This paper highlights one aspect of the enormous but little-exploited potential of the Demographic and Health Surveys programme, namely, the use of data on siblings. Such data can be used to control for familylevel unobserved heterogeneity that might confound the relationship of interest, and to study correlations in sibling outcomes. It also discusses potential problems associated with the sibling data being derived from the retrospective fertility histories of mothers.
NATIONAL FAMILY HEALTH SURVEY-3Economic & Political Weekly EPW november 29, 200839Sonia Bhalotra (s.bhalotra@bristol.ac.uk) is at the Department of Economics and Centre for Market and Public Organisation, University of Bristol.Sibling-Linked Data in the Demographic and Health SurveysSonia BhalotraThis paper highlights one aspect of the enormous but little-exploited potential of the Demographic and Health Surveys programme, namely, the use of data on siblings. Such data can be used to control for family-level unobserved heterogeneity that might confound the relationship of interest, and to study correlations in sibling outcomes. It also discusses potential problems associated with the sibling data being derived from the retrospective fertility histories of mothers.Iwish to highlight and illustrate an aspect of the enormous and little-exploited potential of the National Family Health Survey (NFHS) and its sister, the Demographic and Health Surveys (DHS) programme. My objective is to encourage research-ers to develop a range of uses of the DHS data in other contexts and across academic disciplines. 1 IntroductionTheNFHS is one of a family of about 200DHS studies conducted in some 75 developing and transition economies (see www.meas-uredhs.com). The unexploited potential that I focus on in this paper pertains to the use of sibling-linked data.1 These can be extracted from the retrospective birth histories of mothers, a cen-trepiece of the DHS surveys. Section 2 describes in more detail how the data can be constructed and what use they can be put to. Sections 3 and 4 illustrate two applications of these data, both of which use the NFHS. One uses sibling data to control for the (endogenous) composition of births in studying the effects of business cycle fluctuations on infant mortality. The other is designed to understand the clustering of infant mortality amongst siblings and, in particular, to identify the extent to which this reflects causal processes like birth spacing over which policy interventions may have some sway, as opposed to predetermined and thereby less amenable family-level traits. Section 5 discusses potential limitations of the data and Section 6 concludes. 2 Data Structure and UsesTheDHS interview women of reproductive age, most often, age 15 to 49. The women record a complete retrospective history of their births and siblings are easily identified by virtue of having a com-mon mother. Since there are sequential observations of births for each mother, there is effectively a “panel” of children within the mother. We may think of the mother as the cross-sectional unit and of her children, born in different years, as presenting the time dimension of the panel. The panel is “unbalanced” in the sense that the time window of the panel, which is the span of years in which children were born, varies across mothers. These sorts of data are immensely useful in identifying causal effects.A significant advantage is that they can be used to net out the effects of potentially confounding unobservables at the mother (or family) level such as ability, frailty, tastes or attitudes (unobserved heterogeneity). The statistical model can be speci-fied to use either random effects or fixed effects (for a discussion of these alternatives, see Hsiao 2003). As these procedures rely on variation that is common to siblings, they do not use informa-tion from mothers in the sample who record just one child. It therefore becomes relevant to confirm that the effective selection
NATIONAL FAMILY HEALTH SURVEY-3november 29, 2008 EPW Economic & Political Weekly40of mothers with at least two children is not biasing the results. Developing countries, which dominate the DHS sample, have rel-atively low age at first birth and relatively high fertility rates. Pooling information on births across the available DHS surveys, I find that the average age at first birth is 19.52 (sd 3.93) and that 80.5% of mothers have more than one child. The mean number of children in the sample of families with at least two children is 4.26. As a result, panel data estimates are relatively well deter-mined. I am aware of three studies that have exploited the DHS data to control for mother-level heterogeneity. Bhalotra (2007a) uses the NFHS to identify the effects of aggregate income shocks on infant mortality; a more detailed discussion follows. Kudamatsu (2008) uses the African DHS data sets to identify the effect of democracy on infant mortality. Bhalotra and Steele (work in progress) use a handful of DHS data sets to examine the effects of business cycle variation on the timing of fertility. The first two studies use mother fixed effects and the third uses mother random effects.A further advantage of linked data on siblings is that they can be used to investigate correlations in outcomes amongst siblings. Rajan and James (in this issue) look at the correlation in nutri-tional outcomes of siblings as an indicator of data quality. Bhalotra (2008) studies the correlation of gender of births within mothers. Oettinger (2000) identifies causal effects of an individ-ual’s schooling on the schooling attainment of his or her younger sibling, after allowing for shared traits amongst siblings. In a series of papers, Arulampalam and Bhalotra (2006, 2008) and Bhalotra and van Soest (2008) analyse causal and correlated effects in the clustering of death amongst siblings. They estimate dynamic models with unobserved heterogeneity that involve iso-lating from the correlated traits that siblings share on account of having the same mother and environment (unobserved heteroge-neity), and the causal effect of an outcome for one child on the outcome for his or her succeeding sibling (this dynamic effect is referred to as state dependence or scarring). In consistent estima-tion of such models, an especially useful feature of the sibling linked data in theDHS is that they include information on every birth, including the first. This is relevant because it helps researchers address the “initial conditions problem” (Heckman 1981). This is best explained with an illustration (see section 4). 3 RemovingConfoundingMother-LevelUnobservablesThis section illustrates the usefulness of mother fixed effects esti-mation in analysing the effects of economic fluctuations on infant survival (for more details, see Bhalotra 2007a). The common expectation is that infant mortality is higher in recessions (when incomes fall) and lower in booms (when incomes rise). But what if some people anticipate this and avert birth in recessions? If at all, it seems that the people most likely to behave in this way are poor people whose ability to buffer a dip in income is limited. Indeed richer women may behave in the opposite way, preferring to give birth in recessions, when job opportunities turn scarce. In a scenario in which recessions are associated with a decline in the share of births contributed by poor women, the average birth will face a lower risk of infant death, other things being equal.2 As a result, infant mortality may appear to decline in recessions, contrary to popular belief. This contrary effect is a compositional (not causal) effect. There is another more autonomous reason that the data may show a decline in infant mortality in reces-sions. To understand this, note that infant mortality is typically measured as the proportion of live births in a year that do not survive to the age of one. If recessions affect maternal health and result in an increased risk of miscarriage or stillbirth, and if this effect is largest amongst poorer women, then again poor women will contribute disproportionately fewer births in a recession.The upshot is that if we want to identify the causal effect of income changes on infant mortality, part of the estimation proce-dure must involve controlling for the composition (or selection) effects described. To the extent that fertility preferences or mis-carriage risk are constant (fixed) for a given woman, this can be done by use of linked sibling data. In particular, we can compare the risk of infant death of siblings, one born in a recession and one not. A boom (recession) is defined as a positive (negative) annual change in net state domestic product per capita deflated by the consumer price index for agricultural workers. The reason to study the effects of annual changes in income is that longer-range growth is often entangled with a lot of other change—in techno-logy, education, infrastructure, and political and social institu-tions. These other things evolve sluggishly; so by looking at annual changes, we can expect to distinguish them. I constructed data on infant mortality for linked siblings using data from the secondNFHS (NFHS-2), conducted in 1998-99. The sample I ana-lysed contains information on more than 152,000 children born to around 50,000 mothers in 15 states during 1970-97. I merged these data by state and birth-year of the child with time-series data on state income, social expenditure, rainfall, etc. The length of the state panel (28 years, 1970-97) aids identification, making it more likely that there are independent macroeconomic fluctua-tions within states; indeed the standard deviation of within-state income variation in these data is almost identical to that of the between-state variation.In the rural sample, 12% of mothers have only one child, and these children account for 3.7% of all children in the sample. Using a specification that includes observable mother traits, I compared the effect of income on mortality risk in the full sample with the effect obtained in a sample restricted to mothers with at least two births, and found that they were identical. This con-firms that the effective selection of mothers with at least two children is not, in this case, biasing the results.I find some evidence that births among high-risk women—in particular, uneducated and scheduled tribe women in rural areas —are under-represented in recessions. Controlling for the com-position of births by using mother fixed effects, I find that reces-sions increase rural infant mortality in India. The estimates imply that a negative income shock of median size (4.4%) will raise infant mortality by 0.136 percentage points. This is almost half the total annual decline in mortality in India in 1970-99 (which I estimate at about 0.3 percentage points per annum). The median positive income shock in these data, at 5.8%, is larger, so the sim-ulated beneficial effects of booms on infant survival chances are accordingly larger. The effects of income shocks on lifetime
NATIONAL FAMILY HEALTH SURVEY-3Economic & Political Weekly EPW november 29, 200841health will tend to be even greater since (where children survive income shocks in childhood) early exposure to poor living condi-tions has lasting adverse effects on health (van der Berg et al 2006; Banerjee et al 2007; Deaton 2007; Bhalotra 2007b). The effects of recessions are not evenly distributed. The most vulnerable are rural households in which the mother is unedu-cated or had her first birth when she was a teenager. Within households, girls are much more likely to die in a downturn than their brothers, reinforcing previous findings that girls’ welfare is put second to that of boys in lean times (Behrman and Deolalikar 1989; Rose 1999). I investigated potential mechanisms using data on recent births in the first two rounds of the NFHS. Less than a fifth of mothers in these data contribute more than one child to the sam-ple, and they are a select group who have shorter birth intervals than average. For this leg of the analysis, I therefore simply pool the cross-sectional data and cluster the standard errors by mother to account for non-independence of the residuals for the subset of siblings in the data. I find that delivery outside the home, antena-tal care, child vaccinations and the probability of treatment for infectious diseases among children are lower in economic down-turns. This holds even after I control for declines in the supply of public services (proxied by data on state health and development expenditure) in downturns, suggesting that the demand for health services is lower. This is consistent with lower earnings in a downturn, but I show that it is also because mothers are work-ing harder and do not have as much time to seek healthcare. So it seems that households use maternal labour supply as an insur-ance mechanism (with most of the additional work taken on by women being in agriculture). This imposes a cost in terms of the health of their children that has not been sufficiently recognised. There is an interesting contrast here with recent results for richer countries, where women’s work is thought to be procycli-cal (higher in upturns), and this is hypothesised to contribute to the seemingly counter-intuitive finding that infant mortality is procyclical (see Dehejia and Lleras-Muney 2004). Overall, the results suggest a need for mechanisms that shield the vulnerable from temporary falls in wages or increases in unemployment (of main earners), both conditions that we find have irreversible con-sequences. They show that (temporary) increases in labour force participation among relatively poor women may signal distress rather than “liberation” and have unintended consequences for child survival and health. The analysis illuminates a question of long-standing policy interest: To what extent is income (or poverty) an ultimate cause of childhood death in poor countries? This question is informa-tive of the welfare effects of economic growth, and of the effec-tiveness of cash transfers made to poor households. We may expect growth in aggregate income (GDP) to lower mortality if it (a) raises private incomes, especially of the poor, so that they can acquire more nutrition and other health inputs, and (b) raises state social expenditure. However, the evidence on the effective-ness of income in improving survival is not overwhelming. His-torical evidence suggests that secular improvements in medical technology, public services and education were more important than income growth in bringing about sustained mortality decline (Cutler et al 2006). And as discussed, recent studies of theUS and other developed countries show that mortality risks – for adults and children – are lowerin recessions (Ruhm 2000). Against this backdrop, the analysis of income and substitution effects in the poorer setting of India is pertinent.4 IdentifyingDynamicEffects In this section, I consider an application in which the question of interest concerns the effect on a child of an outcome for his or her preceding sibling (what was earlier referred to as state depend-ence or scarring). The outcome of interest is infant or neonatal death. The phenomenon of interest is the clustering or concentra-tion of death risk among siblings (Curtis et al 1993; Zenger 1993). To illustrate the importance of this phenomenon, consider the following figures derived from the NFHS-2 for Uttar Pradesh. Among second and higher-order children, the average probabil-ity of neonatal death is 5.2% in the sub-sample in which the pre-vious sibling survived the neonatal period. In contrast, in the sub-sample in which the previous sibling suffered neonatal death, this probability is a remarkable 18.80%. So the death of a preced-ing sibling is associated with a more than threefold increase in mortality risk.A part of the reason for sibling death clustering is no doubt that families are different, with some being more effective at averting child death than others. For example, mothers in the more effec-tive families are probably more educated, more aware and innately healthier than in the less effective families. TheDHS (and most data sets) record education but they do not record awareness or frailty. To comprehensively capture these “unob-servables”, we can use mother-level fixed or random effects, and we choose to use random effects (see Arulampalam and Bhalotra (2006, 2008); Bhalotra and van Soest 2008). Now consider how and why dynamics come in to this picture. An interesting question for policy is the extent to which the event of death of a child causes a higher risk of death for his or her younger sibling. A potential mechanism is as follows. The death of a neonate or infant results in the cessation of breastfeeding. Since breastfeeding delays the return of fecundity following birth, this tends to reduce the birth interval to the next child. And there is a lot of evidence to suggest that short preceding birth intervals elevate mortality risk. So in an appropriate model, neo-natal or infant mortality risk for child j of mother k will depend on the realised neonatal or infant mortality of child j-1 of mother k (state dependence) as well as on all traits specific to mother k (mother-level unobserved heterogeneity). A classical problem with consistent estimation of this sort of model is that the real-ised mortality of the preceding sibling is necessarily correlated with mother-level heterogeneity. This is called the initial condi-tions problem (Heckman 1981), and a good way to attempt to address it is to use information on the first-born child in each family. Intuitively, the problem is that the model describes a dynamic process and we need to allow for how it starts. The death risk of the third child in a family depends on whether the second child died and the death risk of the second child depends on whether the first child died, while the first child presumably has no history. So it is important to use information on whether the
NATIONAL FAMILY HEALTH SURVEY-3november 29, 2008 EPW Economic & Political Weekly42first child died or not. In this respect, the availability of complete birth histories of married women in the DHS is a valuable resource. In other applications of dynamic models with unobserved heterogeneity, data on the start of the process are often unavail-able. For example, in studying unemployment spells of indi-viduals, researchers would ideally like to have data on school-leavers but must often make do with truncated retrospective data, that is, data that do not include the first spell of unemploy-ment for each individual. The main findings of this research are as follows. Scarring explains about 40% of the clustering observed in the data in Uttar Pradesh; the corresponding proportions being 14% for West Bengal and 21.5% for Kerala. In a model that allows for scarring, the proportion of the error variance attributable to mother-level unobservables is estimated to be 11% in Uttar Pradesh, 21% in West Bengal and 7% in Kerala. We estimate that eliminating scar-ring would reduce the incidence of infant mortality (among chil-dren born after the first child) by 9.8% in Uttar Pradesh, 6.0% in West Bengal and 5.9% in Kerala (Arulampalam and Bhalotra 2006). Scarring is large and significant in 13 of the 15 major states of India. The two states in which evidence of scarring is weak are Punjab, the richest, and Kerala, the socially most progressive. The size of the scarring effect depends on the gender of the previ-ous child in three of the 15 states, in a direction consistent with the preference for sons (Arulampalam and Bhalotra 2008). The only other covariate with an effect of similar magnitude is mother’s (secondary or higher) education. While there is consid-erable evidence of the effects of maternal education on infant survival, the literature has paid scarce attention to scarring effects. Evidence of scarring implies that policies targeted at reducing infant mortality will have social multiplier effects by helping avoid the death of subsequent siblings. Mechanisms underlying scarring were further investigated for Uttar Pradesh in Bhalotra and van Soest (2008), who exploit the panel nature of the data to model birth spacing and fertility with (neonatal) mor-tality. The tendency for neonatal death to shorten the birth inter-val to the next child explains about a fourth of the total scarring effect. We speculate that some of the residual effect may involve maternal depression. We find direct evidence of replacement behaviour: a child death results in a shortening of the interval to the next birth, and also increases the probability of a next birth. Model simulations imply that accounting for direct and indirect effects, 37 in 100 children who die during the neonatal period are replaced by new births. Of these, about 30 survive. There is no evidence that couples practise hoarding, that is, that they have higher fertility in anticipation of the risks of neonatal death. The estimates of fertility behaviour are consistent with the preference for sons. The estimates suggest that fertility decline in India started in 1981. Despite this, birth intervals have become shorter since about then.5 PotentialProblemsHaving elucidated some of the less recognised or less used advan-tages of the retrospective fertility histories in the DHS, it is relevant to point out some of their limitations. We have highlighted the advantage of having information on all births to a mother, including her first, as this helps address the initial conditions problem. A consequence of this is that the range of birth years in the data is wide, and varies by mother. The long time span for which cohort infant mortality is available is high-lighted as an advantage because it assists in identifying the effects of within-state changes in income on mortality. However, we may be limited in the range of variables that can be associated with information on births in a time-consistent fashion. For instance, even if we have data on infant mortality rates in 1970 from a survey conducted in 1992, we do not have detailed infor-mation on the mother’s (or family’s) circumstances (for example, assets, location, labour force participation) over time, rather, this information pertains to the year of the survey. Location is of particular interest if we want to match state-level factors like political or macroeconomic indicators to outcomes for children born across long periods of time. If the mother migrated between births, we cannot assume that all her children were born in the state that she was in when interviewed. This problem can be assessed if not addressed by using answers to a question in theDHS that asks how long the mother has lived in her current place of residence (place at the time of survey). Using the NFHS-2, I estimate that 86% of children born in 1970-97 in the 15 major states were born in the mother’s current place of residence. This refers to her local location and will record, for example, move-ment between villages within a state. So it is a much stricter con-dition than required. If migration is exogenous to the question of interest, the parameter of interest can be re-estimated on the sub-sample of children for whom the state of birth is known with certainty, and the problem is solved. However, if migration is endogenous, it is not, although this exercise may nevertheless be informative. And it may be useful to estimate the model on the full sample and the restricted sample (of non-migratory mothers), and compare the estimates.Another potential problem with retrospective data is recall bias: the concern that mothers are more likely to forget the incidence or dates of events the further back in time they are. This may be espe-cially pertinent if mothers with different characteristics (for exam-ple, age, and education) do this to different extents. The DHS do have numerous checks built in to ensure the quality of birth his-tory data (see ORC Macro 2006, p14) but researchers might never-theless assess the extent of this problem by, for example, checking the sensitivity of their results to truncation of the data, that is, to restricting the sample to windows of more recent years. This is done, for example, in Arulampalam and Bhalotra (2006) and in Bhalotra (2007a). Another recall-related issue is rounding off in age. For example, data on age at death for children reveal age-heaping at six-month intervals. If one is studying a discrete event such as death by the age of one (infant mortality), a natural sensi-tivity check would be to conduct the analysis with mortality defined inclusive and exclusive of deaths in the 12th month.There are other problems relating to how representative the retrospective sample is (Rindfuss et al 1982). For example, as we go back in calendar time, the births captured in the sample are not only fewer but also disproportionately of relatively young mothers. A woman who gave birth at age 15 in 1965 will be 49 in
NATIONAL FAMILY HEALTH SURVEY-3Economic & Political Weekly EPW november 29, 200839Sonia Bhalotra (s.bhalotra@bristol.ac.uk) is at the Department of Economics and Centre for Market and Public Organisation, University of Bristol.Sibling-Linked Data in the Demographic and Health SurveysSonia BhalotraThis paper highlights one aspect of the enormous but little-exploited potential of the Demographic and Health Surveys programme, namely, the use of data on siblings. Such data can be used to control for family-level unobserved heterogeneity that might confound the relationship of interest, and to study correlations in sibling outcomes. It also discusses potential problems associated with the sibling data being derived from the retrospective fertility histories of mothers.Iwish to highlight and illustrate an aspect of the enormous and little-exploited potential of the National Family Health Survey (NFHS) and its sister, the Demographic and Health Surveys (DHS) programme. My objective is to encourage research-ers to develop a range of uses of the DHS data in other contexts and across academic disciplines. 1 IntroductionTheNFHS is one of a family of about 200DHS studies conducted in some 75 developing and transition economies (see www.meas-uredhs.com). The unexploited potential that I focus on in this paper pertains to the use of sibling-linked data.1 These can be extracted from the retrospective birth histories of mothers, a cen-trepiece of the DHS surveys. Section 2 describes in more detail how the data can be constructed and what use they can be put to. Sections 3 and 4 illustrate two applications of these data, both of which use the NFHS. One uses sibling data to control for the (endogenous) composition of births in studying the effects of business cycle fluctuations on infant mortality. The other is designed to understand the clustering of infant mortality amongst siblings and, in particular, to identify the extent to which this reflects causal processes like birth spacing over which policy interventions may have some sway, as opposed to predetermined and thereby less amenable family-level traits. Section 5 discusses potential limitations of the data and Section 6 concludes. 2 Data Structure and UsesTheDHS interview women of reproductive age, most often, age 15 to 49. The women record a complete retrospective history of their births and siblings are easily identified by virtue of having a com-mon mother. Since there are sequential observations of births for each mother, there is effectively a “panel” of children within the mother. We may think of the mother as the cross-sectional unit and of her children, born in different years, as presenting the time dimension of the panel. The panel is “unbalanced” in the sense that the time window of the panel, which is the span of years in which children were born, varies across mothers. These sorts of data are immensely useful in identifying causal effects.A significant advantage is that they can be used to net out the effects of potentially confounding unobservables at the mother (or family) level such as ability, frailty, tastes or attitudes (unobserved heterogeneity). The statistical model can be speci-fied to use either random effects or fixed effects (for a discussion of these alternatives, see Hsiao 2003). As these procedures rely on variation that is common to siblings, they do not use informa-tion from mothers in the sample who record just one child. It therefore becomes relevant to confirm that the effective selection 
NATIONAL FAMILY HEALTH SURVEY-3november 29, 2008 EPW Economic & Political Weekly40of mothers with at least two children is not biasing the results. Developing countries, which dominate the DHS sample, have rel-atively low age at first birth and relatively high fertility rates. Pooling information on births across the available DHS surveys, I find that the average age at first birth is 19.52 (sd 3.93) and that 80.5% of mothers have more than one child. The mean number of children in the sample of families with at least two children is 4.26. As a result, panel data estimates are relatively well deter-mined. I am aware of three studies that have exploited the DHS data to control for mother-level heterogeneity. Bhalotra (2007a) uses the NFHS to identify the effects of aggregate income shocks on infant mortality; a more detailed discussion follows. Kudamatsu (2008) uses the African DHS data sets to identify the effect of democracy on infant mortality. Bhalotra and Steele (work in progress) use a handful of DHS data sets to examine the effects of business cycle variation on the timing of fertility. The first two studies use mother fixed effects and the third uses mother random effects.A further advantage of linked data on siblings is that they can be used to investigate correlations in outcomes amongst siblings. Rajan and James (in this issue) look at the correlation in nutri-tional outcomes of siblings as an indicator of data quality. Bhalotra (2008) studies the correlation of gender of births within mothers. Oettinger (2000) identifies causal effects of an individ-ual’s schooling on the schooling attainment of his or her younger sibling, after allowing for shared traits amongst siblings. In a series of papers, Arulampalam and Bhalotra (2006, 2008) and Bhalotra and van Soest (2008) analyse causal and correlated effects in the clustering of death amongst siblings. They estimate dynamic models with unobserved heterogeneity that involve iso-lating from the correlated traits that siblings share on account of having the same mother and environment (unobserved heteroge-neity), and the causal effect of an outcome for one child on the outcome for his or her succeeding sibling (this dynamic effect is referred to as state dependence or scarring). In consistent estima-tion of such models, an especially useful feature of the sibling linked data in theDHS is that they include information on every birth, including the first. This is relevant because it helps researchers address the “initial conditions problem” (Heckman 1981). This is best explained with an illustration (see section 4). 3 RemovingConfoundingMother-LevelUnobservablesThis section illustrates the usefulness of mother fixed effects esti-mation in analysing the effects of economic fluctuations on infant survival (for more details, see Bhalotra 2007a). The common expectation is that infant mortality is higher in recessions (when incomes fall) and lower in booms (when incomes rise). But what if some people anticipate this and avert birth in recessions? If at all, it seems that the people most likely to behave in this way are poor people whose ability to buffer a dip in income is limited. Indeed richer women may behave in the opposite way, preferring to give birth in recessions, when job opportunities turn scarce. In a scenario in which recessions are associated with a decline in the share of births contributed by poor women, the average birth will face a lower risk of infant death, other things being equal.2 As a result, infant mortality may appear to decline in recessions, contrary to popular belief. This contrary effect is a compositional (not causal) effect. There is another more autonomous reason that the data may show a decline in infant mortality in reces-sions. To understand this, note that infant mortality is typically measured as the proportion of live births in a year that do not survive to the age of one. If recessions affect maternal health and result in an increased risk of miscarriage or stillbirth, and if this effect is largest amongst poorer women, then again poor women will contribute disproportionately fewer births in a recession.The upshot is that if we want to identify the causal effect of income changes on infant mortality, part of the estimation proce-dure must involve controlling for the composition (or selection) effects described. To the extent that fertility preferences or mis-carriage risk are constant (fixed) for a given woman, this can be done by use of linked sibling data. In particular, we can compare the risk of infant death of siblings, one born in a recession and one not. A boom (recession) is defined as a positive (negative) annual change in net state domestic product per capita deflated by the consumer price index for agricultural workers. The reason to study the effects of annual changes in income is that longer-range growth is often entangled with a lot of other change—in techno-logy, education, infrastructure, and political and social institu-tions. These other things evolve sluggishly; so by looking at annual changes, we can expect to distinguish them. I constructed data on infant mortality for linked siblings using data from the secondNFHS (NFHS-2), conducted in 1998-99. The sample I ana-lysed contains information on more than 152,000 children born to around 50,000 mothers in 15 states during 1970-97. I merged these data by state and birth-year of the child with time-series data on state income, social expenditure, rainfall, etc. The length of the state panel (28 years, 1970-97) aids identification, making it more likely that there are independent macroeconomic fluctua-tions within states; indeed the standard deviation of within-state income variation in these data is almost identical to that of the between-state variation.In the rural sample, 12% of mothers have only one child, and these children account for 3.7% of all children in the sample. Using a specification that includes observable mother traits, I compared the effect of income on mortality risk in the full sample with the effect obtained in a sample restricted to mothers with at least two births, and found that they were identical. This con-firms that the effective selection of mothers with at least two children is not, in this case, biasing the results.I find some evidence that births among high-risk women—in particular, uneducated and scheduled tribe women in rural areas —are under-represented in recessions. Controlling for the com-position of births by using mother fixed effects, I find that reces-sions increase rural infant mortality in India. The estimates imply that a negative income shock of median size (4.4%) will raise infant mortality by 0.136 percentage points. This is almost half the total annual decline in mortality in India in 1970-99 (which I estimate at about 0.3 percentage points per annum). The median positive income shock in these data, at 5.8%, is larger, so the sim-ulated beneficial effects of booms on infant survival chances are accordingly larger. The effects of income shocks on lifetime 
NATIONAL FAMILY HEALTH SURVEY-3Economic & Political Weekly EPW november 29, 200841health will tend to be even greater since (where children survive income shocks in childhood) early exposure to poor living condi-tions has lasting adverse effects on health (van der Berg et al 2006; Banerjee et al 2007; Deaton 2007; Bhalotra 2007b). The effects of recessions are not evenly distributed. The most vulnerable are rural households in which the mother is unedu-cated or had her first birth when she was a teenager. Within households, girls are much more likely to die in a downturn than their brothers, reinforcing previous findings that girls’ welfare is put second to that of boys in lean times (Behrman and Deolalikar 1989; Rose 1999). I investigated potential mechanisms using data on recent births in the first two rounds of the NFHS. Less than a fifth of mothers in these data contribute more than one child to the sam-ple, and they are a select group who have shorter birth intervals than average. For this leg of the analysis, I therefore simply pool the cross-sectional data and cluster the standard errors by mother to account for non-independence of the residuals for the subset of siblings in the data. I find that delivery outside the home, antena-tal care, child vaccinations and the probability of treatment for infectious diseases among children are lower in economic down-turns. This holds even after I control for declines in the supply of public services (proxied by data on state health and development expenditure) in downturns, suggesting that the demand for health services is lower. This is consistent with lower earnings in a downturn, but I show that it is also because mothers are work-ing harder and do not have as much time to seek healthcare. So it seems that households use maternal labour supply as an insur-ance mechanism (with most of the additional work taken on by women being in agriculture). This imposes a cost in terms of the health of their children that has not been sufficiently recognised. There is an interesting contrast here with recent results for richer countries, where women’s work is thought to be procycli-cal (higher in upturns), and this is hypothesised to contribute to the seemingly counter-intuitive finding that infant mortality is procyclical (see Dehejia and Lleras-Muney 2004). Overall, the results suggest a need for mechanisms that shield the vulnerable from temporary falls in wages or increases in unemployment (of main earners), both conditions that we find have irreversible con-sequences. They show that (temporary) increases in labour force participation among relatively poor women may signal distress rather than “liberation” and have unintended consequences for child survival and health. The analysis illuminates a question of long-standing policy interest: To what extent is income (or poverty) an ultimate cause of childhood death in poor countries? This question is informa-tive of the welfare effects of economic growth, and of the effec-tiveness of cash transfers made to poor households. We may expect growth in aggregate income (GDP) to lower mortality if it (a) raises private incomes, especially of the poor, so that they can acquire more nutrition and other health inputs, and (b) raises state social expenditure. However, the evidence on the effective-ness of income in improving survival is not overwhelming. His-torical evidence suggests that secular improvements in medical technology, public services and education were more important than income growth in bringing about sustained mortality decline (Cutler et al 2006). And as discussed, recent studies of theUS and other developed countries show that mortality risks – for adults and children – are lowerin recessions (Ruhm 2000). Against this backdrop, the analysis of income and substitution effects in the poorer setting of India is pertinent.4 IdentifyingDynamicEffects In this section, I consider an application in which the question of interest concerns the effect on a child of an outcome for his or her preceding sibling (what was earlier referred to as state depend-ence or scarring). The outcome of interest is infant or neonatal death. The phenomenon of interest is the clustering or concentra-tion of death risk among siblings (Curtis et al 1993; Zenger 1993). To illustrate the importance of this phenomenon, consider the following figures derived from the NFHS-2 for Uttar Pradesh. Among second and higher-order children, the average probabil-ity of neonatal death is 5.2% in the sub-sample in which the pre-vious sibling survived the neonatal period. In contrast, in the sub-sample in which the previous sibling suffered neonatal death, this probability is a remarkable 18.80%. So the death of a preced-ing sibling is associated with a more than threefold increase in mortality risk.A part of the reason for sibling death clustering is no doubt that families are different, with some being more effective at averting child death than others. For example, mothers in the more effec-tive families are probably more educated, more aware and innately healthier than in the less effective families. TheDHS (and most data sets) record education but they do not record awareness or frailty. To comprehensively capture these “unob-servables”, we can use mother-level fixed or random effects, and we choose to use random effects (see Arulampalam and Bhalotra (2006, 2008); Bhalotra and van Soest 2008). Now consider how and why dynamics come in to this picture. An interesting question for policy is the extent to which the event of death of a child causes a higher risk of death for his or her younger sibling. A potential mechanism is as follows. The death of a neonate or infant results in the cessation of breastfeeding. Since breastfeeding delays the return of fecundity following birth, this tends to reduce the birth interval to the next child. And there is a lot of evidence to suggest that short preceding birth intervals elevate mortality risk. So in an appropriate model, neo-natal or infant mortality risk for child j of mother k will depend on the realised neonatal or infant mortality of child j-1 of mother k (state dependence) as well as on all traits specific to mother k (mother-level unobserved heterogeneity). A classical problem with consistent estimation of this sort of model is that the real-ised mortality of the preceding sibling is necessarily correlated with mother-level heterogeneity. This is called the initial condi-tions problem (Heckman 1981), and a good way to attempt to address it is to use information on the first-born child in each family. Intuitively, the problem is that the model describes a dynamic process and we need to allow for how it starts. The death risk of the third child in a family depends on whether the second child died and the death risk of the second child depends on whether the first child died, while the first child presumably has no history. So it is important to use information on whether the 
NATIONAL FAMILY HEALTH SURVEY-3november 29, 2008 EPW Economic & Political Weekly42first child died or not. In this respect, the availability of complete birth histories of married women in the DHS is a valuable resource. In other applications of dynamic models with unobserved heterogeneity, data on the start of the process are often unavail-able. For example, in studying unemployment spells of indi-viduals, researchers would ideally like to have data on school-leavers but must often make do with truncated retrospective data, that is, data that do not include the first spell of unemploy-ment for each individual. The main findings of this research are as follows. Scarring explains about 40% of the clustering observed in the data in Uttar Pradesh; the corresponding proportions being 14% for West Bengal and 21.5% for Kerala. In a model that allows for scarring, the proportion of the error variance attributable to mother-level unobservables is estimated to be 11% in Uttar Pradesh, 21% in West Bengal and 7% in Kerala. We estimate that eliminating scar-ring would reduce the incidence of infant mortality (among chil-dren born after the first child) by 9.8% in Uttar Pradesh, 6.0% in West Bengal and 5.9% in Kerala (Arulampalam and Bhalotra 2006). Scarring is large and significant in 13 of the 15 major states of India. The two states in which evidence of scarring is weak are Punjab, the richest, and Kerala, the socially most progressive. The size of the scarring effect depends on the gender of the previ-ous child in three of the 15 states, in a direction consistent with the preference for sons (Arulampalam and Bhalotra 2008). The only other covariate with an effect of similar magnitude is mother’s (secondary or higher) education. While there is consid-erable evidence of the effects of maternal education on infant survival, the literature has paid scarce attention to scarring effects. Evidence of scarring implies that policies targeted at reducing infant mortality will have social multiplier effects by helping avoid the death of subsequent siblings. Mechanisms underlying scarring were further investigated for Uttar Pradesh in Bhalotra and van Soest (2008), who exploit the panel nature of the data to model birth spacing and fertility with (neonatal) mor-tality. The tendency for neonatal death to shorten the birth inter-val to the next child explains about a fourth of the total scarring effect. We speculate that some of the residual effect may involve maternal depression. We find direct evidence of replacement behaviour: a child death results in a shortening of the interval to the next birth, and also increases the probability of a next birth. Model simulations imply that accounting for direct and indirect effects, 37 in 100 children who die during the neonatal period are replaced by new births. Of these, about 30 survive. There is no evidence that couples practise hoarding, that is, that they have higher fertility in anticipation of the risks of neonatal death. The estimates of fertility behaviour are consistent with the preference for sons. The estimates suggest that fertility decline in India started in 1981. Despite this, birth intervals have become shorter since about then.5 PotentialProblemsHaving elucidated some of the less recognised or less used advan-tages of the retrospective fertility histories in the DHS, it is relevant to point out some of their limitations. We have highlighted the advantage of having information on all births to a mother, including her first, as this helps address the initial conditions problem. A consequence of this is that the range of birth years in the data is wide, and varies by mother. The long time span for which cohort infant mortality is available is high-lighted as an advantage because it assists in identifying the effects of within-state changes in income on mortality. However, we may be limited in the range of variables that can be associated with information on births in a time-consistent fashion. For instance, even if we have data on infant mortality rates in 1970 from a survey conducted in 1992, we do not have detailed infor-mation on the mother’s (or family’s) circumstances (for example, assets, location, labour force participation) over time, rather, this information pertains to the year of the survey. Location is of particular interest if we want to match state-level factors like political or macroeconomic indicators to outcomes for children born across long periods of time. If the mother migrated between births, we cannot assume that all her children were born in the state that she was in when interviewed. This problem can be assessed if not addressed by using answers to a question in theDHS that asks how long the mother has lived in her current place of residence (place at the time of survey). Using the NFHS-2, I estimate that 86% of children born in 1970-97 in the 15 major states were born in the mother’s current place of residence. This refers to her local location and will record, for example, move-ment between villages within a state. So it is a much stricter con-dition than required. If migration is exogenous to the question of interest, the parameter of interest can be re-estimated on the sub-sample of children for whom the state of birth is known with certainty, and the problem is solved. However, if migration is endogenous, it is not, although this exercise may nevertheless be informative. And it may be useful to estimate the model on the full sample and the restricted sample (of non-migratory mothers), and compare the estimates.Another potential problem with retrospective data is recall bias: the concern that mothers are more likely to forget the incidence or dates of events the further back in time they are. This may be espe-cially pertinent if mothers with different characteristics (for exam-ple, age, and education) do this to different extents. The DHS do have numerous checks built in to ensure the quality of birth his-tory data (see ORC Macro 2006, p14) but researchers might never-theless assess the extent of this problem by, for example, checking the sensitivity of their results to truncation of the data, that is, to restricting the sample to windows of more recent years. This is done, for example, in Arulampalam and Bhalotra (2006) and in Bhalotra (2007a). Another recall-related issue is rounding off in age. For example, data on age at death for children reveal age-heaping at six-month intervals. If one is studying a discrete event such as death by the age of one (infant mortality), a natural sensi-tivity check would be to conduct the analysis with mortality defined inclusive and exclusive of deaths in the 12th month.There are other problems relating to how representative the retrospective sample is (Rindfuss et al 1982). For example, as we go back in calendar time, the births captured in the sample are not only fewer but also disproportionately of relatively young mothers. A woman who gave birth at age 15 in 1965 will be 49 in 
