WPS7171 Policy Research Working Paper 7171 Vulnerability to Malnutrition in the West African Sahel Federica Alfani Andrew Dabalen Peter Fisker Vasco Molini Poverty Global Practice Group January 2015 Policy Research Working Paper 7171 Abstract This study estimates marginal increase in malnutrition for of rainfall and temperature to link a child to the shock children ages 1–3 years from exposure to an extreme shock experienced in-utero. The study finds that while around in the West African Sahel. The study uses knowledge of 20 percent of the children in the sample are stunted or a child’s birth and high resolution spatial and temporal underweight, more than 30 percent of the children in the distribution of shocks, calculated from the Normalized sample are highly vulnerable to either form of malnutrition. Difference Vegetation Index and satellite-based measures This paper is a product of the Poverty Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at vmolini@ worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Vulnerability to Malnutrition in the West African Sahel Federica Alfani, Andrew Dabalen, Peter Fisker and Vasco Molini Key words: Malnutrition, Vulnerability, Shocks, Sahel JEL Classification: I31, I32, I140 F. Alfani, Food and Agriculture Organization of the United Nations (FAO); A. Dabalen, The World Bank; P. Fisker, University of Copenhagen, Changing Disaster; Vasco Molini, The World Bank. Acknowledgement: We would like to thank Yeon Soo Kim for excellent support with data analysis. This study was funded by the Regional Studies Program of the Chief Economist of the Africa Region. 1. INTRODUCTION A large literature has documented how households in low income settings suffer short and long run welfare losses from uninsured risk, especially in rural settings where agricultural production risk is prevalent and markets are thin or nonexistent (Townsend, 1994; Karlan et al. 2013; Udry, 1994; Jacoby and Skoufias, 1997). While the short run welfare losses are bad enough, it is now widely acknowledged that the long run losses which typically manifest in foregone investments – in human capital, enterprises, high yielding crops, and so on – are especially damaging (Dercon and Christiaensen, 2011; Alderman, Hoddinott and Kinsey, 1996; Morduch, 1990; Hill, 2009). The actions that are taken by households in these contexts to avoid high risk but high return activities are motivated by their desire to reduce their vulnerability to shocks. The concept of vulnerability has gained currency in recent studies of well-being because the static analysis of poverty has been found to be too limiting in capturing the dynamic reality of poor populations: focusing only on the poor leaves out a significant portion of the population who live at a constant risk of becoming poor. Vulnerability is an ex-ante statement about future poverty, before the veil is lifted and the uncertainty is replaced by the knowledge of the actual facts. However, it has proven a lot easier to define vulnerability conceptually than to measure it. Empirically, since it is a prediction about the future, the ideal data sets – which would involve panel data over several years for each individual (or household) and shocks s/he experienced, responses to the shocks, and the outcomes (e.g. welfare) - rarely exist. Therefore, alternative models that exploit the most commonly available data sets have been proposed in the literature. 2 One group of authors defined it as the probability of falling into poverty (Christiaensen and Boisvert, 2000; Christiaensen and Subbarao, 2005; Chaudhuri, 2001; Pritchett et al. 2000), and more recently by Dang and Lanjouw (2014). An alternative definition given by Ligon and Schechter (2003) defines vulnerability as the difference between expected utility and a level of consumption that is assured (a level of consumption where there is no risk), while Dercon and Krishnan (2003) propose vulnerability as uninsured exposure to risk. Most of these measures of vulnerability share three elements in common. First, there is a basic acceptance that vulnerability involves exposure to a bad event – that is, a negative shock – that has not yet been realized. Second, there is a non-negligible probability that in the event of the shock there will be a loss (income, consumption, health, and so on). Finally, they all define income thresholds that classify households into vulnerable and non-vulnerable. However, there is no consensus on what the threshold income that assigns households into vulnerable or not vulnerable should be. For instance, most of the papers that define vulnerability as future or expected poverty assume an income threshold at which a household has 50% probability of falling into poverty, although Dang and Lanjouw propose a 10% probability. This study places itself in the category of estimating vulnerability to expected poverty. Unlike the existing studies, this paper examines vulnerability to malnutrition induced by rainfall shocks in the Sahel belt of the West African drylands. Five countries are included in the study: Burkina Faso, Ghana, Mali, Nigeria, and Senegal. For Ghana and Nigeria, only territories in the north of these countries lie in the Sahel belt, so the statistics and evidence on welfare losses will apply to households resident in those areas. For other countries, we look at the entire sample. For the rest of the paper, Sahel will refer to these countries. 3 Our approach follows the methods proposed by Anttila-Hughes and Hsiang (2013). First, we estimate the impact of shocks on child health using spatial and historical variation of a measure of drought that is not affected by anthropogenic activities. Next we use the historical and spatial distribution of drought to obtain a distribution of the “expected loss”. This is obtained by multiplying the average effect of a shock with values of our drought measure for each cluster and point in time. This allows us to evaluate the probability that a child in a given location will be malnourished in a hypothetical future period. We find that approximately 20% of children on average, in the five countries, are malnourished and that the uncertainty about the future in combination with the effects of negative weather shocks means that the fraction that is vulnerable to malnutrition lies between 30% and 40 % of the children in the sample. The rest of the paper is organized as follows. Section 2 provides a short description of the risk environment. Section 3 lays down a simple model to estimate the damage to child health (malnutrition) and build on that to estimate vulnerability of children to droughts. Section 4 describes the data we use to estimate welfare impacts of and vulnerability to droughts, while section 5 discusses the results. In section 6 we conclude. 2. RISK CONTEXT OF THE SAHEL Households in the Sahel belt face many risks. Some of these risks, when realized, diminish human capital of the households on a frequent basis (e.g. a subset of idiosyncratic health risks), and thereby affect their productivity. In the most extreme, physical destruction of human capital happens when there is a large scale conflict or high mortality epidemic. Other 4 risks affect production directly, such as pests, droughts and floods, when they destroy crops and livestock assets. One persistent risk that has come to be associated with the belt is rainfall failure. The drying of the Sahel, and subsequent changes to social organizations and livelihoods, has been a steady process that began in the 1950s, and constitutes one of the most consequential changes in observed global precipitation in the 20th century (Nicholson, 1993; Dai et al., 2004). Figure 1 shows the now familiar evolution of precipitation for the Sahel region, dating back a century, characterized by three distinct periods. The first period, between 1900 and 1930, was characterized by substantial variability in rainfall relative to the long term average rainfall. This was followed by a relatively long wet period in the 1950s and 1960s, when the rainfall averages were above the long run mean (positive values). Then beginning around the 1970s, and through the 1990s, things turned for the worse when there occurred a prolonged period of unusually dry spell (negative values). While there has been some recovery since the 1990s, the variability remains high and more importantly, during the decade of the 2000s, there is yet no return to the relatively favorable precipitation trends of pre-1970. Figure 1: Sahel precipitation, 1900-2013 5 Source: The Sahel precipitation index: doi:10.6069/H5MW2F2Q But rainfall failure (drought) is just one of many shocks that erode households’ capabilities. In fact, part of the reason why the shocks in the Sahel tend to be devastating is because they are highly correlated. Even in normal times, when rainfall has not failed and harvests have been good, the stock of quality land, and the existing production technology, is not enough to meet the nutritional needs of large fractions of the rural population. These correlations are very well established for rainfall and price shocks. Figure 2 plots rainfall and price trends for the six study countries in the 2000s. The rainfall values are averages for each month, over small areas (grids) collected by satellite data. Prices were obtained from markets in each country by the Famine Early Warning System (FEWS) and Vulnerability Assessment and Mapping (VAM) of the World Food Program (WFP). Although we have the rainfall data since 2000, the price data collection does not go back that far for each country, so for each country we plotted the periods when both data sets are available. Figure 2 shows the resulting trends for Burkina and Northern Ghana. The maps for the other countries are in Annex 1. 6 Figure 2: Rainfall and price Graphs here Burkina Faso Ghana 0 0 0 0 0 2 5 3 2 250 *) 0 0 t 5 n G i 5 ) 1 u 2 K S/ 0 0 F H / 0 0 O G )2 2 X m e( e( m m ) c c m lpri 0 i l( 0 lp r 0 0 al ( a 1 2 l a a f l in in 0 n m nf a 5 m R i a 1 o i 0 R o n 5 n 1 e e g g a a 0 0 e r 5 r 5 e v 1 A v 0 A 10 100 0 0 10 0 7 7 7 7 7 m m m m m 5 8 9 0 1 2 0 0 1 1 1 7 7 7 7 7 7 7 7 7 7 7 7 7 0 0 0 0 0 m m m m m m m m m m m m m 2 2 2 2 2 0 1 2 3 4 5 6 7 8 9 0 1 2 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 Average monthly rainfall Cassava Average monthly rainfall Maize Maize Plantain_Apentu Millet Sorghum Rice_Local Yam Source: Authors’ own calculation using VAM and TRMM data. As the graphs show, the rainy season for all these countries falls around the same time of year – June to September. The exception is northern Ghana where the wet season starts one month early, and ends a month later – that is, the rainy season lasts from May to October. The graphs show that for many of the countries, price volatility is very high and highly correlated with the growing season. There is a tendency for prices to fall sharply immediately following the harvest – probably because there is a huge surge in supply in the market – and then rise slowly and reach a peak just before the rainy season, especially for certain key staples. There are also two additional observations. One is that periods of droughts are notable for the sharp increases in prices, not surprisingly. This is the case in 2002, 2005 and 2010, especially in Burkina Faso. The other is that the global food price shocks in 2008 and 2009 show up in these price data for countries where there is substantial dependence on imports to meet the main staple, as is evident in Nigeria and Senegal for rice. But even in a country like Mali, where domestic rice production is substantially higher, there was visible upward pressure on domestic prices from international transmission of shocks. 7 Finally, the interactions of known and “new” shocks have raised the scope of potential welfare losses. One of those new shocks is open communal conflict. To be clear, the Sahel and especially these six countries, like most of Sub-Saharan Africa, have had their periods of political instability, characterized by coups and low scale violence. But as Figure 3 shows, there has been a major spike in conflict in some of the countries in recent years. Furthermore, these conflicts seemed to be more common during the dry season – perhaps because mobility of combatants is easier – when the populations are already under stress. Figure 3: Conflict days and weather in Sahel West Africa Total conflict days Total conflict days Dry season Dry season Mali Nigeria Senegal Burkina Faso Ghana Niger 0 100 200 300 400 60 40 20 0 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 Year Year Rainy season Rainy season Mali Nigeria Senegal Burkina Faso Ghana Niger 50 100 150 200 10 2 0 30 40 0 0 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 Year Year Note: Different scale Note: Different scale Source: Authors’ own calculation using ACLED data. A look at self-reported shocks from recent surveys confirms the multiplicity of shocks that matter to households. More than half the households consider price and weather shocks as the most important shocks in their lives. Furthermore, the share of households who report experiencing two or more shocks ranges from a low of 6 percent in Senegal to 50 percent in 8 Burkina Faso. In this paper we look at the impact of rainfall shocks in six countries. In the next section we explain the empirical strategy. Figure 4: Share of households reporting two or more shocks The next section describes the data we use and how it is spatially linked. The two most important data sets for this particular analysis are the Demographic and Health Surveys (DHS) and our measure of shocks. 3. DATA DESCRIPTION 3.1 Demographic and Health Surveys (DHS) The DHS are cross sectional surveys, funded primarily by USAID and implemented until recently by Macro International1 and local statistical agencies, which are designed to collect information mostly on maternal and child health across many countries in the world. Our analysis in this paper includes five countries in West Africa, each with at least two rounds of geo-referenced data for the period between 2000 and 2012. The data are representative at the national level as well as at rural and urban locations. Our unit of analysis is children who 1 The implementation of the DHS is now overseen by ICF, which recently bought Macro International. 9 were born 1 to 3 years before each survey because our focus is on child nutrition outcomes which are especially crucial at these ages. The numbers of children in the sample in each country and survey year are shown in Table 1. The earliest survey was conducted in 2001 and the most recent was completed in 2010. For both Ghana and Nigeria, we pick only the regions or states in the northern parts of the countries because they are the ones with the closest resemblance to the Sahel environment. The final sample size is 55,772 observations. Table 1: Sample sizes of children between 1-3 years, by country and survey 2001 2003 2005 2006 2008 2010 Total Burkina Faso 0 5,733 0 0 0 8,421 14,154 Ghana 0 970 0 0 739 0 1,709 Mali 6,076 0 0 7,727 0 0 13,803 Nigeria 0 2,212 0 0 10,837 0 13,049 Senegal 0 0 5,498 0 0 7,559 13,057 Total 6,076 8,915 5,498 7,727 11,576 15,980 55,772 Source: Authors’s calculation using DHS, various years Although maternal and child health are the primary concerns of the surveys, information on household demographics, nutrition, asset holding, migration and employment is also collected. Table 2 contains summary statistics for all variables used in the regressions. The top part of the table shows the anthropometric Z-scores, and for rounds one and two surveys. Note that the survey rounds are different from country to country, so that e.g. 2005 may be round 1 in one country and round 2 in another. The second part of the table summarizes the shocks, the construction of which is described in detail in the subsection that follows. The summary statistics for some of the control variables which explain a large part of the variation in standards of living are shown in the third section, while the bottom part shows the distribution of observations among countries in the database. About 30 percent of the sample is from Northern Nigeria, 10 while only 6 percent is from Northern Ghana. The share of the sample from the other three countries is roughly the same – about 20 percent each. Table 2: Summary statistics Variable Observations Mean Std. Dev. Min Max Height-for-age 32958 -119.36 182.99 -600 600 Weight-for-age 32958 -130.47 151.11 -592 600 Height-for-age (round 1) 13166 -115.61 174.75 -599 600 Height-for-age (round 2) 19792 -121.84 188.24 -600 600 Weight-for-age (round 1) 13166 -132.55 150.86 -592 600 Weight-for-age (round 2) 19792 -129.08 151.27 -591 599 Shock 54366 -.092 .217 -.79 .351 Rainfall levels (avg. mm/h) 62981 0.11 0.05 0.01 0.40 Urban 63911 0.26 0.44 0 1 Number of HH members 63911 8.73 5.96 1 74 Primary education 63911 0.17 0.37 0 1 Secondary education 63911 0.12 0.33 0 1 Higher education 63911 0.02 0.13 0 1 HH has toilet 63911 0.60 0.49 0 1 HH has electricity 63911 0.26 0.44 0 1 HH owns radio 63911 0.74 0.44 0 1 HH owns TV 63911 0.25 0.43 0 1 HH owns Refrigerator 63911 0.08 0.28 0 1 HH owns bicycle 63911 0.48 0.50 0 1 HH owns motorcycle 63911 0.25 0.43 0 1 HH owns car 63911 0.04 0.21 0 1 HH owns phone 63911 0.04 0.19 0 1 Dwelling has good floor 63911 0.44 0.50 0 1 Age of HH head 63777 43.29 13.66 13 97 Male headed HH 63911 0.89 0.31 0 1 Twin 63911 0.04 0.18 0 1 Male 63911 0.51 0.50 0 1 Burkina Faso 14154 0.22 0.42 0 1 Ghana 3874 0.06 0.24 0 1 Mali 13803 0.22 0.41 0 1 Nigeria 19023 0.30 0.46 0 1 Senegal 13057 0.20 0.40 0 1 Source: Authors’ calculation using DHS data (various years), TRMM and NDVI data. 3.2 Predicted Normalized Difference Vegetation Index as a measure of shocks 11 The shock indicator is the predicted greenness, that is, our best estimate of the deviation from long-run average of Normalized Difference Vegetation Index based on accumulated anomalies in rainfall and temperature. Since it is a predicted anomaly, it is distributed around zero when considering all years, but without fixed endpoints. The measure combines monthly information on the NDVI, rainfall, temperatures at night and temperatures at daytime to predict NDVI before aggregating to yearly averages. Figure 5: Lagged monthly correlations between year-on-year changes in rainfall, temperature and NDVI globally Source: Authors’ calculation using NDVI and TRMM data. For each DHS-cluster, monthly rainfall is estimated using the four nearest weather data observations, as pictured in Figure 6. The estimated rainfall, NDVI, drought etc. in the cluster is calculated as the weighted average of the measurement in these four points, which represent the four corners of the world of the gridded cell to which the cluster belongs. The weights of the equation are the inverse distances between the cluster and the weather observation, so that more weight is assigned to data points, the nearer they are to the DHS cluster. 12 Figure 6: Linking DHS clusters with gridded weather data Figures 7 and 8 provide an overview of the key variables used for identification of drought in the sample. The graphs show respectively the monthly and yearly variation in the weather data employed. The growing season spans the summer months with average rainfall peaking around a month earlier than greenness. Daytime temperatures drop during the rainy season and again in the winter months before rising sharply throughout the dry season. Looking at the graph of yearly variation in climate indicators, it is for instance noted that a drop in rain and greenness in 2011 led to relatively dry conditions. Otherwise the period from 2006 to 2010 was characterized by relatively little variation in climatic conditions when observing the region as a whole. But it is important to note that these averages may mask significant differences between countries. Figure 7: Monthly variation in weather data, all countries 13 Note: Z-scores (or anomalies – deviations from long run average) Figure 8: Yearly variation in weather data, all countries Note: Z-scores (or anomalies - deviations from long run average) are on the vertical axis. Shock is the predicted greenness variable. Negative values correspond to relative drought, while positive correspond to relative green conditions. The levels of rainfall, greenness and daytime temperatures are shown in figure 9. As we expected, there is a strong negative correlation between extreme temperatures and NDVI and rainfall. Coastal Ghana and Nigeria are the wettest, the most green and coolest of the group, whereas Sahelian Mali and Niger experience the driest and hottest conditions. 14 Figure 9: NDVI, rainfall and temperature; deviation from regional average by country Note: Regional average=1. There is a strong negative relationship between rainfall/greenness and daytime temperatures across the region. To identify the impact of the shocks on child malnutrition, we link the child to the shock that s/he experienced while in-utero. Since DHS collects information on the date of birth of each child, we can calculate when a child was likely to be in-utero. We use that information together with the knowledge of the spatial and temporal distribution of shocks to assign to a child the most likely shocks that s/he faced. Therefore, this is in keeping with a large body of literature that uses such natural experimental conditions to identify causal impacts (Neugebauer, Hoek, Susser, 1999; Ravelli et al., 1998; Mu and Zhang, 2011; Stanner and Yudkin, 2001; Almond, 2006). Table A9 in the annex illustrates how growing periods and cohorts are matched. 4. EMPIRICAL MODEL 15 Our main objective is to estimate vulnerability to child nutritional deficiency in the West African Sahel. However, this involves several steps. The first step is to estimate the welfare costs of shocks. Second, we need to obtain the probability distribution of the shock or the likelihood of the shock that a household faces. Third, we need to estimate expected loss from a shock given its average impact and probability distribution. Finally, we need to evaluate the resulting welfare relative to a standard. We now describe how we obtain values we need in each of these steps. Our first objective is to estimate household welfare losses from shocks. To identify the impact of shocks on household outcomes, we exploit spatial and historical variation of shocks in each location using a difference-in-difference approach. Although drought affects many households at once, it tends to have strong spatial patterns. We will use spatial variation of monthly historical rainfall recorded in the Sahel since 1998 to estimate the average impact of rainfall on child nutrition. Droughts lead to large scale crop and livestock losses, which in turn lead to high food prices and reduce access to food for many households. This is the first obvious channel for the close link between drought and child nutrition. However, even when households can protect the calories of children, they may do so by forgoing dietary diversity, which denies children essential nutrients for their growth. We therefore estimate the following difference in difference model. = + + + + + + (1) Where W is a welfare outcome (child malnutrition, measures of dietary diversity, consumption, food security, and so on), and h, r, c, and t indexes household, region, country and time; S is the shock, µ is a regional fixed effects, δ is a country fixed effects, θ is a time fixed effect and ε is a household level error term. 16 For identification, we use random year-to-year variation in exposure to shocks by adding regional and country fixed effects. Such controls allow us to identify the impact of the shock because they will be able to absorb the unobservable reasons why, on average, some locations may have higher or lower child nutritional deficiency. One concern is whether our shock – which relies on rainfall and temperature anomalies (see the data section above) is completely random. We would be concerned in particular if households were to forecast the arrival of drought and move away from the place most affected by it. However, while it is impossible for households to forecast the rainfall risk of a location, it is possible that they can respond to drought incidence by moving away for long periods of time. This will affect the composition of households, and if the movers are households that are richer, for instance, then the estimates will be biased upwards. Knowing the migration status of households will allow us to exclude migrants and run the regression on non-migrants. To avoid spurious correlations between outcomes and drought incidence, we introduce time fixed effects in all the models. We run model (1) for all six countries in a pooled regression. Our coefficient of interest is γ, which indicates the average impact of the shock on the welfare of interest: how much nutritional deficits worsen, how much reduction in dietary diversity occurs, how much consumption is foregone, and so on. We also run the model for subsamples of households: rural versus urban, whether the head is male or female, and by education of the head of household. Notice that an alternative equivalent specification is to run equation (1), but adding interaction of these variables defining subsamples – urban/rural, female/male head, and education categories - with the shock variable. The latter is the model we adopt. The estimates in equation (1) above will provide us with the average impact of shocks on outcomes. But in order to estimate the resulting welfare losses, we need to also obtain the 17 probability distribution of shocks for each household. For each location, we can obtain the distribution of shocks from historical data on drought incidence. In our case we have monthly rainfall and temperature data for well-defined spatial grids from 1998 to 2012. Therefore we can obtain the historical distribution of shocks that a household is likely to encounter at a location. We exploit this knowledge in combination with our knowledge on the average effect of a shock to calculate for each location the “expected loss” occurring from exposure to weather shocks for each time period in our historical distribution of the drought measure. In some periods the drought measure is positive, and the expected loss will be zero. Based on this, it is possible to evaluate how many (if any) observations will fall below a specified outcome measure threshold in a hypothetical future period, thus indicating vulnerability rates at different risk-levels. We apply this empirical strategy to measuring vulnerability to malnutrition for young children in West Africa. We use DHS data from five West African countries in order to capture the incidence and prevalence of underweight and stunting that can be attributed to droughts. In the next section we take up a discussion of the results. 5. RESULTS AND DISCUSSION Table 3 shows the first stage results – the impact of shocks. The dependent variables are the two most common measures of child malnutrition. Both variables are standardized relative to the global reference median for children of the same age. The shock variable is also normalized using the distribution over time and across space in our sample. The dependent variable is scaled by 100. Therefore, the results suggest that for a standard deviation change in shocks, stunting changes by around an eighth of a standard deviation. The impact on underweight is just slightly 18 smaller, around a tenth of a standard deviation. These are average values across 6 countries and control for time, country and province fixed effects. In addition to the greenness index – which is used as the shock variable – we also control for rainfall levels. Rainfall does not have any additional impact on stunting once the shock variable is controlled for, but it does influence underweight. Plausibly, places with better rainfall have better harvests and that is likely to lead to children with higher weights. Children are comparatively healthier in richer households, as implied by those with TV, good floor, and have more education. Table 3: Impact of shocks on stunting (height/aage) and underweight (weight/age) (1) (2) ht/a standard deviations wt/a standard deviations Shock 13.39*** 10.06** Rainfall levels (avg. mm/h) 55.97 242.08*** HH has toilet 8.56*** 6.93*** Number of household members -0.95*** -0.75*** Primary education 15.92*** 14.87*** Secondary education 27.97*** 33.42*** Higher education 49.79*** 49.88*** HH has radio 4.29 3.19 HH has TV 15.61*** 14.91*** HH has refrigerator 13.43** 9.99** HH has bicycle -5.72** -7.42*** HH has car 14.42** 11.53** Dwelling has good floor 15.61*** 10.63*** Age of HH head 0.19** 0.10 Male headed HH -6.24* 2.08 Current age of child 9.19** 15.56*** Observations 31,995 31,995 R-squared 0.176 0.217 Robust standard errors in parentheses. Clustered standard errors in parentheses. Both columns included interaction between shock variable and covariates as well as country, province and year fixed effects.*** p<0.01 ** p<0.05 * p<0.1 19 We now turn to our measure of vulnerability. Recall that to calculate vulnerability we need the impact of shocks on welfare, and a probability distribution of a shock occurring. We use the historical distribution of the predicted greenness index for each cluster and the average impact obtained from the pooled regression to estimate the average welfare loss for individuals living in each of the clusters. Our interest is to estimate the number of children who are vulnerable to negative shocks within different probability intervals. The steps required in order to achieve an estimate of the number of vulnerable children are as follows: First, using the monthly distribution of our shock measure in the period 2000- 2012, we calculate the “expected anthropometric loss” in each month by multiplying the specific value on the predicted greenness index by the coefficients on shocks obtained in the regression above. The validity of this method rests on the assumption that the effect of a shock is linear in values of the predicted greenness index. We also impose the assumption of “no positive gains” (or alternatively no negative loss); meaning that all cases where the drought-measure is 0 or above are seen as normal years and the expected loss is therefore set to equal 0. Secondly, we follow the common practice in empirical vulnerability studies to calculate the predicted value of our outcome measure for each child instead of using the actual values. This is because vulnerability status is an ex-ante statement about a future scenario that has not yet been revealed. Third, for each child we then proceed to calculating the share of months where the predicted outcome measure (stunting or underweight) minus the expected loss lies below a certain threshold. This share is a rough indication on the probability that a child will be 20 stunted/underweight in a hypothetical future period. Based on this, we calculate the size of different groups characterized by their risk of falling into malnutrition (50%, 25% 10% 5%). The results are shown in Table 4a and 4b. As a point of reference, the tables also show the fraction of children who are stunted and underweight. Roughly 20% of the children ages 1-3 in the West African Sahel belt are stunted and the same figure applies to underweight. The highest shares of children with nutritional deficiencies are found in Northern Nigeria, Northern Ghana and Mali. Senegal has the lowest and the malnutrition rates are lower in urban areas than in rural areas as we would expect. We find that vulnerability to malnutrition is considerably more widespread than actual malnutrition. For instance, around a third of the child population face a 50% risk of becoming stunted in the near future compared to the 20% who are already stunted. For underweight the proportion increases to 35%. The places with the largest difference between vulnerability and actual malnutrition are Northern Nigeria for stunting (24 percentage points) and Burkina Faso for underweight (28 percentage points). Table 4a: Vulnerability to stunting, Sahel West Africa. Stunted Vulnerable Vulnerable Vulnerable Vulnerable at 50 % risk at 25 % risk at 10 % risk at 5 % risk Full sample 0.190 0.326 0.345 0.364 0.374 Burkina Faso 0.179 0.364 0.389 0.411 0.423 Ghana 0.245 0.301 0.333 0.371 0.387 Mali 0.237 0.269 0.293 0.321 0.334 Nigeria 0.286 0.527 0.540 0.548 0.553 Senegal 0.054 0.152 0.164 0.180 0.187 Rural 0.207 0.366 0.386 0.407 0.417 Urban 0.138 0.200 0.215 0.231 0.240 No primary education 0.196 0.342 0.361 0.381 0.390 Primary education 0.155 0.232 0.251 0.267 0.279 21 Female headed household 0.143 0.204 0.220 0.240 0.250 Male headed household 0.195 0.338 0.358 0.377 0.387 Table 4b: Vulnerability to underweight, Sahel West Africa. underweight Vulnerable Vulnerable Vulnerable Vulnerable at 50 % risk at 25 % risk at 10 % risk at 5 % risk Full sample 0.197 0.355 0.380 0.404 0.415 Burkina Faso 0.219 0.494 0.518 0.541 0.549 Ghana 0.251 0.359 0.395 0.437 0.453 Mali 0.256 0.320 0.357 0.393 0.412 Nigeria 0.241 0.407 0.425 0.443 0.453 Senegal 0.058 0.182 0.198 0.214 0.220 Rural 0.216 0.406 0.432 0.457 0.468 Urban 0.138 0.195 0.216 0.239 0.248 No primary education 0.204 0.380 0.404 0.427 0.438 Primary education 0.155 0.211 0.239 0.266 0.277 Female headed household 0.149 0.260 0.284 0.305 0.315 Male headed household 0.202 0.365 0.390 0.414 0.425 Finally, we compute the share of the children in each cluster who can be considered vulnerable and plot the results on the map. Figure 10 is a vulnerability map, or cluster level vulnerability estimates. The vulnerability rates range from zero to almost 100%, the latter denoted by red dots. As is evident from the map, and as the tables above show, Senegal has the lowest vulnerability, while the northern Sahel belt – Burkina and Mali – has a substantially higher number of clusters with high vulnerability. Northern Nigeria also has a large number of clusters with high levels of vulnerability. 22 Figure 10: The cluster level vulnerability maps, height for age (left) and weight for age (right) Source: produced using GADM v.2 (gadm.org) and GPS coordinates from DHS 6. CONCLUSION In this paper we show that households in the West African Sahel experience multiple shocks which lead to large welfare losses. We use a combination of household surveys and a high resolution spatial and temporal measure of relative drought to estimate the average impact of a shock on a child’s nutritional deficiency. We find that on average a one standard deviation change in the shock leads to a change in nutritional deficiency of between an eighth and a tenth of a standard deviation. However, these welfare losses potentially hide large variations across individual countries in the sample. While we see our results as a rough average of the effect, more rigorous (panel data) estimation techniques could possibly yield a more precise indication of the true effect of weather shocks on measures of malnutrition. 23 The study then used the estimated impacts of the shocks and the historical and spatial distribution of shocks to calculate how many children are vulnerable to malnutrition under different circumstances. We estimate that around a third of the children in our sample face a 50% risk of falling into malnutrition in the near future, partly as a consequence of exposure to weather shocks. REFERENCES Anttila-Hughes, Jesse and Solomon Hsiang. 2013. “Destruction, Disinvestment, and Death: Economic and Human Losses Following Environmental Disaster.” UC, Berkeley. Alderman, H., J. Hoddinott, and B. Kinsey. 2006. Long term consequences of early childhood malnutrition." Oxford Economic Papers, 58(3): 450-474.Morduch, 1990; Almond D. Is the 1918 influenza pandemic over? Long-term effects of in-utero influenza exposure in the post-1940 U.S. population. Journal of Political Economy 2006; 114: 672-712. doi: 10.1086/507154 Chaudhuri, S. 2001. “Assessing vulnerability to poverty: concepts, empirical methods and illustrative examples.” Discussion paper series, Department of Economics, Columbia University. Chen, Xi. 2014. “Fetus, fasting, and festival: the persistent effects of in-utero social shocks.” International Journal of Health Policy Management, 3(4), 165–169 Christiaensen and Boisvert. 2000. “On Measuring Household Food Vulnerability: Case Evidence From Northern Mali.” Working Papers 127676, Cornell University, Department of Applied Economics and Management. Christiaensen, Luc J. and Kalanidhi Subbarao, 2005. "Towards an Understanding of Household Vulnerability in Rural Kenya," Journal of African Economies, Centre for the Study of African Economies (CSAE), vol. 14(4), pages 520-558, December. Dang, Hai-Anh and Peter Lanjouw. 2014. "Welfare dynamics measurement: two definitions of a vulnerability line and their empirical application," Policy Research Working Paper Series 6944, The World Bank. Dercon, Stefan and Pramila Krishnan, 2003. "Risk Sharing and Public Transfers," Economic Journal, Royal Economic Society, vol. 113(486), pages C86-C94, March. Dercon, S., and L. Christiaensen. 2011. “Consumption risk, technology adoption and poverty traps: evidence from Ethiopia." Journal of Development Economics, 96(2): 159-173. 24 Fisker, P. 2014. “Green Lights: quantifying the economic impact of drought”. IFRO working paper series, University of Copenhagen. Hill, R.V. 2009. “Using stated preferences and beliefs to identify the impact of risk on poor households." The Journal of Development Studies, 45(2): 151-171. Jacoby, Hanan G. and Emmanuel Skoufias. 1997. "Risk, Financial Markets, and Human Capital in a Developing Country," Review of Economic Studies, Wiley Blackwell, vol. 64(3), pages 311-35, July. Ligon, Ethan and Laura Schechter, 2003. "Measuring Vulnerability," Economic Journal, Royal Economic Society, Royal Economic Society, vol. 113(486), pages C95-C102, March. Mu R, Zhang X. Why does the Great Chinese Famine affect the male and female survivors differently? Mortality selection versus son preference. Economics and Human Biology 2011; 9: 92-105. doi: 10.1016/j.ehb.2010.07.003 Neugebauer R, Hoek HW, Susser E. Prenatal exposure to wartime famine and development of antisocial personality disorder in early adulthood. JAMA 1999; 281: 455-62. doi: 10.1097/00006254-200001000-00005 Pritchett, L., Asep Suryahadi, Sudarno Sumarto. 2000. Quantifying Vulnerability to Poverty: A Proposed Measure, Applied to Indonesia. Policy Research Working Paper, World Bank. http://dx.doi.org/10.1596/1813-9450-2437 Ravelli AC, van der Meulen JH, Michels RP, Osmond C, Barker DJ, Hales CN, et al. Glucose tolerance in adults after prenatal exposure to famine. Lancet 1998; 351:173-7. doi: 10.1016/s0140-6736(97)07244-9 The Sahel precipitation index: doi:10.6069/H5MW2F2Q Stanner SA, Yudkin JS. Fetal programming and the Leningrad Siege study. Twin Research 2001; 4: 287-92. doi:10.1375/1369052012498 Tiwari, Sailesh, Hanan G. Jacoby and Emmanuel Skoufias, 2013. "Monsoon babies : rainfall shocks and child nutrition in Nepal," Policy Research Working Paper Series 6395, The World Bank. Townsend, R.M. 1994. Risk and Insurance in Village India." Econometrica: Journal of the Econometric Society, 62(3): 539{591.Karlan et al. 2013; Udry, C. 1994. “Risk and insurance in a rural credit market: An empirical investigation in northern Nigeria." The Review of Economic Studies, 61(3): 495-526. 25 Rai n a f ll( m m) Rainfall(mm) 2 0 50 1 00 150 20 0 100 200 300 00 3 00 m m 7 20 7 2 0 01 04 m m 7 20 7 02 2 m 00 5 20 7 m 7 03 m 20 20 7 0 6m 04 7 m Maize 20 7 Maize 2 00 05 7 m m 7 Sorghum Sorghum 20 7 06 20 m 0 8m 20 7 7 Rainy season Rainy season 07 Rice_Imported 2 m 0 0 Mali 20 7 9 m 08 7 m ANNEX 1: FIGURES AND TABLES 20 7 20 1 09 0 m m 7 Burkina Faso 20 7 2 10 0 m 11 m 20 7 7 11 2 m 01 20 7 Millet 2 Millet m 7 12 m 7 Dry season Rice_Local Dry season 100 200 300 400 500 50 100 150 200 250 Ae v r a genominalpri c eX ( OFK / G) Average nominal price(XO F/KG) Rainfall(mm ) Figure A2: Rainfall and prices, Mali and Niger Ran i f a m ll( m) 20 0 20 40 60 80 100 0 50 100 150 200 250 0 0m 7 26 2 0 20 01 m 08 7 m 2 7 00 2 m 7 20 0 3 Figure A1: Rainfall and prices, Burkina Faso and Ghana m 7 20 2 0 09 0 m Source: Authors’ calculation using VAM and TRMM data 4 m 7 Yam 7 Millet 2 00 Beans 5 m 7 Cassava 2 Sorghum 0 0 20 6 m 10 7 m 2 0 7 0 Rainy season Rainy season 7 m 7 2 Plantain_Apentu 0 08 Niger m Ghana 7 20 2 00 11 9 m m 7 7 20 10 m 7 2 0 11 20 m 7 12 Maize 2 m 01 7 2 Maize m 7 Dry season Rice_Local Dry season Rice_Imported 100 200 300 400 500 600 0 50 100 150 200 Ae v a r genominalpri c eX ( OFK / G) Average nominal price(GHS/unit*) Source: Authors’ calculation using VAM and TRMM data Figure A3: Rainfall and prices, Nigeria (North) and Senegal Nigeria Senegal 0 0 0 0 0 0 3 3 3 400 0 G) 25 G) /K K OF 0 F/ 0 O 0 X 0 2 0 X ( 0 2 e m) 0 3 ( ric e 0 m) m ric 2 lp m ll( lp a fa ll( in a fa in m in in a m R o a 0 en R o 5 0 en 1 0 0 g 0 0 1 a 0 2 g 1 r a e r v e A Av 100 00 0 1 0 20 7 7 7 20 7 7 7 7 20 7 7 20 7 7 20 7 7 0 m m m m m m m m m m m m m 5 00 01 02 03 04 05 06 07 08 09 10 11 12 7 7 7 7 7 7 7 7 7 7 7 20 20 20 20 20 20 20 20 m m m m m m m m m m m 02 03 04 05 06 07 08 09 10 11 12 20 20 20 20 20 20 20 20 20 20 20 Rainy season Dry season Rainy season Dry season Maize Maize_Imported Maize Millet Millet Rice_Imported Sorghum Sorghum Source: Authors’ calculation using VAM and TRMM data Figure A4: Rainfall and prices, Burkina Faso and Ghana (rainy season months) Burkina Faso Ghana 0 0 20 0 30 25 250 nit*) 0 0 15 ) 25 G /u /K S 0 0 F ) 20 H 20 O e(G e(X m ) ric m infall(m ric infall(m lp 0 0 lp 10 20 ina ina 0 m a 15 m R a e no 0 R e no 15 erag erag 0 50 15 Av v 0 A 10 100 0 0 10 50 7 7 7 7 7 m m m m m 08 09 10 11 12 20 7 20 7 20 7 20 7 20 7 20 7 20 7 20 7 20 7 20 7 20 7 20 7 7 20 20 20 20 20 m m m m m m m m m m m m m 00 01 02 03 04 05 06 07 08 09 10 11 12 20 Average monthly rainfall Cassava Average monthly rainfall Maize Maize Plantain_Apentu Millet Sorghum Rice_Local Yam Source: Authors’ calculation using VAM and TRMM data 27 Rainfall(mm) 0 50 Rainfall(mm) 100 150 20 100 150 20 0 250 300 20 02 03 m m 7 7 20 20 03 04 m m 7 7 20 04 20 m 05 7 m 20 7 05 20 Millet m 06 7 Millet m 20 7 06 20 m 7 07 Rice_Local m 20 7 07 m 20 7 08 20 m 08 7 m Mali 20 7 09 20 m 7 Nigeria 09 m 20 7 Average monthly rainfall 20 10 m Average monthly rainfall 10 7 m 7 20 20 11 11 m m 7 7 20 Maize 20 12 12 m m 7 7 Sorghum Maize Sorghum Rice_Imported 100 200 300 400 500 50 100 150 200 250 300 A verage nom inalp rice(XOF/KG) Averagenominalp rice(X OF/KG) Rainfall(m m ) Rainfall(mm ) 20 0 20 40 60 80 100 20 50 100 150 20 0 250 300 00 00 m m 20 7 28 20 7 01 01 m m 20 7 20 7 02 02 m m 20 7 20 7 03 03 m m 20 7 20 7 04 Source: Authors’ calculation using VAM and TRMM data 04 Maize m m 20 7 05 20 7 05 m 20 7 m 06 20 7 m 06 20 7 m Rice_Imported 07 20 7 m Rice_Imported 07 20 7 m Maize_Imported Niger 08 20 7 m Figure A5: Rainfall and prices, Mali and Niger (rainy season months) 08 20 7 m 09 Senegal 20 7 m Average monthly rainfall 09 20 7 m 10 Average monthly rainfall 20 7 m Figure A6: Rainfall and prices, Nigeria and Senegal (rainy season months) 10 20 7 m 11 20 7 m 11 20 7 m 12 20 7 m 7 Millet 12 m Beans 7 Millet Maize Sorghum Sorghum 100 200 300 400 500 600 100 200 300 400 Average nominalprice(XOF/KG) Averagenominalprice(XOF/KG) Source: Authors’ calculation using VAM and TRMM data Figure A7: Reported shocks from latest household surveys Burkina Faso Mali Price shock 72.4 Natural/weather shocks 27.0 Natural/weather shocks 65.3 Other 20.9 Illness, accident 43.1 Loss of assets 18.1 Loss of assets 30.7 Crime and conflict 13.6 Illness, accident 12.7 Death of HH member 9.2 Death of HH member 12.6 Income shock 5.9 Price shock 9.9 Other 1.2 Income shock 3.9 0 20 40 60 80 0 10 20 30 Share of households (%) Share of households (%) Niger Nigeria Price shock 30.9 Income shock 12.0 Natural/weather shocks 28.3 Natural/weather shocks 10.5 Other 17.2 Death of HH member 9.6 Loss of assets 12.7 Price shock 7.9 Income shock 7.5 Illness, accident 6.2 Illness, accident 6.5 Loss of assets 4.2 Death of HH member 6.4 Crime and conflict 3.3 Crime and conflict 2.7 Other 2.5 0 10 20 30 0 5 10 15 Share of households (%) Share of households (%) Senegal Natural/weather shocks 15.3 Illness, accident 11.5 Income shock 5.7 Death of HH member 5.0 Loss of assets 0.9 0 5 10 15 Share of households (%) Source: Authors’ calculation using recent household surveys, various. Figure A8: Calculation of NDVI using Infrared radiation from satellite data. Source: NASA (earthobservatory.nasa.gov) 29 Table A9: Which growing seasons are included in shock-measure for different cohorts? Born last year Born 2 years ago Born 3 years ago Shock year t-1 X Shock year t-2 X Shock year t-3 X Annex 2: Description of NDVI data The images used in this analysis are so-called monthly maximum value composites. Since all atmospheric influence lowers NDVI, NASA stores only the highest greenness-value for each pixel over the period, where most pixels are recorded daily. This way cloud cover is filtered out in almost all cases. It is not straightforward to use NDVI as a proxy for drought, however. Year-on-year variation in greenness might be caused by factors other than climatic changes. As an example, deforestation quickly reduces the greenness of an area without being associated with drought. On the contrary, deforestation is often a sign of increased economic activity in a region. Broadly speaking, all factors that are non-climatic but affect the greenness of the planet will create noise in the picture of NDVI anomalies as a drought indicator. Most of these factors would be anthropogenic and, apart from deforestation, include changes in cultivation, irrigation and urban expansion. We use predicted NDVI, which takes greenness into consideration, but importantly leaves out all anthropogenic causes of change in “greenness”. The details of the construction of the index can be found in Fisker (2014). 30 From space it is possible to observe the surface of the earth and measure the light that is emitted at different wavelengths. Vegetation indexes such as the NDVI translate visible red and near infrared radiation into a decimal number between -1 and 1 which describes the greenness of a specified geographical area. In order to use NDVI as a proxy for drought, it is common to calculate the anomaly, i.e. the deviation from a long-run average for a specific time of the year. Figure A8 shows how NDVI is calculated as the ratio between near infrared radiation and visible red radiation; a higher index value is related to a greener land surface. NDVI data is obtained from the MODIS Terra satellite. It has been orbiting Earth daily since 2000, and here we employ a pre-processed product made publicly available by NASA that has a temporal resolution of one month and a spatial resolution of 0.05 degrees (3 arc minutes or around 5.8 km at the equator). It is later aggregated to 0.25 degrees in order to match the resolution of the rainfall data and reduce the number of observations. In the end we have a data frame with 1440 x 720 observations over 180 months for every location. Like NDVI, land surface temperature is measured from space globally using the MODIS Terra satellite, and again, the product in use has a spatial resolution of 0.05 degrees. Year-on- year changes in both daytime and night time temperatures are included in the model (see Table A2 in the Annex). On average, it is expected that day time temperatures affect greenness negatively since hotter means drier in most parts of the world. Night time temperatures are likely to affect greenness positively, however, since cold also becomes a serious constraint for plant growth when moving away from the equator. While greenness is best seen from above, rainfall is harder to measure using satellites. This study uses data from the Tropical Rainfall Measuring Mission (TRMM) which to our 31 knowledge is the most precise and valid remote sensing estimate of rainfall for the relevant period. In terms of spatial extent and resolution, the TRMM data is not as good as our measures of greenness and land surface temperature. It includes pixels of 0.25 degrees, which seems sufficient for our purpose. The link between year-on-year change in NDVI and the climatic background variables for every month is modeled using up to 11 lags so that it is only what has happened during the preceding year that is included. The technical aspects regarding the estimation of predicted greenness is described in Fisker (2014). Table A2.1: Predicting NDVI using rainfall and temperatures 32