utilities only for prospects with positive outcomes, while displaying increasing marginal utilities for prospects with negative outcomes, suggesting that individuals are risk-averse for gains, but risk-seekers for losses. The key element of this argument is that individuals engage in decision making first by identifying a reference point from which people tend to be risk-averse for gains and risk-loving for losses. Therefore, risk aversion is a function not only of the riskiness of an option, but also of its desirability (see McDermott, 2001, for a comprehensive review). The way in which individuals interpret their choices, as gains or as losses, influences how much risk they will take. As has been found in several experiments on framing (Druckman, 2001a, 2001b, 2001c), the way in which information is framed influences individuals'judgment as well, in that it affects how they interpret their choices. Kahneman and Tversky (1979) found that framing a policy as a gain (for example, by describing a 10% rate of unemployment as an employment rate of 90%) induces individuals to consider their choices in a domain of gain, while framing the same policy as a loss induces them to consider their choices in a domain of losses. As regards the second question, Alvarez and Franklin (1994) suggest that no consensus exists in the literature on how uncertainty affects the voter's evaluation of the candidates. Most of the literature focuses on the uncertainty about the policies that the candidates would put in place once elected: Shepsle (1972), Enelow and Hinich (1981), Bartels (1998) and Palfrey and Poole (1987). consider how the uncertainty about candidate locations enters voters' expected utility. Alternatively, Quattrone and Tversky (1988) focus on how uncertainty about candidates' performance affects voting choice. They test a voting choice between an incumbent and a challenger with identical policy preferences but with different degrees of likelihood to implement the policy. Quattrone and Tversky (1988) confirm Kahneman and Tversky (1979)'s predictions in a social choice domain: voters are averse to voting for risky candidates 6 in the domain of gains but they do seek out the riskier candidate in the domain of losses. This prediction has been extended to data from real-world elections in Mexico by Morgenstern and Zechmeister (2001), who found risk attitudes to be a strong determinant of voter behavior when deciding between an incumbent and a challenger. As the objective of the paper is to analyze the way in which weather influences voting decisions via a "mood-risk tolerance" channel, the paper's design builds on the Quat-trone and Tversky (1988) classical design, in which respondents vote between two candidates with identical policy preferences. This assumption, though not natural in all situations, provides a necessary preliminary to a more general analysis and may be reasonable in some circumstances (for instance in parties' primaries, in which policy differences might be negligible). With this design, the impact of background personal characteristics on individual decision-making in voting choices is controlled for, and the effect of weather on individual choices can be imputed to its effect on the individual level of risk aversion caused by mood. 3 Expectations and Conjectures The expectations about the experimental results can be described by the following hypotheses. 1. The vote share for a risk-free candidate is larger than the vote share for a risky candidate who yields the same expected utility in a positive prospect. I expect this finding, because according to both the standard expected utility theory and prospect theory, agents are risk-averse in the positive domain. So, when an individual is presented with two candidates who are identical in all dimensions but for the degree of risk that they yield, individuals are expected to choose the one candidate who carries the least risk. 2. The vote share for a risk-free candidate is larger in the domain of gains (positive 7 prospect) than in the domain of losses (negative prospect). According to prospect theory, agents are more risk-averse in the positive domain. So, voters are expected to choose the one candidate who carries the least risk more frequently in the positive prospect than in the negative one. 3. The vote share for a risk-free candidate is smaller than the vote share for a risky candidate who yields the same expected utility in a negative prospect. According to prospect theory, agents are not only more risk-tolerant in the negative domain, but they actually seek risk. Voters are therefore expected to choose the one candidate who carries the most risk, but who can yield a positive or higher outcome. 4. The vote share for a risk-free candidate is larger in bad weather days than in good weather days. Regardless of the prospect framing, I expect bad weather conditions to positively affect the likelihood of voting for the riskier candidate and viceversa. 5. Exposure to bad weather is positively correlated with bad mood and vice versa. This is the standard psychological prediction, and I expect results consistent with those of Denissen et al. (2008). 6. Good (bad) mood is positively correlated with risk tolerance (aversion) behavior. This is the conventional mood-risk channel prediction, positing that anxiety and other negative mood states lead to higher risk-aversion while positive mood states produce a higher risk-tolerance (Eisenberg et al., 1998). To test these expectations, I implemented a controlled laboratory experiment, which I describe in the following section. 8 4 The Experimental Design This analysis investigates whether subjects' voting decisions, when candidates' performance is uncertain, differ when the weather conditions are perceived as favorable or poor. To examine whether risk preferences are associated with objective and/or perceived weather conditions, I ran a controlled experiment in which subjects were exposed to different weather treatments. To operationalize the weather treatments and control the assignment of subjects, I scheduled twin pairs of experimental sessions per week in days with diametrically opposed weather forecasts. Subjects could register to participate in the experiment only by registering for both of the twin sessions. Subjects were told that they would be ultimately selected to participate in one of the twin sessions, but that they could not choose which one. Subjects were randomly allocated by the experimenter to one of the two sessions. The experiment was conducted by paper and pencil in a large classroom of the Kenan-Flagler Business School that allowed for exposure to the outside weather conditions. The same classroom was used in all experimental sessions and all sessions were run at the same time of the day (from 2:00 pm to 3:30 pm). A total of 199 participants has been recruited from December 2011 to January 2013, with 95 participants allocated to the bad weather treatment and 104 to the good weather treatment. Out of this pool, 166 subjects actually participated in the experiment, with 81 subjects in the bad weather treatment and 85 in the good weather treatment. The participation rate was very similar across weather treatments, at 85.2% and 81.7% for the bad and good weather, respectively.1 The overall design consists of a lottery choice experiment with a 2 x 2 design, with two weather treatments (good and poor weather) and two prospect treatments (pos- 1Tables Al and A2 in the on-line Appendix report the demographics of the sample on which the experiment was conducted. Even though the majority of the participants were college or graduate students, the sample appears to be evenly distributed with regard to age, gender, racial group, income, political leaning and religiousness. This makes this experiment an ideal laboratory to test our hypothesis about the effect of weather on voting, after controlling for other personal characteristics. 9 itive and negative prospect). A between-subjects design has been used for the weather treatments: subjects were randomly allocated to participate in only one of the weather treatments. Furthermore, a within-subject design has been used for the prospect treatments: every subject participated in both prospect treatments sequentially. The order of the prospect treatments has been randomized to eliminate any order effect. Before computing payoffs, subjects were asked to complete an affect scale form about their mood (PANAS-X), and at the end of the experiment and before being paid, subjects were asked to complete a questionnaire about socioeconomic characteristics such as: i) general information about the subject; ii) individual and family income and education; hi) health; iv) religion; v) political views; and vi) economic assessments. The Appendix provides specific details about the experimental procedures and the instructions distributed to the subjects. 4.1 Prospects and Decisions The decision problems in this experiment build on those of the classical experimental study by Quattrone and Tversky (1988). The design focuses on framing political decisions in a domain of gains or losses. A reference point is induced by providing the subjects with the level of wealth - measured through the Standard of Living Index (SLI)- of other comparable countries. In this way the wealth of other countries appears to the subjects to be a feasible and reasonable goal to achieve or surpass: if they are satisfied with the domestic level of wealth, they would consider the decision in a domain of gain; otherwise, they would consider the decision to be in a domain of loss.2 In the experimental surveys conducted by Quattrone and Tversky (1988), respondents are asked to imagine facing a voting decision between two candidates, who 2Heath et al. (1999) claim that goals serve as reference points and systematically alter the value of outcomes, as described by the psychological principles in the prospect theory's value function. 10 are known to favor different economic policies which would affect the respondents' wealth. Experts provide different forecasts of the SLI in case the two candidates should win the election: forecasts for one candidate are mostly consistent among economic experts, while forecasts for the other candidate are more diverse. This feature implies that although both candidates are risky, the latter is more risky than the former. 4.2 Risk Preference Elicitation The experimental design improves on the Quattrone and Tversky (1988) design in two ways. First, experimental data are collected not by means of hypothetical questions, but by means of actual decisions that affect the remuneration of the subjects. This approach improves the salience, and thus the internal validity, of the experiment. Second, the experimental design is not limited to the analysis of a reversal in the voting decisions of the respondents as a function of a reversal of the reference point: rather, the design aims to analyze voting decision as a function of the risk associated with the two candidates in both domains. The elicitation procedure used to ascertain voting decisions as a function of candidates associated risk in the experimental laboratory is a variation of the Multiple Price List (MPL) design.3 The original MPL design entails giving the subject an ordered array of binary lottery choices to make all at once. The MPL requires the subject to pick one of the lotteries on offer; then the experimenter plays that lottery out for the subject to be rewarded. The design used in this experiment departs from the original MPL in two ways. First, the respondents are not given the binary lotteries (elections between two candidates) in an array; instead, they make each binary lottery choice in a sequence. In this way, consistency and monotonicity of voting de- 3The earliest use of the MPL design is by Miller et al. (1969). Later used by Schubert et al. (1999), and by Holt and Laury (2002), the method became the standard procedure to elicit and measure risk attitudes. 11 Table 1 Payoff Tables - Risk Aversion Elicitation Candidate C's projected sli Candidate Ps projected SLI Election 1 : Election 2 : Election 3 : Election 4 : Election 5 : Election 6 : Election 7 : Election 8 : Election 9 : Election 10 $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $66,000 w.p 50% , $42,000 w.p 50% $42,000 w.p 50% , $42,000 w.p 50% $46,000 w.p 50% , $46,000 w.p 50% $48,000 w.p 50% , $48,000 w.p 50% $50,000 w.p 50% , $50,000 w.p 50% $52,000 w.p 50% , $52,000 w.p 50% $54,000 w.p 50% , $54,000 w.p 50% $56,000 w.p 50% , $56,000 w.p 50% $58,000 w.p 50% , $58,000 w.p 50% $60,000 w.p 50% , $60,000 w.p 50% $62,000 w.p 50% , $62,000 w.p 50% cisions is not produced as a procedural artifact of the design. Second, as one of the options entails no risk, the variant of the MPL instrument that is adopted (as developed by Schubert et al., 1999) entails that the respondents choose between a risk-free option ("candidate I", characterized as the incumbent), whose payoffs are fixed, and a risky option ("candidate C", characterized as a challenger), whose payoffs are determined by his chances of success or failure. Thus, risk aversion in this context can be equivalently interpreted as status quo bias. While the payoffs attached to candidate C remain fixed in each election, the "secure" payoff attached to candidate I increases monotonically moving from one election to the next. As the amount of the secure payoff grows, choosing candidate C looks less attractive to a risk-averse respondent, who would switch to the safer candidate. Table 1 describes the variant of the MPL design used. This experiment departs from a classical political economy experiment in that it aims to analyze behavioral responses to environmental conditions, rather than context-free theoretical expectation. The use of a context in this case is desirable, in that it provides clues to help subjects interpret the task as a voting decision rather than as an abstract choice (Woon, 2012a, discusses the desirability of contextualized scripts in voting experiments). To contextualize the voting decision in a setting where one can- 12 Table 2 Risk Aversion Elicitation - Prospects Positive Prospect Other 4 nations' sli Negative Prospect Other 4 nations' sli Election 1 $42,000 $66,000 Election 2 $42,000 $66,000 Election 3 $42,000 $66,000 Election 4 $42,000 $66,000 Election 5 $42,000 $66,000 Election 6 $42,000 $66,000 Election 7 $42,000 $66,000 Election 8 $42,000 $66,000 Election 9 $42,000 $66,000 Election 10 : $42,000 $66,000 didate is considered more risky than the other, the two candidates have been labeled as "challenger" and "incumbent", respectively4 Payoffs were decided by a fair throw of a twenty-sided die and a coin. In Election 1, the first (second) payoff is paid if the subjects has chosen candidate C and the coin has landed heads up (down); the third (fourth) payoff is paid if the subject has chosen candidate I and the coin has landed heads up (down). As one proceeds down the matrix, the payoffs and attached probabilities of candidate C remain the same, but the payoffs attached to candidate I change. The matrix of ten election scenarios for each prospect is designed in such a way that only extremely risk-seeking subjects choose candidate C in the last row, and only extremely risk-averse subjects choose candidate I in the first few rows.5 A risk-neutral subject would choose candidate C as long as the expected utility of candidate C is higher than the expected utility of candidate I (in the first five elections), and candidate I otherwise 4a manipulation check treatment with abstract candidate labels has been run to test for possible framing effects. The results show no significant difference between the framed and the abstract treatment. 5Notice that rational players should always choose candidate c in the first election, because it weakly dominates the alternative option: it never yields a lower payoff and it yields a strictly larger payoff with a positive probability. Thus the first election is considered as a control that the subject has understood the instructions. 13 (last 4 elections). Table 2 describes the design used for the prospect treatments. While payoffs remain exactly identical to the ones described in Table 1, gain and loss domains are created by contextualizing the voting decision internationally The subjects are given information about the relative wealthiness of other comparable countries by making the other countries appear to be the reference point. 5 Prospect treatment results I now investigate the effect of prospects on voting decisions (e.g., hypotheses 1, 2 and 3). Figure 1 reports the raw results of the prospect treatment. Consistent with the extant literature, subjects display a considerable degree of risk aversion in the positive prospect, in that they switch from the risky to the safer candidate much earlier than would a risk neutral agent. As shown by Figure 1, in election 6, around 75% of the subjects vote for the safer candidate, as against 25% for the risky candidate, even though the two candidates are identical in terms of expected utility (e.g., hypothesis 1). Even when the safer candidate yields a lower expected utility (election 5), subjects still prefer to vote for him rather than for the risky candidate (around 65% of the subjects vote for the safer candidate). Subjects seem, however, to also be risk-averse in the negative prospect. Figure 1 shows that in election 6, around 70% of the subjects vote for the safer candidate, contradicting hypothesis 3. However, the likelihood of choosing the safer candidate seems to be larger in the gain domain than in the loss domain in nearly every election (e.g., hypothesis 2), even though the domain's effect seems to be more modest, as compared to the previous studies of Quattrone and Tversky (1988). This variance is caused by a difference of incentives that the subjects face between this experiment and those in Quattrone and Tversky As reported by Laury and Holt (2008)6, the use of real 6Despite the widespread references to prospect theory in theoretical and experimental work, few 14 Elections FlG. 1 - The effect of prospects: the vertical axis reports the percentage of votes for candidate I. The horizontal axis reports the election number. The blue line refers to observations in the domain of positive prospects, while the red line refers to the domain of negative prospects. incentives dramatically reduces the incidence of reflection behavior around the reference point. However, considering that the positive or negative domain does not affect the subjects' earnings, by inducing only a psychological framing, the results provide support for the prospect theory: subjects are less risk-averse in the loss treatment than in the gain treatment. To test whether the effect of prospect (e.g., hypothesis 2) shown in Figure 1 is statistically significant, I calculated the average frequencies of votes for candidate I (the safer choice) across all subjects for the two prospect treatments. Across all ten elections, subjects vote for the safer candidate on average 5.96 times in the positive prospect, and 5.55 times in the negative prospect. The difference is statistically sig- studies have tested the theory with incentivized tasks (Kahneman and Tversky, 1979 and Tversky and Kahneman, 1992 are based on hypothetical payoffs). Laury and Holt (2008) use a simple tool to measure risk preferences directly, based on a series of lottery choices with significant money payoffs in parallel gain and loss treatments. 15 nificant according to a Welch test, which tests the null hypothesis that the average number of votes for the safer candidate in the negative prospect (fineg) exceeds the average number of votes in the positive prospect (fipos)- Denoting as