Do Natural Disasters Lead to More Migration? Evidence from Indonesia

Do Natural Disasters Lead to More Migration? Evidence from Indonesia Chun-Wing Tse November 2011 Abstract Using Indonesian panel datasets, I examine how earthquakes, volcanic eruptions and floods affect household migration. The study separately analyzes the impact of these natural disasters on the tendency of entire households to migrate, as well as for part of the household to split off and migrate. Contrary to conventional wisdom, I find that all three types of disasters significantly reduce migration rates. Nevertheless, the channels of impact are quite different. Earthquakes reduce household size, earnings and non-business assets, each of which tends to reduce migration rates. Volcanic eruptions on the other hand raise the value of farmland, which, in turn, reduces migration. Floods have no significant impacts on household assets or earnings, and their effect remains unexplained. Keywords: Indonesia, natural disasters, migration JEL codes: O15, Q54 I would like to thank Dilip Mookherjee for all his guidance and support. I also wish to thank Daniele Paserman and Michael Manove for their advice and comments. I am also grateful to Ye Li, Jie Hou, Julian Chan, Hyo-Youn Chu, Saori Chiba and seminar participants at Boston University. All errors are my own. Department of Economics, Boston University, 270 Bay State Rd., Boston, MA 02215 (winghk@bu.edu) 1

1 Introduction Given the rising losses from environmental calamities across the globe (Cameron and Shah (2010), UNISDR (2007) ), the study of natural disasters has never been more crucial in our time. In the year 2010 alone, natural disasters of various types have killed at least a quarter million people, which exceeds the number of people killed in terrorist attacks in the past 40 years combined (U.S. Federal Emergency Management Agency). The research on environmental risk is even more important in development economics given the fact that poor households have limited resources to deal with natural disasters, which are highly unpredictable and aggregate in nature (Cavallo and Noy (2009), Noy and Bang (2010), Ebeke and Combes (2010), Cavallo et al. (2010)). Based on a descriptive study of cross-country analysis, the International Organization for Migration (IOM) in 2009, suggests that rising natural disasters can drive toward more migration because poor households in developing countries resort to migration to stay away from disaster-prone areas. Soaring climate change exacerbates the problem of water shortages and agricultural failures. Seismic activities destroy industrial establishments or deter prospective investors from investing in quake zones. Having lost their livelihoods after natural disasters, households need to make a living elsewhere and thus move out (IOM 2009). Natural disasters, however, can actually lower migration. With their assets destroyed by disasters, households become financially constrained which discourages migration. Disasters also present an aggregate shock that hurts most households in affected villages, and therefore, it becomes more difficult to borrow from others to finance migration (Yang 2008). They are not forced to migrate but forced to stay. Disasters can also alter household asset composition to lower migration; eruptions and floods enrich soil fertility with lava ash and alluvial deposits (Ahmad, 2011) respectively, which causes households to be less willing to dispose all the farm assets and move out. Furthermore, the marginal product of labor of some sectors can rise after disasters and lead to less migration. Wages in the construction sector increase after earthquakes due to a higher labor demand for village rebuilding. With respect to eruptions and floods, both make farmlands more fertile and hence drive up agricultural wages. Finally, some other social factors such as stronger family ties and enhanced community bonding after disasters can all lower migration. 2

Using Indonesia as the case country, this paper attempts to understand the link between natural disasters and migration. The study relies on two nationally representative datasets. Given the panel nature of the datasets, I am able to conduct a longitudinal study to account for household fixed effects and measure how time variation of disasters alters migration at both the household and individual level. Moreover, disasters of various types occur in Indonesia frequently and I do not treat each type of disaster alike, but rather separately analyze the impacts of various types of disasters on different geographic levels of moving. Specifically, this paper studies the three most common types of disasters in Indonesia, earthquakes, volcanic eruptions and floods, to find out how these disasters affect migration across provinces, districts and subdistricts. In particular, this paper concentrates on household migration. I separately examine the effects of these disasters on migration of the entire household and split-household migration, which is defined as part of the household splitting from the original households to move to a new location. This study then examines the different economic channels through which disasters operate to shape household-moving decisions. The baseline empirical results show that all three types of disasters drive down migration at both the household and individual level. Nevertheless, the channels of impact on household migration are quite different. Eruptions push up the values of farm assets which can be due to enrichment of soil fertility by lava ash. Households with greater farm business assets are also less likely to move out. Hence, eruptions suppress household migration by increasing farm assets. On the other hand, evidence shows that migration is less likely to occur if a household is smaller, has lower earnings or non-business assets. Earthquakes decrease household size, earnings and non-business assets, which explains why they lower migration rates. Floods, however, do not lead to reductions in household assets or earnings and the negative impacts of floods on household migration remain unexplained. The economics literature on the link between natural disasters and migration is relatively new. Naude (2008) adopts a panel analysis at the country level and finds that environmental shocks drive up migration through increasing conflicts. The cross-household studies by Halliday (2007), Ó Gráda (1997) and Attzs (2008) analyze specific one-off deadly disasters and relate cross-section disaster exposure with migration at the individual level. They find that individual migration goes up after massive disasters. 3

This study is most related to Yang s (2008) paper, which is a panel study on El Salvador, examining how the massive one-off event of earthquakes in 2001 affects household migration. Specifically, he highlights how the earthquakes affect household access to credit to finance migration and discovers that the earthquakes present a large aggregate shock across households in quake-affected villages. Hence, it becomes more difficult to borrow from other households to pay fixed migration costs as most households in the villages are financially impaired, which actually drives down migration. This finding is similar to the overall result of my paper. I investigate a wide range of natural disasters and a broad range of channels of impact, such as household size, earnings, value of farmland and non-business assets. The same outcome is obtained despite heterogeneous impacts of different types of disasters on different kinds of assets. Similar to Yang s study, this paper also adopts a longitudinal analysis and focuses on household migration. However, I treat different disasters as heterogeneous shocks. Different disasters cause differential changes in household asset composition and marginal product of labor in various sectors, which result in different migration decisions. Furthermore, I separately examine the impacts of these natural disasters on split-household and whole-household migration, which are two contrasting decisions. In split migration, the household aims to reduce risk by sending members to other locations to diversify sources of income in anticipation of receiving remittances in the future. However, whole-household migration is a risk-taking strategy which involves displacing the entire household to a new location. Hence, the above facts point toward the need to treat different types of disasters as heterogeneous shocks and to separately analyze various forms of household migration. Finally, I do not just consider specific one-off events but account for time variation of disasters. Given the fact that disasters of various types occur in some developing countries regularly, such as the Philippines, Bangladesh and Pakistan, a longitudinal study of time variation of natural disasters is important. The paper is organized as follows. Section 2 provides background information on Indonesia, illustrating the demographics and disaster occurrence in the country. Section 3 outlines the data used and gives some descriptive statistics. Section 4 discusses the relationship between disasters and household migration. Section 5 describes the empirical strategy, and Section 6 presents the 4

main findings. Section 7 checks for robustness and Section 8 concludes. 2 Background According to the UN Office for the Coordination of Humanitarian Affairs, Indonesia is the most disaster-prone country of the world. Most parts of Indonesia are on the fault line of volcanic origin, which gives rise to frequent outbreaks of massive earthquakes and volcanic eruptions. The country is also regularly hit by floods due to its large scale deforestation. In 2009 alone, it experienced 469 earthquakes with a magnitude of 5 or higher. Sumatra, Java and Papua were especially hard hit. According to the government data (BPS Indonesia), floods have accounted for about 40 percent of Indonesia s disasters in the past few years. Figure 1 shows the time-series patterns of earthquakes and floods and Figures 2 to 4 provide geographic snapshots on where earthquakes, eruptions and floods occurred in the country between 1988 and 2000. Java and Sumatra Islands have always been the black spots of disasters. However, people do not stay away from disasters but continue to live with the risk of increasing environmental calamities. Figure 5 depicts the population density of Indonesia in 2000, with most dwellers crowded in Java and Sumatra where disasters of different types frequently occur. West Javanese people need to face the regular occurrence of floods and earthquakes. The volcanoes in Yogyakarta pose a constant threat to the inhabitants there, where the eruption in 2010 destroyed numerous villages and killed more than 390 people (New York Times 2010). However, population density of West Java well surpasses 1000 per km square and Yogyakarta has more than 980 people per same size of area (SEDAC) in 2000. Using simple cross-province regressions, the results show that population density in 1993 is not negatively correlated with disasters within 50 years before 1993. This implies that people are not driven away by disasters, but stay with the environmental risk instead. It has always been claimed that communities in Indonesia stay near volcanic areas regardless of the constant threat of eruptions. The regression of rice yield on eruptions at the province level within the last 50 years shows that provinces with more eruptions can produce higher rice yield. Lava ash from volcanoes enhances soil fertility and boosts the farm yield, which explains why people 5

settle and stay near volcanic areas. 3 Data and descriptive statistics This paper uses two datasets for the empirical analysis. The first one is a panel dataset from the Indonesian Family Life Survey (IFLS), a nationally representative survey covering both rural and urban areas. This dataset gives a nation-wide sample of households spreading across 13 provinces in the first wave of the survey in 1993 (IFLS1) with three more waves conducted in 1997 (IFLS2), 2000 (IFLS3) and 2007 (IFLS4). 1 One prominent feature of this longitudinal survey is the very high tracking rate. The survey did not just attempt to re-interview original households sampled in 1993, but also all the migrant households and those split off from the original households. In IFLS4, 94% of IFLS1 households were re-contacted and this rate is as high as, or even higher than, most longitudinal surveys in the United States and Europe. High re-interview rates contribute significantly to data quality because this lessens the attenuation bias due to nonrandom attrition, which is a critical issue of concern for studies on migration and natural disasters. 2 A dummy variable indicating whether a household migrates between two successive survey years is the main outcome of interest in the empirical study. But first, we need a clear definition of household migration. In this paper, I define two forms of household migration: (1) split-household migration and (2) whole-household migration. For split-household migration, one or more household members, but not including the head of household, leave and establish a new household in the new location. On the other hand, if the whole household including the household head moves to a new place, I call this whole-household migration. Apart from a detailed section of household migration history, IFLS also asks several comprehensive sets of questions to obtain the economic variables of the sample households. Specifically, I focus on household size, aid received, remittances, total household earnings and levels of different assets, to study how natural disasters alter these variables to shape the two forms of household migration. The second dataset is the Indonesian DesInventar Database (DesInventar) administered by Data 1 IFLS2+ was also carried out in 1998 to measure the impact of financial crisis starting from 1997. Yet only about 20% of the households of IFLS2 were re-interviewed in IFLS2+. 2 I also test whether there exists non-random attrition and the results are not sensitive to the treatment of households which dropped out from the samples. 6

& Informasi Bencana Indonesia. The aim of DesInventar is to record every disaster happening in Indonesia from the early 20th century. The details include location, date, fatalities, financial losses, damage of infrastructure and other relevant information of the disasters. This paper looks at earthquakes, volcanic eruptions and floods, which are the three most common types of natural disasters occurring in Indonesia. DesInventar adopts a method of counting natural disasters different from the traditional practice. First, a disaster is defined as the set of effects caused by an event on human lives and economic infrastructure on a geographic unit of minimum resolution. (DesInventar) It imposes no thresholds on the amount of damage for an environmental shock to be regarded as a disaster. Furthermore, instead of treating a single event of environmental shock as one disaster, DesInventar counts the number of minimal geographic units, referred to as kecamatan (subdistrict), affected in the event. Thus, DesInventar counts an earthquake event of extensive geographic coverage as multiple earthquake disasters. Thus, this makes statistics recorded by DesInventar look inflated compared with statistics kept under the traditional practice. Yet such a method is desirable for this study, as disaster of extensive coverage should receive more weight. DesInventar defines earthquakes, eruptions and floods as follows: Earthquakes - All movements in the earth s crust causing any type of damage or negative effect on communities or properties. Volcanic eruptions - eruptions with disastrous effects: eruption and emission of gas and ashes, stone falls (pyroclast), flows of lava, etc. Floods - Water that overflows river-bed levels ( riverine floods ) and runs slowly on small areas or vast regions in usually long duration periods (one or more days). This study just retains households and their split-offs which exist in all four waves of the survey. I only keep households with clear migration history between 1993 and 2007. Households without information of some economic variables such as household size, earnings and assets are discarded. This finally leaves the study with 8,217 households. Table 1 presents the descriptive statistics of the sample households. The disaster statistics record the annual average number of each type of disaster at province level occurring between 1988 and 7

2000. Households, on average, experience 0.099 earthquakes and 0.20 eruptions annually between 1988 and 2000. Floods are more prevalent in Indonesia and the sample households are exposed to more than two floods every three years. Table 1 also presents the migration figures between 1993 and 2007. I first consider the annual rate of migration in general, combining both split-household migration and whole-household moving. On average, 1.4% of the households annually move across provinces. The corresponding rates across kabupatens (districts) and kecamatans (subdistricts) are respectively 3.2% and 4.6%, which are considerably high. Yet when we examine the two forms of household moving separately, the statistics shows that most household migration is in the form of splits. More than 3.8% of households have split-off households located in a new province. On the contrary, whole-household migration is much less frequent. Annually, just less than 0.1% of households move to a new province as a whole on average. Table 1 also shows that there is a generally even proportion of urban and rural households. Most of the household heads have only finished elementary education and about 15% of the households are headed by females. Table 2 links household economic well-being in 2000, with household migration between 2000 and 2007. I separate the entire samples into three groups: (1) households with no migration, (2) households which split and migrate across provinces between 2000 and 2007 but do not move out as a whole, (3) households which move to a new province as a whole. Households which split have a bigger size with higher earnings and assets of various kinds. On the other hand, households which migrate as a whole are smaller and have less non-business assets. The median figures illustrate a much clearer picture. 50% of households which move out as a whole have non-business assets less than 4.7 million rupiah. Yet the corresponding figure for split migrant households is 17 million rupiah. In general, households which migrate as a whole have less farm and non-business assets compared with the other two groups. The above descriptive analysis portrays the disparity in asset composition between migrant and non-migrant households, which illustrates how household asset composition links with migration. Before presenting the empirical analysis, the paper first explains how disasters may drive down migration. Since the empirical study emphasizes the disparity between split migration and wholehousehold migration, the following section also describes how the two forms of migration differ. 8

4 How natural disasters drive down household migration Migration can increase with natural disasters because households want to stay away from the risk of disasters or they need to make a living elsewhere if their livelihoods are eliminated. Yet households exposed to natural disasters can actually be less likely to move out. There exist three possible reasons: (1) increase in marginal product of labor, (2) decrease in financial resources to pay for migration and (3) strengthened social bonding and mutual insurance. (1) Increase in marginal product of labor (MPL) Natural disasters can cause recession, higher unemployment and lower wages. Yet the affected areas with infrastructure and houses destroyed have a high demand for labor to rebuild villages. The MPL of the reconstruction sector rises, which induces households to stay for better employment opportunities. Regarding the agricultural sector, soil fertility can be enriched by lava ash in eruptions and alluvial deposits in floods, which increase the productivity of farming. Hence, households may choose to stay. (2) Decrease in financial resources to pay for migration With assets destroyed and earnings reduced, households are less capable of affording migration. Thus, they are not forced to migrate but forced to stay. Moreover, households find it more difficult to borrow from others to finance migration as disasters present an aggregate shock and hurt most households living in the affected villages (Yang, 2008). Disasters pose liquidity constraints and, as a result, lower migration. (3) Strengthened social bonding and mutual insurance Disasters can boost family ties and strengthen social bonding, especially in developing countries since social capital plays a significant role in less developed economies. Households may choose to cope with disaster shock by accumulating social capital instead of moving out. Thus, they are less likely to migrate. Furthermore, households with houses destroyed by disasters need members to stay to rebuild houses. Law and order may also break down after disasters and households should 9

remain to protect property and land rights. This paper will empirically examine the first two reasons and leave out the third due to data limitations. 4.1 Split-household migration and whole-household migration Split-household migration is a rarely studied concept, which involves not just household splits but the split-off households moving to a new location. In this study, the migration of a single individual to set up a single-member household is also classified as split-household migration. Splithousehold migration differs from individual migration in various aspects: (1) in individual moving, the migrants may just move out and enter another household in the new location; (2) individual migration tends to be temporary and migrants may return after some time; and (3) individual migrants are, in general, more attached to the original household. Yet split-off households are considered separate from the original household. Households may also consider split migration as an insurance strategy. Considering household members, especially the young and educated groups, as human asset, the head of household can diversify risk by spreading the asset to various locations. The remittances received from split-off households is also an important source of income, which enables the original household to better mitigate the risk of future economic shocks. Whole-household migration is a completely different concept, which is defined as the moving of the entire household at a certain geographic level. The insurance factor is much less significant when the head of household decides to relocate the entire household. The decision is based on the push factors of the origin and the pull factors of the destination, taking into account the total migration cost. This paper separately examines how disasters shape these two forms of household migration. The rest of the subsection will discuss the following economic determinants of split and whole-household migration: household size, total earnings, external transfer and household assets. Household size: A bigger household will be more likely to split and migrate, as it has more human asset to allocate to various locations for the purpose of diversifying risk. On the other hand, a household with more members is less likely to move out as a whole because migration cost 10

increases with the size of household. Total earnings: Households with more earnings have a higher likelihood to split and migrate because they have more financial resources to support the splits. Furthermore, considering splithousehold migration as a risky investment, the risk of the investment decreases with the income of the households. Thus, higher earnings lower the risk for split-off households to move out and make an even higher income elsewhere. On the contrary, higher earnings can drive down whole-household migration because the opportunity cost of moving increases with the current earnings. External transfer: External transfer such as remittances and government aid, is a positive factor for split migration. Similar to the theory related to household earnings, households with more external transfer have more financial resources to pay for split migration. Yet remittances can have totally different effects from government aid on whole-household migration. Households receiving more remittances can better afford migration. Government aid received, however, induces people to stay in order to obtain more aid money, which points toward Samaritan s Dilemma. Household assets: Households with more assets are better endowed financially to support splits. Similar to the theory on total earnings, the risk of split-household migration falls with the wealth of the households. Greater assets of various kinds lower the risk and hence, drive up split migration. On the other hand, households with more assets are less likely to migrate as a whole because it is costly for households to sell and dispose of their assets to move out. Cost for whole-household migration rises with the amount of assets. The above discussion suggests how disasters operate through a variety of economic channels to shape different household-migration decisions. The following sections empirically examine the above claims. 5 Empirical Strategy The empirical analysis first starts with equation (1): M it = α 0 + α 1 D ct + θ i + ρ t + ɛ it (1) The LHS variable M it is the migration dummy indicating whether household i moves out at a 11

given geographic level between time t and t + 1. The three different geographic levels are across provinces, across kabupatens (districts) and across kecamatans (subdistricts). The most important RHS variable is D ct, which uses the definition given by DesInventar to count the annual average number of disasters happening in province c, where household i resides in between time t 1 and t. The panel survey spans from 1993 to 2007, with a total of four waves. The regression specification includes t =1993, 1997 and 2000. I take t 1 =1988 for t =1993 and t + 1 =2007 when t =2000. In Indonesia, earthquakes, eruptions and floods occur regularly in different provinces across time. This environmental context provides a sufficient degree of dispersion for the RHS disaster variable, D ct. Equation (1) also controls for household fixed effect, θ i and ɛ it captures idiosyncratic errors. ρ t denotes time dummies, which is essential because the panel dataset is unevenly spaced. However, I first run a regression on equation (1) without including the household fixed effect and conduct a simple OLS analysis. The OLS results tell us how the cross-household variation of disasters correlates with migration in the following period. Such correlation gives the causal impact of disasters on household migration only when D ct is uncorrelated with the combined error term, θ i + ɛ it. This assumption is arguably plausible given the random nature of disasters. However, it could be possible that people with a high unobserved propensity to migrate tend to live in a disaster prone province, which will render the coefficients from a simple cross-section regression biased. Thus the paper takes advantage of the panel nature of the IFLS dataset and includes household fixed effects, θ i, in the equation. Household migration, M it, consists of split-household migration and whole-household moving. Equations (2) and (3) give the regression specification, respectively, on these two different forms. Split it = β 0 + β 1 D ct + δ i + η t + e it (2) All it = γ 0 + γ 1 D ct + µ i + π t + ε it (3) Split it in equation (2) counts how many new households are formed between time t and t + 1 12

by splitting and moving. In equation (3), All it is a migration dummy, indicating whether the entire household i migrates to a new location. The above empirical analysis enables us to measure the total effects of disasters on these two forms of migration, making up the first part of the analysis. The second part explains through which channels disasters operate, to bring about such effects. To do this, I modify equations (2) and (3) to include controls for different economic variables, as shown in equations (4) and (5). Split it = β 0 + β 1D ct + β 2Y it + δ i + η t + e it (4) All it = γ 0 + γ 1D ct + γ 2Y it + µ i + π t + ε it (5) Y it, consists of a list of economic variables, which includes household size, total earnings, external transfer and household assets, of household i at time t. By comparing the coefficients on disaster variables, D ct, in equations (2) and (4), and also the coefficients on economic variables, Y it in equation (4), we can tell through which economic channels disasters operate to affect split migration. We can also use the same approach to discover the channels for whole-household migration. 6 Results Table 3 presents the results of the linear probability model. The dependent variable is the household-migration dummy between time t and t + 1, combining both split-household migration and moving of an entire household. The explanatory variables are the annual average number of earthquakes, eruptions and floods, happening between time t 1 and t. All specifications allow for clustering of standard errors at the province-time level. The first three columns do not control for household fixed effects, which give the cross-household analysis. The results show that the probability for households to move out goes down with more disasters. Furthermore, floods significantly drive down household moving across provinces and kabupatens (districts). The effect of eruptions on all three geographic levels of migration is negatively significant at the 0.01 level. With respect to eruptions, the probability of moving to another province 13

falls by 0.024. Given the overall migration rate across provinces as 0.06, an additional eruption annually drives down cross-province migration by 40%. By the similar token, one more flood each year leads to a fall of 29% in cross-province migration. Yet as suggested in Section 5, simple OLS cannot account for unobserved household-migration propensity. From now on, I control for household fixed effects in all specifications to address this possible endogeneity. In columns (4) to (6), the results present an even more negative impact of disasters. Besides all the coefficients being negative, the effect of earthquakes is much greater for all geographic levels of migration. Time variation of all three types of disasters does significantly lower household migration. Earthquakes cause cross-province migration to fall by 0.024, which amonts to 134%. Similarly, an annual additional eruption and flood also reduce cross-province migration by 18% and 24% respectively, even though the impacts of eruptions are not significant. I now separately consider the two different forms of household moving. Columns (1) to (3) of Table 4, list the results for split-household migration. The dependent variable counts the number of new households formed by splitting and moving to a new location. Columns (7) to (9) list the results for whole-household migration. The dependent variable is a dummy indicating whether the entire household relocates to a new residence. The main analysis uses a count variable for splithousehold migration and dummy for whole-household migration. To enhance comparability, I also include columns (4) to (6), which use a split-migration dummy as the dependent variable. Table 4 shows a clear difference between the two forms of migration. Earthquakes significantly reduce split-migration at all geographic levels. Splits to a new province fall by 0.068, which is 120% in percentage terms. Eruptions also decrease cross-province splits by 31%. However, the effects of floods are not statistically significant except for splits across provinces. On whole-household migration, earthquakes do not significantly reduce the moving of the entire households, as shown in columns (7) to (9). These findings contrast with the results of split migration. Yet eruptions cause an entire household to move out less. Cross-district migration falls by 32%. Furthermore, household moving decreases significantly at all geographic levels when one more flood occurs each year, with cross-province migration decreasing by 64%. Table 4 shows that floods do not significantly reduce split-household migration, but reduces 14

whole-household migration at all geographic levels. Earthquakes lower splits at all levels but have no effect on moving of the entire households. Eruptions cause both forms of household migration to fall. The negative impacts of disasters do not just apply to household moving, but also migration at the individual level. From Table 5, all types of disasters decrease individual migration even though the effects of floods are not significant. Earthquakes reduce cross-province migration by 121%, or 12% for an additional earthquake in every 10 years. Similarly, when one more eruption takes place, cross-province migration decreases by 15%. Thus, the analysis on migration at both the household and individual level shows that disasters make people move out less. From now on, the paper will shift the focus back to household migration because I will explain how disasters operate through economic variables to shape migration. The datasets provide economic variables at the household level rather than at the individual level. To understand why disasters lower migration, it is necessary to first examine how different disasters affect a variety of economic variables. This can be done by running an auxiliary regression on the following equation. Regression on equation (6) tells us the impacts of disasters on different economic variables of household i at time t controlling for household fixed effects, ϑ i, and time dummies, ν t. Y it = λ 0 + λ 1 D ct + ϑ i + ν t + e it (6) Table 6 lists all the economic variables, which are measured in natural logs of real values except household size. The stock variables include household size, non-business assets, farm assets and nonfarm-business assets, which are recorded at time t. Non-business assets are further categorized into land holdings, housing and financial assets. On the other hand, the flow variables include total household earnings, remittances and financial aid received within one year before time t. It would be ideal to have the average annual measures of flow variables between time t 1 and t, which is not feasible due to data limitations. Table 6 shows that earthquakes significantly lower economic well-being on various measures. An additional earthquake each year reduces household size by 0.35. Earthquakes also slash non-business 15

assets by 69%. Financial asset, a category of non-business asset, declines by 79% when one more earthquake takes place annually. This implies that households may drain financial resources to cope with earthquakes. Earthquakes damage housing assets, decreasing the values by 14% if one more earthquake happens in every 10 years. Households also suffer from losses in farm and nonfarmbusiness assets but the effects are not significant. One more earthquake in every decade also lowers total household earnings by 13%. One possible explanation is the worsening of macroeconomic conditions or destruction of factories, which may reduce the employment prospects. On the other hand, remittances and aid received do not go up significantly with more earthquakes. While earthquakes have negative impacts on household economic status, eruptions increase different measures of economic variables. An additional eruption raises the amount of farm assets by 55%. Lava ash in eruptions can highly enrich soil fertility which plausibly increases the value of farm assets. Eruptions also increase housing assets, which can be due to the fact that relief money runs into affected areas for house rebuilding and consequently helps boost the housing market. Furthermore, households receive significantly more remittances with the rise of 48%. Such significant increase, however, is not observed for earthquakes and floods. One possible explanation is that the impacts of eruptions may only be geographically confined to the areas near volcanoes. Hence, most households in the province are largely unaffected and they are still financially intact to remit money to affected households. However, the damage of floods and earthquakes can be much more far reaching, adversely affecting most households in the province. Earthquakes and floods may constitute aggregate shocks, causing households to not receive more financial support, as nonhousehold members are also financially impaired. To recap, earthquakes reduce non-business assets and specifically, the values of financial and housing assets fall. Total household earnings and household size also decrease with more earthquakes. Eruptions raise farm assets and the amount of remittances received. Floods, in general, do not affect any measure of household economic well-being. Given the results of Tables 4 and 6, we can now explore the channels through which disasters operate to affect the two different forms of household migration. Table 7 presents the findings for split-household migration. I put the regression results without 16

controls and with controls for economic variables, side by side. By including controls for economic variables, the magnitude of coefficients on earthquakes has dropped for all three geographic levels of moving. From column (1), earthquakes reduce household splits to a new province by 0.068 (120% in percentage terms), but the magnitude falls to 0.058 (102%) after adding economic variables as shown in column (4). The drop in magnitude is even more noticeable for splits to a new kabupaten (district). Furthermore, the coefficients on split migration to kecamatan (subdistrict) are no longer significant after adding controls. This suggests that earthquakes operate through some of the included economic variables to reduce split migration. Table 7 also shows that household size and total earnings are significant positive factors for split migration. An additional household member increases cross-province splits by 42%. One percentage increase in household earnings also raises the number of new household formed in a new province by 0.00085, which amounts to an elasticity of 1.5%. We know from Table 6, that earthquakes significantly reduce household size and total earnings. Combining these findings, I conclude that earthquakes decrease household earnings and household size to drive down split migration. However, the findings on non-business assets do not give us a clear conclusion. Table 7 tells us that non-business assets do not significantly increase split migration and that the coefficient of crosssubdistrict splits is even negative. However, I also consider farm and nonfarm business assets and both types of business assets significantly raise split migration. As shown in table 6, earthquakes lower the two types of business assets, although insignificantly. Thus, the results suggest that earthquakes decrease split migration through reducing business and non-business assets. Table 7 presents different findings for eruptions. All the negative signs remain and the coefficients are even more negative after controlling for economic variables. Households with more farm assets split more, and farm assets rise with eruptions, which explains why eruptions cause more household splits and the coefficients on eruptions in columns (4) to (6) of Table 7 are even more negative. Hence, I reject all the economic variables listed in Table 6, as the channels through which eruptions operate to suppress household splits. We now shift our focus to whole-household migration. Table 8 shows how disasters and economic variables affect the moving of an entire household. We need only to consider the effects of eruptions 17

and floods because earthquakes are not significant in affecting whole-household migration. After adding economic variables, there is a substantial drop in magnitude for the coefficients on eruptions. Coefficients for cross-district moving falls from -0.0094 to -0.0073, and including controls completely eliminates the significant impacts on migration across subdistricts. Households with more farm assets are less mobile to move as a whole. The likelihood of moving to a new district drops by 4% when farm business asset goes up by 1%. Eruptions raise the amount of farm assets as shown in Table 6, which explains why eruptions drive down whole-household migration. Table 8 shows that the magnitude and significance of coefficients on floods do not change substantially, which implies that the suggested economic variables are not the channels through which floods operate to reduce migration. From Table 6, floods do not cause significant impacts on any of the economic variables. Thus, I conclude that the reduction of whole-household moving because of floods is not related to household asset or earnings. Tables 7 and 8, together, show some contrasting impacts of economic variables on split migration and moving of an entire household. The size of the household has totally opposite effects on these two forms of moving. Households with more members split and migrate more, but are less likely to move out as a whole. Similarly, more assets enable households to split and move to a new location, but reduce migration of an entire household. These results are in line with the discussion in Section 4. Households with more assets have greater ability to support splits. However, most forms of assets, such as land and house, are illiquid, accumulating assets makes the entire household more rooted in its village and less mobile to move out. As a summary, when earthquakes, eruptions and floods occur, households move out less in the following period. But after breaking down the analysis into two different forms of migration, we observe that earthquakes only reduce household splits, and floods have negative impacts only on migration of an entire household. Eruptions drive down both forms of migration. For the channels of impacts, earthquakes lower household splits through decreasing household size, household earnings and non-business assets. On the other hand, eruptions increase farm assets and consequently make households move out less as a whole. Reduction of whole-household migration due to floods is not related to any of the suggested economic variables. 18

I conduct a simple back-of-the-envelope calculation to quantitatively assess the impacts of disasters on household migration through economic variables. From Tables 6 and 7, earthquakes decrease household size by 0.35, and an additional household member drives up splits across provinces by 0.024. Thus, earthquakes reduce cross-province splits by 0.0085 (0.35*0.024), which amounts to 15%. Using the similar method, earthquakes lower earnings to decrease cross-province splits by 0.0011 or 1.9%. For whole-household migration, eruptions increase farm assets by 55% and consequently drives down moving of an entire household across provinces by 0.00022 (0.55*0.00041), or 2.1%. We can also tell to what extent the suggested economic variables explain the negative impacts of disasters on household migration. From Table 7, the coefficient on earthquakes for cross-province migration drops from 0.068 to 0.058, which is a 15% fall. Thus, 15% of the negative impacts of earthquakes is explained by economic variables. Similarly, economic variables explain the 25% and 36% decline in cross-district and cross-subdistrict splits respectively. We use the same method to explain the decrease in whole-household migration due to eruptions. According also to Table 8, including economic variables explain 22% and 17% of the cross-district and cross-subdistrict moving of the entire households, respectively. 7 Robustness checks First, to affirm the negative impacts of disasters on the two forms of migration, the study takes a placebo test on the migration data. The analysis alters the time interval for the disaster variables. Instead of using the yearly average number of disasters within the immediate last period, I push the time period 14 years backward to set up a placebo time frame. For instance, the regression of migration between 1997 and 2000, the time period for disaster variables, is from 1983 to 1986. Hence, the specification uses the number of disasters in the placebo time frame and checks whether disasters in that period have any effects on the two forms of migration. Table 9 shows that the coefficients on disasters in the placebo time frame are mostly insignificant. Earthquakes have only barely significant effects on split migration at the district level and wholehousehold moving at province level. Floods are just marginally significant in affecting cross-province 19

splits. Hence, the placebo test affirms the negative relationship between migration and disasters within the immediate last period. The surveys are not conducted at a regular time interval and there is a seven-year gap between the last two waves, IFLS3 (2000) and IFLS4 (2007). This time period is so long that the effects of disasters in the previous period (1997-2000) have substantially diminished well before 2007. Furthermore, a huge tsunami happened in the province of Aceh in 2004, and resulted in massive fatalities. Although the samples do not include any households from Aceh, the tsunami could have forced Acehnese households to relocate to neighboring provinces, which may cloud the estimates. To address this problem, I set a cut-off point in year 2004, and discarded all the sample households which moved after 2004. Only households moving before 2004 are considered migrants. Tables 10 and 11 present the results of the revised specification. With respect to split migration, most of the negative coefficients still remain, but the magnitude and significance drop. Earthquakes still primarily reduce split migration. The number of cross-province splits decrease by 91%. An additional eruption also causes significantly less splits to districts and subdistricts. Furthermore, the conclusions drawn in Section 6 still stand. The coefficients on earthquakes fall in magnitude after adding economic variables. Furthermore, the size of household and total earnings are still significant to raise household splits. Household assets also have marginally significant impacts on increasing splits. Thus, earthquakes suppress household splits by reducing household size, earnings and assets. Such results are similar to the findings in Table 7. Regarding whole-household migration, eruption is no longer a significant negative factor at all, after controlling for economic variables. The coefficients either become less negative or even positive. Following the results that more farm assets lower whole-household moving, we can conclude that eruptions reduce the likelihood of migration by increasing farm assets. Tables 10 and 11 show us some contrasting results which are also observed in Tables 7 and 8. The size of household, on the one hand, increases household splits, but on the other hand, it suppresses the migration of an entire household. More assets of different kinds enhance the likelihood of household splits, but at the same time lowers the likelihood for an entire household to move out. All the above specifications use annual average number of disasters as the explanatory variables. 20

However, number by itself cannot fully gauge the severity of disasters. A single massive deadly catastrophe has far greater effects than a series of small-scale disasters of mild intensity. Hence, I use other disaster measures in the specification, which include number of deaths, injuries, people missing and houses destroyed. These variables count the average annual number of respective losses at the province level between time t 1 and t. The list also includes the logged value of financial losses and tonnes of crop damage due to disasters in the last period. Table 12 shows some mixed findings. On the front of human losses, earthquakes and eruptions are just marginally significant to reduce the two forms of household migration. More deaths due to floods raise split migration and whole-household moving, but more injuries from floods make an entire household less likely to migrate. The number of missing people caused by floods is another important factor for lowering both forms of household migration. For economic losses, the effects of disasters on migration are mostly negative. Households residing in the province with more houses destroyed by earthquakes, are significantly less likely to migrate. Similarly, when floods damage more houses in a province, households are less likely to relocate. Financial losses and crop damage by floods also lower split-household migration. 7.1 Extension: Heterogeneous effects of disasters on household migration The main analysis in Section 6 tells us how disasters affect household migration on average. Yet when disasters happen, different households can make different migration decisions, depending on their economic status at time t (Y it ). To empirically analyze the heterogeneous impacts of disasters, I add some interaction terms between disasters and economic variables to equations (4) and (5). Split it = κ 0 + κ 1 D ct + κ 2 Y it + κ 3 Y it D ct + ψ i + ν t + ω it (7) All it = ϕ 0 + ϕ 1 D ct + ϕ 2 Y it + ϕ 3 Y it D ct + χ i + ϖ t + ξ it (8) The coefficients on the interaction term, κ 3 and ϕ 3 denote how disasters in the previous period interact with economic variables at time t to shape the household-migration decision in the next 21

period. A positive significant coefficient implies that households with higher values of economic variables are more likely to move out in the following period after disasters. The disaster and economic variables in the interaction terms are first grand-mean centered such that the results are comparable to the main results in Tables 7 and 8. Table 13 presents the findings. In general, the heterogeneous impacts are minimal and households with different economic status do not have significantly different migration responses. There exist negative significant impacts for eruptions interacting with household receipt of aid. Given the average number of eruptions, split migration to a new district goes down by 0.0034 with a percentage increase of aid received. On the other hand, the probability for an entire household to move to a new district falls by 0.0011. Furthermore, floods interacting with non-business assets lead to contrasting impacts on the two forms of migration. Households with more non-business assets will split and move out more, but are less likely to move out as a whole. Households can rely on non-business assets to finance splits when floods occur. Yet non-business assets also act as a buffer against the environmental shock, which lowers the need for an entire household to relocate. However, the above-mentioned effects are just barely significant statistically and we can conclude that disasters do not cause substantially different responses in migration for households with different levels of economic well-being. 8 Conclusion Using Indonesia as the case country, this study examines whether natural disasters lead to more migration. It discovers that more disasters actually result in less migration. The three most common types of disasters in Indonesia, earthquakes, volcanic eruptions and floods, all lower household and individual migration. Regarding household migration, the paper separately considers split migration and whole-household migration, and finds that disasters have negative impacts on both. Specifically, earthquakes reduce migration primarily through suppressing household splits, and floods drive down whole-household migration. Eruptions lower both forms of migration at all geographic levels. The above results invalidate the claim that disasters cause more migration. The paper then moves on to explain this negative relationship. For split migration, earthquakes significantly reduce 22

household size, total earnings and holding of non-business assets. Smaller households are less likely to split, as are the households with less earnings and non-business assets, which explains why earthquakes cause less split migration. For whole-household migration, eruptions increase the values of farm business assets possibly by enhancing soil fertility through lava ash. Evidence shows that households with more farm assets are less mobile to move out as a whole, which explains why eruptions lower whole-household migration. Finally, the reductions of whole-household migration due to floods cannot be traced to household assets or earnings. I also quantitatively assess the explanatory power of the economic variables for the negative impacts of disasters. For earthquakes, the economic variables explain 15% of the fall in crossprovince splits. The economic channels can also account for 25% and 36% of the decline of crossdistrict and cross-subdistrict splits, respectively. In case of eruptions, economic variables explain 22% and 17% of the reduction of cross-district and cross-subdistrict migration of an entire household, respectively. This paper shows that the claim of more migration after natural disasters is not valid for Indonesia. The hypothesis of the claim ignores two important facts: (1) disasters can alter household economic well-being, which may consequently lower people s propensity to migrate as described in the study; and (2) given the regular occurrence of disasters, households may resort to a variety of adaptation mechanisms instead of simply moving out of disaster-prone areas (IOM 2009). However, after adding economic variables in the regression, the negative coefficients still remain, and the significance has not been fully eliminated. In terms of the effects of eruptions on household splits, the magnitude of the coefficients goes up. Indonesian people develop their communities near volcanic areas which may give rise to a positive correlation between eruptions and population density. This reason, however, cannot explain the findings given the empirical specification of this paper. The regression controls for household fixed effects, hence, the coefficients measure how the variation of the eruptions across time alters the household migration pattern. Increasing eruptions should not induce households to stay. Furthermore, none of the suggested economic variables can explain how floods drive down whole-household migration. Thus, the most plausible explanation is that the specification has not captured some other 23

variables through which disasters operate to affect migration. Since the regression has controlled for time-invariant household fixed effects, those other possible variables should be time varying which may include degree of risk aversion, health status of household heads, accumulation of social capital and other social factors as described in Section 4. The negative causal relationship between disasters and household migration warrants further research to better study how households in developing countries determine migration decisions in our time of surging environmental calamities. 24

References [1] Ahmad, Z. (2011), Impact of Alluvial Deposits on Soil Fertility during the Floods of 2010 in Punjab, Pakistan, Research Findings, International Potash Institute. [2] Attzs, Marlene (2008), Natural Disasters and Remittances: Exploring the Linkages between Poverty, Gender, and Disaster Vulnerability in Caribbean SIDS, WIDER Research Paper, UNU-Wider. [3] Cameron, Lisa and Manisha Shah (2010), Risk Taking Behavior in the Wake of Natural Disasters, Working Paper. [4] Cavallo, Eduardo and Ilan Noy (2009), The Economics of Natural Disasters: A Survey, IDB Working Paper Series, No. IDB-WP-124. [5] Cavallo, Eduardo, Sebastian Galiani, Ilan Noy and Juan Pantano (2010), Catastrophic Natural Disasters and Economic Growth, IDB Working Paper Series, No. IDB-WP- 183. [6] DesInventar, http://dibi.bnpb.go.id/desinventar/main.jsp?countrycode=id&lang=en [7] Ebeke, Christian and Jean-Louis Combes (2010), Do remittances dampen the effect of natural disasters on output growth volatility in developing countries?, Working Paper Series, No.: 201031, CERDI [8] Halliday, Timothy J. (2007), Migration, Risk and the Intra-Household Allocation of Labor in El Salvador, Working Paper. [9] Hsiao, Cheng (1986), Analysis of Panel Data, Second Edition, Chapter 4. [10] International Organization of Migration (2009), Migration, Environment and Climate Change: Assessing the evidence, IOM publication. [11] Naude, Wim. (2008), Conflict, Disasters, and No Jobs: Reasons for International Migration from Sub-Saharan Africa, Research Paper No. 2008/85, UNU-Wider. 25

[12] New York Times (Nov 02, 2010), http://www.nytimes.com/2010/11/02/world/asia/02indo.html [13] Noy, Ilan and Tam-Bang Vu (2010), The economics of natural disasters in a developing country: The case of Vietnam, Working Paper series 200903, University of Hawaii at Manoa, Department of Economics. [14] Ó Gráda, Cormac (1997), The Great Irish Famine : Winners and Losers, No 97-23, Discussion Papers from University of Copenhagen, Department of Economics. [15] SEDAC - Gridded population of the World, http://sedac.ciesin.columbia.edu/ [16] UN International Strategy for Disaster Risk Reduction, (2007), World experts unite to confront growing risks of disasters, Press release UN/ISR 2007/8, Geneva. [17] US Federal Emergency Management Agency (2011), http://www.columbiamissourian.com/stories/2011/01/01/2010s-world-gone-wildquakes-floods-blizzards/ [18] Wooldridge, Jeffrey (2001), Econometric Analysis of Cross Section and Panel Data, MIT press. [19] Yang, Dean (2008) Risk, Migration, and Rural Financial Markets: Evidence from Earthquakes in El Salvador, Social Research Vol 75 : No 3. 26

Fig. 1: Yearly occurrence of earthquakes and floods in Indonesia no. of earthquakes 12 no. of floods 50 45 10 40 8 6 35 30 25 Eathquake Flood 4 20 15 2 10 5 0 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 0 Source: DesInventar Database 27

Fig. 2: Spatial variation of number of earthquakes, 1988-2000 Fig. 3: Spatial variation of number of eruptions, 1988-2000 Source: DesInventar Database 28

Fig 4: Spatial variation of number of floods, 1988-2000 Source: DesInventar Database Fig. 5 Population density of Indonesia in 2000 Source: Gridded Population of the World (GPWv3) Socio-Economic Data and Application Center 29