WPS4883 P olicy R eseaRch W oRking P aPeR 4883 Fertility Response to Natural Disasters The Case of Three High Mortality Earthquakes Jocelyn E. Finlay The World Bank Sustainable Development Network Vice Presidency Global Facility for Disaster Reduction and Recovery Unit March 2009 Policy ReseaRch WoRking PaPeR 4883 Abstract The event of a natural disaster, and being directly affected 2001; North-West Frontier, Pakistan, in 2005; and Izmit, by it, brings a large shock to life-cycle outcomes. In Turkey, in 1999. There is evidence of a positive fertility addition to the replacement effects of higher fertility response to exposure to these large-scale natural disasters following a disaster that caused high mortality, a positive in addition to the response to child mortality. The results fertility response may be induced as children can be used in this study are consistent with those of other studies to supplement household income. This paper analyzes that also find a positive fertility response following three high mortality earthquakes: Gujarat, India, in exposure to a disaster. This paper--a product of the Global Facility for Disaster Reduction and Recovery Unit, Sustainable Development Network Vice Presidency--is part of a larger effort in the network to disseminate the emerging findings of the forth coming joint World Bank-UN Assessment of the Economics of Disaster Risk Reduction. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The team leader--Apurva Sanghi--can be contacted at asanghi@worldbank. org, and the author of this paper at jfinlay@hsph.harvard.edu. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Fertility Response to Natural Disasters: The Case of Three High Mortality Earthquakes Jocelyn E. Finlay1 Harvard School of Public Health Prepared for Global Facility for Disaster Reduction and Recovery at the World Bank 2 1 Address for correspondence: Jocelyn Finlay, Center for Population and Development Studies, 9 Bow St, Cambridge, MA 02138 USA. Email: jfinlay@hsph.harvard.edu. Telephone: +1-617-496-8040. 2 This paper was prepared as a background paper to the joint World Bank - UN Assessment of the Economics of Disaster Risk Reduction. Funding from the Global Facility for Disaster Reduction and Recovery is gratefully acknowledged. Helpful comments from Apurva Sanghi, S. Ramachandran, Martina Tonizzo, Javier Baez, Indhira Santos, and David Canning, and participants in the World Bank's Economics of Disasters Seminar Series are also acknowledged. 1. Introduction The event of a natural disaster can cause large-scale losses to a household. As a result of a natural disaster, households may experience a loss of income, assets, or life. In societies where formal insurance mechanisms are not available, and children have the ability to work to supplement household income, and we may expect that the event of a natural disaster will have a positive effect on fertility. In this paper, I analyze the effect exposure to a natural disaster has on the number of children born into a household. In a developing country context, children are used to assist the household in smoothing consumption over time (Guarcello, Mealli and Rosati 2002). A fertility response to a large-scale disaster can be motivated by the household's need to use children as a mechanism to compensate for income and asset loss as a result of the disaster. Children can earn income to supplement the support provided by the main income earner of the household, they can provide care for young siblings enabling the parents to participate in the work force, and they can provide care to their old-age parents. Following the experience of a large-scale disaster, households may have the incentive to increase the number of children. A positive fertility response, in addition to the child mortality response, can indicate that these families utilize children to supplement diminished income and asset wealth. Children can be used as a form of insurance. In a risky environment, where there is a degree of uncertainty over the household's consumption stream, the risk of income loss is insured by having additional children who may be able to support the family when income and consumption needs are compromised. This use of children is in response to risk and uncertainty. Children are used as an insurance mechanism and the decision to have extra children is based on the probability of a catastrophic event occurring and potential loss of income, assets and life. In this paper, however, the fertility response is a response to a shock. Thus the decision for extra children is ex post rather than ex ante in the case where children are used as insurance. Children can be used to smooth consumption over time in response to risk, in which case the decision to have children will come before the event of a disaster (if it occurs at all). Children can also be 2 used to smooth consumption over time in response to a shock, in which case the decision to have children will come after the event of a disaster. In the demography literature there are insurance and replacement effects in response to child mortality (Schultz 1997). The insurance mechanism is such that households who live in an environment where the risk of child mortality is high will insure against the probability loss of a child by having more children than desired. The replacement effect is an ex post response: following the death of a child another child is born to replace the one who was lost. This of course, relies on the mother being in her fertile years when the loss of the child occurs. The event of a large-scale natural disaster means that among those killed will be children. This shock to human life may induce the replacement effects. In response to an unanticipated natural disaster, fertility will increase due to the replacement effects. In this study, if children are also used as an instrument to smooth household income and consumption over time, we would expect there to be a fertility response in excess of the replacement effects. The aim of this study is to use the cases of three large-scale natural disasters to test the hypothesis of a positive fertility response that is additional to the child mortality response. In this paper I examine three large disasters: Gujarat, India 2001 earthquake (20,000 killed); North- West Frontier, Pakistan 2005 earthquake (73,338 killed); and Turkey 1999 earthquake in the western region (17,127 killed). The paper is structured as follows. In the next section background information regarding the previous findings, the conceptual framework, and facts of the earthquakes are given. In section three the data are discussed. In section four the identification strategy is outlined and the estimated model, results and robustness checks are given. In section five, discussion and conclusions are offered. Tables are presented at the back of the paper. 2. Background The three countries chosen for this case study are chosen for the practical analytical reason of data availability and the scale of the disaster in terms of the number of people killed. The 3 behavioral response to natural disasters is not a new field of study, although it is seldom explored due to the lack of suitable data. In this study, the focus is on the use of children as a form of insurance to smooth life-cycle outcomes over the long term. In this section, clarification over the framework of this analysis is given; differentiating risks from shocks, and the distinction between fertility response to risks, shocks, and mortality. 2.1 Previous Findings 2.1.1 Disasters: Risks and Shocks Disasters can occur with variation over type, magnitude and location. Often in the study of natural disasters, it is assumed that disasters occur with some degree of anticipation and thus pose risk and uncertainty over future outcomes. Volcanic activity, earthquakes and tsunamis occur along fault lines, bushfires occur in forested areas, floods occur along rivers, and hurricanes occur in the tropics. The elements of a disaster that are random, however, are the exact epicenter of the disaster zone and the magnitude of the event. While households may be able to prepare for small anticipated or seasonal disasters, for example annual flooding levels near the average seasonal mean, large-scale disasters that strike without warning make affected individuals' unprepared for the force that is upon them. Thus a natural disaster has two elements to it: risk and shock. This risk is associated with your understanding of the probability of a disaster occurring and one would adjust life cycle decisions according to the exposure to the risk. Natural disasters could also impose a shock, and decisions over savings and the number of children, and other life-cycle decisions determined before the disaster may alter decisions over the allocation of resources following the disaster. Children can supplement depleted household resources as older children can work effectively in either agrarian or newly industrialized zones. Disasters can result in the death or disability of economically active household members, cause damage to farming land, or cause the loss of workplace ­ each case reduces the income and consumption stream and asset base of the household. With more children available to work or provide care to young and old in times of 4 need, the family can attempt to smooth their consumption and other life-cycle outcomes following a disaster. Thus to supplement income following a negative shock to income and asset wealth as a result of exposure to a natural disaster, there may be an increase in the number of children to ensure that the future income stream is not compromised ­ conditional on the child bringing great benefit to the household than cost. In this study, three large-scale earthquakes are the focus. The earthquake in India occurred on January 26, 2001 in Gujarat. According to EM-DAT data (Centre for Research on the Epidemiology of Disasters (CRED) 2008), 20,005 people were killed as a result of this earthquake and nearly 6.5 million people were affected. The earthquake in Pakistan on October 8, 2005 occurred in North West Frontier in the north-west region of the country. It had a devastating effect and killed 73,338 people and 5.1 million people were affected. The earthquake in Turkey hit on August 17, 1999 in Izmit in the western region of the country. It killed 17,127 people and affected 1.3 million people. Each of these earthquakes led to large-scale loss of life and property. Publicity surrounding each of the earthquakes was widespread and images of the damage were televised around the world and remain imprinted in the viewers' minds. Data on earthquakes since 1900 (Centre for Research on the Epidemiology of Disasters (CRED) 2008) indicate that India, Pakistan and Turkey are exposed to the risk of the earthquake events. What remains unanticipated is the magnitude of the earthquake and the exact location of the earthquake, and the timing of the event. Studies that consider the fertility response to natural disasters often account for risk (Cain 1981; Cain 1983; Cain 1986; Pörtner 2008): the seasonality of flooding and hurricanes makes the concept of risk the key instrument. In the case of earthquakes, which are neither seasonal nor predictable in magnitude and epicenter, the concept of a shock is the dominant instrument. In India, according to the EM-DAT data, the 2001 Gujarat earthquake was the largest recorded in terms of casualties since data are available in 1900. The second largest was in 1905 in Kangra, some 1000km north of Gujarat along the Himalya that killed 20,000 people. Gujarat is subject to seismic activity, but it is not along the Himalaya mountain range which is the region most vulnerable to seismic activity. The last large earthquake in Gujarat was in 1956, more than a 5 generation ago, when 113 people died. Thus the force of the earthquake that hit in January 2001 was unanticipated and unprecedented for the Gujarat population. In Pakistan, the earthquake the struck in the North West Frontier was the largest on record in that country in terms of the number of people killed. The second largest earthquake by this measure was in 1935 in Quetta , 1000km south on a different mountain range to that in the North West. The North West Frontier earthquake had its epicenter in the Himalaya Mountain Range. Unsurprisingly, this region is subject to a lot of seismic activity. But the magnitude of the earthquake on October 8, 2005 was unprecedented in terms of lives lost. Thus while the event of an earthquake may be anticipated to some degree, the magnitude of the Kashmir earthquake had a greater effect on human lives than any experienced in this region before. In Turkey, the Izmit earthquake was the second largest since 1900 in this country, and the largest was in Erzincan in 1939 (1000km due east of the Izmit earthquake). While there has been seismic activity around Izmit the scale of the earthquake in 1999 was unprecedented. While each of these countries is clearly subject to seismic activity the timing, exact location, and magnitude were unanticipated, and thus I analyze the event of the earthquakes as an exogenous shock. 2.1.2 Fertility Following a Shock In the US, there is evidence that following a disaster birth rates increase. Following the Okalahoma bombings Rogers, John and Coleman (2005) find that there was a positive fertility response. Given the site of this disaster, they attributed the increase in fertility to the return to traditional family values. In a developing country context, however, the increase in fertility following a disaster may be attributed to the desire to supplement income, and generate and insurance mechanism against future shocks to income and consumption. The return to family values is also a viable option in the developing country context. 6 2.1.3 Children as Insurance in Risky Environments Pörtner (2001) writes in detail about the use of children as insurance in less developed countries. The model developed by Pörtner (2001) incorporates uncertainty over future income, and discusses the benefit children can bring to the household as a substitute for insurance. Children can work at home or provide labor for wages, older children can provide transfers to parents, children expand family ties through marriage, and children can be "sold." In Pörtner (2008) and Cain (1981; 1983; 1986) children are found to be used as insurance in risky environments. The decision to have extra children is based on the potential need for support during a crisis. Support from family can be relied upon to a greater extent than external assistance during a crisis as informal insurance mechanisms at the village level, and contractual agreements, may break down. The use of extended family, and not just children, is explored by Rosenzweig (1988), who finds that in the event of a shock family ties are called upon ex post to assist in smoothing consumption. 2.1.4 A Positive Fertility Response in Addition to the Child Mortality Response The fertility response to child mortality is either an insurance against expected reduction in family size (an i decision), or a replacement effect in which case additional children are born to replace those who were lost (an ex post decision). The strike of a disaster may insight replacement effects which is consistent with the response to child mortality (Schultz 1997), however, exposure to a disaster may change a household's preference for children through a changed perception of community spirit, or the return the traditional family values (Rogers, John and Coleman 2005). Alternatively, the event of a disaster may insight households to supplement future household income with income from child labor to ensure life-cycle outcomes are smoothed over the lifetime of the parents and thus the demand for children increases. 7 In this paper, the core question explored is: Does an unanticipated random shock generated by a natural disaster instill a fertility response over and above the mortality effect? 2.2 Conceptual Framework In this paper I analyze the effect of exposure to a large-scale earthquake on the number of children ever born, and test the hypothesis that exposure to an earthquake will have a positive effect on the number of children ever born in addition to the mortality effect. In the following I present a model to illustrate why we would expect fertility to increase in an area affected by a natural disaster, relative to those in areas unaffected, over and above the mortality response. Consider an individual who has utility over consumption in three periods, period 1, 2 and 3, and that utility in the period specific consumption is additively separable and marginal utility of consumption in each period is increasing at a decreasing rate. To illustrate this I adopt the log function for the felicity. Such that the individual maximizes lifetime utility, U u c1 u c2 u c3 (1) U ln c1 ln c2 ln c3 For simplicity there is no intertemporal discounting. The individual works and earns an exogenously determined income in periods one, Y1 . In period one she will have n1 children who provide benefit to the household consumption possibilities of amount w per child but cost k per child to raise. In period one, she will also save an amount s1 and consume c1 such as to satisfy the budget constraint in the first period of, Y1 w k n1 s1 c1 (2) In period two, the individual will work for income Y2 , she will have n2 children who cost k and provide benefit w , and she will save and amount s2 . Children from period one continue to work and provide benefit to her, but do not impose a cost of care as they are older and able to look after themselves. Savings from period one are brought into period two having earned an interest rate of r . The second period budget constraint such that, 8 Y2 w k n2 wn1 1 r s1 c2 s2 (3) In period three, the representative individual is dependent on the returns from decisions over the number of children and savings in the previous periods. She will have income from their second period children and savings are brought forward from the second period to cover consumption needs, such that, wn2 1 r s2 c3 (4) Rearranging (2), (3), and (4), the utility function in (1) can be written as, U ln Y1 w k n1 s1 ln Y2 w k n2 wn1 1 r s1 s2 (5) ln wn2 1 r s2 First order conditions with respect to s1 , s2 , n1 , n2 are, U u ' c1 1 r u ' c2 0 s1 U u ' c2 1 r u ' c3 0 s2 (6) U w k u ' c1 wu ' c2 0 n1 U w k u ' c2 wu ' c3 0 n2 For this paper, I turn particular focus to second period fertility. Continuing with the logarithmic functional form, this implies that from the fourth first order condition in (6) that, wk w (7) Y2 w k n2 wn1 1 r s1 s2 wn2 1 r s2 Assuming s1 , s2 , and n1 are fixed, n2 can be expressed as the following function, 1 w k 1 r w s n2 Y2 1 r s1 wn1 2 (8) 2w k w The comparative statics of a change in the number of children born in period two, given exogenous changes in income, asset wealth and number of children are, 9 n2 1 0 if w k (9) Y2 2 w k n2 1 r 0 if w k (10) s1 2 w k n2 w 0 if w k (11) n1 2 w k Notice that fertility increases in the event of a natural disaster if the benefit from the children is higher than the cost of caring for the children. If children are costly to care for, and provide little or no contribution towards household consumption possibilities, then fertility in the second period will fall as a result of exposure to an exogenous shock that reduces income, assets and child survival. In period two, the representative individual may be subject to a random shock ­ exposure to a natural disaster. As a result of being exposed to the natural disaster, her income will be lower, there will be loss of accumulated assets, and there will be child mortality. This implies that while the budget constraint in period two takes the same form as in (3), the levels of the variables will be lowered exogenously and the comparative statics in (9), (10) and (11) imply that fertility in the second period will increase. In the next sub-sections, I discuss in detail what would motivate an individual to increase fertility following the event of a natural disaster, broadening the interpretation of the model presented in this section. 2.2.1 Explaining the Fertility Increase: Using Children to Smooth Consumption Children can be a form of insurance. But we think of insurance as a financial instrument that a household would take out ex ante based on the probability of an event occurring that has a negative impact on life-cycle outcomes. The justification for children as insurance is elaborate, and even though in this paper I explore the use of children to smooth consumption as an ex post decision, the motivation for having children can translate from the risk analysis to the analysis of the response to an exogenous shock. 10 In the context of a developing country, formal insurance mechanisms are not readily available and there may be use of alternative mechanisms to smooth life-cycle factors in the face of fluctuations in income. The life-cycle factors that could be affected by a large shock such as an earthquake are: consumption; asset accumulation; and family formation. An earthquake can disrupt a household's income stream and thus jeopardize consumption possibilities. Assets such as housing, land, livestock, small business, can be damaged or disappear as a result of the earthquake. Family formation, another life-cycle decision, may also be adversely affected by an earthquake, with mortality of the parents or children bringing differential effects. The loss of the key breadwinner will have obvious repercussions over consumption possibilities of the household; the loss of the key care giver will require a diversification of support for the family unit; and the loss of children may induce a replacement response. Where formal insurance is available, individuals can make payments on a premium on a regular basis and receive a pay-out when in need to ensure that life-cycle outcomes are smoothed over an adverse shock. In the case of formal insurance, the premium paid is monetary and the pay-out is monetary. This is also the case for some informal insurance mechanisms that operate at the village level. In the case of consumption and asset loss insurance, we can hypothesize that children can be used as a form of insurance in these cases. The premium cost is the cost of raising the child, and the pay-off comes from the children when needed. Children can supply labor when young (or adult-child labor with intergenerational transfers to elderly parents) to provide for the family unit so that the family does not rely on a sole breadwinner. Children can also provide care to their siblings, and adult children can care for elderly parents. In crude terms, the premium cost of children as insurance is different to the formal and informal insurance mechanisms that involve monetary exchange. Children are multi-use, versatile insurance, thus the cost of raising the child is the premium cost for many forms of insurance: property damage, job loss, consumption, or savings for old-age support. The cost of raising the child may be less than the sum of premiums one would pay for all the types of insurance in the formal or informal monetary insurance markets. 11 With the notion that children can be a mechanism for insurance, we may hypothesize that in countries or regions that have little access to monetary forms of insurance (formal or informal) then children may be utilized as a form of insurance (Pörtner 2001). Children can be used as a form of insurance, a decision made ex ante, but the decision to have more children can also occur ex post ­ both with the goal of smoothing consumption over time. 2.2.2 Explaining the Fertility Increase: Other Possibilities When conducting this analysis using a difference-in-difference estimation technique, one of the key issues relates to omitted variables. In the reduced form difference-in-difference estimation a positive and significant difference between treatment and control indicates that between the pre- and post-period an event occurred in the treatment area that induced a positive fertility response in the treatment area relative to the control area. This positive and significant fertility response in the treatment area could be due to a shock that affected the treatment but not the control group. Alternatively, it could be due to differential trends between factors in the treatment and control groups. In the regression analysis we control for education and mortality, consistent with the model developed above (education proxies income and assets), however, there may be other factors that change differentially between treatment and control groups between the pre- and post-periods that insights a positive and significant fertility response in the treatment area. One of the factors that may have varied differentially across treatment and control groups is access to contraception. For example, following the large-scale earthquake, contraceptive access may have become limited relative to what it was before. Thus following the earthquake, the probability of becoming pregnant may have increased. Another option is that if in the event of a high mortality earthquake, loss of a spouse may then encourage remarrying and the new couple has the desire to have children of their own. Another alternative is that there is a return to family values as a result of exposure to the earthquake and this increases a couple's preference for the number of children they desire. All of the above suggestions however are second order to the event of the disaster. 12 Another possibility is that some other shock affected the treatment area relative to the control area during the same time period. There may have been a policy change in each of the three treatment areas relative to the control areas that had a positive effect on fertility. The likelihood of a policy change in the three affected regions, within the time period in this study, is, however, reasonably low given there are three treatment groups across three countries. Another possibility is that there is some factor that is changing differentially between the treatment areas and the control areas that affects fertility. I test for the differential trend between treatment and control groups in the included independent variables, but there may be omitted variables that change differentially between treatment and control groups over time. With this is mind, results should be interpreted with caution. 3. Data The data I use for this investigation is from the Demographic and Health Survey (DHS) (Demographic and Health Surveys 2005). The DHS is a nationally representative (state representative for the case of India) household survey that subsumed the World Fertility Survey. Data are available online free of charge on request to the distributors (Macro International). The survey has been conducted over 76 developing countries since 1985, and there are repeated surveys in most of the countries. The data however are not a panel, but rather a repeated cross section. Following a particular family is not possible, and clusters change between each wave, but we can observe changes in the average behavior at the regional and country level. To conduct the analysis for this project, I chose to focus on India, Pakistan and Turkey. These three countries satisfied two key criteria for selection: they each had a natural disaster that in terms of the number of people killed were in the global top 20 since 1990 (see Table 1) according to EM-DAT (Centre for Research on the Epidemiology of Disasters (CRED) 2008). Secondly, DHS data were available before and after the earthquakes in each country. In this investigation, the fertility response in addition to the response to mortality is of particular interest. Thus, for there to be a mortality response, high mortality natural disasters must be the focus. The reason for requiring data before and after the natural disaster is to conduct difference- 13 in-difference estimation to observe changes in the average over time of people living in those regions affected by the disasters relative to those living in unaffected regions. Data for India are taken from the 1998 (International Institute for Population Sciences (IIPS) and ORC Macro 2000) and 2005 (International Institute for Population Sciences (IIPS) and Macro International 2007) surveys, and the earthquake occurred in Gujarat 2001. Data for Pakistan are taken from the 1990 and 2006 surveys, and the earthquake occurred the North West Frontier in 2005. Data for Turkey are taken from the 1998 and 2003 (Hacettepe University Institute of Population Studies 2003) surveys, and the earthquake occurred in Izmit in 1999. As the data for each of the countries are not a panel, we cannot observe changes in behavior at the household level and rely on changes in the regional average over time. There are a number of issues that come about because of this, each of which ameliorates the effect of the earthquake on the observed behavioral response. The response to the earthquake may not be limited to those who were directly affected by it, and thus there are spillovers into the control groups (regions other than those affected by the earthquake). If there is a response in the control group, then in the difference-in-difference estimation the change in fertility over time in the treatment group compared to the control group may appear to be smaller than what it actually is in response to the earthquake. The second issue that arises from not having a household panel survey is that out- migration following the earthquake may induce a selection bias: those household with preferences for large families before the earthquake remain in the treatment area and those with smaller families have the capability to move away from the affected region. This then means that when the post earthquake data are observed the regional average family size for the affected area is larger as the smaller families moved away from the treatment area. To control for this in a robustness check, the sample is limited to respondents who have never moved. The third issue to recognize is the dampening effects emergency foreign aid may have on a household's behavioral response. While this is a welcome response for those people affected by the disaster, it distorts household decisions that may include a fertility response. \ 14 4. Identification For this investigation, to observe changes in fertility preferences following the event of exposure to a large-scale natural disaster a difference-in-difference estimation technique is used. The change in the difference between treatment and control averages is identified using this strategy. With adequate control variables, we can attribute any change in the difference between treatment and control groups to the shock (that is, the earthquakes) experienced by the treatment groups. As the dataset was not generated for the specific purpose of analyzing the fertility response to the earthquakes 3 other variables must be controlled for to isolate other factors that may have changed differentially between the treatment and control groups over time. In addition to controlling for these variables (education of the mother and father, age at first birth, age of the mother, urban/rural place of residence, number of children who have died), a robustness check is also conducted. These controls are interacted with the difference-in-difference dummy variables to see if there is a placebo response and differences in fertility may be generated by differences in another variable (e.g. education of the mother) that is different over time across the treatment and control group. 4.1 Equation for Estimation From equation (8), the number of children born is a function of income, assets, the number of children who survived from the previous period, and a contemporaneous residual. An exogenous shock, that is the event of an earthquake, will reduce income, assets and the number of children and increase the number of children born after the earthquake as illustrated in (9), (10), and (11). To analyze the effect of the shock on the number of children born, consider the difference-in- difference approach where the difference across time between the treatment and control groups is compared. 3 Indeed, if the earthquake is a random unanticipated shock conducting a purpose built survey would imply that this assumption is violated as the location and timing of the earthquake would be known to the survey designers. 15 N irt 1 postt 2 nearr 3 postt * nearr X ' irt (12) Where N irt is the number of children woman i in region r has ever had at time t. The difference- in-difference dummies are postt which is equal to one in the time period after the earthquake and is equal to zero before the earthquake. The value of the dummy variable is common across all women and regions within a time period. The dummy variable labeled nearr is equal to one if the region was affected by the earthquake (Gujarat for India, North West Frontier (North) for Pakistan, and Izmit (West) for Turkey) and zero otherwise. The value of this dummy is common across time. The interaction term, postt * nearr hosts the coefficient of interest, and informs us of any difference between the treatment and control groups in the number of children born that is attributable to the earthquake. The vector of controls is taken from equation (8). Changes in income, assets the number of children who survived from the first period are control variables derived from the model. One of the objectives of this paper is to isolate the fertility response following a natural disaster from the mortality response, thus, the key control variable is the number of children of mother i who have died. The coefficient on this variable should be positive and significant, consistent with replacement effects (Schultz 1997). Income and assets are proxied for using education of the husband. There are other socio-economic factors that we may expect to affect the number of children a woman has. Women who have attained higher levels of education may have fewer children as the opportunity cost of having an extra child may be higher if she is earning a skilled wage in the workforce (other reasons as well such as a greater understanding of reproductive health if she has access and the capability to read information, moreover if she is educated she may value her children's education and thus employ the quality/quantity tradeoff with fewer more highly educated children). Living in an urban area, may imply fewer children ever born as space may be limited and childcare more difficult compared to a rural setting where the work (on the farm) and childcare duties can be conducted jointly. In addition to these socio-economic variables, demographic variables such as the age of the mother and the mother's age at first birth are controlled for. 16 In conducting a difference-in-difference estimation, we may be concerned that other factors may change differentially between the treatment and control groups over time. In which case, it would be these other factors driving the results of interest, and not the exogenous shock of the earthquake. To control for this, the difference-in-difference dummies are interacted with the control variables that vary across time to test whether these variables vary differential in treatment and control area. In this case, the following equation is estimated, N irt 1 postt 2 nearr 3 postt * nearr X ' (13) 10 postt * xirt 11nearr * xirt 12 postt * nearr * xirt irt The xirt is either the number of children born, education of the mother, age of the mother, age at first birth, urban, or the education of the father. If coefficient 12 is significant, then this implies that that particular variable changes differentially between the treatment and control groups over time, and this would contaminate the estimated coefficient of 3 in equation (12). The a list of control variables, including regional fixed effects, can never be definitive, and thus the risk of omitted variables means that the difference-in-difference parameter is picking up another factor that changed differentially between treatment and control groups over the same period. To rule out this possibility, a further test is conducted. The core regression in equation (12) is run for a randomly selected group of regions (one region per country) to be the treatment groups and the remainder to be the control groups. 4.2 Results In Table 2 the pre-shock summary by treatment status statistics are presented. In the total sample across the three countries, we see that before each country's respective earthquake, the average age of the respondent was 31 years. The average number of children who died per woman is 0.36 on average, but with the stratification by urban/rural we see that the number of children who died per woman is 0.42 in rural areas and 0.24 in urban areas. The average number of years of 17 education of the women surveyed is 3.9 years. In rural areas it is lower at 2.6 years on average and in urban areas it is 6.2 years. The education of the husband is higher than that of the women on average, with 6.4 years as the total average, and 5.3 years in rural areas and 8.5 years in urban areas. The age at first birth is 19.3 years on average. The pre-shock summary statistics for each of the countries is reported in Table 2. For many of the variables there is a significant difference between the treatment and control groups. These variables are controlled for in the regression analysis, and thus differences in variables in the pre- disaster period (and any subsequent difference in the changes over time) are accounted for. Variation in the pre-shock summary statistics may raise doubt with regard to what the difference- in-difference dummies are picking up. Changes in education, age at first birth, or other time varying factors that may change differentially between the treatment and control groups may contaminate the effect of the earthquake on the number of children born. To check this concern, in Table 5 time variant variables are interacted with the difference-in-difference dummy variables. Further discussion on this is to follow, but the results indicate that there were no differential changes in education, age, age at first birth, father's education and urban living, between the treatment and control groups. Thus while the treatment and control groups many differ over some of the control variables in the pre-shock period (as shown in Table 2), these socio-economic and demographic factors do not change differentially over the treatment and control groups over time. In Table 3 the difference-in-difference reduced form estimation results of the impact of the earthquake on the number of children ever born are presented. In the first set of reduced form regressions (the total sample), the results indicate that the average number of children born to a woman has declined over time in both the treatment and control groups. However, in the total sample, we see that the decline in the number of children is greater in the control group than in the treatment group. The positive difference-in-difference coefficient in the reduced form regression analysis in Table 3 implies that some factor (potentially the earthquake, but without sufficient controls we cannot 18 yet make this assertion) has had a positive effect on fertility so that the difference between the treatment and control has widened over time. In Table 4a and Table 4b the main regression results are presented. Regional fixed effects are controlled for in each of the regressions. Observing the coefficient on `post' the regression results indicate that the number of children ever born has declined between the two periods. With the full controls in column (8) of Table 4a we see that the total number of children born is higher in the treatment group. In column (8), we also see that even after controlling for other explanatory variables, the event of the earthquakes in the three countries has had a positive effect on the average total number of children in the treatment group compared to the control group (coefficient on post*near is 0.08). If we feel confident that sufficient controls have been included in the regression analysis, and we can attribute any remaining change in the number of children born to the event of the earthquake, then the coefficient of 0.08 indicates that the earthquakes had a positive and significant effect on the number of children ever born, albeit it a modest marginal effect. Viewing across columns in Table 4, we see that the introduction of each control lowers the marginal effect on the post*near variable. The controls are all highly significant. The number of children who have died has a positive effect on the number of children born consistent with replacement theory. In column (8) with full controls, we see that this marginal effect of 0.8 implies that replacement is slightly less than one for one. The age of the mother has a positive effect on the number of children ever born, with a coefficient of around 0.1 this indicates that over a 30 year fertile life (age 15-45) a woman would have 3 children if age were the only factor influencing fertility. The number of years of schooling of the mother has a negative effect on the number of children ever born. Fewer children are born in urban areas. The coefficient on father's education is negative. If father's education adequately proxies household income, then the decline in income leading to an increase in fertility is consistent with the theory. In this investigation, one of the key insights is to isolate a fertility response to the earthquake that is additional to a mortality effect. Thus, to ensure that the mortality response does not change differentially over the treatment and control groups, the number of children who have died is 19 interacted with the difference-in-difference variables. In column (8) with the full set of controls we see that the marginal effect of number of children who have died is not different across the treatment and control groups over time. In Table 4b, column (8) from Table 4a is presented again to serve as a point of reference. In columns (9), (11) and (12), the results for the individual countries are presented. From column (9) we see that in India there was no differential fertility response in Gujarat relative to the rest of India following the earthquake in Gujarat. In Pakistan (column (11)) and in Turkey (column (12)), however, we see that there was a positive fertility response following the event of the earthquake in addition to the response to child mortality. For both countries, the response to child mortality is close to one (0.77 children for every child who died in Pakistan, and one for one in Turkey) and that in the treatment areas there was a positive fertility response in addition to the mortality response of around 0.2 children. In the whole sample (column (8)) we may be concerned that the differential fertility response is driven by migration effects. For example, following a disaster, it may be that individuals with smaller families have a higher propensity to move away from the affected region than larger families. Quite simply, smaller families take up less space, and friends and relatives in other parts of the country are more easily able to assist a smaller family than a larger one. To overcome this issue using the DHS, I utilize information on the length of time the respondent has been living in the same village or city. With a sample of individuals who have always lived in the same village or city, I run the same analysis as that in column (8). As data on length of stay in the same village or city does not exist for Pakistan, the analysis is conducted for India (column (10)) and Turkey (column (13)). This restriction of the sample, however, is not a perfect control for migration. Individuals who moved as a result of the earthquake are excluded from the sample. The benefit of this sample restriction however, is that we are comparing like with like across treatment and control and within this sub-sample migration is not an issue. Taking Turkey as the example, column (13), we see that for the sample of individuals who have always lived in the same place, there is a positive fertility response in addition to the fertility response to mortality. The coefficient measuring this response, post*near, is larger for the sample 20 that has never moved than for the whole sample (0.214 in column (12) for the whole sample, compared to 0.293 in column (13) for the sample who have never moved). In Table 4a and Table 4b the potential for differential effects of the number of children who have died over the treatment and control groups over time were ruled out (insignificance of post*near*dead in columns (7) and (8). The same concern may exist for other control variables. In Table 5 time variant variables are interacted with the difference-in-difference dummy variables. We see that for each of the control variables interacted with the difference-in- difference dummy variables that none of the variables change differentially between the treatment and control groups over time. In the difference-in-difference estimation, the regions that are affected by the three earthquakes are assigned as the treatment groups, and other regions are control groups. After taking into account various control variables, education of the mother and father and age at first birth for example, differences in the number of children born between the two regions is attributed to the incidence of the earthquake. The incidence of the earthquake in the treatment areas is unique to those treatment regions, and the resultant fertility response exhibited as a result of being exposed to the earthquake should be unique. If, for example, the treatment region is randomly assigned to any region (other than the true treatment region) then we should not observe a differential fertility response in that region relative to the remaining regions. In Figure 1, the t-statistics of the post*near variable are given for all the regressions with every possible combination of "treatment" regions. In 19 percent of the regressions, the co-efficient on post*near is statistically significant at the 5 percent level. Thus with 81 percent confidence, the regression analysis reported in Table 4a and Table 4b, can be attributed to the incidence of the earthquakes in Gujarat, North West Frontier and Izmit. 5. Discussion and Conclusion In this paper I have explored the hypothesis that there is a positive fertility response to exposure to natural disasters in addition to the child mortality response. Children can be used to smooth consumption over time. Once children have grown they can work to supplement the income of 21 the main breadwinner, they can provide care to siblings, and when older they can make intergenerational transfers to their elderly parents and provide them with care. The event of an earthquake can make individuals exposed to the disaster realize the vulnerability of their life- cycle outcomes, and they can elect to increase their family size as a mechanism to smooth consumption over time. If indeed this is the reason for the increase in children, then the parents are not myopic in their decision making. The children must mature and reach an age (around seven) when they can be relied on to provide for the family unit in some way. The results found in this paper, that is the increase in fertility following a negative shock to income, assets and child survival, are consistent with other studies of exogenous shocks (Rosenzweig 1988; Rogers, John and Coleman 2005), but inconsistent with others Pörtner (2008). The role of children as an insurance mechanism in an environment subject to risk and uncertainty over income and consumption is well explored as discussed above. The use of children to smooth consumption following an exogenous negative shock is consistent with this insurance literature ­ only that the decision to have children is ex post rather than ex ante. In this paper, I found the fertility response to exposure to natural disasters to be positive and significant. This response is additional to the response to child mortality ­ the well known replacement effect. In the data, we observe a positive response to child mortality just under a one-for-one replacement, and in areas affected by the earthquake there is an additional positive fertility response. If we are convinced that there are adequate controls in the regression analysis with regional fixed effects and other socio-economic and demographic variables influencing the number of children born to a woman, then the results suggest that exposure to an earthquake does induce a positive fertility response over and above the mortality response. Using data that is not purposefully built for the question at hand will always introduce irregularities into the analysis. The timing of the response is an issue that I have not been able to properly reconcile. Each of the three disasters occur within a different lead and lag to the before and after surveys. While discounting the years following the disaster may be one approach, the degree of discounting is unknown and thus has been excluded. The second issue that cannot be properly dealt with in this data set is that of migration. In the robustness checks the sample is 22 restricted to people who have never moved, but this does not capture individuals who moved as a result of the disaster. The rise in fertility following exposure to a natural disaster found in this paper, coupled with results found by Baez and Santos (2007), which indicates that childhood outcomes over education, nutrition, and child labor worsen, implies that there will be more children under relatively worse conditions following the event of the disaster. With lower investments in children following a disaster, this may weaken the capability of using children as an insurance mechanism over the long run. With lower education attainment following the disaster (Baez and Santos 2007), for example, the children will grow up to earn a lower wage. Thus their income earning capabilities will be lower, and if the children were had with the idea of supporting the family or supplementing the main breadwinner then the probability of achieving this goal is lower. The results imply that helping communities subject to natural disasters to gain access to alternative forms of insurance could have positive implications for childhood outcomes. 23 References Baez, J. E. and I. V. Santos (2007). "Children's Vulnerability to Weather Shocks: A Natural Disaster as a Natural Experiment." Mimeo. Cain, M. (1981). "Risk and Insurance: Perspectives on Fertility and Agrarian Change in India and Bangladesh." Population and Development Review 7(3): 435-474. Cain, M. (1983). "Fertility as an Adjustment to Risk." Population and Development Review 9(4): 688-702. Cain, M. (1986). "Risk and Fertility: A Reply to Robinson." Population Studies 40(2): 299-304. Centre for Research on the Epidemiology of Disasters (CRED). (2008, 2008). "EM-DAT: Emergency Events Database." Retrieved 6th October 2008, from http://www.emdat.be/. Demographic and Health Surveys (2005). Guarcello, L., F. Mealli and F. C. Rosati (2002). "Household Vulnerability and Child Labor: The Effect of Shocks, Credit Rationing and Insurance." Working Paper, UCW. Hacettepe University Institute of Population Studies (2003). Turkey Demographic and Health Survey. Ankara, Turkey, Hacettepe University Institute of Population Studies, Ministry of Health General Directorate of Mother and Child Health and Family Planning, State Planning Organization and European Union. International Institute for Population Sciences (IIPS) and Macro International (2007). National Family Health Survey (NFHS-3), 2005-06: India: Volume I. Mumbai: IIPS. International Institute for Population Sciences (IIPS) and ORC Macro (2000). National Family Health Survey (NFHS-2) 1998­99. India. Mumbai: IIPS. Pörtner, C. (2001). "Children as Insurance." Journal of Population Economics 14: 119-136. Pörtner, C. (2008). "Gone with the Wind? Hurrican Risk, Fertility and Education." University of Washington, Department of Economics, Working Papers UWEC-2006-19. Rogers, J. L., C. A. S. John and R. Coleman (2005). "Did Fertility Go Up After Oklahoma City Bombing? An Analysis of Births in Metropolitan Counties in Oklahoma, 1990-1999." Demography 42(4): 675. Rosenzweig, M. (1988). "Risk, Implict Contracts and the Family in Rural Areas of Low-Income Countries." The Economic Journal 98: 1148--1170. Schultz, T. P. (1997). Demand for Children in Low Income Countries. Handbook of Population and Family Economics. M. R. Rosenzweig and O. Stark. Amsterdam, North-Holland. 1A: 643. 24 Appendix 1: Tables Table 1: Top 20 Natural Disasters since 1990 by number of people killed. Estimated Damage, US$ Start Date End Date Country Type of Disater Number Killed Total Affected milllion 26/12/2004 26/12/2004 Indonesia Wave / Surge 165,708 532,898 4,452 29/04/1991 10/5/1991 Bangladesh Wind Storm 138,866 15,438,849 1,780 8/10/2005 8/10/2005 Pakistan Earthquake 73,338 5,128,000 5,200 21/06/1990 21/06/1990 Iran Islam Rep Earthquake 40,000 710,000 8,000 26/12/2004 26/12/2004 Sri Lanka Wave / Surge 35,399 1,019,306 1,317 15/12/1999 20/12/1999 Venezuela Flood 30,000 483,635 3,160 26/12/2003 26/12/2003 Iran Islam Rep Earthquake 26,796 267,628 500 16/07/2003 15/08/2003 Italy Extreme Temperature 20,089 4,400 26/01/2001 26/01/2001 India Earthquake 20,005 6,321,812 2,623 1/8/2003 20/08/2003 France Extreme Temperature 19,490 4,400 17/08/1999 17/08/1999 Turkey Earthquake 17,127 1,358,953 20,000 26/12/2004 26/12/2004 India Wave / Surge 16,389 654,512 1,023 1/8/2003 11/8/2003 Spain Extreme Temperature 15,090 880 25/10/1998 8/11/1998 Honduras Wind Storm 14,600 2,112,000 3,794 28/10/1999 30/10/1999 India Wind Storm 9,843 12,628,312 2,500 29/09/1993 29/09/1993 India Earthquake 9,748 30,000 280 00/08/2003 00/08/2003 Germany Extreme Temperature 9,355 1,650 26/12/2004 26/12/2004 Thailand Wave / Surge 8,345 67,007 1,000 18/08/1991 18/08/1991 Peru Epidemic 8,000 6/5/1991 6/5/1991 Nigeria Epidemic 7,289 10,000 Source: EM-DAT, http://www.emdat.be/ 25 Table 2: Pre-shock summary statistics by treatment status Total Rural Urban Total Treatment Control Difference Total Treatment Control Difference Total Treatment Control Difference Age 31.180 31.284 31.172 0.112 30.728 31.024 30.710 0.313 32.013 31.535 32.069 -0.534*** 0.027 0.103 0.028 0.104 0.034 0.146 0.035 0.148 0.045 0.145 0.047 0.146 105,490 7,680 97,810 68,349 3,773 64,576 37,141 3,907 33,234 Dead 0.360 0.318 0.363 -0.044*** 0.422 0.382 0.424 -0.0420*** 0.245 0.257 0.243 0.013 0.003 0.008 0.003 0.010 0.003 0.013 0.003 0.015 0.003 0.011 0.004 0.011 105,490 7,680 97,810 68,349 3,773 64,576 37,141 3,907 33,234 Education 3.904 4.327 3.871 0.456*** 2.633 2.762 2.625 0.136** 6.243 5.838 6.290 -0.453*** 0.014 0.053 0.015 0.055 0.015 0.062 0.015 0.064 0.026 0.079 0.028 0.086 105,450 7,679 97,771 68,327 3,772 64,555 37,123 3,907 33,216 Husband's education 6.397 6.497 6.389 0.108* 5.297 5.151 5.305 -0.153* 8.484 7.876 8.551 -0.675*** 0.016 0.058 0.016 0.062 0.018 0.076 0.019 0.081 0.027 0.083 0.028 0.089 102,797 7,125 95,672 67,318 3,605 63,713 35,479 3,520 31,959 Urban 0.352 0.509 0.340 0.169*** 0.001 0.006 0.002 0.006 105,490 7,680 97,810 Age at first birth 19.323 19.903 19.280 0.624*** 18.841 19.391 18.809 0.581*** 20.227 20.435 20.204 0.230*** 0.012 0.046 0.012 0.047 0.014 0.062 0.014 0.061 0.022 0.067 0.023 0.073 92,355 6,373 85,982 60,243 3,243 57,000 32,112 3,130 28,982 India India Rural India Urban Total Treatment Control Difference Total Treatment Control Difference Total Treatment Control Difference Age 31.364 31.957 31.338 0.619*** 30.806 31.268 30.790 0.479** 32.602 32.866 32.585 0.281 0.029 0.141 0.030 0.144 0.035 0.188 0.036 0.192 0.050 0.211 0.052 0.213 90,303 3,845 86,458 62,248 2,188 60,060 28,055 1,657 26,398 Dead 0.365 0.355 0.366 -0.011 0.425 0.423 0.425 -0.002 0.232 0.264 0.230 0.034** 0.003 0.012 0.003 0.013 0.004 0.017 0.004 0.019 0.004 0.016 0.004 0.016 90,303 3,845 86,458 62,248 2,188 60,060 28,055 1,657 26,398 Education 3.896 4.461 3.871 0.590*** 2.657 2.839 2.650 0.189** 6.647 6.602 6.650 -0.048 0.016 0.080 0.016 0.077 0.015 0.086 0.016 0.084 0.031 0.129 0.032 0.131 90,265 3,844 86,421 62,228 2,187 60,041 28,037 1,657 26,380 Husband's education 6.478 6.945 6.457 0.488*** 5.394 5.506 5.390 0.115 8.884 8.847 8.886 -0.039 0.017 0.081 0.017 0.084 0.019 0.098 0.019 0.103 0.030 0.121 0.031 0.126 90,086 3,843 86,243 62,117 2,187 59,930 27,969 1,656 26,313 Urban 0.311 0.431 0.305 0.126*** 0.002 0.008 0.002 0.008 90,303 3,845 86,458 Age at first birth 19.205 19.436 19.195 0.241*** 18.741 18.803 18.738 0.065 20.223 20.278 20.220 0.057 0.013 0.055 0.013 0.062 0.014 0.069 0.014 0.076 0.024 0.086 0.025 0.103 80,872 3,443 77,429 55,560 1,966 53,594 25,312 1,477 23,835 continued... 26 Table 2 continued... Pakistan Pakistan Rural Pakistan Urban Total Treatment Control Difference Total Treatment Control Difference Total Treatment Control Difference Age 31.270 31.250 31.276 -0.026 30.906 30.690 30.992 -0.302 31.616 31.936 31.526 0.410 0.103 0.211 0.119 0.238 0.151 0.287 0.178 0.336 0.141 0.310 0.159 0.340 6,611 1,665 4,946 3,227 916 2,311 3,384 749 2,635 Dead 0.480 0.453 0.489 -0.036 0.505 0.442 0.530 -0.088** 0.457 0.467 0.454 0.014 0.012 0.023 0.014 0.028 0.018 0.031 0.022 0.040 0.017 0.033 0.019 0.040 6,611 1,665 4,946 3,227 916 2,311 3,384 749 2,635 Education 1.807 1.213 2.007 -0.794*** 0.518 0.456 0.542 -0.086 3.036 2.139 3.291 -1.152*** 0.045 0.076 0.054 0.103 0.032 0.059 0.038 0.071 0.077 0.147 0.088 0.183 6,609 1,665 4,944 3,225 916 2,309 3,384 749 2,635 Husband's education 4.745 4.785 4.732 0.053 3.167 3.827 2.904 0.923*** 6.256 5.962 6.339 -0.377* 0.063 0.128 0.073 0.146 0.075 0.157 0.084 0.166 0.094 0.203 0.106 0.226 6,576 1,659 4,917 3,216 915 2,301 3,360 744 2,616 Urban 0.512 0.450 0.533 -0.083*** 0.006 0.012 0.007 0.014 6,611 1,665 4,946 Age at first birth 19.897 19.910 19.892 0.017 19.952 19.964 19.947 0.017 19.845 19.845 19.845 0.000 0.054 0.113 0.062 0.126 0.080 0.152 0.094 0.178 0.074 0.169 0.082 0.179 5,905 1,472 4,433 2,845 800 2,045 3,060 672 2,388 Turkey Turkey Rural Turkey Urban Total Treatment Control Difference Total Treatment Control Difference Total Treatment Control Difference Age 29.175 30.117 28.855 1.262 28.822 30.682 28.258 2.424*** 29.352 29.865 29.169 0.697** 0.105 0.207 0.121 1.262 0.186 0.380 0.211 0.437 0.127 0.246 0.148 0.287 8,576 2,170 6,406 2,874 669 2,205 5,702 1,501 4,201 Dead 0.207 0.151 0.226 -0.075*** 0.260 0.166 0.288 -0.122*** 0.181 0.144 0.194 -0.050*** 0.007 0.011 0.008 0.016 0.014 0.021 0.017 0.032 0.008 0.013 0.010 0.018 8,576 2,170 6,406 2,874 669 2,205 5,702 1,501 4,201 Education 5.601 6.478 5.303 1.175*** 4.492 5.667 4.136 1.531*** 6.159 6.840 5.916 0.924*** 0.043 0.081 0.050 0.098 0.059 0.115 0.067 0.138 0.056 0.104 0.066 0.126 8,576 2,170 6,406 2,874 669 2,205 5,702 1,501 4,201 Husband's education 6.977 7.187 6.902 0.285** 5.687 6.020 5.574 0.446*** 7.594 7.711 7.551 0.160 0.051 0.096 0.059 0.115 0.070 0.144 0.079 0.160 0.065 0.120 0.077 0.146 6,135 1,623 4,512 1,985 503 1,482 4,150 1,120 3,030 Urban 0.665 0.692 0.656 0.036*** 0.005 0.010 0.006 0.012 8,576 2,170 6,406 Age at first birth 20.423 21.001 20.219 0.783*** 20.143 20.851 19.895 0.956*** 20.561 21.074 20.378 0.696*** 0.051 0.099 0.059 0.116 0.086 0.159 0.101 0.195 0.063 0.125 0.073 0.144 5,578 1,458 4,120 1,838 477 1,361 3,740 981 2,759 27 Table 3: Difference-in-difference reduced form estimation of the impact of the earthquake on the number of children ever born Total Rural Urban Before After Difference Before After Difference Before After Difference Treatments 2.812 2.594 -0.218 Treatments 3.078 2.918 -0.160 Treatments 2.555 2.307 -0.247 0.026 0.025 0.037 0.039 0.041 0.056 0.036 0.031 0.047 Controls 2.990 2.213 -0.777 Controls 3.130 2.440 -0.690 Controls 2.718 1.951 -0.767 0.007 0.006 0.009 0.009 0.009 0.012 0.011 0.008 0.014 D at a point in time -0.178 0.381 D at a point in time -0.052 0.478 D at a point in time -0.164 0.356 0.026 0.025 0.038 0.039 0.036 0.031 D-D 0.560*** D-D 0.530*** D-D 0.520*** (0.036) (0.054) (0.047) Observations 247,973 Observations 144,065 Observations 103,908 India India Rural India Urban Before After Difference Before After Difference Before After Difference Treatments 2.859 2.111 -0.748 Treatments 3.057 2.308 -0.750 Treatments 2.597 1.847 -0.750 0.031 0.031 0.044 0.043 0.043 0.061 0.043 0.042 0.060 Controls 2.983 2.063 -0.920 Controls 3.122 2.296 -0.826 Controls 2.667 1.789 -0.878 0.007 0.006 0.009 0.009 0.009 0.012 0.011 0.008 0.014 D at a point in time -0.124 0.048 D at a point in time -0.064 0.012 D at a point in time -0.070 0.059 0.035 0.034 0.048 0.048 0.047 0.047 D-D 0.172*** D-D 0.0765 D-D 0.128* (0.049) (0.068) (0.066) Observations 214,688 Observations 129,672 Observations 85,016 Pakistan Pakistan Rural Pakistan Urban Before After Difference Before After Difference Before After Difference Treatments 4.150 4.090 -0.059 Treatments 4.009 4.160 0.152 Treatments 4.322 3.952 -0.370 0.073 0.069 0.100 0.100 0.085 0.131 0.107 0.116 0.158 Controls 4.137 3.852 -0.285 Controls 4.056 3.930 -0.126 Controls 4.207 3.730 -0.477 0.042 0.032 0.052 0.062 0.042 0.075 0.057 0.047 0.073 D at a point in time 0.013 0.239 D at a point in time -0.048 0.230 D at a point in time 0.115 0.222 0.084 0.074 0.116 0.095 0.121 0.118 D-D 0.226** D-D 0.278* D-D 0.108 (0.11) (0.15) (0.17) Observations 16,634 Observations 9,420 Observations 7,214 Turkey Turkey Rural Turkey Urban Before After Difference Before After Difference Before After Difference Treatments 1.701 2.171 0.470 Treatments 1.871 2.256 0.384 Treatments 1.626 2.156 0.531 0.038 0.032 0.049 0.071 0.075 0.112 0.044 0.035 0.055 Controls 2.201 3.026 0.825 Controls 2.383 3.608 1.225 Controls 2.105 2.771 0.665 0.031 0.031 0.044 0.058 0.066 0.088 0.036 0.033 0.049 D at a point in time -0.500 -0.855 D at a point in time -0.511 -1.352 D at a point in time -0.480 -0.615 0.057 0.053 0.113 0.151 0.065 0.053 D-D -0.355*** D-D -0.841*** D-D -0.135 (0.078) (0.19) (0.084) Observations 16,651 Observations 4,973 Observations 11,678 28 Table 4a: Regression results (1) (2) (3) (4) (5) (6) (7) (8) Dependent variable: Number of children ever born Post -0.777*** -0.540*** -0.403*** -0.226*** -0.249*** -0.133*** -0.146*** -0.146*** (0.017) (0.012) (0.014) (0.011) (0.0100) (0.010) (0.010) (0.010) Near -0.178*** -0.103** -0.134*** -0.0922*** -0.416*** 1.659*** 1.647*** 1.674*** (0.056) (0.044) (0.047) (0.034) (0.038) (0.051) (0.052) (0.052) Post * near 0.560*** 0.468*** 0.279*** 0.165*** 0.158*** 0.0760** 0.0749** 0.0803** (0.076) (0.061) (0.063) (0.046) (0.033) (0.035) (0.035) (0.035) Number of children ever died 1.689*** 1.290*** 1.131*** 1.088*** 0.889*** 0.873*** 0.864*** (0.0079) (0.0070) (0.0061) (0.0058) (0.0053) (0.0070) (0.0070) Age of mother 0.118*** 0.112*** 0.115*** 0.107*** 0.107*** 0.107*** (0.00059) (0.00055) (0.00055) (0.00067) (0.00067) (0.00067) Number of years of schooling of the mother -0.106*** -0.0896*** -0.0514*** -0.0513*** -0.0380*** (0.00082) (0.00081) (0.00089) (0.00088) (0.00097) Urban dummy -0.0665*** -0.0865*** -0.0866*** -0.0731*** (0.0098) (0.011) (0.011) (0.011) Age at first birth -0.137*** -0.137*** -0.137*** (0.0011) (0.0011) (0.0011) Number of years of schooling of the father -0.0200*** (0.00090) Post*number of children ever died 0.0345*** 0.0341*** (0.011) (0.011) Near*number of children ever died 0.0190 0.0187 (0.031) (0.031) Post*near*number of children ever died 0.0125 0.00479 (0.047) (0.046) Regional fixed effects yes yes yes yes yes yes yes yes Countries within sample India, India, India, India, India, India, India, India, Pakistan, Pakistan, Pakistan, Pakistan, Pakistan, Pakistan, Pakistan, Pakistan, Turkey Turkey Turkey Turkey Turkey Turkey Turkey Turkey Observations 247,973 247,973 247,973 247,921 247,921 193,076 193,076 191,910 R-squared 0.03 0.32 0.54 0.59 0.61 0.59 0.59 0.59 Notes: Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1 Post is the year after the earthquake; near is the region in which the earthquake occurred. Clusters are unique across countries and within countries across years. Regions are unique within a country and consistent across time within a country. Regional fixed effects are controlled for in each regression. Clusting of the standard errors at the cluster level. Table 4b: Regression results (8) (9) (10) (11) (12) (13) India, India, Never Turkey, Never Pakistan, India Pakistan Turkey Moved Moved Sample Turkey Dependent variable: Number of children ever born Post -0.146*** -0.0510*** -0.142*** -0.232*** -0.184*** -0.232*** (0.010) (0.013) (0.024) (0.043) (0.042) (0.059) Near 1.674*** 1.033*** 0.234*** 2.142*** 2.410*** -0.589*** (0.052) (0.052) (0.081) (0.093) (0.085) (0.094) Post * near 0.0803** -0.0920 0.0221 0.184** 0.214*** 0.293*** (0.035) (0.057) (0.12) (0.082) (0.061) (0.10) Number of children ever died 0.864*** 0.908*** 0.872*** 0.777*** 1.078*** 1.085*** (0.0070) (0.0075) (0.023) (0.027) (0.041) (0.072) Age of mother 0.107*** 0.112*** 0.0953*** 0.209*** 0.100*** 0.105*** (0.00067) (0.00067) (0.0014) (0.0024) (0.0024) (0.0036) Number of years of schooling of the mother -0.0380*** -0.0487*** -0.0302*** -0.0422*** -0.0819*** -0.0794*** (0.00097) (0.0011) (0.0023) (0.0039) (0.0040) (0.0080) Urban dummy -0.0731*** -0.0563*** -0.0941*** -0.0384 -0.193*** -0.322*** (0.011) (0.014) (0.024) (0.035) (0.043) (0.059) Age at first birth -0.137*** -0.0771*** -0.124*** -0.226*** -0.129*** -0.122*** (0.0011) (0.0016) (0.0023) (0.0040) (0.0037) (0.0063) Number of years of schooling of the father -0.0200*** -0.0148*** -0.0164*** -0.0115*** -0.0375*** -0.0364*** (0.00090) (0.00098) (0.0022) (0.0032) (0.0038) (0.0069) Post*number of children ever died 0.0341*** -0.0251** -0.0172 -0.0365 0.181*** 0.176* (0.011) (0.012) (0.029) (0.036) (0.059) (0.10) Near*number of children ever died 0.0187 -0.160*** -0.0357 0.0341 -0.0660 -0.329** (0.031) (0.031) (0.097) (0.065) (0.087) (0.15) Post*near*number of children ever died 0.00479 0.0709 -0.0326 -0.0333 -0.261** 0.00598 (0.046) (0.059) (0.20) (0.086) (0.12) (0.27) Regional fixed effects yes yes yes yes yes yes Countries within sample India, India India Pakistan Turkey Turkey Pakistan, Turkey Observations 191,910 164,367 29,231 14,637 12,906 3,981 R-squared 0.59 0.88 0.53 0.90 0.89 0.65 Notes: Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1 Post is the year after the earthquake; near is the region in which the earthquake occurred. Clusters are unique across countries and within countries across years. Regions are unique within a country and consistent across time within a country. Regional fixed effects are controlled for in each regression. Clusting of the standard errors at the cluster level. 30 Table 5: Robustness check ­ differential trends between treatment and control (1) (2) (3) (4) (5) (6) Number of years Number of years Number of Age of mother of schooling of Urban dummy Age at first birth of schooling of children ever died Interaction variable: x the mother the father Dependent variable: Number of children ever born Post -0.146*** 0.338*** -0.112*** -0.0899*** -0.173*** -0.0829*** (0.010) (0.036) (0.015) (0.013) (0.042) (0.017) Near 1.674*** 1.712*** 1.693*** 1.665*** 1.843*** 1.717*** (0.052) (0.14) (0.054) (0.056) (0.12) (0.057) Post * near 0.0803** -0.203 0.111** 0.0794 0.136 0.109* (0.035) (0.17) (0.052) (0.054) (0.16) (0.063) Number of children ever died 0.864*** 0.878*** 0.881*** 0.880*** 0.880*** 0.881*** (0.0070) (0.0053) (0.0053) (0.0053) (0.0053) (0.0053) Age of mother 0.107*** 0.115*** 0.107*** 0.107*** 0.107*** 0.107*** (0.00067) (0.00094) (0.00067) (0.00067) (0.00067) (0.00067) Number of years of schooling of the mother -0.0380*** -0.0386*** -0.0342*** -0.0381*** -0.0382*** -0.0380*** (0.00097) (0.00097) (0.0013) (0.00097) (0.00097) (0.00097) Urban dummy -0.0731*** -0.0731*** -0.0737*** -0.0132 -0.0731*** -0.0736*** (0.011) (0.011) (0.011) (0.016) (0.011) (0.011) Age at first birth -0.137*** -0.137*** -0.137*** -0.137*** -0.137*** -0.137*** (0.0011) (0.0011) (0.0011) (0.0011) (0.0015) (0.0011) Number of years of schooling of the father -0.0200*** -0.0197*** -0.0201*** -0.0201*** -0.0200*** -0.0155*** (0.00090) (0.00090) (0.00090) (0.00090) (0.00090) (0.0012) Post*x 0.0341*** -0.0144*** -0.00538*** -0.111*** 0.00198 -0.00771*** (0.011) (0.0012) (0.0017) (0.020) (0.0019) (0.0016) Near*x 0.0187 -0.000919 -0.0111** 0.00751 -0.00805 -0.00672 (0.031) (0.0041) (0.0045) (0.050) (0.0055) (0.0045) Post*near*x 0.00479 0.00864 -0.00569 0.0234 -0.00271 -0.00375 (0.046) (0.0055) (0.0061) (0.070) (0.0073) (0.0062) Regional fixed effects yes yes yes yes yes yes Countries within sample India, Pakistan, India, Pakistan, India, Pakistan, India, Pakistan, India, Pakistan, India, Pakistan, Turkey Turkey Turkey Turkey Turkey Turkey Observations 191910 191910 191910 191910 191910 191910 R-squared 0.59 0.59 0.59 0.59 0.59 0.59 Notes: Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1 Post is the year after the earthquake; near is the region in which the earthquake occurred. Clusters are unique across countries and within countries across years. Regions are unique within a country and consistent across time within a country. Regional fixed effects are controlled for in each regression. Clusting of the standard errors at the cluster level. 31 Figure 1: Robustness check: Randomizing the placement of the earthquake ­ t-statistic on the D-D variable, post*near 5.00 4.00 3.00 2.00 1.00 t-statistic 0.00 -1.00 -2.00 -3.00 -4.00 32