On the source of risk aversion in Indonesia using micro data 2007

Sanjaya, Muhammad Ryan Working Paper On the source of risk aversion in Indonesia using micro data 2007 Economics Discussion Papers, No. 2013-33 Provided in Cooperation with: Kiel Institute for the World Economy – Leibniz Center for Research on Global Economic Challenges Suggested Citation: Sanjaya, Muhammad Ryan (2013) : On the source of risk aversion in Indonesia using micro data 2007, Economics Discussion Papers, No. 2013-33, Kiel Institute for the World Economy (IfW), Kiel This Version is available at: https://hdl.handle.net/10419/76660 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/3.0/ Received May 14, 2013 Accepted as Economics Discussion Paper June 26, 2013 Published June 26, 2013 © Author(s) 2013. Licensed under the Creative Commons License - Attribution 3.0 Discussion Paper No. 2013-33 | June 25, 2013 | http://www.economics-ejournal.org/economics/discussionpapers/2013-33 On the Source of Risk Aversion in Indonesia Using Micro Data 2007 Muhammad Ryan Sanjaya Abstract Many conventional economic analyses assume that risk preference is taken as given and do not give much scrutiny on it. However, empirical studies show that risk preference is not random: shocks and predetermined characteristics can determine risk preference. This study tried to see if these potential determinants together affect risk aversion in Indonesia using 2007 micro data. The author found that there is limited evidence that shocks and predetermined characteristics can affect risk preference. There is a preliminary indication that risk preference was not only driven by the individual’s wealth and demographic factors (that can be easily controlled), but also by the individual’s time preference. JEL O12 D81 Keywords Risk aversion; preference; Indonesia; microeconometrics Authors Muhammad Ryan Sanjaya, Faculty of Business and Economics, Universitas Gadjah Mada, Indonesia, [email protected] The author completed this paper during his postgraduate study at the Australian National University in 2012. Citation Muhammad Ryan Sanjaya (2013). On the Source of Risk Aversion in Indonesia Using Micro Data 2007. Economics Discussion Papers, No 2013-33, Kiel Institute for the World Economy. http://www.economics-ejournal.org/ economics/discussionpapers/2013-33 1 1 Introduction Many conventional economic analyses assume that risk preference is taken as given and do not give much scrutiny on it. In microeconomic theory, for example, a utility maximiser individual is assumed to have a stable preference, either with regard to risk or non-risk preference. Otherwise, she will violate the axioms of consumer choice— especially the transitivity axiom—and analyses that are derived from this unstable preference will be inconsistent. In addition to that, risk preference is also thought to be one of the key ingredients in tastes formation, and tastes are mostly assumed as stable (Stigler and Becker, 1977). These arguments, however, does not imply that stable preference should hold overtime. It means that an individual’s inconsistent behaviour can be attributed to random preference rather than unstable preference. Nonetheless, some empirical studies suggest that risk preference is not random. For example, one of the most common assumptions when people are making decisions under uncertainty is that absolute risk aversion is decreasing with wealth, which implies that individuals are willing to pay less for insurance if their wealth increases (Pratt, 1964).1 This assumption is proven empirically in lab experiment and in household survey as well (Guiso and Paiella, 2008, Holt and Laury, 2002). In addition to the role of wealth in determining risk aversion, several studies have found that shocks such as natural hazards make people less willing to take risk in disaster prone countries such as Peru, Nicaragua, and Indonesia (Cameron and Shah, 2011, Dang, 2012, van den Berg et al., 2009). Other than natural hazards, economic shocks can also have a positive relationship with risk aversion as observed from the effect of the 1930’s Great Depression on individual’s unwillingness to take financial risk (Malmendier and Nagel, 2011). These findings are psychologically intuitive: individuals update their information when there is an abrupt change (shocks) in their environment, and this new information changes their risk behaviour. The question is, of course, if this relationship between shocks experienced and risk-taking attitude is consistent and perpetual. Besides these shocks or temporary events, several studies argue that some predetermined characteristics such as genetic heritability can explain risk preference. Rubin and Paul (1979), for example, developed an evolutionary economics theory that links economic goods and “inclusive fitness”, a biological utility function that is maximised by the individual as a result of natural selection. This biological utility function “punishes” individuals who are not willing to take risk in the form of having no offspring (genetically). Hence, this theory predicts that only those who are willing to take risk that will survive. This theoretical prediction is then developed by Ball et al. (2010) by arguing that the taste for risk should co-evolve with superior physical prowess (and indeed they found that a physically stronger individual tend to be more risk loving). This argument is 1 Not only decreasing with wealth, but the shape of the curve is also important. See Figure A1 in the Appendix. 2 also supported by a finding in the US that shows that twins who are not genetically identical tend to have lesser similarity in risk preference than genetically identical twins (Cesarini et al., 2009). Psychology can also explain the role of physical attributes. For example, taller people tend to get positive reinforcement from their environment and this translates into greater engagement in leadership role that required willingness to make risky choice (Korniotis and Kumar, 2012). Using data from the US and Europe, they found that taller people with normal weight are having greater likelihood to engage in the financial market and take risky portfolios. Across the Atlantic, in Germany, two studies also show that height could explain some of the variations in risk preference (Dohmen et al., 2009, Hu!bler, 2012). Another possible determinant of risk preference is parental education, in which the more educated parents tend to have children who are less risk averse (Dohmen et al., 2009, Hu!bler, 2012, Hryshko et al., 2011). This is probably because the more educated parents are, on average, having better knowledge about risk, and this knowledge is passed on to their child. However, it should be acknowledged that there is a likelihood that there are unobserved traits of the parent—other than their education achievement—that can explain children’s attitude toward risk. This essay tried to answer the following question: do these potential determinants of risk preference significantly affect individual’s risk aversion in Indonesia? Indonesia—with more than 240 million people with wide array of diversity in its demographic, geographic and economic background—is an interesting subject for studying the determinant of risk preference. Cameron and Shah had done a similar study for Indonesia in 2011, but their contribution is limited to the impact of natural disaster on risk preference in rural area (especially East Java). My study took a broader look on any possible determinant of risk preference, which includes both the impact of shocks (such as natural disaster) and of individual’s predetermined characteristics (such as physical attributes and parental education), in both rural and urban area in Indonesia. This is my main contribution in this subject area. My second contribution is in giving more understanding on the exogeneity of risk preference. First, there are studies that tried to observe the impact of risk preference on individual behaviour (Cramer et al., 2002, Dow and Werlang, 1992, Gaduh, 2012, Guiso and Paiella, 2005) or earnings (Bonin et al., 2007, Le et al., 2011). Bonin et al., for example, found that people who are less willing to take risk tend to choose low-earning job. However, if an individual’s risk preference is endogenously determined by wealth or income—as had been found in the regression results in this essay—then the estimated coefficients will be invalid. If this is the case, these studies might, for example, overestimate the impact of someone’s risk preference on occupational choice if we exclude the fact that the person just recently experienced natural disaster. With regard to the policy implication, one of the results from Cameron and Shah (2011) study is that they suggest a policy that can increase the access for a natural disaster 3 related insurance. This follows from the finding that people who lived in villages that experienced disaster are more likely to engage in self-insurance. However, given the limited information outside East Java, this policy recommendation cannot be generalized for the whole Indonesia. Therefore my study adds to the debate on the importance of natural disaster insurance policy by taking a more general observation on Indonesia. I used data from the latest wave of the Indonesia Family Life Survey (IFLS4) that was surveyed in 2007. The preliminary result shows that, except for time preference and father’s education, only the usual demographic characteristics such as age, education, and sex that correlated with risk preference. Several subsample regressions resulted in the significance of height and disaster, but the pattern is scanty. There is also limited supporting evidence for disaster-related insurance promotion. The organisation of this study is as follow: Section 2 discussed data descriptions, variable constructions, and estimation methodology. Section 3 discussed estimation results, robustness checks, and a simple investigation on the policy implication. Finally, last section concludes. 2 Estimation design 2.1 Data I used data from the Indonesian Family Life Survey (IFLS) to construct a measure of risk aversion. The IFLS was conducted by RAND cooperated with local research institutions in Indonesia and available for free at the RAND website.2 While the respondents for the IFLS only come from 13 (out of 26) provinces in Indonesia but they represent around 83% of Indonesia due to the heavy population distribution in these selected provinces. The first wave of the IFLS was in 1993 and it has been repeated in 1997, 2000, and 2007. The IFLS consists of two blocks: household block and community block. The household block measures individual’s and household’s life such as consumptions, welfare, and health level, while the community block measures community/village life such as the availability of health facilities and school. Combined, there are 290 data files from these two blocks, each with specific information on the individual/household/community. While the IFLS is a panel dataset rich with information on households and individual’s behaviour, it is unfortunate that only in the latest available round (IFLS4) that it incorporates the questions on risk-taking behaviour. Nonetheless, I use information from IFLS2 (1997) and IFLS3 (2000) as well to construct several variables that I need in this essay. In addition to the IFLS, I also use poverty rate data in 1996 and 1999 at district level later on in the sensitivity regression.3 2 See http://www.rand.org/labor/FLS/IFLS.html 3 I would like to thank to Robert Sparrow for providing me with this data. 4 2.2 Variable construction Risk aversion In IFLS4 there are questions that can be used to measure risk aversion under the “Risk and Time Preference” section. There are two games in this section, Game 1 and Game 2, in which they differ only in the amount of hypothetical money involved.4 In this section, the respondent will be asked to choose between two gamble and if he/she chose the risky one then he/she will move to the next question (which gives different payoffs). In every question there is a “Don’t Know” option that can be used to rule out respondent who do not understand the question. Here’s an example of the gamble (see the Appendix for the full set of questions and description): In Option 2 you have an equal chance of receiving either Rp1.6 million per month or Rp400 thousand per month, depending on how lucky you are. [On the other hand,] Option 1 guarantees you an income of Rp800 thousand per month. Which option will you choose? There are several methods that have been applied to construct risk aversion from the IFLS dataset: 1) Ordering based on the riskiness of the choice (Cipollone, 2011, Gaduh, 2012). 2) Binary variable, which simplifies risk choice into either risk loving or risk averse (Cameron and Shah, 2011). 3) Estimates the Arrow-Pratt index of Absolute Risk Aversion (ARA) (Permani, 2011). By construction, Option 1) and 2) forced us to make two regressions based on Game 1 and Game 2. Option 1) is probably the second best option albeit difficulties in interpreting the coefficient if we use standard OLS to do the estimation. Option 2) is the simplest one in its construction, but it fits with Cameron and Shah experimental method since they do not use ordinal variable in the main part of their paper. By and large, Option 3) gives the best option due to the following reasons: first, ARA took information from both of Game 1 and Game 2. Second, this measure is also linked directly with the theoretical underpinning of risk aversion (Pratt, 1964). Third, as can be seen in equation (1) below, ARA is a nonlinear, continuous variable that gives more variation in risk aversion. Therefore, I used ARA in the main regression where a higher value indicates a more risk-averse behaviour. ARA is constructed based on the expected utility of an individual’s participation in the gamble (after considering his/her initial wealth endowment as well). Taking the second order Taylor expansion of the expected utility around the initial wealth endowment 4 This is probably the biggest drawback of using IFLS4 to construct risk aversion. With no stake involved, there is a chance that the respondent will choose randomly and even exaggerating their risk preference. However, IFLS is the most feasible dataset today in Indonesia that represents the largest population sample of Indonesia. 5 resulted in the following formula (where 𝑍! is the high payoff (Rp1.6 million in the example above) and 𝑍! is the low payoff (Rp400 thousand)): 𝐴𝑅𝐴 =!!!!! !! !!(!!!!!)!!!!(!!!!!) (1) From 10 questions on risk preference, I found eight possible payoff combinations of 𝑍! and 𝑍! that translated into eight values of ARA. The frequency distribution of ARA is skewed toward those who are very risk averse (ARA = 0.25): 11,641 out of 27,717 observations (42%) are very risk-averse (with mean value of 0.15 and standard deviation of 0.09). Figure 1: Absolute Risk Aversion frequency distribution In addition to this measure of risk aversion, I also used Cameron and Shah’s method in order to see how the regression result change if we use different methods to measure risk aversion (variable RL1 for Game 1 and RL2 for Game 2). RL1 and RL2 are binary variables that take the value of 1 if the respondent is risk loving. However, since this method forced us to make two regressions based on Game 1 and Game 2 then we cannot really make a fair comparison with the main regression (that use information from both games to make a single regression). In order to overcome this problem, I developed an alternative way of measuring risk aversion: variable RA that takes the value of 0 (very risk averse) to 4 (very risk loving). This was done by giving a point for every risky decision that a respondent took. See the Appendix for further details on how to calculate these various measure of risk aversion. Table 1: Cross-correlations of various measure of risk aversion ARA RA RL1 RL2 ARA 1.00 RA -0.76 1.00 RL1 -0.51 0.81 1.00 RL2 -0.39 0.63 0.35 1.00 6 Table 1 above shows that the cross-correlation between ARA, RA, RL1, and RL2 is quite strong (particularly between ARA and RA that has -0.76 correlation coefficient). With regard to alternative measures of risk aversion, the mean for RL1 is 0.16 (SD 0.36) and 0.05 (SD 0.22) for RL2, indicates that a great majority of the respondents are risk-averse. On the other hand, the average value of RA is 0.64 (SD 1.07) and 67% of total observation are very risk averse (RA=0). See Table 2 below. Table 2: Descriptive statistics Variable Observations Mean Std. Dev. Measures of risk aversion ARA 27717 0.15 0.09 RA 27717 0.64 1.07 RL1 27717 0.16 0.36 RL2 27717 0.05 0.22 Predetermined characteristics (PC) Height (cm) 27717 155 12 Weight (kg) 27717 54 11 Ideal (=1) 27717 0.62 0.49 Tall (=1) 27717 0.49 0.50 Father’s education 27717 0.75 0.96 Mother’s education 27717 0.53 0.79 Temporary events (TE) Disaster (number disaster experienced) 27717 0.15 1.70 Log of amount lost 27717 0.82 3.25 Log of assistance received 27717 0.57 2.71 Ecshock (=1 if in construction/financial sector in 1997) 8965 0.06 0.24 Change in poverty rate 27717 .58 .66 Ecshock × Change in poverty rate 8965 0.04 0.22 Other control variables (X) Log of assets 27717 17.18 1.84 Log of past assets 27717 16.12 2.48 Islam (=1) 27717 0.90 0.30 Javanese (=1) 27717 0.43 0.49 Rural (=1) 27717 0.48 0.50 Age (year) 27717 37 15 Male (=1) 27717 0.48 0.50 Married (=1) 27717 0.70 0.46 Dependency ratio (0-1, higher more independent) 27717 0.36 0.23 Time preference (1-5, higher more impatient) 27717 4.44 1.02 Education (0-4, higher more educated) 27717 2.00 1.15 Cognitive ability (0-1, higher smarter) 10642 0.74 0.24 Numerical ability (0-1, higher smarter) 10642 0.42 0.31 In this essay I categorise possible determinants of risk aversion into two main groups: individual predetermined characteristics and temporary shocks. Predetermined characteristics variables Variables in individual predetermined characteristics are height and parental education. I use height (in centimetres) as the main physical attributes variable and adding weight as a complement in the regression. The average height is 155cm (SD 12cm) while the average weight is 54kg (SD 11kg). Parent’s education is straightforward to observe and I made a categorical variable based on the highest (but not necessarily completed) educational level. Moreover, around half 7 of the parents were never been in school, which might be attributed to the fact that these uneducated parents were, on average, born around 1944 when Indonesia as a nation was not even born.5 Temporary events/shocks variables I simply included the number of natural disaster experienced by the household, which comprises more than just earthquake and flood as in Cameron and Shah’s paper.6 While there are data on the number of householder that was injured or killed because of the disaster but the variation is very small: more than 99% of the observation did not have their household member killed or injured due to the disaster. Including this in the regression will lead to large standard errors. IFLS also reports the amount of household’s belongings (business and non-business related belongings) that was lost due to the disaster. Many of the disaster victims also received financial assistance. I took the natural log of these and included as additional control variables. Other control variables The construction of other control variables such as wealth and education is standard and relatively straightforward. Nevertheless, there are several control variables worth discussed. First, it is possible that the observed risk loving behaviour is due to cohort’s impatience to get an immediate reward. Under the “Time Preference” section the respondents were asked to answer a series of questions regarding to hypothetical money won in a lottery. There are two games in this section that differs in the time when the respondent will get the money (in 1 year in Game 1 and in 5 year in Game 2). Then I constructed a categorical measure of time preference which values range from 1 (very patient) to 5 (very impatient). Here is an example (see the Appendix for the full set of questions and rules to generate this variable): You have won the lottery. You can choose between being paid: 1. Rp1 million today or 2. Rp2 million in 1 year. Which do you choose? Second, in addition to the wealth variable I also enter a lagged of wealth variable based on the information from IFLS3 (2000). This variable is included to take into account any possible correlation between past endowments on current risk behaviour. For example, if two people have the same level of wealth in 2007 but the first person had lost much of his wealth (while the second person not), then the first person might become more risk averse than the second person. 5 The average might be born before 1944 since the IFLS only asked about the age of the parent at the time of the survey was conducted or the age when they died. 6 Still, earthquake and flood contribute for about 87% of all disasters in Indonesia. U5206521 IDEC8011 14 Table 6: Subsample regressions by quintiles of assets and by education level (dependent variable: ARA) By quintile of assets By education level Bottom quintile Second quintile Third quintile Fourth quintile Fifth quintile Not/never school Basic education Higher education (1) (2) (3) (4) (5) (6) (9) (10) Height 0.0001 -0.0001 -0.0003** -0.0002 -0.0001 -0.0004* -0.0005*** -0.0003*** (0.0001) (0.0002) (0.0001) (0.0001) (0.0001) (0.0002) (0.0001) (0.0001) Weight 0.0001 0.0001 -0.0002 0.0000 -0.0000 0.0001 0.0000 -0.0003** (0.0001) (0.0002) (0.0001) (0.0001) (0.0001) (0.0002) (0.0001) (0.0001) Father’s education Elementary -0.0033 -0.0019 -0.0016 -0.0021 0.0026 -0.0235** -0.0024 0.0014 (0.0036) (0.0028) (0.0034) (0.0033) (0.0034) (0.0075) (0.0018) (0.0027) Junior high -0.0091 -0.0015 0.0057 0.0034 -0.0005 0.0086 -0.0052 0.0025 (0.0061) (0.0073) (0.0062) (0.0061) (0.0051) (0.0201) (0.0043) (0.0036) Senior high -0.0024 -0.0005 0.0033 -0.0024 0.0034 -0.0766*** -0.0066 0.0029 (0.0067) (0.0072) (0.0085) (0.0064) (0.0050) (0.0164) (0.0057) (0.0035) University -0.0070 0.0062 -0.0144 -0.0059 -0.0069 0.0014 -0.0070 (0.0134) (0.0140) (0.0165) (0.0105) (0.0071) (0.0183) (0.0047) Mother’s education Elementary 0.0014 -0.0034 -0.0022 0.0022 -0.0021 0.0110 -0.0015 0.0015 (0.0036) (0.0035) (0.0038) (0.0036) (0.0035) (0.0138) (0.0020) (0.0027) Junior high -0.0076 -0.0046 -0.0028 -0.0047 0.0021 0.0082 -0.0048 (0.0069) (0.0082) (0.0083) (0.0071) (0.0055) (0.0062) (0.0036) Senior high 0.0101 -0.0006 0.0025 -0.0047 0.0008 -0.0016 0.0081 -0.0027 (0.0088) (0.0108) (0.0108) (0.0089) (0.0060) (0.0203) (0.0093) (0.0042) University 0.0084 -0.0163 0.0123 0.0075 -0.0264* -0.0466 -0.0105 (0.0186) (0.0221) (0.0162) (0.0182) (0.0130) (0.0748) (0.0073) Disaster -0.0008 0.0011* 0.0001 0.0000 0.0020 -0.0068*** 0.0001 0.0003 (0.0006) (0.0006) (0.0005) (0.0001) (0.0043) (0.0010) (0.0003) (0.0003) Log lost 0.0004 0.0001 -0.0004 0.0004 0.0005 0.0028* -0.0002 0.0003 (0.0008) (0.0008) (0.0008) (0.0008) (0.0006) (0.0013) (0.0005) (0.0005) Log assistance -0.0003 -0.0024* 0.0010 -0.0007 -0.0001 0.0025 -0.0008 0.0003 (0.0011) (0.0012) (0.0011) (0.0014) (0.0010) (0.0014) (0.0006) (0.0007) F 10.63 10.88 10.38 8.46 18.59 . 13.95 17.01 R2 0.05 0.05 0.06 0.06 0.08 0.05 0.04 0.04 N 5550 5539 5556 5536 5536 1882 15101 10734 Notes: The regressions include all variables within PC, TE, and X except assets (column (1) to (5)) and education (column (6) to (8)). Variables in X are not displayed for reading convenience. Robust standard error is in parentheses. *** statistically significant at 1% level, ** at 5% level, * at 10% level. The estimations include subdistrict fixed effects and the standard errors are clustered at subdistrict level. U5206521 IDEC8011 15 The second part of Table 6 is for regression by education level. A person is categorised as having “Basic education” if that person is educated at elementary or junior high level as mandated by the Government Regulation 47/2008, and “Higher education” if educated at senior high school and above. I found many anomalies here especially with regard to those who never/not been in school, that might be attributed to the respondent’s lack of understanding about the questions on risk aversion. Interestingly, height is significantly correlated with being risk-loving in all specifications, but this result might be caused by the omission of education from the regressions. This means that there is a positive correlation between education level and height. It would be more interesting to see how the interaction between various levels of assets and education can have different impact on risk preference. One can logically infer that education and endowment level should move in the same direction and the findings in Table 6 should also hold. But when I made another four subsamples based on the combination of education (those educated at higher level) and assets level (those within the fifth quintile assets), there is still no significant impact of variables in PC and TE on ARA.8 One might suspect also that there is a reverse causality between ARA and time preference and married. There is another possibility as well that assets, lag of assets, rural, and impatience are influenced by the shock variables. I ran another regression that excludes those variables and found that while the estimated coefficients for height became significant, but the role of temporary events remains insignificant. Overall, the regressions in Table 4, 5, and 6 show the greater importance of demographic characteristics over predetermined characteristics or temporary events in explaining the variations in ARA. Still, there are limitations in these such as the sensitivity over different methods of measuring risk aversion, different ways to incorporate physical characteristics, possible impact of past economic shock, and the impact of abilities. Section 3.3 below will take a closer look over these potential problems. 3.3 Robustness check First, I checked for the sensitivity on the choice of the dependent variable by running full regressions as in equation (2), but using RA, RL1, and RL2 instead of ARA as the dependent variable. I also ran an ordered logit model as another specification since RA is ordinal.9 Recall that these alternative measures of risk aversion are in reverse direction of ARA, which means that higher value of RA, for example, is associated with being riskloving (rather than being risk-averse as in the case of ARA) The results are summarised in Table 7. 8 I do not display the tables of the subsample regressions that follow due to the large size of the table. The tables, however, are available upon request. 9 We cannot directly interpret the estimated coefficient, rather we can only comment on the sign of the coefficient and see if it fits our hypothesis. U5206521 IDEC8011 16 The OLS part of Table 7 shows that almost all predetermined characteristics and temporary events are not significant, supporting the results from the main regressions. Nonetheless, father’s education at the university and mother’s education at junior high school are significant in some of the regressions. Other variables such as age, age-square, higher degree education, and being very impatient remain significant and exhibiting the same direction as in the main regressions. In addition to that, except for being very impatient, other category of impatience loses its significance. Surprisingly, the constants seem to be not significant in all of these OLS specifications. The result from the ordered logit gives more interesting findings. First, the cut-off points (equivalent to the constant in the OLS) are significant. Second, I found that religiosity and ethnic background play a significant role in explaining risk aversion (RA). In particular, being a muslim and non-Javanese is correlated positively with being riskloving. I also redid subsample regressions based on assets and education and the results are fairly similar. While RL2 provides support for a positive relationship between height and risk loving behaviour for people on the third quintile, but in general the evidence that PC and TE can explain variations in risk aversion is limited. U5206521 IDEC8011 17 Table 7: Sensitivity in the dependent variable OLS Ordered logit Dependent variable RA RL1 RL2 RA (1) (2) (3) (4) Predetermined characteristics (PC) Height 0.0002 0.0000 -0.0001 0.0003 (0.0006) (0.0002) (0.0001) (0.0014) Weight 0.0011 0.0002 0.0003* 0.0007 (0.0007) (0.0002) (0.0001) (0.0016) Father’s education Elementary 0.0105 0.0052 -0.0041 0.0582 (0.0168) (0.0058) (0.0037) (0.0360) Junior high -0.0248 -0.0065 -0.0063 -0.0418 (0.0310) (0.0101) (0.0065) (0.0616) Senior high -0.0101 -0.0051 -0.0054 0.0078 (0.0343) (0.0116) (0.0079) (0.0663) University 0.1683* 0.0326 0.0299 0.3252** (0.0681) (0.0235) (0.0160) (0.1056) Mother’s education Elementary 0.0234 0.0099 0.0069 -0.0306 (0.0186) (0.0063) (0.0037) (0.0420) Junior high 0.0938* 0.0258* 0.0114 0.1392* (0.0366) (0.0118) (0.0083) (0.0699) Senior high 0.0515 0.0228 -0.0029 0.0695 (0.0492) (0.0164) (0.0111) (0.0877) University 0.1485 0.0180 0.0122 0.1126 (0.0999) (0.0314) (0.0256) (0.1577) Temporary events (TE) Disaster 0.0024 0.0016 0.0016 0.0153 (0.0086) (0.0028) (0.0024) (0.0181) Lost (ln) -0.0019 -0.0010 -0.0004 -0.0055 (0.0041) (0.0013) (0.0009) (0.0086) Assistance (ln) 0.0028 0.0004 0.0001 -0.0030 (0.0056) (0.0018) (0.0013) (0.0123) Other control variables (X) Assets (ln) 0.0155** 0.0033* 0.0022* 0.0126 (0.0047) (0.0016) (0.0010) (0.0095) Lagged assets (ln) 0.0027 0.0018 -0.0001 -0.0002 (0.0033) (0.0011) (0.0008) (0.0066) Islam 0.0046 0.0011 0.0034 0.2374*** (0.0403) (0.0125) (0.0079) (0.0703) Javanese -0.0104 -0.0028 -0.0045 -0.3299*** (0.0259) (0.0080) (0.0061) (0.0727) Rural 0.0095 0.0030 -0.0064 0.1276 (0.0370) (0.0125) (0.0076) (0.0730) Age 0.0095*** 0.0043*** 0.0010* 0.0145** (0.0023) (0.0008) (0.0005) (0.0049) Age^2 -0.0001*** -0.0001*** -0.0000* -0.0002*** (0.0000) (0.0000) (0.0000) (0.0001) Sex 0.2411*** 0.0669*** 0.0294*** 0.4560*** (0.0161) (0.0058) (0.0033) (0.0335) Married -0.0115 0.0003 0.0011 -0.0304 (0.0171) (0.0062) (0.0034) (0.0366) Dependency -0.0028 0.0073 -0.0107 -0.0385 (0.0348) (0.0121) (0.0069) (0.0748) Time preference Patient 0.0681 0.0031 0.0156 0.0931 (0.0533) (0.0198) (0.0127) (0.0914) Somewhat impatient -0.0369 -0.0574** -0.0078 0.0092 (0.0557) (0.0206) (0.0123) (0.0960) Impatient -0.0149 -0.0367 -0.0157 0.0820 (0.0548) (0.0197) (0.0121) (0.1005) Very impatient -0.2283*** -0.0586** -0.0163 -0.5248*** (0.0501) (0.0181) (0.0114) (0.0905) Education Elementary -0.0377 0.0003 -0.0119* -0.0151 (0.0273) (0.0104) (0.0056) (0.1016) Junior high -0.0289 -0.0084 -0.0000 0.0545 (0.0311) (0.0122) (0.0066) (0.1100) Senior high 0.0040 0.0003 0.0028 0.1212 (0.0333) (0.0131) (0.0072) (0.1151) University 0.1680*** 0.0396* 0.0273** 0.4509*** (0.0423) (0.0154) (0.0096) (0.1229) Constant 0.1399 -0.0151 -0.0050 (0.1445) (0.0510) (0.0293) F 21.852 10.137 7.668 χ2 705.455 R2 0.04 0.02 0.01 N 27717 27717 27717 27717 Notes: robust standard error is in parentheses. *** statistically significant at 1% level, ** at 5% level, * at 10% level. OLS estimations include subdistrict fixed effects and the standard errors are clustered at subdistrict level. Cut off points for ordered logit are not shown. Stata does not have a built-in command to accommodate subdistrict fixed effects for ordered logit estimation, which is unfortunate because we may suspect a time-invariant omitted variable bias. U5206521 IDEC8011 18 Another robustness check is by using a dummy variable Ideal as a proxy for physical prowess that is derived from the body mass index (BMI). BMI is simply the ratio between the weight (kg) and the square of height (meter). The variable Ideal equals to 1 if the BMI is at normal range (between 18.5 to 25 as defined by the WHO).10 Another alternative measure is relative height, which is a dummy variable Tall, which equals to 1 if the person is taller than the median of other respondents of the same sex living in the same district.11 As can be seen in column (1) and (2) of Table 8, the use of either Ideal or Tall as an alternative measure of physical attribute cannot help explaining variations in ARA. While economic shock is relevant for Indonesia (the country experienced the 1997/1998 Asian economic crisis) and there are studies that shows the impact of the crisis on different households or economic sectors (Fallon and Lucas, 2002, Waters et al., 2003, Wie, 2000), but the information on individual risk preference is only available in 2007. There are also various factors affecting the individual within that 10-year gap that might not be observed. It is also difficult to identify the impact of the crisis for different individuals or to know if an individual’s observed behaviour is due to the crisis. Nonetheless, I tried to control for the crisis by adding three variables: Ecshock, change in the poverty rate, and the interaction between these two. Ecshock is a dummy variable that equals to 1 if the respondent worked in the construction and financial sector in 1997 by utilising data from IFLS2. These two economic sectors got the hardest hit (based on the drop in real GDP growth) during the crisis (Wie, 2000). In Table 8 column (3) we can see that there is no observed impact of past crisis on current risk preference. It should be noted that since the number of respondent increased between IFLS2 and IFLS4 and not all respondent worked during the IFLS2 survey, the final number of observation is severely limited. 10 See http://apps.who.int/bmi/index.jsp?introPage=intro_3.html 11 I use median rather than mean to avoid measurement error due to the outliers. U5206521 IDEC8011 19 Table 8: Ideal posture, economic crisis, and abilities Dependent variable: ARA (1) (2) (3) (4) Predetermined characteristics (PC) Ideal -0.0015 (0.0011) Tall -0.0003 (0.0011) Height -0.0002* -0.0001 (0.0001) (0.0001) Weight 0.0000 0.0000 (0.0001) (0.0001) Temporary events (TE) Ecshock 0.0086 (0.0055) Change in poverty rate -0.0029 (0.0044) Shock -0.0053 (0.0069) Other control variables (X) Education Elementary 0.0061** 0.0060** 0.0040 0.0080 (0.0023) (0.0023) (0.0032) (0.0122) Junior high 0.0035 0.0035 -0.0023 0.0058 (0.0025) (0.0025) (0.0042) (0.0123) Senior high -0.0028 -0.0027 -0.0041 0.0022 (0.0027) (0.0027) (0.0045) (0.0125) University -0.0145*** -0.0145*** -0.0112* -0.0074 (0.0033) (0.0033) (0.0052) (0.0128) Cognitive ability 0.0014 (0.0049) Numerical ability -0.0161*** (0.0035) Constant 0.1887*** 0.1875*** 0.2381*** 0.1936*** (0.0092) (0.0091) (0.0234) (0.0296) F 44.82 44.79 14.46 15.78 R2 0.06 0.06 0.06 0.05 N 27717 27717 8965 10642 Notes: The regressions also include all variables within PC, TE, and X. Variables in X are not displayed for reading convenience. Robust standard error is in parentheses. *** statistically significant at 1% level, ** at 5% level, * at 10% level. The estimations include subdistrict fixed effects and the standard errors are clustered at subdistrict level. Again, subsample regressions cannot explain variations in ARA when I varied the measure for physical attributes (Tall and Ideal) or when I control for the impact of past shock. Finally, I controlled for cognitive ability and numerical ability in Table 8 column (4) because I also used education as one of the explanatory variables in X. Excluding ability will bias the estimated coefficient of education. However, question on ability is limited only to respondent age 15-24, which reduces the number of observation. The estimation shows that education variable became insignificant and numerical ability is strongly and negatively correlated with ARA, indicating that people with high mathematical ability tend to be more risk loving. This result is confirmed when I used subsample regressions where the numerical ability is significant and negatively associated with risk averseness for people in the third and fifth endowment quintiles. This is somewhat an important U5206521 IDEC8011 20 result because we observe that the coefficients for elementary and higher degree education are statistically significant throughout all specification in the main regression (Table 4). 3.4 Insurance policy Cameron and Shah (2011) observed that people who lived in disaster-prone area in East Java tend to self-insure through a rotating saving mechanism (Arisan) and they also found that receiving remittance offset some of the impact of natural disaster on risk aversion. In order to test this I included a dummy for the participation in Arisan and the amount of transfer received from outside the household (Transfer, in ln). Table 9 shows that people who experience disaster are, on average, have higher transfer and involve more in Arisan. Table 9: Self-insurance and natural disaster Disaster No disaster Difference Arisan 0.3865 (0.0113) 0.2230 (0.0026) 0.1635*** Transfer (ln) 8.6102 (0.2116) 7.7545 (0.0566) 0.8557*** N 1868 25849 Note: *** significant at 1% level I then interacted these variables with how often the individual experienced disaster (Arisan × Disaster and Transfer × Disaster) and included these in the full regression (equation (2)). If the estimated coefficient for Transfer × Disaster is negative and significant, it means that the larger the transfer, the less risk averse the individual when there is a shock (disaster). Hence, these additional variables can be seen as an informal proxy for the demand for a disaster-related insurance. Table 10: Self-insurance (dependent variable: ARA) Full sample Subsample Not Arisan Arisan (1) (2) (3) Arisan -0.0030* (0.0014) Arisan × disaster 0.0008* (0.0003) Transfer (ln) -0.0002** -0.0002* -0.0002 (0.0001) (0.0001) (0.0001) Transfer × disaster -0.0001 -0.0001 -0.0002 (0.0001) (0.0001) (0.0001) Constant 0.2046*** 0.2074*** 0.1825*** (0.0117) (0.0117) (0.0117) F-test 38.78 32.53 12.22 R2 0.06 0.06 0.07 N 27707 21220 6487 Notes: The regressions also include all variables within PC, TE, and X. Variables in PC, TE, and X are not displayed for reading convenience. Robust standard error is in parentheses. *** statistically significant at 1% level, ** at 5% level, * at 10% level. The estimations include subdistrict fixed effect and the standard error is clustered at subdistrict level. U5206521 IDEC8011 21 In Table 10 column (1), I found that while Arisan is negatively correlated with ARA but the coefficient for Arisan × Disaster is positive and significant. This means that after controlling for the direct impact of the Arisan, an individual tend to be more risk averse when he/she experienced (more) disaster. On the other hand, only the coefficient for Transfer is negative and significant, which suggests that only the direct effect of Transfer that drives risk aversion. Overall, these results give less support for a natural disasterrelated insurance policy. Nonetheless, we might suspect that Arisan has reverse causality with ARA: risk-averse individuals tend to involve more in such rotating saving mechanism to smooth their consumption. Therefore, I made subsample regressions by Arisan participation in column (2) and (3). The estimated coefficients do not differ much from those in column (1), thus support the previous claim that only Transfer that determines ARA. 4 Conclusion Several studies point out to the important role of temporary shocks and predetermined characteristics on determining an individual’s risk preference. My observation using IFLS4 data for Indonesia shows that this is not necessarily the case: only father’s education at higher level that exhibits the expected sign and significance. The impact of natural disaster as found in Cameron and Shah (2011) diminished when I use full sample of both the rural and urban area. Physical attributes were showing significance and correlates negatively with ARA in regressions that contain predetermined characteristics and shock variables, but then fell down when I control for demographic variations and other variables. Nonetheless, there is a strong correlation as well between being impatient with low degree of ARA (risk-loving). These give preliminary indication that variations in risk preference are indeed random. From the policy perspective, a simple proxy for the demand of a disaster-related insurance shows that only the direct effect of the transfer that drives risk aversion, which means larger transfer for people who experience disaster does not reduce the risk averseness of the individual. In other words, there is no observed demand for natural disaster-related insurance. Nonetheless, the absence of evidence is not necessarily an evidence of absence. There has been a great concern on the use of utility function to reveal risk preference and on how the framing of the question, information processing, and reference point can affect risk preference (Schoemaker, 1993). The survey design itself does not elicit any use of real monetary payoff to the respondent, which might underestimate the observed degree of risk aversion of the respondent. It is possible also that the individual gives nonlinear probability on gain and loss, which explains why many people are risk-averse. Finally, this study is just a brief introduction to studies on risk preference in Indonesia. A way forward is to take a closer look on how sensitive the result is if we observe that people see gain and loss differently. The construction of ARA rests on the expected utility U5206521 IDEC8011 22 theory that assign linear probabilities on gain and loss, but prospect theory—one of the cornerstone in behavioural economics—suggests that people give nonlinear weighting in the probability of gain and loss (Kahneman and Tversky, 1979), in which people tend to value loss more than they value gain. An excellent applied research in this topic is by Tanaka et al. (2010) where they found that poor villagers in Vietnam are not always fear of uncertainty in income variation, but they also fear of loss. This will be the future direction of this study. U5206521 IDEC8011 23 Appendix Risk-averse individual Consider an individual that has a von Neumann-Morgenstern (VNM) utility function over wealth 𝑢𝑤. Consider also that there is a simple gamble g that has an expected value of 𝐸𝑔=𝑝!𝑤!, where 𝑝! is the probability of winning wealth 𝑤!. Suppose that the person is asked to choose to either: (1) engaged in a gamble g, or (2) getting an amount 𝐸𝑔 with certainty. A risk-neutral individual will have a linear utility function and sees these two options indifferently because the expected value from engaging in the gamble is simply equal to 𝐸𝑔. However, for a person who is not risk-neutral, he/she should consider the utility for each possible wealth resulted from the gamble. Therefore, he/she compared 𝑢𝑔=𝑝!𝑢𝑤! of Option (1) and 𝑢𝐸𝑔=𝑢𝑝!𝑤! of Option (2). Figure A1: A risk averse utility function A risk-averse individual is someone who choose (2) over (1), that is if 𝑢𝐸𝑔>𝑢𝑔, as shown in Figure A1 above. This is because a risk-averse individual will choose a certain amount of wealth 𝐶𝐸 that generates the same level of utility as 𝑢𝑔, even though the gamble’s expected value 𝐸𝑔>𝐶𝐸. E(g)!CE!w1!w2! u(E(g))! u(g)! u(w)! w! u! U5206521 IDEC8011 30 KAHNEMAN, D. & TVERSKY, A. 1979. Prospect Theory: An Analysis of Decision under Risk. Econometrica, 47, 263-292, http://www.jstor.org/stable/1914185?origin=JSTOR-pdf Accessed on 5 February 2011. KORNIOTIS, G. & KUMAR, A. 2012. Stature, Obesity, and Portfolio Choice. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1509173 Accessed on 30 August 2012. LE, A. T., et al. 2011. Attitudes towards economic risk and the gender pay gap. Labour Economics, 18, 555-561, http://genepi.qimr.edu.au/contents/p/staff/Le_LabEconomics_555-561_2011.pdf Accessed on 9 August 2012. MALMENDIER, U. & NAGEL, S. 2011. Depression Babies: Do Macroeconomic Experiences Affect Risk Taking. The Quarterly Journal of Economics, 126, 373–416, http://qje.oxfordjournals.org/ Accessed on 6 August 2012. PERMANI, R. 2011. Revisiting the Link between Maternal Employment and School-Aged Children Health Status in Developing Countries: An Instrumental Variable Approach. The University of Adelaide School of Economics, Research Paper No. 2011-21 http://www.economics.adelaide.edu.au/research/papers/doc/wp2011-21.pdf Accessed on 29 August 2012. PRATT, J. W. 1964. Risk Aversion in the Small and in the Large. Econometrica, 32, 122-136, http://www.jstor.org/stable/1913738 Accessed on 30 October 2012. RUBIN, P. H. & PAUL, C. W. 1979. An Evolutionary Model of Taste for Risk. Economic Inquiry, 17, 585-596, http://onlinelibrary.wiley.com/doi/10.1111/j.1465- 7295.1979.tb00549.x/abstract Accessed on 17 September 2012. SCHOEMAKER, P. J. H. 1993. Determinants of Risk-Taking: Behavioral and Economic Views. Journal of Risk and Uncertainty, 6, 49-73, http://imap.whartonmackcenter.com/files/0/4/443/Schoemaker, Paul J.H. Determinants of Risk-Taking-Behavioral and Economic Views.pdf Accessed on 30 October 2012. STIGLER, G. J. & BECKER, G. S. 1977. De Gustibus Non Est Disputandum. American Economic Review, 67, 76-90, http://www.jstor.org/stable/1807222 Accessed on 17 September 2012. TANAKA, T., et al. 2010. Risk and Time Preferences: Linking Experimental and Household Survey Data from Vietnam. American Economic Review, 100, 557-571, http://www.aeaweb.org/articles.php?doi=10.1257/aer.100.1.557 Accessed on 23 September 2012. VAN DEN BERG, M., et al. 2009. Natural Hazards And Risk Aversion: Experimental Evidence From Latin America. Paper prepared for the International Association of Agricultural Economists Conference, Beijing, China, 2009, http://ideas.repec.org/p/ags/iaae09/51394.html Accessed on 7 August 2012. WATERS, H., et al. 2003. The impact of the 1997–98 East Asian economic crisis on health and health care in Indonesia. Health Policy and Planning, 18, 178-181, http://heapol.oxfordjournals.org/content/18/2/172.short Accessed on 20 September 2012. WIE, T. K. 2000. The Impact of the Economic Crisis on Indonesia's Manufacturing Sector. The Developing Economies, 420-53, http://onlinelibrary.wiley.com/doi/10.1111/j.1746- 1049.2000.tb00886.x/abstract Accessed on 17 September 2012. Please note: You are most sincerely encouraged to participate in the open assessment of this discussion paper. You can do so by either recommending the paper or by posting your comments. Please go to: http://www.economics-ejournal.org/economics/discussionpapers/2013-33 The Editor © Author(s) 2013. Licensed under the Creative Commons Attribution 3.0.