77545 Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines Dean Yang and HwaJung Choi Do remittances sent by overseas migrants serve as insurance for recipient households? In a study of how remittances from overseas respond to income shocks experienced by Philippine households, changes in income are found to lead to changes in remittances in the opposite direction, consistent with an insurance motivation. Roughly 60 percent of declines in household income are replaced by remittance inflows from overseas. Because household income and remittances are jointly determined, rainfall shocks are used as instrumental variables for income changes. The hypothesis cannot be rejected that consumption in households with migrant members is unchanged in response to income shocks, whereas consumption responds strongly to income shocks in households without migrants. JEL codes: D81, F22, F32, O12, O15 Several facts motivate this study. First, life in developing countries is prone to many kinds of risk, such as crop and income loss due to natural disasters (weather, insect infestations, �re) and civil conflict. Second, international migration and remittance flows are substantial and growing. Between 1965 and 2000, individuals living outside their country of birth grew from 2.2 to 2.9 percent of the world population, totaling 175 million people in 2000.1 The remittances that these migrants send to their countries of origin are an impor- tant but poorly understood type of international �nancial flow. In 2002, remit- tance receipts of developing countries totaled $79 billion.2 This amount Dean Yang (corresponding author) is an assistant professor at the Gerald R. Ford School of Public Policy and in the Department of Economics at the University of Michigan; his email address is deanyang@umich.edu. HwaJung Choi is a PhD candidate in economics at the University of Michigan; her email address is hwajungc@umich.edu. The authors are grateful for comments and suggestions from Becky Blank, John Bound, Jaime De Melo, Claudio Gonzalez-Vega, Caglar Ozden, Maurice Schiff, three anonymous referees, and seminar participants at Ohio State University’s Rural Finance Program. The World Bank’s International Migration and Development Research Program provided important research support. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. 1. Estimates of the number of individuals living outside their country of birth are from United Nations (2002), whereas data on world population are from United States Bureau of the Census (2002). 2. The remittance �gure is the sum of the “workers’ remittances,� “compensation of employees,� and “migrants’ transfers� items in the International Monetary Fund’s International Financial Statistics database for all countries not listed as high income in the World Bank’s income groupings. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 2, pp. 219 –248 doi:10.1093/wber/lhm003 Advance Access Publication 9 May 2007 # The Author 2007. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 219 220 THE WORLD BANK ECONOMIC REVIEW exceeded the total of�cial development aid ($51 billion) and equaled roughly 40 percent of foreign direct investment (FDI) inflows ($189 billion) received by developing countries that year.3 Understanding the functions of remittances for recipient households is necessary for weighing the bene�ts to origin countries of developed country policies liberalizing inward migration [as proposed by Rodrik (2002) and Bhagwati (2003), for example].4 What connection, if any, is there between the pervasiveness of risk in deve- loping countries and international remittance flows? Do remittances from over- seas migrants serve as insurance for relatives back home? To shed light on this question, this article examines how income shocks experienced by households in the Philippines affect their receipt of remittances from overseas. To break the simultaneity between income and remittances, rainfall shocks are used to instrument for changes in household income. In households with members who are overseas migrants, changes in income from domestic sources lead to changes in remittances in the opposite direction of the income change: remit- tances fall when income rises and remittances rise when income falls. In such households, the amount of insurance is large: roughly 60 percent of exogenous declines in income are replaced by remittance inflows from overseas. In con- trast, in households without overseas migrants, changes in income from dom- estic sources have no effect on remittance receipts. As a result, the hypothesis cannot be rejected that consumption in households with migrant members is unchanged in response to income shocks, whereas consumption responds strongly to income shocks in households without migrants. Numerous studies have examined the mechanisms through which house- holds in developing countries cope with risk. Among others, Townsend (1994), Udry (1994), Ligon, Thomas and Worall (2002), and Fafchamps and Lund (2003) have documented risk-pooling arrangements among households intended to smooth consumption in response to shocks. Households may also autonomously build up savings or other assets in good times and draw down these assets in hard times (Paxson 1992; Rosenzweig and Wolpin 1993; Udry 1994), increase their labor supply when shocks occur (Kochar 1999), or take steps (such as crop and plot diversi�cation) to reduce the variation in their incomes (Morduch 1993). This article examines a mechanism for coping with shocks ex post on which previous micro-level studies have not focused: remittances from family members overseas. At the international level, it is commonly posited that remit- tance flows from overseas buffer economic shocks in migrants’ home countries (e.g., Ratha 2003), but there have been relatively few empirical tests of this 3. Aid and FDI �gures are from World Bank (2004). Although the �gures for of�cial development aid and FDI are likely to be accurate, by most accounts national statistics on remittance receipts are considerably underreported (see, e.g., Ratha 2003), so the remittance �gure may be taken as a lower bound. 4. Borjas (1999) argues that the investigation of bene�ts accruing to migrants’ source countries is an important and virtually unexplored area in research on migration. Yang and Choi 221 claim with micro-level household data.5 Related research on the role of domes- tic migration in pooling risk within extended families includes Lucas and Stark (1985), Rosenzweig and Stark (1989), and Paulson (2000). A key distinguishing facet of this article is its emphasis on credible identi�- cation of the effect of income shocks on international remittances. Studies of the impact of household income on remittance receipts use cross-sectional data, and so are subject to potentially severe biases in directions that are not obvious a priori. Reverse causation is a major concern: productive investments funded by migrant remittances can raise household income, leading to positive correlations between household income and remittances. Alternately, remit- tances may reduce households’ need to �nd alternative income sources, leading to a negative relationship between remittances and domestic-source income. Even if reverse causation from remittances to income in migrants’ source households were not a problem, it would be dif�cult to separate the cross- sectional relationship between income and remittances from the influence of unobserved third factors affecting both income and remittances (e.g., the entre- preneurial spirit of household members). Two aspects of the empirical strategy are the key in resolving these identi�- cation problems. First, the focus is on income changes due to shocks that are credibly exogenous—changes in local rainfall—so that bias due to reverse cau- sation is not a concern.6 But the estimated impact of economic shocks in cross- sectional data is still likely to be biased, because the likelihood of experiencing a shock may be correlated with time-invariant household characteristics (in other words, omitted variables are still a concern). For example, if shocks occur more frequently in poorer areas, and more remittances generally flow to poor areas, estimates of the impact of income on remittances will be biased in a negative direction. So the second crucial aspect of this article is its use of panel data, so that estimates of the impact of income shocks can be purged of the influence of unobserved time-invariant household characteristics that are jointly related to remittances and to the likelihood of experiencing a shock. Estimation of the impact of shocks focuses on how shocks are related to changes in remittances rather than on the level of remittances. Section I considers the theoretical role of international remittance flows in sharing risk across family members in different countries. Section II describes the data used and provides empirical results. Section III discusses some of the policy implications of the �ndings and recommendations for future research. Further details on the household data sets are provided in the supplemental appendix, available at http://wber.oxfordjournals.org/. 5. On the international macroeconomic level, Yang (2006a) documents that international �nancial flows (including remittances) at the country level respond positively to economic losses due to hurricanes. For studies with micro-data, see Brown (1997) and Gubert (2002). 6. Other research using rainfall shocks as instruments includes Paxson (1992), Munshi (2003), Miguel (2005), and Maccini and Yang (2006). 222 THE WORLD BANK ECONOMIC REVIEW I. INCOME SHOCKS AND R E M I T TA N C E S IN THEORY When a household experiences a negative income shock, how should we expect remittance receipts from overseas to change? A basic theoretical result is that if there is a Pareto-ef�cient allocation of risk across individual entities (in this case, individual household members) in a risk-sharing arrangement, individual consumption should not be affected by idiosyncratic income shocks. Consider households consisting of two members, indexed by i [ f1,2g. Let one household member be located in the origin household in the Philippines and the other household member be located overseas. Assume that both house- hold members work and are able to send funds back and forth to each other. i Individuals have an uncertain income in each period t, ys t , depending on i the state of nature st [ S. Household member i consumes cst, and experiences i within-period utility of Ui (cst ) at time t. Let utility be separable over time, and 00 let instantaneous utility be twice differentiable with Ui0 . 0 and U i , 0. For the allocation of risk across household members to be Pareto-ef�cient, the ratio of marginal utilities between members in any state of nature must be equal to a constant: 0 1 U1 ðcst Þ v2 ð1Þ 0 ¼ ; for all st and t; U2 ðc2 st Þ v1 where v1 and v2 are the Pareto weights of members 1 and 2. Household members’ marginal utilities are proportional to each other, and so consumption levels between members move in tandem. Let utility be given by the following constant absolute risk aversion func- tion: i ÀeÀucst ð2Þ Ui ðcist Þ ¼ : u Then, following Mace (1991), Cochrane (1991), Altonji, Hayashi, and Kotlikoff (1992), and Townsend (1994), a relationship between individual household member i’s consumption and average consumption across the house- hold members, c¯ st, is obtained by: ln vi À 1=2ðln v1 þ ln v2 Þ ð3Þ cist ¼  c st þ u Ef�cient risk sharing implies that an individual’s consumption level depends here only on mean consumption in the household, c ¯ st, and an effect determined by the individual’s Pareto weight relative to the other’s. Because this latter term is constant over time, changes in consumption for each individual will depend Yang and Choi 223 only on the change in mean household consumption. Said another way, indi- viduals face only household-level risk. How might this within-household (but cross-country) risk sharing be carried out in practice? It is simple to imagine that an individual sends remittances to the other household member when that member experiences a negative shock.7 Adapting Fafchamps and Lund (2003), let consumption of individual i in i i state st be the sum of income ys t and net inflows of remittances rst : cist ¼ yist þ rist : So then equation (3) can be rewritten as: ln vi À 1=2ðln v1 þ ln v2 Þ ð4Þ rist ¼ Àyist þ  c st þ : u This equation can be transformed into an empirically testable speci�cation as i follows. First, separate income yst into: ð5Þ yi þ ziSt ; yiSi ¼ ~ where y ˜ i is the permanent component of income and zs i t is the transitory com- ponent of income. Only the transitory component depends on the state of the world. The function of Pareto weights and the permanent income component y ˜ i can be captured by an individual �xed effect, gi. The mean household consumption level, c ¯ st, can be represented by a time effect, ft. Also, allow a random com- ponent, 1it, a mean-zero error term. Then equation (4) becomes: ð6Þ rist ¼ ÀZist þ gi þ ft þ 1it : The empirical test of this article is based on equation (6), where the outcome variable is remittances received from overseas. The focus is on a par- i ticular type of transitory shock, zs t , changes in income from domestic (Philippine) sources, using rainfall shocks as the instrumental variable. There are two key questions of interest. First, is the coef�cient on remit- i tances with respect to domestic income zs t less than zero? If yes, then this is the evidence that at least some insurance is taking place. Second, can the null i hypothesis of full insurance be rejected, that is, that the coef�cient on zs t is equal to negative one? 7. Micro-economic studies among households of the insurance role of gifts and remittances include Lucas and Stark (1985), Ravallion and Dearden (1988), Rosenzweig and Stark (1989), Platteau (1991), and Cox, Eser, and Jimenez (1998). 224 THE WORLD BANK ECONOMIC REVIEW I I. E MP IRICAL A N A LY S I S This section �rst describes the data and sample constructions and provides descriptive statistics on the sample households. It then discusses the regression speci�cation and some empirical issues and presents empirical results. Finally, tests are conducted of potential violations of the instrumental variable exclu- sion restriction and of an important omitted variable concern. Data and Sample Construction The empirical analysis uses data from linked household surveys conducted by the Philippine National Statistics Of�ce covering a nationally representative household sample: the Labor Force Survey, the Survey on Overseas Filipinos, the Family Income and Expenditure Survey, and the Annual Poverty Indicators Survey. The Labor Force Survey is administered quarterly to inhabitants of a rota- ting panel of dwellings in January, April, July, and October. The other three surveys are administered with lower frequency as riders to the Labor Force Survey. Usually, one-fourth of dwellings are rotated out of the sample in each quarter, but the rotation was postponed for �ve quarters starting in July 1997, so that three-quarters of dwellings included in the July 1997 round were still in the sample in October 1998 (one-fourth of the dwellings had just been rotated out of the sample). The analysis here takes advantage of this postponement of the rotation schedule to examine changes in households over the 15-month period from July 1997 to October 1998. Survey enumerators note whether the household currently living in the dwelling is the same as the household surveyed in the previous round; only dwellings inhabited continuously by the same household from July 1997 to October 1998 are included in the sample for analysis. Because the impact of domestic income shocks on remittance receipts is likely to vary according to whether households had migrant members, households that reported having one or more members overseas in June 1997 and households that did not are analyzed separately. The comparison between migrant and non-migrant house- holds should be taken as merely suggestive because households with and without migrants differ in many ways. Thus, while differences in results between these two groups of households could be due to whether they have migrants, it could also be due to other factors such as differential access to other risk-coping mechanisms (savings, credit, informal and formal insurance). Rainfall data used in constructing instrumental variables for household domestic income were obtained from the Philippine Atmospheric, Geophysical, and Astronomical Services Administration. Daily rainfall data are available for 47 weather stations, often as far back as 1951. Rainfall variables are con- structed by station separately for the two distinct weather seasons in the Philippines: the dry season from December through May and the wet season from June through November. Monthly rainfall is calculated by summing daily Yang and Choi 225 rainfall totals, with missing daily values replaced by the average daily totals in the given station-month for stations that had 20 or more daily rainfall records. For station-months with less than 20 daily rainfall records, monthly rainfall for the station is taken to be the monthly rainfall recorded in the nearest station with 20 or more daily rainfall records. Seasonal total rainfall for each station in each year is obtained by summing monthly rainfall for the months in each wet and dry season. Rainfall shock variables for a given season are constructed as rainfall in that season (in thousands of millimeters) minus rainfall in the same season the year before. Households are assigned the rainfall data for the weather station geo- graphically closest to their local area (speci�cally, the major city or town in their survey domain), using great circle distances calculated using latitude and longitude coordinates. Because some stations are never the closest station to a particular survey domain, a total of 38 stations are represented in the empirical analysis. The supplemental appendix provides other details on the household surveys and the construction of the sample for analysis (available at http://wber.oxford- journals.org/). Characteristics of Sample Households Table 1 presents summary statistics for the 27,881 households used in the empirical analysis, separately for migrant and non-migrant households. Migrant households are those with overseas workers in June 1997. The 1,655 migrant households represent 5.9 percent of the sample households. The table also presents rainfall data to provide a sense of the instrumental variables used. The rainfall data are deviations (in thousands of millimeters) from the historical mean of each station for the dry and wet seasons. The dry season immediately before the �rst observation for each household (year 1, with income from January to June 1997) runs from December 1995 to May 1996, and the wet season is from June to November 1996. Correspondingly, the dry season for the second observation for each household (year 2, with income from April to September 1998) is December 1996 to May 1997, and the wet season is from June to November 1997. For migrant households, the dry season in year 1 was on average wetter than normal, with a mean deviation across households of 0.28. In year 2, dry season rainfall was more typical, with a mean deviation of 0.03. Therefore, the mean household experienced a decline in dry season rainfall between years 1 and 2: the mean change in the dry season deviation across households is 2 0.25. The wet season for year 1 was only slightly wetter than normal, with a mean devi- ation across households of 0.07. In year 2, wet season rainfall was much dryer than normal, with a mean of 2 0.48. The mean household thus experienced a substantial decline in wet season rainfall between years 1 and 2 of 2 0.55. These declines in rainfall between the 2 years have been attributed to the 1997 El Nin ˜ o weather phenomenon. The mean of the rainfall variables for the T A B L E 1 . Characteristics of Sample Households, 1997 226 Migrant households Non-migrant households Standard 10th 90th Standard 10th 90th Mean deviation percentile Median percentile Mean deviation percentile Median percentile Rainfall variables (thousands of millimeters) Dry season year 1 0.28 0.40 2 0.04 0.10 0.55 0.38 0.50 2 0.03 0.20 0.90 Dry season year 2 0.03 0.21 2 0.17 0.00 0.32 2 0.01 0.23 2 0.33 2 0.01 0.29 Change between dry season 2 0.25 0.56 2 0.79 0.02 0.23 2 0.39 0.66 2 0.93 2 0.28 0.23 years 1 and 2 Wet season year 1 0.07 0.47 2 0.35 2 0.03 1.16 0.01 0.40 2 0.35 2 0.03 0.54 Wet season year 2 2 0.48 0.33 2 0.98 2 0.48 2 0.12 2 0.41 0.31 2 0.85 2 0.42 2 0.05 Change between wet season 2 0.55 0.67 2 2.00 2 0.39 0.18 2 0.42 0.58 2 1.33 2 0.23 0.18 years 1 and 2 THE WORLD BANK ECONOMIC REVIEW Outcome variable Change in household domestic 0.03 0.64 2 0.50 2 0.04 0.57 0.03 1.44 2 0.59 2 0.13 0.68 income as share of initial household income Change in household 0.07 0.67 2 0.50 2 0.02 0.71 0.01 0.22 0.00 0.00 0.00 remittance receipts as share of initial household income Change in household 0.00 0.61 2 0.55 2 0.14 0.67 2 0.04 1.06 2 0.55 2 0.16 0.55 expenditure as share of initial household expenditures Change in indicator for 2 0.38 0.49 2 1.00 0.00 0.00 0.02 0.15 0.00 0.00 0.00 overseas worker in household Household �nancial statistics (January – June 1997) Total expenditure 73,576 66,443 24,568 57,691 132,793 47,437 54,159 13,657 32,493 93,522 Total income 94,189 92,636 28,093 71,012 175,000 56,059 77,676 13,513 35,908 113,460 Total income per capita in 20,204 21,356 5,510 15,206 39,166 11,857 15,116 2,864 7,625 24,100 household Domestic income 58,067 80,815 7,971 38,310 120,317 54,170 75,912 13,076 34,800 109,760 Remittance receipts 36,122 46,752 0 26,000 87,000 1,888 13,182 0 0 0 Remittance receipts as share of 0.39 0.31 0.00 0.37 0.85 0.02 0.10 0.00 0.00 0.00 total income Household size, including 6.2 2.4 3.0 6.0 9.0 5.2 2.3 3.0 5.0 8.0 overseas members, July 1997 Located in urban area 0.68 0.58 Household head characteristics July 1997 Age 49.9 13.9 32.0 50.0 68.0 46.7 14.1 30.0 45.0 67.0 Highest education level (indicators) Less than elementary 0.17 0.28 Elementary 0.20 0.22 Some high school 0.10 0.11 High school 0.22 0.18 Some college 0.16 0.11 College or more 0.14 0.09 Occupation (indicators) Agriculture 0.23 0.38 Professional job 0.08 0.06 Clerical job 0.13 0.11 Service job 0.05 0.07 Production job 0.14 0.26 Other 0.38 0.12 Does not work 0.00 0.00 Marital status is single (indicator) 0.03 0.03 Number of households 1,665 26,126 Note: Rainfall variables are deviations from historical average of each station in corresponding season. For year 1, dry season is December 1995– May 1996 and wet season is June 1996– November 1996. For year 2, dry season is December 1996– May 1997 and wet season is June 199– November 1997. Rainfall data are collected from 38 stations. Total income includes both domestic-source income and remittances from overseas. Remittance receipts are Yang and Choi from overseas only. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. 227 228 THE WORLD BANK ECONOMIC REVIEW non-migrant households are generally quite similar to those for migrant households. The changes between years 1 and 2 for the wet and dry seasons are used as the instrumental variables for the change in domestic income in the empirical analysis that follows. The geographic distribution of the rainfall shocks is depicted graphically in �gures 1 and 2. F I G U R E 1. Dry Season Rainfall Shocks, Philippines Note: Rainfall shocks are change in rainfall (in thousands of millimeters) between the December 1996– May 1996 season and the December 1996– May 1997 season. Each number in the �gure is centered at coordinates of a rainfall station. Source: Authors’ calculations based on data described in the text. Yang and Choi 229 F I G U R E 2. Wet Season Rainfall Shocks, Philippines Note: Rainfall shocks are changes in rainfall (in thousands of millimeters) between the June 1996– November 1996 and the June 1997– November 1997 season. Each number in the �gure is centered at coordinates of a rainfall station. Source: Authors’ calculations based on data described in the text. Total expenditure and total income in the �rst period (January–June 1997) are higher in migrant households than in non-migrant households. Average total expenditure is 73,576 pesos ($2,830) for migrant households and 47,437 pesos ($1,825) for non-migrant households.8 Average total income is 94,189 pesos ($3,623) for migrant households and 56,059 pesos ($2,156) for 8. Peso �gures were converted to US dollars at the January–June 1997 rate of 26 pesos per dollar. 230 THE WORLD BANK ECONOMIC REVIEW non-migrant households. Remittances have a mean of 36,122 pesos ($1,389) in migrant households and 1,888 pesos ($73) in non-migrant households. Remittances amounted to 39 percent of household income for migrant house- holds and 2 percent for non-migrant households.9 Average migrant household size is 6.2 members (including overseas members), whereas non-migrant household size is 5.2 members. Overall, 68 percent of migrant households are located in survey-de�ned urban areas, com- pared with 58 percent of non-migrant households.10 Heads are more educated in migrant households than in non-migrant households: around in 30 percent of heads in migrant households have at least a college degree, compared with only 20 percent in non-migrant households. Fewer household heads worked in agriculture in 1997 in migrant households (23 percent) than in non-migrant households (38 percent). Heads in migrant households are also slightly older (mean age of 49.9) than heads of non-migrant households (46.7). Identi�cation Strategy Overseas remittance responses to exogenous changes in household income from domestic sources are examined to determine whether households use remittances as insurance. The following identi�cation strategy is used. The remittance amount received by each household at time t is determined by household characteristics that are constant over time (such as completed education of household adults), time-variant household characteristics (such as household size), time effects common to all households (such as changes in remittance regulations or the nationwide economic situation), and time-varying household income from domestic sources. In addition, there may be time effects that vary systematically according to household characteristics, as when a nationwide economic shock has differential effects on better-educated and less-educated households. For household h at time t, the remittance equation is as follows: ð7Þ Rht ¼ a þ bYht þ u 0 Xht þ d 0 Wh þ gt þ xt0 ðTt Wh Þ þ 1ht ; where Rht is the household remittance receipts from overseas, Yht the house- hold income from domestic sources, Xht a vector of time-variant household characteristics, Wh a vector of time-invariant characteristics, gt the time effect for period t, Tt a dummy variable for each time period, the Tt Wh term allows the time effect to vary systematically with household time-invariant character- istics, and 1ht a mean-zero error term. 9. Remittance receipts of non-migrant households are not zero because households can receive remittances from non-household members (such as distant relatives or friends). 10. Although these may seem to be high urban percentages, the Philippine National Statistics Of�ce appears to use a broad de�nition of an urban area, and many areas classi�ed as “urban� are likely to be closely linked to adjacent agricultural areas. Yang and Choi 231 The coef�cient of interest is b, the coef�cient on domestic income Yht. If remittances help insure households from losses of domestic income, this coef�- cient should be negative. Its magnitude represents the replacement rate of dom- estic income by remittances from overseas. Although a rich variety of information is available on household character- istics that might be included in the vector Xht, serious problems remain with obtaining an unbiased estimate of b. First, there is reverse causation: domestic income can itself be a function of remittances, when remittances help fund household entrepreneurial investments. Alternately, households receiving insur- ance through remittances could exert less effort and thus earn lower incomes (Azam and Gubert 2005), leading to a negatively biased estimate of the effect of income on remittances. Such endogeneity concerns motivate this article’s empirical strategy—the use of panel data and the use of rainfall shocks as instruments for income. In this context, the focus can be on the impact of exogenous changes in income on changes in remittances. There should there- fore be no concern that the changes in income are endogenous with respect to remittance receipts due to moral hazard or other reasons. Another concern is omitted-variable bias: unobservable household character- istics (say, the entrepreneurial spirit of household members) are likely to jointly determine domestic income and remittances. The identi�cation strategy focuses on reducing bias generated from simultaneity and omitted variables. With two observations for each household, �rst differences can be used to control for the influence of unobservable household characteristics. Rewriting equation (7) separately for each of the 2 years (1997 and 1998) yields: Rh97 ¼ a þ bYh97 þ u 0 Xh97 þ d 0 Wh þ g97 ð8Þ 0 þ x97 ðT97 Wh Þ þ 1h97 Rh98 ¼ a þ bYh98 þ u 0 Xh98 þ d 0 Wh þ g98 ð9Þ 0 þ x98 ðT98 Wh Þ þ 1h98 : To eliminate the influence of unobservable household time-invariant charac- teristics, Wh, �rst differences can be taken by subtracting equation (8) from equation (9), and rearranging to obtain: DRh98 ¼ ðg98 À g97 Þ þ bDYh98 þ u 0 DXh98 ð10Þ þ ðx98 À x97 Þ 0 Wh þ ð1h98 À 1h97 Þ: It still remains to deal with time-variant heterogeneity, DXh98, and with reverse causation. To do so, the change in household domestic income, DYh98, is instrumented by the change in rainfall over the study period. The change in rainfall should be a valid instrument, as it is likely to have an important effect 232 THE WORLD BANK ECONOMIC REVIEW on household income in a country such as the Philippines, where most house- holds owe their livelihoods either directly or indirectly to agriculture. In addition, it is also plausible that rainfall affects remittances primarily through the change in household income (the instrumental variable exclusion restric- tion).11 The sample is not limited to households in rural areas, since (as dis- cussed earlier) the de�nition of an urban area used in the surveys is quite broad, and many households classi�ed as urban (43 percent) do report non-zero agricultural income. Nor would it be desirable to limit the sample to households with agricultural income. Negative agricultural income shocks should reduce demand on the part of agricultural households for non- agricultural goods and services, so that negative rainfall shocks should also affect income in non-agricultural households. The �rst stage regression is: DYh98 ¼ p0 þ p1 DRAIN DRYh98 ð11Þ þ p2 DRAIN WETh98 þ m 0 Wh þ vh98 ; where DRAIN_DRYh98 and DRAIN_WETh98 are the changes in rainfall in the dry and wet seasons relevant for the change in income between 1997 and 1998, and vh98 is a mean-zero error term. The inclusion of Wh in the regression allows for heterogeneity in the time trend from 1997 to 1998 across house- holds depending on time-invariant characteristics. The predicted change in income from equation (11), DY ˆ h98, can be substi- tuted for DYh98 in equation (10), and various terms rewritten to obtain: ð12Þ ^ h98 þ v 0 Wh þ h ; DRh98 ¼ j þ bDY h98 where j, a constant term, substitutes for the change in year effects, n for the change in the vector of coef�cients (x98 2 x97), and the new error term hh98 for the remaining terms from equation (10), 1h98 2 1h97 þ u0 D Xh98. (Now that the change in household income is instrumented by rainfall, it is plausible to assume that shocks to other household outcomes, D Xh98, are orthogonal to DYˆ h98 and so can safely be included in the error term.) Equation (12) is the estimating equation used in the regression analysis. The variables included in the vector of controls, Wh, are a set of household charac- teristics in the �rst period (January–June 1997): an indicator for urban location; �ve indicators for the household head’s highest level of education completed (elementary, some high school, high school, some college, and college or more; less than elementary omitted); six indicators for head’s occu- pation ( professional, clerical, service, production, other, not working; agricul- tural omitted); and log per capita household income. 11. Robustness checks, conducted later, examine and reject the existence of important alternative channels (other than household income) for rainfall’s effects on remittances. Yang and Choi 233 Regression Results This section describes the impact of rainfall shocks on changes in household domestic income. It then presents the impact of changes in household domestic income (instrumented by rainfall shocks) on changes in household remittance receipts from overseas. It also looks at the impact of instrumented domestic income on total household expenditures and at the number of migrant members in the household. Impact of Rainfall on Domestic Income (First-Stage Estimates) Regression results from the �rst stage—predicting changes in domestic income using rainfall shocks, as in equation (11)—are presented for two speci�cations in table 2. The dependent variable in both regressions is the change in house- hold income from domestic sources between the January–June 1997 and April–September 1998 reporting periods, divided by initial (January–June 1997) total household income. (For example, a change amounting to 10 percent of initial income is expressed as 0.1.12) The mean of the dependent variable is 0.03 for both migrant and non-migrant households, indicating that both types of households experienced increases in domestic-source income on average between the two time periods. Spatial correlation in the outcome variables is likely to be a problem in this analysis, biasing ordinary least squares (OLS) standard error estimates down- ward (Moulton 1986). In particular, the concern is correlation among error terms of households associated with the same weather station, because the rain- fall instrumental variables vary only at this level. So standard errors allow for an arbitrary variance–covariance structure within the coverage areas of 38 weather stations (standard errors are clustered by weather station coverage area). The �rst column in table 2 presents the coef�cient estimates where the rain- fall shock variables are changes in rainfall in the dry and wet seasons. The coef�cient on the dry season rainfall shock is positive and statistically signi�- cant at the 5% level. The coef�cient on the wet season shock is negative but is not statistically signi�cant. A decline of 500 mm of rainfall in the preceding dry season leads to a 3.3 percentage point decline in initial household domestic income. The F-statistic for the test of joint signi�cance of the rainfall variables is 3.150, with a P-value of 0.054. 12. Dividing by pre-crisis household income achieves something similar to taking the log of an outcome: normalizing to take account of the fact that households in the sample have a wide range of income levels and allowing coef�cient estimates to be interpreted as fractions of initial household income. We choose to normalize outcome variables in this way (rather than taking the log) because some second-stage outcome variables (in particular, remittances) often take on zero values. Results are robust to express the dependent variable as the level of income (in pesos) rather than as shares of initial income. 234 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Impact of Rainfall Shock on Domestic Income, 1997–98: OLS Estimates, First Stage of Instrumental Variable Regression Regression Variable (1) (2) Dry season rainfall shock (thousands 0.065** (0.027) 0.198*** (0.054) of millimeters) Square of dry season rainfall shock 0.065*** (0.018) (thousands of millimeters) Wet season rainfall shock (thousands 2 0.034 (0.028) 0.043 (0.066) of millimeters) Square of wet season rainfall shock 0.059* (0.030) (thousands of millimeters) Household head characteristics Highest education level (indicators) Elementary 0.030 (0.027) 0.023 (0.028) Some high school 0.069** (0.034) 0.065* (0.034) High school 0.100*** (0.025) 0.089*** (0.026) Some college 0.189*** (0.036) 0.183*** (0.037) College or more 0.381*** (0.047) 0.385*** (0.047) Occupation (indicators) Professional job 0.195*** (0.037) 0.188*** (0.038) Clerical job 0.151*** (0.033) 0.144*** (0.031) Service job 0.115*** (0.027) 0.107*** (0.028) Production job 0.039 (0.026) 0.032 (0.026) Other job 0.254*** (0.029) 0.246*** (0.027) Does not work 0.159* (0.080) 0.154** (0.074) Household characteristics Log income per capita in household 2 0.355*** (0.031) 2 0.365*** (0.030) Located in urban area 0.126*** (0.018) 0.115*** (0.017) F-statistic: joint signi�cance of 3.150 6.780 rainfall shock variables P-value 0.054 0.000 Number of observations 27,781 27,781 R2 0.03 0.03 *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Dependent variable: change in household domestic income as share of initial household income. Each column reports the results of a �rst-differenced regression. Numbers in parentheses are standard errors, clustered by rainfall station. Domestic income (January– June 1997/April – September 1998) is household total income excluding remittances from overseas expressed as a fraction of initial (January– June 1997) total household income. Rainfall shocks are changes in rainfall between �rst and second period. Omitted occupation indicator is agricultural job. Omitted education indicator is less than elementary. See table 1 for other variable de�nitions. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. The effect of rainfall on household income may be non-linear, so column 2 of the table presents three results of a regression that also includes the square of the wet and dry season rainfall shocks. The dry season shock and its square are both now positive and statistically signi�cantly. The wet season shock and Yang and Choi 235 its square are also both positive, but only the squared term is statistically sig- ni�cant. The four rainfall shock variables jointly appear to be quite strong as instrumental variables: the F-statistic for the test of the joint signi�cance of the rainfall variables is 6.78, with a P-value of 0.000. In addition, the coef�cients on the main effect and squared terms imply fairly substantial effects on income: a decline in dry season rainfall of 500 mm (roughly the standard devi- ation of the dry season change variable) leads to a 11.5 percentage point decline in household income. Because of the relative strength of the instruments as speci�ed in column 2, this regression is used to create the �rst-stage prediction of the change in house- hold income, DY ˆ h98, in the instrumental variables analyses. Instrumental Variables Estimates The instrumental variables estimates are based on regression equation 12, using the regression results in column 2 of table 2 to generate the predicted income change. Instrumental variable standard errors are calculated using a bootstrap procedure that takes into account the variation induced by the gener- ated regressor as well as geographic clustering of observations by rainfall station. OLS standard errors simply account for clustering of observations at the level of the rainfall station. REMITTANCE RECEIPTS. Table 3 presents OLS and instrumental variable regression results where the outcome variable is the change in household remittance receipts from overseas between the January–June 1997 and April– September 1998 reporting periods, expressed as a share of initial (January– June 1997) household income. On average, both migrant and non-migrant households saw increases in remittances: the mean of the dependent variable was 0.07 for migrant households and 0.01 for non-migrant households. For migrant households, the OLS estimate of the impact of the change in household domestic income on the change in remittances is negative and stat- istically signi�cant at the 10 percent level, but is small in magnitude ( 2 0.080). In contrast, the corresponding instrumental variable estimate is negative, large in magnitude, and statistically signi�cant at the 1 percent level: 62.9 percent of income declines are replaced by new inflows of remittances to the household. That said, the standard error on the 2 0.629 instrumental variable point esti- mate is large enough that the null hypothesis of full insurance (that the coef�- cient is equal to negative one) cannot be rejected. The OLS and instrumental variable estimates of the impact of changes in household domestic income on changes in remittances are dramatically differ- ent, highlighting the importance of the instrumental variable approach. A number of factors are likely to help explain this difference. First, measure- ment error in domestic household income will attenuate the OLS coef�cient ( particularly as this is a regression in �rst differences). Second, reverse causa- tion may be at work. For example, increases in remittances may reflect T A B L E 3 . Impact of Domestic Income Shock on Remittances, 1997–98: OLS and Instrumental Variable Estimates 236 Migrant households Non-migrant households Ordinary least Instrumental Ordinary least Instrumental Variable squares variable squares variable Change in domestic household income as share of 2 0.080* (0.042) 2 0.629** (0.216) 2 0.003 (0.002) 0.056** (0.022) initial household income Household head characteristics Highest education level (indicators) Elementary 0.034 (0.048) 0.055 (0.046) 0.007** (0.003) 0.005** (0.002) Some high school 0.045 (0.063) 0.079 (0.066) 0.003 (0.003) 2 0.001(0.003) High school 0.105** (0.048) 0.170*** (0.046) 0.013** (0.004) 0.007* (0.004) Some college 0.159*** (0.046) 0.271*** (0.061) 0.019*** (0.007) 0.008 (0.008) College or more 0.291*** (0.088) 0.514*** (0.130) 0.025*** (0.007) 0.003 (0.010) Occupation (indicators) THE WORLD BANK ECONOMIC REVIEW Professional job 0.003 (0.076) 0.115 (0.082) 0.006 (0.009) 2 0.005 (0.010) Clerical job 2 0.064 (0.063) 0.018 (0.057) 0.008 (0.005) 2 0.001 (0.004) Service job 2 0.003 (0.054) 0.053 (0.063) 2 0.001 (0.005) 2 0.007 (0.005) Production job 2 0.028 (0.061) 2 0.003 (0.061) 0.006 (0.004) 0.004 (0.004) Other job 0.018 (0.049) 0.155** (0.064) 0.010 (0.008) 2 0.005 (0.008) Does not work 0.538 (0.639) 0.632 (0.673) 2 0.182* (0.103) 2 0.192** (0.094) Household characteristics Log income per capita in household 2 0.282*** (0.028) 2 0.490*** (0.075) 2 0.014*** (0.003) 0.006 (0.007) Located in urban area 0.068 (0.046) 0.148*** (0.051) 0.007** (0.003) 2 0.001 (0.004) Number of observations 1,655 1,655 26,126 26,126 *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Dependent variable: change in household remittance receipts as share of initial household income. Each column reports results of a separate �rst-differenced regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (�rst stage is column 2, table 2). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors are bootstrapped. Migrant households are de�ned as those with an overseas worker in June 1997. See table 2 for other notes and table 1 for vari- able de�nitions. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. Yang and Choi 237 increased investment in household entrepreneurial enterprises, leading to increased domestic income. This would lead the OLS coef�cient to be biased in a positive direction. Finally, there may be omitted variables positively corre- lated with both the change in remittances and the change in income. For example, a need to accumulate resources for a large household purchase (such as a vehicle) or some other lump-sum payment (tuition, medical expenses) might lead to both increased remittances, increased domestic labor supply, and increased domestic income. Omitted variable stories such as these would also cause positive bias in the OLS coef�cient compared with the instrumental variable. The contrast with the results for the non-migrant households (the last two columns of table 3) is striking. The OLS coef�cient is essentially zero, whereas the instrumental variable coef�cient is positive (0.056). Although the instru- mental variable coef�cient is small, it is statistically signi�cantly at the 5 percent level: exogenous increases in household income raise remittance receipts from overseas. Further analyses (described later) provide an expla- nation for this result. How valuable is the insurance provided by remittances for Philippine house- holds? One way to gage the welfare gain is to ask what fraction of income households would be willing to give up to reduce rainfall-driven income shocks, both positive and negative, by the amount indicated in column 2 of table 3 (62.9 percent). The actual distribution of rainfall shocks observed in the data is used to calculate predicted 1998 income due solely to this rainfall variation for all households in the data set (only dry season shocks are used because wet season shocks are not statistically signi�cant in table 2). The distri- bution of predicted 1998 income shocks (relative to 1997 income) observed across households is then used to represent the underlying risk to be insured. The calculation assumes constant relative risk aversion utility, U(c) ¼ c 1 2 g/ (1 2 g).13 A household with a reasonable risk aversion parameter ( g ¼ 1.5) should be willing to give up 0.24 percent of income to achieve income with this degree of smoothness. Although this number may seem small, it is large relative to the fraction of income such a household would be willing to give up to achieve complete smoothness, which is 0.28 percent under the same assumptions.14 HOUSEHOLD EXPENDITURES. If remittances serve as insurance for migrant house- holds, changes in household expenditures should be relatively unresponsive to changes in household domestic income, because remittances respond so strongly (and in the opposite direction) to changes in household domestic 13. The theoretical section assumed constant absolute risk aversion utility for tractability of the empirical derivation, but constant relative risk aversion is typically thought to better characterize individual behavior under uncertainty. 14. In an analogous calculation, Lucas (1987) �nds relatively small welfare gains from elimination of aggregate consumption risk in the United States. 238 THE WORLD BANK ECONOMIC REVIEW income. It is also of interest to explore whether expenditures are smoother in migrant households than in non-migrant households in the face of domestic income shocks. Table 4 presents the results from OLS and instrumental variable regressions where the outcome variable is the change in household expenditures between the January–June 1997 and April–September 1998 reporting periods, expressed as a share of initial (January–June 1997) household expenditures. The mean of the dependent variable is 0.00 for migrant households and –0.04 for non-migrant households (on average, expenditures are roughly stable or slightly declining between the periods). The regression speci�cations are other- wise the same as those reported in table 3. The OLS results indicate that household domestic income is highly positively correlated with total expenditures for both migrant and non-migrant house- holds. For example, for migrant households, a 10 percentage point increase in domestic household income is associated with a 5.0 percentage point increase in total expenditure; the magnitude of the OLS coef�cient is similar for non- migrant households. In the instrumental variable speci�cation, however, the income coef�cient for migrant households declines by roughly half (from 0.499 to 0.248) and also declines somewhat for non-migrant households (from 0.623 to 0.508). That the coef�cient on the change in domestic income in the migrant regression declines substantially in the instrumental variable speci�cation (and is not stat- istically signi�cant) is consistent with remittances playing an important role in helping these households maintain their expenditure levels when they experi- ence income shocks. That said, standard errors in the instrumental variable regressions are quite large (the equality of the OLS and instrumental variable coef�cients cannot be rejected), so these results should be taken only as suggestive. The relative decline in the coef�cient on the change in domestic income is larger for migrant households than for non-migrant households, although again standard errors are too large to allow strong conclusions. This is most appropriately taken as merely suggestive evidence that migrant households are better able to smooth expenditures in the face of exogenous income shocks. The comparison between migrant and non-migrant households should be taken as suggestive because households with and without migrants differ across many observed and unobserved characteristics. Differences in results between these two groups of households could be due to the presence or absence of migrants, but it could also be due to other differences such as variations in access to other risk-coping mechanisms (savings, credit, other types of informal and formal insurance). EFFECT OF SHOCK ON OVERSEAS MIGRATION FROM THE HOUSEHOLD. Do exogenous income shocks driven by rainfall also affect whether a household has a member working overseas? Part of the insurance provided by migrants could T A B L E 4 . Impact of Domestic Income Shock on Total Expenditure, 1997–98: OLS and Instrumental Variable Estimates Migrant households Non-migrant households Ordinary least Instrumental Ordinary least Instrumental Variable squares variable squares variable Change in domestic household income (as share of 0.499*** (0.071) 0.248 (0.171) 0.623*** (0.087) 0.508*** (0.156) initial household income) Household head characteristics Highest education level (indicators) Elementary 2 0.084* (0.046) 2 0.111*** (0.043) 2 0.008 (0.010) 2 0.003 (0.021) Some high school 2 0.069 (0.054) 2 0.070 (0.055) 2 0.012 (0.013) 2 0.004 (0.033) High school 2 0.019 (0.042) 2 0.022 (0.064) 2 0.026* (0.015) 2 0.013 (0.027) Some college 0.009 (0.047) 0.002 (0.062) 2 0.041* (0.022) 2 0.016 (0.046) College or more 0.065 (0.054) 0.030 (0.097) 2 0.146*** (0.052) 2 0.090 (0.065) Occupation (indicators) Professional job 2 0.042 (0.058) 2 0.052 (0.059) 2 0.055** (0.024) 2 0.028 (0.036) Clerical job 0.004 (0.045) 2 0.002 (0.050) 2 0.033 (0.024) 2 0.012 (0.032) Service job 0.053 (0.057) 0.060 (0.063) 2 0.038* (0.020) 2 0.020 (0.026) Production job 0.066 (0.061) 0.042 (0.067) 2 0.020 (0.014) 2 0.010 (0.018) Other job 0.045 (0.048) 0.063 (0.064) 2 0.020 (0.025) 0.008 (0.047) Does not work 0.015 (0.150) 2 0.076 (0.165) 2 0.226* (0.083) 2 0.144* (0.078) Household characteristics Log income per capita in household 2 0.113*** (0.026) 2 0.110* (0.064) 0.069** (0.034) 0.019 (0.053) Located in urban area 0.033 (0.034) 0.044 (0.037) 2 0.012 (0.015) 0.006 (0.027) Number of observations 1,655 1,655 26,126 26,126 *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Dependent variable: change in household expenditure as share of initial household expenditures. Each column of table is a separate �rst- differenced regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for �rst- stage regression). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors Yang and Choi are bootstrapped. See table 2 for other notes and table 1 for variable de�nitions. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. 239 240 THE WORLD BANK ECONOMIC REVIEW take the form of delayed return and extended periods of high overseas earnings if their origin households experience negative income shocks. This section shows whether income shocks affect whether a household has a member working overseas. The outcome variable is the change in an indicator for a household having an overseas worker between the July 1997 and October 1998 surveys. For migrant households, this indicator was equal to 1 in the �rst period and could equal 0 or 1 in the second period. The mean of the outcome variable is 2 0.38 for migrant households, meaning that in 38 percent of house- holds with a migrant member in July 1997, all migrant members had returned by October 1998. For non-migrant households, this indicator was equal to 0 in the �rst period. The mean of the outcome variable for non-migrant households is 0.02, meaning that 2 percent of initially non-migrant households had become migrant households by the second period. Table 5 presents the results from OLS and instrumental variables regressions. Speci�cations are the same as in tables 3 and 4. For migrant house- holds, both the OLS and instrumental variable coef�cients on the change in domestic income are negative, but neither is statistically signi�cant at conven- tional levels. There is no indication for migrant households that remittance responses to income shocks are in part explained by migrants’ changing their return decisions. For non-migrant households, the OLS coef�cient is close to zero and is not statistically signi�cant. The instrumental variable coef�cient is positive and statistically signi�cant. The instrumental variable coef�cient (0.075) indicates that a 10 percent increase in domestic income leads to a 0.75 percentage point increase of in the household’s likelihood of having an overseas migrant. This is a large effect, given that the mean of the outcome variable among all initially non-migrant households is 2.0 percentage points. This positive causal impact of income on overseas migration among initially non-migrant households helps explain the positive impact of income on remit- tances in these households (table 3, last column). This may reflect the fact that international migration requires �xed up-front costs (such as fees to recruit- ment agencies), so that households facing credit and savings constraints become more willing or able to pay the �xed costs when current income increases. Robustness Checks This section discusses the evidence against alternative channels to income for rainfall’s effects and against an important potential confounding factor— exchange rate changes in migrants’ overseas locations. POTENTIAL VIOLATIONS OF EXCLUSION RESTRICTION. An important concern when instrumenting for changes in household income using rainfall variation is that rainfall shocks affect all households in a local area. Because of this, at least part of the effects found may be due to changes in locality-level economic T A B L E 5 . Impact of Domestic Income Shock on Indicator for Overseas Worker in Household, 1997–98: OLS and Instrumental Variable Estimates Migrant households Non-migrant households Ordinary least Instrumental Ordinary least Instrumental Variable squares variable squares variable Change in domestic household income as share of 2 0.031 (0.021) 2 0.068 (0.178) 2 0.001 (0.001) 0.075** (0.020) initial household income Household head characteristics Highest education level (indicators) Elementary 0.048 (0.047) 0.051 (0.054) 0.002 (0.002) 2 0.000 (0.002) Some high school 0.079 (0.057) 0.082 (0.054) 0.006*** (0.003) 0.001 (0.004) High school 0.037 (0.049) 0.043 (0.051) 0.006*** (0.003) 2 0.002 (0.004) Some college 0.066 (0.043) 0.076 (0.054) 0.013** (0.005) 2 0.001 (0.006) College or more 0.017 (0.066) 0.038 (0.100) 0.004 (0.007) 2 0.024*** (0.011) Occupation (indicators) Professional job 2 0.112* (0.066) 2 0.102 (0.065) 2 0.006 (0.007) 2 0.020** (0.006) Clerical job 2 0.170*** (0.061) 2 0.163** (0.062) 2 0.004 (0.004) 2 0.015** (0.005) Service job 2 0.045 (0.070) 2 0.040 (0.075) 2 0.008** (0.003) 2 0.017** (0.004) Production job 2 0.108* (0.059) 2 0.104* (0.054) 2 0.003 (0.003) 2 0.006*** (0.003) Other job 2 0.037 (0.050) 2 0.025 (0.062) 0.020** (0.004) 0.001 (0.006) Does not work 2 0.082 (0.225) 2 0.070 (0.253) 0.236* (0.139) 0.224 (0.144) Household characteristics Log income per capita in household 0.010 (0.023) 2 0.008 (0.072) 0.014** (0.001) 0.040** (0.007) Located in urban area 2 0.013 (0.034) 2 0.006 (0.044) 2 0.001 (0.002) 2 0.011** (0.003) Number of observations 1,655 1,655 26,126 26,126 *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Dependent variable: change in indicator for overseas worker in household. Each column is a separate �rst-differenced regression. Instrumental Yang and Choi variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for �rst-stage regression). Numbers in parenth- eses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors are bootstrapped. See table 2 for other notes and table 1 for variable de�nitions. 241 Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. 242 THE WORLD BANK ECONOMIC REVIEW conditions (such as wage rates), rather than merely to changes in household income.15 This would be a violation of the instrumental variable exclusion restriction, the assumption that the rainfall instruments affect household remit- tances only through their effect on household income. This section tests for potential violations of the exclusion restriction. One way in which rainfall might affect remittances is through changes in the relative returns to various types of work, which could induce households to change their labor supply. This could be problematic if changes in household labor supply lead to changes in remittances independent of their effects on household income. For example, if adults in the household spend more time working, households may hire maids or nannies to provide child care, and remittances may rise to help pay for such help. Or households may invite older relatives to live with them and look after children, and remittances may rise to help support the larger number of household members. If such responses are empiri- cally important, the instrumental variable regression estimates of the impact of the change in domestic income on the change in remittances will be biased in directions that cannot be predicted in advance. To test whether such concerns have any basis, it is useful to test the stability of the instrumental variable regression coef�cients to the inclusion of control variables for the change in various alternative channels. In particular, control variables are included for the change in total household hours worked and for the change in household size.16 Any substantial change in the instrumental variable estimates when these control variables are included would cast doubt on the assumption that the effects of rainfall variability work primarily through changes in domestic income. Table 6 presents the results of this exercise. The coef�cient estimates for regressions where the outcome variable is the change in remittances are very similar to those in table 3. For example, the coef�cient in the instrumental vari- able speci�cation for migrant households is 2 0.569 in table 6 (and is statisti- cally signi�cant at the 1 percent level), compared with 2 0.629 in table 3. There appears to be little reason for concern that rainfall affects remittances through changes in household labor supply or household size independently of rainfall’s effects on income. The results for the change in household expendi- ture (row 1) and for the change in the indicator for having an overseas migrant (row 3) are not substantially different from the previous results (tables 4 and 5). The same is true for non-migrant households. AN OMITTED VARIABLE CONCERN: CHANGES IN EXCHANGE RATES. Another general identi�cation concern arises because 1997–98 was a time of substantial 15. Rosenzweig and Wolpin (2000) raise concerns from using weather events as instrumental variables. 16. Hours worked in the past week are reported for all household members above the age of 10. The change is from July 1997 to October 1998. The change in household size is over the same time period and includes overseas members. Yang and Choi 243 T A B L E 6 . Impact of Domestic Income Shock on All Outcomes, 1997–98: Fixed Effect OLS and Instrumental Variable Estimates, Controlling for Change in Household Size and Labor Supply Migrant households Non-migrant households Ordinary least Instrumental Ordinary least Instrumental Outcome squares variable squares variable Total 2 0.084* (0.043) 2 0.562** (0.204) 2 0.003 (0.002) 0.063** (0.022) remittance Total 0.480*** (0.083) 0.305 (0.160) 0.621*** (0.090) 0.521** (0.160) expenditure Overseas 2 0.030 (0.021) 2 0.006 (0.172) 2 0.001 (0.001) 0.079*** (0.018) worker indicator Number of 1,655 1,655 26,126 26,126 observations *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Each cell presents coef�cient estimate on change in domestic household income in a sep- arate regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for �rst-stage regression). Each regression includes control variables for the change in number of household members and the change in hours worked by household members between 1997 and 1998, as well as other control variables included in tables 3– 5 (coef�cients not shown). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instrumental variable standard errors are boot- strapped. See table 2 for other notes and table 1 for variable de�nitions. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. economic fluctuation in the Philippines (and in other Asian countries) due to the Asian �nancial crisis. The Philippine economy experienced a decline in economic growth after the onset of the crisis in mid-1997. Annual real GDP contracted by 0.8 percent in 1998, following growth of 5.2 percent in 1997 and 5.8 percent in 1996 (World Bank 2004). The urban unemploy- ment rate (unemployed as a share of total labor force) rose from 9.5 percent in 1999 to 10.8 percent in 1998, whereas the rural unemployment rate went from 5.2 percent to 6.9 percent (Philippine Yearbook 2001, table 15.1). Of course, any effects of the domestic economic downturn common to all households are not an issue, because the regressions here use �rst-differenced variables, so that common economic shocks are captured in the constant term. In addition, the control variables for households’ 1997 characteristics included in all regressions (education, occupation, income, and urban indicator) will help account for any differential effects of the 1997–98 crisis that differ across households by socioeconomic status. However, there is another important dimension of heterogeneity that is par- ticularly relevant for migrant households: fluctuations in the exchange rates 244 THE WORLD BANK ECONOMIC REVIEW faced by migrant members. The devaluation of the Thai baht in June 1997 set off a wave of speculative attacks on national currencies, primarily in East and Southeast Asia. Overseas Filipinos work in dozens of foreign countries, includ- ing many countries most affected by exchange rate shocks due to the 1997 Asian �nancial crisis, such as the Republic of Korea and Malaysia and, to a lesser extent, Taiwan, China, Singapore, and Japan.17 An omitted variable concern arises if the 1997–98 exchange rate shocks experienced by households in particular areas happen to be correlated with the rainfall shocks in the same areas over the same period. If, for example, areas with greater declines in dry season rainfall (and thus greater declines in income) also had exchange rate shocks that allowed migrants to send more remittances, then the negative relationship between income and remittances would be overstated. To test whether such concerns are empirically important, the main regressions are repeated for migrant households with the change in the exchange rate (Philippine pesos per unit of foreign currency) experienced by the households’ migrants included as a control variable (table 7). The change in the exchange rate is the average of the 12 months leading to October 1998 minus the average of the 12 months leading to June 1997, divided by the second number.18 None of the coef�cients is substantially different from the corresponding coef�cients in tables 3–5. The exchange rate shocks experienced by household migrants appear to be orthogonal to the rainfall shocks experi- enced by their origin households. There is no evidence that omitted variables bias due to correlation between exchange rate and rainfall shocks is a cause for concern. III. CONCLUSION The incomes of households in developing countries are often highly exposed to environmental risk factors such as weather. At the same time, government- sponsored social insurance is generally poor or non-existent. How do house- holds in poor countries shield themselves from environmental risk? This article documents empirically that some households are able to insure themselves without direct government involvement by sending members to work overseas. Their remittances serve as insurance in times of negative income shocks. In households with overseas migrants, exogenous changes in income lead to changes in remittances of the opposite sign, consistent with an insurance motivation for remittances. In such households, the results show a replacement rate of household domestic income by remittances of roughly 60 percent. The null hypothesis of full insurance cannot be rejected. In contrast, changes in 17. Yang (2006b), examines the impact of these heterogeneous exchange rate shocks on return migration and on investment behavior in migrants’ origin households. 18. For further discussion of the exchange rate shock measure, see Yang (forthcoming). Yang and Choi 245 T A B L E 7 . Impact of Domestic Income Shock on All Outcomes, 1997–98: Fixed Effect OLS and Instrumental Variable Estimates, Controlling for Exchange Rate Shock, Migrant Households Only Migrant households Ordinary least Instrumental Outcome squares variable Total 2 0.080* (0.042) 2 0.639** (0.219) remittance Total 0.500*** (0.071) 0.256 (0.169) expenditure Overseas 2 0.032 (0.021) 2 0.107 (0.176) worker indicator Number of 1,655 1,655 observations *Signi�cant at 10 percent level; **Signi�cant at 5 percent level; ***Signi�cant at 1 percent level. Note: Each cell presents coef�cient estimate on change in domestic household income in a sep- arate regression. Instrumental variables for change in domestic household income are rainfall shocks in dry and wet seasons (see table 2 for �rst-stage regression). Each regression includes control variable for the exchange rate shock experienced by migrant members between 1997 and 1998, as well as other control variables included in tables 3– 5 (coef�cients not shown). Numbers in parentheses are standard errors. OLS standard errors are clustered by rainfall station; instru- mental variable standard errors are bootstrapped. See table 2 for other notes and table 1 for vari- able de�nitions. Source: Authors’ analysis based on Philippine National Statistics Of�ce surveys described in text. household income have no effect on remittance receipts in households without overseas migrants. A key question is whether remittance responses to income shocks depend on the performance or availability of alternative methods of coping with risk, such as asset sales, credit markets, and reciprocal transfer networks. In particular, the availability of other risk-coping mechanisms may depend on whether shocks are aggregate (shared by other households) or idiosyncratic (on average uncorrelated with other households). By focusing on income shocks driven by local weather changes, this article assesses the role of remittances as insurance in the face of aggregate shocks to local areas. One reason for the �nding of such large responses of remittances to rainfall-driven income shocks could be that such shared shocks make it more dif�cult to access credit or interhousehold assistance networks that nor- mally help households cope with risk. For example, when a large fraction of households in a local area experiences a negative shock, the demand for credit may rise, pushing up local interest rates. Some substantial fraction of house- holds needing loans may thus be priced out of the credit market. In addition, there may be dif�culties in smoothing consumption through asset sales when 246 THE WORLD BANK ECONOMIC REVIEW there are aggregate shocks, because other households simultaneously seek to sell their assets, driving down prices.19 If local risk-coping mechanisms break down under aggregate shocks, remittance inflows from migrant household members may be used more heavily as a smoothing device. Whether remittances exhibit such large responses to income shocks when the shocks are idiosyncratic, or speci�c to given households, is therefore an important avenue for future research. An idiosyncratic shock to a given house- hold, if truly uncorrelated on average with shocks experienced by other house- holds, should have negligible effects on the quality of local risk-coping mechanisms, and so households should be better able to use such mechanisms than if the shock were aggregate. Remittances might not respond nearly as much to idiosyncratic shocks precisely because households should still have access to alternative local risk-pooling arrangements. These results provide additional justi�cation for government policies facili- tating international migration and remittance flows. For migration-origin countries, greater opportunities for international migration and improvements in the ease of sending remittances should expand the extent to which remit- tances can serve as social insurance. Policies to ease international migration include provision of information and social services for migrants and their families left behind and oversight of recruitment agencies for overseas jobs. Policies to facilitate remittances include strengthening �nancial infrastructure and payment systems to lower the cost and broaden the reach of formal remit- tance channels. Migration destination countries also have numerous policy options, such as expanding immigration quotas and loosening restrictions on formal remittance flows to developing countries. To the extent that immigra- tion policies in the rich world remain relatively restrictive, however, in most countries, remittances will be an important source of insurance for only a small minority of households. Thus, the results reported here do not support the wholesale dismantling of existing systems of social protection in migration origin countries, which will remain important for the majority of households. REFERENCES Altonji, J.G., F. Hayashi, and L.J. Kotlikoff. 1992. “Is the Extended Family Altruistically Linked? Direct Tests Using Micro Data.� American Economic Review 82(5):1177 –98. Azam, Jean-Paul, and Flore Gubert. 2005. “Those in Kayes: The Impact of Remittances on their ´ conomique 56(6):1331–58. Recipients in Africa.� Revue E Bhagwati, Jagdish. 2003. “Borders Beyond Control.� Foreign Affairs 82(1):98–104. Borjas, George. 1999. “The Economic Analysis of Immigration.� In Orley Ashenfelter, and David Card eds., Handbook of Labor Economics. Vol. 3A. Amsterdam: North-Holland. Brown, Richard P.C. 1997 “Estimating Remittance Functions for Paci�c Island Migrants.� World Development 25(4):613–26. 19. This point has been made by Rosenzweig and Wolpin (1993), Fafchamps, Udry, and Czukas (1998), and Lim and Townsend (1998). Yang and Choi 247 Cochrane, John. 1991. “A Simple Test of Consumption Insurance.� Journal of Political Economy 99 (5): 957– 76. Cox, Donald, Zekeriya Eser, and Emmanuel Jimenez. 1998. “Motives for Private Transfers over the Life-Cycle: An Analytical Framework and Evidence for Peru.� Journal of Development Economics 55(1):57–80. Fafchamps, Marcel, and Susan Lund. 2003. “Risk-sharing Networks in Rural Philippines.� Journal of Development Economics 71(2):261 –87. Fafchamps, Marcel, Christopher Udry, and Katherine Czukas. 1998. “Drought and Saving in West Africa: Are Livestock a Buffer Stock?� Journal of Development Economics 55(2):273– 305. Gubert, Flore. 2002. “Do Migrants Insure Those Left Behind? Evidence from the Kayes Area (Western Mali).� Oxford Development Studies 30(3):267 – 87. Kochar, Anjini. 1999. “Smoothing Consumption by Smoothing Income: Hours-of-Work Responses to Idiosyncratic Agricultural Shocks in Rural India.� Review of Economics and Statistics 81(1):50–61. Ligon, Ethan, Jonathan P. Thomas, and Tim Worall. 2002. “Informal Insurance Arrangements with Limited Commitment: Theory and Evidence from Village Economies.� Review of Economic Studies 69(1):209– 44. Lim, Y., and Robert M. Townsend. 1998. “General Equilibrium Models of Financial Systems: Theory and Measurement in Village Economies.� Review of Economic Dynamics 1(1):58–118. Lucas, Robert E. 1987. Models of Business Cycles. New York: Basil Blackwell. Lucas, Robert E.B., and Oded Stark. 1985. “Motivations to Remit: Evidence from Botswana.� Journal of Political Economy 93(5):901–18. Maccini, Sharon, and Dean Yang. 2006. “Under the Weather: Health, Schooling, and Socioeconomic Consequences of Early-Life Rainfall.� University of Michigan. Mace, B.J. 1991. “Full Insurance in the Presence of Aggregate Uncertainty.� Journal of Political Economy 99(5):928 –56. Miguel, Edward. 2005. “Poverty and Witch Killing.� Review of Economic Studies 72(4):1153 –72. Morduch, Jonathan. 1993. “Risk, Production, and Saving: Theory and Evidence from Indian Villages.� Harvard University. Moulton, Brent R. 1986. “Random Group Effects and the Precision of Regression Estimates.� Journal of Econometrics 32(3):385– 97. Munshi, Kaivan. 2003. “Networks in the Modern Economy: Mexican Migrants in the U.S. Labor Market.� Quarterly Journal of Economics 118(2):549– 99. Paulson, Anna. 2000. “Insurance Motives for Migration: Evidence from Thailand.� Northwestern University, Kellogg Graduate School of Management. Paxson, Christina. 1992. “Using Weather Variability to Estimate the Response of Savings to Transitory Income in Thailand.� American Economic Review 82(1):15–33. Platteau, Jean-Philippe. 1991. “Traditional Systems of Social Security and Hunger Insurance: Past Achievements and Modern Challenges.� In E. Ahmad, J. Dreze, J. Hills, and A. Sen eds., Social Security in Developing Countries. Oxford: Clarendon Press. Ratha, Dilip. 2003. “Workers’ Remittances: An Important and Stable Source of External Development Finance.� In Global Development Finance 2003: Striving for Stability in Development Finance. Washington, D.C.: International Monetary Fund. Ravallion, Martin, and L. Dearden. 1988. “Social Security in a ‘Moral Economy’: An Empirical Analysis for Java.� Review of Economics and Statistics 70(1):36–44. Rodrik, Dani. 2002. Feasible Globalizations. NBER Working Paper 9129. Cambridge, Mass.: National Bureau of Economic Research. Rosenzweig, Mark, and Oded Stark. 1989. “Consumption Smoothing, Migration, and Marriage: Evidence from Rural India.� Journal of Political Economy 97(4):905–26. 248 THE WORLD BANK ECONOMIC REVIEW Rosenzweig, Mark, and Kenneth Wolpin. 1993. “Credit Market Constraints, Consumption Smoothing, and the Accumulation of Durable Production Assets in Low-Income Countries: Investments in Bullocks in India.� Journal of Political Economy 101(2):223–44. Rosenzweig, Mark, and Kenneth Wolpin. 2000. “Natural ‘Natural Experiments’ in Economics.� Journal of Economic Literature 38(4):827– 74. Townsend, Robert. 1994. “Risk and Insurance in Village India.� Econometrica 62(3):539–591. Udry, Christopher. 1994. “Risk and Insurance in a Rural Credit Market: An Empirical Investigation in Northern Nigeria.� Review of Economic Studies 61(3):495– 526. United Nations. 2002. “International Migration Report.� Department of Economic and Social Affairs, Population Division, New York. United States Bureau of the Census. 2002. “International Data Base.� Washington, D.C. World Bank. 2004. World Development Indicators 2004. CD-ROM. Washington, D.C. Yang, Dean. 2006a. Coping with Disaster: The Impact of Hurricanes on International Financial Flows, 1970– 2002. NBER Working Paper 12794. Cambridge, Mass.: National Bureau of Economic Research. ——. 2006b. “Why Do Migrants Return to Poor Countries? Evidence from Philippine Migrants’ Responses to Exchange Rate Shocks.� Review of Economics and Statistics 88(4):715 –35. ——. Forthcoming. “International Migration, Remittances, and Household Investment: Evidence from Philippine Migrants’ Exchange Rate Shocks.� Economic Journal.