The long run impact of severe shocks in childhood: Evidence from the Ethiopian famine of 1984

Abstract

In 1984, the world was shocked at the scale of a famine in Ethiopia that caused up to a million deaths. The crisis was brought on by drought from repeated rainfall failure, poor economic policies and civil war. But what of the survivors? This paper estimates the long-term impact of the famine twenty years later, on young men and women who experienced this severe shock as small children during the crisis. The data show considerable heterogeneity in experiences in famine intensity-around three quarters of our sample had to cut back on quantities of food eaten, and of those, a third ate only one meal a day. We model adult health outcomes as a function of childhood inputs using a 2-period model of childhood development grounded in the human capital and nutrition literature, and identify the impact of the shock during "critical" or sensitive periods in childhood. We then examine height-for-age z scores at an intermediate stage ten years after the famine as an indicator for nutrition progress. At that point, children exposed to the famine aged 1-3 are shorter in comparison to those exposed as older children, and those born after the famine. The results indicate the possibility of limited catchup. We discuss the robustness of the results, including endogeneity of the shock, and positive and negative sample selection from survival and migration respectively.

Full text

Live aid revisited: long-term impacts of the 1984
Ethiopian famine on children
Stefan Dercon
University of Oxford
Catherine Porter
Heriot-Watt University∗
September 15, 2013
Abstract
In 1984, the world was shocked at the scale of a famine in Ethiopia that
caused over half a million deaths, making it one of the worst in recent
history. The mortality impacts are clearly significant. But what of the
survivors? This paper provides the first estimates of the long-term impact
of the famine twenty years later, on the height of young adults aged 19–
22 years who experienced this severe shock as infants during the crisis.
An innovative feature of the analysis is that famine intensity is measured
at the household level, while impacts are assessed using a difference-in-
differences comparison across siblings, and compared with an IV cross-
section, using rainfall as an instrument for the shock. We find that by
adulthood, affected children who were aged of 12-36 months at the peak of
the crisis are significantly shorter than the older cohort, and their unaffected
peers, by at least 5cm. There are no significant effects on those in utero
during the crisis, although we cannot rule out that for this cohort, the
selection effect dominates scarring. Indicative calculations show that for
the affected group such height loss may lead to income losses of around
5% per year over their lifetime. The evidence also suggests that the relief
operations at the time made little difference to those who survived.
Keywords: Famine, human development, Ethiopia
JEL Codes: I12, O12, J13, O15.
∗Email for correspondence: Catherine.Porter@hw.ac.uk. The data used in this paper were
collected by the University of Addis Ababa, the International Food Policy Research Institute
(IFPRI), and the Centre for the Study of African Economies (CSAE). Funding for the ERHS
survey was provided by the Economic and Social Research Council (ESRC), the Swedish Interna-
tional Development Agency (SIDA) and the United States Agency for International Development
(USAID). Thanks also to Marcel Fafchamps, John Hoddinott, Alexander Moradi, Francis Teal,
Fabrizio Zilibotti and two anonymous reviewers for useful comments on earlier versions. Any
errors and omissions remain our own.
1

1 Introduction
In October 1984 Ethiopia came to the developed world’s attention in a dramatic
BBC news broadcast from Tigray province in the Northern Highlands. The report
showed pictures of starving people on a massive scale and galvanised citizens
in Europe and the US into donating millions of pounds to relief agencies, and
putting unprecedented pressure on their governments to send humanitarian relief.
Banerjee and Duflo (2011) note “.. no single event affecting the world’s poor has
captured the public imagination and prompted collective generosity as much as
the Ethiopian famine of the early 1980’s...”(p19). Up to a million people may have
died, and many more were left destitute, making it one of the worst famines in
recent history, and on par with the Chinese famine of 1959–61 in terms of mortality
as a proportion of the population (O Grada, 2007). This paper examines what
has happened to a sample of young people who experienced this extreme shock
as infants by following up on their height attainment and other socio-economic
outcomes twenty years on from the crisis.
A small but growing economics literature documents the long-term impact of
severe shocks and famines on subsequent human development. Several papers
investigate effects of China’s Great famine (e.g. Chen and Zhou (2007)). Neel-
son and Stratmann (2010) provide recent evidence on educational impacts of the
Greek famine of 1941. Alderman et al. (2006) find negative nutritional and school-
ing impacts of civil war and drought in Zimbabwe. Examining a positive rather
than negative ‘shock’, Maluccio et al. (2009) show lasting improvements from an
experimental nutrition intervention in Guatemala.
This paper is the first study to quantify the long-term consequences of one of
Africa’s most severe famines. We have access to longitudinal data of 550 young
adults aged 17-27 years old in 2004. Our empirical strategy exploits the natural
experiment inherent in the drought crisis underlying the famine, by comparing
2

affected and non-affected siblings across cohorts both in-utero and in the first few
years after birth with their older and younger siblings.
Methodologically, our paper is an innovation on earlier work as we have access
to a measure of crisis intensity at the household level, though our sample size is
smaller. Other studies of the impact of drought or famine in other countries have
relied on shocks specified at the covariate level which, while exogenous, increases
measurement error in the measure of the crisis.1The standard geographical iden-
tification of the shock also limits their ability to isolate the impacts of the drought
from other factors. Our measure of the crisis at the household level is likely to
increase precision and offer more convincing causal attribution. However, it is
self-reported and could suffer from endogeneity bias. We deal with this in three
ways. First, we employ a battery of tests to determine whether the reported
shocks are correlated with obervables such as height and/or having a child of the
relevant age group. Second, we use a household fixed effects model to eliminate
correlations between measured famine exposure and unobservable fixed household
characteristics. Finally, we use rainfall data at the village level as an instrument
for the drought shock at the household level in order to isolate the exogenous
variation in household-level drought impacts.
We find that by adulthood, children who were under the age of 36 months
at the peak of the crisis are significantly shorter than an older cohort who were
at a less vulnerable age, by 5 cm. This is a relatively high estimate, but rep-
resents the impact on those who lived in affected households (as opposed to an
“average treatment effect”). Besides providing evidence on the impact of a spe-
cific famine on height, the paper contributes to the broader empirical literature
on the importance of early childhood nutrition for subsequent childhood devel-
opment (Almond and Currie, 2011), and future adult outcomes such as health,
1We note however two recent working papers that use SHARELIFE data on adult individual
recall of hunger periods in their childhood in wartime Europe, (Kesternich et. al. (2012), van
den Berg et. al. (2011)).
3

schooling and earnings (Deaton and Arora, 2009). We find suggestive evidence of
the presence of some of these other socio-economic effects from the famine period:
affected children may be less likely to have completed primary school, and more
likely to have experienced recent illness. Using a Mincerian framework, we can
show that the loss of human capital is economically important for these young
adults, likely leading to income losses of at least 5% per year over their lifetime.
In the next section we give more background on the 1984 famine, and introduce
the data used in the paper. Section 3 offers a conceptual backdrop, and a brief
review of the existing evidence on the impact of serious shocks and famine. Section
4 presents the econometric strategy, section 5 discusses the results and robustness
checks, section 6 concludes.
2 Background and Data: The 1984 Famine
Ethiopia has a long and troubled history of famines (Pankhurst, 1986) including
prolonged droughts and frequent severe rainfall failure. Since 1984 Ethiopia has
appeared frequently in the worldwide media because it was on the verge of famine.
The economy has experienced growth in the past decade, but seasonal hunger
continues to be an endemic feature of life in many rural areas.
Even against this difficult backdrop, the 1984 famine is still classed as one
of the worst famines ever to have hit Ethiopia, and ranks amongst the worst in
recent world history (O Grada, 2007). The impact of the famine in 1984 was
deep and broad, though as is often the case in a complex emergency, statistics are
sparse and unreliable. There is still no firm consensus on the number of deaths
that it caused, though estimates are up to a million (the upper bound being the
most popular media quotation). The main regions affected were Tigray, Eritrea
and Wollo in the North of the country, although effects were felt countrywide.
Warfare played a key role in causing famine in Tigray even before the drought
4

occurred. Military offensives, aerial bombardment of markets, destruction of cattle
and grain stores, burning of crops and tight controls on movements of migrants
and traders combined to prevent the normal redistribution of grain and livestock
surpluses in Northern Ethiopia, as documented by the Africa Watch Committee
(1991). Kidane (1990) suggests a total mortality estimate of 700,000 as reasonable.
He further notes that the famine affected all socio-economic groups equally, a
suggestion that is also made by Kumar (1990), who contrasts this with the 1974
famine that had a greater effect on less well-off groups.
Relief in the form of food aid during 1984 was delayed due to political factors: a
Marxist regime and international suspicion that it was overstating crisis statistics.
Historical reports by the Ethiopia Relief and Rehabilitation Commission (RRC
(1984)) and other accounts (Africa Watch Committee (1991), Webb et al. (1992))
contextualise the development of the crisis. From these sources, we also conclude
that 1982 was considered a ‘normal year’ of production in most regions, though
no surveys exist to corroborate this. In April 1983, however the RRC report
was alarming, and the main Meher crop2season of 1983 showed widespread crop
failure. 1984 was by all accounts a year of severe drought; in almost all regions the
Belg (minor crop) season rains failed. Segele and Lamb (2005) analysed rainfall
at a regional level over a period of 38 years (1961–99) and find 1984 is extremely
distinctive as the driest overall year – the Kiremt rains started early but then
dried up quickly, leading to an impossibly short effective growing season. They
cite rainfall deficits of up to 94% in Wollo and the Rift Valley. The drought can
be said to have lasted through late 1983 into 1984, 1985 and officially ended with
the Kiremt rains in 1986.
For the analysis we use a sample of 550 young adults (aged 17–27) from the
2The Meher is the main crop of the year, harvested after the main Kiremt rains. The Kiremt
rains account for 65–95 per cent of total annual rainfall in Ethiopia and tend to fall between
June and September. In some regions (especially south of the Rift Valley) there are two sets of
rain, the minor rainy season is the Belg rain in March/April though this varies by region.
5

sixth round of the Ethiopian Rural Household Survey (ERHS) in 2004. Their
households have been surveyed by the University of Addis Ababa and the Centre
for the Study of African Economies (CSAE) at the University of Oxford, as well
as the International Food Policy Research Institute (IFPRI) since 1994. It builds
on a survey in 1989 conducted by IFPRI to study the impact of the famine (Webb
et al. (1992)). The survey covers eighteen villages in fifteen communities3from
five regions.4Within each village, random sampling of households was used.
The attrition rate for households is low, at around three per cent per round,
reflecting the very low levels of migration in Ethiopia. Data on height are from
the sixth round of the survey in 2004, twenty years after the famine.5In 1995,
the famine information collected in 1989 was supplemented by a careful recall
module on droughts at household level. These data allow us to create a household
level variable describing the intensity of the famine, and match it with the young
adults, something that is unique in this literature.6
77% of all households in the sample cite drought as one of the shocks that has
affected them in the past 20 years, in a question that included a variety of potential
crises such as; too much rain, pest and diseases, harvest losses in storage, frost
and hailstorms. Respondents were asked to cite the three worst crises of the last
twenty years. Our drought variable is equal to one if the household (1) suffered
a substantial loss of harvest through drought between 1975 and 1995 and (2)
the household named the worst year of the crisis as Sept 84–Sept 85 (Ethiopian
3These communities are called Woredas, or the equivalent of a county in the UK. They are
further divided into what we term villages, officially called Peasant Associations (PAs), the
lowest administrative unit.
4Although representative, 18 villages is clearly not enough to make strong inference about
Ethiopia as a whole.
5A seventh round has recently been completed in end 2009, however anthropometric data
were not collected. Note that we use age data from 1994, as there is less heaping when reporting
children’s ages than adult ages.
6Specifically, the question asked in the third round of the ERHS is “In the last 20 years
[EC] has the household suffered a substantial loss of harvest through any of the following [list
of potential crises]?” The households that responded affirmatively for drought were asked to list
the three worst crisis years. EC refers to Ethiopian calendar which is eight years behind the
Gregorian calendar. It was used in phrasing ERHS household survey questions. For example
EC77 runs September 11, 1984 to September 10, 1985.
6

calendar year 1977). Of those who cite drought, the majority (72%) mentioned
the years Sept 83–Sept 85 (1975–77 EC) and 60% cite Sept 84–Sept 85 as their
worst year when asked to name the worst three years.
Table 1 column (2) shows the region of each village. Column (3) shows that
over 80% of people mention drought as a problem in approximately two thirds
(eleven) of the villages, and the fourth column shows that in these villages Sept
83–Sept 85 is the clear mode. We include the Sept 84–Sept 85 specific drought
shock at the household level (column 5). Note that the shock variable is collected
at the household level, ten years prior to the final height measurement we use
as the dependent variable in 2004. The question used refers to the households’
experiences, not those of any one child, limiting respondent bias correlated with
the height potential of one child relative to another. Despite being based on recall,
the underlying question is simple, and refers to one of the most serious crises
in these families’ life time, rendering measurement error less likely. Also, even
though drought is a covariate shock, it does not affect all individual households
in a community in the same way, depending on the specific livelihood, type of
crops grown, access to alternative water sources for water harvesting, and, given
the mountainous nature of many of the villages and its implications for the micro-
climate, the exact geographical location. In a few villages, all villagers reported
to be affected but this is a plausible outcome.7
We supplement the household measure with rainfall data as an instrumental
variable. We compile data on rainfall during the period from three sources: the
Ethiopian Meteorological office, the Food and Agricultural Organisation of the
United Nations (FAO), and the US National Climactic Data Centre (NCDC)
Global Historical Climatology Network (GHCN) database.8We use data on the
7In the village fixed effects estimates, these villages drop out due to lack of variation so we
exclude them from the analysis, similarly for the IV estimates.
8FAO data available from http://geonetwork3.fao.org/climpag/agroclimdb_en.php,
accessed July 17th 2011. Many thanks to Andreas Georgiadis for the alerting us to the lat-
ter two updated datasets, and for the use of his NCDC-GCHN data.
7

nearest rainfall station to the village which has non-missing data for the relevant
period (there is not information on all villages from any one of the time series).
We calculate mean rainfall during the period in which the comparison cohort
children were young (see below for description of cohorts: 1978 to 1987, omitting
Jan 84- Dec 85), and rainfall in 1984-85, and use the shortfall in 1984-85 compared
to the ten-year mean as our measure of the drought shock. The lowest year on
average for which we have all villages is 1984, by quite some margin. There is
variation across villages, though no village experienced rainfall above the long
term mean. Haresaw had the worst rainfall, recording only 58.6% of the 10-year
mean: in table 1, final column we calculate this as a 41.4% deficit. Two villages
had rainfall slightly above the mean (1978-87), Geblen and Yetemen (of around
1-2%). A time series of rainfall, and self-reported shocks are shown in Figure A.2
(appendix).
A further discussion of the data, including the identification of the child cohorts
used in the empirical analysis, is included in the econometric strategy.
3 Existing evidence
Almond and Currie (2011) review recent evidence on how a severe shock in a
critical period of development (usually under the age of five) may lead to persis-
tent lower levels of achievement in human capital, and this is the backdrop to our
empirical analysis. Cunha et. al’s (2006) review chapter on the economics of hu-
man capital formation provides a theoretical framework that echoes the nutrition
literature and the focus on ‘critical period programming’, showing why shocks
in childhood may have persistent impacts. A key preoccupation of the economic
literature has thus been to try and find exogenous sources of variation in nutri-
tion inputs, and drought is such an example. We provide here a necessarily brief
review of key evidence.
8

A broad literature in medicine and epidemiology contains a large number of
articles on the correlation between childhood characteristics and adult anthropo-
metric outcomes, though they are less successful at documenting causality - see
Strauss and Thomas (2008) for details. In the economics literature there is a small
but growing body of evidence on the long-term impacts of early childhood shocks.
In a seminal study, Almond (2006) found a substantial long-run effect of the 1918
influenza pandemic on US data of those in utero during the crisis, though notes
that it is impossible to separate the effect of the illness from other macroeconomic
events of the time.
Extreme negative shocks do not have the luxury of the presence of experimental
designs. At best, we can use natural experiments such as those offered by large
droughts and other crises. Strauss and Thomas (2008) note that relatively few
economic studies have identified a long–term causal impact of an extreme event
experienced in childhood, given the high data demands for such an exercise. Chen
and Zhou (2007) investigate the China famine of 1959-61 on those born between
1959 and 1962 using a difference-in-differences estimator of birth cohorts using
outcome data collected as part of a cross-section survey in 1991, and death rates
at the province level as a measure of shock severity. They use individuals born
five years before and after as the control group. Their results show a height deficit
of just over 3cm for those born in 1959. Luo et al. (2006) find increased obesity
in later life of the same cohort.
Neelson and Stratmann (2010) find educational attainment was affected by the
Greek famine of 1941, especially for those who experienced the crisis as infants.
Maccini and Yang (2009) show that even non-extreme variability in rainfall during
early life has a significant effect on a large number of future adult outcomes in
Indonesia.
While generally persuasive, these and other studies are not without their po-
tential problems (for a review and critique of several of the studies quoted, see
9

Strauss and Thomas (2008)). For example, many rely on variables defined over rel-
atively large geographical entities for distinguishing the affected and non-affected
within a particular cohort, contributing to measurement error and attenuation
bias, as well as risking that confounding factors cannot be isolated. In our study,
we define the famine shock at the household level rather than relying on a geo-
graphically defined shock.
A further issue (Strauss and Thomas, 2008) is that the impact of some of the
crises studied above likely lasted longer than the time-period specified (e.g. as
infrastructure needed to be rebuilt after a civil war). This is definitely an issue
for our study, as the Ethiopian famine took place within the context of a mostly
relatively localized but nevertheless intense civil war, so that life in any case did
not return to normal immediately after the ‘end’ of the famine. In any case, this
would bias us against finding any results and thus our findings would be a lower
bound.
Finally, we consider mortality and fertility selection. Dyson (1991) identifies
some demographic regularities of five large south Asian famines between 1876 and
1975, and finds that fertility is significantly affected by famine, and at an earlier
stage of the crisis than mortality, with birth rates well below normal before peak
mortality. We also discuss the implications of these for our analysis below.
4 Econometric strategy
We aim to test whether the Ethiopian Famine of 1984 affected children who were at
a vulnerable age – in utero or newly born, up to the age of 36 months – at the time
of the drought shock. The empirical analysis uses a reduced form specification,
given that we do not have information on other inputs in early childhood; our key
question is to identify the long-term impact of the famine. Equation 1 provides
the basis for our test.
10

Hi=
C
X
c=2
βfcf amhρc+
C
X
c=2
βcρc+γh+ei(1)
Our main focus is height Hifor each individual iat adulthood in 2004. We
compare the height attainment of a number of age cohorts cdefined across all the
young adults aged 17 to 27 in 2004. In particular, as different age groups are likely
to have been differentially affected, we specify a spline function across Ccohorts;
ρcthus identifies the cohort fixed effect. We consider four 24 month cohorts, with
the oldest group as the base group.9Table 2 shows the construction of the cohorts,
and also notes the key events in the timeline of the crisis. The oldest cohort were
over the age of four during the peak of the crisis (born between Sept 76 and
Sept 81) and therefore beyond the critical stage of development as reviewed in
the literature above, though to the extent that they were at all affected this will
reduce our estimate again towards a lower bound.10 Two cohorts are most likely
to have been affected: first, those aged 12-36 months in the crisis period (born
between Sept 81 and Sept 83). We might a priori expect these children to be the
most affected for two reasons: they were at one of the most vulnerable ages to
experience a nutrition shock (having probably stopped breastfeeding around the
time of the peak crisis), and also, in some of the villages (particularly the three
in the North of the country) the food crisis was ongoing from around 1981 due to
civil war, therefore they may have lived through a longer period of suffering and
nutritional deficit. The second cohort of concern is those in-utero or born during
the crisis (i.e. born Sept 83–Sept 85). The youngest cohort were conceived after
the peak of the crisis, so are less exposed (though it is possible that they were,
through prolonged malnutrition of the mother- see discussion below). Also, they
9Alderman, Hoddinott, and Kinsey (2006) identified the group aged 12-36 months as the
most vulnerable to a drought in Zimbabwe (though did not include a cohort fixed effect), and
we expand in both directions to cover the in-utero and infant cohort as identified e.g. by Almond
(2006).
10We include five years in this cohort in order to maximise the sample size, but start only in
1976 as another large famine in 1974 may have affected children born before 1976.
11

are aged 17–18 at the time of measurement in 2004, and therefore possibly still
growing relative to the other cohorts considered.11 We used the age data collected
during the first round of the longitudinal study in 1994. There is considerable
rounding; for example 46% of children are reported to have a rounded age (e.g.
12 years and zero months) with heaping around age 10 and 12. This will add to
measurement error, causing attenuation bias that reduces the chances of finding
an impact of the famine. Finally, as noted above, mortality rates are usually
higher for infants during crises, so we may expect the selection effect of mortality
to be highest for the ‘in utero’/infant cohort - we examine this below.
We observe a household level famine shock famhwhich is interacted with the
age (cohort) of the individual at the time of the shock. This was defined above. βf c
will measure the impact of the famine on children of cohort cliving in a household
directly affected by the drought shock. If the impact of the famine went well
beyond cohorts living in families directly affected by the famine, then this would
be picked up by values of βc, defined relative to the base group cohort that was
born well before the famine, and thus past their critical first few years of life. Our
identification strategy for the famine effects depends crucially on two elements:
a household-specific drought shock and the appropriate identification of cohorts
for comparison as discussed above. We discuss our estimating strategies below.
First, the use of household fixed effects exploiting the differences in birth timings
within households, and second an instrumental variable to purge the household
drought shock of endogeneity due to correlation with unobservable household
characteristics.
11As adolescents may still have growth spurts up to the age of 18, we will have to be cautious
in interpreting the findings, while heterogeneity in the end year of growth will increase the
variance of the estimates of the famine impact for this cohort.
12

4.1 Identification strategy: household fixed effects
Our first identification strategy is based on the inclusion of γhin equation 1 cap-
turing all household effects, ensuring that comparisons are done between cohorts
of siblings living within the same family.12 Crucially, it allows us to disentangle
famine exposure of a particular cohort from unobserved household heterogene-
ity correlated with famine exposure. However, and if anything, we still expect
that our effects may be biased downwards, giving lower bound estimates of the
long-term impact of the famine.
A standard problem in inference on the impact of large crises is positive selec-
tion into the sample of those who survived. The drought and famine will have had
a permanent impact on a large number of children through their early mortality.
Stronger, healthier children with better genetic health endowments are more likely
to have survived (Deaton et al., 2008). To the extent that such characteristics
are correlated with household fixed effects, the sibling difference model consid-
erably limits this selection bias as well. In addition, these children could have
subsequently benefited from reduced cohort sizes due to lower fertility during the
famine, or high levels of child mortality at that time. This positive selection would
lead to a downward bias in the estimated impact. It is important, however, to
note that evidence from epidemiological studies during famines and other crises
in Ethiopia and Guinea-Bissau highlight the strong predictive power of household
and parental characteristics such as wealth and parental education in explain-
ing excess mortality during such crises (Kiros and Hogan (2001), Nielsen et. al.
(2006)), so that the use of household fixed effects will reduce the likely bias.
As well as mortality selection, we must also be aware of fertility selection.
In, and just after a crisis, fertility may be affected implying our younger cohorts
12In Ethiopia, households are typically organised as nuclear families, with parents living with
their children, and married children leaving the homestead. The result is that despite high
fertility, households are relatively small, on average below 6 members per household.
13

(born after the event) may be considered as endogenous. For example, richer
or less affected families may have more children (either through biological chan-
nels, whereby their better nutrition has a positive impact on the probability of
a successful pregnancy and childbirth, or, because they choose to have children
since their circumstances have not worsened). Conversely, richer households may
deliberately plan to have fewer children in response to the adverse environment,
whereas poorer households may be less likely to actively plan fertility decisions.
Below, we test the relevance of these considerations. We also drop the younger
cohort to check robustness of the results to its exclusion.
4.2 Identification strategy: Instrumental variable
Several problems remain with household fixed effects: the differencing may in
fact increase the noise to signal ratio, and exacerbate measurement error leading
to greater attenuation bias (Bound and Solon, 1999). Further, the self-report
of the drought shock may be driven by households’ capacity to respond. There
may be ‘high capacity’ and ‘low capacity’ households, and if the shock is more
likely to happen to the latter, using evaluation language we are capturing a local
average treatment effect, rather than an average treatment effect - where the
impact includes the households’ inability to remediate the effect of the shock
(and indeed any possible reinforcing of the shock by parents who invest in other
siblings). If future famines are likely to affect certain groups then this is not
necessarily an estimate with little use - rather it gives the potential impact on
those who are likely to be affected. A further important concern is that the self-
report of the drought shock may be driven by the very fact of having a young
child in the household at the time of the crisis- leading to reverse causality.
We check the exogeneity of the drought shock and do not find any evidence of
selective reporting based on observable household characteristics. Finally, there is
14

a possibility that differences across siblings are somehow driven by macroeconomic
(or village level) factors that are experienced by each cohort.
To allay the concern that the self-reported shock is correlated with unobserv-
ables, we use an instrumental variables strategy to isolate the exogenous compo-
nent of the shock. Rainfall deficits at the village level are used as the exogenous
instrument (as discussed in the data section). We estimate a cross-section model
separately in order to rule out confounding the results with any other macroe-
conomic events or circumstances that may have affected cohorts differently over
their life course. Regressing outcomes in 2004 on the famine shock variable in a
cross-section means that we leave γhunobserved and a potential cause of endo-
geneity if correlated with famh. We therefore instrument famhwith rainv(a
village level variable) in the first stage.
4.3 Descriptive Statistics
To measure the stock of nutritional achievement of the cohorts when they are
young adults in 2004 we use their height in centimetres.
Basic summary statistics are reported in table 4. Average heights of each
cohort, separated into the affected and unaffected groups is reported in table
A.1(appendix), and we find no significant differences. Affected children are smaller
only in the cohort born 1981-83. This result does not control for unobserved
endogeneity, and our empirical strategy pursues this further.
5 Results and Discussion
5.1 Self-reported shock tests
We begin with results investigating the exogeneity of the self reported drought
shock. This variable has the advantage of measuring impact at the household
15

level. However, the concern is that it may be endogenous in ways that cannot
be controlled for by using household fixed effects, including potential unobserved
ability bias, fertility and mortality effects. To address these, in table 3 we regress
the self-reported drought shock on various variables of concern, using all house-
holds in the survey, including those with no children in the sample.13 We repeat
the analysis restricting to only our sample households in table A.3 to confirm the
results. Column (1) includes predetermined characteristics of the household head
such as gender, age, height, schooling as well as village fixed effects. None are
significant, and the village fixed effects are almost all significant, and an F-test
shows joint significance. We include a dummy variable for the presence of a child
in each of our study cohorts on the right hand side in all specifications. None of
these has a significant coefficient, allaying the concern that having a child in any
of the cohorts causes an increase in the probability of households reporting the
drought shock. In column two we replace the village fixed effects with rainfall
deviation in the drought years. This is significant and large in magnitude. A one
standard deviation deficit from mean rainfall increases the probability of reporting
the drought shock by eight percentage points (the mean number of households re-
porting the shock is 54 per cent). In this specification the height of the household
head is also significant, though its effect is around half that of rainfall - an increase
in height of one standard deviation (8cm) reduces the probability of reporting the
drought by just under four percent.14 When we include average height of the
village in column (3), the individual height impact disappears completely - which
may be picking up for example ethnic differences in height, though could also be
proxying other unobserved village factors. To investigate mortality, in column
(3) we include a variable for number of infants who died that would be in the
sample, which is significant for the larger sample but not for our subsample. We
13Descriptive statistics at household level for this regression are shown in table A.2 (appendix).
14In table A.3 column two we actually find no significant impact of height on the self-report.
16

also include child mortality in column (4) and this is not significant. A further
check is to construct a full sample of babies born in all four cohorts (including 192
children who died), and regress a binary variable indicating that the child died
before the survey on the cohort and famine interactions. We present the results
in table A.4 (appendix) which show no impact of the famine on mortality of any
of our cohorts. We test the hypothesis that the famine affected fertility differen-
tially across household characteristics by running a probit on the dummy variable
indicating that the household gave birth to a child that is in one of the study
cohorts born during or after the famine (ie born Sept 83–Sept 87, regressing each
year separately). We include age, gender, height and schooling of the household
head, the drought shock, and their interaction. Results in table A.5 (appendix)
do not show any significant impacts, except that in Sept 83–Sept 84, educated
heads who report the famine shock are 8% less likely to have a child in this year
(marginal effect). This could mean that for children born in Sept 83–Sept 84,
there may be downward bias in our estimates of famine impacts. The children
that are most affected according to the results are born in Sept 81–Sept 83 and
we find no evidence of selection through differential fertility. We also remove the
younger cohorts in some of our robustness checks. Some results show female heads
who experience the shock have lower births in several years, but there are fewer
than 20 female heads experiencing the shock in any given year so we believe this
is not a major cause for concern.15
The tests presented are exhaustive, but are based on observables, so we also
report IV regressions later in the results using the rainfall deviation as our instru-
ment, to allay concern about endogeneity caused by correlation between the self
reported shock and unobservables.
15See also an extended discussion in Appendix B about possible downward trends in child
height, and fertility selection issues.
17

5.2 Main Height Results
Table 5 shows the impact of the famine on the 550 young adults aged 17–27 in
2004. Table 4 shows means and standard deviations of the included variables. We
include the age cohorts, and their interactions with the shock variable, as well
as controls for sex of the child and its birth order. Standard errors are adjusted
for cluster-village specific heterogeneity. Column (1) displays OLS (village fixed-
effects) estimates. The cohort of children born in Sept 81–Sept 83, who were aged
12-36 months at the peak of the crisis, are 5.3cm shorter than their unaffected
cohort peers. There are no significant shock impacts on the other cohorts. We
would have expected an impact on the next cohort down, who were in utero or
infant during the crisis - the coefficient is negative, but insignificant. However, we
cannot reject that the coefficient is the same for both groups, given the standard
errors. One possible explanation fertility or mortality selection not picked up in
our tests in tables A.5 or A.4(appendix). In column (2) we restrict the sample to
households with two or more siblings in the sample and show the household fixed
effects estimate. It is slightly higher, though does not appear to be different from
OLS given the coefficients and standard errors in both equations, lending some
support to the idea that the self-reported drought estimate is exogenous.16 As
discussed above, there is a concern that, given a strong and heterogeneous fertility
response to the drought, younger children (i.e. conceived just after the crisis)
may be endogenous and therefore an unsuitable comparison group to analyse
the impact of the crisis. We therefore re-estimate the OLS village fixed effects
model with only those children already conceived prior to the crisis (our oldest
three cohorts) in column (3). This leads to a reduction in sample size to 379
individuals. The point estimate goes up by just under one cm, though this is not
16Alternatively, it is possible that any downward bias in OLS may be exactly offset by the
downward bias potentially caused by the exacerbated measurement error problem in the fixed-
effects model.
18

likely to be substantively different from that of the larger sample, given the point
estimates and standard errors. In the fourth column we repeat the analysis using
household fixed-effects, and the point estimate is also very similar.17
The results in table 3 confirmed no correlation between observable household
characteristics and the drought shock. We check the robustness of the results
to any endogeneity driven by unobservables by implementing an IV strategy, as
outlined above, using village rainfall deviation in 1984-85 as an instrument for
the self-reported drought shock. Column (5) of table 5 shows the reduced form
specification, confirming the self-reported results. We present results for each of
the three oldest cohorts separately in table 6 to remove the possibility of any
unobserved macroeconomic events/trends driving the between-cohort results, and
estimate first an OLS cross-section and then an IV. IV estimation with weak in-
struments and small sample size could lead to bias in our results towards OLS,
and standard errors may be biased downwards, (Stock and Yogo, 2002; Murray,
2006). We report the Stock and Yogo (2002) critical values for strength of the in-
struments, and Fuller IV estimates, a weak-IV robust estimator (Murray, 2006).18
In the OLS we do not include village fixed effects in the model (as we cannot in-
clude these in the following IV estimates, due to having a village level instrument).
Columns (1-2) are the youngest cohort (born Sept 83–Sept 85), columns (3-4) are
the cohort aged 12-36 months during the crisis (born Sept 81–Sept83) and the fi-
nal two columns are the oldest cohort (born prior to Sept 81). We have a weak IV
for the oldest cohort, however for the other cohorts the IV is fairly strong, given
17Note that table 5 shows that coefficients also for the younger cohort are negative (even
though they were not in utero or born during the famine) and it may be suggested that drought
affected areas may be facing a downward trend, rather than a clear famine effect. See web
appendix for more details, in summary we ran a year-by-year fixed effects regression, with no
controls and the findings suggest that this explanation could not be discounted. We explored
further by adding extra cohorts, and figure A.1 (appendix) shows the results, which qualitatively
confirm the lack of a downward trend. Many thanks to an anonymous referee for pointing out
the relevance of exploring this issue further.
18We also compared the results between the Fuller estimates and Limited Information Max-
imum Likelihood which is an alternative estimator that is also robust in the presence of weak
instruments. The results were effectively the same.
19

the small sample size. Our first stage results are reported in table A.6(appendix).
A significant impact is again found for the 12-36 month old cohort, confirming the
pooled results. The IV has a Kleibergen-Paap F-statistic of 19.92, which is above
the 10% maximal bias in the Stock and Yogo (2005) critical values table.19 The
point estimate is slightly higher (9cm), which could suggest some downward bias
in the fixed effects model, as discussed above.20 Is it totally implausible that the
effect be so large? First, this was a famine of historical scale. Second, we find that
even in areas quite strongly affected by the drought not everybody was affected.
So the IV estimate gives the impact on those affected by the drought (implying
a local average treatment effect on the 60% who were affected). In other studies,
famine impacts have been estimated at the geographical level so they calculate the
average impact across sufferers and non-sufferers. As a benchmark, the reduced
form estimate for our instrument on height(table 5) shows an average of 5.1cm at
the mean rainfall deficit. We also note that Van Den Berg et al (2012) find much
higher impacts of European famine periods near WWII using an IV strategy (for
example, 7cm height loss for males).
In sum, using several different estimation techniques we find a strong and
significant effect of the drought and famine shock of Sept 84–Sept 85 on children
2–3 years of age at the time of the famine. In terms of the literature, many
studies have found significant impacts of infection or disease on children in the
womb (Almond, 2006), though nutrition studies (e.g. Alderman et al. (2006))
have also found impacts on children of a similar age to those in our results.
19We did attempt to also estimate across the pooled model. However, instrumenting the
interaction of the cohort and the drought shock with the interaction of the cohort and the
instrument resulted in weak instruments.
20Though the standard errors are somewhat larger, and thus we do not necessarily claim that
the estimate is different.
20

5.3 Robustness and Further Interpretation
We also explored whether the famine had long term impacts on any other human
capital outcomes for these cohorts, such as health, BMI or schooling (see table
A.7, appendix). For schooling, we ran a probit on whether the child had any
school at all (28% have no schooling). We found that all of the cohorts were
around 10% less likely to have schooling if their households were impacted by
the famine, and we cannot reject the equality of the coefficients for all cohorts.
However there are too few households with variation across schooling to explore
the fixed-effects strategy. We also ran a probit on whether or not the young person
was ill in the four weeks prior to the 2004 survey. The coefficient on the drought
interaction term for children born in 1982-83 is significant at 10% but the marginal
effects reported in the table were just insignificant (p=0.14). Even this result is
quite interesting, given our relatively small sample size, as morbidity measures are
generally weak discriminators for health and morbidity is measured 20 years after
the event. We also ran regressions on BMI, but there were no significant effects.
For all the specifications, we also explored whether there were any gender-specific
effects across all the results reported. We find that there is never a jointly or
cohort-specific significantly different effect for girls and boys due to the famine
(see table A.8 (appendix)).
The results on height are not driven by opportunistic definition of the relevant
period for the famine shock. For example, repeating the analysis by replacing the
famine shock as defined here referring to the peak period (in Sept84–Sept85) to a
broader period Sept83–Sept85 (i.e. households that also reported this period as
their worst period) made no difference to the results (See table A.10, appendix).
Qualitative work in some of the communities by Bevan et al. (1994) has suggested
that the crisis may have started earlier in Tigray, and the communities covered
by the survey in these regions, as rains may have failed in 1981–82, and also
21

because this was the specific region most affected by the conflict.21 The result
is that for these communities (Haresaw, Geblen), the use of the older cohort as
the base group may not be providing an appropriate counterfactual, potentially
underestimating the impact the famine as a result. As a robustness test, we
interacted the all variables with a Tigray dummy and the interaction terms on
the drought interactions were insignificant, both individually and jointly. This is
shown in table A.9 (appendix).
As the data in 2004 come from a longitudinal data set, we pursue the likely
impact of any other, post-famine, attrition in the data. For example, even if
children survived, differential strength and intellectual ability may have affected
migration or marriage of those affected and those not affected differentially.22
Relative to 1994, we have considerable attrition for the age-specific sibling
sample used in this study. Of the initial sample in 1994, 21% of the children
moved away and about 3 percent died. In our fixed-effects model, we lose another
fifth of the observations, as our identification requires at least 2 observations per
household. Missing height measurements of children or of relevant siblings adds
further to attrition. However, we find no evidence that this attrition post-1994 is
systematically affecting our results. We estimated a probit on remaining in the
sample using the same correlates as in table 5. We found those who were younger
predictably being more likely to be in the sample, and females less likely to be
observed. Crucially, however, for the famine interactions, there were no significant
differences either individually or jointly (p=0.35); those affected were not more
or less likely to be observed in the sample. We conclude that despite significant
21Intensive qualitative research in the other communities covered by the survey has suggested
that the civil war and conflict in Ethiopia was not affecting them until much later, in the late
1980s, contrary to the communities in Tigray.
22In 1994, the sample would have been too young to live independently away from home. In
recent times, children often live away from home for educational reasons, as secondary schools
are only in towns. However, in 1994, few did so in rural Ethiopia. Net primary school enrolment
rates in the sample were only around 20%; and below 10% for secondary education. Even so,
these children would have been recorded as in the household.
22

attrition, there is nothing systematic that would bias our analysis.
All the long-term impact estimates are net of coping strategies that households
may have undertaken in order to mitigate the short-term impact of the shock on
household consumption. They are also net of any food aid or other relief that may
have taken place. We have some information on whether food aid was received
during this period, as well as how the household responded to the crisis. In
1994, detailed data were collected recalling this traumatic period. Contrary to
impressions created by the media during famines, few made it to feeding camps
(under 5%). Food aid targeting is always difficult, as the general equilibrium
effects of a famine mean that all face some problems, such as those linked to
rising food prices. Unsurprisingly, in view of what happened, we find relatively
imprecise targeting: during Sept84-Sept85, when food started being distributed,
41% of those affected by the drought shock obtained some food aid, compared to
25% of unaffected. Still, amounts received per capita were considerably higher for
those reporting the shock compared to the others. Median receipts for both groups
is zero, and a mean 97kg per affected household is relatively little for a whole year.
Imperfect targeting of aid, and the scale of the crisis meant that households had
to resort to costly coping strategies. Of those affected by the crisis, 91% cut back
on meal sizes, compared to 65% of those not reportedly affected; 61% of those
affected ate wild foods they would not normally eat (29% of those unaffected),
and 48% sold valuable assets (24% of those unaffected).23
Is there any evidence that the food aid reduced the long-term impact of the
famine on our cohorts? Non-random placement of famine relief makes this a
difficult topic, so at most, we can only provide some indicative evidence. We in-
teracted the cohort and cohort-famine shock variables with a dummy for receiving
food aid (table A.11, appendix). None of the interaction effects were significant,
23The relatively high percentages of those not reporting to be directly affected by drought but
still using particular coping mechanism is a reflection of the likely general equilibrium effects,
making our estimates of impact again more likely to be a lower bound.
23

neither individually, nor jointly across the specific groups affected by the famine
(p=0.67) or across the cohorts (p=0.41). Similarly, using amounts of food aid
received, there were no significant interaction effects. We should be cautious in
interpreting these results: it could mean that food aid was irrelevant on average;
for example, it was just too little to make any difference. By lack of an appropri-
ate counterfactual, it could also have been so well targeted that those receiving
food aid would have been worse off than those not receiving it, and that now they
have equal opportunities. In view of the evidence on targeting described above,
and the key findings of this paper, this would seem rather unlikely. On the basis
of our evidence, it would be hard to defend that this was a well-handled successful
famine relief operation.
What are the implications of this loss in human capital for the affected cohort?
A number of studies have found a positive and significant relationship between
height and earnings, in many country contexts. We ran some simple Mincerian
regressions of total annual income, using a sample of household heads from the
2004 round of the ERHS (i.e. the full sample and a larger sample than the young
adults we could trace back to the famine). We included village fixed-effects, gender
of the head, age of the head and its square, schooling of the head and household
size and composition variables.24 We found that a 1 cm increase in height results
in approximately 1% increase in annual income (See table A.13 (appendix)).
24Note that in the ERHS there are very few people who work for wages (approximately
350 with non-missing values), as most people are occupied on the family farm. Whilst this is
therefore clearly an incomplete model of income generation, our aim is to provide some basic
correlations of the relationship between height of the household head and income generated by
the household as a whole. As a robustness check we also estimated the same equation for crop
income, and the results were very similar. Table A.12 (appendix) shows descriptive statistics
for the sample of household heads.
24

6 Conclusion
This paper contributes to a small body of economic evidence on long-term impacts
of extreme events experienced in early childhood, by providing the first estimates
of the impact of one of the biggest famines to have hit Africa, the case of Ethiopia
in 1984.
The use of a family-specific famine shock allows us to distinguish this effect
from a more general macro-effect affecting all in particular communities, while the
within-household estimator ensures that any household-level heterogeneity corre-
lated to famine exposure is not biasing our results. The effects are large, but
could still even be only a lower bound. First, there could still be selection into
the sample of probably stronger children, if mortality rates were higher for shorter
children. While mortality is considerably determined by family and community
characteristics, and therefore controlled for by the household fixed effects regres-
sor, within family effects could have led to relatively stronger children surviving.
Second, the comparison group is other children who were alive at the time of a
severe famine, despite their household not reporting it as the worst year– it is
highly likely that every village in Ethiopia was affected by this to a greater or
lesser extent, and if so this will make the difference between affected and non-
affected children smaller. Third, we have to contend with some heaping of age,
and the drought instrument is relatively ‘blunt’ in the sense that we cannot be
more precise (e.g. to the month) about the length of the shock. These measure-
ment issues cause attenuation bias in the estimates, making it less likely that we
find any effects.
The results presented show that children who experienced but survived a large
scale and severe nutritional shock at a critical period in their development are
discernibly smaller than their peers when measured twenty years later. We find
that those in the particularly vulnerable age of 12 to 36 months at the height of
25

the famine were about 5 cm shorter due to the famine. We cannot reject that
all those in utero and those below the age of 36 months were all similarly and
significantly affected. The loss can be compared it to the findings summarised in
Strauss and Thomas (2008) that developing countries gained an average of 1 cm
in height per decade. However the famine impact in Ethiopia is also in line with
findings from other serious famines. For example, the results on China quoted in
Chen and Zhou (2007) suggest on average a height reduction of 3.03 cm due to
the 1959–61 famine, with further effects on labour supply and earnings. Van den
Berg, Pinger, and Schoch (2011) find impacts of around 7cm for post-war europe,
when looking at individual shock measures. We also find some tentative evidence
that those vulnerable and affected at the time of the famine may be more likely to
be ill and less likely to have schooling in adulthood. Indicative calculations show
that the observed height deficit could lead to reduced income of around 5% per
annum. Our analysis also suggests that famine relief in the form of food aid did not
appear to have been effective in reducing impacts on the most vulnerable children
(who survived the crisis), despite massive aid efforts. This study thus adds to a
body of knowledge on the long-term impact of severe shocks, and underlines the
importance of swift nutritional interventions in complex emergencies, specifically
targeting children who are at a critical stage in their development.
References
Africa Watch Committee (1991): Evil days : 30 years of war and famine in
Ethiopia, Africa Watch report. Human Rights Watch, New York ; London.
Alderman, H., J. Hoddinott, and B. Kinsey (2006): “Long term conse-
quences of early childhood malnutrition,” Oxford Economic Papers, 58(3), 450.
Almond, D. (2006): “Is the 1918 Influenza Pandemic Over? Long-Term Effects
26

of In Utero Influenza Exposure in the Post-1940 U.S. Population,” Journal of
Political Economy, 114(4), 672–712.
Almond, D., and J. Currie (2011): “Human Capital Development before Age
Five,” in Handbook of Labor Economics, ed. by O. Ashenfelter, and D. Card,
vol. 4, Part B of Handbook of Labor Economics, pp. 1315 – 1486. Elsevier.
Banerjee, A. V., and E. Duflo (2011): Poor economics : a radical rethinking
of the way to fight global poverty. PublicAffairs, New York.
Bevan, P., B. Kebede, and A. Pankhurst (1994): “Ethiopian Village Stud-
ies,” Discussion paper, Centre For the Study of African Economies.
Bound, J., and G. Solon (1999): “Double trouble: on the value of twins-based
estimation of the return to schooling,” Economics of Education Review, 18,
169–182.
Chen, Y., and L.-A. Zhou (2007): “The long-term health and economic conse-
quences of the 1959-1961 famine in China,” Journal of Health Economics, 26(4),
659–681.
Cunha, F., J. J. Heckman, L. Lochner, and D. V. Masterov (2006):
“Interpreting the Evidence on Life Cycle Skill Formation,” in Handbook of the
Economics of Education, ed. by E. Hanushek, and F. Welch, vol. 1. Elsevier,
North-Holland.
Deaton, A., and R. Arora (2009): “Life at the top: the benefits of height,”
Economics and Human Biology, 7(2), 133–136.
Deaton, A., C. Bozzoli, and C. Quintana-Domeque (2008): “Adult Height
and Childhood Disease,” Demography, 46(4), 647–669.
Dyson, T. (1991): “On the Demography of South Asian Famines: Part 1,”
Population Studies, 45(1), 5–25.
27

Kesternich, I., B. Siflinger, J. P. Smith, and J. K. Winter (2012): “The
Effects of World War II on Economic and Health Outcomes across Europe,” IZA
Discussion Papers 6296, Institute for the Study of Labor (IZA).
Kidane, A. (1990): “Mortality estimates of the 1984-85 Ethiopian famine,” Scan-
danavian Journal of Social Medicine, 18(4), 281–6.
Kiros, G.-E., and D. P. Hogan (2001): “War, famine and excess child mor-
tality in Africa: the role of parental education,” International Journal of Epi-
demiology, 30(3), 447–455.
Kumar, B. (1990): “Ethiopian famines 1973-1985: a case study,” in The Political
Economy of Famine, Vol. II: Famine Prevention, ed. by J. Dr`eze, and A. Sen,
vol. II, pp. 173–216. Clarendon Press, Oxford.
Luo, Z., R. Mu, and Z. Xiaobo (2006): “Famine and Overweight in China.,”
Review of Agricultural Economics, 28(3), 296 – 304.
Maccini, S., and D. Yang (2009): “Under the Weather: Health, Schooling,
and Socioeconomic Consequences of Early-Life Rainfall,” American Economic
Review, 99(3), 1006–1026.
Maluccio, J. A., J. Hoddinott, J. R. Behrman, R. Martorell, A. R.
Quisumbing, and A. D. Stein (2009): “Effect of a nutritional intervention
during early childhood on economic productivity among Guatemalan adults,”
The Lancet, 119(537), 734763.
Murray, M. P. (2006): “Avoiding Invalid Instruments and Coping with Weak
Instruments,” Journal of Economic Perspectives, 20(4), 111–132.
Neelson, S., and T. Stratmann (2010): “Effects of Prenatal and Early Life
Malnutrition: Evidence from the Greek Famine,” Working Paper 2994, CESifo.
28

Nielsen, J., H. Jensen, P. K. Andersen, and P. Aaby (2006): “Mortality
patterns during a war in Guinea-Bissau 1998-99: changes in risk factors?,” Int
J Epidemiol, 35, 438 – 446.
Norton, E. C., H. Wang, and C. Ai (2004): “Computing interaction effects
and standard errors in logit and probit models,” Stata Journal, 4(2), 154–167.
O Grada, C. (2007): “Making Famine History,” Journal of Economic Literature,
45(1), 5–38.
Pankhurst, R. (1986): The history of famine and epidemics in Ethiopia : prior
to the twentieth century. Relief and Rehabilitation Commission, Addis Ababa.
RRC (1984): Review of the current drought situation in Ethiopia. Relief and
Rehabilitation Commission, Addis Ababa.
Segele, Z. T., and P. J. Lamb (2005): “Characterization and variability
of Kiremt rainy season over Ethiopia,” Meteorology and Atmospheric Physics,
89(1), 153–180.
Stock, J. H., and M. Yogo (2002): “Testing for Weak Instruments in Linear
IV Regression,” Discussion paper, National Bureau of Economic Research, Inc.
Strauss, J., and D. Thomas (2008): “Health over the life course.,” in Handbook
of Development Economics, Volume IV, ed. by T. P. Schulz, and D. Thomas,
vol. 4. Elsevier Press.
Van den Berg, G. J., P. Pinger, and J. Schoch (2011): “Instrumental
Variable Estimation of the Causal Effect of Hunger Early in Life on Health
Later in Life,” IZA Discussion Papers 6110, Institute for the Study of Labor
(IZA).
29

Webb, P., J. von Braun, and Y. Yohannes (1992): “Famine in Ethiopia:
Policy Implications of Coping Failure at National and Household Levels,” Dis-
cussion Paper 92, International Food Policy Research Institute (IFPRI).
30

A. Tables
Table 1: Famine and Drought, by Village
Village Region % Drought % Drought % Worst Rain
affected in 83-5 84-5 deficit
Haresaw Tigray 1.00 1.00 0.77 .414
Geblen Tigray 1.00 1.00 0.15 -.024
Dinki Shewa 1.00 1.00 1.00 .032
Yetemen Gojjam 0.59 0.41 0.41 -.012
Shumsha Wollo 0.93 0.71 0.64 .289
Adele Keke Harerghe 0.68 0.54 0.46 .476
Korodegaga Arssi 1.00 1.00 0.78 .255
Trurufe Ketchema Shewa 0.87 0.76 0.65 .206
Imdibir Shewa 0.83 0.83 0.83 .227
Aze Deboa Shewa 1.00 1.00 1.00 .193
Adado Sidamo 0.00 0.00 0.00 .335
Gara Godo Sidamo 1.00 1.00 0.76 .354
Do’oma Gama Gofa 1.00 0.86 0.61 .176
Debre Berhan-Milki Shewa 1.00 1.00 1.00 .085
D.B. -Kormargefia Shewa 0.25 0.25 0.13 .085
D.B. -Karafino Shewa 0.20 0.20 0.20 .085
D.B. -Bokafia Shewa 0.00 0.00 0.00 .085
Notes: EC refers to Ethiopian calendar which runs from September to September used in
household survey questions, e.g. EC77 runs September 11, 1984 to September 10, 1985.
1) Specifically, responded ‘drought’ to the question “In the last 20 years has the household
suffered a substantial loss of harvest through any of the following [list of potential crises]?”
2) Households were asked to list the three worst crises, this entry is positive if the household
responds EC75, EC76 or EC77 (Sept1982–Sept1985) 3) In the list from (2), household ranks
EC77 (1984–85) as the worst crisis. 4) Rainfall deficit is that of Jan84-Dec85 compared to the
average from 1978-87 (excluding 84-85). A positive number indicates a deficit.
Rainfall data from the Ethiopian Meteorological Agency. We supplement missing data using two
other sources: FAO database at http://geonetwork3.fao.org/climpag/agroclimdb_en.php
and National Climatic Data Center (NCDC)-Global Historical Climatology Network (GHCN)
Monthly data (received as stata dataset from Andreas Georgis, University of Oxford). The
relevant average is the decade around the birth of the cohort children 1978-1988 (Sept84-Sept85).
We compute the rainfall in Sept84-Sept85 as a proportion of the decade around it. In robustness
checks we also compute the total for the years only prior to 1984 when we use the older cohort.
Villages which have imputed data: Haresaw: Mekele (15K) FAO data before 1988. Aze Deboa:
Mar, May, Jun 84 missing. We compute the long term average for all years based on the 9
non-missing months for 1984. Debre Birhan: 1980-83 missing. We impute using the percentage
change in the next nearest station (Alem Ketema (52Km away and 1000m lower in altitude),
using the assumption that whilst levels of rainfall are quite different, the deviations year by
year are correlated. Gara Godo: Use FAO data Sodo (31KM). 1983 is missing and there is no
rainfall data available on any station within 100KM. Therefore we calculate averages excluding
this year. Do’oma: Imputed as for Debre Birhan 1982-85. Do’oma is a resettlement village. We
exclude it in some of the robustness checks.
31

Table 2: Dates of famine and cohort ages
Dates Famine related
events
Cohort birth
and age in
famine
Age 2004
survey
Sept 76 - Sept 80 Relatively normal har-
vest
“Born before
Famine” (Oldest
cohort aged
37-104 months
in famine)
23–27
Sept 80 - Sept 81 Relatively normal har-
vest
Sept 81 - Sept 82 Bumper rains and har-
vests in many places
during this year
“Just before
famine” Age
12-36 months in
famine
21–22
Sept 82 - Sept 83 Drought, crop failure,
war spread in the
north
Sept 83 - Sept 84 Peak of the famine
continues through
two years- widespread
hunger, death.
“Born during
Famine” Born
and in-utero
during severe
shock period
19–20
Sept 84 - Sept 85 Famine peaks in many
areas
Sept 85 - Sept 86 Normal Meher
rains (Aug/Sept
86) marked the end of
drought
“Born af-
ter Famine”
(Youngest co-
hort)
17–18
Sept 86 - Sept 87 Normal year (but
some problems in
certain villages this
year)
Notes: Dates run from September to September to match the Ethiopian calendar which begins
on 11th September - which is how households refer to key events in the survey. Information on
dates is discussed and referenced in the main text.
32

Table 3: Determinants of self-reported drought shock
(1) (2) (3) (4)
Rainfall dev. 1984-1985 .627 .462 .484
(.118)∗∗∗ (.121)∗∗∗ (.121)∗∗∗
Age of hh head .002 .0006 .0005 .0004
(.001) (.001) (.001) (.001)
Household head female -.278 -.591 -.619 -.617
(.824) (.490) (.452) (.456)
Height of HH head -.002 -.006 -.003 -.003
(.003) (.003)∗∗ (.003) (.003)
Head height*female head .002 .004 .005 .005
(.006) (.006) (.006) (.006)
HH head any schooling -.034 7.48e-06 -.014 -.016
(.044) (.040) (.041) (.041)
Mean height in village -.050 -.049
(.009)∗∗∗ (.009)∗∗∗
Have child born Sep85-Sep87 .048 .002 .004 .006
(.045) (.043) (.044) (.044)
Have child born Sep83-Sep85 .017 -.020 -.027 -.019
(.048) (.047) (.047) (.047)
Have child born Sep81-Sep83 .046 .038 .037 .032
(.055) (.053) (.054) (.054)
Have child before Sep 81 .020 .031 .029 .038
(.038) (.036) (.037) (.036)
No. infants died cohort age .035 .039
(.022) (.021)∗
No. infants died in famine .024
(.056)
Village F.E. Yes No No No
Obs. 985 985 985 985
Notes: Probit estimates. Dependent variable=1 if household self-reports
Sep84-Sep85 as drought “worst year” (see table 1 for full definition).
Rainfall deviation is from 10 year mean. All coefficients reported as
marginal effects, robust standard errors in brackets. *significant at 10%, **
significant at 5%, *** significant at 1%. Mortality estimates are from 1997
survey mother fertility history, based on recall of all births and deaths and
their dates. ”Died would be in sample” means the child was born in one of
the included cohorts but subsequently died. ”Died in famine” means died
during Sep84-Sep85. Included but not reported in column (2, 3, 4) are
average and variance of village rainfall and distance to town.
Table 4: Summary statistics: Child Level
Variable Mean Std. Dev. N
Height (CM) 2004 159.83 11.87 550
BMI 2004 19.15 2.59 548
Illness or injury 0.09 0.29 544
Any schooling 0.73 0.45 442
Female 0.41 0.49 550
Birth Order 2.98 1.40 550
Born Sep85-Sep87 0.31 0.46 550
Born Sep83-Sep85 0.25 0.43 550
Born Sep81-Sep83 0.18 0.39 550
Born before Sep 81 0.26 0.44 550
Drought shock 0.59 0.49 550
Rainfall dev. 1984-85 0.23 0.14 550
Age of hh head 47.48 11.10 550
Household head female 0.16 0.37 550
Height of HH head 165.91 8.84 550
Head ever attend school 0.21 0.41 547
Number died cohort age 0.49 0.89 550
Number died in famine 0.07 0.26 550
Notes: Sample is children of the household head aged 17–27 years old in
2004, definition of cohorts in table 2, definition of drought shock in table 1.
33

Table 5: Drought Impact on 2004 Height: OLS and Fixed-Effect estimates
(1) (2) (3) (4) (5)
Drought shock 2.074 1.592
(1.492) (1.426)
Drought*Born Sep85-Sep87 -.429 -3.811 -21.929
(3.091) (3.026) (7.163)∗∗∗
Drought*Born Sep83-Sep85 -2.454 -2.002 -2.949 -4.167 -9.826
(2.336) (2.892) (2.193) (3.188) (6.974)
Drought*Born Sep81-Sep83 -5.272 -6.287 -5.990 -5.707 -18.505
(2.640)∗∗ (2.672)∗∗ (2.547)∗∗ (2.657)∗∗ (10.681)∗
Born Sep85-Sep87 -7.377 -7.490 -2.498
(2.444)∗∗∗ (3.092)∗∗ (1.460)∗
Born Sep83-Sep85 -2.146 -4.957 -2.321 -5.668 -1.338
(1.774) (2.140)∗∗ (1.695) (2.532)∗∗ (1.294)
Born Sep81-Sep83 -.643 1.950 -.059 .488 .646
(2.187) (1.898) (2.261) (1.997) (3.029)
Age of hh head .067 -.001 .072
(.046) (.047) (.049)
Household head female .918 .838 .957
(1.743) (1.993) (1.737)
Height of HH head .187 .141 .189
(.097)∗(.103) (.097)∗
Head ever attend school 2.566 1.368 2.612
(1.220)∗∗ (1.294) (1.212)∗∗
Female -6.948 -8.493 -8.579 -8.745 -6.891
(1.211)∗∗∗ (1.340)∗∗∗ (1.124)∗∗∗ (1.554)∗∗∗ (1.138)∗∗
Obs. 550 369 379 199 550
Notes: (1) and (3) include village fixed effects, (2) and (4) include household fixed
effects. (2) restricts the sample to only those with siblings. (3) and (4) repeat the
analysis dropping the youngest cohort. Omitted cohort is the oldest, born 1978-81.
In column (5) the drought is defined as the rainfall deficit Jan84-Dec85, as in column
(5) of table (1), and used as the instrumental variable in table 6.
Table 6: IV estimates of height - older cohorts
(1) (2) (3) (4) (5) (6)
Drought shock -1.133 -16.014 -4.403 -9.781 1.746 8.002
(1.519) (109.956) (2.646)∗(5.567)∗(1.400) (7.206)
Obs. 136 136 100 100 143 143
Notes: * significant at 10%, ** significant at 5%, *** significant at 1%.Robust
standard errors in brackets these are not clustered at village level due to small
number of villages (and note that clustering reduced standard errors).Omitted group
is those born before 1982. Columns (1, 3, 5) are OLS, for cohorts born in
Sept83–Sept85, Sept81–Sept83 and pre-Sept81 (ie from youngest to oldest). Column
(2,4,6) are IV Fuller(1) estimates on the same samples, with the same controls as in
table 5.
34

Online Appendix A: Additional Tables
-6
-4
-2
0
2
4
6
Relative height by birth cohort (extended to 1992)
Figure A.1: Year by year height differences- extended
Table A.1: Height by cohort and famine shock
Cohort Unaffected Affected T-test diff N
p-value
Youngest group 155.102 156.3061 0.57 (0.56) 171
(Born Sep 85–Sep 87)
In utero-12months in famine 160.241 159.693 0.54 (0.58) 136
(Born Sep83–Sep 85)
Aged 13-36 months in famine 163.356 159.367 1.51 (0.14) 100
(Born Sep 81–Sep 83)
Aged 36 months + in famine 163.450 164.412 0.63 (0.52) 143
(Born Sep 75-Sep 81)
Notes: Sample is children aged 17–27 years old in 2004, definition of cohorts in table 2, definition
of drought shock in table 1. P-value of two-sided t-test in brackets.
A-1

0
10
20
30
40
50
60
70
80
600
700
800
900
1000
1100
1200
1300
Percent report shock
Rainfall (mm)
Figure A.2: Rainfall time series and self-reported drought shock
Notes: Rainfall total in mm as average for all 15 ERHS villages as described in Table 1, and
refers to the year preceding the self-report. Proportion reporting drought shock in each year
also as described in Table 1.
Table A.2: Summary statistics:Household Level
Variable N Mean Std. Dev. Min Max
Variable Obs Mean Std. Dev. Min Max
Household drought shock 985 .544 .498 0 1
Age of household head 985 45.975 15.184 15.875 101
Household head female 985 .229 .421 0 1
Height household head 985 164.726 8.637 80 186
Household head went to school 985 .256 .437 0 1
Have a child born Sep85-Sep87 985 .190 .392 0 1
Have a child born Sep83-Sep85 985 .157 .364 0 1
Have a child born Sep81-Sep83 985 .117 .321 0 1
Have a child born before Sep 81 985 .461 .499 0 1
Rainfall dev. 1984-85 from 10 year mean 985 .215 .149 -.024 .476
No. infants died who would be in sample 985 .394 .838 0 6
No. infants died in famine 985 .082 .293 0 2
Notes: Sample is household level including all households used in table 3. Definition of cohorts
in table 2, definition of drought shock in table 1.
A-2

Table A.3: Determinants of self-reported drought shock
(1) (2) (3) (4)
Rainfall dev. 1984-1985 from 10 year mean 1.090 .931 .939
(.214)∗∗∗ (.213)∗∗∗ (.215)∗∗∗
Age of hh head .002 .002 .001 .001
(.003) (.003) (.003) (.003)
Household head female .097 -.305 -.525 -.491
(1.731) (1.612) (1.157) (1.262)
Height of HH head -.002 -.005 -.003 -.003
(.004) (.004) (.004) (.004)
Head height*female head -.00004 .002 .004 .004
(.011) (.011) (.011) (.011)
HH head any schooling .032 .093 .060 .055
(.074) (.067) (.070) (.070)
Mean height in village -.049 -.050
(.016)∗∗∗ (.016)∗∗∗
Have a child born Sep85-Sep87 .078 .038 .045 .049
(.064) (.061) (.062) (.062)
Have a child born Sep83-Sep85 .063 .033 .032 .039
(.061) (.060) (.060) (.060)
Have a child born Sep81-Sep83 .047 .028 .030 .033
(.063) (.060) (.060) (.060)
Have a child before Sep 81 -.025 .010 -.001 .009
(.066) (.065) (.065) (.065)
No. infants died who would be in sample -.006 -.004
(.031) (.029)
No. infants died in famine .119
(.100)
Village F.E. Yes No No No
Obs. 357 357 357 357
Notes: Probit estimates. This table replicates results of table 3, but using as the sample only
those households which are included in the main regressions for height. Dependent variable=1 if
household self-reports Sep84-Sep85 as drought “worst year” (see table 1 for full definition). All
coefficients reported as marginal effects, robust standard errors in brackets. *significant at 10%,
** significant at 5%, *** significant at 1%. Mortality estimates are from 1997 survey mother
fertility history, based on recall of all births and deaths and their dates. “Died would be in
sample” means the child was born in one of the included cohorts but subsequently died. “Died
in famine” means died during Sep84-Sep85. Included but not reported in column (2, 3, 4) are
average and variance of village rainfall and distance to town.
A-3

Table A.4: Determinants of child mortality
(1)
Born Sep85-Sep87 -.187
(.039)∗∗∗
Born Sep83-Sep85 -.144
(.040)∗∗∗
Born Sep81-Sep83 -.146
(.042)∗∗∗
Drought*Born Sep85-Sep87 .024
(.063)
Drought*Born Sep83-Sep85 -.003
(.060)
Drought*Born Sep81-Sep83 -.007
(.070)
Drought*Born before Sep 81 .019
(.044)
Obs. 885
Notes: Probit estimates. Dependent variable=1 if child died before the survey (192 children).
All coefficients reported as marginal effects, robust standard errors in brackets. *significant at
10%, ** significant at 5%, *** significant at 1%. Mortality estimates are from 1997 survey
mother fertility history, based on recall of all births and deaths and their dates.
A-4

Table A.5: Determinants of fertility
Borns83s84 Borns84s85 Borns85s86 Borns86s87 Borns87s88
(1) (2) (3) (4) (5)
Drought shock -.541 .013 .079 -.354 -.758
(.665) (.593) (.606) (.628) (.515)
Age of hh head .0002 -.0009 -.001 -.001 -.002
(.001) (.001) (.001) (.001) (.0009)∗∗
Household head female .988 -.465 .135 -.967 -.088
(.040)∗∗∗ (.377) (1.514) (.088)∗∗∗ (.378)
Height of HH head -.003 -.001 .002 -.003 -.0009
(.003) (.003) (.003) (.003) (.002)
Head height*female head -.009 .006 -.002 .018 .0001
(.005)∗(.005) (.008) (.008)∗∗ (.004)
HH head any schooling -.059 .015 -.070 -.085 -.019
(.038) (.045) (.040)∗(.038)∗∗ (.034)
HH head age*shock -.001 -.001 .0007 -.002 .001
(.001) (.001) (.002) (.001) (.001)
HH head female*shock -.312 .570 .930 .993 .973
(.243) (1.927) (.160)∗∗∗ (.016)∗∗∗ (.053)∗∗∗
HH head height*shock .003 .0008 -.0007 .003 .004
(.004) (.004) (.004) (.003) (.003)
HH head height*female*shock .008 -.002 -.006 -.018 -.005
(.008) (.007) (.010) (.010)∗(.006)
HH head school*shock .029 -.088 .011 .074 -.012
(.063) (.034)∗∗∗ (.062) (.074) (.045)
Obs. 985 985 985 985 985
Notes: Dependent variable is a dummy whether the household gave birth in the year specified
in the column name, where s83s84 refers to the year from Sept 83 to Sept 84 (for consistency
with the Ethiopian Calendar). Marginal effects. Robust standard errors. *** Significant at 1%,
** Significant at 5%, * Significant at 10%.
A-5

Table A.6: IV estimates : First stage regressions
Borns83s85 Borns81s83 Borns76s81
(1) (2) (3)
Rainfall dev. 1984-1985 from 10 year mean 1.153 1.694 .480
(.301)∗∗∗ (.347)∗∗∗ (.345)
Female -.038 .026 .075
(.081) (.095) (.089)
Age of hh head -.003 -.005 -.004
(.004) (.005) (.005)
Household head female .033 -.010 .094
(.120) (.152) (.141)
Height of HH head .002 -.005 -.005
(.004) (.008) (.007)
Head ever attend school .159 .103 .127
(.094)∗(.120) (.119)
Obs. 136 100 143
Notes: robust standard errors in brackets. * significant at 10%, ** significant at 5%, ***
significant at 1%. Dependent variable is household self-reported drought shock. Columns labeled
as in previous table
A-6

Table A.7: Famine Impacts: Illness, schooling, BMI
bmi1 bmi2 ill1 ill2 sch1 sch2
(1) (2) (3) (4) (5) (6)
Drought shock -.235 -.286 -.012 -.017 .052 .059
(.362) (.358) (.050) (.058) (.070) (.069)
Drought*Born Sep85-Sep87 -.363 -.314 .056 .036 -.133 -.151
(.592) (.593) (.087) (.087) (.122) (.123)
Drought*Born Sep83-Sep85 .132 .130 .047 .040 -.135 -.134
(.868) (.868) (.088) (.093) (.120) (.124)
Drought*Born Sep81-Sep83 -.017 .017 .192 .186 -.057 -.048
(.802) (.810) (.147) (.150) (.125) (.119)
Born Sep85-Sep87 -1.122 -1.319 -.027 -.026 .105 .117
(.499)∗∗ (.502)∗∗∗ (.048) (.052) (.071) (.063)∗
Born Sep83-Sep85 .080 -.060 -.012 -.013 .054 .061
(.455) (.446) (.051) (.056) (.073) (.069)
Born Sep81-Sep83 .877 .804 -.041 -.046 .076 .076
(.600) (.586) (.047) (.049) (.075) (.069)
Age of hh head -.007 -.019 .001 .001 -.004 -.004
(.012) (.013) (.001) (.001) (.002)∗∗ (.002)∗∗
Household head female 1.325 1.236 .007 .012 -.192 -.191
(.449)∗∗∗ (.463)∗∗∗ (.040) (.044) (.071)∗∗∗ (.074)∗∗∗
Height of HH head .049 .043 .0007 .0004 -.002 -.001
(.020)∗∗ (.017)∗∗ (.002) (.002) (.003) (.003)
Head ever attend school -.088 -.093 -.006 .011 .122 .060
(.360) (.369) (.029) (.036) (.045)∗∗∗ (.048)
Female 1.237 1.233 .018 .018 -.216 -.237
(.319)∗∗∗ (.318)∗∗∗ (.024) (.026) (.042)∗∗∗ (.044)∗∗∗
Obs. 548 548 544 463 525 525
Notes: Dependent variable columns (1) and (2) is illness (a binary variable whether child was ill
in the past 4 weeks prior to 2004 survey). Dependent variable columns (3) and (4) is schooling
(binary variable whether child attained any schooling), and (5) and (6) is body mass index
(BMI). Columns 1-4 are probit estimates, marginal effects. Marginal effects for interaction
terms are calculated as in Norton, Wang, and Ai (2004). * significant at 10%, ** significant at
5%, *** significant at 1%.
A-7

Table A.8: Robustness check: Girls and Boys separate
Boys Girls
(1) (2)
Drought shock 3.733 -1.066
(1.923)∗(2.162)
Drought*Born Sep85-Sep87 .983 -1.044
(4.604) (3.416)
Drought*Born Sep83-Sep85 -1.071 -2.281
(2.610) (3.220)
Drought*Born Sep81-Sep83 -5.443 -3.160
(3.595) (3.744)
Born Sep85-Sep87 -10.880 -5.620
(3.230)∗∗∗ (3.119)∗
Born Sep83-Sep85 -2.842 -3.789
(2.016) (2.306)
Born Sep81-Sep83 -.938 -1.133
(3.022) (3.156)
Age of hh head .042 .012
(.047) (.071)
Household head female -.338 5.342
(2.404) (2.205)∗∗
Height of HH head .127 .374
(.080) (.112)∗∗∗
head ever attend school 3.201 1.159
(1.499)∗∗ (1.625)
Obs. 323 227
Notes: robust standard errors in brackets. * significant at 10%, ** significant at 5%, ***
significant at 1%.
A-8

Table A.9: Robustness check: Is Tigray different?
(1)
Tigray dummy -160.457
(85.719)∗
Drought shock 1.533
(1.575)
Interact Tigray 5.732
(7.015)
Drought*Born Sep85-Sep87 -2.441
(2.279)
Interact Tigray 20.216
(14.328)
Drought*Born Sep83-Sep85 -2.754
( 2.359)
Interact Tigray 5.137
(7.327)
Drought*Born Sep81-Sep83 -4.718
(3.066)
Interact Tigray -1.037
(8.212)
Born Sep85-Sep87 -5.646
(1.866)∗∗∗
Interact Tigray -17.107
(14.114)
Born Sep83-Sep85 -2.044
(1.644)
Interact Tigray -.555
(5.373)
Born Sep81-Sep83 -1.216
(2.288)
Interact Tigray 3.411
( 6.456)
Age of hh head .068
( .042)
Interact Tigray -.162
(.275)
Household head female .779
(1.788)
Interact Tigray 9.153
(7.650)
Height of HH head .159
(.092)∗
Interact Tigray .951
(.550)∗
Obs. 550
Notes: robust standard errors in brackets. * significant at 10%, ** significant at 5%, ***
significant at 1%. A-9

Table A.10: Robustness check: Drought shock definition
(1) (2) (3) (4)
Drought shock 2.310 1.274
(1.199)∗(1.184)
Drought*Born Sep85-Sep87 -1.982 -6.379
(3.494) (3.133)∗∗
Drought*Born Sep83-Sep85 -4.243 -3.173 -3.938 -5.124
(2.239)∗(2.900) (1.907)∗∗ (2.524)∗∗
Drought*Born Sep81-Sep83 -6.097 -4.317 -6.664 -3.929
(1.940)∗∗∗ (2.505)∗(1.632)∗∗∗ (2.446)
Born Sep85-Sep87 -6.220 -5.145
(2.768)∗∗ (3.265)
Born Sep83-Sep85 -.660 -3.915 -1.359 -4.831
(1.796) (2.318)∗(1.504) (2.337)∗∗
Born Sep81-Sep83 .629 .959 1.217 -.494
(1.915) (1.933) (1.807) (1.945)
Age of hh head .064 -.004
(.047) (.047)
Household head female .705 .460
(1.747) (2.004)
Height of HH head .186 .134
(.097)∗(.103)
Head ever attend school 2.437 1.287
(1.245)∗(1.249)
Female -6.877 -8.463 -8.496 -8.630
(1.226)∗∗∗ (1.332)∗∗∗ (1.132)∗∗∗ (1.521)∗∗∗
Obs. 550 369 379 199
Notes: Analysis as in table 5, but drought shock defined as: household identifies 1983-85 as
worst year, rather than Sep84-Sep85 only. robust standard errors in brackets. * significant at
10%, ** significant at 5%, *** significant at 1%.
A-10

Table A.11: Food aid and drought interactions
(1)
Drought shock .638
(1.681)
Interact foodaid 2.236
(1.743)
Born Sep85-Sep87 -8.203
(2.970)∗∗∗
Interact foodaid 3.809
(3.768)
Born Sep83-Sep85 -.810
(1.531)
Interact foodaid -4.211
(2.564)
Born Sep81-Sep83 1.421
(1.604)
Interact foodaid -4.464
(5.732)
Drought*Born Sep85-Sep87 1.491
(3.316)
Interact foodaid -6.253
(4.630)
Drought*Born Sep83-Sep85 -2.770
(2.458)
Interact foodaid 1.137
(3.755)
Drought*Born Sep81-Sep83 -4.967
(2.577)∗
Interact foodaid -.991
(6.930)
Obs. 549
Notes: Analysis as in table 5, but age cohort famine variables interacted with dummy “received
any food aid in Sep84-Sep85”. robust standard errors in brackets. * significant at 10%, **
significant at 5%, *** significant at 1%.
A-11

Table A.12: Descriptive statistics: Household heads, 2004
Variable Mean Std. Dev. N
Height in centimetres 163.83 9.281 1144
Gender (female=2) 1.295 0.456 1144
Age in years 50.557 15.121 1144
Highest school grade attained 3.877 6.283 1144
Dummy: attended school up to primary 0.141 0.348 1144
Dummy: finished primary school (or above) 0.252 0.434 1144
Household Size 5.746 2.534 1144
Female children 5-15 0.921 0.989 1144
Female children under 5 0.339 0.563 1144
Male children under 5 0.32 0.567 1144
Male children 5-15 0.92 0.998 1144
Female elderly 0.167 0.405 1144
Male elderly 0.199 0.49 1144
Log total annual income 7.513 0.974 1144
Descriptive statistics for sample of household heads in 2004 used for the regressions of household
income on human capital of household heads.
Table A.13: Human capital regressions, 2004
(1) (2)
Height, centimetres .009 .010
(.003)∗∗∗ (.003)∗∗∗
Gender (female=2) -.162 -.176
(.074)∗∗ (.075)∗∗
Age, years -.002 -.006
(.009) (.009)
Square of age, years .00006 .00008
(.00009) (.00009)
Highest school grade attained .075
(.019)∗∗∗
School squared -.004
(.001)∗∗∗
Any school but not complete primary .078
(.075)
Completed primary school .171
(.059)∗∗∗
Obs. 1144 1144
Notes: *** Significant at 1%, ** Significant at 5%, * Significant at 10%. Both regressions OLS
with village fixed-effects. Dependent variable in both columns is log of household income in
2004.
A-12

7 Appendix B: Investigating possible downward
trend in child height
In some of our main results presented in table 5, coefficients on the drought shock are
also negative for the youngest cohort in the sample (even though they were not in utero
or born during the famine) and we may be concerned that drought affected families are
facing a downward trend in their children’s height, rather than a clear famine effect. For
this reason, and because children born after the famine may be endogenously chosen,
we omit the younger cohorts in columns (3) and (4) of table 5, and find that the results
do not change.
We further address the concern of whether height of those born to drought-affected
households is on the decline relative to those born to non-affected households for some
unobservable reasons in this section.25 To search for a trend, we extend the analysis to
include more (younger) children by adding cohorts beyond the youngest group used in
the main analysis. The new children are born up to September 1992 and are therefore
12-16 years old in the 2004 survey, and not fully grown.26
In figure A.1 of the main text, (reproduced below, as figure B.1), we plot the graph
of a simple fixed effects regression on cohort [year of birth] dummies interacted with
the drought shock, to incorporate younger children born in Sept 88–Sept 92, as well as
extending the control group back to children born Sept 75–Sept 76 in order to examine
that the (within household) trend in height does not appear to be downward. The
regression can also be seen as a placebo test, in which we check each year of birth
for a famine effect where we would not expect one. Equation (2) below shows the
specification:
Ht=
1992
X
t=1979
βt
f∗famineh∗Y eardummyt+
1992
X
t=1979
βt∗Y eardummyt+γh+ǫi(2)
Where ǫiis the error term, and γhare household fixed-effects. In the graph, a simple
linear trend and a quadratic trend are shown. The coefficients are somewhat “noisy”
but on this extended data set, the affected group born Sept 81–Sept 82 clearly stands
out as the worst individual year. Whilst the coefficients are in general insignificant,
that on Famine*Born Sep81–Sep82 is significant, including when gender is included as
a control. This is as expected, given the main results presented in the test. In table B.1
below, we use the cohort methodology as was used in table 5, but with added cohorts.
Again, the affected cohort is the same as in our main results, with no other significant
result.
The concern remains however, that for some reason post-drought fertility may be
systematically different for famine-affected households. If fewer drought affected house-
holds are giving birth, and these are also households with greater ability to raise taller
children (i.e. parents are taller or better schooled,) then the sample may spuriously
appear be taller due to such selection. We therefore extend the analysis of fertility
selection explored in table A.5, to double-check whether the height differences for the
latter-added cohorts are driven by fertility selection on observables. Table B.2 shows
the results. There is definitely some significant variation in fertility for the individual
25Many thanks to an anonymous referee for alerting us to this issue.
26Unfortunately, anthropometrics were not included in the latest, 2009, round of the ERHS.
B-1

years, by drought experiences, and for households of different types, however the coef-
ficients move in different directions, in a way that does not appear to be related to the
height differential as plotted above in the graph of post-famine heights. For example, in
Sept 88–Sept 89 the drought significantly reduces births, but less so for taller household
heads, older heads, female heads (and in this year, drought affected households’ children
are around 2cm taller). The converse is true in Sept 90–Sept91 (drought affected house-
holds’ children are on average 4cm taller). We therefore combine the “post–1988” years
to estimate a probit on whether a household gave birth in any of the years after 1988,
and the coefficient on the drought shock is negative, but insignificant (column 1). We
take this as evidence that the heights of children born post-1988 are not systematically
affected by fertility.
-6
-4
-2
0
2
4
6
Relative height by birth cohort (extended to 1992)
Figure B.1: Year by year height differences-reproduced
B-2

Table B.1: Cohort height regressions, extended
Height 2004, cm Coef. Std. Err. T P>t
Drought* bornSep87-Sep89 -3.59 3.01 -1.19 0.23
Drought* bornSep85-Sep87 -1.13 2.49 -0.45 0.65
Drought* bornSep83-Sep85 -0.63 2.52 -0.25 0.80
Drought*bornSep81-Sep83 -4.72 2.78 -1.70 0.09
Drought* bornSep79-Sep81 -1.69 4.07 -0.42 0.68
Drought* born Sep77-Sep79 -1.44 4.35 -0.33 0.74
Notes: Fixed effects (within household) regressions. N=1365. Cohorts used in the analysis in
table 5 of the main text are born Sep81-Sep87, with those born Sep76-Sep81 used as the control
group. We extend the control group back to those born Sep67-Sep77. Robust standard errors.
*** Significant at 1%, ** Significant at 5%, * Significant at 10%.
Table B.2: Extended fertility table
bornafter88 borns88s89 borns89s90 borns90s91 borns91s92
(1) (2) (3) (4) (5)
Drought shock -.221 -.999 -.415 .918 -.622
(.808) (.004)∗∗∗ (.677) (.213)∗∗∗ (.648)
Age of hh head -.008 -.005 -.002 -.003 -.003
(.002)∗∗∗ (.0009)∗∗∗ (.0008)∗∗ (.0008)∗∗∗ (.0008)∗∗∗
Household head female .512 -.794 -.191 1.000 .842
(1.198) (.315)∗∗ (.262) (.002)∗∗∗ (.865)
Height of HH head -.002 -.006 .001 .001 -.0006
(.004) (.002)∗∗ (.002) (.002) (.002)
Head height*female head -.004 .009 .002 -.008 -.003
(.008) (.005)∗(.004) (.004)∗(.004)
HH head any schooling .003 -.012 -.041 .034 .020
(.057) (.033) (.029) (.036) (.031)
HH head age*shock .001 .003 .0005 -.0003 .0009
(.002) (.001)∗∗∗ (.001) (.001) (.001)
HH head female*shock -.276 .996 .993 -.657 -.095
(.746) (.009)∗∗∗ (.014)∗∗∗ (.305)∗∗ (.163)
HH head height*shock .001 .007 .002 -.005 .002
(.005) (.003)∗∗ (.003) (.003)∗(.002)
HH head*female*shock .002 -.009 -.010 .013 .001
(.010) (.006)∗(.006) (.006)∗∗ (.005)
HH head school*shock .064 .045 .096 -.035 -.007
(.079) (.055) (.069) (.032) (.035)
Obs. 985 963 985 935 985
Notes: As in other tables, dates are consistent with the Ethiopian Calendar, e.g. s88s89 means
September 11, 1983 to September 10, 1984. Table shows estimation result for probit model
where dependent variable=1 if household gives birth to a child in that year. Column one
dependent variable=1 if household gives birth to a child post 1988 (i.e. in any of the individual
years). Marginal effects reported, robust standard errors, village fixed-effects included.
B-3