Do Risk Preferences Change? Evidence from the Great East Japan Earthquake Chie Hanaoka

Hitoshi Shigeokay

Yasutora Watanabez

May, 2017

Abstract We investigate whether individuals’ risk preferences change after experiencing a natural disaster, speci…cally, the 2011 Great East Japan Earthquake. Exploiting the panels of nationally representative surveys on risk preferences, we …nd that men who experienced greater intensity of the Earthquake became more risk tolerant a year after the Earthquake. Interestingly, the e¤ects on men’s risk preferences are persistent even …ve years after the Earthquake at almost the same magnitude as those shortly after the Earthquake. Furthermore, these men gamble more, which is consistent with the direction of changes in risk preferences. We …nd no such pattern for women. Key words: Risk preference, Persistence, Gender di¤ erence, Great East Japan Earthquake, Risk-taking behavior, Panel data JEL codes: D81, Q54, C23, J16

Faculty of Economics, Kyoto Sangyo University, Kamigamo-Motoyama, Kita-ku, Kyoto, JAPAN. Email: [email protected] y Department of Economics, Simon Fraser University, 8888 University Drive, Burnaby, BC, V5A1S6, CANADA and NBER. Email: [email protected] z Department of Economics, HKUST Business School, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, HONG KONG. Email: [email protected]

1

1

Introduction

Risk preferences are fundamental determinants of individual decision making regarding economic behavior, such as saving, investment, consumption, technology adoption, and migration. Standard economic models assume that individual risk preferences are stable across time (Stigler and Becker, 1977).1 More recently, a growing number of studies have suggested that individuals’risk preferences and thus, risk-taking behavior, can be altered by various negative shocks such as …nancial crises, trauma from con‡ict or violence, and natural disasters (e.g., earthquakes, hurricanes, ‡oods, and tsunamis). However, there is little consensus about the direction in which such negative shocks a¤ect risk preferences. Past studies …nd con‡icting patterns even within the same domain. For example, regarding the impact of natural disasters, some …nd increased risk aversion (van den Berg et al., 2009; Cassar et al., 2011; Reynaud and Aubert, 2014; Cameron and Shah, 2015), while others …nd decreased risk aversion (Eckel et al., 2009; Bchir and Willinger, 2013; Willinger et al., 2013; Page et al., 2014).2 Furthermore, even if such e¤ects exist, little is known about whether they are transitory or persistent.3 In this study, we investigate whether individuals’ risk preferences change after they have experienced a natural disaster, speci…cally, the 2011 Great East Japan Earthquake (hereinafter, the Earthquake), which is by far the largest earthquake in the country’s history since the modern measurement of earthquakes began. In particular, we test whether individuals who live in locations that endured a higher intensity of the Earthquake become either risk averse or risk tolerant and examine whether the e¤ects, if any, are transitory or permanent. Our data is the panels of nationally representative surveys which follow risk preferences of the same individuals before and one and …ve years after the Earthquake.4 In fact, Chuang and 1

“One does not argue over tastes for the same reason that one does not argue over the Rocky Mountains — both are there, will be there next year, too, and are the same to all men” Stigler and Becker (1977). 2 See Haushofer and Fehr (2014) and Chuang and Schechter (2015) for an excellent review. Similarly, studies on the impact of con‡icts or violence also …nd con‡icting results: while some show increased risk aversion (Moya, 2012; Kim and Lee, 2014; Callen et al., 2014; Jakiela and Ozier, 2015; Brown et al., 2017), others …nd decreased risk aversion (Voors et al., 2012). By contrast, studies on the impact of …nancial crisis or hardship are relatively more consistent than those on non-…nancial shocks: Cohn et al. (2015) document that negative …nancial shocks make …nancial professionals more risk averse in laboratory experiments and Necker and Ziegelmeyer (2016) indicate that the Great Recession decreased risk tolerance. Imas (2016) suggests that these inconsistent …ndings can be explained by individuals’di¤erential responses to realized versus paper losses. 3 For example, Malmendier and Nagel (2011) show that early life …nancial experiences are associated with more conservative investment behavior in later life, whereas recent life events have the greatest impact on willingness to take risks. Eckel et al. (2009) show that hurricane Katrina evacuees become risk loving shortly after they are evacuated, but such e¤ects diminish in a year. Gallagher (2014) also shows that ‡ood insurance take-up spikes the year after a ‡ood but then steadily declines to the baseline. 4 Notable exceptions that use panel data include Brunnermeier and Nagel (2008), Sahm (2012), and Guiso et al. (2013), who examine shocks in …nancial domain. Our study di¤ers from theirs as our shocks fall within the non-…nancial domain (i.e., more physical). In addition, our data are nationally representative survey, and we can

2

Schechter (2015), the most recent and comprehensive survey on this topic, point out that data on preferences is usually only available after the event and not before. This is especially the case of natural disasters because researcher cannot anticipate where the natural disasters will occur. Panel structure of our data enables us to examine the potential bias of the cross-sectional studies due to selective attrition or migration, that is, more risk averse (or risk tolerant) individuals are more likely to be excluded from the sample or migrate after the negative shock. How the Earthquake, or negative shocks in general, potentially alters individuals’ risk preferences is not well understood. The psychology literature emphasizes that a particular type of emotion triggered by the negative shock can a¤ect individuals’ risk preferences (e.g., Leith and Baumeister, 1996; Lerner and Keltner, 2001; Loewenstein et al., 2001). Furthermore, Fessler et al. (2004) show that the e¤ects of emotions on risk preferences depend on not only the type of emotion but also gender. They show that anger makes male subjects risk loving, while disgust makes female subjects risk averse. If the Earthquake triggered di¤erent types of emotions in men and women, the susceptibility of risk preferences and their direction may also di¤er by gender (Lerner et al., 2003). Our …ndings are summarized as follows. We …nd that men who are exposed to higher intensities of the Earthquake become more risk tolerant one year after the Earthquake. As for women, we occasionally …nd opposite patterns (i.e., they become more risk averse) in high intensity locations, although the results for women are not very robust. Our …ndings suggest that men’s risk preferences are more likely to change than those of women and men become less risk averse in response to a negative shock, despite their already low level of risk aversion. While previous studies show that men are less risk averse than women (level )5 , to the best of our knowledge, ours is the …rst study that documents gender di¤erences in the susceptibility of risk preferences and their direction to a negative shock (changes) in a non-laboratory setting. In terms of magnitude, as the intensity of the Earthquake increases by 2, which corresponds to a 10-fold increase of ground acceleration, the relative risk aversion for men decreases by 8.1 percent from the mean before the Earthquake. Put di¤erently, the size of the reduction is roughly half of the mean di¤erence in relative risk aversion between men and women before the Earthquake. Furthermore, we investigate whether the e¤ects on men’s risk preferences are merely transitory or persistent. We …nd that these e¤ects are persistent: men who are exposed to higher intensities of the Earthquake are still risk tolerant even …ve years after the Earthquake. Surprisingly, the magobserve both short- and long-term e¤ects. 5 The extant literature shows that men are less risk averse than women. See, for example, Eckel and Grossman (2008) and Croson and Gneezy (2009) for surveys of economics, and Byrnes et al. (1999) for a survey of psychology on gender di¤erences in preferences.

3

nitude of the e¤ects remains almost unchanged at the level observed shortly after the Earthquake. This result suggests that a large physical negative shock such as our case can fundamentally alter individual’s risk preferences and thus, might have a long-lasting impact on one’s decision making on economic behavior. Our results are robust to a number of potential concerns. First, we verify that selective attrition or migration does not occur in our setting: the pre-Earthquake risk preferences are not correlated with both migration and attrition, suggesting that the more risk averse or risk tolerant individuals are not more likely to be excluded from the sample or migrate. Then, we conduct various speci…cation checks. Our results are hardly a¤ected after controlling for changes in income, assets, property price, and home ownership. Moreover, our results are robust to di¤erent ways of constructing the intensity measure of the Earthquake and risk preference measures. Throughout our study, the intensity measure used is the seismic intensity of the Earthquake (Shindo in Japanese), which is a metric of an earthquake’s strength at a speci…c location. We do not use other metrics, such as the level of radiation and fatality rate, because both these measures are concentrated in a small number of municipalities and there is little variation among the municipalities covered in our survey. Moreover, we add radiation level and fatality rate as controls and con…rm that our results on risk preferences are robust to these controls. Finally, in addition to the change in risk preferences, we investigate whether risk-taking behavior is a¤ected by the Earthquake. We …nd corroborative evidence that men who live in the hardest-hit locations gamble more as the intensity of the Earthquake increases. While these results are consistent with the direction of changes in risk preferences, because risk-taking behavior is a¤ected by not only risk preferences but also many other factors (e.g., risk perception, time preferences, and peer e¤ects), we need to view the results on risk-taking behavior with caution. The remainder of this paper is organized as follows. Section 2 describes the data and Section 3 presents our identi…cation strategy. Section 4 reports our …ndings and Section 5 discusses their implications. Section 6 concludes.

2 2.1

Data Intensity of the Great East Japan Earthquake

Although Japan has a long history of coping with earthquakes, the 2011 Earthquake is by far the largest in the country’s history since modern measurement of earthquakes began. It is the fourth largest earthquake on record in the world with a magnitude of 9.0 on the Richter scale. The Earthquake occurred in the afternoon of March 11, 2011, triggering a tsunami and causing

4

more than 15,800 deaths and 3,000 cases of missing people (Fire and Disaster Management Agency, 2013). One of the Earthquake’s features is that its e¤ects spread over a very wide area of East Japan in various ways. About 130,000 homes were destroyed. Both cellular and landline phones were dysfunctional over a wide area for a few days (Ministry of Internal A¤airs and Communications, 2011). In addition, approximately 8.6 million households experienced power outages, and 2.3 million households, disruption in water supply (Ministry of Education, Culture, Sports, Science and Technology, 2012). Further planned power outages were inevitable because of the accident at the Fukushima Daiichi Nuclear Power Station. The degree of negative shock di¤ers signi…cantly, depending on location. As our interest is to understand how risk preferences are a¤ected by negative shocks, an ideal explanatory variable would be one that captures the wide variation of negative shocks for people who su¤ered most severely to people who are not a¤ected at all. One straightforward variable would be distance from the epicenter of the Earthquake. This variable, however, is not necessarily ideal because it may not necessarily capture local di¤erences in the intensity of negative shocks— the severity with which an earthquake hits a particular location depends heavily on subsurface structure. Instead, the main explanatory variable we use is seismic intensity of the Earthquake, which is a metric of the strength of an earthquake at a speci…c location (see, e.g., Scawthorn, 2003).6 The seismic intensity of earthquakes (Shindo) is constructed by the Japan Meteorological Agency (JMA).7 The JMA operates more than 1,700 monitoring stations to measure the intensity of earthquakes at each location. People in Japan are very familiar with this intensity measure, Shindo. In fact, Shindo is regularly used in media coverage for intensity of an earthquake at each location (it is similar to a weather report). We focus on seismic intensities measured for the main Earthquake on March 11, 2011. Approximately, an increase of the seismic intensity by two means 10-fold of acceleration (see Table 1 for details of the JMA’s intensity scale). Our intensity measure is the weighted average of seismic intensities from the three closest monitoring stations, where weight is the inverse of the distance between the city hall of the municipality and each monitoring station. Our results are robust to di¤erent ways of constructing an intensity measure as shown in Subsection 5.1. In addition, we collect data on the level of fatalities (which were largely due to the tsunami following the Earthquake) and radiation (following an accident at the Fukushima Daiichi Nuclear Power Plant). However, we use these two measures only as complementary explanatory variables for a robustness check (in Subsection 5.1) mainly because radiation and fatalities are too concen6

The magnitude of an earthquake is a metric of the energy released by the earthquake, and hence, takes a single value for each earthquake, while seismic intensity varies by location. 7 Torch (2011) uses seismic intensity as a proxy for maternal stress to study the e¤ects of stress on birth outcomes.

5

trated in a small number of municipalities and little variation exists for the municipalities covered in our nationally representative survey data (see Appendix Section A for details of these variables). Figure 1 shows the intensity of the Earthquake measured by Shindo in quintiles, together with the location of the epicenter. The darker color indicates higher levels of intensity. As the intensity of an earthquake depends not only on distance from the epicenter but subsurface structure, there are reasonable variations in intensity measure, even within the same distance from the epicenter.8 Of Japan’s 1,724 municipalities as of April 1, 2011, the survey covers 227 municipalities (as described in Subsection 2.2), which are shown with black outlines in Figure 1. The …gure shows that our survey data cover throughout Japan, and thus, there is considerable variation in intensity level among our surveyed municipalities. In addition, the …gure demonstrates that our sample includes very few coastal municipalities in the high intensity regions, where the vast majority of casualties were caused by the tsunami. In fact, of our 227 municipalities, only 6 (2.6 percent) report some population living in ‡ooded areas. Furthermore, none of the 11 municipalities directly a¤ected by the accident at the Fukushima Daiichi Nuclear Power Plant are covered in the survey.

2.2

Panel survey on risk preference

Our measure of risk preferences is elicited directly from a hypothetical lottery question in the Japan Household Panel Survey on Consumer Preferences and Satisfaction (JHPS-CPS). The JHPS-CPS is a nationally representative annual panel survey of the resident population in Japan. The sample is strati…ed according to two criteria: geographical area and city size.9 The data are collected using self-administered paper questionnaires, which are hand-delivered to and picked up from the houses of participating households.10 The surveys are conducted in January and February each year. Since the Earthquake occurred on March 11 in 2011, the data before the Earthquake are from the 2011 survey and the data after the Earthquake are from the 2012 survey for short-term analysis. Thus, the data before the Earthquake were collected 1 or 2 months before the Earthquake, and the data after the Earthquake were collected 10–11 months after the Earthquake. Although the survey was terminated, it resumed in January and February 2016, which is roughly …ve years after the Earthquake in March 2011. While the sample is limited to a randomly selected 80 percent of the original sample owing to budget cutbacks, we use the risk preferences measured in the follow-up 8 The correlation between the distance from the epicenter of the Earthquake and our seismic intensity measure at the municipality level is –0.896 (N=227) in all locations, –0.662 (N=79) in locations with intensity of 4 and higher, and –0.835 (N=148) in locations with intensity lower than 4. 9 All municipalities are classi…ed into 40 stratums: 10 geographical areas and 4 categories corresponding to population size. The number of sample subjects in each stratum is distributed in proportion to the resident population aged 20–69 years. The unit of sampling spot in each stratum is the census unit and is selected by random systematic sampling. 10 All respondents are given a JPY 1,500 (USD 15) cash voucher by completing the survey.

6

survey to examine the long-term e¤ects. The JHPS-CPS asks respondents about their willingness to pay for a hypothetical lottery with a 50 percent chance of winning JPY 100,000 (USD 1,000) or nothing otherwise. This approach to elicit risk preferences using a hypothetical lottery is also taken by Cramer et al. (2002), Hartog et al. (2002), and Guiso and Paiella (2008).11 Appendix Section A shows the exact format of the survey question. A respondent is presented with 8 di¤erent prices in ascending order, from a price of JPY 10 (USD 0.1) in the …rst row of the survey question to a price of JPY 50,000 (USD 500) in the last row. For each row, starting from the …rst, a respondent is required to choose one of two options: buy a lottery ticket at the price (option A) or not buy the ticket (option B). Thus, the reservation price ( ) should lie in the interval between the two prices. This is the price at which a respondent switches from option A to option B and the price in the row immediately before the switch. In this study, we de…ne the reservation price as the midpoint of the two prices.12 This provides an implicit point estimate of the measure of an individual’s risk preference. Note that the question we use is the only survey question eliciting risk preference in the JHPS-CPS that is available as panel data.13 Unfortunately, other important measures, such as time preferences and social preferences, are not available in the panel fashion, and thus, we do not examine them in this study. Respondents who switch more than once (multiple switches) comprise 6.2 percent of the sample and are eliminated. For those who choose option A in all choices (1 percent of the sample), risk preference is estimated using the lower bound of the reservation price, that is, the price in the last decision row. Similarly, for those who choose option B in all choices (4 percent of the sample), we use the upper bound of the reservation price, that is, the price in the …rst decision row. We take two approaches to convert the reservation price ( )— obtained from a respondent’s choice in hypothetical lottery— to a measure of risk aversion, following Cramer et al. (2002). We denote Z as a prize of the lottery and JPY 100,000 (USD 1,000), and

as the probability of winning the prize. In our case, Z is

is 0.5, and thus, the expected value Z is JPY50,000.

11 See Holt and Laury (2002) for a standard approach in an incentivized setting to elicit risk preferences through choices between two lotteries. 12 We treat the discrete variable for our risk aversion measure as continuous. Because our risk aversion measure is censored above at the price of JPY 50,000 and below at the price of JPY 10, the econometric methods should account for the censoring of the risk aversion measure. As a robustness check, we use the estimation method proposed by Honoré (1992), which explicitly accounts for the censoring of a dependent variable with …xed e¤ects. The results are nearly identical to those using our basic speci…cation, suggesting that our …ndings are likely una¤ected by treating the discrete variable for our risk aversion measure as continuous (the results are available upon request). 13 Because this is the only question we can use as panel data, we cannot disentangle risk preference and loss aversion as Callen et al. (2014) did, based on Andreoni and Sprenger (2011).

7

The …rst approach is a simple transformation of the reservation price: Transformed price = 1 = 1

= Z =50; 000

(1)

Note that the greater a respondent’s value of transformed price, the more risk averse the respondent is. In our setting, the values of transformed price only take values between zero and one, where the value of zero corresponds to the case of perfectly risk-neutral preference.14 Appendix Table C1 shows the values of transformed price in our settings. The second approach is to use the Arrow–Pratt measure of absolute risk aversion (Pratt, 1964), =

00

0

U =U , where U (W ) is a standard concave utility function of wealth, W . In expected utility

theory, the utility of wealth without participation in the lottery is equal to expected utility when participating at reservation price : U (W ) = (1

)U (W

) + U (W + Z

). By taking a

second-order Taylor expansion around U (W ), we obtain an estimate of the Arrow–Pratt measure of absolute risk aversion as follows )=[(1=2)( Z 2

Absolute risk aversion = ( Z

2 Z +

2

)]

(2)

See Cramer et al. (2002) for details on derivation. Note that the greater a respondent’s value of absolute risk aversion, the more risk averse the respondent is. In our setting, the values of absolute risk aversion are bounded below and above in a similar manner to the transformed price, where the value of zero corresponds to the case of perfectly risk-neutral preference. Appendix Table C1 reports the values of absolute risk aversion. As discussed in Subsection 5.1, our results are robust to the choice of two risk aversion measures.15 Our measure of risk preference has several advantages. First, one cannot easily isolate the change in risk preference from the change in risk perception because risk perception is likely to change when risk preference changes. However, we can identify changes in risk preferences owing to the fact that we present survey respondents with the same explicit stakes and probabilities both 14

The lottery’s expected value is JPY 50,000 and, thus, risk-averse individuals should be willing to pay less than JPY 50,000 while risk-loving individuals should opt to pay more than JPY 50,000. Because the maximum ticket price in our survey is indeed JPY 50,000, risk-loving individuals can only choose up to JPY 50,000, possibly underestimating risk tolerance. However, this may be of little concern because only a small fraction of respondents chooses JPY 50,000 (1.4 percent in 2011, 1.3 percent in 2012, and 1.6 percent in 2016). 15 Because both risk aversion measures are bounded above and below, we alternatively use the logit transformation of each risk preference measure as outcomes. This method is conventionally used for regression analysis when a dependent variable is a fractional variable bounded between zero and one, such as fractions (McDowell and Cox, 2001). The results using logit-transformed values are quantitatively the same (the results are available upon request). Here, we replace the 0 values in risk preference by 0.0001 when they are logit transformed.

8

before and after the Earthquake. Second, because of low complexity of the lottery question in the JHPS-CPS, the nonresponse rate is quite low— 2.4 percent of the original data— in contrast with a high nonresponse rate observed in previous studies with more complex questions.16 Third, while the samples in previous studies are highly restricted (e.g., to investors or clients), the measure in this study is obtained using a large nationally representative sample. One concern of self-reported measures based on a non-incentivized hypothetical question is whether they actually re‡ect an individual’s underlying risk traits. Several studies document that risk measures obtained by hypothetical survey questions are reliable predictors of actual risk-taking behavior (e.g., Barsky et al., 1997; Donkers et al., 2001; Anderson and Mellor, 2008; Dohmen et al., 2011). In order to check the validity of our risk measures, described in detail in what follows, we simply run regressions of risky behavior on our risk measures. One expects a negative correlation between risk-aversion measures and risk-taking behavior. Appendix Table C3 con…rms the validity of our risk measures; the results show a signi…cant correlation between our risk measures and risky behavior (gambling, drinking, and smoking) with the expected signs.17 We also conduct analysis that does not depend how we construct these two cardinal riskpreference measures in Subsection 5.1. Speci…cally, we exploit the panel nature of the data, and create an ordinal risk measure that takes 1 if the choice of risk category after the Earthquake is higher (i.e., more risk-averse) than the one before the Earthquake, –1 if opposite, and 0 if there is no change. Our results are robust to using this alternative outcome. The sample selection for short-term analysis is conducted as follows. We start with 3,829 subjects with no missing values for the lottery question, risk-taking behavior, age, gender, and household income in 2011 and 2012 surveys.18 We then eliminate 198 respondents who switched more than once in the lottery question.19 Next, 263 respondents who were dropped from the sample in 2012 and 16 respondents who moved to a di¤erent municipality between surveys in 2011 and 2012 are eliminated. The attrition and migration rates are fairly low at 7 percent and 1 percent.20 The reason for low migration rate is because the respondents whose municipality of residence di¤ers between the two years but were dropped from the 2012 sample are counted not 16

For example, Guiso and Paiella (2008) report a 27 percent nonresponse rate. Another concern is that our risk measures capture only risk preferences in the …nancial domain. However, Dohmen et al. (2011) and Vieider et al. (2015) show that risk preferences in di¤erent domains or contexts are not perfectly but strongly correlated to each other. 18 All results shown hereinafter remain unchanged even after including observations with missing values (the results are available upon request). 19 As a robustness check for multiple switches, we use the …rst switch as the reported willingness to pay and re-estimate the models. The estimates are quantitatively unchanged (the results are available upon request). 20 The reasons for attrition in 2012 and 2016 surveys are as follows (however, breakdown by gender is not available): refused (51.3 percent and 50.3 percent), moved away (22.9 percent and 18.2 percent), temporarily absent (7.2 percent and 4.9 percent), and others (18.6 percent and 18.9 percent). 17

9

as migration but as attrition. We demonstrate in Subsection 4.3 that attrition and migration do not seem to a¤ect our results. Our …nal sample comprises 3,352 respondents located across 227 municipalities. The sample selection for long-term analysis is conducted in a similar way.21 We merge the JHPS-CPS data with the earthquake intensity data at the municipality level because the municipality is the most detailed unit of location available in the JHPS-CPS. Table 2 presents the summary statistics of our …nal sample. The mean of transformed price, which takes a value between 0 and 1, is 0.8. The intensity measure of the Earthquake takes a value between 0 and 6.06, with a mean of 2.81. The mean age of the individuals is 52.0 years, 47 percent are male, and about 70 percent are employed.

3

Identi…cation strategy

We examine the e¤ect of the Earthquake on risk preference using panel data before and after the Earthquake. Our identi…cation strategy exploits the variation in the intensity of the Earthquake, while controlling for time-invariant individual characteristics using the individual …xed e¤ects model. Our basic idea is similar to di¤erence-in-di¤erence: we capture the e¤ect of the Earthquake by comparing individuals with zero intensity (control group) and non-zero intensity (treatment group), assuming that the response would have been the same in the absence of the Earthquake. By using panel data, we can isolate the e¤ects of the exogenous treatment (the Earthquake) by comparing the di¤erences before and after the Earthquake across individuals who experienced di¤erent levels of intensity of the treatment. The departure from di¤erence-in-di¤erence is that our treatment is not binary and we can exploit the variation in intensity of the exogenous treatment. More formally, the basic model to test whether the Earthquake in‡uences risk preference can be written as follows Yijt = where

t

t

+ Xjt + Zijt + Wi + "ijt

(3)

is a year e¤ect and Yijt is a measure of risk preference for individual i at location j at

time t. Xjt is intensity of the Earthquake at location j at time t and takes the value of zero before the Earthquake. Zijt represents time-varying individual characteristics, Wi represents unobserved time-invariant individual characteristics, and "ijt is a random shock. The coe¢ cient the e¤ect of the Earthquake that is common to all individuals and

captures

measures the e¤ect of the

Earthquake on risk preference, depending on the intensity of the Earthquake. 21

We start with 3,040 subjects with no missing values for the lottery question, risk-taking behavior, age, gender, and household income in both 2011 and 2016 surveys. We then eliminate 139 respondents who switched more than once in the lottery question. Next, 431 respondents who were dropped from the sample in 2016 and 11 respondents who moved to a di¤erent municipality between surveys in 2011 and 2016 are eliminated. The attrition and migration rates are 17 percent and 0.5 percent. Our …nal sample comprises 2,076 respondents located across 227 municipalities.

10

An econometric issue is the possible presence of unobserved individual …xed e¤ects, Wi , such as di¤erential risk preference formation in response to the riskiness of one’s environment. For example, individuals that have experienced severe natural disasters in the past might form di¤erent attitudes toward risk, which results in unobservable di¤erences in risk preferences. Such di¤erences in the baseline level of risk preference can generate a bias in the estimates if we estimate the above model with cross sectional data ignoring the unobserved …xed e¤ect. In fact, Figure 2 shows that before the Earthquake, risk-averse individuals indeed tended to live in locations with a lower probability of future catastrophic earthquake, which is predicted by a Japanese government agency using the criteria for the probability of a large earthquake in the next 30 years (see Appendix Section A for details of this variable).22 To the extent that such a residential choice and/or the formation of risk preference drive the baseline levels of risk preferences, the estimates based on the cross-sectional data after the natural disaster can be biased, at least in this setting. In order to test our claim further, we test the null hypothesis that the explanatory variables are uncorrelated with unobserved individual-speci…c error. If the null hypothesis were rejected, then it would be necessary to consider a model that accounts for the correlation between unobserved individual-speci…c error and the explanatory variables. We use a cluster-robust version of the Hausman test based on the di¤erence between the …xed e¤ect estimator and the random e¤ect estimator, proposed by Wooldridge (2002). The statistics of a Wald test show the null hypothesis is rejected: 10.0 with p-value of 0.0069 (for short-term analysis between 2011–2012).23 Again, this …nding supports the notion that the estimates obtained from a cross-sectional approach are likely to be biased. Fortunately, we have panel data and can take advantage of the …xed e¤ects estimator to isolate the e¤ects of the unobserved individual characteristics by considering the following speci…cation Yijt = where

+ Xj +

Zijt +

"ijt

(4)

indicates the di¤erence of variables before and (either one year or …ve year) after the

22

Interestingly, such a pattern is more apparent among women than men. Of course, it is di¢ cult to separate whether this result re‡ects residential sorting (i.e., more risk-averse people migrate to safer locations) or di¤erential formation of risk preferences at each location. On one hand, regarding residential sorting, Jaeger et al. (2010) and Bauernschuster et al. (2014) show that individuals who are more willing to take risks are more likely to migrate. In addition, Sahm (2012) …nds that sorting— less risk averse individuals take jobs with higher risk of job loss— is more important than the response to the shock (e.g., job loss, health shock) itself, using ten-years of panel data on hypothetical gambles. On the other hand, regarding local formation of risk preferences, Dohmen et al. (2012) document that risk preference among children is signi…cantly related to the prevailing risk preferences in the region, controlling for parental risk preferences. 23 The test is conducted using the “xtoverid” command in Stata software, which is developed by Scha¤er and Stillman (2006).

11

Earthquake. We denote

Xjt as Xj for notational convenience (given that Xjt = 0 for all ob-

servations before the Earthquake). The speci…cation is di¤erence-in-di¤erence, with Xj taking a continuous variable (Xj takes a value of zero if the intensity of the Earthquake is zero.) In addition to the issue of unobserved …xed e¤ects, another econometric issue is possible nonlinearity in the e¤ects of intensity of the Earthquake. The de…nition of the JMA’s intensity scale as well as its description of di¤erent levels of intensity is documented in Table 1. According to this table, JMA classi…es an intensity level of four as a level at which many people are frightened. In fact, Maruyama et al. (2001) provide …eld evidence that mental health is di¤erentially a¤ected by the level of seismic intensity below and above four. Thus, our main speci…cation considers the possibility that the e¤ect of the Earthquake can be kinked as follows Yijt = where the coe¢ cient

+ Xj + I[Xj

4](Xj

4) +

Zijt +

"ijt

captures the additional linear e¤ect of the Earthquake intensity for locations

with intensity higher than four. In addition, we consider speci…cations in which we use I[Xj and I[Xj

(5)

5] instead of I[Xj

4:5]

4], and most of our results are robust to the change in the cut-o¤

between 4, 4.5, and 5, as shown in Subsection 4.1.24 We also adopt a speci…cation with X squared instead of I[Xj

4] to account for the nonlinearity, and most of our results also remain unchanged.

We estimate Equation (5) using ordinary least squares (OLS).25 We cluster the standard error at the municipality level to allow for arbitrary serial correlation within each municipality (Bertrand et al., 2004). The key assumption underlying the di¤erence-in-di¤erence identi…cation strategy is that trends in outcome would be the same between the treatment and control groups in the absence of a treatment. Even though such an assumption is not directly testable, we can investigate whether the pre-trend is similar for both the treatment and control group. While we wish to go back as further as we can to examine the pre-trend, the best we can trace back to is 2009 because a large number of new subjects have been added to the 2009 sample although the initial wave of the survey 24

Furthermore, we draw on Hansen (1999) to estimate a regression kink model with an unknown location of the kink. We …nd that the location of the kink is X = 3.30 for a default trimming parameter of 0.15 and X = 5.09 for the parameter 0.20 for the short-term analysis (2011–2012). In addition, Wang (2015), which extended Hansen (1999), reject the models with two or three kinks, and only the model with single kink is not rejected at the conventional level (the results are available upon request). These results are consistent with the JMA classi…cation and Maruyama et al.’s (2001) …nding that the e¤ects change at the threshold of X = 4. 25 Alternatively, one could use estimation models that account for the discrete, ordinary nature of our risk aversion measure as the dependent variable. However, in short panels, it is known that the …xed e¤ect estimators in discrete models appear to be biased substantially (e.g., Greene, 2004). Therefore, we rely on …xed e¤ect estimation in linear models. One concern with our approach is that the risk aversion measure is treated as a continuous variable by using the midpoint of the interval between the two risk aversion categories. As a robustness check, we construct an ordinal risk aversion measure that does not rely on the interval of the risk aversion categories described in Subsection 5.1.

12

was conducted in 2003.26 Figure 3 plots the relationship between the intensity of the Earthquake and changes in risk preference between 2009–2011, both of which are before the Earthquake. Here, we plot the changes in residual of transformed price regressed on a constant term (year …xed e¤ect). Each dot in the graph represents the mean of observations within each bin of 0.2 in intensity measure and the size of the dot re‡ects the number of observations in each bin. The solid …tted line is the lowess curve with a bandwidth of 0.5. Panels (a)–(c) show the …gures for the full sample, men, and women, respectively. Given that a vast body of literature provides evidence that gender plays an important role in decision making under risk (e.g., Lerner et al., 2001, Croson and Gneezy, 2009), we divide the sample by gende r. All the …gures show that there is no systematic change in risk aversion with respect to the intensity of the Earthquake.

4

Results

4.1

Short-term e¤ects on risk preferences

Figure 4 shows the relationship between the intensity of the Earthquake and short-term changes in risk preference between 2011–2012. Again, we plot the changes in residual of transformed price regressed on a constant term (year …xed e¤ect). Each dot represents the mean of observations within each bin of 0.2, and the solid …tted line is the lowess curve with a bandwidth of 0.5.27 Panels (a)–(c) show the …gures for the full sample, men, and women, respectively. Panel (a) shows that there is no systematic change in risk aversion with respect to the intensity of the Earthquake for the full sample. However, Panel (a) masks striking gender di¤erences shown in Panels (b) and (c) for men and women, respectively. Panel (b) shows that men who live in locations above the intensity level of roughly four become more risk tolerant as the intensity of the Earthquake increases (recall that a higher number indicates higher risk aversion).28 Interestingly, Panel (c) shows that the same pattern is not observed among women. Unlike men, we do not observe much change in risk preference after the Earthquake, except a slight increase in risk aversion at very high intensity locations. Here, note that risk aversion is much higher among women than men at the level before the 26 In addition, we need to interpret Figure 3 with caution as the survey questions on risk preferences between 2009 and 2011 di¤er in the formats and the monetary amounts of lottery prizes. More speci…cally, the 2011 survey asks respondents to answer the question using multiple choices, while in the 2009 survey, respondents are required to write a monetary value of their willingness to pay. 27 Each bin includes 12.4 municipalities on average, ranging from 1 to 30 municipalities, and 173.5 individuals on average, ranging from 6 to 400 individuals. 28 See Appendix Table C2 for the transition matrix of risk aversion category before and after the Earthquake (men only).

13

Earthquake. In fact, the mean of risk aversion (transformed price) before the Earthquake is 0.737 for men and 0.869 for women. This observation is consistent with the large body of literature that documents that men are less risk averse than women in the vast majority of environments and tasks (for reviews, see Eckel and Grossman, 2008, and Croson and Gneezy, 2009). However, the literature is silent on whether men’s risk preference is more “malleable” to the experience of negative events than that of women, and to which direction risk preferences may change. Our …ndings suggest that men’s risk preferences are more likely to change than those of women and, furthermore, that men become less risk averse, despite their already low level of risk aversion, at least in this setting. Table 3 con…rms our …ndings in the …gures. Table 3 summarizes the estimates for running speci…cation (5) in which the outcome is the transformed price. Note that a negative coe¢ cient implies a decrease in the risk aversion as the intensity of the Earthquake increases. Columns (1)– (4) show no observable or discernable relationship between the intensity of the Earthquake and risk aversion among the full sample, as can be seen in Panel (a) in Figure 4. However, Column (5) shows that men become more risk tolerant at high intensity locations. With regard to magnitude, as the intensity of the Earthquake increases by 2— which corresponds to 10-fold increase of acceleration— above the threshold of 4, our risk aversion measure decreases by 0.060 [= 2

(–0.038 + 0.008)]. This corresponds to 8.1 percent reduction from the mean risk

aversion of 0.737 before the Earthquake for men. Put another way, the size of the reduction is roughly half of the mean di¤erence in risk aversion between men and women (0.132 = 0.869–0.737) before the Earthquake. Columns (6) and (7) show that our estimates are robust to the alternative intensity thresholds of 4.5 and 5; the estimates on the interaction term are negative and statistically signi…cant at the 1 percent level. Finally, Column (8) shows estimates from a more ‡exible form of speci…cation. Speci…cally, we do not impose kink and instead include the quadratic of intensity measures. It is reassuring that the patterns of estimates are similar to other columns that impose kink. Interestingly, women show the opposite pattern to men. The result in Column (9) suggests that women become more risk averse at high intensity locations, which is consistent with Panel (c) in Figure 4.29 Note, however, that the result is barely statistically signi…cant in Columns (9) and (10), and is not signi…cant in Column (11). While we view the contrast between men and women as interesting, the results for women are much less robust than those for men. Thus, henceforth, we focus only on men. 29

We estimate the speci…cation (5) fully interacted with a gender dummy and …nd that the slope is di¤erent between men and women at 1 percent level (the results are available upon request).

14

4.2

Long-term e¤ects on risk preferences

Thus far, we show that men who experienced greater intensity of the Earthquake become more risk tolerant. However, we only observe risk preferences a year after the Earthquake, and thus, the e¤ect may be short lived. Understanding whether the e¤ect is transitory or persistent is important because persistence implies that a large negative shock such as our case can fundamentally alter individual’s risk preferences and thus, have a long-lasting impact on one’s decision making regarding economic behavior. Therefore, it might have a broad implication for post-disaster assistance programs. There is limited and mixed evidence for persistence in the literature, which is understandable because panel data that is long enough after a negative shock to test persistence is rarely available.30 A notable exception is Malmendier and Nagel (2011), who show that individuals who experienced low real stock market returns early in their life demonstrate a more conservative investment behavior in their later life. Eckel et al. (2009) also show that hurricane Katrina evacuees became risk-loving shortly after they were evacuated but such e¤ects disappear within a year. Indeed, we can investigate whether the e¤ects on men’s risk preferences we document above are simply transitory or persistent. Figure 5 presents the same graphs as Figure 4, except that the di¤erence in risk preferences between 2011 (before the Earthquake) and 2016 are plotted instead of the di¤erence between 2011 and 2012. Panels (a), (b), and (c) show the graphs for the full sample, men, and women, respectively. Panel (b) in Figure 5 shows a pattern similar to that in Figure 4: men who live in locations with an intensity level of roughly four become risk tolerant as the intensity of the Earthquake increases. This result suggests that these men are still risk tolerant even …ve years after the Earthquake. Here as well, panel (c) shows that the same pattern is not observed for women. Table 4 presents the estimates for the speci…cation (5) on the long-term e¤ects. Columns (5)– (8) report the estimates for the corresponding long-term e¤ects by comparing men’s risk preferences between 2011 and 2016. Table 4 con…rms our …ndings in the Figure 5. Consistent with the panel (b) in Figure 5, Columns (5)–(8) show that the sign for the interaction term is negative, thus suggesting a decline in the slope in high intensity locations. To our surprise, the long-term estimates in Columns (5)–(8) in Table 4 are comparable or even slightly larger than the corresponding short-term estimates in Columns (5)–(8) in Table 3, indicating that the e¤ects of the Earthquake on men’s risk preferences are persistent at least for …ve years. In addition, it is reassuring since our short-term estimates may not be merely picking 30

In a laboratory setting, some studies examine the stability of risk aversion over time (e.g., Harrison, et al., 2005).

15

up certain noises immediately after the Earthquake.31 Again, we do not …nd much e¤ects among women as shown in Columns (9)–(12) in Table 4.

4.3

Selective attrition and migration

The selective attrition of particular types of individuals can be a potential confounder that may bias the estimate when post-disaster data are used. For example, risk-averse individuals may be more likely to leave an area after a disaster, and hence, may not be observed in the cross-sectional data collected after the negative shock. One of the biggest advantages of panel data is that we can examine whether selective attrition takes place. We start with the short-term data between 2011–2012. First, we test the null hypothesis that the mean risk preference of those who were dropped from the sample after the Earthquake (N = 147) and those who were not (N = 1566) are the same. We …nd that the null could not be rejected with a p-value of 0.89, implying that the risk preferences of those who were dropped from the sample after the Earthquake do not statistically di¤er from those who were not. Second, we regress a dummy for attrition (1 if the individual is not observed after the Earthquake and 0 otherwise) on the same set of variables as in equation (5) to check if the attrition is correlated with the Earthquake’s intensity. Columns (1)–(4) in Table 5 show that the attrition is not systematically correlated with the Earthquake’s intensity. Finally, and most importantly, Column (5) directly investigates this question by regressing an attrition dummy on the risk preferences before the Earthquake to examine if risk preferences of those who are not observed after the Earthquake systematically di¤er from those who continue to be observed. Column (5) shows that past risk preference does not predict attrition, suggesting that there is no selective attrition at least in our case. All these exercises provide no evidence of selective attrition after the Earthquake in our setting. This is probably because our data do not include municipalities that were most severely hit by the tsunami or directly a¤ected by the accident at the Fukushima Daiichi Nuclear Power Plant. Furthermore, selective migration may plague our results in a manner similar to selective attrition (as emphasized by Callen, 2015). As mentioned earlier, the migration rate is low in our case since attrition includes most of the migration. Nonetheless, we conduct the same procedure to verify that migration is not a concern in our data. We regress an indicator for migration on the same variable sets as we do for selective attrition. Columns (6)–(10) in Table 5 demonstrate 31

With regard to magnitude, as the intensity of the Earthquake increases by 2 above the threshold of 4, our risk aversion measure decreases by 0.114 [= 2 (–0.063 + 0.006)]. This corresponds to 15.5 percent reduction from the mean risk aversion of 0.734 before the Earthquake for men.

16

that migration is also not systematically associated with the intensity of the Earthquake and pre-Earthquake risk preferences.32 In our main analysis, we exclude from the …nal sample individuals who are not observed after the Earthquake because it is di¢ cult to assign the Earthquake’s intensity to such individuals. Alternatively, we construct bounds on the estimated treatment e¤ects using a range of assumptions for attriters following an approach based on Horowitz and Manski (2000), Kling and Liebman (2004), and Lee (2009). Table 6 reports the results. For the ease of comparison, Column (4) replicates the unadjusted results from Column (5) of Table 3 when the threshold of the intensity is four. Columns (1) and (7) of Table 6 report lower and upper bounds under the worst-case assumptions. We compute worst-case lower bounds by imputing missing values for each observation in the treatment group (Xj

4) as the minimum value of risk preferences and in the control group

(Xj < 4) as the maximum value of risk preferences. The worst-case upper bounds are computed under the opposite scenario. Since these adjusted estimates are the most extreme ones, and thus, arguably based on implausible assumptions, we construct more reasonable bounds using milder assumptions, following Kling and Liebman (2004). Column (2) constructs a lower bound by imputing missing values for the treatment group as the mean of the treatment group minus 0.25 standard deviations of its distribution, while those for the control group are imputed as the mean of the control group plus 0.25 standard deviations of its distribution. In a similar vein, Column (6) constructs an upper bound by imputing missing values for the treatment group as the mean plus 0.25 standard deviations, while those for the control group are imputed as the mean minus 0.25 standard deviations. Columns (3) and (5) repeat the same procedure as Columns (2) and (6) but use 0.10 standard deviations. Table 6 shows that our main results are barely a¤ected by di¤erent assumptions for the attrition patterns, except for the most extreme case in Column (7), which reports the upper bound of the worst-case scenario. We also change the threshold of the intensity of the Earthquake for the treatment and control group to 4.5 and 5 and obtain similar results (the results are available upon request). We do the same exercise using the long-term data between 2011–2016. It is reassuring that selective attrition and migration do not seem to occur even …ve years after the Earthquake, as 32

We also estimate the main speci…cation (5) including the individuals who relocated between 2011 and 2012 in which we assign the intensity measures from the municipalities reported in 2011. Since the migration rate is low, the results remain essentially unchanged (the results are available upon request).

17

shown in Table 7.33 We also construct bounds on the estimated treatment e¤ects using a range of assumptions for attriters. Table 8 shows that the long-term estimates reported in Table 4 are also robust using a range of assumptions for attriters.

5

Discussion

Thus far, we show that men who experienced greater intensity of the Earthquake become more risk tolerant both in the short term as well as in the long term. In this section, we …rst examine the robustness of our main results on risk preferences (Subsection 5.1). Then, we investigate whether risk-taking behavior is a¤ected by the Earthquake in addition to the change in risk preferences (Subsection 5.2). Finally, we attempt to examine a potential mechanism of our …ndings on risk preferences (Subsection 5.3). To save space, we present the men’s short-term results between 2011–2012 hereafter (the long-term results between 2011–2016 are available upon request).

5.1

Robustness of e¤ects on risk preferences

In addition to the issues of selective attrition and migration in the previous section, we perform a series of robustness checks on our main results for risk preferences presented in Table 3. Here, we present one of the main robustness checks for income and assets in Table 9, and leave the detail descriptions of other robustness checks to Appendix Section B. Table 9 investigates whether our results are robust to changes in income and assets. We additionally control for income, expected future income, and assets, and the estimates remain unchanged. In Appendix Section B1, we further control for a dummy for industry, house ownership as well as assets and property price; however, here as well, the estimates are barely a¤ected. We also directly estimate main speci…cation (5) where outcome is replaced by employment, income, and expected income at the municipality-year level. It is reassuring that none of the estimates are statistically signi…cant at the conventional level. Hereinafter, we brie‡y summarize other robustness checks. See Appendix Section B for the details. First, one may argue that radiation due to the accident at the Fukushima Daiichi Nuclear Power Station may be another factor that a¤ects people’s risk preferences. An alternate intensity measure, which is often used in the literature (especially in violent con‡icts), is fatality rate. However, we do not use these as our intensity measure because both measures are concentrated 33

The attrition rate for men between 2012 and 2016 is 8 percent. When combined with the attrition rate of 9 percent between 2011 and 2012, the overall attrition rate between 2011 and 2016 is 17 percent. The migration rate in the male sample is 0.5 percent between 2011 and 2012, and 0.4 percent between 2011 and 2016. Again, the reason for low migration rate is because the respondents whose municipality of residence di¤ers between the …ve years but were dropped from the 2016 sample are counted not as migration but as attrition.

18

in a small number of municipalities and there is little variation in our sample. Nonetheless, Appendix Section B2 adds radiation level and fatality rate as controls. We …nd that our main estimates remain unchanged. Second, another potential concern are outliers, and thus, we address this issue using Mestimation (Huber, 1964), which places less weight on residuals that are more likely to be outliers. Appendix Section B3 shows that our estimates seem not to be driven by outliers. In addition, the mean reversion of people with high-risk aversion can pose a problem. We re-estimate the main speci…cation by excluding individuals whose risk categories are the highest and the second highest before the Earthquake. Appendix Section B3 shows that our estimates remain unchanged. Third, our results might be driven by a particular form of risk preference measure. To account for this concern, we consider two ways to measure risk preferences: another cardinal measure of risk preference de…ned by equation (2) in Subsection 2.2 (absolute risk aversion) and an ordinal measure of risk preference that takes the value of 1 if the choice of risk category after the Earthquake is higher (i.e., more risk-averse) than that before the Earthquake, –1 if the opposite is the case, and 0 if there is no change between the two surveys. Appendix Section B4 shows that our results are not driven by particular forms of risk preference measures. Finally, Appendix Section B5 demonstrates the robustness of our results to di¤erent ways of constructing the intensity measure. Thus far, we have used the weighted average of the three closest monitoring stations. Instead, we construct the intensity measure using the two closest monitoring stations, the simple average of intensity at the three closest stations, and the only closest station. The estimates remain quantitatively unchanged. In sum, our main results for men’s risk preferences are robust to additional controls, di¤erent speci…cations, and alternative way of constructing both explanatory and outcome variables.

5.2

Risk-taking behavior

So far, our focus has been on how the Earthquake alters individuals’ risk preferences. In this subsection, we examine whether risk-taking behavior, such as drinking, also changes with the intensity of the Earthquake. It is important to mention that because risk-taking behavior can be a¤ected by many factors other than risk preferences (e.g., time preferences and peer e¤ects), we have to view the results on risk-taking behavior with caution.34 Table 10 summarizes the estimates from running speci…cation (5) for the following three forms of risk-taking behavior for men: gambling, drinking, and smoking.35 Note that all the variables 34

Unfortunately, the survey we use does not have questions on time preference and social preference in a panel manner. 35 Even though our risk measure is based on the …nancial question, our analysis on risk-taking behavior do not

19

we construct here capture the most extreme form of behavior that we can observe in the data.36 Speci…cally, a gambling dummy takes one if the person is engaged in gambling (such as horse racing and Japanese pinball, or pachinko) at least once a week. A drinking dummy takes one is if the person drinks 5 or more 12 oz-cans of beer or its equivalent almost every day. Similarly, a smoking dummy takes 1 if the person smokes more than 30 cigarettes a day. The mean of gambling, drinking, and smoking dummies among men before the Earthquake is quite low— 14.4, 2.4, and 2.3 percent, respectively. Column (1) shows the estimates for gambling. The interaction term is positive and statistically signi…cant at the 5 percent level. This result suggests that men who live in locations hit by an intensity level higher than four become more engaged in gambling as the intensity of the Earthquake increases. This result is consistent with our results that men become more risk tolerant after the Earthquake. Column (2) shows that men in these locations become more engaged in heavy drinking, even though the estimate on the interaction term is marginally not statistically signi…cant (p-value = 0.155). Column (3) shows that we do not observe a similar pattern for smoking. An alternative interpretation of our results is that men who lost their jobs because of the Earthquake have spare time for risk-taking behavior, such as gambling and drinking. However, job loss does not seem to drive our results. First, we control for the change in employment status, recognizing its endogeneity. Second, we limit the sample to men whose employment status does not change before and after the Earthquake. The point estimates are quantitatively una¤ected in either case (the results are available upon request).37 For completeness, we present the result for women in Appendix Table C4. The results do not show any e¤ects on risk–taking behavior for women.38

5.3

A potential mechanism: Emotional responses

Thus far, we show evidence that men who are exposed to higher intensities of the Earthquake become more risk tolerant, while such a pattern is not observed among women. However, how does the experience of high intensity of the Earthquake alter men’s risk preferences? The psychology literature points out that emotional responses to negative shocks such as fear and anger may alter examine …nancial risk-taking behavior (e.g., saving, investment, and portfolio choice). This is because …nancial risktaking behavior is confounded with the change in risk perception and risk preferences. Moreover, our data includes only a very coarse measure of such behavior. 36 In fact, we do not observe any statistically signi…cant change in the mean number of drinks, gambling, and cigarettes (the results are available upon request). 37 In addition, we attempt to limit the sample to retired men (i.e., those whose income and time available are less subject to change) but the sample size is too small (N = 144) to gain any meaningful estimate. 38 Note that the mean of each outcome among women before the Earthquake is by order of magnitude smaller than that of men, making it hard to detect any changes, if any.

20

risk preferences (e.g., Leith and Baumeister, 1996; Lerner and Keltner, 2001; Loewenstein et al., 2001). In particular, Fessler et al. (2004) …nd that di¤erences in emotional reactions by gender explain the di¤erence in change in risk preferences in a laboratory setting, which coincides with our …nding. Although it is beyond the scope of our study to fully examine the mechanism underlying our …ndings, we attempt to examine a potential channel through emotions as suggested in the literature. Fortunately, the survey collects three variables that may be related to the emotional stakes of respondents. Speci…cally, the survey asks respondents to indicate how well each of the following applies to them: feeling stressed lately (“stressed”); feeling depressed lately (“depressed”); or have not been sleeping well lately (“sleep problems”). Each variable is answered on a scale of 1–5, where 1 means “particularly true for me” and 5 means “does not hold true at all for me.” Therefore, a higher score indicates that respondents have less mental/emotional issues. Because these variables are highly positively correlated, we construct a summary index measure by taking the unweighted average of the standardized values of three emotional variables. For each variable, the standardized value is calculated by subtracting the mean and dividing by standard deviation using two years of data, so that each component of the index has a mean of zero and standard deviation of one. The aggregation improves the statistical power to detect e¤ects that are in the same direction among similar outcomes (Kling et al., 2007).39 One limitation of this approach is that we implicitly weight each outcome equally, which may not be appropriate. Table 11 summarizes the estimates from running speci…cation (5), in which the outcome is the summary index measure. Note that the higher the score, the lower the mental issues of the individual (i.e., he or she is more mentally healthy). Column (1) for men shows that the interaction term is negative and statistically signi…cant at the 5 percent level. This result suggests that men who live in locations hit by an intensity level higher than four become less mentally healthy.40 This pattern is consistent with our results that men become more risk tolerant and engage more in risk-taking behavior after the Earthquake at high intensity locations. An alternative explanation for these results is that men who lost jobs due to the Earthquake have greater mental issues. In fact, some studies document the link between job loss and mental health (e.g., Burgard et al., 2007; Schaller and Stevens, 2015). Therefore, we control for change 39

We also adopt two additional approaches to show that the aggregation method we apply for these variables does not play a crucial role. The …rst approach is to calculate the average standardized treatment e¤ect, in which the regression coe¢ cient for each outcome is divided by the standard deviation of each outcome to be standardized, and a simple average of all standardized coe¢ cients is taken (see Kling and Liebman, 2004; Finkelstein et al., 2012). The second approach is to estimate the principal component score instead of aggregating these three variables. Both approaches yield similar estimates as in Table 11 (both results are available upon request). 40 Some studies document the impact of natural disasters on short- and medium-run mental health (Frankenberg et al., 2008; Paxson et al., 2012; Rhodes et al., 2010).

21

in employment status, once again recognizing its endogeneity, in Column (2). The estimate is essentially unchanged. Interestingly, Columns (3) and (4) show that the estimate on the interaction term for women is not statistically signi…cant at the conventional level.41 We cannot conclude that emotional responses drive our results on changes in risk preferences because there are many alternative explanations. However, it is interesting that the emotional response patterns correspond to those of risk preferences and risk-taking behavior, and that we observe such a strong pattern only among men, while the pattern among women is much weaker.

6

Conclusion

Many studies have documented the relationship between risk preferences and various economic decision making. Standard economic models assume that risk preferences are stable across time but recent studies suggest that risk preferences may be altered by various shocks. However, little is yet known how risk preferences can be altered. Furthermore, very little is known about whether the changes are transitory or persistent. To shed light on this research question, we test whether experience of the Great East Japan Earthquake alters risk preferences of individuals. We exploit a unique panel dataset that enables us to track changes in risk preference of the same individuals before and after the Earthquake, unlike previous studies that use cross-sectional data collected only after the negative shocks occurred. Furthermore, the panel structure of our data allows us to address the issue of selective attrition and migration. We …nd that men who experienced higher intensity of the Earthquake become more risk tolerant. Furthermore, we …nd that the e¤ects on men’s risk preferences are persistent even …ve years after the Earthquake at almost the same magnitude as that shortly after the Earthquake. Finally, we …nd corroborative evidence that these men become more engaged in gambling. This study may be especially important because natural disasters are becoming increasingly prevalent all over the world, including in developed countries. Past studies have predominantly examined cases in developing countries. Our results show that the risk preferences of people in Japan— a developed country with a history of coping with frequent natural disasters— are also a¤ected by a very large negative shock. Our study has some limitations. First, we can examine the e¤ects of the Earthquake on only a limited number of risk-taking behavioral forms, such as gambling. Second, while we show that 41

Note that we fail to reject the null hypothesis that the estimate on the interaction term I[Xj 4](Xj 4) are the same for male and female ( 2 (1) = 0.66 (p-value = 0.416)). The estimates come from Columns (2) and (4) in Table 11.

22

such changes in risk preferences are at least persistent for …ve years after the Earthquake, we cannot examine how long such e¤ects persist because such data do not exist yet. Third, we cannot fully understand the mechanism of how experience of high intensity of the Earthquake alters individuals’ risk preferences. The intensity measures of the Earthquake we use indeed capture the degree of shock physically felt by each individual, and thus, such shocks can plausibly a¤ect people’s risk preferences. While we show some suggestive evidence of emotional responses stories, it is impossible to identify whether our results are driven entirely by the emotional responses, by other channels, or by combinations of these channels. These questions are beyond the scope of our study but clearly remain as an avenue for future research.

References Anderson, Lisa R., and Jennifer M. Mellor. 2008. “Predicting Health Behaviors with an Experimental Measure of Risk Preference.” Journal of Health Economics, 27(5): 1260–1274. Andreoni, James, and Charles Sprenger. 2011. “Uncertainty Equivalents: Testing the Limits of the Independence Axiom.” Unpublished Manuscript. Barsky, Robert B., F. Thomas Juster, Miles S. Kimball, and Matthew D. Shapiro. 1997. “Preference Parameters and Behavioral Heterogeneity: An Experimental Approach in the Health and Retirement Survey.” The Quarterly Journal of Economics, 112(2): 537–579. Bauernschuster, Stefan, Oliver Falck, Stephan Heblich, Jens Suedekum, and Alfred Lameli. 2014. “Why are educated and risk-loving persons more mobile across regions?”Journal of Economic Behavior & Organization, 98: 56–69. Bchir, Mohamed A., and Marc Willinger. 2013. “Does the exposure to natural hazards a¤ect risk and time preferences? Some insights from a …eld experiment in Perú.”Unpublished Manuscript. Bertrand, Marianne, Esther Du‡o, and Sendhil Mullainathan. 2004. “How Much Should We Trust Di¤erences-in-Di¤erences Estimates?”The Quarterly Journal of Economics, 119 (1): 249–275. Brown, Ryan, Verónica Montalva, Duncan Thomas, and Andrea Velásquez. 2017. “Impact of Violent Crime on Risk Aversion: Evidence from the Mexican Drug War.” NBER Working Paper No. 23181. Brunnermeier, Markus K., and Stefan Nagel. 2008. “Do Wealth Fluctuations Generate TimeVarying Risk Aversion? Micro-Evidence on Individuals’Asset Allocation.”American Economic Review, 98(3): 713–736.

23

Burgard, Sarah, Jennie Brand, and James House. 2007. “Toward a Better Estimation of the E¤ect of Job Loss on Health.” Journal of Health and Social Behavior, 48: 369–384. Byrnes, James P., David C. Miller, and William D. Schafer. 1999. “Gender Di¤erences in RiskTaking: A Meta-analysis.” Psychological Bulletin, 125(3): 367–383. Callen, Michael. 2015. “Catastrophes and Time Preference: Evidence from the Indian Ocean Earthquake.” Journal of Economic Behavior & Organization, 118: 199–214. Callen, Michael, Mohammad Isaqzadeh, James D. Long, and Charles Sprenger. 2014. “Violence and Risk Preference: Experimental Evidence from Afghanistan.”American Economic Review, 104(1): 123–148. Cameron, Lisa, and Manisha Shah. 2015. “Risk-Taking Behavior in the Wake of Natural Disasters.” Journal of Human Resources, 50(2): 484–515. Cassar, Alessandra, Andrew Healy, and Carl von Kessler. 2011. “Trust, Risk and Time Preferences after a Natural Disaster: Experimental Evidence from Thailand.” Unpublished Manuscript. Chuang, Yating, and Laura Schechter. 2015. “Stability of experimental and survey measures of risk, time, and social preferences: A review and some new results.” Journal of Development Economics, 117: 151–170. Cohn, Alain, Jan Engelmann, Ernst Fehr, and Michel Maréchal. 2015. “Evidence for countercyclical risk aversion: an experiment with …nancial professionals.”American Economic Review, 105(2): 860–885. Cramer, Jan S., Joop Hartog, Nicole Jonker, and Mirjam van Praag. 2002. “Low Risk Aversion Encourages the Choice for Entrepreneurship: An Empirical Test of a Truism.” Journal of Economic Behavior & Organization, 48(1): 29–36. Croson, Rachel, and Uri Gneezy. 2009. “Gender Di¤erences in Preferences.” Journal of Economic Literature, 47(2): 448–474. Dohmen, Thomas, Armin Falk, David Hu¤man, Uwe Sunde, Jürgen Schupp, and Gert G. Wagner. 2011. “Individual Risk Attitudes: Measurement, Determinants and Behavioral Consequences.” Journal of the European Economic Association, 9(3): 522–550. Dohmen, Thomas, Armin Falk, David Hu¤man, and Uwe Sunde. 2012. “The Intergenerational Transmission of Risk and Trust Attitudes.” Review of Economic Studies, 79(2): 645–677. Donkers, Bas, Bertrand Melenberg, and Arthur van Soest. 2001. “Estimating Risk Attitudes Using Lotteries: A Large Sample Approach.” Journal of Risk and Uncertainty, 22(2): 165–195.

24

Eckel, Catherine C., Mahmoud A. El Gamalb, and Rick K. Wilson. 2009. “Risk Loving After the Storm: A Bayesian-Network Study of Hurricane Katrina Evacuees.” Journal of Economic Behavior & Organization, 69: 110–124. Eckel, Catherine C., and Philip J Grossman. 2008. “Men, women and risk aversion: Experimental evidence.”Chapter 113. In Handbook of Experimental Economics Results, Charles R. Plott and Vernon L. Smith (eds.), Vol. 1, New York, Elsevier Press. Fessler, Daniel M.T., Elizabeth G. Pillsworth, and Thomas J. Flamson. 2004. “Angry men and disgusted women: An evolutionary approach to the in‡uence of emotions on risk taking.” Organizational Behavior and Human Decision Processes, 95(1): 107–123. Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber, Joseph P. Newhouse, Heidi Allen, Katherine Baicker, and the Oregon Health Study Group. 2012. “The Oregon Health Insurance Experiment: Evidence from the First Year.” The Quarterly Journal of Economics, 127(3): 1057–1106. Fire and Disaster Management Agency. 2013. Heisei 23 Nen Touhoku Chihou Taiheiyou Oki Jishin Ni Tsuite (Report on the Great East Japan Earthquake),

Vol 149.

http://www.fdma.go.jp/bn/higaihou/pdf/jishin/149.pdf (in Japanese) [accessed April 3, 2014]. Frankenberg, Elizabeth, Jed Friedman, Thomas Gillespie, Nicholas Ingwersen, Robert Pynoos, Iip Umar Rifai, Bondan Sikoki, Alan Steinberg, Cecep Sumantri, Wayan Suriastini, and Duncan Thomas. 2008. “Mental Health in Sumatra after the Tsunami.” American Journal of Public Health, 98(9): 1671–1677. Gallagher, Justin. 2014. “Learning about an infrequent event: Evidence from ‡ood insurance takeup in the United States.” American Economic Journal: Applied Economics, 6(3): 206–233. Greene, William. 2004. “The Behaviour of the Maximum Likelihood Estimator of Limited Dependent Variable Models in the Presence of Fixed E¤ects.” Econometrics Journal, 7(1): 98–119. Guiso, Luigi, and Monica Paiella. 2008. “Risk Aversion, Wealth, and Background Risk.” Journal of the European Economic Association, 6(6): 1109–50. Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2013. “Time Varying Risk Aversion.” Booth Working Paper No. 13–64. Hansen, Bruce E. 1999. “Threshold e¤ects in non-dynamic panels: Estimation, testing, and inference.” Journal of Econometrics, 93(2): 345–368. Harrison, Glenn W., Eric Johnson, Melayne M. McInnes, and E. Elisabet Rutström. 2005. “Temporal stability of estimates of risk aversion.”Applied Financial Economics Letters, 1(1): 31–35. 25

Hartog, Joop, Ada Ferrer-i-Carbonell, and Nicole Jonker. 2002. “Linking Measured Risk Aversion to Individual Characteristics.” Kyklos, 55(1): 3–26. Haushofer, Johannes, and Ernst Fehr. 2014. “On the psychology of poverty.” Science, 344(6186): 862–867. Holt, Charles A., and Susan K. Laury. 2002. “Risk Aversion and Incentive E¤ects.” American Economic Review, 92(5): 1644–1655. Honoré, Bo E. 1992. “Trimmed Lad and Least Squares Estimation of Truncated and Censored Regression Models with Fixed E¤ects.” Econometrica, 60(3): 533–565. Horowitz, Joel L., and Charles F. Manski. 2000. “Nonparametric Analysis of Randomized Experiments with Missing Covariate and Outcome Data.” Journal of the American Statistical Association, 95 (449): 77–84. Huber, Peter J. 1964. “Robust estimation of a location parameter.” Annals of Mathematical Statistics, 35: 73–101. Imas, Alex. 2016. “The Realization E¤ect: Risk-Taking After Realized Versus Paper Losses.” American Economic Review, 106(8): 2086–2109. Jaeger, David A., Thomas Dohmen, Armin Falk, David Hu¤man, Uwe Sunde, and Holger Bonin. 2010. “Direct Evidence on Risk Attitudes and Migration.”Review of Economics and Statistics, 92: 684–689. Jakiela, Pamela, and Owen Ozier. 2015. “The Impact of Violence on Individual Risk Preferences: Evidence from a Natural Experiment.” Unpublished Manuscript. Japan Meteorological Agency. 2015. Tables explaining the JMA Seismic Intensity Scale. Available at http://www.jma.go.jp/jma/en/Activities/inttable.html [assessed July 30, 2014]. Kim, Young-Il, and Jungmin Lee. 2014. “The Long-Run Impact of Traumatic Experience on Risk Aversion.” Journal of Economic Behavior & Organization, 108: 174–186. Kling, Je¤rey, and Je¤rey Liebman. 2004. “Experimental Analysis of Neighborhood E¤ects on Youth.”’ Unpublished Manuscript. Available at http://www.nber.org/mtopublic/483.pdf [accessed April 3, 2014]. Kling, Je¤rey, Je¤rey Liebman, and Lawrence Katz. 2007. “Experimental Analysis of Neighborhood E¤ects.” Econometrica, 75(1): 83–119. Lee, David. S. 2009. “Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment E¤ects.” Review of Economic Studies, 76: 1071–1102.

26

Leith, Karen Pezza, and Roy F. Baumeister. 1996. “Why do bad moods increase self-defeating behavior? Emotion, risk tasking, and self-regulation.” Journal of Personality and Social Psychology, 71(6): 1250–1267. Lerner, Jennifer S., and Dacher Keltner. 2001. “Fear, Anger, and Risk.” Journal of Personality and Social Psychology, 81: 146–159. Lerner, Jennifer S., Roxana M. Gonzalez, Deborah A. Small, and Baruch Fischho¤. 2003. “E¤ects of Fear and Anger on Perceived Risks of Terrorism: A National Field Experiment.”Psychological Science, 14(2): 144–150. Loewenstein, George F., Elke U. Weber, Christopher K. Hsee, and Ned Welch. 2001. “Risk as Feelings.” Psychological Bulletin, 127(2): 267–286. Malmendier, Ulrike, and Stefan Nagel. 2011. “Depression Babies: Do Macroeconomic Experiences A¤ect Risk Taking?” The Quarterly Journal of Economics, 126(1): 373–416. Maruyama, Soichiro, Young-Sook Kwon, and Kanehisa Morimoto. 2001. “Seismic Intensity and Mental Stress after the Great Hanshin-Awaji Earthquake.”Environmental Health and Preventive Medicine, 6: 165–169. McDowell,

Allen,

and Nicholas J. Cox. 2001. “Logit Transformation” Available at:

http://www.stata.com/support/faqs/statistics/logit-transformation/ [accessed April 3, 2014]. Ministry per

of on

Internal Information

A¤airs and

and

Communications.

Communication

in

2011. Japan

White Available

http://www.soumu.go.jp/johotsusintokei/whitepaper/eng/WP2011/2011-index.html

Paat [ac-

cessed April 3, 2014]. Ministry of Education, Culture, Sports, Science and Technology. 2012. White Paper on Science and Technology. Available at http://www.mext.go.jp/english/whitepaper/1323541.htm [accessed April 3, 2014]. Moya, Andrés. 2012. “Violence, Mental Trauma and Induced Changes in Risk Attitudes among the Displaced Population in Colombia.” Unpublished Manuscript. Necker, Sarah, and Michael Ziegelmeyer. 2016. “Household risk taking after the …nancial crisis.” The Quarterly Review of Economics and Finance, 59: 141–160. Page, Lionel, David A. Savage, and Benno Torgler. 2014. “Variation in risk seeking behaviour following large losses: a natural experiment.” European Economic Review, 71: 121–131.

27

Paxson, Christina, Elizabeth Fussell, Jean Rhodes, and Mary Waters. 2012. “Five Years Later: Recovery from Post Traumatic Stress and Psychological Distress among Low-income Mothers A¤ected by Hurricane Katrina.” Social Science & Medicine, 74(2): 150–157. Pratt, John W. 1964. “Risk Aversion in the Small and in the Large.” Econometrica, 32: 122–136. Reynaud, Arnaud, and Cécile Aubert. 2014. “Do Natural Disasters Modify Risk Preferences? Evidence from a Field Experiment in Vietnam.” Unpublished Manuscript. Rhodes, Jean, Christian Chan, Christina Paxson, Cecilia E. Rouse, Mary Waters, and Elizabeth Fussell. “The Impact of Hurricane Katrina on the Mental and Physical Health of Low-Income Parents in New Orleans.” 2010. American Journal of Orthopsychiatry, 80(2): 237–247. Sahm, Claudia. 2012. “How Much Does Risk Tolerance Change?” Quarterly Journal of Finance, 2(4). Scawthorn, Charles. 2003. “Earthquakes: Seismogenesis, measurement, and distribution.”Chapter 4. In Earthquake Engineering Handbook, W. Chen, and C. Scawthorn (eds.), Boca Raton, FL, CRC Press. Scha¤er, Mark E., and Steven Stillman. 2006. “XTOVERID: Stata module to calculate tests of overidentifying restrictions after xtreg, xtivreg, xtivreg2, xthtaylor.” Statistical Software Components S456779, Boston College Department of Economics, revised Nov 2, 2011. Schaller, Jessamyn, and Ann Hu¤ Stevens. 2015. “Short-run E¤ects of Job Loss on Health Conditions, Health Insurance, and Health Care Utilization.” Journal of Health Economics, 43: 190–203. Stigler, George J., and Gary S. Becker. 1977. “De Gustibus Non Est Disputandum.” American Economic Review, 67(2): 76–90. Torch, Florencia. 2011. “The E¤ect of Maternal Stress on Birth Outcomes: Exploiting a Natural Experiment.” Demography, 48:1473–1491. van den Berg, Marrit, Ricardo Fort, and Kees Burger. 2009. “Natural hazards and risk aversion: Experimental evidence from Latin America.” Unpublished Manuscript. Vieider, Ferdinand M., Mathieu Lefebvre, Ranoua Bouchouicha, Thorsten Chmura, Rustamdjan Hakimov, Michal Krawczyk, and Peter Martinsson. 2015. “Common components of risk and uncertainty attitudes across contexts and domains: Evidence from 30 countries.” Journal of the European Economic Association, 13(3): 421–452.

28

Voors, Maarten J., Eleonora E. M. Nillesen, Philip Verwimp, Erwin H. Bulte, Robert Lensink, and Daan P. Van Soest. 2012. “Violent Con‡ict and Behavior: A Field Experiment in Burundi.” American Economic Review, 102(2): 941–964. Wang, Qunyong. 2015. “Fixed-e¤ect panel threshold model using Stata.” Stata Journal, 15(1): 121–134. Willinger, Marc, Mohamed A. Bchir, and Carine Heitz. 2013. “Risk and time preferences under the threat of background risk: A case-study of lahars risk in central Java.”Unpublished Manuscript. Wooldridge, Je¤rey M. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA, MIT Press.

29

Figure 1. Seismic Intensity of the 2011 Great East Japan Earthquake

Notes: The epicenter of the Earthquake (38.322°N 142.369°E) is marked with a cross. Of the 1,724 municipalities in Japan on April 1, 2011, 227 municipalities in our survey are shown with black outline. The intensity of the earthquake comes from the seismic intensity of the Earthquake (Shindo in Japanese), which is a metric of the strength of the earthquake at a specific location, and is constructed by the Japan Meteorological Agency (JMA). The definition of the JMA’s intensity scale as well as its description of different levels of intensity is documented in Table 1. See the main text for construction of our measure of the intensity of the earthquake and Appendix Section A for the data source.

30

3 2 1 0

Risk Aversion (before earthquake)

Figure 2. Risk Preferences and Earthquake Hazard Prediction before the Earthquake

0

.2

.4

.6

.8

1

Pre-Earthquake Hazard Prediction

Notes: Pre-Earthquake hazard prediction on the x-axis is the probability of future catastrophic earthquake, predicted by a Japanese government agency with the criterion being the probability of experiencing a large earthquake in the next 30 years (see Appendix Section A for details of the variable). Risk aversion on the y-axis is the transformed price before the Earthquake in 2011. See the main text for construction of the risk aversion measure. The line is a lowess curve with a bandwidth of 0.3. When we plot the same graph separately for men and women, we see the same patterns for both gender (the results are available upon request).

31

Figure 3. Pre-trend of Changes in Risk Preferences and the Intensity of the Earthquake Outcome: Risk Aversion Measure 1 (Transformed Price)

0

1

2

3

4

Seismic Intensity

5

6

.4 .2 0 -.4

-.2

Risk Aversion

.2 0 -.4

-.2

Risk Aversion

.2 0 -.2 -.4

Risk Aversion

(c) Women

.4

(b) Men

.4

(a) Full Sample

0

1

2

3

4

Seismic Intensity

5

6

0

1

2

3

4

Seismic Intensity

5

6

Notes: The data are from the JHPS-CPS in 2009 and 2011, both of which are before the Earthquake. Risk aversion on the y-axis is the changes in transformed price. See the main text for construction of the variable. The seismic intensity of the Earthquake (Shindo in Japanese) on the x-axis is a metric of the strength of the earthquake at a specific location. We plot the changes in residual of transformed price regressed on a constant term (year fixed effect). Each dot in the graph represents the mean of observations within each bin of 0.2 in intensity measure and the size of the dot reflects the number of observations in each bin. The solid fitted line is a lowess curve with a bandwidth of 0.5. The vertical dotted lines correspond to seismic intensity of four. In total, there are 4,420 individuals (2,066 men and 2,354 women) and 227 municipalities.

32

Figure 4. Short-Term Changes in Risk Preferences and the Intensity of the Earthquake Outcome: Risk Aversion Measure 1 (Transformed Price)

(b) Men

0

1

2

3

4

Seismic Intensity

5

6

.15 .05 0 -.05

Risk Aversion

-.15

-.1

.05 0 -.05 -.15

-.1

Risk Aversion

.1

.1

.15 .1 .05 0 -.05 -.1 -.15

Risk Aversion

(c) Women

.15

(a) Full Sample

0

1

2

3

4

Seismic Intensity

5

6

0

1

2

3

4

Seismic Intensity

5

6

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). Risk aversion on the y-axis is the transformed price. See the main text for construction of the variable. The seismic intensity of the Earthquake (Shindo in Japanese) on the x-axis is a metric of the strength of the earthquake at a specific location. We plot the changes in residual of transformed price regressed on a constant term (year fixed effect). Each dot in the graph represents the mean of observations within each bin of 0.2 in intensity measure and the size of the dot reflects the number of observations in each bin. The solid fitted line is a lowess curve with a bandwidth of 0.5. The vertical dotted lines correspond to seismic intensity of four. In total, there are 3,352 individuals (1,566 men and 1,786 women) and 227 municipalities. We address the potential concern of outliers in Subsection 5.1 and Appendix Section B3.

33

Figure 5. Long-Term Changes in Risk Preferences and the Intensity of the Earthquake Outcome: Risk Aversion Measure 1 (Transformed Price)

0

1

2

3

4

Seismic Intensity

5

6

.1 0 -.2

-.1

Risk Aversion

.2 .1 0 -.2

-.1

Risk Aversion

.1 0 -.1 -.2

Risk Aversion

(c) Women .2

(b) Men

.2

(a) Full Sample

0

1

2

3

4

Seismic Intensity

5

6

0

1

2

3

4

Seismic Intensity

5

6

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2016 (5 years after the Earthquake). Risk aversion on the y-axis is the transformed price. See the main text for construction of the variable. The seismic intensity of the Earthquake (Shindo in Japanese) on the x-axis is a metric of the strength of the earthquake at a specific location. We plot the changes in residual of transformed price regressed on a constant term (year fixed effect). Each dot in the graph represents the mean of observations within each bin of 0.2 in intensity measure and the size of the dot reflects the number of observations in each bin. The solid fitted line is a lowess curve with a bandwidth of 0.5. The vertical dotted lines correspond to seismic intensity of four. In total, there are 2,076 individuals (976 men and 1,100 women) and 185 municipalities.

34

Table 1. JMA Seismic Intensity Scale (Shindo) Seismic intensity

Human perception and reaction

Indoor situation

Imperceptible to people, but recorded by seismometers. Felt slightly by some people keeping quiet in buildings.

-

2

Felt by many people keeping quiet in buildings. Some people may be awoken.

Hanging objects such as lamps swing slightly.

3

Felt by most people in buildings. Felt by some people walking. Many people are awoken. Most people are startled. Felt by most people walking. Most people are awoken.

Dishes in cupboards may rattle.

5 Lower (4.5 – 5)

Many people are frightened and feel the need to hold onto something stable.

Hanging objects such as lamps swing violently. Dishes in cupboards and items on bookshelves may fall. Many unstable ornaments fall. Unsecured furniture may move, and unstable furniture may topple over.

5 Upper (5 – 5.5)

Many people find it hard to move; walking is difficult without holding onto something stable.

6 Lower (5.5 – 6)

It is difficult to remain standing.

6 Upper (6 – 6.5)

It is impossible to remain standing or move without crawling. People may be thrown through the air.

Dishes in cupboards and items on bookshelves are more likely to fall. TVs may fall from their stands, and unsecured furniture may topple over. Many items of unsecured furniture move and may topple over. Doors may become wedged shut. Most items of unsecured furniture move and are more likely to topple over.

0 1

4

7

-

Hanging objects such as lamps swing significantly, and dishes in cupboards rattle. Unstable ornaments may fall.

Most items of unsecured furniture move and topple over or may even be thrown through the air.

Notes: The seismic intensity scale (Shindo) is used to measure the degree of shaking at a specific location in Japan. It is computed using acceleration data for each monitoring station by Japan Meteorological Agency (JMA). After adjusting the raw digital acceleration data to the adjusted acceleration (a gal), the JMA seismic intensity scale (I) can be obtained by I=2log10a+0.94. Thus, the measure can be considered essentially as the logarithm of the acceleration. In other words, an increase of the JMA intensity scale corresponds to an exponential increase in acceleration. Approximately, an increase of JMA seismic intensity scale by two means 10-fold of acceleration. This table comes from JMA’s descriptions on seismic intensity for human perception and reaction as well as indoor situations (Japan Meteorological Agency, 2015).

35

Table 2. Summary Statistics Variables

N

Mean

SD

Min

Max

A. Individual-Level Variables (before the Earthquake) Risk Preferences Risk aversion measure 1 (transformed price)

3,409

0.8070

0.2140

0

0.9998

Risk aversion measure 2 (absolute risk aversion)

3,409

1.8224

0.4230

0

2.0000

Gambling (once or more a week)

3,409

0.0883

0.2838

0

1

Drinking (5 or more cans of beer, almost every day)

3,409

0.0135

0.1154

0

1

Smoking (more than 30 cigarettes per day)

3,409

0.0123

0.1103

0

1

Age (in years)

3,409

52.0

12.6

22

78

Male

3,409

0.47

0.50

0

1

Annual household income (in JPY million)

3,409

6.37

3.80

1

20

Employed

3,409

0.71

0.45

0

1

Percentage change in expected income 1 year

3,240

-0.009

0.04

-0.09

0.09

Percentage change in expected income 5 year

3,211

-0.025

0.09

-0.2

0.2

Assets (in JPY million)

3,101

13.5

17.4

2.5

100

Stress

3,367

2.79

1.1

1

5

Depression

3,367

3.34

1.1

1

5

Sleep problems

3,367

3.81

1.1

1

5

X (seismic intensity)

227

2.81

1.94

0

6.06

Radiation (µSv/h)

227

0.10

0.24

0

2.40

Behavior

Individual Characteristics

B. Municipal-Level Variables

Fatality rate (per 1,000 population) 227 0.25 2.42 0 26.9 Notes: See Appendix Section A for construction of each municipal-level variable. The data are from the JHPS-CPS in 2011 (a year before the Earthquake), 2012 (a year after the Earthquake), and 2016 (5 years after the Earthquake). The values of risk aversion measure 2 (absolute risk aversion) are multiplied by 1000. Note that the number of observations for percentage change in expected income 1 year, percentage change in expected income 5 year, assets, stress, depression, and sleep problems differs slightly because of missing values .

36

Table 3. Short-Term Changes in Risk Preferences Outcome: Risk Aversion Measure 1 (Transformed Price) Full Sample X (X – 4) * 1[X ≥ 4]

Women

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

0.002

0.001

0.001

0.003

0.008*

0.006

0.004

0.018*

-0.004

-0.003

-0.002

-0.011**

(0.002)

(0.002)

(0.002)

(0.005)

(0.005)

(0.004)

(0.003)

(0.010)

(0.002)

(0.002)

(0.002)

(0.005)

-0.010

-0.038**

0.017*

(0.009)

(0.017)

(0.009)

(X – 4.5) * 1[X ≥ 4.5]

-0.015

-0.056***

0.024*

(0.013)

(0.021)

(0.012)

(X – 5) * 1[X ≥ 5]

-0.032*

-0.090***

0.027

(0.019)

(0.031)

(0.018)

X-squared Constant

Men

-0.001

-0.003*

0.002**

(0.001)

(0.002)

(0.001)

0.005

0.005

0.005

0.005

0.007

0.008

0.011

0.006

0.003

0.002

0.000

0.005

(0.005)

(0.005)

(0.005)

(0.005)

(0.011)

(0.010)

(0.010)

(0.011)

(0.005)

(0.005)

(0.005)

(0.006)

×

×

×

×

×

×

×

×

×

×

×

×

Mean of outcome (before)

0.807

0.807

0.807

0.807

0.737

0.737

0.737

0.737

0.869

0.869

0.869

0.869

N of individuals

3,352

3,352

3,352

3,352

1,566

1,566

1,566

1,566

1,786

1,786

1,786

1,786

Individual FE

R-squared 0.000 0.000 0.001 0.000 0.003 0.004 0.005 0.002 0.002 0.002 0.001 0.002 Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. Outcome is the short-term changes in transformed price between 2011 and 2012. See the main text for construction of the variable. Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

37

Table 4. Long-Term Changes in Risk Preferences Outcome: Risk Aversion Measure 1 (Transformed Price) Full Sample X (X – 4) * 1[X ≥ 4]

Women

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

0.002

0.001

-0.001

0.011

0.006

0.004

-0.001

0.025*

-0.002

-0.003

-0.002

-0.003

(0.003)

(0.003)

(0.003)

(0.008)

(0.006)

(0.005)

(0.005)

(0.015)

(0.004)

(0.003)

(0.003)

(0.010)

-0.024*

-0.063***

0.014

(0.012)

(0.023)

(0.016)

(X – 4.5) * 1[X ≥ 4.5]

-0.034**

-0.098***

0.031

(0.016)

(0.029)

(0.021)

(X – 5) * 1[X ≥ 5]

-0.052*

-0.148***

0.055*

(0.027)

(0.047)

(0.029)

X-squared Constant

Men

-0.003*

-0.006**

0.001

(0.001)

(0.003)

(0.002)

-0.031*** -0.030*** -0.028*** -0.034***

-0.015

-0.014

-0.007

-0.018

-0.044*** -0.043*** -0.045*** -0.046***

(0.009)

(0.008)

(0.008)

(0.009)

(0.016)

(0.016)

(0.016)

(0.018)

(0.011)

(0.010)

(0.010)

(0.012)

×

×

×

×

×

×

×

×

×

×

×

×

Mean of outcome (before)

0.806

0.806

0.806

0.806

0.734

0.734

0.734

0.734

0.870

0.870

0.870

0.870

N of individuals

2,076

2,076

2,076

2,076

976

976

976

976

1,100

1,100

1,100

1,100

Individual FE

0.002 0.002 0.002 0.002 0.010 0.013 0.013 0.008 0.001 0.002 0.003 0.000 R-squared Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2016 (5 years after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. Outcome is the long-term changes in transformed price between 2011 and 2016. See the main text for construction of the variable. Standard errors clustered at the municipality are reported in parentheses. The sample size in this table is smaller than that in Table 3 because the sample in 2016 is limited to randomly selected 80 percent of the original sample. Note that the estimates in Table 3 are quantitatively similar even if we restrict the sample to the same one as Table 4 (the results are available upon request). Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01. 38

Table 5. Short-term Selective Attrition and Migration (Men Only) Outcome X (X – 4) * 1[X ≥ 4]

Attrition Dummy (1)

(2)

(3)

(4)

-0.004 (0.006) 0.007 (0.018)

-0.002 (0.005)

-0.002 (0.004)

-0.019 (0.012)

(7)

(8)

(9)

-0.000 (0.001) 0.007 (0.006)

0.000 (0.001)

0.001 (0.002)

-0.002 (0.003)

-0.004 (0.006) 0.003 (0.002)

0.001 (0.001)

-0.004 (0.025) 0.096*** 0.093*** 0.093*** 0.105*** 0.089*** (0.013) (0.013) (0.012) (0.014) (0.020) 1,713 0.000

1,713 0.000

(10)

0.008 (0.007)

X-squared

N of individuals R-squared

(6)

-0.013 (0.033)

(X – 5) * 1[X ≥ 5]

Constant

(5)

-0.009 (0.025)

(X – 4.5) * 1[X ≥ 4.5]

Risk aversion measure 1(in 2011)

Migration Dummy

1,713 0.000

1,713 0.001

1,713 0.000

0.004 (0.003)

0.003 (0.003)

0.001 (0.003)

0.003 (0.003)

-0.007 (0.012) 0.010 (0.009)

1,574 0.002

1,574 0.002

1,574 0.001

1,574 0.002

1,574 0.001

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. An attrition dummy takes one if the individual is not observed after the Earthquake, and 0 otherwise. The number of short-term attrition between 2011–2012 is 147 (8.6 percent). Similarly, a migration dummy takes one if the individual lives in different municipalities after the Earthquake. The number of short-term migration between 2011–2012 is 8 (0.5 percent). Note that the attrition includes most of the migrations. Risk aversion measure 1 is the transformed price. Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01. 39

Table 6. Bounds Estimation with Short-term Sample Attrition (Men only) Outcome: Risk Aversion Measure 1 (Transformed Price)

Unadjusted effect

Lower bounds Worst case

0.25SD

0.10SD

(1)

(2)

(3)

X

-0.022*** (0.008)

0.005 (0.004)

(X – 4) * 1[X ≥ 4]

-0.077*** (0.028)

Constant Individual FE R-squared

Upper bounds 0.10SD

0.25SD

Worst case

(4)

(5)

(6)

(7)

0.006 (0.004)

0.008* (0.005)

0.008* (0.004)

0.009** (0.004)

0.037*** (0.007)

-0.038** (0.015)

-0.036** (0.015)

-0.038** (0.017)

-0.035** (0.015)

-0.033** (0.015)

0.007 (0.026)

0.127*** (0.018)

0.015 (0.010)

0.011 (0.010)

0.007 (0.011)

0.006 (0.010)

0.002 (0.009)

-0.114*** (0.015)

×

×

×

×

×

×

×

0.046

0.004

0.004

0.003

0.003

0.003

0.039

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. For the ease of comparison, Column (4) replicates the unadjusted results from Column (5) of Table 3. Columns (1) and (7) report lower and upper bounds under the worst-case assumptions, respectively. We compute worst-case lower bounds by imputing missing values for each observation in the treatment group (X≥4) as the minimum value of risk preferences, and as the maximum value of risk preferences in the control group (X<4). Worst-case upper bounds are computed under the opposite scenario. Column (2) constructs a lower bound by imputing missing values for the treatment group as the mean of the treatment group minus 0.25 standard deviations of its distribution, while those for the control group are imputed as the mean of the control group plus 0.25 standard deviations of its distribution. In a similar vein, Column (6) constructs an upper bound by imputing missing values for the treatment group as the mean plus 0.25 standard deviations, while those for the control group are imputed as the mean minus 0.25 standard deviations. Columns (3) and (5) repeat the same procedure as Columns (2) and (6) but instead use 0.10 standard deviations. The number of attrition is 147 (8.6 percent). Results for specifications with (X – 4.5) * 1[X ≥ 4.5], (X – 5) * 1[X ≥ 5], and X-squared are very similar (the results are available upon request). Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

40

Table 7. Long-term Selective Attrition and Migration (Men Only) Outcome X (X – 4) * 1[X ≥ 4]

Attrition Dummy (1)

(2)

(3)

(4)

-0.004 (0.009) 0.016 (0.035)

-0.000 (0.009)

0.003 (0.008)

-0.006 (0.021)

(7)

(8)

(9)

-0.001 (0.001) 0.012 (0.007)

-0.001 (0.001)

0.001 (0.002)

-0.006 (0.004)

0.005 (0.009) 0.001 (0.004)

0.001* (0.001)

0.036 (0.050) 0.179*** 0.174*** 0.169*** 0.178*** 0.147*** (0.023) (0.023) (0.022) (0.025) (0.038) 1,181 0.000

1,181 0.000

(10)

0.014 (0.009)

X-squared

N of individuals R-squared

(6)

-0.063 (0.054)

(X – 5) * 1[X ≥ 5]

Constant

(5)

0.001 (0.045)

(X – 4.5) * 1[X ≥ 4.5]

Risk aversion measure 1(in 2011)

Migration Dummy

1,181 0.001

1,181 0.000

1,181 0.001

0.004 (0.004)

0.003 (0.004)

0.001 (0.004)

0.005 (0.005)

0.007 (0.004) -0.001 (0.002)

980 0.005

980 0.004

980 0.001

980 0.005

980 0.001

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2016 (5 years after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. An attrition dummy takes one if the individual is not observed after the Earthquake, and 0 otherwise. The number of long-term attrition between 2011–2016 is 205 (17 percent). Similarly, a migration dummy takes one if the individual lives in different municipalities after the Earthquake. The number of long-term migration between 2011–2016 is 4 (0.4 percent). Note that the attrition includes most of the migrations. Risk aversion measure 1 is the transformed price. Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

41

Table 8. Bounds Estimates on Long-term Sample Attrition (Men Only) Outcome: Risk Preference Measure 1 (Transformed Price)

Unadjusted effect

Lower bounds Worst case

0.25SD

0.10SD

(1)

(2)

(3)

X

-0.034*** (0.012)

0.002 (0.005)

(X – 4) * 1[X ≥ 4]

-0.215*** (0.048)

Constant Individual FE R-squared

Upper bounds 0.10SD

0.25SD

Worst case

(4)

(5)

(6)

(7)

0.004 (0.005)

0.006 (0.006)

0.006 (0.005)

0.007 (0.005)

0.045*** (0.012)

-0.067*** (0.019)

-0.061*** (0.019)

-0.063*** (0.023)

-0.052*** (0.020)

-0.046** (0.020)

0.109* (0.058)

0.202*** (0.030)

0.001 (0.014)

-0.008 (0.014)

-0.015 (0.016)

-0.019 (0.013)

-0.027** (0.013)

-0.228*** (0.026)

×

×

×

×

×

×

×

0.123

0.019

0.013

0.010

0.008

0.005

0.077

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2016 (5 years after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. For the ease of comparison, Column (4) replicates the unadjusted results from Column (5) of Table 4. Columns (1) and (7) report lower and upper bounds under the worst-case assumptions, respectively. We compute worst-case lower bounds by imputing missing values for each observation in the treatment group (X≥4) as the minimum value of risk preferences, and as the maximum value of risk preferences in the control group (X<4). Worst-case upper bounds are computed under the opposite scenario. Column (2) constructs a lower bound by imputing missing values for the treatment group as the mean of the treatment group minus 0.25 standard deviations of its distribution, while those for the control group are imputed as the mean of the control group plus 0.25 standard deviations of its distribution. In a similar vein, Column (6) constructs an upper bound by imputing missing values for the treatment group as the mean plus 0.25 standard deviations, while those for the control group are imputed as the mean minus 0.25 standard deviations. Columns (3) and (5) repeat the same procedure as Columns (2) and (6) but instead use 0.10 standard deviations. The number of attrition is 205 (17 percent). Results for specifications with (X – 4.5) * 1[X ≥ 4.5], (X – 5) * 1[X ≥ 5], and X-squared are very similar (the results are available upon request). Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

42

Table 9. Robustness Checks: Control for Income and Wealth Effects (Men Only) Outcome: Risk Aversion Measure 1 (Transformed Price)

Additional Control Variables

X (X – 4) * 1[X ≥ 4] Constant Individual FE Income Expected Income Asset Mean of outcome (before) No. of individuals R-squared

Baseline

Income

Expected Income

Asset

(1) 0.008* (0.005) -0.038** (0.017) 0.007 (0.011)

(2) 0.008* (0.005) -0.038** (0.017) 0.007 (0.011)

(3) 0.008* (0.005) -0.038** (0.017) 0.008 (0.011)

(4) 0.008 (0.005) -0.038** (0.017) 0.008 (0.011)

×

× ×

× × ×

× × × ×

0.737 1,566 0.003

0.737 1,566 0.004

0.737 1,566 0.006

0.737 1,566 0.006

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. Column (2) controls for self-reported income measured in brackets of JPY 2 million (roughly USD 20,000 in 2012). Column (3) controls for the expected future income, using two survey measures for the percentage change in expected income 1 year or 5 years after the survey. Column (4) controls for assets measured in brackets of JPY 2.5 million (roughly USD 25,000 in 2012). For Columns (3) and (4), we replace the missing value for expected income and assets with 0, and add a dummy for such observations in the estimation. We also drop these observations, but the estimates are essentially identical. The results for specifications with (X – 4.5) * 1[X ≥ 4.5], (X – 5) * 1[X ≥ 5], and X-squared are very similar (the results are available upon request). Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

43

Table 10. Changes in Behavior (Men Only) Outcomes

Gambling

Drinking

Smoking

(1)

(2)

(3)

-0.012*

-0.003

-0.001

(0.006)

(0.002)

(0.003)

0.042**

0.014

0.001

(0.018)

(0.010)

(0.010)

0.040**

0.008

0.014

(0.017)

(0.006)

(0.009)

× ×

× ×

× ×

Mean of outcome (before)

0.144

0.024

0.023

N of individuals

1,566

1,566

1,566

R-squared

0.003

0.002

0.000

X (X – 4) * 1[X ≥ 4] Constant Individual FE Income

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. A gambling dummy takes one if the person is engaged in gambling once or more a week. A drinking dummy takes 1 if the person drinks 5 or more cans of beer (12 oz. per can) or its equivalent a day almost every day. A smoking dummy takes 1 if the person smokes more than 30 cigarettes per day. See Appendix Section A for the survey question for each outcome. Results for specifications with (X – 4.5) * 1[X ≥ 4.5], (X – 5) * 1[X ≥ 5], and X-squared are very similar (the results are available upon request). Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01.

44

Table 11. Emotional Responses Men (2) 0.010 (0.015) -0.117** (0.050) -0.044 (0.100) -0.004 (0.039)

(3) 0.012 (0.014)

(4) 0.012 (0.014)

-0.058 (0.049)

-0.059 (0.049)

-0.020 (0.031)

-0.020 (0.031)

× ×

× × ×

× ×

× × ×

N of individuals

1,544

1,544

1,768

1,768

R-squared

0.005

0.005

0.001

0.001

X (X – 4) * 1[X ≥ 4]

(1) 0.010 (0.015) -0.118** (0.050)

Women

Employment Status Constant Individual FE Income

-0.004 (0.040)

Employment status

-0.012 (0.064)

Notes: The data are from the JHPS-CPS in 2011 (a year before the Earthquake) and 2012 (a year after the Earthquake). X is the seismic intensity of the Earthquake (Shindo in Japanese), a metric of the strength of an earthquake at a specific location. Outcome is the simple average of the standardized values of three emotion variables: “stress,” “depression,” and “sleep problems.” All emotion variables are based on respondents indicating on a 5-point scale how well each of the questions applies to them. The “Stress” question asked whether a respondent has been feeling stressed lately. The “Depression” question asked whether a respondent has been feeling depressed lately. The “Sleep problems” question asked whether a respondent has not been sleeping well lately. A scale of 1 means “particularly true for me” and 5 means “does not hold true at all for me.” Therefore, a higher score indicates that the respondent has less mental/emotional issues (i.e., is mentally healthier). Columns (2) and (4) further control for employment status in addition to income. The number of individuals is slightly smaller in this table than other tables (1,544 vs. 1,566) due to the missing values for emotional outcomes (N = 22). Results for specifications with (X – 4.5) * 1[X ≥ 4.5], (X – 5) * 1[X ≥ 5], and X-squared are very similar (the results are available upon request). Standard errors clustered at the municipality are reported in parentheses. Significance levels are *p < 0.10, **p < 0.05, and ***p < 0.01

45