Motivating Risky Choices Increases Risk Taking


 We study the impact of the mode of cognition on risk taking. In an online experiment we ask participants to make a simple decision involving risk. In the control group no manipulation is made, while in the treatment group we exogenously manipulate the mode of cognition by requiring subjects to write down a text that motivates their risky choice before any action is actually taken. Such motivation treatment is meant to induce more reflection upon the action to be taken. Our results show an effect of the motivation treatment on risk taking, suggesting that higher reflection makes subjects more prone to risk taking. The effect is stronger if we consider only subjects who imperfectly understand the probability distribution implied by the simple choice task. Based on our experimental findings, we suggest that reflection and comprehension might be substitutes when individuals make decisions involving risk.


12
In many real-life situations people make decisions intuitively with barely no effort, while in other situations 13 they exert a substantial effort to make more conscious and reflected decisions. Given the pervasiveness of choices involving risk to be more in line with risk aversion 3 . In particular, in the domain of gains, people tend to be more risk averse if their choices are more intuitive. Although this prediction might be consistent with 23 the finding that subjects' Cognitive Reflection Test scores are inversely related with risk aversion 4 , results 24 from other researches focusing on the causal effect of the mode of cognition on risk taking are not conclusive. 25 Time pressure on decision-making leads to more risk aversion 2, 5 , and cognitive load is associated with more 26 risk-averse behaviors 1 ; however, reliance on intuition appears to increase risk tolerance 6 , and arousal increases 27 risk taking 7 , suggesting that reliance on intuition is not necessarily associated with more risk aversion. Finally, 28 depleting self-control seems to have negligible effects 8 . Further, there is evidence that lower reflection is 29 correlated with higher probabilities to participate in risky activities 9 . Thus, it is fair to say that the actual effect 30 of cognition on risk taking seems to depend on the method which is employed to manipulate the mode of 31 cognition. 32 We contribute to this literature by investigating how risk taking is affected by greater reflection induced 33 by the request to motivate one's decision. Such method has never been applied to study this issue. We 34 run an online experiment where we ask participants to make a simple decision involving risk: the "Bomb 35 Risk Elicitation Task" (BRET) 10 , which has been recently used in a number of studies to measure risk 36 taking behavior 11-14 . We attempt to manipulate the extent of reflection by means of a motivation treatment: 37 participants are required to motivate their choice with a written text before any decision is actually made.

38
Online experiments are characterized by shorter procedures and lower stakes with respect to laboratory 39 experiments, which reasonably increase the likelihood that participants make quick and intuitive decisions 15 .

40
On the one side, this suggests that inducing greater reflection -as we attempt to do -can produce greater 41 effects than in typical laboratory experiments. On the other hand, there is the risk that experimental subjects 42 put too little effort in the experimental task for any effect to emerge. For instance, experimental subjects 43 on Amazon Mechanical Turk have been shown to devote limited attention while performing online tasks 16 .

44
To avoid this, our experiment was designed to be quick, graphically informative, and engaging, in order to 45 minimize the the risk that experimental subjects put little attention. For this reason we applied the BRET with 46 graphical representation 14 instead of other measures of risk preferences applied in the literature which involve 47 Ours is the first experiment on risk taking behavior where cognition is manipulated by means of the 57 recently developed method where subject are required to motivate their decision with a written text 21 . While it 58 has yet to be established whether such motivation treatment is more or less effective in inducing reflection than 59 traditional ones, like time delay 2, 22-26 or priming 6, 24, 27, 28 , it does have been proved to be easily implementable 60 and to work properly in an online setting 15, 29 .

62
Out of 398 participants, 9 are excluded from the dataset before the analysis, as 2 subjects failed the control 63 question (1 in the control group and 1 in the treatment group), and 7 subjects decided to open 100 boxes (4 64 in the control group and 3 in the treatment group), which is a dominated strategy and therefore likely to be samples. Mann-Whitney tests cannot reject the hypothesis that the control variables have the same distribution 73 in the two groups (gender z = 0.034, p=0.97; age z = 0.789, p=0.43; self-assessment of risk preferences, 74 z = 0.564, p=0.57). Also, the fraction of correct answers to the comprehension question -concerning the 75 probability of getting the bomb in the BRET -is not substantially different in the control group (37.5%) and 76 in the treatment group (41.3%) (Fischer's exact test, p=0.468).

77
The treatment effect 78 To assess risk taking behavior we use the number of boxes opened by experimental subjects in the BRET, 79 ranging between 0 and 100 (also referred to simply as "boxes"). The greater the number of boxes opened, the 80 greater the risk taken.  in the group of those who gave a correct answer (right bar). No appreciable difference is found between the two groups (statistics reported for Mann-Whitney tests of equal distributions).

4/12
To check if experimental subjects understood the probabilities involved in the BRET, at the end of the 87 experiment we asked them "If you have collected 35 boxes, what is the probability of getting the bomb?" 88 and they had to enter manually a number (see Slide 5A in the Supplementary Information). The answers 89 to this comprehension question were not appreciably different between the control and treatment groups. 90 Moreover, as shown by Figure 2, the average number of boxes opened was not appreciably different between 91 experimental subjects who answered correctly to the comprehension question and those who did not.

92
Yet, we found appreciable differences in the average number of boxes opened in the control group ( Figure   93 3, left chart) when we compare experimental subjects who answered correctly to the comprehension question 94 and those who did not (Mann-Whitney test, z=2.54, p=0.011). In particular, experimental subjects who gave a 95 wrong answer opened, on average, less boxes than those who did answer correctly. In contrast, we found no 96 difference when we do the same comparison for the treatment group (Mann-Whitney test, z=1.10, p=0.274).

97
This finding suggests that the treatment and the comprehension of the probabilities involved in the BRET 98 may have interacted in some way. To explore this possibility we also looked at the treatment effects within 99 the group of experimental subjects who answered correctly to the comprehension task and those who did not.

100
Consistent with this idea we find that the treatment effect is appreciable only for the latter group (see Figure 3,  In order to assess the joint statistical significance of what can be inferred from the previous non-parametric 108 analysis, we run a series of regressions (reported in Table 1). We use linear regressions (OLS) where the de- No appreciable difference is found between the two groups for the treatment group, while in the control group we find that more boxes are opened by those who answered correctly to the comprehension question (statistics reported for Mann-Whitney tests of equal distributions). Right chart. Number of boxes opened by experimental subjects split in the group of those who gave a wrong answer to the comprehension question regarding the probabilities involved in the BRET (left bar) and in the group of those who gave a correct answer (right bar), further divided by control and treatment groups. No appreciable treatment effect is found for the group of those who answered correctly to comprehension question, while a strong treatment effect is found (from 34.5 to 45.7 boxes) for those who answered wrongly (statistics reported for Mann-Whitney tests of equal distributions). Motivation is a dummy variable taking value 1 if the subject is in the treatment group; Comprehension is a dummy variable taking value 1 if the subject has correctly answered the question about the probability implied by the BRET; Female is a dummy variable taking value 1 if the subject is a woman; Age is equal to the number of years of the subject; Self-reported willingness to take risk is a variable between 0 and 10 where 10 is the maximum willingness to take risk. Robust standard errors are reported in parenthesis. Statistical significance is indicated as follows: * p < 0.1, * * p < 0.05, * * * p < 0.01.

7/12
becomes statistically significant. Moreover, the estimated coefficient of interaction between treatment and 120 comprehension is about -9.6 and statistically significant. Overall, these estimates confirm that the treatment 121 effect is quite stronger among experimental subjects who did not answer correctly to the comprehension 122 question and, further, that in the control group comprehension led to take more risk. These findings confirm 123 our main result that the motivation treatment effectively increases risk taking in the BRET and, moreover, they 124 suggest that the motivation treatment and the comprehension of the probabilities involved in the BRET are, at 125 least to some extent, substitutes.

126
In Model (3) we add as controls gender, age and the self-reported willingness to take risk. Results

127
concerning the variables included also in Model (2) are substantially the same as in Model (2). It is worth 128 noting that while gender does not seem to play any role, more aged subjects tended to open more boxes (one 129 more for 3 additional years of age) as well as subjects who declared greater willingness to take risk (three 130 more boxes for each level of willingness, ranging from 0 to 10).

132
In this paper we explored experimentally the effects of inducing greater reflection on risk taking. Although 133 this has been investigated in previous studies, the evidence collected so far is mixed and suggests that much 134 depends on the method applied for manipulating cognition. We add to the ongoing discussion by providing 135 evidence from an online experiment where we attempt to manipulate cognition by means of a motivation 136 treatment, namely by requiring experimental subjects to write down a text that motivates their choice before 137 they can actually take action.

138
Our main finding is that the motivation treatment induces more risk taking, as measured by the number of 139 boxes opened in the BRET 10 . This suggests that greater reflection makes subjects more prone to take risks 3 .

140
However, we also find that the treatment effect is sizeable for the subjects who did not answer correctly to the 141 comprehension question regarding the probability distribution implied by the BRET, while the effect almost 142 disappears for subjects who gave the correct answer. Moreover, while the motivation treatment does not 143 appear to affect comprehension of the probability distribution implied by the BRET, the latter seems to go 144 with more risk taking only for the control group. Overall, these findings suggest that the motivation treatment 145 and the comprehension of the probabilities involved may be substitutes in promoting risk taking.

146
Our results could perhaps be explained with reference to the reduction of ambiguity brought about by 147 greater reflection. Subjects with imperfect probability understanding who also reflect little on their decision

164
The experiment was conducted using oTree 37 . Data were collected in a single session in March 2019.

165
On Prolific, the experiment was labeled "An experiment on decision making", and was described as follows: 166 "This is an experiment on decision making. We will ask you to complete a quick task, which may allow you to 167 earn additional payments, and a short questionnaire". The sample was restricted to subjects from the UK and 168 the US, in an age between 18 and 35. A minimum of two submissions in previous studies, with at least a 50% 169 approval rate, was also imposed. We gave an estimate of three minutes for the time needed to complete the 170 experiment, while we set to 10 minutes the maximum time for completion. Subjects received a show up fee of 171 0.30 GBP.

172
All participants gave their informed consent at the beginning of the experiment, and they were given 173 instructions about the task to be performed. Payoffs were automatically converted in USD for participants 174 from the US.

175
To measure risk taking behavior we employed the BRET 10 which is increasingly applied in the experimen-176 tal literature 11-14 . In our implementation of the BRET (for which we used a pre-programmed tool for oTree 38 ) 177 subjects had to choose how many boxes to collect from a 10x10-grid containing 100 boxes. They were told control question to verify data validity (subjects were asked: "If you're reading this check 'Agree little"', and 189 they had to check as indicated; see Slide 6 in the Supplementary Information).

190
Informed consent was obtained by all experimental subjects. In particular, the participants were informed 191 that data would be used anonymously for scientific purpose only. The experiment was conducted online  Number of boxes opened by the experimental subjects splitted in the group of those who gave a wrong answer to the comprehension question regarding the probabilities involved in the BRET (left bar) and in the group of those who gave a correct answer (right bar). No appreciable difference is found between the two groups (statistics reported for Mann-Whitney tests of equal distributions).

Figure 3
Left chart. Number of boxes opened in the control and treatment groups, splitting experimental subjects in the group of those who gave a wrong answer to the comprehension question regarding the probabilities involved in the BRET (left bar) and in the group of those who gave a correct answer (right bar). No appreciable difference is found between the two groups for the treatment group, while in the control group we nd that more boxes are opened by those who answered correctly to the comprehension question (statistics reported for Mann-Whitney tests of equal distributions). Right chart. Number of boxes opened by experimental subjects split in the group of those who gave a wrong answer to the comprehension question regarding the probabilities involved in the BRET (left bar) and in the group of those who gave a correct answer (right bar), further divided by control and treatment groups. No appreciable treatment effect is found for the group of those who answered correctly to comprehension question, while a strong treatment effect is found (from 34.5 to 45.7 boxes) for those who answered wrongly (statistics reported for Mann-Whitney tests of equal distributions).

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. bibospSI.pdf