Food Policy 36 (2011) 647655

Contents lists available at ScienceDirect

Food Policy
journal homepage: www.elsevier.com/locate/foodpol

The effect of prime age adult mortality on household composition and consumption in rural Ethiopia
Suneetha Kadiyala a,, Beatrice Rogers b, Agnes Quisumbing a, Patrick Webb b
a b

International Food Policy Research Institute, 2033 K Street, Washington, DC 20006, USA Gerald J. and Dorothy R. Friedman School of Nutrition Science and Policy, Tufts University, 150 Harrison Avenue, Boston, MA 02111, USA

a r t i c l e

i n f o

a b s t r a c t
Using panel data from Ethiopia covering 19941997, we estimate the impact of prime age adult mortality on household composition, household expenditures and dietary diversity. We employed propensity score matching with a difference-in-difference estimator to control for endogeneity of mortality to the outcomes of interest. Households losing a productive adult did not replenish the lost labor, regardless of economic status, sex or status of the deceased adult. With the exception of non-poor households, adult mortality resulted in increased dependency ratios, but did not adversely affect households expenditure patterns (total, food and non-food expenditures) regardless of the sex and position of the deceased and the economic status of the households. Although food expenditures were protected, a decline in dietary diversity, especially among the poorest households, reected increased nutrition insecurity associated with adult mortality. 2011 Elsevier Ltd. All rights reserved.

Article history: Received 10 July 2009 Received in revised form 19 May 2011 Accepted 6 June 2011 Available online 6 July 2011 Keywords: Adult mortality Household composition Household consumption Dietary diversity Propensity score matching Ethiopia

Introduction With the death of an adult member of a household, that household must adjust to the consequent loss of labor and related income, to loss of management skills, and to loss of acquired human capital investment. Loss of an individuals labor and related income may result in changes in the demographic composition of the household and in its expenditure patterns. Although there are recent contributions to this literature (Mason et al., 2010; Linnemayr, 2010; Grimm, 2010; Hosegood, 2009; Chapoto and Jayne, 2008; Beegle et al., 2008), there remains a need for rigorous empirical studies on the magnitude of the effects of prime age adult mortality (PAM) on household and individual welfare outcomes, their magnitude, and variations across different contexts. Understanding these effects is critical to developing effective ways of responding to adult mortality, especially with the persistence of

Corresponding author. Present address: IFPRI New Delhi Ofce, CG Block, NASC Complex, PUSA, New Delhi 110 012, India. Tel.: +91 11 2584 6565/66/67; fax: +91 11 2584 8008/6572. E-mail address: s.kadiyala@cgiar.org (S. Kadiyala).
0306-9192/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.foodpol.2011.06.003

the HIV/AIDS epidemic in Africa. Building on the literature, this paper evaluates the impact of PAM on household composition, total expenditures, food expenditures, non-food expenditures, and dietary diversity in rural Ethiopia. This paper draws on two strands of literature. The rst ows from the broader literature on consumption smoothing. The theory of full insurance initiated by Arrow (1964) in this literature indicates that the value of consumption at the margin for each household is the same, provided there are institutions that pool risks and that can make Pareto-optimal allocation possible. This implies that the growth in household consumption will respond to the growth in aggregate level consumption but not to idiosyncratic shocks such as illness or transient variation in income. This means that well-functioning risk-sharing institutions will mitigate the consumption effects of idiosyncratic shocks across households within a village. There is evidence that households are able to smooth consumption over short time horizons such as over agricultural seasons (Townsend, 1994). When a family member falls ill, households incur direct costs of illness such as medical expenditures and indirect costs such as loss of labor, income (in case of a productive member falling ill) and lost productivity of those providing care. Studies investigating the impact of illnesses show that households, on average, are able to protect income in response to illnesses (Townsend,

648

S. Kadiyala et al. / Food Policy 36 (2011) 647655

1994; Kochar, 1995; Wagstaff, 2007). But Gertler and Gruber (2002) in Indonesia and Cochrane (1991) in the United States nd that minor illnesses (that is, illnesses of short duration) are insured while major illness of long duration are not fully insured. In Ethiopia, using the 19941997 panel, Skouas and Quisumbing (2005) found that illness did not have a signicant impact on food and non-food expenditures. Using more recent rounds of the ERHS from 1999 to 2004 that covered a longer interval, Dercon et al. (2005) found that illness reduced consumption expenditures by 9%; it is possible that the longer duration panel was able to capture longer-term impacts of illness, especially if impacts of illness expenses such as depletion of assets or forgone income are compounded over time. This strand of literature explicitly acknowledges that households adopt a variety of risk-minimizing and risk-coping strategies to protect themselves from consumption uctuations when faced with shocks (for example, Skouas and Quisumbing, 2005). These strategies include informal transfers and remittances (in cash and in kind) from relatives and friends (Morduch, 1999), labor-sharing, selling assets (Deaton, 1992), shifting to low-risk, low-return agriculture (Rosenzweig and Binswanger, 1993), or obtaining loans from the nancial sector (Udry, 1994). Ex-post responses also include availing of targeted food assistance and formal social-protection interventions. Some of these coping strategies may unintentionally jeopardize the familys future food security, such as divesting assets, incurring debts and pulling children out of school. Rural households in Ethiopia have developed their own risksharing arrangements. In addition to the very common extended family network, iddir (funeral societies), equib (credit associations), senbete (some social gatherings), debo (labor-sharing arrangements), and mahiber (religious gatherings) are some of the most common institutions developed by the communities themselves to mitigate their vulnerability to the effects of shocks (see von Braun et al., 1999). The second strand refers to the specic literature on the impact of PAM on farm households. A general assumption in some of this literature and in popular discussion on PAM, especially related to AIDS, is that it results in higher dependency ratios, severe labor constraints, and increased medical expenditures in case of chronic illness preceding death, funeral costs, thereby leading to increased food insecurity. The net effect of PAM on per capita consumption depends on the extent to which costs and income loss associated with PAM are offset by the reduction of consumption units due to the death of the productive adult. Studies that examine households coping strategies in response to PAM show that households adapt in a number of ways. These include demographic changes such as attracting new productive adults or sending children to other households; relying on informal remittances and transfers; altering production activities to adjust to the labor loss such as reducing the area cultivated; and divesting assets (Gillespie and Kadiyala, 2005; Kadiyala and Chapoto, 2010). PAM may be correlated with individual and household characteristics such as wealth and education that are themselves important determinants of household composition, agricultural production and consumption patterns. As Chapoto and Jayne observe (2008), although the few longitudinal empirical studies measuring impacts of adult mortality on rural households welfare acknowledge that adult death may be endogenous to outcomes, PAM was treated mostly as an exogenous event (Ainsworth and Semali, 2000; Beegle, 2005; Yamano and Jayne, 2004). Kadiyala et al. (2009), Chapoto and Jayne (2008), Donovan and Bailey (2006), and Gertler et al. (2004) explicitly address the problem of endogeneity, further discussed in the following sec-

tions. Despite the popular assertion that PAM increases food insecurity, to date only a handful of studies have investigated the impact of PAM on household consumption patterns while taking into account issues of endogeneity in Ethiopia or elsewhere. The present study uses four rounds of panel data from Ethiopia to investigate the impacts of PAM on household composition, value of total consumption, value of food and non-food consumption and dietary diversity using data on expenditures. It seeks to take into account endogeneity issues and uses PSM with difference-in-difference (DID) estimators to control for observable and (time invariant) unobservable characteristics that might bias the estimated impacts. The term prime age in this paper refers to adults between 15 to 54 years of age, 54 years being the life expectancy at age 15 years in Ethiopia.1

Data We used a three year panel (19941997) of the Ethiopian Rural Household Survey (ERHS),2 a collaborative endeavor by the Economics Department of Addis Ababa University, the Centre for the Study of African Economies (CSAE), Oxford University and the International Food Policy Research Institute (IFPRI). The history of ERHS, details of sampling strategy and data are elaborated elsewhere (Dercon et al., 2005; Dercon and Krishnan, 2003).3 Briey, since 1994, ERHS surveyed a total of 1477 households from 15 villages. Between 1994 and 1997, the sample households were interviewed four times: in the rst part of 1994, later in the same year, in the rst part of 1995 and 1997. The 15 ERHS villages are representative of Ethiopias major agro-ecological zones, with the exception of pastoralists, who are underrepresented. One to three villages were selected per stratum. Within each village, households were selected using random sampling, stratied by female and male headed households. In addition, to ensure that landless households were not excluded, the sample was stratied within each village to ensure that a representative number of landless households were included. Sample sizes in each village were chosen so as to approximate a self-weighting sample, when considered in terms of the farming systems. That is, as of 1994, each person represents the same proportion of persons found in the main sedentary farming systems. The resulting sample can be considered broadly representative of the households in nonpastoralist farming systems in the country as of 1994. All the households represented by a complete set of information at the baseline and the nal round were included in the analysis.

1 The term prime age refers to those adults who are economically productive and of reproductive age. The age range used for prime age differs across various studies. The high end point for dening prime age varies from 49 to 59 depending on the country and purpose of the study: for example, 1559 years in Zambia (Chapoto and Jayne, 2008); 1554 years for men and 1549 years for women in Kenya (Yamano and Jayne, 2004). Even though only 10% of those people in the ERHS data set are above 49 years of age, individuals could continue to make economic contributions past this age. Therefore, for the purposes of this dissertation research, the upper bound of the prime age range is the average life expectancy of Ethiopians at the age of 1554 years (WHO, 2002). 2 Ethiopia Rural Household Survey Dataset, 19892004. 2009. Washington, DC: International Food Policy Research Institute (IFPRI). 3 For further details of ERHS: http://www.ifpri.org/dataset/ethiopian-rural-household-surveys-erhs-1989-2004.

S. Kadiyala et al. / Food Policy 36 (2011) 647655

649

Thus the nal ERHS sample for the analysis consisted of 1408 households.4

Estimation strategy Endogeneity PAM may be correlated with individual and household characteristics such as wealth and education that are themselves important determinants of the outcomes of interest. Failure to control for these characteristics may generate biased estimates of the impact of PAM on these outcomes. The key analytical challenge is to determine how households affected by death would have fared if they had not had such traumatic events and thereby to understand the impact of the events. This paper seeks to add insight into both this research question and the use of a specic research method propensity score matching (PSM)when applied in the context of PAM. Most panel studies (Ainsworth and Semali, 2000; Deininger et al., 2003; Yamano and Jayne, 2004) used xed-effects methods to account for endogeneity. Fixed-effects methods impose strong restrictions on functional forms. These restrictions are an important limitation of fully parametric estimation techniques, even after conditioning for a wide range of key variables (Jalan and Ravallion, 2003; Gertler et al., 2004). This is because parametric regressions require a single estimated effect to be calculated from a data set of (in this case) children or households whose characteristics may differ widely. The ideal experiment compares the outcome of two identical individuals/households exposed to the
4 We examined the possibility of attrition bias in the following three ways: (a) There were no signicant differences in means of several variables (demographic variables, value of assets and so on) between 1477 households at baseline and 1408 that we used for the analysis. (b) We checked for attrition bias (in this case bias introduced due to exclusion of 69 households) using inverse probability weighting (IPW) method developed by Fitzgerald, Gottschalk, and Moftt 1998. It requires estimating two probit regressions, one with and one without variables that are signicantly associated with attrition, and using the ratio of predicted probabilities from these regressions to reweight the observations. F-tests on the joint signicance of variables signicantly associated with attrition and Wald tests of the joint signicance of the probability of attrition and its interaction with all other variables both led to the acceptance of the null hypothesis that attrition is random. Results of the attrition analysis are available from the authors upon request. (c) We further investigate if there is a signicant difference in propensity scores unadjusted and adjusted for attrition before matching. As the graph below shows, the propensity score kernel density plots overlap perfectly with no statistically signicant difference in the means. The above analyses suggest that dropping the 69 households from the analytical sample is unlikely to cause attrition bias.

same treatment. The PSM method approximates this experiment. Kadiyala et al. (2009) employ propensity score matching (PSM) with a difference-in-difference (DID) estimator to estimate the impact of PAM on child survival and growth using the ERHS dataset. As will be explained below, DID estimation is equivalent to household xed-effects analysis employed by previous authors because unobservable time-invariant household characteristics are differenced out. Donovan and Bailey (2006) and Gertler et al. (2004) use PSM methods to study the impact of PAM on farm production and childrens schooling respectively. The present study builds on such earlier work on impact evaluation using PSM in general (Gilligan and Hoddinott, 2007; Jalan and Ravallion, 2003) and in the context of adult mortality in particular to estimate the impact of PAM on household composition and consumption. Propensity score matching (PSM) with difference-in-difference (DID) estimation To understand the true effect of death on a household, we need (a) the outcome of a household with a death (hereafter, a PAM household) and (b) the outcome of that same household under the same conditions, but without death (non-PAM households). Since the latter is unobservable, we need to construct a proxy counterfactual for the missing data. PSM allows us to match each PAM household with similar non-PAM households for the same time period and use the outcome of the unaffected households as a proxy for the outcome of the affected household if it had not had a death (Heckman et al., 1997). This involves estimating a probit or a logit model that predicts the probability of each household suffering PAM as a function on pre-PAM household and community characteristics. The better the set of characteristics in accounting for outcome differences in the absence of treatment, the better chance of eliminating confounding factors that affect the outcome (Becker and Ichino, 2002). For each PAM household, matching involves nding a non-PAM household with the closest estimated propensity score. The results presented in the following section are estimated using kernel matching, following Gilligan and Hoddinott (2007). The kernel is a function that weights the contribution of each non-treated household according to distance of that household to the estimated propensity score. Exact matches are given a large weight and poor matches are given a small weight (Dehejia and Wahba, 2002; Becker and Ichino, 2002). Abadie and Imbens (2006) show that using the bootstrap after nearest neighbor matching, until recently a common approach to estimating standard errors in evaluation studies, does not yield valid estimates. Bootstrapping standard errors for kernel matching estimators is not subject to this criticism because the number of observations used in the match increases with the sample size. We implemented this procedure only for PAM households with common support in the sample of potential comparison observations by omitting any observations on PAM households from the matched sample that had propensity scores above the maximum propensity score of the non-PAM households (Smith and Todd, 2001). We also tested the balancing properties of the data by testing that PAM and non-PAM households had the same distribution of propensity scores and of control variables within quantile groupings of the propensity score. All results presented below are based on specications that passed the balancing tests. The study estimated the effect of PAM on household outcomes for the cases of affected households, not for all households in the general population. This is known as the average treatment effect on the treated (ATT). The ATT is obtained by comparing the means of changes of outcomes of interest between the matched PAM and

.2

.4 Estimated propensity score

.6

.8

Attrition unadujsted

Attrition adujsted

Finally, since the propensity score matching programs (pscore and psmatch2) for kernel matching with boot strapping in STATA do not support the use of weights, it is not possible to estimate ATT with attrition weights. However, since we have demonstrated that attrition is random, this does not affect the validity of our PSM estimates.

650

S. Kadiyala et al. / Food Policy 36 (2011) 647655

non-PAM households. It is plausible that unobservables that are uncorrelated to observables may still bias the impact estimations after PSM. In order to further improve the impact estimations, we employ difference-in-difference (DID) estimation. DID estimation accounts for treatment level group xed effects, which is one important component of the bias from unobservables. Similar to xed-effects methods, PSM with double difference allows for time invariant, but not time variant, unobservables to be controlled for. Through comparisons with experimental estimators, Heckman et al. (1997) show that propensity score matching provides reliable, low-bias estimates of treatment impact provided that the same data source is used for PAM and non-PAM households and the data include meaningful X variables capable of identifying the PAM status. The ERHS surveys clearly meet the rst criterion. The ERHS data also provide a very rich set of variables to identify PAM status, as required by the second criterion. Variable description A PAM household is a household that has had an adult between 15 and 54 years of age die between 1994 and 1997 (=1, 0 otherwise). The main outcome variables of interest pertain to household demographic changes and expenditure patterns. In PSM, it is desirable to over-parameterize the logit from the point of view of a model of determinants in order to nd the closest match possible. Therefore, we used a broad set of variables in constructing the propensity scores in matching PAM and non-PAM households. In all subsequent discussions, we refer to lagged characteristics as those referring to the previous survey round (lagged by one round), and the baseline as the rst round (1994) or before the occurrence of PAM. We estimated a logit, by pooling the observations across all the four rounds, to compute the propensity score using the following model:

Log oddsPAMt 1 g Ht1 ; Ht ; V R

where Log odds (PAMt = 1) was the log odds of a household having a prime age adult die. The dataset does not have information on the exact time of death and the information elicited in the survey was since the last round. Thus identifying the precise timing of death is not feasible. Since we used all four rounds of data, the impacts can be considered as the average impacts on outcome variables since death. Ht1 is a vector of lagged household characteristics, lagged by one round (log household size; fraction of males and females by age groups; log total monthly expenditures per capita and its square; log monthly expenditures on food per adult equivalent; share of expenditures on food, health, education, alcohol and tobacco and ceremonies; value of assets (excluding land); square root of total land holdings in hectares; a dummy variable for whether any prime age adult in the household was ill and 12 month annual rainfall one year before survey). Ht is a vector of household characteristics in the baseline year, dened as the rst round in 1994 (gender of the household head, highest grade of the household head, highest grade attained by any female in the household; a dummy variable whether the household had livestock in the house; a dummy variable for whether it boiled water for drinking and/or if piped water was the main source of drinking water; a dummy variable for whether any prime age adult migrated in or out before the rst round in 1994; and a dummy variable for whether the household had suffered any PAM before the rst round (using a ve year recall period); distance to the nearest government hospital; distance to the nearest town). V R is a set of village and round interaction terms to control for aggregate or

covariate shocks common to all households in the village within any given round.5 The outcome variables for the impact of PAM on demographic composition of the households are change in (a) household size and (b) dependency ratios (dened here as the persons under age 15 plus 54 years and above to those between the age of 15 and 54 years of age). The outcome variables for the impact outcomes of PAM on consumption are: change in (a) total monthly log per capita expenditures; (b) monthly log food expenditures per adult equivalent; (c) monthly log non-food expenditures per capita and (d) dietary diversity. When poor households, which already devote a very high proportion of their expenditures to food, are faced with shocks they tend to maintain the value of food expenditures by buying cheaper foods to maintain energy consumption and/or consume less diverse diets. Monetary measures of food consumption alone do not reveal the adaptations that households make in food intake patterns when faced with shocks. Indicators of dietary diversity provide more information than monetary measures of food consumption and have been shown to be associated with a number of welfare outcomes. Studies from multiple developing countries have documented positive associations between dietary diversity and nutrient adequacy (Hatloy et al., 1998; Ogle et al., 2001; Ruel, 2003). A more varied diet is associated with a number of improved outcomes such as birth weight (Rao et al., 2001) and child anthropometric status (Onyango et al., 1998; Hatloy et al., 2000). Finally, evidence from a multicountry analysis by Hoddinott and Yohannes (2002) suggests that household-level dietary diversity is strongly associated with per capita energy availability. Dietary diversity is usually measured using a simple count of foods (Drewnowski et al., 1997; Hatloy et al., 1998) or food groups over a given reference period (Kant et al., 1995; Hatloy et al., 1998; Swindale and Ohri-Vachaspati, 1999). Hoddinott and Yohannes (2002) point out the advantages and disadvantages of both these approaches. Knowing, for example, that a household consumes four food groups, as opposed to four different types of cereals, is more indicative of a more diverse diet. Conversely, changes in food consumption resulting from higher incomes may be evidenced by improved quality of foods rather than consumption of different food groups. Consequently, this study used both the count of food groups and the number of unique foods consumed in the last week to measure dietary diversity using expenditure and consumption data.6 Based on the ERHS data we grouped food into a total of seven groups: cereals, lentils/nuts, dairy, animal products (excluding dairy), fats, fruits and vegetables, and processed food such as bread and spaghetti. The maximum number of unique foods consumed by any household in the last week was 19. Results and discussion Table 1 presents the characteristics of unmatched PAM (120) and non-PAM (1288) households. There were signicant baseline differences between the groups for most of the variables of interest. Dietary diversity of rural Ethiopian households was alarmingly low, with households consuming, on average, fewer than three food groups (out of seven) and approximately seven unique food items per week.
5 Alternate models that employ characteristics at the baseline instead of lagged variables; and alternate specications of the migration variables (in migration and out migration lagged by one round) to estimate propensity scores produced similar results. 6 ERHS households were asked if the households consumed any of the following foods (list of foods provided to the eld researcher) in the last one week. Then households were then asked to provide information on the amount and source of the food consumed (purchased, consumed from own harvest/stock., received as a gift, wage in kind or as barter).

S. Kadiyala et al. / Food Policy 36 (2011) 647655 Table 1 Baseline characteristics of ERHS households by prime age adult mortality (PAM) status before matching. Baseline characteristics Total number of households (=1408) Household demographic variables Mean household size Mean dependency ratio Households with male household head at baseline (%) Mean number of children under ve years of age Mean number of girls (614 years) Mean number of boys (614 years) Mean number of prime age females(15 54 years) Mean number of prime age males(15 54 years) Mean number of the elderly persons (over 55 years of age) Mean highest grade of females in the household Mean highest grade of household head Household economic variables Mean household monthly expenditures Total monthly expenditures per capita Monthly food expenditures per adult equivalent Monthly non-food expenditures per capita Value of assets (in Birr), excluding land Square root of total land in hectares PAM households 120 7.71** 1.12*** 78.33 0.9 1.03 1.21** 2.12** 1.91 0.5 2.6
*** a

651

Table 2 Logit estimates for households suffering PAM between 1994 and 1997 in Ethiopia Rural Household Survey (Pooled Modelstacking all four rounds). Variables Household characteristics Lagged household characteristics (lag by one round) Log household size <5 year, girls <5 year, boys 514 years, girls 514 years, boys 1524 years, females 1524 years, males 2554 years, females >55 years, females >55 years, females Total household expenditure per capita (in Birr) Total household expenditure per capita squared Log food expenditure per adult equivalent Share of food expenditures Share of health expenditures Share of education expenditures Share of alcohol and tobacco expenditures Share of ceremonies expenditures Value of assets, excluding land (in Birr) Square root of total land holdings(hectares) Prime age adult ill (=1, 0 otherwise) PAM households (=1, 0 otherwise)

Non-PAM households 1288 5.96 1.32 78.80 0.9 0.86 0.87 1.46 1.36 0.53 3.6 2.91

0.464 0.107 .0175 0.161 0.112 0.132 0.23* 0.21 0.05 0.025 0.003 0.000 0.16 0.001 0.065** 0.118 0.026* 0.008 0.001 0.12 0.313

1.53***

58.47* 50.75* 15.29 165.73 1.13

71.95 65.77 18.37 167.18 1.06

Mean monthly household expenditure shares of selected expenditure groups Food 69.35** 74.97 Health 2.53* 1.60 Education 0.450 0.327 * Ceremonies 5.44 3.73 Alcohol and tobacco 5.35 5.04 Dietary diversity based on monthly expenditure data Mean number of food groups 2.9 Mean number of unique foods 6.8 Other variables Households with piped water (%) Households with livestock in the house (%) Households reporting adult migration (%) Households reporting illness of prime age adult (%) Households reporting prime age adult death before baseline (%) Mean distance to the nearest government hospital (km) Mean distance to the nearest town (km)
* ** ***

2.9 6.5 18.8 48.37 8.70 36 12.06 39.16

***

15.9 47.06 13.3*** 51** 18.33* 38.11 12.48

Household characteristics at baseline year (19941995) Male household head (=1, 0 otherwise) 0.192 Highest grade of household head 0.22** Highest grade of females at baseline .012 Household keeps livestock in the house(=1, 0 1.01** otherwise) Boils drinking water(=1, 0 otherwise) 0.012 Piped water for drinking(=1, 0 otherwise) 0.66*** Migration of member before baseline(=1, 0 0.742* otherwise) Prime-age adult mortality before baseline(=1, 0 0.725** otherwise) Distance to nearest government hospital (km) 0.496** Distance to nearest town (km) 4.38** 12 month annual rainfall one year before 0.006 survey, lag Observations 4073 Note: Regression includes village and round interaction dummies and a constant tem. * Differences signicant at 5%. ** Differences between PAM and non-PAM households signicant at 1% level. *** Differences signicant at 10%.

11.54

Differences signicant at 5%. Differences between PAM and non-PAM households signicant at 1% level. Differences signicant at 10%.

We obtained propensity scores using the psmatch2 procedure in STATA (Leuven and Sianesi, 2003). Here, the focus was on nding a set of conditioning variables that, on theoretical grounds and according to information in the ERHS survey, should have been highly associated with the probability of PAM and outcomes of interest. Table 2 presents the logit model predicting a household suffering PAM. Although the parameter estimates from this model were not intended to show causality, the estimates demonstrate association (Table 2). Based on the propensity scores from the model in Table 2 and using the kernel matching method with common support as described earlier, we generated a sample of matched PAM and nonPAM households. Graphs 1 and 2 present the kernel density plots of the estimated propensity scores before and after matching. Here, we observe neg-

ligible difference in the mean propensity scores and their distribution of the two groups after kernel matching. In addition, few statistically signicant differences between covariates in Table 2, between matched households, were detected (results provided upon request). Table 3 presents the difference-in-difference means of outcomes of interest by PAM status. As the effects of adult death differ depending on the economic status of households and the gender and position of the deceased adult (Chapoto and Jayne, 2008), we estimated the overall and heterogeneous effects of PAM by the gender and position of the deceased adult and by the poverty status of the household (lagged median expenditures) on changes in household composition and consumption patterns. Impact of PAM on household composition Given the importance of household labor in agricultural production and other household reproduction activities such as childcare, the analysis initially focused on the effects of PAM on household composition. Household age and sex composition is typically

652
4

S. Kadiyala et al. / Food Policy 36 (2011) 647655 Table 4 Average treatment effect (ATT)a of prime age adult mortality (PAM) on change in household composition between 1994 and 1997 in rural Ethiopia. ATT Change in household size Overallb
1

Change in dependency ratio 0.32 (3.18)** 0.62 (3.39)** 0.05 (0.40) 0.31 (1.92)*** 0.3 (1.72)***

1.33 (7.30)**

0 -.2 0 .2 .4

Estimated propensity score

Non-PAM households PAM households

Graph 1. Kernel density of propensity score before matching by adult mortality status.

Position and gender of the deceased adult Death of head/spouse 1.26 (4.14)** Death of non-head/ 1.6 spouse (7.6)** Death male adult 1.46 (6.24)** Death female adult 1.22 (4.66)**

Household poverty status (lagged median expenditure per capita) Poor 1.68 0.39 (7.13)** (2.62)** Non-poor 1.44 0.11 (1.1) (5.66)**
a Absolute value of t-statistics on ATT, in parentheses, are based on bootstrapped standard errors using 1000 replications of the sample. b For analysis on change in overall household size, number of treated observations is 161 and number of matched controls observations is 4061. For analysis on change in overall dependency ratio, number of treated observations is 150 and number of matched controls observations is 3886. Differences signicant at 5%. ** Differences between PAM and non-PAM households signicant at 1% level. *** Differences signicant at 10%.

0 -.2 0 .2 .4

psmatch2: Propensity Score

Non-PAM households PAM Households

Graph 2. Kernel density of propensity score after kernel matching by adult mortality status.

Table 3 Overall sample means of changes between 1994 and 1997 in household composition and consumption outcomes by prime age adult mortality (PAM) status after kernel matching in rural Ethiopia. PAM households Changes in household composition Change in household size Change in dependency ratio Changes in monthly log expenditures Total, per capita Food, adult per equivalent Non-food, per capita Change in dietary diversity Number of food groups Number of unique foods 1.63 0.25 0.27 0.29 0.13 0.07 0.30 Non-PAM 0.29 0.07 0.19 0.08 0.03 0.05 0.16

dynamic as households and their members progress through the normal lifecycle. PAM shocks, while severe for the households concerned, must be understood in the context of these dynamics. The observed changes in household composition are also a function of the duration of the study. Most panel studies examining the effect of PAM on household composition are between two to four years. Because the death of a prime age adult potentially reduces the households labor supply, we expect that households affected by PAM would respond by trying to replace adult labor and/or possibly by sending away dependents to maintain their household dependency ratios. Results from the few empirical studies on the effects

of adult mortality on household composition suggest varied responses. For example, Ainsworth et al. (1995) found that rural households in Tanzania were able to maintain their household size and dependency ratios even after suffering a prime age death. However, studies in Thailand and Uganda found that household size declined by about one person following a prime age death, suggesting that, on average, affected households in these areas were unable to attract new members (Janjaroen, 1998; Menon et al., 1998). Studies from Kenya (Yamano and Jayne, 2004) and Zambia (Chapoto and Jayne, 2008) show that households were able to partially replenish their household size or adjust their dependency ratios by attracting members or sending some of their members away. Table 4 shows that death of a prime age adult resulted in a statistically signicant decline in household size by 1.33 members within the three-year survey period. These ndings suggest that households were unable to respond to PAM by replenishing, even partially, their household size, and in fact, reduced their household size by even more than the deceased member. Table 4 also shows PAM resulting in a statistically signicant increase in dependency ratio. A caveat in interpreting changes in dependency ratios is that some households have incurred the death of their only adult member. The number of households that did not have a prime age adult in the subsequent rounds was signicantly higher among the PAM households (5.3%) compared to the non-PAM households (2.7%). These households were not considered in estimating changes in the dependency ratio (as the denominator is zero for such households). So the estimates of dependency ratios should be considered as an underestimate of the impact of PAM.

Heterogeneous impact of PAM on household composition By position and gender of the deceased Among the ERHS households, the pattern of more than a oneperson decline in household size due to PAM persisted regardless of the sex and position and the economic status of the deceased

S. Kadiyala et al. / Food Policy 36 (2011) 647655 Table 5 Average treatment effect (ATT)a of prime age adult mortality (PAM) on change in household consumption patterns between 1994 and 1997 in rural Ethiopia. ATT Change in log monthly expenditures Total (per capita) Overallb 0.08 (1.27) Food (adult equivalent) 0.07 (1.1) Non-food (per capita) 0.09 (0.83) Changes in dietary diversity Number of food groups 0.02 (0.15) 0.38 Number of unique foods 0.47 (1.8)*** 0.88

653

Position and gender of the deceased adult 0.04 0.02 0.04 Death of head/ spouse (1.34) (0.15) (0.21) 0.11 0.14 .11 Death of nonhead/ spouse (1.2) (.64) Death male 0.10 0.17 0.19 adult (0.88) (1.6) (1.1) 0.06 0.09 0.16 Death female adult (0.61) (0.13) (1.1) Household poverty status (lagged median expenditure Poor 0.14 0.15 0.004 (1.39) (1.5) (0.03) Non-poor .04 0.03 0.23 (0.42) (0.28) (1.7)***

(1.8)*** .26

(1.94)*** 0.36

(1.76)*** 0.01 (0.05) 0.08

(0.64) 0.8 (1.4) 0.05

these corresponding coefcients is perhaps surprising. However, mean real per capita consumption of sample households grew by a third. Using the same panel data set, Bigsten et al. (2003) found that poverty fell from 41% to 36% during the survey period. It is plausible that, helped by the restoration of peace, good weather, and changes in macroeconomic policies during the mid-1990s, overall improvements in consumption may have enabled households to better withstand the shock of PAM than such losses during a period of growing poverty or food insecurity. Earlier studies examining the impact of health shocks on household consumption using the ERHS data set have shown similar results. Skouas and Quisumbing (2005) did not nd a signicant relationship between days of illness and household consumption patterns between 1994 and 1997. Using the rst and third rounds of the ERHS data set, Asfaw and von Braun (2004) also observed similar patterns for total expenditures and food expenditures. PAM reduced the number of unique foods, but not the number of food groups, consumed (Table 5). Given that the mean number of food groups consumed per week was very low at 2.9, the results suggest that households may not have had much scope to reduce the number of food groups consumed in response to PAM. Heterogeneous impacts of PAM on household consumption patterns By position and gender of the deceased Neither the gender nor position of the deceased had a signicant impact on changes in expenditure outcomes. Death of a head/spouse resulted in a signicant decline in number of food groups and number of unique foods consumed. Death of a nonhead/spouse resulted in a signicant decline in number of food groups only. Gender of the deceased did not show a statistically signicant relationship with changes in dietary diversity (Table 5). By economic status of the household Similar to the ndings of Dercon (2000) in Ethiopia, this study found that poor households had a higher consumption growth while non-poor households had negative growth, suggesting convergence of consumption across economic groups. The mean change in the log value of total consumption per capita in Birr between 1994 and 1997 for poor and non-poor households was 0.501 and 0.151 respectively (Table 5). Table 5 shows that PAM did not signicantly alter total expenditures per capita or food expenditures per adult equivalent in poor or in non-poor households. Death of a prime age adult among nonpoor households resulted in marginally signicant increase in value of non-food consumption. This relationship could also be due to the higher death related expenditures such as funerals and transfers that the relatively well-off may have received in response to the PAM shock. Earlier evidence indicates that, compared to poor households, richer households are usually richer, not only in physical and human capital, but also in social capital. In a panel study in Tanzania, for example, richer households were more likely to receive private assistance, and they received more assistance overall than poor households when a prime age adult died (Lundberg et al., 2000). PAM had a signicant negative impact on dietary diversity among the poor. Among the poor households, PAM resulted in a decline in number of food groups consumed as well as unique foods consumed (signicant at 10% level). This relationship was not signicant among non-poor households. Conclusion With 8.5% of ERHS households reporting loss of a productive adult, PAM is an important shock that can affect the consumption

(0.4) per capita) 0.16 (1.8)*** 0.42 (1.6)

(1.1) 0.54 (1.98)*** 0.57 (1.33)

a Absolute value of t-statistics on ATT, in parentheses, are based on bootstrapped standard errors using 1000 replications of the sample. b For analysis on change in overall consumption outcomes, number of treated observations is 156 and number of matched controls observations is 3925. Differences signicant at 5%. Differences between PAM and non-PAM households signicant at 1% level. *** Differences signicant at 10%.

(Table 4). We observe a similar pattern with respect to dependency ratios. Households that reported the death of a head/spouse, had a signicant increase in dependency ratios compared to the matched households that did not suffer PAM. The death of other adult members (non-head/spouse) did not have a statistically signicant impact on dependency ratios. These results are consistent with the results from earlier studies (Janjaroen, 1998; Menon et al. 1998; Chapoto and Jayne, 2008). PAM resulted in increased dependency ratios regardless of the sex of the deceased. By economic status of the household In Table 4, the rows labeled poor and non-poor (by lagged median expenditures) show that PAM resulted in substantial household size reductions of greater than one among both poor and non-poor households. PAM resulted in a statistically signicant increase in dependency ratios among the poor only. Among the non-poor households, PAM did not result in a signicant increase in dependency ratios. One possible reason for this could be that relatively well-off households may have better social networks, and they may be able to send away some of their dependents to other households. Impact of PAM on value of consumption and dietary diversity Table 5 presents estimates of the average impact of PAM on expenditures and dietary diversity. The results show that, on the whole, PAM did not have a signicant impact on total expenditure, food expenditure or non-food expenditure. The positive sign on

654

S. Kadiyala et al. / Food Policy 36 (2011) 647655

patterns of the rural Ethiopian households. Using panel data for the mid-1990s, this study provides evidence of the impact of prime age adult mortality on household composition and consumption patterns. The study advances research on the impact of PAM on household welfare by utilizing PSM with DID to address endogeneity of PAM and outcomes of interest. As stated earlier, while PSM with DID controls for observable characteristics and time-invariant unobservable characteristics respectively, we cannot rule out that unobservable characteristics may still bias our results if there are time-varying characteristics that affect treated and non-treated observations differently. This study has shown that households were unable to replenish lost labor during the three-year survey period by attracting new members into the households that had suffered PAM, regardless of the gender and position of the deceased and the economic status of the household. With the caveats of changes in dependency ratios being an underestimate (since PAM households were more likely to have no prime age adults), the mean dependency ratios increased signicantly when poor households, but not non-poor households, suffered PAM and regardless of the position and gender of the deceased. These results are in accordance with the widespread view that the death of productive members of households causes labor shortages and that these impacts tend to be most severe among poor households and households with death of head/ spouse. Given that households did not replace the lost labor of the deceased, one would expect an increase in the households economic constraints reected in a decline in value of total consumption or at least some of the components of total consumption (especially non-food consumption). Nevertheless, the results suggest that, regardless of the economic status and sex of the deceased, Ethiopian households were largely able to protect total expenditure, including food expenditure from the PAM shock. However, we did nd a strong indication of economic constraints imposed by PAM based on its heterogeneous negative effect on dietary diversity. In other words, while the food expenditures were maintained, the composition of foods consumed at the household level changed. In resource poor settings, when households face shocks, they often adapt by trading- off quality to quantity (to protect total calorie intake), tending to shift their consumption away from nutrient rich foods such as milk, meat and fruits to cheaper staples. The reduction in dietary diversity due to PAM was driven by the poor; this result is consistent with our expectation of worsening nutrition insecurity due to PAM, especially among the relatively poor in our sample. There are numerous examples in the vulnerability literature on households increasing or maintaining expenditure on food in the face of real or potential loss of total income resulting from some kind of shock. Such households typically seek to stabilize or buffer their consumption of staple foods (usually grains or tubers), with a resultant reduction in diet quality as higher value (more nutrient dense) foods drop off the menu (Webb, 2010). For example, Block et al. (2004) showed that during the 1990s nancial crisis in Indonesia, rural households in Java sought to maintain basic levels of rice consumption despite escalating prices, which forced them to consume fewer higher-value products such as eggs, vegetables, meat and dairy foods. The result was reduced diet diversity and a concomitant serious increase in childhood iron deciency anemia. This study, the rst to our knowledge to do so, showed that PAM further worsened the already low quality of diet of the Ethiopian rural households. The results presented here should be interpreted in the context of the following: Households make several adjustments, including alterations in household consumption and composition, to cope with direct and indirect costs associated with illness preceding PAM. To the extent to which households engage in adjustment measures to cope with illness preceding PAM in our sample, the

measured impact of death would be an underestimate of the extent to which the household was initially affected due to morbidity and subsequent mortality. Protecting consumption when faced with PAM comes with many apparent as well as hidden costs, including depleting assets and increasing labor supply. Asset depletion may allow households to get through crisis periods but may erode their productivity and food security in later years (Yamano and Jayne, 2004; Mather et al., 2004; Donovan and Bailey, 2006; Chapoto and Jayne, 2008). Some coping strategies might have severe negative intergenerational effects something to be explored in the context of future research using longer periods of panel data. Many studies found a negative impact of parental mortality on childrens schooling (see, e.g. Gertler et al., 2004; Yamauchi et al., 2008). A previous analysis of the impact of PAM on child mortality in Ethiopia showed that PAM increases the mortality rates of both boys and girls below eight years of age (Kadiyala et al., 2009). Dercon et al. (2005) report that although prime-age adult death did not signicantly affect consumption, the ERHS (19992004) households often reported that the death of a husband, wife or another person had an other effect. Moreover, our results, which examine dietary diversity at the household level, do not adequately reveal whether the reduction in dietary diversity is worse for some individuals within the household. Dercon and Krishnan (2000), using the same data but setting anthropometric indicators as the outcome, examined the ability of individuals to smooth their consumption over time, relative to other members within the household. The authors found that there was a great variation in outcomes, especially for women in the southern parts of the country. Future research should further assess the impact of PAM on intra-household reallocation of resources and how such reallocation differentially affects the welfare outcomes of various members of the household. This new research will contribute to the development of social protection schemes and risk management that are better designed and targeted toward more vulnerable individuals within the household. Acknowledgements The authors are grateful to Daniel Gilligan, International Food Policy Research Institute, for his signicant contribution to this paper. This article draws from Kadiyalas PhD dissertation and has beneted from the involvement of Stuart Gillespie, Futoshi Yamauchi and Yisehac Yohannes from International Food Policy Research Institute (IFPRI) and Antony Chapoto from Michigan State University. The authors are grateful for the funding from the Regional Network on AIDS, Livelihoods and Food Security (RENEWAL), coordinated by IFPRI, to conduct this work. All errors and omissions are ours. References
Abadie, A., Imbens, G.W., 2006. On the failure of the bootstrap for matching estimators. NBER Working Paper No. T0325. Ainsworth, M., Ghosh, S., Semali, I., 1995. The Impact of Adult Deaths on Household Composition in Kagera Region, Tanzania. The World Bank, Washington, DC. Ainsworth, M., Semali, I., 2000. The Impact of Adult Deaths on Childrens Health in Northwestern Tanzania. Policy Research Working Paper, The World Bank, Washington, DC. Arrow, K., 1964. The role of securities in the optimal allocation of risk bearing. Review of Economic Studies 31 (86), 9196. Asfaw, A., von Braun, J., 2004. Is consumption insured against illness? Evidence on vulnerability of households to health shocks in rural Ethiopia. Economic Development and Cultural Change 53 (1), 115129. Becker, S., Ichino, A., 2002. Estimation of Average Treatment Effects Based on Propensity Scores. STATA Journal 2 (4), 358377. Beegle, K., 2005. Labor effects of adult mortality in Tanzanian households. Economic Development and Cultural Change 53 (3), 655684.

S. Kadiyala et al. / Food Policy 36 (2011) 647655 Beegle, K., De Weerdt, J., Dercon, S., 2008. Adult mortality and consumption growth in the age of HIV/AIDS. Economic Development and Cultural Change 56 (2), 299326. Bigsten, A., Kebede, B., Shimeles, A., Taddesse, M., 2003. Growth and poverty reduction in Ethiopia: evidence from household panel. World Development 31 (1), 87106. Block, S.A., Kiess, L., Webb, P., Kosen, S., Moench-Pfanner, R., Bloem, M.W., Timmer, C.P., 2004. Macro shocks and micro outcomes: child nutrition during Indonesias crisis. Economics and Human Biology 2 (1), 2144 Chapoto, A., Jayne, T.S., 2008. Impact of AIDS-Related mortality on farm household welfare in Zambia. Economic Development and Cultural Change 56 (2), 327 374. Cochrane, J.H., 1991. A simple test of consumption insurance. Journal of Political Economy 99 (5), 957976 Deaton, A., 1992. Understanding Consumption. Oxford University Press, Oxford. Dehejia, R.H., Wahba, S., 2002. Propensity score-matching methods for nonexperimental causal studies. Review of Economics and Statistics 84 (1), 151161 Deininger, K., Garcia, M., Subbarao, K., 2003. AIDS-induced orphanhood as a systemic shock: Magnitude, impact, and program interventions in Africa. World Development 31 (7), 12011220. Dercon, S., Krishnan, P., 2000. In sickness and in health: risk shaving within households in rural Ethiopia. Journal of Political Economy 108 (4), 688727 Dercon, S., Krishnan, P., 2003. Changes in Poverty in Rural Ethiopia 19891995. In: Booth, A., Mosley, P. (Eds.), The New Poverty Strategies. Palgrave MacMillan, London. Dercon, S., 2000. Growth and Shocks: Evidence from Rural Ethiopia. Center for the Study of African Economies (CSAE), Oxford University. Dercon, S., Hoddinott, J., Woldehanna, T., 2005. Shocks and consumption in 15 Ethiopian Villages, 19992004. Journal of African Economies 1 (4), 559585 Donovan, C., Bailey, L.A., 2006. Understanding Rwandan agriculture household strategies to deal with prime age illness and death: a propensity score matching approach. In: Gillespie, S. (Ed.), AIDS, Poverty, and Hunger: Challenges and Responses. Highlights of the International Conference on HIV/AIDS and Food and Nutrition Security, Durban, South Africa, April 1416, 2005. International Food Policy Research Institute, Washington DC. Drewnowski, A., Henderson, S.A., Driscoll, A., Rolls, B.J., 1997. The dietary variety score: assessing dietary diversity in healthy young and older adults. Journal of American Dietetic Association 97 (3), 266271. Gertler, P., Gruber, J., 2002. Insuring consumption against illness. American Economic Review 92 (1), 5170 Gertler, P., Levine, D.I., Ames, M., 2004. Schooling and parental death. Review of Economics and Statistics 86 (1), 211225. Gillespie, S., Kadiyala, S., 2005. HIV/AIDS Food and Nutrition Security: From Evidence to Action. Food Policy Report 7. International Food Policy Research Institute, Washington, DC. Gilligan, D., Hoddinott, J., 2007. Is there persistence in the impact of emergency food aid? Evidence on consumption, food security and assets in rural Ethiopia. American Journal of Agricultural Economics 89 (2), 225242. Grimm, M., 2010. Mortality shocks and survivors consumption growth. Oxford Bulletin of Economics and Statistics 72 (2), 146171 Hatloy, A., Torheim, L.E., Oshaug, A., 1998. Food variety a good indicator of nutritional adequacy of the diet? A case study from an urban area in Mali, West Africa. European Journal of Clinical Nutrition 52 (12), 891898. Hatloy, A., Hallund, J., Diarra, M.M., Oshaug, A., 2000. Food variety, socioeconomic status and nutritional status in urban and rural areas in Koutiala (Mali). Public Health Nutrition 3 (1), 5765. Heckman, J.J., Ichimura, H., Todd, P.E., 1997. Matching as an econometric evaluation estimator: evidence from evaluating a job training programme. Review of Economic Studies 64 (4), 605654. Hoddinott, J., Yohannes, Y., 2002. Dietary Diversity as a Food Security Indicator. Food Consumption and Nutrition Division Discussion Papers No. 36, International Food Policy Research Institute, Washington, DC. Hosegood, V., 2009. The Demographic Impact of HIV and AIDS Across the Family and Household Life-Cycle: Implications for Efforts to Strengthen Families in Sub-Saharan Africa AIDS Care-Psychological and Socio-Medical Aspects Of AIDS/HIV, 21, pp: 1321 (Supplement 1). Jalan, J., Ravallion, M., 2003. Does piped water reduce diarrhea for children in rural India? Journal of Econometrics 112 (1), 153173. Janjaroen, W., 1998. The impact of AIDS on household composition and consumption in Thailand. In: Ainsworth, M., Fransen, L., Over, M., (Eds.), Confronting AIDS: Evidence from the Developing World. The European Commission, Brussles. Kadiyala, S., Quisumbing, A., Rogers, B.L., Webb, P., 2009. The impact of prime age adult mortality on child survival and growth in rural Ethiopia. World Development 37 (6), 11161128.

655

Kadiyala, S., Chapoto, A., 2010. The AIDS epidemic, nutrition and food security, and livelihoods: Review of evidence in Africa. In: Sahn, D.E. (Ed.), The Socioeconomic Dimensions of the HIV/AIDS Epidemic in Africa: Challenges, Opportunities and Misconceptions. Cornell University Press, Ithaca, pp. 74109. Kant, A., Schatzkin, A., Ziegler, R., 1995. Dietary diversity and subsequent causespecic mortality in the NHANES I epidemiologic follow-up study. Journal of the American College of Nutritionists 14, 233238. Kochar, A., 1995. Explaining household vulnerability to idiosyncratic income shocks. The American Economic Review 85 (2), 159164. Leuven, E., Sianesi, B., 2003. PSMATCH2: Stata Module to Perform Full Mahalanobis and Propensity Score Matching, Common Support Graphing, and Covariate Imbalance Testing. <http://ideas.repec.org/c/boc/bocode/s432001.html>. Linnemayr, S., 2010. Consumption smoothing and HIV/AIDS: The case of two communities in South Africa. Economic Development and Cultural Change 58 (3), 475507. Lundberg, M., Over, M., Mujinja, P., 2000. Sources of nancial assistance for households suffering an adult death in Kagera. The South African Journal of Economics 68 (5), 139. Mather, D., Donovan, D., Jayne, T., Weber, M., Mazhangara, E., Bailey, L., Yoo, K., Yamano, T., Mghenyi, E., 2004. A cross-country analysis of household responses to adult mortality in rural sub-Saharan Africa: implications for HIV/AIDS mitigation and rural development policies. In: International AIDS Economics Network Pre-Conference, Bangkok, Thailand. Mason, N.M., Jayne, T.S., Chapoto, A., 2010. A test of the new variant famine hypothesis: panel survey evidence from Zambia. World Development (3), 356 368. Menon, R., Wawer, J.M., KOnde-Lule, K.J., Sewankembo, N.K., Chuanjun, L., 1998. The economic impact of adult mortality on households in Rakai district, Uganda. In: Ainsworth, M., Fransen, L., Over, M. (Eds.), Confronting AIDS: Evidence from the Developing World. The European Commission, Brussles. Morduch, J., 1999. Between the state and the market: can informal insurance patch the safety net? World Bank Research Observer 14 (2), 187207. Ogle, B.M., Hung, P.M., Tuyet, H.T., 2001. Signicance of wild vegetables in micronutrient intakes of women in Vietnam: an analysis of food variety. Asia Pacic Journal of Clinical Nutrition 10 (3), 249. Onyango, A., Koski, K.G., Tucker, K.L., 1998. Food diversity versus breastfeeding choice in determining anthropometric status in rural Kenyan toddlers. International Journal of Epidemiology 27 (3), 484489. Rao, S., Yajnik, C.S., Kanade, A., Fall, C.H.D., Margetts, B.M., Jackson, A.A., Shier, R., Joshi, S., Rege, S., Lubree, H., Desai, B., 2001. Intake of micronutrient-rich foods in rural Indian mothers is associated with the size of their babies at birth: Pune maternal nutrition study. Journal of Nutrition 131 (4), 12171224. Rosenzweig, M.R., Binswanger, H.P., 1993. Wealth, weather risk and the composition and protability of agricultural investments. Economic Journal 103 (416), 5678. Ruel, M., 2003. Is Dietary Diversity an Indicator of Food Security or Dietary Quality? A Review of Measurement Issues and Research Needs. Food Consumption and Nutrition Division Discussion Paper No. 140. International Food Policy Research Institute, Washington, DC. Skouas, E., Quisumbing, A.R., 2005. Consumption insurance and vulnerability to poverty: a synthesis of the evidence from Bangladesh, Ethiopia, Mali, Mexico and Russia. The European Journal of Development Research 17 (1), 2458. Smith, J.A., Todd, P.E., 2001. Reconciling conicting evidence on the performance of propensity-score matching methods. American Economic Review 91 (2), 112 118. Swindale, A., Ohri-Vachaspati, P., 1999. Measuring Household Food Consumption: A Technical Guide. Food and Nutrition Technical Assistance, Title 2 Indicator Guides, Academy of Educational Development, Washington, DC. Townsend, R.M., 1994. Risk and insurance in village India. Econometrica 62 (3), 539591. Udry, C., 1994. Risk and insurance in a rural credit market an empiricalinvestigation in Northern Nigeria. Review of Economic Studies 61 (3), 495526. Von Braun, J., Teklu, T., Webb, P., 1999. Famine in Africa: Causes, Responses, and Prevention. Johns Hopkins University Press, Baltimore Wagstaff, A., 2007. The economic consequences of health shocks: Evidence from Vietnam. Journal of Health Economics 26 (1), 82100. Webb, P., 2010. Medium to long-run effects of high food prices on nutrition. Journal of Nutrition 140 (1), 140S147S. WHO, 2002. Life Tables for 191 Countries: World mortality in 2000. WHO, Geneva. Yamano, T., Jayne, T.S., 2004. Measuring the impacts of working-age adult mortality on small-scale farm households in Kenya. World Development 32 (1), 91111. Yamauchi, F., Buthelezi, T., Velia, M., 2008. Impact of prime-age adult mortality on labor supply: Evidence from adolescents and women in South Africa. Oxford Bulletin of Economics and Statistics 70 (3), 375398.