1
The effect of IQ on risk taking in a financial
context: Evidence from Bolivia
2 Abstract
Our paper examines the effect of IQ on risk taking in a financial context by using a 1876 household dataset from Bolivia. We approach risk taking both as a one step process and a two-step process by using a Tobit model as well as an Heckman model. We find that if treated as a one-step process IQ significantly and positively influences risk taking. For a two-step process, we conclude that IQ is significant and positive for the engagement in risky behavior, but insignificant for the amount of risk taken.
Introduction
Our paper examines the effect of IQ on risk taking in a financial context. We use a 1876 household dataset from Bolivia to test whether IQ significantly influences the engagement in risk taking as well as the amount of risk taking. The unique survey features a risk game which reveals risk preferences as well as an IQ test. We combine this data with control variables on gender, age, income, education and health.
We take two different approaches with respect to testing if IQ influences risk taking. We start with a one step process, with the amount invested in the risk game as the dependent variable .We assume here that the participant faces one decision, namely how much money they want to invest in the risk game. Secondly we approach this question from a different angle, by dividing the process up in two steps. We argue that participants make two decisions. They first decide whether they want to engage in risky behavior, whether they play the risk game or not. Once participants decide to play, the amount of risk is adjusted to the personal risk preferences and participants decide how much they invest in the risk game. To accommodate this setup we use a Heckman model (Heckman 1976), with the risk game as the dependent variable in the first equation, and the amount invested in the game as the dependent variable for the second equation.
Three distinguishing features between previous research and this paper are the use of a new extensive dataset from Bolivia, approaching risk taking as a two-step process and using IQ as a risk driving variable. The dataset, obtained by field researchers in rural areas of Bolivia contains around 1900 households. Data is collected on many topics, including an IQ test with the use of Raven’s matrices (Raven 1936). This richness of data provides the basis for thorough analysis. Besides this, we are one of the first in this field to treat the risk taking process as a two-step process where one first decides whether or not to engage in risky behavior before the amount of risk is chosen. Finally we are one of the first papers to our knowledge to focus on IQ as a potential risk driving variable. We contribute new information to the vast body of literature on risk taking.
3
The paper continues as follows: We provide an overview on theories and literature with respect to our setup for IQ as well as the control variables age, income and intelligence. We state hypotheses in the methodology section, where we elaborate on the working of the risk game as well as the choice for the Tobit model and the Heckman selection model. The data section follows. We then present results to both approaches and the corresponding models. We end this paper with conclusions, limitations of our work and suggestions for further research.
Literature Review
In order to get an overview on the question how IQ influences risk taking behavior in our financial setting, we look at the literature from multiple perspectives. We discuss the seminal views with respect to IQ and risk taking, but also discuss the literature on our control variables gender, age and income. The possible influence of different factors combined with different views ensures that a broad range of literature will be reviewed in order to draw meaningful conclusions. We are deliberately choosing to cover multiple possible influential factors. In the first place, we want to create a framework to test the influence of IQ and arrive at a hypothesis. Secondly, we cover literature on control variables to get an idea of the possible significance and coefficients of these variables. The relation of our results compared to the results predicted by accepted theories in the literature might tell information about our model and the quality of it. All in all, the aim of this review is to provide a framework from which we can perform our analysis. We will start with a definition of risk taking, continue with literature on the effect of IQ on risk taking, and end with the review of literature on the control variables age, gender and income.
4 Intelligence
Another topic of interest for our research and institutions concerned with lending is the intelligence of borrowers. Although information on this factor can be relatively easily obtained, through diplomas or a simple IQ test, we lack a large body of literature focusing on the relation between risk aversion and intelligence.
Dohmen et al. (Dohmen et al. 2010) use a sample of roughly 1000 German adults to test whether there is a relation between cognitive ability (IQ) and risk aversion. The randomly selected participants made two cognitive ability test, filled in a survey on background information and were randomly selected to participate in either a lottery experiment, where they decide between a safe bet or a lottery, or an intertemporal choice experiment, which revolves around finding out about the patience of participants. The lottery experiment is relevant for our research, as it reveals risk preferences. 20 questions with safe payoffs and risky bets have to be answered. The results reveal that cognitive abilities and risk preferences are related. Higher cognitive abilities result in less risk aversion and vice versa. The results hold when robustness checks are performed. (Dohmen et al. 2010) Another finding of the paper is that higher cognitive abilities result in more patience. Although there is no direct link to our research, it might point in the direction of future decisions from borrowers.
In a later study, Dohmen et al. (Dohmen et al. 2011) look among other factors in a large study based on 22000 German individuals at the effect of the parental level of education on risk aversion. They find that a passed Abitur by a mother or father, which is the German equivalent of a high school exam that allows students to go to university if passed, has a positive effect on the risk taking by children. Although this is not a direct measurement of one’s own cognitive abilities, there is a correlation between the cognitive abilities of parents and children. It is another indication that the relation between risk aversion and IQ is negative.
Zuckerman (Zuckerman 1994) finds a significant correlation between the Wechsler Adult Intelligence Scale, a full scale IQ test, and sensation seeking derived by the sensation seeking scale (Zuckerman et al. 1964). The correlation for a sample of 138 high school students is 0.22, and for 80 drug abusers 0.29. Both are significant at a 1% significance level. However, although we find it worthwhile to mention these results we do not want to attach too much value to them. First of all, the correlation exists between IQ and sensation seeking, which is not the same as risk taking, although the two are related. Furthermore, the significant correlations are small, and correlation in itself is not a strong enough foundation to draw conclusions. We do conclude that there is no evidence that a higher level of intelligence leads to less risk seeking.
5
more in response to the changing likelihood of a win, although the latter effect does not survive correction for multiple comparisons.” (Deakin et al. 2004). The correlation coefficient for IQ and Deliberation time is -0.23 and between IQ and risk adjustment the coefficient is 0.15. Both are significant at a 95% confidence interval. Again we refrain from drawing strong conclusions, but more intelligent participants appear to be more capable at understanding the mechanics of the gamble and adjusting to outcomes.
A theory related to these results is the cognitive load theory (Sweller 1988). A part of the theory revolves around differences in the ability to solve problems between experts and novices. The main difference in these abilities lies in knowledge about the specific domain of the problem. Experts know the type of problem at hand, and are able to apply so called “schemes”, basically structures to solve the problem in an efficient way. Novices do not possess the knowledge to immediately recognize appropriate structures to solve the problem, which causes a higher cognitive load, or mental capacity used, to solve the problem. If a heavy cognitive load occurs, the task at hand is performed worse (Paas 1992). An unknown task , limited cognitive abilities or a combination of these factors can lead to heavy cognitive loads, which can worsen the performance of the task. For our research, participants unknown with the risk game could perform worse. This can result in not playing the game, or playing the game with a suboptimal choice of the amount of risk.
Although the body of work on the specific relationship between intelligence and risk aversion is small, most papers focus on closely related topics, and some papers only address the topic with a sample size of young students or high schoolers, we do see a pattern. There do not appear to be any results that point in the opposite direction, being that a lower IQ leads to more risk taking. Therefore we expect with some reservation that IQ has a significant positive effect on risk taking.
Gender
A variable that comes up many times in risk taking literature is difference in risk taking behavior between men and women, or gender differences. Being easily applicable in the risk assessment of financial institutions and microfinance firms it is a potential valuable control variable. A starting point is the meta-analysis performed by Byrnes et al. (Byrnes et al. 1999) that looked at over 150 studies in the field of gender differences regarding risk taking. Most of these studies contained more than one comparison, resulting in a total of 322 effects. A coding of the studies in 3 categories is performed, hypothetical choice, self-reported behavior and observed behavior. With respect to our research the observed behavior is the most interesting. It represents studies were participants were observed performing behavior first hand, similar to our setup. Finally the studies were classified according to the type of tasks that the participants perform.
6
with respect to certain subjects ,such as driving, stayed relatively small across all ages for others, including smoking and decreased with a higher age level for others. All in all this leads to the following conclusion: Generally speaking males take more risk, but care is required when applying this rule to specific situations and specific age levels. Besides these conclusion the authors also found opposite behavior in risky situations between men and women. The data shows that men appear to take risks even when the outcome is likely to be negative, whereas women tend to avoid risks even when the outcome is relatively favorable. This would not only imply a loss of welfare on the side of men as the result of taking too much risks, but also a loss of welfare for women due to a lack of risk taking.
On a more specific level we use gender difference in a financial setting. Both Powell and Ansic (Powell & Ansic 1997) and Dwyer et al. (Dwyer et al. 2002) address this topic. Powell and Ansic use two computerized experiments to test whether risk taking differences stay significant when accounted for task type and task familiarity. Dwyer et al. use a different approach and perform analysis on data acquired from 2000 mutual fund investors. Powell and Ansic test by using an insurance experiment with n=126, which revolved around choosing an insurance cover, and a currency market study with n=101 and participants deciding on staying or leaving the currency market . The insurance study was a task that the participants were familiar with, while the currency market study was unfamiliar to the participants. They conclude that in a management setting there is indeed more risk taking behavior by men than by woman, controlled for task type, familiarity with the task on hand and framing. This paper in the setting of financial decision making finds similar results as find in their study. It has to be noted that the sample was highly educated. Dwyer et al. reach a similar conclusion, in which their analysis shows that among mutual fund managers men take significantly more risk than women. However, their regression which included a “financial investment knowledge” variable shows more dampened results, implying that the gender difference partly disappears if financial knowledge is taken into account.
All in all the literature points strongly in the direction of a general gender difference in risk taking, with women taking less risk than men. However, this generality cannot be applied to any task and the difference does not show a clear pattern across task type and familiarity of the task with the participants.
Age
7
a sample performing the tests in the presence of a peer. Firstly the authors conclude that there is indeed a strong decline in risk taking between the adolescents and the adults. Besides this adolescents have also a higher risk preference than adults, and take more risky decisions. Another relevant conclusion is that peer presence does increase risky behavior in general, but significantly more during adolescence. In relation to our research, we should both take the increased risk taking of younger participants into account as well as the heightened influence of peer pressure. Adults take less risk and base their decisions largely on their own preferences, rather than being influenced by peers.
Caufmann and Steinberg (Cauffman & Steinberg 2000) base their research around the question how mature participants are in their judgements, based on assessments that included hypothetical decisions with possible anti-social and risky outcomes. The sample contained more than 1000 participants aged between 12 and 48. Caufmann and Steinberg view maturity of judgements as a
mix of cognitive and psychosocial factors. The sample contains 5 age brackets, with the 8th, 10th
and 12th graders being around 13.5, 15.5 and 17.5 years old. The other two brackets contain college
students of over and under 21 years old. A point of concern in this study is the gender bias in the college bracket, with 83% and 73% being females. Other brackets do not suffer from this issue. The results reveal that antisocial behavior decreases as the age of the respondents increases. Furthermore across all age brackets the amount of antisocial behavior is highest when there are no perceived consequences. Interestingly, the study also finds a significantly higher level of anti-social decision making for males than females, which supports the previous literature covered. Another finding was that higher levels of responsibility, temperance and perspective resulted in lower levels of antisocial behavior. This psychosocial maturity is also significantly related to age. The two final conclusions that are relevant for our research, is that antisocial decision making depends more strongly on the psychosocial maturity than on age. Besides this, social behavior does not increase in the college students brackets over and under 21. This would imply that age is not the focus point of risk, but psychosocial maturity, and that the issue of increased antisocial behavior disappears around adulthood. A point to note here is that in the highest age bracket the mean age is 25, so the effects of a high age level are not tested here.
8
from decisions from description is that there is no significant age difference in risk taking. All in all we conclude that there is too little evidence for differences in risk taking among adults.
Summarizing the findings of the above papers we conclude the following: There is a heightened level of antisocial behavior and risk taking among children and adolescents compared to adults, but once adulthood is reached the level of risk taking appears to be stable. Besides this, peer pressure and cognitive differences also play a smaller role once adulthood is reached. Since the vast majority of our participants can be classified as adults, we do not expect the age variable to significantly influence the level of risk displayed in our game.
Income
The possible effect of income on the amount of risk people are willing to take is often explained by common sense. There is a general consensus that more wealthy individuals can afford to engage in financially risky behavior since this group can incur a loss with smaller consequences than a poor person. Engaging in financially risky behavior would be more difficult if one cannot afford to lose money due to financial constraints. If we take the perspective of participants at the risk game, a similar logic is applicable. The 20 bolivianos are more needed by the participant with the lower income than the participant with a high income. This would result in a higher invested amount for participants with a higher income, since the consequences of losing the gift partly are limited. These consequences of a loss would be potentially bigger for a low income individual, resulting in a lower investment. In other words, the value of the gift is larger for a poor person than a rich person, and therefore the poorer person will be more careful than the rich person with investing the money in a risky bet. Absolute risk aversion increases with poverty.
9
income uncertainty and liquidity constraints. We expect that the income has a positive and significant sign with respect to our research question.
Methodology
In order to test our research question in an empirical manner we use data from the field granted to us by mister Cecchi and mister Lensink from the university of Groningen. The use of empirical data has the advantage of being a real world representation of behavior, rather than a computerized or hypothetical experiment. Another advantage with respect to our risk game is the use of real monetary payoffs. Since participants keep the money that is left after the risk game has been played, their observed behavior is an actual representation of their risk preference. Within the large amount of survey data available we use variables relevant for our research questions. These can be directly observed for certain variables, such as IQ, gender, age, income, years of education, whether or not the risk game is played and the invested amount in the risk game. For other variables we transform data into binary answers and cluster questions into proxies. Examples of variables created in this fashion are financial literacy, long and short term risk engagement and mental and physical happiness. A more detailed explanation of each variable and the creation of these variables can be found in the data section. We look at the data and variables for the person within a household that played the risk game. A large proportion of the data acquired remains unused in our research. However, this does not hamper the validity of our research, since data is acquired on a large variety of topics, often unrelated with the research question addressed here.
“ The Risk Game”
10
with, which is either the national emblem of Bolivia on one side and the number 2 on the other side.
The payoff structure is set up in such a way that a risk neutral person would invest 20 Bolivianos, with an outcome of 40 if the coin toss is won and a outcome of 10 if it is lost, leading to an expected payoff of 0.5*40+0.5*10=25. This is higher than not investing anything which guarantees a payoff of 20 Bolivianos, and therefore it is under the assumption of risk-neutrality optimal to invest 20 Bolivianos. X can range from 0 to 20. Every investment smaller than 20 but larger than 0 leads to an expected payoff larger than 20 but lower than 25. The formula is linear, so an investment of for example 10 Bolivianos leads to an expected payoff of 22.5 Bolivianos. Equation 1 shows the mathematical relation between investment and expected payoff. The expected payoff for every amount invested can be found in appendix A.
𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑦𝑜𝑓𝑓 = (0.5 ∗ 2𝑥) + (0.5 ∗ 0.5𝑥) + (20 − 𝑥) (1)
The participants are allowed to practice as many times as they want to understand the inner and outer workings of the game and decide whether they want to play or not and which side of the coin they want to play. Furthermore multiple numerical examples of the payoff structure are given to the participant until they are comfortable with how the game works and what risks they are taking. When everything is clear to the participant, the amount invested into the game is chosen as well as the coin side, the game is played one time, with either a win or a loss for the participant. After the real coin toss is played the game has ended.
The setup of the game is in line with the Gneezy and Potters method as described in Experimental
Methods: Eliciting risk preferences (Charness et al. 2013). It implies a setup that reveals
11 Alternative Methods
When using real life subjects in order to generate data there are multiple options to obtain this data besides the Gneezy and Potters method. Firstly there is the Balloon Analogue Risk Task or BART method (Lejuez et al. 2002). “The Balloon Analogue Risk Task measures risk preferences by presenting individuals with a computer simulation of pumping air into a series of balloons” (Charness et al. 2013). The participants decide at each step whether they want to pump an additional pump of air in the balloon, and with each pump the balloon grows, as well as the reward for the participant. They can stop at any time and if the balloon did not explode add the reward to their bank account. The amount of pumps added to the balloons reveals the risk preferences of the participant. This is played with a total of 90 balloons, with different breaking points. “This procedure was designed to model situations where excessive risk-taking leads to diminishing
returns and greater hazards.” (Charness et al. 2013) This method is not used in our research since a
computer is needed, which is impractical in rural areas, and the method is time consuming, which would greatly increase the difficulty of gathering a large sample size.
Another option would have been the well-known questionnaire where respondents answer to hypothetical statements. For example the willingness to participate in risky behavior. Although this method is simple, easy to understand and useable within our setup, we choose to not use this method. The reason is that the statements are hypothetical and there are no actual risks involved, which implies that the answers might not reflect the true risk preferences. However, no consensus has been reached in literature on whether financial incentives influence human behavior or not (Camerer & Hogarth 1999). Economist often take the view that people do not work for free, and therefore exhibit different behavior if financial rewards are involved. Psychologist on the other hand take the view that intrinsic motivation matters mostly, which devaluates the use of financial rewards (Camerer & Hogarth 1999). We follow the economist view in this matter, and to ensure that we can draw meaningful conclusions from our data, we do not use this measure.
Lastly the Eckell and Grossman method can be used (Eckel & Grossman 2002). This simple method presents different possible bets to participants, increasing in risk and return. Participants are allowed to only choose one of the options, which reveals their risk preference. We deliberately choose not to implement this measure, because it can only place participants in certain risk brackets and fails to create a continuous scale of risk taking. The method lacks the finesse that we try to achieve when interpreting our results.
12 Bolivian Income
The Worldbank (2016) states that the 2014 GNI per capita of Bolivia is 2870 US dollars a year. Converting this amount of Dollars to Bolivianos gives an annual mean income of roughly 20000 Bolivianos. To obtain the daily mean income, we calculate 20000/365, which gives a mean daily income of 54.80 Bolivianos. The reward of 20 Bolivianos represents 36.5% of the mean daily income. The US GNI per capita is 55230 US dollars a year, or 151.32 dollars a day (Worldbank 2016). A comparable reward in the US for such a survey would be 55.23 dollars. We will leave the relative value of this reward up to the reader of this paper. However in our opinion, the reward is substantial enough to carry value for the participant, and therefore a valuable variable to assess risk taking behavior.
Raven’s Progressive Matrices
We conduct our IQ test by using 19 Raven’s matrices (Raven 1936). Developed in 1936, it is a proven and widely used measure of IQ. The test involves pictures of geometrical patterns, that follow a certain pattern. The final picture of the pattern is unfinished, and from multiple answers the right pattern has to be chosen. We apply the Raven’s matrices by showing laminated matrixes and six possible answer pieces for the unfinished picture. One of the six answer pieces fits the matrix best at each question. Only one answer to each question is correct. The questions become gradually more difficult, in order to differentiate between levels of cognitive abilities. A higher level of correctly answered questions correspond to a higher IQ. One of the advantages of using this method is that once explained it does not involve the use of language nor writing, decreasing the chance of misinterpretation. Besides this, it is easy to use and carry for the researchers, and is relatively easy to explain.
Tobit Model
13
𝐴𝑚𝑜𝑢𝑛𝑡 𝐼𝑛𝑣𝑒𝑠𝑡𝑒𝑑 = 𝛼 + 𝛽1𝐼𝑄 + 𝛽2𝐴𝑔𝑒 + 𝛽3𝑌𝑒𝑎𝑟𝑠 𝑜𝑓 𝑒𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝐺𝑒𝑛𝑑𝑒𝑟 +
𝛽5𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽6𝑀𝑜𝑛𝑡ℎ𝑙𝑦 𝑖𝑛𝑐𝑜𝑚𝑒 + 𝛽7𝑃ℎ𝑦𝑠𝑖𝑐𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ + 𝛽8𝑀𝑒𝑛𝑡𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ +
𝛽9𝑆ℎ𝑜𝑟𝑡 𝑡𝑒𝑟𝑚 𝑟𝑖𝑠𝑘 𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡 + 𝛽10𝐿𝑜𝑛𝑔 𝑡𝑒𝑟𝑚 𝑟𝑖𝑠𝑘 𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡 (2)
Two step
Another possibility in approaching the risk game is splitting the process up in two steps. We split the risk game into two sub sections, firstly regressing if participants decide to play the risk game or not, and secondly conditional on playing the risk game, the amount of Bolivianos invested. We make this distinction because we take the view that the participant makes two choices. To begin the participant has to decide, after the explanation of the game and one or more trials with the coin toss, whether or not they want to engage in the act of risky behavior. After this first decision has been made, the participant that decides to take risk has to adjust the amount of risk to their personal preferences. We can perform specific test on risk preferences once the decision to engage in risky behavior has been made.
Heckman Model
The model we use to incorporate the two step process is the Heckman Model. This model relies on a differentiating variable between the two regressions in order to give meaningful results and function. The differentiating variables we use are loan offer 1 and loan offer 2. Loan offer 1 asks respondents whether they would take on a loan of 2000 Bolivianos that has to be paid back in one month. Due to the fact that the loan has to be paid back within a short amount of time, the question functions as a short term risk engaging proxy. Loan offer 2 poses the same question, but with a loan of 50000 Bolivianos and a payback time of 1 year. It represents a long term risk engaging proxy. Dropping the risk engaging proxies in the second equation is justified because the answers to these questions only influence the engagement in risk in a two-step setup. The effect of the risk engagement proxies will be captured by the first equation. The binary nature of the question, makes it effective ask a risk engaging proxy but invaluable in differentiating between levels of risk taken, since the questions are unable to “scale” the amount of risk. The following analogy applies. Participants answering yes to the risk engaging proxies will be more likely to play the risk game. Since the sample playing the risk game is likely to contain the people answering yes to the risk engaging proxies, there is no distinguishing value anymore in the second equation, since no additional information is obtained on the maximum size of the loan they would be willing to accept. One could argue that we should leave it out of the Tobit model. However, since no distinction is made there between risk engagement and the amount of risk taken once engaging in risky behavior in the Tobit model, it can be a significant variable.
14
are IQ, gender, age, years of education, regular income, financial literacy, a long term and short term risk engaging proxy, physical health and mental health. This leads to equation 3.
The second regression is conditional on participants playing the risk game. Therefore the sample size is smaller than in the first regression. The dependent variable here is the amount of money invested in the risk game. The independent variables are identical to the first regression except the risk engaging proxies, being IQ, gender, age, years of education, income, financial literacy, physical health and mental health. Equation 4 represents the above.
The Heckman model automatically regresses the two steps, first starting with the full sample and regressing the second step with the participants who opt to play. It corrects for the binary nature of the dependent variable in the first equation and incorporates the selection bias as an omitted variable bias, thereby adjusting for it. Additionally an inverse mills ratio is calculated, which if significant suggests that there is correlation between the residuals of the two regressions.
𝑅𝑖𝑠𝑘 𝑔𝑎𝑚𝑒 = 𝛼 + 𝛽1𝐼𝑄 + 𝛽2𝐴𝑔𝑒 + 𝛽3𝑌𝑒𝑎𝑟𝑠 𝑜𝑓 𝑒𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝐺𝑒𝑛𝑑𝑒𝑟 + 𝛽5𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽6𝑀𝑜𝑛𝑡ℎ𝑙𝑦 𝑖𝑛𝑐𝑜𝑚𝑒 + 𝛽7𝑃ℎ𝑦𝑠𝑖𝑐𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ + 𝛽8𝑀𝑒𝑛𝑡𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ + 𝛽9𝑆ℎ𝑜𝑟𝑡 𝑡𝑒𝑟𝑚 𝑟𝑖𝑠𝑘 𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡 + 𝛽10𝐿𝑜𝑛𝑔 𝑡𝑒𝑟𝑚 𝑟𝑖𝑠𝑘 𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡 (3) 𝐴𝑚𝑜𝑢𝑛𝑡 𝐼𝑛𝑣𝑒𝑠𝑡𝑒𝑑 = 𝛼 + 𝛽1𝐼𝑄 + 𝛽2𝐴𝑔𝑒 + 𝛽3𝑌𝑒𝑎𝑟𝑠 𝑜𝑓 𝑒𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛 + 𝛽4𝐺𝑒𝑛𝑑𝑒𝑟 + 𝛽5𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽6𝑀𝑜𝑛𝑡ℎ𝑙𝑦 𝑖𝑛𝑐𝑜𝑚𝑒 + 𝛽7𝑃ℎ𝑦𝑠𝑖𝑐𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ + 𝛽8𝑀𝑒𝑛𝑡𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ (4) Community bias
15 Hypothesis
The hypothesis follow the accepted theories in the economic and psychological literature as described in the literature section. For our main variable of interest IQ we state the hypothesis below.
Hypothesis 1: We expect that IQ positively and significantly influences both the engagement in risky behavior and the chosen risk level.
The hypothesis with regard to the control variables age, gender, income and mental and physical health are stated below.
Hypothesis 2: We expect that men engage significantly more in risk taking and exhibit significantly higher risk preferences.
Hypothesis 3: We expect that age does not significantly influence the engagement in risky behavior and the level of risk taken.
Hypothesis 4: We expect that income has significantly positive influences on both the engagement in risky behavior and the level of risk exhibited.
Hypothesis 5: We expect that the influence of mental health and physical health on engagement in risky behavior and the level of risk taken is insignificant.
Data
In the following section we explain the data used by showing summary statistics and explaining the context and meaning of several data sections.
All of the data used in this research stems from information that is acquired by a single large survey
that can be received if requested. The research is a collaboration between the University of
16
Table 1. Distribution of participants position within the household
Respondent Freq. Percent Cum.
Head 1,430 71.75 71.75 Wife / husband 460 23.08 94.83 Son / daughter 91 4.57 99.4 Son-in-law / daughter-in-law 2 0.1 99.5 Grandchild 2 0.1 99.6 Parent / Parent-in-law 4 0.2 99.8 Brother / sister / -in-law 1 0.05 99.85
Other 3 0.15 100
Total 1,993 100
Source: Stata
17
Table 2: Summary statistics for independent variables of the regression models.
Count Mean Variance Std. Dev. Min Max
IQ 1878 100.00 225.00 15.00 88 188 Gender 1878 0.40 0.24 0.49 0 1 Age 1876 47.18 198.38 14.08 17 95 Years of Education 1877 5.78 17.66 4.20 0 23 Monthly Income 1878 2161.21 9386608 3063 0 60000 Financial Literacy 1878 1.08 0.67 0.82 0 4 Mental Health 1878 8.35 4.44 2.11 2 12 Physical Health 1878 2.16 1.70 1.30 0 4
Short Term Risk Engaging Proxy 1878 0.43 0.25 0.50 0 1
Long Term Risk Engaging Proxy 1878 0.50 0.25 0.50 0 1
Source: Stata
IQ is one of the independent variables that we use consistently in our regressions and the core variable of interest in our research. As explained in the methodology section, we use 19 questions with Raven’s matrices to conduct the IQ test. In order to prevent frustration and the possibility of getting too many right answers by chance, the participant will not continue after three consecutive wrong answers. As a result, participants do only play until the last question if they do not make three consecutive mistakes before this question. We count the amount of questions answered correct instead of counting the amount of questions played before three consecutive mistakes are made in order to determine the IQ score. This represents the true skills of the participants , because it only counts the correctly answered questions, whereas only one in three questions need to be correct to be able to continue playing. If a participant manages to play all 19 questions and answers 12 questions correct, the IQ score will be 12 instead of 19, representing the amount of questions answered correctly. Summary statistics can be found in appendix B. The average number of questions answered correctly is 2.08, with a standard deviation of 2.55. This clearly shows that the logic skills of most respondents are low. However, as the maximum of 17 correct answers proves, the test is manageable, and around 5% of the participants answer 7 or more questions correctly, with 1% managing to answer more than 10 questions correctly. However, the vast majority of 80% does not answer more than 3 questions correctly.
18
abilities, this is only true within the sample. Compared to the world population, these scores could be significantly lower, reflecting the weak cognitive abilities.
Table 2 highlights the average age of the respondents, which is 47 years, with the youngest respondent being 17 and the oldest respondent being 95. Graph 1 gives a more intuitive idea of the age distribution, with the majority of respondents being between 35 and 55 years old. The graph also shows similarities to a normal distribution, with the two sides of the graph being relatively similar but a slightly flatter peak than a normal distribution. Although the majority of the data lies between 35 and 55, ages ranging from 25 to 70 years all occur more than 50 times. This creates a broad bandwidth to use in research and ensures that enough data is available to draw robust conclusions.
Graph 1. Age distribution of survey participants
19
Table 2 also presents summary statistics on the years of education of participants. The mean is 5.78 years and the standard deviation is 4.20 years. The level of education often does not exceed primary schooling, which is low compared to western countries. The different background can partially explain these results. The participants are from rural areas, and the educational levels in rural areas are significantly lower than urban areas in Bolivia (Worldbank 2006).
Besides the general IQ levels we also want to test specifically for Financial literacy, which we proxy by summing the right answers of 4 financial literacy questions in the survey. The first question is whether participants received any formal financial literacy training, followed by 3 numerical questions testing whether participants understand finance in general. Each question that is answered correctly gets a score of 1, which yields a financial literacy score between 0 and 4. A similar pattern as with IQ and years of education is shown, with over 70% of the respondents only scoring 1 point out of 4. Only 4.5% manages to score more than 2 points. Furthermore Appendix C shows that none of the cognitive related variables, IQ, years of education and financial literacy is highly correlated with each other. This ensures that none of the variables is redundant.
Two of the independent variables are mental health and physical health. The idea behind these variables is that we test whether physical and mental health influences the decision making process, and specifically if there is a difference in effect on the decision making process between mental and physical health. We face a lack of time and an inconvenient setup for a full mental and physical test. However, we do have a module with 16 questions on physical and mental health that are being used as proxies for mental and physical health. We assign each question to either the mental or physical health proxy, to ensure that no question is regressed by two different variables. At one question this proves to be difficult. The question asks whether respondents had a lot of energy during the past four weeks. Energy is obtained by a good physique and an optimistic and healthy state of mind, but we decide to assign this question to mental health.
Questions that are not binary in their setup, but with answers ranging for example from 1 to 4, are restructured as binary questions. The reason for this transformation lies in the fact that we want the count the correctly answered questions with 1 point, and this only works if all the questions have an equal value. Furthermore the positive answer for specific questions correspond to a 0, which would then not be added to the score of mental and physical health. These answers are transformed so that a positive answer always gets a score of 1. This has the effect that for both mental and physical health a higher score corresponds to a higher level of health.
20
The short and long term risk engaging proxies have similar means and variances. The mean value is 0.43 for the short term proxy and 0.50 for the long term proxy. The variance is identical with a value of 0.25. This might give rise to the belief that the variables are highly correlated. However, appendix D shows that the correlation coefficient between the two proxies is 0.35. This is within the limits of what is deemed normal. Therefore we do not drop one of the two proxies.
The risk game is split up in two sub-sections. First of all the table 4 shows whether respondents decide to play the risk game, i.e. invest an amount larger than 0 in the risky bet, or decide to keep the 20 Bolivianos and not play the risk game. It shows that a majority of 59.34% of the respondents decide to keep the 20 Bolivianos and not participate in the risk game. This is a sign of relative strong risk-aversion since the expected payoff of investing in the risk game is higher than not investing. 40.66% of the respondents decide to play the risk game, implying that they invested at least 1 Boliviano in the risk game.
21
Table 3. Investment in the risk game conditional on playing the risk game
Amount
invested Freq. Percent Cum.
2 200 26.18 26.18 4 97 12.7 38.87 6 88 11.52 50.39 8 16 2.09 52.49 10 258 33.77 86.26 12 6 0.79 87.04 14 1 0.13 87.17 16 5 0.65 87.83 18 4 0.52 88.35 20 89 11.65 100 Total 764 100 Source: Stata
Table 4 shows there are more male than female respondents. Males account for 60% of the respondents and females for 40% of the respondents. Besides this, table 5 shows that males also participate more often in the risk game, with 43% playing the game, as opposed to about 37% of females playing the risk game. Both in absolute and relative terms men play the risk game more often than women. Ideally there would be an equal share of male and female respondents, ensuring that there would be no bias at all. However, with females still accounting for 40% and a total of 749 respondents we do not see any problems regarding a lack of female respondents.
Table 4. Respondents distribution by gender
Gender Freq. Percent Cum.
22
Table 5. Participation in the risk game by gender
Gender Risk game Male Female Total Don't play 640 474 1,114 Play 490 274 764 Total 1130 748 1,878 Source: Stata Results
Table 6. Tobit regression adjusted for clustering showing the effect of variables on the amount
invested in the risk game with N=1876 for a 95% confidence interval.
Amount invested Coef. Std. Err. t P>t Conf. Interval
IQ 0.078 0.019 4.09 0 0.041 0.12 Gender -1.206 0.703 -1.72 0.086 -2.59 0.17 Age 0.002 0.026 0.07 0.944 -0.05 0.05 Years of education 0.386 0.089 4.35 0 0.21 0.56 Financial literacy 0.732 0.417 1.76 0.079 -0.08 1.55 Monthly income 0.000 0.000 2.38 0.017 0.00 0.00
Short term risk
engagement 1.511 0.681 2.22 0.027 0.17 2.85
Long term risk
23
Table 6 shows the results of the Tobit regression with the amount invested in the risk game as the dependent variable, and IQ, gender, age, years of education, income, financial literacy, short and long term risk engaging proxies, mental and physical health and a constant as independent variables. For a 95% confidence interval IQ, years of education, financial literacy, income, both risk engaging proxies and the constant are significant. The other variables are insignificant at the 5% significance level. Although the minimum amount invested in the risk game is 0, by not playing the risk game, The constant has a large negative coefficient value of -9.59, with a standard deviation of 2.27. This is possible due to the nature of the Tobit regression, which takes into account that the limits of investing, being 0 and 20, gives rise to a possible sample bias. According to this regression, people show strong signs of risk-averseness, which shows through the negative constant, implying that people would invest less than 0 would it be possible.
IQ is both highly significant and positive. The coefficient is 0.078 and the standard deviation is 0.019. The coefficient is large enough to have economic meaning as well. The positive sign shows that a higher IQ leads to more risk taking. This positive relationship is in line with what we expect based on the papers and theories reviewed.
Years of education, another cognitive variable is highly significant as well, with a positive coefficient of 0.39 and a standard deviation of 0.09. This relatively large coefficient suggests that the variable is economically significant as well. The positive nature implies that more years of education lead to more risk taking. Appendix C shows that the correlation coefficients between IQ and years of education is 0.35. For IQ and Financial literacy the coefficient is 0.17. Years of education and Financial Literacy yields a correlation coefficient of 0.27. There is positive correlation between the variables, but the correlation coefficients do not suggest that one of the variables is redundant. The distinction is that IQ represents measured intelligence, while years of education represents the level of schooling of participants. An interesting result is that these two variables are both valuable. While it is accepted that IQ and level of education are related, a different factor might be at play here. Since many of the subjects grew up in rural and poor circumstances, there might not have been an opportunity to attend school due to a lack of money or working obligations. Therefore the actual intelligence may not be fully in line with the years of education a participant has.
24
A variable that is insignificant at a 5% level but significant at a 10% level is gender. With a p-value of 0.086 it clearly falls out of the 5% level but easily within the 10% boundaries. The strong negative coefficient of -1.20 does imply that we should not completely disregard the variable. It indicates that woman take less risk than men do, since men are coded as 0 and women as 1, a result that is often found in other academic studies. The standard deviation of 0.70 is large which makes the estimate not reliable enough to draw strong conclusions. However, this result does point in the direction of a men taking more risk than women. A similar explanation holds for the financial literacy variable. Not significant at a 5% level but significant at a 10% level it hints at a positive relation between financial literacy and risk taking. Although the coefficient is relatively large with a value of 0.73 , we refrain from putting a lot of value on this result due to the p-value of 0.079. It does hint at the fact that financially trained subjects engage in more risk taking.
The two health related control variables, mental health and physical health, are insignificant at both a 5% and 10% level, with a much higher p-value for mental health than physical health. Both variables do not appear to be of any influence on the level of risk participants engage in. Since there is relatively little known about this relationship we cannot place this result in context to other studies. The state of mind and body is known to have an effect on people , but based on these results, the two variables are of no use in determining risk profiles in a financial context.
Both proxies of engagement in risk are highly significant and strongly positive. Although this risk game clearly revolves around engaging in risky behavior in the short term, the coefficient of the long term proxy is larger than that of the short term proxy. We do not find a reasoning for this in the literature. The issue is also not related to correlation between the two proxies. The correlation between the short and long term risk engaging proxies is 0.35, which is within the normal range. Another remark is that the independent variable in this case is the amount invested in the risk game, which revolves around the amount of risk taken. The proxies determine whether participants want to engage in risky behavior. However, the proxies are still highly significant with respect to the amount of risk taken in this regression, which could point at a weakness in the setup of this regression. This weakness could be that we treat the risk game as a one step process. People immediately decide how much they want to invest in the risk game. However, we argue that treating the risk game as a two-step process where subjects first decide whether to engage in the game before deciding the amount to invest might reveal more realistic results.
25
26
Table 7. Heckman selection model showing the effect of selected variables on the engagement in
risky behavior and the amount of risk taking with N=1876
Coef. Std. Err. z P>z 95% Conf. Interval Amount invested IQ -0.005 0.015 -0.32 0.747 -0.04 0.02 Gender -0.971 0.451 -2.15 0.031 -1.85 -0.09 Age 0.038 0.018 2.18 0.029 0.00 0.07 Years of education 0.178 0.063 2.82 0.005 0.05 0.30 Financial literacy -0.113 0.274 -0.41 0.679 -0.65 0.42 Monthly income 0.000 0.000 1.31 0.191 0.00 0.00 Physical health 0.166 0.184 0.90 0.366 -0.19 0.53 Mental health -0.089 0.110 -0.80 0.422 -0.30 0.13 Constant 7.404 2.420 3.06 0.002 2.66 12.15 Risk game IQ 0.009 0.002 4.06 0 0.00 0.01 Gender -0.063 0.064 -0.98 0.327 -0.19 0.06 Age -0.002 0.003 -0.94 0.349 -0.01 0.00 Years of education 0.026 0.009 2.89 0.004 0.01 0.04 Financial literacy 0.074 0.039 1.90 0.057 0.00 0.15 Monthly income 0.000 0.000 1.23 0.219 0.00 0.00
Short term risk
engagement 0.167 0.064 2.61 0.009 0.04 0.29
Long term risk
27
Table 7 shows the results of the Heckman selection model. In the first equation with engagement in the risk game as the binary dependent variable IQ, years of education, the short and long term risk engagement proxies and the constant are significant at a 5% level. The other variables are insignificant at a 5% level, with only the financial literacy variable being close to the 5% level with a p-value of 0.057. No other variables are significant within the 10% level. Based on the significance level of the variables we see a picture that is relatively similar to the Tobit regression. Besides the income variable that is significant at a 5% level in the Tobit regression but insignificant in the first Heckman equation and the gender that is significant at a 10% level in the Tobit regression but insignificant in the first Heckman equation the other variables fall within the same confidence interval.
The constant has the largest coefficient, with a strong negative value of -0.75 and a standard deviation of 0.21 implying that on average the participants are risk averse with respect to the risk game. The Heckman regression takes the binary nature of the dependent variable into account and adjusts for these limits. The intelligence related variables again play a central role in the participation in the risk game. Both IQ and years of education are significant, with positive coefficients of 0.009 and 0.050 and standard deviations of 0.002 and 0.012. The sign of the coefficients is similar to the Tobit regression and in line with literature, implying that a higher level of intelligence relates to more engagement in risky behavior. It can also be plainly explained by rationale in this experiment. In pure economic terms it is beneficial to invest the full 20 bolivianos in order to reach the highest expected payoff. Playing the game always yields a higher expected payoff than not playing the game. Although the game is simple and easy to understand, the underlying economics and the actual implications of these economics might be harder to grasp. This economic rationale could be easier to grasp by higher educated people as suggested by the cognitive load theory, resulting in more engagement in the risk game. The financial literacy variable points in the same direction. Although it just falls out of the 5% level it does have a strong positive coefficient value of 0.074, which is larger than both IQ and years of education. It supports the theory that understanding finances and the mechanics at play behind this simple game increases the engagement in the game.
The income of subjects is not significant, which is not in line with theory. We do not have a clear answer for this result. However, a potential cause could be that although the reward is substantial in relative terms, it is not substantial enough to cause differentiating behavior between participants with different incomes. Both gender and age are insignificant, which is not in line with expectations for gender, where it is generally accepted that men take more risk than woman. However, it has not frequently been tested in a two-step setup like we use here. The two health related variables, mental and physical health, are both highly insignificant and therefore not of any value in determining whether or not participants engage in risky behavior.
28
behavior. The implication is that people who are more willing to accept a hypothetical loan are also more likely to engage in risky behavior in a real life situation.
The second equation in the Heckman model has the amount invested as the independent variable, similar to the Tobit model. However, as stated above, the sample on which the variables are regressed is completely different. This sample only contains participants who opt to play the risk game. The independent variable set is similar to the first equation except for the two risk engaging proxies. The results of the second equation paint a different picture than the first equation. At a 5% level, gender, age, years of education and the constant are significant. No other variables are significant at a 10% level. The constant has a large positive value, which suggests that once participants decide to play, they invest a significant part of the 20 bolivianos. As stated in the data section, a peak occurs at an investment of 10 Bolivianos. We find possible a possible explanation in the fact that to some households the reward was presented in two notes of 10 Bolivianos. Although this did not matter for the payoff structure or final reward, it could have triggered a mental response to invest 10 Bolivianos.
From the intelligence and knowledge related variables only the years of education is significant. IQ becomes highly insignificant, which represents a sharp deviation from the other regressions. The main takeaway is that IQ and to a lesser degree financial literacy influence the engagement in risk but are not influential with respect to the amount of risk taken. So intelligence is a determinant for the engagement in risk, but once this is done not for the amount of risk taken. Years of education on the other hand are significant in both steps of the process. Since actual intelligence is represented by the IQ, it appears that more intelligent people are more likely to engage in risk. This has the effect that the sample of people that play the risk game contains a large amount of intelligent people. Because a large part is relatively intelligent, this coefficient loses the differentiating value, therefore becoming insignificant. The same analogy holds for financial literacy. This leaves the question unanswered why years of education stay significant in the second step of the Heckman regression although it appears to measure a similar skill. Although we do not find an explanation in the literature, the difference possibly lies in an attitude towards risk that is developed during the educational years but is unrelated to the actual level of intelligence. It is possible that more years at school lead to a more lenient attitude towards risk levels.
Another important variable in economic terms is the gender variable. With a strong negative sign of -0.97 and a standard deviation of 0.45 it is evidence that men take more risk than women do. This is in no way a surprise when analyzing previous research as we show in the literature section. There is a general consensus that men in general take more risk than women. There is however no clear evidence in which specific situations and frameworks men take more risk. This also shows in the difference between the first and second equation in the Heckman model. Whereas the gender variable proves to be in no way significant to determine whether people engage in risk, it is very clarifying in dissecting different risk preferences once people engage in risky behavior. We conclude that men do not engage more in risky behavior than women, but do take significantly more risk once they engage in the risky behavior.
29
deviation of 0.018. This result does not correspond to many of the findings in financial literature, which suggests that risk taking does not significantly change once maturity has been reached. On the other hand, there is little research on differences in risk taking across ages once maturity has been reached. The positive effect implies that older subjects take more risk than younger ones. In line with gender, it suggests that age does not influence the decision to engage in risk, but does influence the amount of risk taken. We do note again that these results are specific for our sample. Income is not a significant factor and does not add any value to explaining what drives levels of risk taking. We expect this is again partly due to the limited amount of money that participants receive. The two health proxies are insignificant, a result similar to the other regressions. We conclude that both the mental and physical health are not contributing factors in either the engagement in risk taking and the level of risk taken. A note that we do make is that we created proxies for both mental and physical health, so it is an approximation. However, we experimented with different formats of the proxies and none of these formats proved to be significant. The result is neither surprising nor confirming. There is a little literature on the matter and the effect has not been tested in such a setting before.
The lambda is insignificant at both a 10% and a 5% level which is a further confirmation that the regression and selection of variables is executed in a correct way. There is no correlation between the residuals of the two regressions.
Conclusion
With the goal of exploring the relationship between IQ and risk taking we explore two different approaches to obtain results. Firstly we view the process as one step where subjects directly decide the amount of money invested in the risk game. We use a Tobit model which includes the full sample of 1876 household with the amount invested in the risk game as the dependent variable. The independent variables are IQ, gender, age, years of education, financial literacy, income, a short and a long term risk engaging proxy, and mental and physical health. In our second regression we approach the risk game as a process in which two steps are involved. In the first step participants decide whether or not they want to play the risk game, or engage in risky behavior. In the second step the participants playing the risk game select an amount of risk, reflected by the amount invested in the risk game. To accommodate this setup we use the Heckman selection model. The first equation with the risk game as the dependent variable we use the same set of independent variables as in the Tobit model. In the second equation we drop the two risk engaging proxies. These questions ask whether a participant would engage in risky behavior, but do not provide information on the amount of risk taken. This selection ensures that the Heckman model functions in a correct way.
30
level of education and intelligence result in higher levels of risk taking. This result is an addition to the relatively small body of work that finds a negative relation between risk aversion and intelligence. The negative coefficient for gender at a 10% level suggest that men are less risk averse than women, a commonly accepted finding in financial literature. The positive effect of income implies that a higher income results in a higher level of risk taking. This is again a finding that is in line with financial literature. Lastly the two proxies for risk engaging are significant in this setup where no distinction is made between risk engagement and the amount of risk taken. It shows that if this distinction is not made, it functions as a good proxy for the amount of risk taken. Age, mental and physical health are insignificant. These results are not surprising, since the age effect often stops once adulthood is reached in financial literature and mental and physical health are rather experimental variables, lacking a large body of literature.
The two equations of the Heckman model generate different results. In the first equation, with the risk game as the dependent variable IQ, years of education, and the two risk engaging proxies are significant at a 5% level. At a 10% level financial literacy is significant. The constant is significantly negative, implying a general risk averseness towards playing the risk game. The coefficient is positive for all significant independent variables. A higher level of intelligence and education corresponds to more engagement in risk taking. This is in line with the literature, although the relation is expressed as a decrease in risk-averseness. The literature does not explicitly refer to an increase in engagement in risky behavior. The strong positive coefficient of the risk engaging proxies suggests that these simple questions, namely asking whether respondents would be willing to get a loan of amount x that they have to pay back in time x, do provide a good approximation of how willing subjects are to engage in risk. Gender, age, income, mental and physical health are insignificant. The second equation reflects the amount of risk participants take once they decide to engage in risky behavior. The dependent variable is amount invested, and gender, age and years of education are significant at a 5% level. There is a significant positive constant of 7.41, implying that once participants decide to play they on average invest a substantial amount. Contrary to the first equation, only the years of education are significant, while IQ and financial literacy lost the explanatory power. Although all three variables are knowledge and intelligence based, we suggest that a difference is that in school a culture and attitude towards risk taking can be created that is not completely related to actual intelligence. Gender has a negative coefficient, implying that men take more risk than women, which is as stated before in line with the literature. A surprise given most findings in literature is the fact that age has a positive coefficient. So when participants decide to play, the amount of risk taken increases with a higher age. Income, mental and physical health are insignificant. Additionally, the mills ratio is insignificant, implying that there is no correlation between the residuals of both equations.
31
model. Here age, gender and years of education are significant at a 5% level. We conclude that the distinct differences between the two Heckman equations imply that there are different forces that influence risk engagement and the amount of risk taking, and that the two step process adds explanatory value. The Tobit model shows great similarities to the first Heckman equation, and is more representative as a risk engaging model than explaining what drives the amount of risk subjects take. We do not disregard the Tobit model completely, but the Heckman model shows that there is a two-step nature to the process of playing the risk game and it captures the different significant variables at play better. The overall conclusion across models is that IQ has a strong positive effect the engagement in risk but not on the amount of risk taken. For the control variables, we conclude the following: The overall effect of income with respect to this setup appears to be limited, with the variable only being significant in the Tobit model. Gender is significant with respect to the amount of risk taken, but does not provide a distinguishing factor between the sexes regarding risk engagement. Age is only influential once participants decide to play the risk game, and then risk taking then increases with age. Both mental and physical health are consistently insignificant. We conclude that within this framework and with our approach to approximating the health related variables, there is no additional explanatory value with respect to engagement in risky behavior and the amount of risk taking.
32 Reference List
Byrnes, J., Miller, D. & Schafer, W., 1999. Gender differences in risk taking. Psychological
Bulletin, 125(3), pp.367–383.
Camerer, C.F. & Hogarth, R.M., 1999. The Effects of Financial Incentives in Experiments: A Review and Capital-Labor-Production Framework. Journal of Risk and Uncertainty, 19(1-3), pp.7–42.
Cauffman, E. & Steinberg, L., 2000. (Im)maturity of judgment in adolescence: Why adolescents may be less culpable than adults. Behavioral Sciences and the Law, 18(6), pp.741–760. Charness, G., Gneezy, U. & Imas, A., 2013. Experimental methods: Eliciting risk preferences.
Journal of Economic Behavior and Organization, 87, pp.43–51. Available at:
http://dx.doi.org/10.1016/j.jebo.2012.12.023.
Deakin, J. et al., 2004. Risk taking during decision-making in normal volunteers changes with age. Journal of the International Neuropsychological Society : JINS, 10(4), pp.590–598. Available at: http://www.ncbi.nlm.nih.gov/pubmed/15327737.
Dohmen, T. et al., 2011. Individual risk attitudes: Measurement, determinants, and behavioral consequences. Journal of the European Economic Association, 9(3), pp.522–550.
Dohmen, T.J. et al., 2010. Are Risk Aversion and Impatience Related to Cognitive Ability ?, IZA
Discussion paper no. 2375.
Dwyer, P.D., Gilkeson, J.H. & List, J.A., 2002. Gender differences in revealed risk taking: Evidence from mutual fund investors. Economics Letters, 76(2), pp.151–158.
Eckel, C.C. & Grossman, P.J., 2002. Sex Differences and Statistical Sterotyping in Attitudes Towards Financial Risk. Evolution and Human Behavior, 23(4), pp.281–295.
Furby, L. & Beyth-Marom, R., 1992. Risk taking in adolescence: a decision making perspective.
Developmental Review, 12(1), pp.1–44. Available at:
http://linkinghub.elsevier.com/retrieve/pii/027322979290002J.
Gardner, M. & Steinberg, L., 2005. Peer influence on risk taking, risk preference, and risky decision making in adolescence and adulthood: An experimental study. Developmental
Psychology, 41(4), pp.625–635. Available at:
http://www.ncbi.nlm.nih.gov/pubmed/16060809.
Guiso, L. & Paiella, M.,2008. Risk aversion, wealth, and background risk. Journal of the
European Economic association,6(6), pp.1109-1150.
Gottfredson, Linda, S., 2009. "Chapter 1: Logical Fallacies Used to Dismiss the Evidence on Intelligence Testing" In Phelps, Richard F. Correcting Fallacies about Educational and Psychological Testing. American Psychological Association, Washington.
33
Heckman, J.J., 1976. The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models.
Annals of Economic and Social Measurement, 5(4), pp.475–492. Available at:
http://ideas.repec.org/h/nbr/nberch/10491.html.
Lejuez, C.W. et al., 2002. Evaluation of a behavioral measure of risk taking: the Balloon
Analogue Risk Task (BART). Journal of experimental psychology. Applied, 8(2), pp.75–84. Mata, R. & Josef, A.K., Samanez-Larkin, G.R. & Hertwig, R.,2011. Age differences in risky
choice: A meta-analysis. Annals of the New York Academy of Sciences, 1235(1), p.p. 18-29.
Paas, F., 1992. Training strategies for attaining transfer of problem-solving skill in statistics: A cognitive-load approach. Journal of Educational Psychology, 84(4), pp.429–434.
Powell, M. & Ansic, D., 1997. Gender Differences in Risk Behaviour in Financial Decision-Making: An Experimental Analysis. Journal of Economic Psychology, 18(November 1997), pp.605–628.
Raven, J.C., 1936. Mental tests used in genetic studies: The performances of related individuals in test mainly educationary and mainly reproductive. Master thesis, University of London. Sweller, J., 1988. Cognitive Load During Problem Solving: Effects on Learning - Sweller - 2010
- Cognitive Science - Wiley Online Library. Cognitive science. Available at: http://onlinelibrary.wiley.com/doi/10.1207/s15516709cog1202_4/abstract.
Zuckerman, M., 1994. Behavioral expressions and biosocial bases of sensation seeking.
Cambridge University Press, New York.
Zuckerman, M. et al., 1964. Development of a Sensation-Seeking Scale. Journal of consulting
34 Appendix
Appendix A: Expected payoff for the risk game for every possible amount invested.
Amount Invested Expected Payoff 0 20 1 20.25 2 20.5 3 20.75 4 21 5 21.25 6 21.5 7 21.75 8 22 9 22.25 10 22.5 11 22.75 12 23 13 23.25 14 23.5 15 23.75 16 24 17 24.25 18 24.5 19 24.75 20 25 Source: Stata
Appendix B: Summary statistics for IQ questions answered correct.
Variable Obs Mean
Std.
Dev. Min Max
IQ question answered correct 1878 2.08 2.55 0 17
Appendix C: Correlation coefficients between three variables related to cognitive abilities.
35
Financial Literacy 0.17 1 0.27
Years of Education 0.35 0.27 1
Source: Stata
Appendix D: Correlation coefficients between the short term and the long term risk engaging
