Examining the impact of idiosyncratic andcovariate shocks on Ethiopian households’

consumption and income sources

Catherine Porter∗

August 21, 2008

Abstract This paper investigates whether households in developing coun-tries are able to insure themselves against shocks to their income, and alsowhether such shocks cause changes in the composition of income. In par-ticular, we test whether idiosyncratic and covariate shocks have an impacton household real consumption per adult equivalent, crop income and non-crop earned income using panel data from Ethiopia. We directly includethe shocks in a model of household consumption, controlling for house-hold characteristics and seasonality. We conclude that whilst idiosyncraticshocks such as illness and crop pests do not have a significant impact onwelfare, rainfall levels have a significant and asymmetric effect. Negativerainfall shocks decrease consumption; however positive rainfall shocks ap-pear not to have a significant impact. Exploring the components of incomegenerates some new insights: whilst the agricultural shocks impact nega-tively on income from agriculture as might be expected, they appear also tostimulate non-crop earnings from self-employment and wages. This showsthat households divert their labour to more productive activities ex-postwhen a shock is likely to reduce the marginal product of labour on thehousehold farm, and thus smooth their income and consumption, despitebeing highly dependent on rainfed agriculture. In the case of a covariateshock such as poor rainfall, this smoothing mechanism may break down.

∗St Antony’s College, University of Oxford. Email: catherine.porter@sant.ox.ac.uk. I wouldlike to thank Stefan Dercon for access to and guidance with the use of the Ethiopian RuralHousehold Survey dataset as well as many invaluable comments on earlier drafts. The dataused in this paper were collected by the University of Addis Ababa, the International FoodPolicy 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 International Development Agency (SIDA) and the United States Agencyfor International Development (USAID). Thanks also to Florencia Lopez Boo, Patrick Premand,Natalie Quinn and Francis Teal for useful comments. Any errors and omissions remain my own.

1

Contents

1 Introduction 3

2 Selective review of the literature on risk and shocks 6

2.1 Permanent Income-Life Cycle Hypothesis literature . . . . . . . . 8

2.2 Tests for full insurance and complete markets . . . . . . . . . . . 10

2.3 Literature examining household responses to risk and shocks . . . 11

2.4 Risk, shocks and welfare in the Ethiopian context . . . . . . . . . 13

3 Theoretical Framework 15

4 Empirical specification and econometricstrategy 18

4.1 Empirical specification . . . . . . . . . . . . . . . . . . . . . . . . 18

4.2 Econometric strategy . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Data 25

6 Results 35

6.1 Consumption equations . . . . . . . . . . . . . . . . . . . . . . . . 35

6.2 Specification issues . . . . . . . . . . . . . . . . . . . . . . . . . . 38

6.3 Are households smoothing income? . . . . . . . . . . . . . . . . . 41

7 Conclusion 45

A Appendix One: Tables 57

2

1 Introduction

This paper documents the impact of shocks to income on household welfare and

behaviour in rural Ethiopia. We investigate the impact of some prevalent ad-

verse events on household consumption. We further explore whether households

are engaging in post-shock income smoothing in order to achieve consumption

smoothing, primarily by diverting household labour efforts towards non-crop in-

come generating activities. The majority (90%) of these households are engaged

in agricultural production, and all live in rural areas of a country with low and

erratic rainfall, causing uncertainty over their agricultural output due to weather

fluctuations, as well as sickness in the family, crop and livestock diseases, frost,

pests, and the final selling price of their crops.1

There have been three main approaches to examining the impact of shocks on con-

sumption in the literature. These are i) exploring individual strategies households

adopt in response to the shock; ii) investigating the effectiveness of consump-

tion ‘insurance’; and iii) exploring consumption smoothing over time. This paper

takes the latter route. It begins with the examination of the final consumption

outcomes for households, when all risk management and consumption smoothing

actions have been taken. We do not find conclusive evidence that idiosyncratic

shocks impact on consumption. However, rainfall, a covariate shock, does have a

significant impact on both. As an extension we provide results for sub-components

of household income, and show that the agricultural shocks do indeed impact neg-

atively on crop income. Further, they appear to stimulate non-crop income. In

fact, we argue that a prevelant household survival strategy is that of switching

labour efforts into generating non-crop income as a response to shocks, which

may not suffice when a covariate shock such as bad rainfall depresses the local

1See Dercon, Hoddinott, and Woldehanna (2005) for details on the shocks that householdsself-reported as their main worries in 2004.

3

economy.

Tests of households’ ability to smooth consumption across time look at consump-

tion outcomes in a ‘permanent income’ framework, and consider consumption

outcomes after both village level and self-insurance through other means. Most

empirical research in this literature has been based on tests of the Life-Cycle Hy-

pothesis (LCH) or Permanent Income Hypothesis (PIH), or variations thereof.

The intuition behind this model is very appealing. Households attempt to spread

the consumption of their lifetime earnings throughout their lives, using mecha-

nisms to reduce or mitigate income shocks, or using their formal and informal

savings and insurance to smooth consumption in the event of a shock hitting.

Tests of the life-cycle hypothesis have tended to concentrate on the distribution

of consumption over the longer-term (i.e. comparing different age groups and

how consumption is smoothed, for example during retirement). Deaton (1991)

provides a survey of evidence on the life-cycle hypothesis and concludes that it

tends not to hold in developing countries. Studies investigating the permanent

income hypothesis have investigated whether and how consumption is smoothed

when temporary shocks hit. Examples of this are Dercon and Krishnan (2000a)

on Ethiopia and Jalan and Ravallion (1999) on China; Skoufias and Quisumbing

(2005) synthesise findings from a number of other country studies.

We develop a dynamic model based on a version of the PIH, that specifies con-

sumption as a function of household characteristics, time and season, and idiosyn-

cratic and covariate shocks to income that may have affected the household. These

include deviations of rainfall from the long-term average, crop damage resulting

from pests or livestock trampling, and illness and death in the household. The

dependent variable chosen is the log of real consumption per adult equivalent.

Consumption in this case is a limited basket of goods, mainly food, and some

non-food essential items such as fuel and cooking expenses. This is described

4

further in the data section of the paper.

We use applied techniques for the estimation of dynamic panel data models in

order to consistently estimate the parameters of the research questions. The

empirical analysis uses fixed-effect and dynamic Generalised Method of Moments

(GMM) estimators, suitable for the data used, 1200 households in rural Ethiopia

observed five times over the ten years between 1994 and 2004.

Earlier rounds of the data have been analysed to examine the impact of shocks and

seasonality on consumption (Dercon and Krishnan, 2000b), and this paper uses a

further three rounds of the same survey (1997, 1999 and 2004, data spanning ten

years), to first ask a similar question about whether there is an impact of shocks

on consumption, incorporating dynamic panel data estimation techniques. Our

main finding is that idiosyncratic shocks to income such as illness, crop disease and

other problems, and livestock problems are quite well insured, extreme variation in

rainfall is not. We then go a step further and look at some of the coping strategies

of the household; in particular with regard to alternative income sources, and see if

there are attempts to smooth consumption through smoothing income for example

by increasing non-farm earnings (bearing in mind that for most households, farm

income represents at least three-quarters of total income in any given year).

The remainder of the paper is structured as follows: Section two outlines the theo-

retical framework of modelling consumption over time, and section three develops

the empirical specification and econometric strategy. Section four introduces the

data and provides some preliminary evidence on risk sharing, consumption and

income smoothing. Section five presents the results of the empirical analysis and

robustness tests; section six concludes and proposes some further extensions to

the work.

5

2 Selective review of the literature on risk and

shocks

The impact of adverse events, and the threat of such events, on individual and

household welfare in the developing world is a theme which has gained in both

academic and policy importance over the past ten years. For example, the World

Bank produced its flagship World Development Report (Bank, 2001) that outlined

the role of risk and therefore of policy responses, such as social protection, on wel-

fare. The bottom line is that poor people tend to face greater risks to their already

low incomes (and indeed assets) for example through ill health, bad weather and

job insecurity. Furthermore, the poor are also less equipped to deal with such

risks, having fewer assets, fewer opportunities to diversify income, limited or no

formal social insurance or social protection provisions, and limited access to in-

complete or even missing markets for credit and insurance. The extent of the

impact when a shock materialises will depend on the severity of the shock, the

frequency (or probability) of the shock, and the mechanisms available to house-

holds to mitigate the impact. If uninsured risk has a social cost, then improving

mechanisms for people to manage such risk should lead to welfare improvements.2

The development economics literature has produced a growing number of studies

on the possible impact of adverse events (known as ‘shocks’) on household welfare

(e.g. Skoufias and Quisumbing (2005), Townsend (1994)). There is also a strand

of literature examining the informal mechanisms that have developed in order to

manage risk, and other coping strategies undertaken by households. Risk manage-

ment strategies may involve ex-ante decisions (such as crop production choices) to

minimise the probability of the shock occuring. Some studies note the possibility

2For a review of the policy literature see Dercon (2002) and on health shocks see Gertler andGruber (2002). The World Bank’s Social Risk Management framework is set out in Holzmannand Jorgensen (2000).

6

that households may be undertaking counterproductive strategies (e.g. Morduch

(1995), ? and Chetty and Looney (2006) on income rather than consumption

smoothing), or insurance-type strategies put in place before the event in order to

reduce the impact of the shock (for example social insurance, precautionary saving

or income diversification), or ex-post coping behaviour such selling assets, taking

children out of school (Jacoby and Skoufias, 1997), reducing food consumption

and even more extreme forms of coping (Miguel, 2005). A related strand of liter-

ature attempts to measure the impact of shocks on different individuals (Dercon

and Krishnan, 2000a), and also to quantify vulnerability, variously defined as the

potential welfare loss through the realisation of shocks, the welfare loss due to un-

certainty, or the threat of future poverty.3 Overviews of these important debates

as a whole are provided by Fafchamps (2003) and Dercon (2004b). In this section

we concentrate on the literature that documents the impact of shocks on con-

sumption, and literature examining the risk management strategies of households

in developing countries.

This paper aims to make a contribution to the growing evidence base outlined

above on the impact of such adverse events (shocks) on consumption, and also

makes an attempt to link that to a behavioural response of households in terms

of income diversification strategies and labour supply decisions. We investigate

whether households reallocate labour towards non-farm activities when crop yields

are expected to be especially good or bad, i.e. labour switching as a coping mech-

anism rather than income diversification as an ex-ante risk mitigation strategy

(though Fafchamps (2003) notes that many ex-post strategies in fact require for-

ward planning, such as setting up contacts in order to gain wage employment in

times of need). The following discussion outlines the development of the literature

on consumption smoothing and informal insurance mechanisms, to contextualise

3On policy see Hoddinott and Quisumbing (2003); Alwang, Siegel, and Jorgensen (2001),and more theoretical discussions in Ligon and Schechter (2003) and Calvo and Dercon (2005).

7

the empirical model to be tested in the paper. We also discuss the few papers that

have linked labour supply decisions to shocks and consumption smoothing, and

finally we note related studies on the Ethiopian Rural Household Survey data,

and how the paper adds to this body of knowledge.

2.1 Permanent Income-Life Cycle Hypothesis literature

Various empirical specifications of the Permanent income hypothesis have been

proposed since Milton Friedman’s (1957) initial formulation of the theory (for a

review, see Meghir (2004)), and its insights have become widely used in models

of intertemporal choice. In its purest form, it shows that rational households

with access to perfect markets in insurance and credit will maximise the sum of

expected lifetime discounted utility, constrained only by the sum of initial assets,

and value of their future savings; their ‘permanent income’. Despite the restrictive

assumptions, the intuition that people tend to prefer smoother consumption than

income over time is appealing.

The main insight of the permanent income hypothesis is that the reaction of con-

sumption to an income shock will depend on the nature of the shock (anticipated,

or not) and the revised expectations for future income flows (permanent or transi-

tory). The predictions are that for anticipated transitory changes in income, there

should be no reaction in terms of consumption changes; for unanticipated transi-

tory shocks, the reaction should equal the annuity value of the shock (which should

be quite small, especially for younger people); for permanent shocks (i.e. which

reduce (increase) income earning capacity permanently), the reaction should be

equal to the change in current income. Friedman himself pointed out that the

presence of liquidity constraints would mean that those people without assets

would see their consumption track their income.

8

Tests of the PIH in developed countries on macroeconomic data include Hall

(1978) and also Flavin (1981), who find ‘excess sensitivity’ of consumption (of

non-durables) to current income. The consensus in the literature is overwhelm-

ingly a rejection of the strong version of the PIH. Zeldes (1989) examines whether

liquidity constraints (borrowing constraints) can explain such empirical rejections

of the Permanent Income Hypothesis, and other authors have followed in this

tradition. With respect to the developing world, Deaton (1991) provides a sur-

vey of the available evidence. For example, Paxson (1993) tested for consumption

smoothing amongst Thai rice farmers by separating the permanent and transitory

components of income, and directly looking at savings. She found that the coef-

ficient on transitory income was not statistically significantly different from one,

and also that the coefficient on permanent income was positive and statistically

significantly different from zero, contrary to the PIH – but smaller than that of

transitory income. Consequently her conclusion was that Thai rice farmers were

using savings to smooth consumption, but not to the full extent implied by the

PIH.

Non-compliance with the PIH could be due to a number of factors that have been

examined in the literature. As noted above, liquidity constraints – the absence

of opportunities for saving and borrowing – would lead to consumption tracking

income more closely than the PIH would predict. Also, if marginal utility is not

linear (see below) then there would exist a precautionary motive for saving. (Kim-

ball (1990) outlines measures of ‘prudence’ that differ from risk-aversion, which is

necessary but not sufficient to generate a precautionary motive). Browning and

Lusardi (1996) discuss the evolution of thinking about precautionary saving over

time.

9

2.2 Tests for full insurance and complete markets

An alternative methodology used widely in the literature is to examine whether

consumption smoothing is effective across space by looking at insurance mecha-

nisms at the cohort level. Cochrane (1991) and Mace (1991) are earlier examples

of testing the consumption insurance and more restrictive complete markets hy-

potheses on US data. Townsend (1994) develops this methodology to tests for

full insurance in a developing country context (India) using the village as the co-

hort. The theory is that if complete markets for managing risk exist at the village

level, then households should be able to completely smooth idiosyncratic shocks to

consumption, and only changes in aggregate income should matter (if there is no

access to credit markets or storage). The pareto efficient allocation of risk within

the village is found by maximising the sum of the utilities of all households in the

village, which each have an individual weight (that can be thought of as a house-

hold fixed effect) in the social welfare function. The budget constraint is then

constructed at the village level, and thus theory implies that the marginal utili-

ties (and therefore consumption) of the households will move together. Therefore

the model predicts that only aggregate risk is significant.

The section below discusses the various informal risk sharing arrangements that

have been documented in the literature, and one focus of the results section in the

paper is to discuss the various mechanisms of risk sharing that are involved. Some

authors have gone beyond the village level and argued that it is within sub-groups

of the village that households share risk, such as ethnic groups, credit associations

or other networks (? on network formation in the Phillipines, De Weerdt and

Dercon (2006) on Tanzania, and Rosenzweig and Stark (1989) on marriage as a

consumption smoothing network mechanism). Note also that these networks may

be formed endogenously (De Weerdt (2004), Bold (2007)).

10

Bardhan and Udry (1999) note that the full-insurance-complete-markets frame-

work and the permanent-income-consumption-smoothing framework are concep-

tually quite distinct, however in practice it may be difficult to distinguish the two

empirically. Some degree of inter-household risk sharing may be available, and

indeed some degree of consumption smoothing over time may also be available

(perhaps not in the form of credit and insurance, but for example through live-

stock saving/dissaving). In addition (as discussed below), households may also

smooth income by other methods such as diversifying their income source (in the

case of the Ethiopian households in this paper, by generating non-farm income),

receiving transfers from households outside the village or from government or non-

governmental organisations (NGOs). For this reason, we proceed with a stylised

permanent-income specification of the model, as we wish to assess the impact of

both idiosyncratic and aggregate shocks on consumption. We then discuss some

mechanisms that the households may be using to protect their consumption, and

possibly income, from such shocks.

2.3 Literature examining household responses to risk and

shocks

As noted above, the literature on developed and developing economies tends to

reject the hypothesis that households are able to protect themselves against fluctu-

ations in income, either through peer group insurance, or consumption smoothing.

However, most studies do note at least a partial ability to insure. This section

briefly discusses a number of informal mechanisms that have been documented

thus far, mainly on a case-by-case basis.

Risk and insurance market failures are typical in developing economies due to

moral hazard and adverse selection arising from asymetric information amongst

11

other things (Rothschild and Stiglitz, 1976). Besley (1995) provides a review of

informal institutions that have emerged to provide credit in the context of missing

formal markets for credit and insurance. A further literature models the types of

informal arrangement that have evolved to cope with these market failures (Coate

and Ravallion (1993), Fafchamps (1999) , Fafchamps and Lund (2003) and Udry

(1994) for example).

In the absence of credit and insurance markets, households may choose to hold

precautionary savings in certain contexts. In particular Kimball (1990) shows

that prudent households (defined by having convex marginal utility) would prefer

to hold precautionary savings. Carroll (1997) shows that households may then

engage in ‘buffer-stock’ saving with assets in order to smooth their income (this

methodology is applied by Rogg (2005) on Ethiopian data; see below). Rosen-

zweig and Wolpin (1993) find some support that households in rural India use

bullock stocks to smooth consumption, however Kazianga and Udry (2006) find

that households tend to hold on to their livestock even in times of crisis. This is

because the price of livestock may well fall at a time of crisis, rendering them less

effective as a smoothing mechanism for consumption (see also Dercon (1998)).

Some authors have noted that households’ decisions on income generation do de-

pend on risk considerations. For example Dercon (1996) shows that poorer house-

holds choose to cultivate low-risk, lower return crops in Tanzania. Hoogeveen

(2002) discusses this general issue in more detail. Dercon and Krishnan (1996)

argue that risk factors are less important than structural considerations in terms

of household assets and opportunities to enter into higher return non-farm activi-

ties. Reardon, Berdegue, Barrett, and Stamoulis (2007) have reviewed the growing

importance of non-farm wage income in the developing world, especially for house-

holds with volatile crop income, noting both ‘push’ and ‘pull’ factors as motives

for diversification. Pull factors are seen in a positive light, with households re-

12

sponding to higher returns in the non-farm sector. Diversification may take place

in the form of one or more household members taking up regular salaried work

in the non-farm sector. Push factors include seasonal smoothing, chronic insuf-

ficiency of farm returns to meet subsistence requirements and transitory income

shortages. Lanjouw (2007) also emphasises the role of the non-farm sector as a

potential safety net for households mainly engaged in farming. Fafchamps (1993)

shows that risk is a significant determinant of household labour inputs into own

farm production, and that in bad years, low marginal productivity of labour in

crop production falls so low that households reallocate labour to other activities.

Kochar (1999) examines the labour supply decisions of households in rural India;

in particular, the extent to which labor markets allow households to shift labor

from farm to off-farm employment as a response to adverse shocks. Kazianga and

Udry (2006) also link an analysis of labour supply on-farm to the consumption

smoothing model, in the context of missing markets for paid labour and we shall

outline these in more detail in the theory section as the basis for the discussion

on non-farm income response to shocks.

2.4 Risk, shocks and welfare in the Ethiopian context

Much of the literature on income fluctuations and consumption smoothing has

been generated through analysis of South Asian data.4 The Ethiopian Rural

Household Survey (see data section) due to the rich data available on risk and

shocks has begun to form a basis of a body of evidence documenting the im-

portance of risk and shocks to household welfare and economic growth over the

past few years. Dercon and Krishnan (2000b) find considerable fluctuation in the

4In particular, the ICRISAT survey of villages in rural India in the 1970s has generated aninfluentially large number of papers on seasonality and risk. See Walker and Ryan (1990) foran overview of the villages and earlier research- also the ICRISAT website shows papers thathave been produced using the data- now numbering into the 60s (www.icrisat.org, accessed 15thFebruary, 2008).

13

short-term (a period of 18 months in the first three rounds of the ERHS) that

is driven by idiosyncratic shocks, covariate shocks and seasonality. Dercon, Hod-

dinott, and Woldehanna (2005) examine the impact of shocks on consumption in

subsets of the data used in this study. Dercon (2001) and Dercon (2004a), docu-

ment the impact of shocks on economic growth during the periods 1989–1994 and

1989–1997 respectively, and again find substantial impacts.

Rogg (2005) provides an in-depth study of the asset portfolio responses of the

ERHS households to adverse shocks during the first four rounds of the survey,

and also on portfolio responses to ex-ante uncertainty. He finds that households

in more risky environments hold significantly less livestock. He also shows that

households use buffer-stock assets such as crop/food stocks and some types of

livestock to smooth consumption when income fluctuates.

Groups formed to insure against funeral costs, known locally as Iddir, are very

prevalent in the ERHS data, and these have been studied in some depth, with

a tentative conclusion being that they provide substantial insurance against fu-

neral costs, and also other shocks, though some catastrophic shocks are not fully

insured (Dercon, Bold, De Weerdt, and Pankhurst (2004) and Bold (2007), and

also Ayalew (2003)).

This paper builds on the studies already documenting the impact of shocks on

consumption, and contributes three extensions to the existing literature. First, we

examine a ten-year period with the new rounds available to gain more power in

testing the consumption smoothing hypothesis. Second, we extend the method-

ology to incorporate a more dynamic model of consumption and use new panel

data techniques to control for endogeneity of lagged values of consumption in the

estimation. Third, we incorporate an extension to analyse the linked responses to

shocks, in the form of income diversification.

14

3 Theoretical Framework

This section goes into more formal detail on the consumption smoothing model

used in the empirical analysis.

This study takes as its starting point a standard simple model of consumption

growth that has been much used in empirical investigations. It develops a version

of the permanent income hypothesis that results in a preference for consumption

smoothing and is outlined in Deaton (1992). Households generate risky income yt

in each period, and save in the form of (riskless) assets At. Assets evolve according

to the interest rate, and the proportion of income that is consumed or saved:

At+1 = (1 + r) (At) + yt − ct (1)

The lifetime budget constraint is thus constructed so that the sum of discounted

consumption is equal to initial assets plus discounted future income.

T∑0

ct(1 + r)t

= A0 +T∑0

yt(1 + r)t

(2)

Future time periods’ utility should be discounted by an appropriate rate of time

preference, δ > 0, otherwise known as the discount rate. Households are as-

sumed to maximise utility in the space of consumption over the whole life cycle

(t=1,. . . ,T ). It is standard to incorporate uncertainty into the model by substi-

tuting utility with expected utility5

U = Et

[∞∑t=τ

(1

1 + δ

)t−τ

ut(ct)

](3)

5This implies that consumers have rational expectations, and is subject to some pertinentcriticisms (e.g. Tversky and Kahnemann (1974)). We continue with the standard assumption,whilst noting these criticisms and their potential impact.

15

Maximising multi-period utility subject to this budget constraint will not nec-

essarily lead to the household smoothing consumption. If we assume additive

separability between rounds and concave instantaneous utility then the result

holds. However, this insight has generated a considerable literature debating the

possibility of finding a closed-form solution for this dynamic programme using a

specific utility function.6

We use the most common solution in empirical research: so-called Euler equations

that equalise the first-order conditions of the consumer problem, to specify the

dynamics of the relationship between consumption in one period to the next.

Hall (1978) is considered the seminal contribution to this technique, and Dercon

and Krishnan (2000b); Harrower and Hoddinott (2004) and Skoufias (2003) are

recent examples on developing country data. This avoids solving explicitly the

optimisation problem faced by the consumer, and rather focuses on the specific

first-order conditions. Of course, utility of consumption would also be affected

by tastes of the household, such as household size and composition, expanding

the utility function to ut(zt, ct), and these should be included in the empirical

specification detailed below. Using the T (number of time-periods) first order

conditions for the maximisation, we find that for each period:

λ(ct, zt) = u′t(ct) = Et

[(1 + rt+1)

(1 + δ)λ(ct+1, zt+1)

](4)

If the rate of interest is equal to the rate of time preference, and in the absence of

uncertainty, then the expected ratio of marginal utility of consumption between

any two time periods should be equalised, subject only to changes in preferences.

The underlying assumption in this model is that leisure preferences are separable

from consumption, which is quite restrictive. Heckman (1974) shows that the

6Zeldes (1989) is an exception, using numerical methods to solve the problem using theconstant relative risk aversion (CRRA) utility function.

16

estimates from the Euler equation can be inconsistent if labour supply decisions

are not taken into account.

Kazianga and Udry (2006) take this further and consider the case where leisure

and consumption jointly enter the utility function. Equation 1 showed the budget

constraint of assets evolving according to income, consumption and the interest

rate. We accordingly expand the definition of income to explicitly include the

value of farm profits, which depend on labour inputs, and ‘states of nature’ repre-

sented by annual rainfall. Equation 3 defined utility as depending on consumption,

u(ct). We can now expand this to define utility as depending on both consumption

and leisure u(c, l), and it is concave and increasing in both arguments. Following

Kazianga and Udry (2006) we also assume that ∂2u∂c∂l≤ 0 (marginal utility of con-

sumption is high when the return on labour is high and households are putting

effort into farming). The authors then go on to show that the sign of the coeffi-

cient on both consumption and farm labour demand must be the same, and must

therefore move in the same direction in response to transitory shocks.

Unfortunately, the labour supply data in the ERHS is not directly comparable

across rounds so we take an indirect route by examining the coefficients for the

transitory shocks on farm income and non-farm income respectively. We use

reduced-form specifications, and test to see if a negative (positive) shock leads

households to increase (decrease) their efforts in non-farm activities.

As a further extension, we note that current consumption may very well depend

on past consumption (especially of food) through productivity for example in

the efficiency wage model (see Dasgupta (1993)). Jalan and Ravallion (2001)

include a squared lagged term in their empirical specification in order to proxy

for possible poverty traps caused by nonlinear dynamics of recovery from a shock.

The econometric issues of whether lagged values of consumption do affect current

consumption decisions will be discussed in the econometric section below and we

17

estimate a dynamic version of the model whilst taking into account the uneven

spacing of the dataset.

4 Empirical specification and econometric

strategy

4.1 Empirical specification

Section three above outlined why households may have a preference for smooth

consumption across time. Indeed, it predicts that if households have the avail-

able mechanisms to protect their consumption from temporary fluctuations in

income, then the only reason consumption should change is if permanent income

is expected to change, or if tastes change.

Households in rural Ethiopia are subject to a number of unpredictable shocks

that may occur to individuals, whole villages, regions or countries, and that may

significantly affect income. These include weather (mainly drought/poor rainfall

in Ethiopia), crop failure for various reasons, and illness in the family. The model

outlined above would predict that those shocks that do not affect permanent

income should not affect consumption; rather households would choose to ‘smooth’

away the impact using resources available to them, such as selling assets/dissaving,

borrowing, or migrating. In contrast, shocks to permanent income (those affecting

the households’ human capital permanently; severe illness, or death, or loss of

productive assets) should have a significant impact. Seasonality is predictable,

and temporary, and therefore in theory should not affect consumption.

Using the data available, we can then examine whether consumption is indeed

smoothed by directly including the determinants of the evolution of consumption

18

over time. ‘Consumption’, the dependent variable in the empirical specification,

specifically refers to household real consumption per adult equivalent, which can

be interpreted in the results as a proxy measure of household welfare. Deaton and

Grosh (2000) discuss the merits of using consumption as a welfare measure from

household survey data. These are that it is usually measured with less error than

income, and that it is a better measure of welfare since households seek to stabilise

their consumption over time: the very assumption that this paper is attempting

to test.

This use of consumption can be criticised on the grounds of both what it includes

and excludes; the assumption is that preferences over what is included in the con-

sumption basket can be aggregated in this way7 and that these are also separable

from those over goods that are left out. It is also a very restrictive interpretation of

‘welfare’, which most development practitioners consider to be multidimensional,

in conceptual terms if not always in practice.8 However, we proceed with this

caveat in mind, and consider that ceteris paribus, higher consumption increases

welfare or utility.

Equation 5 below outlines an empirical specification of the log of consumption as

the dependent variable, regressed on vectors of variables that are likely to affect

tastes and permanent income. We overspecify the model by directly including the

shocks, and testing the restriction that the coefficients on them are equal to zero.

ln(ct) = α + β1 ln(ct−1) + β2Hit + β3ISit + β4ASit + β5Tit + β6PIit + µi + εt (5)

Where Hit represents household characteristics that may influence the ‘taste-

shifters’ zt of utility outlined above, and these are various variables on household

7Deaton also provides a full discussion on the implicit assumptions behind the aggregationof consumption baskets.

8See for example Sen (1984) for a seminal critique and numerous other publications in an-thropology and sociology.

19

size and composition. ISit is a vector of idiosyncratic shocks directly observed

in the data (such as individual perceptions of rainfall variation, crop diseases, ill-

ness in the family) and ASit are aggregate shocks, which in this case are basically

variation in rainfall across villages over time. Tit are variables related to time

and seasonality, such as indicator variables for round, for whether the survey was

taken during the post-harvest period, or the peak labour season. PIit represents

a proxy for permanent income changes, which in this case is livestock (lagged 12

months because of potential endogeneity of wealth and consumption) and shocks

to livestock, as these represent an important input into the productivity of the

household farm. We also explore land as an alternative to livestock as a mea-

sure of permanent income. There should be a positive and significant correlation

between permanent income and consumption changes from the theory.

The coefficients (vectors) of interest on the transitory shocks are β2 and β3 which

will be non-zero if the shocks impact on consumption at all. These are all shocks

which in theory should not affect permanent income (with the exception of illness,

which if serious could affect human capital in the long run.9

We also include cit−1 in the dynamic specification, as otherwise we would be

making the assumption that the coefficient of lagged consumption in the levels

equation of current consumption is equal to one, and this would lead to a number

of econometric problems.10 The econometric strategy below shows how we deal

with potential bias in OLS and fixed-effect estimators for this type of model.

9 We may test for this by incorporating past values of illness shocks in the data, as in Derconand Hoddinott (2003).

10We would essentially be assuming that there is a unit root (random walk) and the standarderrors would not be valid (see Wooldridge for further details).

20

4.2 Econometric strategy

As described below, the data used are from a panel of 1220 households, observed

five times over ten years (commonly referred to in the textbooks as ‘large N, small

T’ dataset). The availability of five rounds of data allows us to explore more

efficient panel data estimators, and dynamic estimators as outlined in Wooldridge

(2002). Developing the methodology, we specify a general model of consumption

cit:

ln(cit) = x′itβ + µ′iα + εit (6)

Where the vector xit includes the regressors outlined in equation (5) and µ′iα is an

individual effect of a constant and household-specific variables as in the previous

equation. If these are correlated with our xit then the least squares estimator

of β will be biased and inconsistent, in the usual case of omitted variable bias.

The most common ways to deal with this problem are either to estimate a first-

differenced model, or to use a fixed-effects, or within-group, estimator. If the

number of time periods in the survey is two, then both the first-difference and

fixed-effect estimators yield the same estimates. If T > 2, then the preferred

model depends on the assumptions made about the structure of the error term

(see below).

The intuition behind the fixed-effects approach is that each individual house-

hold will have a different intercept, but the relationship driving the differences

of variables from their means is constant across households. One big advantage

for empirical work is thus that the fixed-effects estimator uses the full amount of

observations, where as the differences approach loses one observation, and there-

fore efficiency (which is significant when there are only six rounds). Wooldridge

(2002) notes that if the error terms are independently and identically distributed,

then the fixed-effects estimator is more efficient. However if the errors are highly

21

autocorrelated (or a random-walk), then first differencing is more efficient. The

problem with the short panel is that it is difficult to detect autocorrelation.

An alternative to fixed-effects or first-differencing is the random-effects model,

where the omitted variables are still presumed to exist, but uncorrelated with the

included variables of the model. The fixed and random effects approaches can also

be empirically tested against each other using a Hausman specification test.11

The basic model outlined above assumes that the explanatory variables are strictly

exogenous. This also rules out any kind of feedback between past and future

values of any of the variables, and so unbiased estimates of lagged values cannot

be derived in the models, in addition the coefficients on the other explanatory

variables may be inconsistent. In sum, these are static models in the sense that

our hypothesis is that current shocks may or may not have an impact on current

consumption. There is no allowing for the influence of shocks in the past to persist

into the future. It also precludes the inclusion of the lagged value of the dependent

variable cit−1.

We wish to allow dynamic evolution of consumption over time by including past

values of consumption the empirical specification.

ln(cit) = ln(cit−1)′γ + x′itβ + z′iα + εit (7)

The econometric issue arising is that the lagged value of cit is positively corre-

lated with εit since it is correlated with the individual unobserved heterogeneity,

and the OLS estimator of the lagged dependent variable would be biased up-

wards. In addition, the fixed effects estimator will be biased downwards (known

as Nickell (1981) bias). Arellano and Bond (1991) show that there is still a cor-

11Hausman (1978). Intuitively, the random effects estimator is more efficient as it has moredegrees of freedom, but it also involves the more restrictive assumption that the omitted variableis uncorrelated with the regressors, and would be inconsistent if this assumption is violated.

22

relation between the transformed lagged dependent variable and the transformed

error, however this time the correlation is negative. This will become smaller as

T becomes large, since it is an average over T-1 periods, but as is the case in

most microeconomic panel datasets, including this one, we are only able to use

asymptotic properties of N (the cross-section) becoming large, rather than T.12

As noted above, first-differencing eliminates the individual (unobserved) effects

from the model, and also does not include all realisations of the disturbance (unlike

the fixed-effects estimator which is based on averages). The second lag of the

dependent variable, i.e. ln(cit−2) is uncorrelated with δεit under the assumption

that the errors are no more than first-order correlated (see below), and so can be

used as an instrument. Further lags can be used as instruments if the number of

time periods allows. We can include more than one lag of the dependent variable

in the right-hand side specification, if we believe that the effects persist through

more than one time period (and instrument it as for the first-period lag).

Thus the model is just identified when T = 3, and can be estimated using two-

stage least squares methods (Anderson and Hsiao, 1982). Arellano and Bond

(1991) show that whilst the Anderson-Hsiao estimator is consistent with T fixed

as N →∞ and are efficient when T = 3. However when T > 3 they show that we

can apply Generalised Method of Moments estimator with an instrument matrix

exploiting more of the lagged values. Since in this paper, T = 5, we have an

overidentified model and we can test the validity of the key assumptions using a

standard Sargan test, and also we can test for first and second-order serial corre-

lation in the first-differenced residuals. Arellano and Bond show that the presence

of first-order autocorrelation does not mean that their estimator is inconsistent,

but if the correlation was second-order this would be a problem. The A-B es-

timator in the results section is then an estimation of 7, but also including the

12This section builds on the overview paper by Bond and Windmeijer (2002).

23

lagged value of change in consumption, instrumented by further lags of consump-

tion. The A-B estimator is sometimes known therefore as the ‘difference GMM’

estimator. An augmented version of the estimator was developed by Arellano and

Bover (1995) and Blundell and Bond (1998), having noted that lagged levels may

be weak instruments for current differences, if the series are close to a random

walk, leading to finite sample bias and imprecision. They therefore combine the

differenced equations with the levels equations and use the lagged changes as in-

struments in the levels equations. The estimator is thus known as ‘system GMM’.

We present results for both cases in the results section.13

Further, the Arellano-Bond and Blundell-Bond estimators can also be used with

endogenous explanatory variables, for example see Jalan and Ravallion (1999) for

a practical application. Endogenous in this case means E[xitεis] 6= 0 for s < t,

but E[xitεis] = 0 for all s > t. The methodology is basically the same as for

the endogenous lagged value of the dependent variable. Instruments are again

lagged values of the endogenous variables. We outlined above that there may be

an endogeneity problem with including wealth (livestock) as a proxy for perma-

nent income, and therefore we include the 12-month lag as an instrument. An

alternative is to use the Arellano-Bond methodology, so in one of the specifica-

tions for the empirical analysis, household wealth as measured by livestock, will

be considered as an endogenous variable in the GMM estimation.

A final issue in the case of the ERHS dataset is the uneven spacing of the panel.

We have data from 1994, 1995, 1997, 1999 and 2004. When including the lagged

dependent variable, we are in fact including for example cit−5 in the case of round

six (2004). We therefore estimate

13Using ‘xtabond2’ (Roodman, 2006) in STATA we can also easily compute standard errorsthat are asymptotically robust to both heteroskedasticity and serial correlation, using a finite-sample correction (Windmeijer, 2005).

24

(ln(cit)− ln(cit−p))/p = ln(cit−p) + [(ln xit − lnxit−p)/p]′β + z′iα + εit (8)

in the robustness checks.14

5 Data

This section contextualises the data and discusses the characteristics of variables

of interest that will be used to develop the empirical specification: household

consumption, income from various sources, seasonality, covariate shocks (rainfall),

and idiosyncratic shocks (health, crops, rainfall).

Data are from the Ethiopian Rural Household Survey (ERHS) collected 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 Re-

search Institute (IFPRI), covering fifteen districts15 from several regions. Seven

villages were originally included in IFPRI’s survey of 1989, which were chosen

primarily because they had suffered hardships in the period 1984-89 (In partic-

ular the 1984-85 famine. For a detailed description see Webb, von Braun, and

Yohannes (1992)). In 1994, 360 of the households in six villages were retraced

and the sample frame was expanded to 1477 households. The nine additional

communities were selected to account for the diversity in the farming systems in

the country. 16 Within each village, random sampling was used. The households

14See also Dercon, Gilligan, Hoddinott, and Woldehanna (2007) for more details and anapplication.

15 These communities are called Woredas; the equivalent of a county in the UK. They arefurther divided into Peasant Associations (PAs), the equivalent of a village, and consist of up toseveral villages (e.g. the ERHS comprises 15 Woredas, and 18 PAs). The administrative systemof the PAs was created in 1974 after the revolution.

16Although representative, clearly 15 villages is not enough to make strong inference aboutEthiopia as a whole. In particular the regions of Afar and Somaliland, primarily pastoralregions (and much poorer than the rest of Ethiopia) are not included, nor are urban areas. Fora critique/alternative view, see Devereux and Sharp (2006).

25

were resurveyed again in 1994 and 1995, and subsequently in 1997 and 1999. The

sixth and latest round of the survey was completed in 2004. Table 18 (Annex)

provides the dates of surveys in each village. The attrition rate is low, less than

one per cent per annum (annualised, or 12.1% in total between 1994 and 1995).

The period which is covered by the data is one of reform and moderate growth

(around 2.5% per capita, per annum17) in Ethiopia as a whole. The protracted

civil war and dictatorship that lasted through the 1980s ended in 199118; in 1992,

the currency was devalued and in 1994, the new government agreed on a more

substantial programme of reforms and structural adjustment with the World Bank

and the IMF. In 2002, the government published its national poverty reduction

strategy, the Sustainable Development and Poverty Reduction Program (SDPRP)

and Ethiopia subsequently qualified for debt relief under HIPC. Recently, there

have been a number of problems regarding governance, in particular after the

elections of 2005, which have led to reductions (and some suspensions) of donor

assistance and ongoing conflicts in the Afar and Ogaden (Somaliland) regions that

are not covered by the ERHS.

Despite some improvements in the rural economy, the country remains prone to

drought, has very low per-capita income and food security is still a critical issue,

with a large number of people needing food aid every year. More than 80% of the

population are engaged in mainly subsistence agriculture. The 2007/8 UN Human

Development Report ranks Ethiopia 169th out of 177 on the Human Development

Index and 105th out of 108 on the Human Poverty Index. Previous work on the

survey (Dercon, 2002) has found that the consumption averages for rounds 1-3

quite similar to the national average, indicating that results of the analysis of this

survey, whilst being interpreted with caution as they are from an aging panel,

17Ethiopia (2002).18The Marxist regime that came into power when the Emperor Haile Sellasie was deposed

in part due to the 1974 famine was known as the Derg (‘Committee’), and headed by GeneralMengistu Haile Mariam, who presided amongst other things over the 1984 famine.

26

and a very small set of villages, are still able to give us some useful and relevant

policy information on the lives of rural households in Ethiopia, and in particular

the impact of risk and shocks on household consumption19.

We now discuss the main variables of interest in the empirical specification. We

begin with a discussion of consumption and income during the ten year period

under discussion, including the breakdown of income into different sources (farm

income, wage employment, self employment and transfers received). We then

examine rainfall, and various idiosyncratic shocks used in the analysis.

The dependent variable we focus on is real household monthly consumption – total

and per adult-equivalent. This is comparable with other studies of consumption

and poverty that have been conducted on the dataset. Data on monthly consump-

tion of food, purchased food and non-investment non-food items (i.e. excluding

durables, as well as health and education expenditure) from a two-week recall

period was divided by adult equivalent units based on World Health Organisa-

tion (WHO) guidelines. This was deflated by a food price index constructed from

data collected for each village at the same time as the household survey. For a

detailed discussion on the construction of the consumption indicator, see ?. Food

represents around eighty per cent of the consumption basket.

Table 1 (annex) shows the average consumption for the six rounds. Growth is

quite volatile, but after ten years of the survey, real consumption is on average

thirty per cent higher than at the beginning; growth of just under three percent

per annum over the ten years. Rounds one two and three were conducted within

18 months of each other in the 1994/5 period, and for this reason we omit round

two from our subsequent analysis. For a discussion of poverty dynamics in this

period, see Dercon and Krishnan (2000b), but it can be seen from the table that

seasonality plays a role. Many villages in the survey have only one harvest, the

19See also Collier, Dercon, and McKinnon (1997) for further detail

27

Meher, which is around December-January after the main kiremt rains, though

some villages have a second harvest. Peak labour, main harvest and survey dates

are also shown in table 18 and we control for seasonality in the empirical analysis

using a dummy variable if the survey takes place during the peak agricultural

labour season.

Examining consumption in more detail, we note that consumption growth between

1994 and 1997 is quite strong overall, as discussed in Dercon (2004a), 1997 being a

year with very good rainfall (see below). Also the timing of the survey was closer

to the main harvest than in the previous three rounds. Average consumption

dropped slightly in 1999, but recovered above its 1997 level by the final survey

in 2004. This is consistent with information on the aggregate economy from the

SDPRP, which notes negative agricultural growth in 1994 and in 1998.

Within this overall trend, there is considerable heterogeneity between villages over

time. Four villages more than doubled consumption levels over the ten years (all

from the lowest end of the distribution), whilst two villages experienced negative

growth. As an illustration of the diverse growth paths compare Adele Keke, which

saw average consumption fall by over thirty per cent in ten years, and Gara Godo,

which more than tripled – average welfare levels in those villages were quite similar

by 2004 in terms of consumption. Imdibir village on the other hand had quite

stagnant, low consumption through the whole time period under consideration.

Poverty in the sample is high, (see Dercon and Krishnan (2000b) for more detailed

discussion). Most people hold relatively little livestock (the average is around two

animals).20 In some of the analysis we are interested in investigating whether

richer or poorer households have different abilities to cope with shocks, so we split

the sample by livestock holdings, our proxy for wealth in the regressions. The

20 We use a measure of scaled livestock units for the purposes of comparing between house-holds, for example one oxen is one unit, or three calves.

28

mean consumption for poor households is 90 Birr per adult equivalent. Mean

consumption for ‘rich’ households is 115 Birr per adult equivalent.

Table 2 shows poverty rates for each year, based on consumption per adult equiv-

alent. Poverty falls overtime, commensurate with the growth in average consump-

tion, but is always higher for those households with low livestock holdings.

Income data are compiled for each household using careful recall of the various

categories of income over the past year (or the two annual harvests) for crop in-

come, and four months for non crop income. The sub-categories are gross crop

income (from both meher and belg seasons if applicable); net annual income from

self employment (including for example selling firewood, drinks, charcoal, crafts

– often this is female income); total wage income; total transfer income (official

and private). Table 3 shows consumption and income data by round. The income

data are far noisier than the consumption data, and at first glance appear incon-

sistent. The mean if divided by twelve is around half that for the consumption

data. This type of inconsistency between consumption and income is not uncom-

mon, for example it occurs in the widely used ICRISAT data (Morduch (2002),

Townsend (1995), Walker and Ryan (1990)). The recall periods are different for

consumption and income (two weeks, and four months respectively), partly due to

design as we would expect smoother consumption than income. A relatively short

four month recall period for annual income could increase the measurement error

from extrapolating to annual (e.g. since income fluctuates seasonally – also any

measurement error would thus be multiplied by four by design). Conversely, recall

error (the likelihood that the respondent of the questionnaire simply doesn’t re-

member clearly) increases as the recall period increases (Deaton, 1991).21 Deaton

(1991) also provides a useful background discussion on the merits of different recall

periods but is inconclusive on which is the better method. Glewwe (2007) notes

21Consumption data has a recall period of two weeks, but we would expect to see somesmoothing and hence less variability.

29

that the recall bias is more likely to affect the dispersion than the mean of the

distribution. We note these issues, and in particular that it may not be possible

to directly use income and consumption data in the same regressions, however we

should be safe to use either in a reduced form analysis with fixed-effects; in case

measurement error or recall error is correlated with the fixed effect. Our later

results do appear to have internal consistency. Income data are quite noisy in

general, and have not been used as extensively for analyis for this reason. We

remove extreme outliers for the estimation and this is also shown in table 3.22

We are interested in income variability, and how households are able to deal with

such variability in particular, how this affects the welfare of household members

as measured by consumption per adult equivalent. As a first step, table 4 shows

the coefficients of variation for consumption and income (both adjusted for the

size and composition of the household). The CV for income at 68.63% is signif-

icantly (at 1%) higher than that of consumption, 58.2%. This could be due to

consumption smoothing, or to increased measurement error in the income data as

discussed above. We would expect monthly income to be more variable according

to economic theory, given household preferences for smoothing- though we are not

at this point sure whether households are able to smooth across years. We also

see that the components of income – crop income and non-crop income – are more

volatile again than the aggregate. We divide the sample into ’rich’ and ’poor’ as

measured by land and livestock separately, and in both cases find that the poor

have significantly more volatile consumption (and income) than the rich – though

the magnitude of the difference is smaller when we use land as the wealth cat-

egory. We find that livestock poor households are able to smooth around eight

percent of their income volatility, compared to ten percent for livestock rich. This

appears to be reversed if we separate the sample according to land holdings.

22Using the ‘iqr’ command in STATA created by Lawrence C. Hamilton, Dept. of Sociology,Univ. of New Hampshire. An extreme outlier is defined as being more than 3 interquartileranges above(below) the 75th(25th) percentile.

30

In table 5 we provide a simple decomposition of risk into its idiosyncratic and

covariate components. We find that on average, 17% of income, and 21% of

consumption risk is covariate. Following Morduch (2002), we regress the log of

household income for each period Yit on a household fixed effect, calculating for

each village a covariate shock (round dummy). The percentage of idiosyncratic

risk is 1 − R2 of this regression as an identity. We cannot distinguish between

measurement error and idiosyncratic shocks, though in our more detailed analysis

we are able to use our direct measures of such shocks in order to gauge their

importance. In this case we can consider this an upper bound of the idiosyn-

cratic risk. In most villages the idiosyncratic income risk is higher than that of

consumption (Columns (1) and (2)), and we note simply in column (3) if this is

significantly different (at one per cent) and if so, we may tentatively conclude that

some consumption smoothing is happening. In terms of whether this is insurance

at the village level, we can also calculate the variance of consumption in the case

of perfect risk-sharing at the village level (though not between time periods), as

in Townsend (1995), if the household were to receive a constant share θ of total

village income. We can express the variance of consumption as a fraction of this

variance as shown below:

T∑t=1

(θiYt − Yi)2

T∑t=1

(Cit − Ci)2

(9)

Where Yi and Ci are the household averages of income and consumption over time

respectively, Yt is the village total income for each period, and θi is defined above.

This varies across villages but the average is 18%, in other words if villages could

achieve perfect risk-sharing, then consumption variance would be only 18 per cent

of its current level.

31

We see in table 6 the pairwise correlations between income and consumption,

and the various components of income. Not surprisingly, they are all positively

correlated (and significantly, apart from transfer income). Rainfall is discussed

below, but here we see that rainfall levels are positively correlated with crop

income and total income, but negatively correlated with consumption and non-

crop income, which we shall see later is a driver of our main results.

In table 8 we can see that the share of income from own-farm agricultural produc-

tion is very high in all rounds, between 65 and 80 per cent of the total.23 As noted

above, more than 90% of the household heads in the sample describe themselves

as farmers. Income from self employment and wage labour (non-farm earned in-

come) accounts for up to 30% of total income, though this varies across years.

Table 9 shows income shares for different quantiles of the consumption distribu-

tion. The upper quintile has the highest share of income from wages, probably

reflecting employment in the formal sector. Conversely the lowest quintile also

have a relatively high share of wage income, though in this case it is more likely

to represent daily wage labour. This pattern reflects lower size and quality of

land holdings for poorer households. At the highest end, non-agricultural income

is significant again, with more educated household members entering the formal

sector. For all quintiles, the share of transfer income rises between 1994 and 2004,

in a decade which did see a considerable increase in social protection and safety

net spending in Ethiopia.24

Turning to the main source of common ‘shocks’ to income, we examine rainfall.

Rainfall is low and erratic in Ethiopia, and the country has suffered significant

droughts in its history, most notably in 1973 and 1984, though with a number of

23 For example Canagarajah, Newman, and Bhattamishra (2001) cites agricultural incomeshare in Ghana and Uganda as being around 40 to 50 per cent.

24See for example Gilligan and Hoddinott (2007). From these data it does not appear thattransfers are progressive. The data on transfers are quite difficult to interpret, since there aremany zeros, with then some quite huge entries. Also we cannot distinguish public from privatetransfers.

32

other droughts since.25 In terms of the time series properties, Seleshi and Zanke

(2004) find that there is no trend in the annual, Kiremt and Belg rainfall totals and

rainy days over central, northern and northwestern Ethiopia in the period 1965

to 2002, however they posit that eastern and southern Ethiopia show a significant

decline since 1982.

Rainfall in the study villages also fluctuates considerably, and table 7 shows the

annual rainfall in the twelve months prior to the surveys in 1994-2004, as well as

a measure of how it deviated from a 30-year average for the rainfall stations used

in the calculations. Rainfall totals were calculated for the 12 months preceding

the survey using data for the nearest rainfall station to the village from the mete-

orological office of Ethiopia. The average over all villages is around 1000mm per

annum, similar to that which Webb, von Braun, and Yohannes (1992) cite as the

average for Ethiopia. Again, there is considerable heterogeneity between villages,

for example Haresaw and Geblen in the Tigray region (North) only average around

500mm per annum. Villages with higher average rainfall are not necessarily better

off, for example Imdibir, which was mentioned earlier as having low and stagnant

average consumption levels, has above average rainfall of 1200mm per annum.

Thus, rainfall relative to the norm for the village may be important (as farmers

grow crops that thrive in different climates with different levels of rainfall). We

include a number of variables in the empirical specification based on the long-

term average for the village. We calculate the quintiles of rainfall distribution for

each village over 30 years, and then include a dummy variable for which quintile

that rainfall fell into 12 months before the survey. This allows us to observe

non-linearities in the relationship between rainfall and consumption; for example

perhaps more rain is better, but only up to a point. Equally, the effects could

be asymmetric; people may wish to smooth consumption against bad rain but

25 2001 was a drought year in many places.

33

consume more when rain is good if they are consuming below their optimum

(e.g. if they are credit constrained). Note that in the regression specification, we

will also include controls for other round-specific effects, so that the rainfall data

does not spuriously pick up any other prevailing round-specific macroeconomic

conditions. In the ERHS, household heads are also asked about their perceptions

of the adequacy of rainfall, and this is included in the analysis of non-crop income,

when it can be considered exogenous.26

The ERHS also contains a unique amount of information on idiosyncratic shocks

facing individual households in the survey, using self-reporting of subjective per-

ceptions. In particular, the regression uses questions in the survey asking about

illness, shocks to crops, and shocks to livestock. The idiosyncratic crop shock

variable is a dummy equal to one if the head reports one of eight shocks related to

the Kiremt27 harvest. These include whether crops were affected by pests, tram-

pled by animals, destroyed by insects etc. A rainfall satisfaction index (mentioned

above) includes the average from four questions regarding the timing and quantity

of rainfall. The illness variable records the number of ill household members in

the last four weeks. We also include ill adults, ill males and ill females in some of

the specification tests.

One caveat to mention is that these are all self-reported variables. Thomas and

Frankenberg (2000) discuss the fact that self-reported illness may suffer from

self-reporting bias when perceptions and responses are correlated with wealth or

education (e.g. richer people tend to report health problems more). Also, in

the fixed-effects regressions we must consider the possibility of chronic illness,

26 For the crop income regressions, we could not prove whether satisfaction with rainfall isdriving the crop income, or whether a good harvest improves perceptions of rainfall- hence it isconsidered endogenous and omitted from the regressions

27 Many of the villages in the sample have only one main harvest, the Meher (in Amharic),which is harvested after the Kiremt (main) rains that usually occur in the summer monthsbetween May and October. In some villages there is a second, lesser harvest, called the Belgaround January-March.

34

which would not show much variation within the household over rounds (i.e. a

high average as part of the fixed-effect). In terms of our shock measurement, the

advantage of the self-reported variables are that they can provide some measure

of idiosyncratic shocks, since we know that households all experience these events

with differing severity. The advantage of the rainfall information is that it is more

clearly exogenous – however, it is of course a village level variable.

Table 10 reports the means and standard deviations of the variables included in

the main analysis and table 18 reports the dates of the survey, and the main

harvest for all of the villages.

6 Results

6.1 Consumption equations

We build up the analysis by first presenting the basic results for the static model

in table 11, with the log of real consumption per household per month as the

dependent variable. Columns (1) and (2) show OLS results, columns (3) and (4)

random effects and (5) and (6) are fixed effects specifications. We note at this point

that the random effects estimator was rejected using a Hausman test (p-value of

0.00) and therefore do not discuss the results further. All of the specifications show

that rainfall variation clearly matters to the consumption patterns of households

in the sample.28 As noted in the data section, the rainfall variable is based on

observations from the nearest rainfall station to the village in the twelve months

leading up to the survey. In column (1) for example in the OLS specification

we see that the elasticity of consumption with respect to rainfall is just under

seventeen per cent at the mean, and this impact is not reduced when we control

28 We include a full set of household size and composition controls in the specifications.

35

for household-level fixed effects (column (5)). In columns (2), (4) and (6) rainfall

is incorporated into the model as a set of four dummies for each quintile of the

30-year rainfall distribution for the village, with the median quintile omitted29.

This allows us to examine non-linearities or asymmetries in the impact of rainfall

on consumption, and also lends itself to quite an intuitive interpretation. The

results are quite consistent; whichever estimator we use, that rainfall in the upper

quintiles significantly increasing consumption (by around 12 per cent) and rainfall

in the lowest quintile reducing consumption by around six per cent. We can reject

the hypothesis that the effect is exactly symmetrical for low or high rainfall, and

cannot reject the hypothesis that the coefficients on the top two quintiles are

the same. None of the specifications shows an effect of a death in the household

between rounds, however, we can interpret the negative significant coefficient on

the sex of the household head in the fixed effects model as a shock: in most cases

the change is from male- to female-headed households through either death or

divorce.30 The crop shock and illness variables have apparently counterintuitive

effects; it appears that both have the effect of marginally increasing consumption.

Turning to the variables included as a proxy for permanent income, it can be seen

that livestock (scaled units, lagged one year31) in the first two specifications has

a positive and significant effect on consumption, suggesting the relationship holds

as in the theory, that if permanent income increases, then so does consumption.

Household size and composition variables are also included within the proxies for

permanent income. They are significant in all specifications though not reported

here. In summary, we find that the large covariate shock of rainfall fluctuation

does affect contemporaneous consumption. Idiosyncratic shocks to crops, deaths

29Various other specifications including deviation of rainfall from the median, rainfall if abovemedian, and rainfall if below median also showed significant effects, consistent with the resultsin table 11.

30We have tried to construct a ‘death of household head’ variable, but the data are patchy.31 To be clear, this is a question in the dataset asking about livestock holdings 12 months

previous to the survey date, rather than the lagged variable from the previous round.

36

and illness do not adversely affect the households average welfare.

We now consider the dynamic model, by including a lagged value of consumption

in the specification as in equation 7 presented in table 12. As noted in the econo-

metric strategy section, in this case we could expect OLS to be upward biased and

the fixed-effects (FE) model to be downward biased. It is indeed the case that

OLS is significant and positive, and the FE model is significant negative (columns

(1) and (2)) In order to correct for the bias in OLS and FE we employ the dynamic

panel data methods outlined in the econometric strategy that allow us to include

lagged values of consumption in the model and obtain consistent estimates of the

other parameters. Specifically we report the results from the Arellano-Bond (A-

B) GMM estimator and the Blundell-Bond (B-B) estimator in columns (3) and

(4) of table 12. Columns (5) and (6) show the same estimators but allowing the

livestock variable to be determined endogeneously (rather than simply using the

one-year lagged value).

Turning directly to the shocks, which are the variables of interest, all of the

dynamic estimators show that it is the negative rainfall shocks that are significant;

if rainfall is in the bottom quintile, consumption falls on average by 10-18 per

cent, a significant shock. The top quintile does not generate higher consumption

except in the Blundell-Bond estimates with exogenous (lagged) livestock (column

(2)), where it is eight per cent higher. This is a significant finding, and would

suggest that households are unable to smooth their consumption against what is

probably the most significant shock to their incomes.32 Round-specific intercepts

have been included to account for both seasonality and growth effects.33 As was

discussed earlier in the empirical specification, the fact that round dummies were

32 Various papers (See for example Dercon, Hoddinott, and Woldehanna (2005)) have providedqualitative information that drought is the shock that most people consider to be the mostsignificant threat to their livelihoods.

33See Dercon and Krishnan (2000b) for a discussion of consumption variability in the firstthree rounds of this dataset.

37

also included in the model means that the rainfall variable is not spuriously picking

up any effects due to round-specific changes in the macroeconomic environment.

The individual shocks results are slightly less consistent, with illness and crop

shocks showing the same counterintuitive sign as in the naive estimates when

we apply the Arellano-Bond difference GMM. Using the system GMM estimator

renders these results insignificant however.34

We tested whether the impact of shocks was greater for households who were

defined as poor by a livestock wealth classification (livestock holdings in round

one below the median for the village) for all the specifications, both by splitting

the sample and by using interaction terms, but found no significant differences.

6.2 Specification issues

As was discussed in the econometric strategy section, the fixed effects estimator

was preferred to ordinary least squares which suffers from omitted variable bias.

A Hausman test rejected the random effects model in all cases.

Incorporating the lagged value of consumption in the dynamic model introduces

more specification issues. As we saw on table 12 the coefficient on the lagged value

changes depending on the model used. We noted above that we would expect the

fixed-effects estimator to be downward biased and the OLS estimate to be upward

biased. The coefficients were -0.18 and 0.11 respectively. The A-B estimator

actually shows a higher coefficient on past values of consumption, allowing for the

wealth variable to be endogenously determined reduces the estimate somewhat.

The system GMM (Blundell-Bond) appears to be more directly in the middle,

and in fact shows past values of consumption to be insignificant in the model. We

34Note: we have also tested separating out adult, male and female illness and it is not signif-icant.

38

might expect this given that the lagged values we use as instruments are actually

the lagged coefficient from more than one period ago, due to the spacing of the

survey (e.g. 2004 lagged value is instrumented with 1999 consumption).

In the four GMM models the test for no autocorrelation of the first order in the

residuals is rejected, but the hypothesis for no second-order correlation cannot be

rejected. Arellano and Bond show that the presence of first-order autocorrelation

does not imply that the estimates are inconsistent, but they would be if second-

order autocorrelation was present. The Wald test that all coefficients are equal

to zero is rejected in all models. Using robust standard errors, the Sargan test is

rejected, though this could be due to heteroskedasticity. Using the two-step (en-

dogenous) model improves the chi-squared statistic, and using the system GMM

with endogenous livestock we cannot reject validity of the instruments. In the

A-B model with endogenous wealth we have a non-rejection of the null hypothesis

at five per cent, though it lies within the one-percent region (p-values are 0.028

and 0.23 respectively and shown in table 12).

The issue of uneven spacing is also a problem in the context of these estimators,

so we also estimate the average growth between periods as outlined in equation 8

and in this case use an instrumental-variables fixed-effects GMM estimator. Ta-

ble 13 shows the results which in terms of the shocks are quite consistent with

the previous results. Illness and death appear to be insignificant determinants

of consumption changes. The agricultural shock also has a counterintuitive sign.

Rainfall (measured in this case as the average growth in rainfall between peri-

ods) is significant, and indeed has quite an impact on growth; a 1% increase in

rainfall growth could improve consumption by up to thirty per cent. Note that

in columns (3) and (4) we drop round two from the analysis, as the Arellano-

Bond and Blundell-Bond estimates start at round three (1997) by necessity. We

therefore check that the difference in results were not driven by the different data

39

inputted into the model. Consequently, the impact of rainfall does drop, though

it is just insignificant (p = 0.15). We note that 1997 was quite an impressive year

in terms of harvest, consumption and rainfall, and is possibly driving this result.

In terms of the specification tests, we note that the Hansen J statistic confirms

the validity of our instruments for the lag of consumption (which are the lag of

livestock and land). Based on the tables from Stock and Yogo (2002) we see that

the Cragg-Donald Statistic shows that our IV estimates have less than five per

cent of the OLS estimates (see Murray (2006) for further discussion).

A further applied econometric issues is measurement error. Despite the careful

nature of the interview process in the surveys, which are carried out by enumera-

tors known to the community, we could still expect a degree of measurement error

in the components of the consumption basket, which are based on recall data.

Greene (2003) notes that measurement error in the dependent variable, as long

as it is not systematic or correlated with the fixed effects, will not cause any bias

to our results, and the measurement error will be absorbed into the disturbance

of the regression. Measurement error in the independent variables, or regressors,

however has more serious implications, and can cause attenuation bias (biasing

the coefficient on the variable towards zero). It is possible that there may also be

issues around reporting of age, severity and timing of shocks; and indeed some of

our shock variables are self-reported. Individuals may have different perceptions

of what makes a shock ‘severe’ and may record this differently (and it is certainly

possible that this could be correlated with the fixed effect). We note this in our

OLS and fixed effects estimators, and note that the use of IV methods (e.g. in the

dynamic models) will have reduced the attenuation bias from the measurement

error that may exist in the regressors.

In conclusion, we find that rainfall is the most significant shock affecting con-

sumption. In all of the specifications used there is a positive relationship between

40

rainfall as compared to the village average, and levels of consumption per adult

in the household. Moreover, it is extremes of rainfall that have the most signif-

icant impact: bad rainfall severely depresses consumption. Using our preferred

dynamic panel data estimates we see that idiosyncratic shocks are not significant

in determining consumption. We do not find significant differences between rich

and poor households.

6.3 Are households smoothing income?

The results above show that households are able to smooth their consumption

when idiosyncratic shocks hit, but appear not to smooth extremes of rainfall (e.g.

in the lowest quintile of the long-term village distribution). This can cause a drop

in levels of consumption per adult of around eleven per cent. In this section we

begin to explore the reasons why this may be the case by examining sources of

income, to see if households are smoothing their earned income in response to

shocks by diverting labour from crop activities to non-farm activities.

We therefore break down household earnings into various components. We would

expect rainfall and crop shocks in particular to affect crop income, despite not

having found evidence that they impact negatively on consumption. There is a

possibility that rainfall, in that it is a covariate shock, could affect the village

economy as a whole, thus depressing demand for non-crop products. On non-crop

earned income, we might see households switching all or part of their labour out

of crops and into another activity, in order to diversify, if crops are expected to

produce less than usual, and therefore smooth income. Conversely, as discussed

above, we might expect there to be a positive relationship between the marginal

product of labour and labour input; therefore if rainfall is better than expected,

the household farm might require more labour in order to reap the greater expected

41

returns (more crops to pick, more weeds to pick etc.). Including rainfall quintile

dummies would allow us to see this asymmetric effect. Our strategy is to present

reduced form estimates in the first instance, and make sure not to include any

endogenous regressors.

We first present the results for the regressions on crop income. We then look at

non-crop earned income. Income shares, by round are shown in table 8, and also

by quintile in table 9. We note from these tables that crop income dominates as

a share of the total, and particularly for poorer households.

We therefore begin by examining crop income in table 14, as it contributes the

largest share (60 to 100 per cent) of households’ total income, and most of the

households in the sample (over 90% of household heads) describe themselves as

‘farmers’. Unsurprisingly, rainfall has a significant impact on crop income (as

specified alternatively by rainfall in the 12 months preceding the survey, and by

the quintile of the long-term village rainfall distribution). Taking the quintiles

results, we see that rainfall in the lowest quintile reduces crop income by 16 per

cent when controlling for fixed effects (slightly less in the OLS specification). In

the fixed effects model (column (3)) we also see an improvement in crop income

when rainfall is at the highest end of the distribution. In column (2) we can see

that the elasticity of crop income with respect to rainfall is approximately 13.5%.

The results on individual non-rainfall crop shocks are interesting compared to the

consumption results: it appears that reported crop shocks such as pests, frosts and

trampling do significantly impact on crop income. This compares to our earlier

evidence that consumption was unaffected by such shocks. From the consumption

regressions we could not differentiate between the alternative hypotheses that this

was due to successful consumption smoothing or that the shocks did not impact

significantly on income.

42

In the OLS specifications we do also find that illness significantly reduces crop

income (by 3.5% per ill household member). However, introducing fixed effects

reduces the impact and renders it insignificant (p = 0.2). We bear in mind two

issues: one is that self-reporting bias exists (and poorer households may therefore

report less illness), but also that if certain households are more prone to illness,

or contain more household members who are chronically ill, this effect may then

be subsumed into the fixed effect.

Our conclusion is that agricultural shocks are significant determinants of crop in-

come, including rainfall variation, crop damage, and illness. One potential coping

strategy for households who do not have access to credit markets or assets is to di-

vert labour resources into other activities if crop yields are expected to be low and

reduce the marginal productivity of farm labour input. We therefore estimate a

reduced form specification for non-farm earned income. This incorporates income

from self employment, such as women’s firewood collection, dung collection and

other artisanal activities, as well as income from wage labour. Table 8 showed

that non-farm income fluctuates quite substantially, which we interpreted as a

preliminary indication that households are using this as a smoothing device. This

concurs with the result of Kochar (1999), who found that in rural India, non-farm

income responded to crop shocks (see above for discussion).

Table 15 shows the results for the variables of interest (household fixed-effects

specifications). Indeed we find the opposite effect of agricultural shocks on non

crop income compared with crop income. Rainfall is negatively correlated with

non-crop earned income. Idiosyncratic shocks to crops appear to stimulate a

significant increase in non-farm income. As outlined in the data section, we

have constructed a variable in which household heads describe their individual

feeling of satisfaction with rainfall. Whilst we omitted that from the crop income

regressions, due to the concern that it may have been endogenous (high crop

43

income may cause farmers to interpret the rainfall as ‘good’), we can include

rainfall satisfaction in this specification as there is no reason for us to suspect

that it can directly affect the returns to non-farm income. We can interpret

this variable as rainfall satisfaction, or if crop income satisfaction enters into the

rainfall satisfaction index it can only strengthen the argument that households are

actively responding to crop shocks by diverting labour supply into other activities.

Can we quantify how much households are smoothing income by diversifying?

For example the coefficient on crop shocks in the farm income regression is -0.12.

Mean annual farm income is 2087.79 Birr, which translates to a drop in income

of 250.53 Birr. The coefficient on non-farm earned income is 0.33, which at the

mean of 604.93 translates into an increase of 199.62 birr. In this way, we might

say that households are able to smooth around 80 per cent of the shock, at the

mean.

Table 15 shows fixed effects estimates for the earned income components. There

are quite a number of households who report zero non-crop income in any one

round (around 30%), therefore in the main results we have used the natural log of

non-crop income +1 to avoid creating missing values. However this means that es-

timation consistent with censored panel data will be more appropriate. ? outlined

a procedure for estimating the fixed-effects Tobit 35 model semiparimetrically (and

for a more intuitive discussion, see ?). Table 17 shows that the individual crop

shocks significantly increase income from self-employment and total non-crop in-

come, as does the rainfall perception index. The crop shock reduces wage income,

which may be if such shocks are correlated across villages, that householders are

unable to seek daily wage labour in agricultural work, which is a common source

of wage income. Illness appears to stimulate wage income (and also total non-

35Honore also designed a program ‘Pantob’in GAUSS to apply the methodology in ? whichis available to download (http://www.princeton.edu/ honore/pantob/). Many thanks to IngoOutes-Leon of Wolfson College who greatly assisted in getting the program to run successfully.

44

crop income). This is slightly puzzling, unless households are generating income

in order to increase health expenditures (not included in our consumption aggre-

gate). We conclude then that the robustness checks confirm the original result

that shocks to crops both reduce crop income and stimulate non-crop income.

7 Conclusion

This paper has investigated the impact of adverse shocks on household welfare as

measured by consumption per adult of a basic basket of food and essential other

expenses. In the descriptive statistics we found some evidence that consumption

is smoother than income. Our first results showed that rainfall shocks in the form

of extreme low or high rainfall can cause significant reductions (improvements) in

consumption. We did not find any evidence that idiosyncratic shocks adversely

impact upon consumption. We went on to examine the impact of the shocks on

earned income and its subcomponents to better understand whether the shocks

were not impacting on income, or whether households are able to smooth con-

sumption through using income smoothing as a coping strategy. Reduced form

fixed-effects regressions on crop income, the highest contributor to household total

income, showed evidence that all of the shocks we measured do have a significant

effect on crop income as theory would expect. Good rainfall improves crop in-

come, bad rainfall depresses it. Crop shocks and illness also decrease crop income.

Turning to non-crop earned income, we found that the agricultural shocks had al-

most exactly the opposite impact – and in terms of actual values at the mean,

households appear to be almost exactly substituting generation of non-crop in-

come for the lost crop income. Whilst we are unable to directly use labour supply

data, we can infer from this that households are diverting their efforts towards

relatively higher return activities in order to smooth income and consumption in

45

the face of shocks.

Some of the consumption specifications had problematic results or did not pass

robustness checks, so we also computed the estimates accounting for the uneven

spacing of the data using instrumental variables approaches, and the results were

consistent. Similarly for the non-crop income, we were faced with a censored

dependent variable, and applied non-parametric methods to generate consistent

estimates.

The findings are consistent with other studies in the literature that have shown

households’ ability to smooth their consumption against idiosyncratic shocks, but

not against covariate shocks. The same is true in Ethiopia. A key feature of

the villages under study is that most households are dependent for their main

livelihood on rainfed agriculture in a country with low and erratic rainfall. It

is unsurprising that their consumption rises and falls with the pattern of the

rain, though not an ideal situation, given that many welfare levels are fluctuating

around or below the very meagre poverty line. What is encouraging is that this

paper provides evidence that households are doing what they can with their main

resource – their labour – in order to smooth income and consumption when a

shock hits. The caveat is that this smoothing mechanism, like selling assets or

receiving intra-village transfers, could break down when the village economy is

depressed in the case of poor rainfall. Thus, there may still be a need for policy

interventions to protect households from falling below subsistence levels in the

face of covariate shocks.

46

References

Alwang, J., P. B. Siegel, and S. L. Jorgensen (2001): “Vulnerability: A

view from Different Disciplines,” Discussion Paper 0115, World Bank.

Anderson, T. W., and C. Hsiao (1982): “Formulation and estimation of

dynamic models using panel data,” Journal of Econometrics, 18(1), 47–82.

Arellano, M., and S. Bond (1991): “Some Tests of Specification for Panel

Data: Monte Carlo Evidence and an Application to Employment Equations,”

Review of Economic Studies, 58(194), 277.

Arellano, M., and O. Bover (1995): “Another look at the instrumental vari-

able estimation of error-components models,” Journal of Econometrics, 68(1),

29–51.

Ayalew, D. (2003): “Risk-sharing networks among households in rural

Ethiopia,” Discussion paper, University of Oxford.

Bank, W. (2001): World Development Report 2000/2001.

Bardhan, P. K., and C. Udry (1999): Development microeconomics. Oxford

University Press, Oxford.

Besley, T. (1995): “Nonmarket Institutions for Credit and Risk Sharing in

Low-Income Countries,” Journal of Economic Perspectives, 9(3), 115.

Blundell, R., and S. Bond (1998): “Initial conditions and moment restrictions

in dynamic panel data models,” Journal of Econometrics, 87(1), 115–143.

Bold, T. (2007): “Risk-Sharing in Insurance Groups in Rural Ethiopia,” .

Bond, S., and F. Windmeijer (2002): “Finite sample inference for GMM

estimators in linear panel data models,” Discussion Paper CWP04/02, Institute

for Fiscal Studies.

47

Browning, M., and A. Lusardi (1996): “Household saving: Micro theories

and micro facts,” Journal of Economic Literature, 34(4), 1797.

Calvo, C., and S. Dercon (2005): “Measuring Individual Vulnerability,” Dis-

cussion paper, Oxford University.

Canagarajah, S., C. Newman, and R. Bhattamishra (2001): “Non-farm

income, gender, and inequality: evidence from rural Ghana and Uganda,” Food

Policy, 26(4), 405–420.

Carroll, C. D. (1997): “Buffer-Stock Saving And The Life Cycle/Permanent

Income Hypothesis,” Quarterly Journal of Economics, 112(1), 1–55.

Chetty, R., and A. Looney (2006): “Consumption smoothing and the welfare

consequences of social insurance in developing economies,” Journal of Public

Economics, 90(12), 2351–2356.

Coate, S., and M. Ravallion (1993): “Reciprocity without commitment :

Characterization and performance of informal insurance arrangements,” Jour-

nal of Development Economics, 40(1), 1–24.

Cochrane, J. H. (1991): “A simple test of consumption insurance,” Journal of

Political Economy, 99(5), 957.

Collier, P., S. Dercon, and J. McKinnon (1997): “Social Sector Review -

Per II, Ministry of Finance.,” Discussion paper, Addis Ababa: Government of

Ethiopia.

Cowell, F. A. (1995): Measuring inequality, LSE handbooks in economics.

Prentice Hall/Harvester Wheatsheaf, London, 2nd ed edn.

Dasgupta, P. (1993): An inquiry into well-being and destitution. Clarendon

Press, Oxford.

48

De Weerdt, J. (2004): “Risk-Sharing and Endogenous Network Formation,” in

Insurance against Poverty, ed. by S. Dercon. Oxford University Press, Oxford.

De Weerdt, J., and S. Dercon (2006): “Risk-sharing networks and insurance

against illness,” Journal of Development Economics, 81(2), 337–356.

Deaton, A. (1991): Household saving in LDC’s : credit markets, insurance and

welfare. Research Program in Development Studies Woodrow Wilson School

Princeton University, Princeton, N.J.

(1992): Understanding consumption, Clarendon lectures in economics.

Clarendon Press, Oxford.

Deaton, A., and M. Grosh (2000): “Consumption,” in Designing Household

Survey Questionnaires for Developing Countries: Lessons from Fifteen Years

of LSMS Experience, ed. by M. Grosh, and P. Glewwe. Oxford University Press,

Washington, DC.

Dercon, S. (1996): “Risk, crop choice, and savings: Evidence from Tanzania*,”

Economic Development & Cultural Change, 44(3), 485.

(1998): “Wealth, risk and activity choice : cattle in western Tanzania,”

Journal of Development Economics, 55, 1–42.

(2001): “Economic reform, growth and the poor : evidence from ru-

ral Ethiopia,” Discussion paper, CSAE Publishing Department of Economics

University of Oxford, Stefan Dercon. tables ; 30 cm.

Dercon, S. (2002): “Income risk, coping strategies and safety nets,” World Bank

Research Observer, 17(2), 141–166.

Dercon, S. (2004a): “Growth and shocks: evidence from rural Ethiopia,” Jour-

nal of Development Economics, 74(2), 309–329.

49

(2004b): Insurance against poverty, Studies in development economics.

Oxford University Press, Oxford.

Dercon, S., T. Bold, J. De Weerdt, and A. Pankhurst (2004): “Group-

based Funeral Insurance in Ethiopia and Tanzania,” Discussion paper.

Dercon, S., D. O. Gilligan, J. Hoddinott, and T. Woldehanna (2007):

“The impact of roads and agricultural extension on consumption growth and

poverty in fifteen Ethiopian villages,” Discussion paper, CSAE.

Dercon, S., and J. Hoddinott (2003): “Health, shocks and poverty persis-

tence,” Discussion Paper 9291903965, United Nations University World Insti-

tute for Development Economics Research, Stefan Dercon and John Hoddinott.

30 cm. ”January 2003.”.

Dercon, S., J. Hoddinott, and T. Woldehanna (2005): “Shocks and Con-

sumption in 15 Ethiopian Villages, 19992004,” Journal of African Economies,

14(Number 4).

Dercon, S., and P. Krishnan (1996): “Income portfolios in rural Ethiopia and

Tanzania: Choices and constraints,” Journal of Development Studies, 32(6),

850.

(2000a): “In Sickness and in Health: Risk Sharing within Households in

Rural Ethiopia,” Journal of Political Economy, 108(4), 688.

(2000b): “Vulnerability, Seasonality and Poverty in Ethiopia,” Journal

of Development Studies, 36(6), 25.

Devereux, S., and K. Sharp (2006): “Trends in poverty and destitution in

Wollo, Ethiopia,” Journal of Development Studies, 42(4), 592 – 610.

50

Ethiopia, F. D. R. o. (2002): “Ethiopia: Sustainable Development and Poverty

Reduction Programme,” Discussion paper, Ministry of Finance and Economic

Development (MOFED).

Fafchamps, M. (1993): “Sequential Labor Decisions Under Uncertainty: An

Estimable Household Model of West-African Farmers,” Econometrica, 61(5),

1173–1197.

(1999): “Risk sharing and quasi-credit,” Journal of International Trade

& Economic Development, 8(3), 257.

(2003): Rural poverty, risk and development. Edward Elgar, Cheltenham.

Fafchamps, M., and S. S. Lund (2003): “Risk-sharing networks in rural Philip-

pines,” Journal of Development Economics, 71(2), 261.

Flavin, M. (1981): “Excess Sensitivity of Consumption to Current Income: Liq-

uidity Constraints or Myopia?,” Discussion paper, NBER.

Friedman, M. (1957): A theory of the consumption function. Princeton Univer-

sity Press for National Bureau of Economic Research, Princeton.

Gertler, P., and J. Gruber (2002): “Insuring Consumption Against Illness,”

American Economic Review, 92(1), 51–70.

Gilligan, D. O., and J. Hoddinott (2007): “Is There Persistence in the

Impact of Emergency Food Aid? Evidence on Consumption, Food Security,

and Assets in Rural Ethiopia,” American Journal of Agricultural Economics,

89(2), 225–242.

Glewwe, P. (2007): “Measurement Error Bias in Estimates of Income and

Income Growth among the Poor: Analytical Results and a Correction Formula,”

Economic Development and Cultural Change, 56(1), 163–189.

51

Greene, W. H. (2003): Econometric analysis. Prentice Hall International, Lon-

don, 5th edn.

Hall, R. E. (1978): “Stochastic Implications of the Life Cycle–Permanent In-

come Hypothesis: Theory and Evidence,” Journal of Political Economy, 86(6),

971.

Harrower, S., and J. Hoddinott (2004): “Consumption Smoothing and

Vulnerability in the Zone Lacustre, Mali,” Discussion paper, International Food

Policy Research Institute.

Heckman, J. (1974): “Life Cycle Consumption and Labor Supply: An Expla-

nation of the Relationship Between Income and Consumption Over the Life

Cycle,” American Economic Review, 64(1), 188–194.

Hoddinott, J., and A. R. Quisumbing (2003): “Methods for Microeconomet-

ric Risk and Vulnerability Assessments,” Discussion paper.

Holzmann, R., and S. Jorgensen (2000): “Social Risk Management: A New

Conceptual Framework for Social Protection and Beyond,” Discussion paper.

Hoogeveen, J. G. M. (2002): “Income Risk, Consumption Security and the

Poor,” Oxford Development Studies, 30(1), 105.

Jacoby, H. G., and E. Skoufias (1997): “Risk, Financial Markets, and Human

Capital in a Developing Country,” The Review of Economic Studies, 64(3), 311–

335.

Jalan, J., and M. Ravallion (1999): “Are the poor less well insured? Evi-

dence on vulnerability to income risk in rural China,” Journal of Development

Economics, 58(1), 61–81.

52

Jalan, J., and M. Ravallion (2001): “Household income dynamics in rural

China,” Discussion paper, World Bank Development Research Group Poverty

Team, Jyostna Jalan and Martin Ravallion. ill. ; 28 cm. ”November 2001.”.

Kazianga, H., and C. Udry (2006): “Consumption smoothing? Livestock,

insurance and drought in rural Burkina Faso,” Journal of Development Eco-

nomics, 79(2), 413–446.

Kimball, M. S. (1990): “Precautionary Saving in the small and in the large,”

Econometrica, 58, 53–73.

Kochar, A. (1999): “Smoothing Consumption By Smoothing Income: Hours-

Of-Work Responses To Idiosyncratic Agricultural Shocks In Rural India,” Re-

view of Economics & Statistics, 81(1), 50–61.

Lanjouw, P. (2007): “Does the rural nonfarm economy contribute to poverty

reduction?,” in Transforming the Rural Nonfarm Economy: Opportunities and

Threats in the Developing World, ed. by S. Haggblade, P. Hazell, and T. Rear-

don, pp. 55–83. Johns Hopkins Press, Baltimore.

Ligon, E., and L. Schechter (2003): “Measuring Vulnerability,” The Eco-

nomic Journal, 113(486), C95–C102.

Mace, B. J. (1991): “Full Insurance in the Presence of Aggregate Uncertainty,”

Journal of Political Economy, 99(5), 928–956.

Meghir, C. (2004): “A Retrospective on Friedman’s Theory of Permanent In-

come,” Economic Journal, 114(496), pF293–F306.

Miguel, E. (2005): “Poverty and witch killing,” Review of Economic Studies,

72(4), 1153–1172.

Morduch, J. (1995): “Income Smoothing and Consumption Smoothing,” Jour-

nal of Economic Perspectives, 9(3), 103.

53

(2002): “Consumption Smoothing Across Space,” in Insurance against

Poverty, ed. by S. Dercon, vol. 1. UN-WIDER, Helsinki.

Murray, M. P. (2006): “Avoiding Invalid Instruments and Coping with Weak

Instruments,” Journal of Economic Perspectives, 20(4), 111–132.

Nickell, S. (1981): “Biases in Dynamic Models with Fixed Effects,” Economet-

rica, 49(6), 1417–1426.

Paxson, C. H. (1993): “Consumption and income seasonality in Thailand,”

Journal of Political Economy, 101(1), 39.

Reardon, T., J. Berdegue, C. B. Barrett, and K. Stamoulis (2007):

“Household Income Diversification into Rural Nonfarm Activities,” in Trans-

forming the Rural Nonfarm Economy: Opportunities and Threats in the De-

veloping World, ed. by S. Haggblade, P. Hazell, and T. Reardon, pp. 115–141.

Johns Hopkins Press, Baltimore.

Rogg, C. (2005): “Precautionary Saving and Portfolio Management in Uncertain

Environments: Evidence from Rural Ethiopia,” Ph.D. thesis, Oxford.

Roodman, D. (2006): “How to Do xtabond2: An Introduction to ”Difference”

and ”System” GMM in Stata,” .

Rosenzweig, M. R., and O. Stark (1989): “Consumption Smoothing, Mi-

gration, and Marriage: Evidence from Rural India,” The Journal of Political

Economy, 97(4), 905–926.

Rosenzweig, M. R., and K. I. Wolpin (1993): “Credit market constraints,

consumption smoothing, and the accumulation of durable production,” Journal

of Political Economy, 101(2), 223.

Rothschild, M., and J. Stiglitz (1976): “Equilibrium in Competitive Insur-

ance Markets,” The Quarterly Journal of Economics, 90(4), 629.

54

Seleshi, Y., and U. Zanke (2004): “Recent changes in rainfall and rainy days

in Ethiopia,” International Journal of Climatology, 24(8), 973–983.

Sen, A. K. (1984): Resources, values and development. Basil Blackwell, Oxford.

Skoufias, E. (2003): “Consumption smoothing in Russia: Evidence from the

RLMS,” Economics of Transition, 11(1), 67.

Skoufias, E., and A. R. Quisumbing (2005): “Consumption Insurance

and Vulnerability to Poverty: A Synthesis of the Evidence from Bangladesh,

Ethiopia, Mali, Mexico and Russia,” European Journal of Development Re-

search, 17(1), 24–58.

Stock, J. H., and M. Yogo (2002): “Testing for Weak Instruments in Linear

IV Regression,” Discussion paper, National Bureau of Economic Research, Inc.

Thomas, D., and E. Frankenberg (2000): “The Measurement and Interpre-

tation of Health in Social Surveys,” Discussion paper, RAND.

Townsend, R. M. (1994): “Risk and Insurance in Village India,” Econometrica:

Journal of the Econometric Society, 62(3), 539–591.

(1995): “Consumption Insurance: An Evaluation of Risk-Bearing Sys-

tems in Low-Income Economies,” Journal of Economic Perspectives, 9(3), 83.

Udry, C. (1994): “Risk and insurance in a rural credit market: An empirical

investigation in Northern Nigeria,” Review of Economic Studies, 61(208), 495.

Walker, T. S., and Ryan (1990): Village and household economies in India’s

semi-arid tropics. Johns Hopkins University Press, Baltimore ; London.

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).

55

Windmeijer, F. (2005): “A finite sample correction for the variance of linear

efficient two-step GMM estimators,” Journal of Econometrics, 126(1), 25–51.

Wooldridge, J. M. (2002): Econometric analysis of cross section and panel

data. MIT Press, Cambridge, Mass. ; London.

Zeldes, S. P. (1989): “Optimal Consumption with Stochastic Income: De-

viations from Certainty Equivalence,” The Quarterly Journal of Economics,

104(2), 275–298.

56

A Appendix One: Tables

Table 1: Consumption per adult 1994-2004, all villages

Peasant Association 1994 1995 1997 1999 2004 N

Haresaw 80.01 75.94 116.35 99.57 84.29 74Geblen 38.36 41.76 104.23 77.97 105.72 58Dinki 64.98 52.78 65.61 80.92 89.64 76Yetemen 134.58 74.53 113.85 86.52 126.92 53Shumsheha 115.68 120.63 117.69 139.49 136.94 109Sirbana Godeti 111.60 109.30 106.33 174.22 175.62 72Adele Keke 112.35 122.43 150.35 90.96 85.61 86Korodegaga 40.89 52.30 72.44 94.64 76.59 88Trurufe Kechema 98.47 73.28 85.95 132.28 94.96 88Imdibir 49.84 37.28 70.22 57.52 55.70 62Aze Deboa 99.06 54.65 90.09 37.45 106.06 56Adado 83.48 60.04 102.26 72.66 56.51 82Gara Godo 33.27 31.73 57.48 56.02 94.56 91Domaa 49.37 111.45 69.67 93.22 107.51 56DB-Milki 115.02 95.84 164.48 149.88 149.95 57DB-Korma 89.76 91.48 160.95 152.87 155.91 53DB-Karafino 108.73 70.05 134.10 139.32 131.17 33DB-Faji Bokafia 123.20 91.66 158.10 148.25 197.09 24

Notes: Data are in Ethiopian 1994 Birr per adult equivalent, per household

57

Table 2: Standard Poverty Measures: By Round

Year Headcount Poverty Gap Sq-Poverty Gap Mean consumptionF-G-T (0) F-G-T (1) F-G-T (2) amongst poor (Birr)

1994 0.39 0.16 0.09 26.841995 0.44 0.19 0.1 25.721997 0.24 0.08 0.04 30.311999 0.28 0.09 0.04 31.132004 0.22 0.07 0.03 30.29

Notes: Source is ERHS data, own calculations. Poverty line is 44.3 Birr per adult (1994 realvalue), per month on average (though is village specific). Measures are weighted by householdsize.

Table 3: Consumption and Income aggregates

All observations Without OutliersYear Consumption Income Consumption Income

1994 86.13 543.92 83.94 505.231995 78.58 524.31 76.37 504.391997 108.14 1010.95 102.66 673.811999 107.57 510.12 101.41 513.502004 116.71 636.61 105.44 586.32

Notes: Source is ERHS data, own calculations.Consumption and Income are both measured inreal 1994 Birr, and are per adult equivalent, per household. Consumption data are monthly,based on two-week recall period, and Income data are annual, based on four-month recallperiod. The first two columns show all observations (n=5989) and the second two columnsremove outliers (n=5736).

Table 4: Coefficients of Variation: Consumption and Income

Consumption Income Crop income Noncrop income

Livestock Poor 62.178 70.466 83.408 130.574Livestock Rich 56.765 66.897 73.883 137.817

Land Poor 60.501 70.547 79.751 133.860Land Rich 58.183 66.397 77.201 134.613

Average 59.454 68.673 78.599 134.201Standard Error 0.102 0.136 0.178 0.520

Notes: Standard Errors calculated following Cowell (1995) as√

1+2(CV )2

2n

58

Table 5: Idiosyncratic versus aggregate shocks: Consumption and Income

Idiosyncratic Idiosyncratic Complete riskConsumption Income Smoothing? sharing vs autarky

Haresaw 0.91 0.84 no 0.06Geblen 0.64 0.87 yes 0.02Dinki 0.89 0.68 no 0.18Yetemen 0.81 0.85 yes 0.13Shumsheha 0.97 0.96 no 0.22Sirbana Godeti 0.78 0.97 yes 0.10Adele Keke 0.84 0.85 yes 0.28Korodegaga 0.71 0.65 no 0.11Trurufe Ketchema 0.9 0.89 no 0.42Imdibir 0.77 0.78 yes 0.26Aze Deboa 0.7 0.9 yes 0.22Adado 0.81 0.82 yes 0.07Gara Godo 0.69 0.72 yes 0.26Domaa 0.8 0.81 yes 0.30Debre Birhan villages 0.8 0.88 yes 0.07

Notes: Column one shows 1 − R2 of the household fixed effects regression on log consumptionper adult equivalent with round fixed effects also included, specified separately for each village.All data are real 1994 Ethiopian Birr. The second column shows the same regression for incomeper adult equivalent. They show the amount of consumption and income variance that is dueto idiosyncratic shocks and measurement error combined. If income shows more idiosyncraticvariance than consumption, then column three assumes smoothing. Those in bold are caseswhere consumption appears more variable than income. Those in italics show no significantdifference between consumption and income variance. The final column shows the ratio of thevariance of consumption under complete risk sharing (a fixed proportion of the village totalincome), to the variance under autarky (actual consumption variance).

Table 6: Simple correlations: Consumption, Income, Rainfall

Consumption Rainfall Tot Inc Crop Inc Noncrop Inc Transfers

Consumption 1Rainfall -0.1068* 1Total Income 0.2805* 0.0198* 1Crop Income 0.2266* 0.0502* 0.8012* 1Noncrop Income 0.1377* -0.0189* 0.5606* 0.0202* 1Transfers 0.0220* -0.0393* 0.1806* -0.0423 -0.0043 1

Notes: Table shows pairwise correlations between variables, asterisk indicatessignificant at 5 per cent. All variables are deflated by the price index and expressedas ‘per adult equivalent’ units.

59

Table 7: Rainfall mean and quintile, all villagesPeasant Association 1994 1995 1997 1999 2004 30 year

mean

Haresaw 393.3 670.5 702.3 519.9 442.3 429.1(3) (5) (5) (4) (3) (49.3)

Geblen 392.3 656.5 722.3 370 320.42 712.6(1) (3) (4) (1) (1) (77.3)

Dinki 1494.8 1653.8 1788.7 1482.9 1520.8 1371.0(4) (4) (5) (4) (4) (32.3)

Yetemen 1540.9 1369.2 1020.2 949 1328.5 1156.7(5) (5) (2) (2) (5) (32.8)

Shumsheha 791 1060 932.75 769.67 794.5 720.0(4) (5) (5) (4) (4) (28.1)

Sirbana godeti 785.5 659.7 756.5 852.7 1033.8 876.8(2) (1) (2) (3) (5) (17.9)

Adele Keke 523.02 971.72 981.5 710.21 880 783.1(1) (5) (5) (2) (4) (35.9)

Korodegaga 936.9 878.4 999.6 1010.28 902.9 874.0(3) (3) (4) (4) (3) (20.7)

Trurufe Kechema 921.8 925.7 1001.9 1002.52 687.7 966.9(3) (3) (3) (4) (1) (14.5)

Imdibir 1606.9 1264.79 1239.33 1440.7 630 1184.5(5) (4) (4) (5) (1) (25.3)

Aze Deboa 852.6 280.6 1024.8 1220.1 981.7 1095.9(1) (1) (2) (4) (2) (28.6)

Adado 1108.7 1553.8 1485 1192 1121.2 1577.3(1) (3) (2) (1) (1) (20.3)

Gara Godo 1118.9 1582.3 1272.5 1332.2 1537.2 1438.3(1) (4) (2) (2) (4) (19.8)

Domaa 492.17 930.69 1434.6 690.7 840.79 834.6(1) (4) (5) (2) (3) (29.0)

Debre Birhan Villages 891.1 806.3 830 817.2 978.8 844.0(3) (2) (3) (2) (4) (22.2)

Notes: Table shows rainfall in mm, in the twelve months preceding the survey, in the nearestrainfall station to the village. If that station had a missing value, we interpolated using thenext nearest rainfall station (with relative movement). Far right hand column shows villagemean over 30 years 1974-2004. The figure below it in brackets is the coefficient of variation ofrainfall over the same period.

60

Tab

le8:

Inco

me

Div

ersi

fica

tion

-so

urc

esof

inco

me,

1994

-20

04

Cro

pin

com

eT

ransf

erin

com

eSel

f-em

plo

ym

ent

Wag

ela

bou

rIn

com

egr

owth

onpre

vio

us

per

iod

1994

0.66

0.05

0.24

0.05

1995

0.72

0.11

0.13

0.04

-10.

9519

970.

730.

030.

200.

0444

.28

1999

0.79

0.07

0.07

0.08

-32.

7420

040.

660.

080.

200.

0513

.41

61

Tab

le9:

Inco

me

sourc

eby

quan

tile

,19

94an

d20

04

Inco

me

sourc

eQ

uin

tile

One

Quin

tile

Tw

oQ

uin

tile

Thre

eQ

uin

tile

Fou

rQ

uin

tile

Fiv

e19

9420

0419

9420

0419

9420

0419

9420

0419

9420

04C

rops

0.64

0.69

0.72

0.67

0.71

0.75

0.63

0.69

0.59

0.57

Sel

f-em

plo

ym

ent

0.03

0.14

0.04

0.16

0.03

0.14

0.09

0.17

0.03

0.32

Wag

es0.

260.

090.

190.

060.

220.

050.

230.

060.

320.

03T

ransf

ers

0.07

0.09

0.05

0.09

0.03

0.06

0.05

0.09

0.05

0.09

62

Table 10: Summary statistics

Variable Mean Std. Dev. NHousehold Size 5.992 2.782 5993Nr female adults 1.57 1.036 5993Nr female 5-15 0.924 1.006 5993Nr female under 5 0.411 0.645 5993Nr female elderly 0.178 0.423 5992Nr male adults 1.499 1.117 5993Nr male 5-15 0.915 0.998 5993Nr male under 5 0.412 0.632 5993Livestock units (last year) 2.983 3.407 5972Land 1.537 1.468 5423Sex of hh head 0.762 0.426 5940Dummy: Peak labour season 0.305 0.46 5993Nr ill hh members last 4 wks 0.541 0.888 5964Dummy: crop shock 0.724 0.447 5993Dummy: any hh member died 0.121 0.326 5993Log annual rainfall 6.827 0.364 5993Village quintile rainfall dummy 3.107 1.329 5993Log real hh consumption 5.737 0.804 5954Log real hh consumption per adult 4.269 0.815 5954Log real annual income 7.325 1.14 5915Log real annual crop income 6.903 1.309 5542Log real annual noncrop income 3.809 3.074 5989Log real annual transfers received 1.594 2.549 5989Round 2.969 1.403 5993

63

Tab

le11

:B

asel

ine

Con

sum

pti

onR

esult

s:Sta

tic

Model

OL

S1

OL

S2

RE

1R

E2

FE

1F

E2

(1)

(2)

(3)

(4)

(5)

(6)

Lag

ged

live

stock

unit

s.0

36.0

36.0

49.0

50.0

23.0

20(.

006)∗∗∗

(.006)∗∗∗

(.006)∗∗∗

(.006)∗∗∗

(.007)∗∗∗

(.007)∗∗∗

Sex

ofH

Hhea

d(m

al=

1,fe

m=

0).1

35.1

35.1

23.1

16.1

36.1

25(.

023)∗∗∗

(.023)∗∗∗

(.028)∗∗∗

(.027)∗∗∗

(.049)∗∗∗

(.047)∗∗∗

Dum

my:

pea

kag

.la

bou

rse

ason

.009

.018

.112

.124

.002

.014

(.022)

(.022)

(.020)∗∗∗

(.020)∗∗∗

(.022)

(.021)

Nr.

Ill

hh

mem

ber

s.0

34.0

36.0

20.0

18.0

30.0

32(.

011)∗∗∗

(.011)∗∗∗

(.011)∗

(.011)

(.012)∗∗

(.012)∗∗∗

Indiv

idual

crop

shock

.044

.043

.042

.030

.049

.045

(.021)∗∗

(.021)∗∗

(.020)∗∗

(.021)

(.022)∗∗

(.022)∗∗

Log

Annual

Rai

nfa

llpas

t12

mth

s.1

79-.

082

.168

(.045)∗∗∗

(.030)∗∗∗

(.044)∗∗∗

Villa

gera

infa

llin

bot

tom

quin

tile

-.04

2-.

071

-.03

8(.

034)

(.033)∗∗

(.032)

Villa

gera

infa

llin

seco

nd

quin

tile

-.00

7.0

43-.

006

(.033)

(.030)

(.032)

Villa

gera

infa

llin

fourt

hquin

tile

.130

.037

.130

(.030)∗∗∗

(.027)

(.029)∗∗∗

Villa

gera

infa

llin

top

quin

tile

.127

.115

.133

(.035)∗∗∗

(.032)∗∗∗

(.034)∗∗∗

Obs.

5186

5186

5186

5186

5361

5186

Not

es:

OL

Sis

the

Ord

inar

yle

ast

squa

res

spec

ifica

tion

,w

ith

villa

gean

dye

arfix

edeff

ects

incl

uded

.R

Ein

clud

esho

useh

old

rand

omeff

ects

.F

Ein

clud

esho

useh

old

fixed

effec

ts.

The

depe

nden

tva

riab

leis

the

log

ofre

alto

tal

hous

ehol

dco

nsum

ptio

npe

rad

ult

equi

vale

nt.

Rai

nfal

lqu

inti

lere

fers

toth

e30

year

rain

fall

dist

ribu

tion

ofth

evi

llage

.A

llre

gres

sion

sin

clud

eho

useh

old

size

and

com

posi

tion

vari

able

s.A

Hau

sman

test

reje

cted

the

rand

omeff

ects

mod

el(p

=0.

000)

.

64

Table 12: GMM Consumption resultsOLS FE DIFF1 SYS1 DIFF2 SYS2(1) (2) (3) (4) (5) (6)

L.lrconsae .111 -.184 .175 -.035 .132 -.012(.016)∗∗∗ (.017)∗∗∗ (.038)∗∗∗ (.031) (.039)∗∗∗ (.033)

Lagged livestock units .032 .021 .011 .041(.006)∗∗∗ (.010)∗∗ (.008) (.008)∗∗∗

Sex of HH head (mal=1, fem=0) .091 .085 .082 .096 .086 .107(.026)∗∗∗ (.053) (.077) (.038)∗∗ (.083) (.041)∗∗∗

Dummy: peak ag. labour season -.039 -.041 -.147 -.012 -.155 .011(.025) (.024)∗ (.033)∗∗∗ (.025) (.036)∗∗∗ (.026)

Nr. Ill hh members .032 .033 .031 .014 .036 .013(.014)∗∗ (.016)∗∗ (.020) (.015) (.020)∗ (.016)

Individual crop shock .051 .029 .071 .042 .070 .037(.023)∗∗ (.024) (.032)∗∗ (.025)∗ (.032)∗∗ (.026)

Any hh member died between rounds .038 .039 .037 .017 .053 .008(.030) (.033) (.049) (.036) (.050) (.037)

Village rainfall in bottom quintile -.074 -.015 -.194 -.076 -.178 -.121(.042)∗ (.041) (.060)∗∗∗ (.044)∗ (.060)∗∗∗ (.047)∗∗

Village rainfall in second quintile .003 -.008 -.032 .013 -.062 -.004(.038) (.033) (.042) (.036) (.044) (.037)

Village rainfall in fourth quintile .104 .090 .031 .016 .030 -.012(.033)∗∗∗ (.032)∗∗∗ (.042) (.034) (.044) (.036)

Village rainfall in top quintile .106 .161 -.075 .072 -.071 .022(.039)∗∗∗ (.039)∗∗∗ (.054) (.041)∗ (.063) (.044)

AR(1) 0.000 0.000 0.000 0.000AR(2) 0.300 0.967 0.188 0.710Sargan 0.001 0.001 0.007 0.001Hansen 0.005 0.043 0.002 0.236Obs. 3860 3860 2712 3827 2752 3827Notes: DIFF is the Arellano Bond (1991) estimator, often referred to as the ‘difference estima-tor’. SYS refers to the Blundell-Bond (1998) estimator, also known as the ‘system’ estimator.In columns (5) and (6) we include livestock as an endogenous variable. The dependent variableis the log of real total household consumption per adult equivalent. Rainfall quintile refers tothe 30 year rainfall distribution of the village. All regressions include household size and compo-sition variables. AR(1) and AR(2) is the Arellano-Bond test for first-order autocorrelation andsecond-order respectively (Null is that such autocorrelation in the errors exists). Hansen andSargen tests are the overidentification restrictions on the instruments. A rejection casts doubton the validity of the instruments. For all of the specification tests, the p-value is reported inthe appropriate column.

65

Table 13: Consumption growth and shocks: GMM IV Fixed Effects estimatesGMMIV1 LIML1 GMMIV2 LIML2

(1) (2) (3) (4)L.lrconsae -.174 -.293 -.175 -.335

(.069)∗∗ (.056)∗∗∗ (.055)∗∗∗ (.049)∗∗∗

Nr ill hh members .009 .010 .009 .009(.009) (.008) (.007) (.005)∗

Agricultural shock .018 .013 .021 .017(.013) (.011) (.010)∗∗ (.008)∗∗

Hh member died .010 .011 .021 .021(.016) (.013) (.015) (.011)∗

Ln rainfall .351 .323 .141 .169(.056)∗∗∗ (.051)∗∗∗ (.079)∗ (.063)∗∗∗

Cragg-Donald F-stat 23.36*** 30.28*** 14.57*** 21.84***Hansen J statistic 0.821 1.08 3.34 1.57Obs. 3671 3762 2710 2822Notes: The dependent variable is the log of real consumption per adult equivalent. GMM-IVis the Generalised Method of Moments estimator. LIML is the Limited-Information MaximumLikelihood Estimator. All estimates include household fixed effects. The second two columnsinclude from round three only, as a comparison with the Arellano- Bond and Blundell-Bondtable previous to this one.

66

Table 14: Crop income regressionsOLS OLS2 FE1 FE2(1) (2) (3) (4)

Land size (hectares) .110 .110 .057 .056(.013)∗∗∗ (.013)∗∗∗ (.013)∗∗∗ (.013)∗∗∗

Lagged livestock units .064 .063 .040 .039(.010)∗∗∗ (.010)∗∗∗ (.011)∗∗∗ (.011)∗∗∗

Sex of HH head (mal=1, fem=0) .341 .339 .275 .276(.043)∗∗∗ (.043)∗∗∗ (.077)∗∗∗ (.078)∗∗∗

Dummy: peak ag. labour season .059 .062 .056 .060(.030)∗∗ (.030)∗∗ (.031)∗ (.031)∗

Nr. Ill hh members -.032 -.033 -.020 -.022(.016)∗ (.016)∗∗ (.017) (.017)

Individual crop shock -.078 -.078 -.123 -.123(.033)∗∗ (.033)∗∗ (.033)∗∗∗ (.033)∗∗∗

Any hh member died between rounds -.054 -.001 .001(.042) (.047) (.047)

Log Annual Rainfall past 12mths .099 .132(.075) (.068)∗

Village rainfall in bottom quintile -.169 -.166(.054)∗∗∗ (.051)∗∗∗

Village rainfall in second quintile -.043 -.040(.051) (.045)

Village rainfall in fourth quintile -.00005 .013(.049) (.045)

Village rainfall in top quintile .066 .099(.064) (.055)∗

Obs. 4679 4679 4679 4679Notes: The dependent variable is the log of total real annual income from cultivating own crops.

67

Table 15: Non-Crop Earned Income regressionsOLS OLS2 FE(1) (2) (3)

Sex of HH head (mal=1, fem=0) -.110 -.085 .002(.111) (.108) (.237)

Dummy: peak ag. labour season -.701 -.717 -.647(.106)∗∗∗ (.105)∗∗∗ (.101)∗∗∗

Nr. Ill hh members .128 .099 .143(.046)∗∗∗ (.045)∗∗ (.049)∗∗∗

Individual crop shock .134 .255 .197(.099) (.097)∗∗∗ (.102)∗

Any hh member died between rounds -.228 -.243 -.016(.145) (.139)∗ (.155)

Individual rain perception index -.389 -.339(.123)∗∗∗ (.127)∗∗∗

Log Annual Rainfall past 12mths -.320 -.397(.187)∗ (.183)∗∗

Village rainfall in bottom quintile -.742(.156)∗∗∗

Village rainfall in second quintile -1.168(.141)∗∗∗

Village rainfall in fourth quintile -1.017(.136)∗∗∗

Village rainfall in top quintile -.858(.165)∗∗∗

Obs. 4659 4922 4747

Notes: The dependent variable is the log (plus 1) of total real annual non-crop earned income.This includes wage earnings, plus income from self-employment (e.g. women’s firewood collectionetc). Robustness checks included running a tobit on the non-logged data, and a Heckman cor-rection model of selection into non-crop income activities with land as the selection instrument.The Heckman model does not adjust for fixed-effects, and the tobit model does not incorporaterobust standard errors. However, the results were consistent with the results presented above.

68

Table 16: Impact of shocks: Earned income componentsCrop Noncrop Self Wage(1) (2) (3) (4)

Nr. Ill hh members -.020 .143 .045 .134(.017) (.047)∗∗∗ (.049) (.043)∗∗∗

Individual crop shock -.123 .202 .335 -.168(.033)∗∗∗ (.099)∗∗ (.092)∗∗∗ (.078)∗∗

Any hh member died between rounds -.001 -.003 -.295 .337(.047) (.152) (.137)∗∗ (.129)∗∗∗

Log Annual Rainfall past 12mths .132 -.419 -.347 -.273(.068)∗ (.174)∗∗ (.167)∗∗ (.145)∗

Individual rain perception index -.271 -.416 .134(.123)∗∗ (.117)∗∗∗ (.106)

Obs. 4679 5175 5175 5175Notes: The table shows the coefficients and standard errors for the reduced form equations forcrop income, non crop income , wage earnings, and income from self-employment. All modelsincorporate fixed-effects and robust standard errors. All dependent variables are expressed asthe log of the real annual income in the category, in 1994 Birr. For non-crop and transfer incomewe add one to the variable before taking the natural logarithm.

Table 17: Impact of shocks: Noncrop earned income componentsNoncrop Self Wage

(1) (2) (3)Nr. Ill hh members 0.169 0.062 0.336

(0.057)*** (0.063) (0.122)***

Individual crop shock 0.276 0.551 -0.748(0.130)** (0.131)*** (0.302)***

Any hh member died between rounds -0.031 -0.420 1.070(0.199) (0.199)** (0.435)***

Ln annual rainfall past 12mths -0.423 -0.268 -0.624(0.219)* (0.225) (0.473)

Individual rain perception index -0.500 -0.777 0.291(0.156)*** (0.160)*** (0.352)

Notes: The table shows the coefficients and standard errors for the for, total non crop income, and income from self-employment and wage earnings. Models calculated using methodologyin ? non-linear semi-parametric fixed effects Tobit. Loss function is quadratic. All dependentvariables are expressed as the log of the real annual income in the category, in 1994 Birr, andwe add one to the variable before taking the natural logarithm.

69

Tab

le18

:D

ates

ofsu

rvey

sin

six

ER

HS

rounds,

and

mai

nhar

vest

Surv

ey

site

Loca

tion

Main

harv

est

Round1

1994

Round3

1995

Round4

1997

Round5

1999

Round

62004

Har

esaw

Tig

ray

Oct

-Nov

Jun-j

ul

Mar

June

Aug

Apr

Geb

len

Tig

ray

Oct

-Nov

Jun-j

ul

Mar

June

Sep

May

Din

ki

N.

Shoa

Dec

Mar

-Apr

Jan

Oct

Aug

May

Deb

reB

erhan

N.

Shoa

Nov

-Dec

Mar

-Apr

Mar

Jul

Jul

Apr

Yet

men

Gojj

amN

ov-D

ecM

ar-A

pr

Mar

Sep

Jun

May

Shum

sha

S.W

ollo

Oct

-Dec

Jun-J

ul

May

Oct

Sep

Apr

Sir

ban

aG

oded

tiShoa

Nov

-Dec

Mar

-Apr

Mar

Jun

Jun

Apr

Adel

eK

eke

Har

argh

eN

ov-D

ecM

ay-J

un

Apr

Oct

Jun

Apr

Kor

odeg

aga

Ars

siO

ct-N

ovM

ay-J

un

May

-Jun

Jun

Aug

May

Turf

eK

echem

ane

S.S

hoa

Dec

Mar

-Apr

Mar

-Apr

Sep

Jun

May

Imdib

irShoa

(Gura

ge)

Oct

-Dec

Mar

-Apr

Mar

Jun

Jun

Apr

Aze

Deb

oaShoa

(Kem

bat

a)O

ct-N

ovM

ar-A

pr

Mar

Oct

Jun

Jul

Addad

oSid

amo

(Dilla

)D

ec-J

anM

ar-A

pr

Mar

Jun

Sep

Apr

Gar

aG

odo

Sid

amo

(Wol

ayta

)A

ug-

Dec

Mar

-May

Mar

Jun

Jun

Apr

Dom

aG

ama

Gof

aSep

-Dec

Apr-

May

May

-Jun

Nov

Sep

May

70