vulnerability to malnutrition in the west african...
TRANSCRIPT
Policy Research Working Paper 7171
Vulnerability to Malnutrition in the West African Sahel
Federica AlfaniAndrew Dabalen
Peter FiskerVasco Molini
Poverty Global Practice GroupJanuary 2015
WPS7171P
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
ed
Produced by the Research Support Team
Abstract
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.
Policy Research Working Paper 7171
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 [email protected].
This study estimates marginal increase in malnutrition for children ages 1–3 years from exposure to an extreme shock in the West African Sahel. The study uses knowledge of a child’s birth and high resolution spatial and temporal distribution of shocks, calculated from the Normalized Difference Vegetation Index and satellite-based measures
of rainfall and temperature to link a child to the shock experienced in-utero. The study finds that while around 20 percent of the children in the sample are stunted or underweight, more than 30 percent of the children in the sample are highly vulnerable to either form of malnutrition.
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.
2
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.
3
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.
4
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
5
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
6
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.
7
Figure 2: Rainfall and price Graphs here
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.
50
100
150
200
250
Average nominal price(XOF/KG)
100
150
200
250
300
Rainfall(mm)
2000m7
2001m7
2002m7
2003m7
2004m7
2005m7
2006m7
2007m7
2008m7
2009m7
2010m7
2011m7
2012m7
Average monthly rainfall Maize
Millet Sorghum
Burkina Faso
050
100
150
200
Average nominal price(G
HS/unit*)
100
150
200
250
Rainfall(mm)
2008m7
2009m7
2010m7
2011m7
2012m7
Average monthly rainfall Cassava
Maize Plantain_Apentu
Rice_Local Yam
Ghana
8
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
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
01
00
20
03
00
40
0
1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014
Mali Nigeria Senegal
Year
Dry season
05
01
00
15
02
00
1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014
Mali Nigeria Senegal
Year
Rainy season
Note: Different scale
Total conflict days
02
04
06
0
1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014
Burkina Faso Ghana Niger
Year
Dry season
01
02
03
04
0
1997 2002 2006 2010 2014 1997 2002 2006 2010 2014 1997 2002 2006 2010 2014
Burkina Faso Ghana Niger
Year
Rainy season
Note: Different scale
Total conflict days
9
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.
10
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,
11
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
12
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.
13
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
14
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.
15
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
16
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.
17
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
18
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
19
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
20
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
21
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
at 50 % risk
Vulnerable
at 25 % risk
Vulnerable
at 10 % risk
Vulnerable
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
22
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
at 50 % risk
Vulnerable
at 25 % risk
Vulnerable
at 10 % risk
Vulnerable
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.
23
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.
24
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.
25
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.
26
ANNEX 1: FIGURES AND TABLES
Figure A1: Rainfall and prices, Burkina Faso and Ghana
Source: Authors’ calculation using VAM and TRMM data
Figure A2: Rainfall and prices, Mali and Niger
50
100
150
200
250
Avera
ge n
om
inal price
(XO
F/K
G)
0100
200
300
Rain
fall(
mm
)
2000
m7
2001
m7
2002
m7
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Rainy season Dry season
Maize Millet
Sorghum
Burkina Faso
050
100
150
200
Ave
rage n
om
inal price(G
HS/u
nit*
)
050
100
150
200
250
Rain
fall(
mm
)
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Rainy season Dry season
Cassava Maize
Plantain_Apentu Rice_Local
Yam
Ghana
100
200
300
400
500
Average nominal price(XOF/KG)
050
100
150
Rainfall(mm)
2003m7
2004m7
2005m7
2006m7
2007m7
2008m7
2009m7
2010m7
2011m7
2012m7
Rainy season Dry season
Maize Millet
Rice_Imported Rice_Local
Sorghum
Mali
100
200
300
400
500
600
Average nominal price(XOF/KG)
020
40
60
80
100
Rainfall(mm)
2000m7
2001m7
2002m7
2003m7
2004m7
2005m7
2006m7
2007m7
2008m7
2009m7
2010m7
2011m7
2012m7
Rainy season Dry season
Beans Maize
Millet Rice_Imported
Sorghum
Niger
27
Source: Authors’ calculation using VAM and TRMM data
Figure A3: Rainfall and prices, Nigeria (North) and Senegal
Source: Authors’ calculation using VAM and TRMM data
Figure A4: Rainfall and prices, Burkina Faso and Ghana (rainy season months)
Source: Authors’ calculation using VAM and TRMM data
50
100
150
200
250
300
Ave
rage nominal price
(XOF/K
G)
0100
200
300
Rainfall(mm)
2002m
7
2003m
7
2004m
7
2005m
7
2006m
7
2007m
7
2008m
7
2009m
7
2010m
7
2011m
7
2012m
7
Rainy season Dry season
Maize Millet
Sorghum
Nigeria
100
200
300
400
Ave
rage nominal price
(XOF/K
G)
0100
200
300
Rainfall(mm)
200
0m7
200
1m7
200
2m7
200
3m7
200
4m7
200
5m7
200
6m7
200
7m7
200
8m7
200
9m7
201
0m7
201
1m7
201
2m7
Rainy season Dry season
Maize Maize_Imported
Millet Rice_Imported
Sorghum
Senegal
50
100
150
200
250
Avera
ge n
om
inal price(X
OF/K
G)
100
150
200
250
300
Rain
fall(m
m)
2000
m7
2001
m7
2002
m7
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Maize
Millet Sorghum
Burkina Faso
050
100
150
200
Avera
ge n
om
inal price
(GHS/u
nit*
)
100
150
200
250
Rain
fall(m
m)
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Cassava
Maize Plantain_Apentu
Rice_Local Yam
Ghana
28
Figure A5: Rainfall and prices, Mali and Niger (rainy season months)
Source: Authors’ calculation using VAM and TRMM data
Figure A6: Rainfall and prices, Nigeria and Senegal (rainy season months)
100
200
300
400
500
Avera
ge n
om
inal price(X
OF/K
G)
050
100
150
Rain
fall(
mm
)
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Maize
Millet Rice_Imported
Rice_Local Sorghum
Mali
100
200
300
400
500
600
Avera
ge n
om
inal price(X
OF/K
G)
020
40
60
80
100
Rain
fall(
mm
)
2000
m7
2001
m7
2002
m7
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Beans
Maize Millet
Rice_Imported Sorghum
Niger
50
100
150
200
250
300
Ave
rage nominal price(X
OF/K
G)
100
150
200
250
300
Rainfall(m
m)
2002
m7
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Maize
Millet Sorghum
Nigeria100
200
300
400
Ave
rage nominal price
(XOF/K
G)
50
100
150
200
250
300
Rainfall(mm)
2000
m7
2001
m7
2002
m7
2003
m7
2004
m7
2005
m7
2006
m7
2007
m7
2008
m7
2009
m7
2010
m7
2011
m7
2012
m7
Average monthly rainfall Maize
Maize_Imported Millet
Rice_Imported Sorghum
Senegal
29
Source: Authors’ calculation using VAM and TRMM data
Figure A7: Reported shocks from latest household surveys
Source: Authors’ calculation using recent household surveys, various.
Figure A8: Calculation of NDVI using Infrared radiation from satellite data.
Source: NASA (earthobservatory.nasa.gov)
1.25.9
9.213.6
30.743.1
65.372.4
0 20 40 60 80Share of households (%)
OtherIncome shock
Death of HH memberCrime and conflict
Loss of assetsIllness, accident
Natural/weather shocksPrice shock
Burkina Faso
3.9
9.9
12.6
12.7
18.1
20.9
27.0
0 10 20 30Share of households (%)
Income shockPrice shock
Death of HH memberIllness, accident
Loss of assetsOther
Natural/weather shocks
Mali
2.76.46.5
7.512.7
17.228.3
30.9
0 10 20 30Share of households (%)
Crime and conflictDeath of HH member
Illness, accidentIncome shockLoss of assets
OtherNatural/weather shocks
Price shock
Niger
2.53.3
4.26.2
7.99.6
10.512.0
0 5 10 15Share of households (%)
OtherCrime and conflict
Loss of assetsIllness, accident
Price shockDeath of HH member
Natural/weather shocksIncome shock
Nigeria
0.9
5.0
5.7
11.5
15.3
0 5 10 15Share of households (%)
Loss of assets
Death of HH member
Income shock
Illness, accident
Natural/weather shocks
Senegal
30
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).
31
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
32
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