WPS4257
Consumption risk, technology adoption and poverty traps: evidence from Ethiopia
Stefan Dercon and Luc Christiaensen1
Abstract
Much has been written on the determinants of input and technology adoption in agriculture,
with issues such as input availability, knowledge and education, risk preferences,
profitability, and credit constraints receiving much attention. This paper focuses on a factor
that has been less well documented: the differential ability of households to take on risky
production technologies for fear of the welfare consequences if shocks result in poor harvests.
Building on an explicit model, this is explored in panel data for Ethiopia. Historical rainfall
distributions are used to identify the counterfactual consumption risk. Controlling for
unobserved household and timevarying village characteristics, it emerges that not just ex
ante credit constraints, but also the possibly low consumption outcomes when harvests fail,
discourage the application of fertiliser. The lack of insurance causes inefficiency in
production choices.
JEL classification: O12, O33, Q12, Q16
Keywords: Technology adoption, Fertiliser, Risk, Poverty trap, Ethiopia
World Bank Policy Research Working Paper 4257, June 2007
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the
exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if
the presentations are less than fully polished. The papers carry the names of the authors and should be cited
accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the
authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries
they represent. Policy Research Working Papers are available online at http://econ.worldbank.org.
1Stefan Dercon, Professor, Oxford University, CEPR and EUDN, stefan.dercon@economics.ox.ac.uk, and Luc
Christiaensen, Senior Economist, World Bank, lchristiaensen@worldbank.org. Financial support from the
NorvegianFinnish Environmentally and Socially Sustainable Development Trust Fund. and the Economic and
Social Research Council, as part of the World Economy and Finance programme (RES 156250034) is
gratefully acknowledged. The authors would also like to thank Harold Alderman for valuable feedback.
1 Introduction
Households in poor developing countries are typically illequipped to cope with large shocks.
Formal insurance schemes are mostly absent and informal risksharing arrangements and
savings offer only partial consumption smoothing (Morduch, 1995; Townsend, 1995, Dercon
2002). Especially the consequences of covariate shocks, such as droughts, are most often hard
felt, often affecting people's welfare many years after the shock (Dercon, 2004). In
anticipation of such outcomes, households, especially poorer ones, may opt for less risky
technologies and portfolios in order to avoid permanent damage. Yet, these often also
generate lower returns on average (Just and Pope, 1979; Rosenzweig and Binswanger, 1993).
This suggests the potential existence of risk induced poverty traps, whereby those
who can insure their consumption against income shocks can take advantage of the more
profitable opportunities and possibly grow out of poverty, while others are stuck with low
return, low risk activities, trapping them into poverty, even though their inherent risk
preferences may fundamentally be the same. Zimmerman and Carter (2003) simulate for
example that, even when assets are divisible and agents fully rational, optimal portfolio
strategies bifurcate in resourcepoor, risky environments with subsistence constraints and
imperfect credit and insurance markets. Initially wealthier agents obtain higher yielding,
higher risk portfolios, while smoothing their consumption. Initially poorer agents revert to
lower yielding and lower risk portfolios, often absorbing the shock by reducing their
consumption to maintain their asset levels.
While theoretically sound2 and supported by anecdotal evidence (Narayan et al.
2000), whether households actually engage in risk avoidance in the face of subsistence
constraints, ineffective selfinsurance strategies and incomplete credit and insurance markets,
and whether such behaviour is quantitatively important in explaining persistent poverty in
2Sandmo (1971), Eswaran and Kotwal (1990), Kurosaki and Fafchamps (2002).
2
poor, risky, agrarian settings is hard to investigate empirically. The limited available evidence
suggests nonetheless that the income and welfare losses associated with risk avoidance can be
significant, especially in drought prone areas.
Rosenzweig and Binswanger (1993) for example find that a one standard deviation
decrease in weather risk would raise average profits by up to 35 per cent among the lowest
wealth quintile of their sample in semiarid India. Similarly, farmers in Shinyanga, a semi
arid district in western Tanzania, with limited options to smooth consumption ex post, were
found to grow more lower return, but safer crops (in this case sweet potatoes) foregoing up to
20 per cent of their income as implicit insurance premium (Dercon, 1996). Adaptation of the
crop portfolio (substituting fodder for Basmati rice production largely in response to
covariant fodder price risk) was also observed in Punjab Pakistan (Kurosaki and Fafchamps,
2002), despite well developed input and product markets, though income and welfare losses
were smaller (2 and 9.4 percent respectively).
Risk avoidance in the face of incomplete insurance may also be key in understanding
limited fertiliser use (Lamb, 2003). Modern input use, including fertiliser, is an important
determinant of agricultural productivity, and continuing low agricultural productivity is an
important contributor to poverty persistence especially in agriculture based countries such as
in Sub Saharan Africa (Christiaensen and Demery, 2007; Morris et al., 2007). If so, there
would be substantial synergies in complementing interventions that foster access to credit
with interventions that help households cope with shocks (e.g. insurance), a critical insight
for the design of effective poverty reducing strategies.
This paper explores the empirical importance of risk avoidance in fertiliser adoption
in Ethiopia, using a four round panel data set of about 1500 rural households. Fertiliser use
in Ethiopia has remained limited despite concerted efforts by the government to promote its
adoption through improved extension services and access to credit. A host of demand and
3
supply side factors have been invoked to explain the limited adoption of fertiliser in Ethiopia3
including limited knowledge and education (Asfaw and Admassie, 2004), risk preferences,
credit constraints (Croppenstedt, Demeke and Meschi, 2003), limited profitability of fertiliser
use (Dadi, Burton, and Ozanne, 2004; World Bank, 2006b), lack of market access (Abrar,
Morrissey, and Rayner, 2004) as well as limited or untimely availability of the inputs
themselves. Carlsson, et al. (2005), the World Bank (2006a) and anecdotal evidence4 have
recently also highlighted the importance of the households' limited expost consumption
coping capacity.
The paper proceeds by introducing a model of risky input choice in section 2. In this
model, the possible impact of seasonal credit constraints on input adoption is distinguished5
from intertemporal constraints related to risk and consumption outcomes, a key contribution
of the paper. An empirical model to test these propositions is presented in section 3. Section
4 describes the data with a particular emphasis on the effect of fertiliser use on profit
variability. The econometric results are discussed in section 5 and section 6 concludes.
2 A theoretical model of risky input choice
Households derive income from agricultural production, which involves determining the
level of risky inputs (such as high yielding varieties and fertiliser) that increase both the mean
and the variance of the net returns to production. The level of input use has to be decided
before the rains have come and the harvest is known, i.e. before uncertainty has been
resolved, and often in the face of imperfect credit and insurance markets.
3 Morris et al. (2007) provide a comprehensive review of the factors affecting fertiliser use in Africa. Feder et al.
(1985) review the international evidence.
4 Bonger et al. (2004). Largely uncorroborated but insistent reports exist that strict enforcement of repayment of
loans that are granted without insurance in case of crop failure may well have discouraged farmers to continue
fertiliser use after local or widespread droughts such as in 2002.
5 Duflo et al. (2006), in an innovative field experiment in rural Kenya, identified access to finance at the right
time as the critical constraint to fertiliser adoption. While they specifically focused on the inability to save over
the agricultural cycle to have sufficient funds when fertiliser needs to be applied, widespread access to credit to
finance fertiliser adoption, as in Ethiopia, is another way of alleviating this constraint.
4
This decision making process, common in many rural settings, is modelled building
on Evans and Jovanovic (1989), Eswaran and Kotwal (1990), Morduch (1990), and Deaton
(1992). We explore in particular the implications of a household's capacity to protect its
consumption from falling in case of a shock for its ex ante risktaking in agriculture (i.e.
assuming income endogenous), beyond the effects from workingcapital related credit
constraints. This means taking into account both limitations on using insurance, credit or
savings expost, i.e. after uncertainty has been resolved, as well as credit constraints exante,
when input decisions have to be taken, i.e. before uncertainty has been resolved.
To highlight the differential impact of input credit market imperfections, and risk and
coping capacity, the adoption of risky inputs is first modelled in a world without uncertainty
but with imperfect credit markets, and then in a world with uncertainty. The level of risky
inputs determines riskiness in production. This suffices to capture the core insights regarding
the dynamic interaction between limited expost consumption smoothing capacity and ex
ante production choices. Abstraction is made from other income risk reducing mechanisms
(such as land and labour allocations to diversify the crop and income portfolio).
Denote gross returns at the end of period t as g(xt ), with xt the quantity of inputs used,
to be decided at the beginning of the period, and g(.) increasing at a decreasing rate in xt. For
now, there is no risk. Input prices are px and inputs have to be paid before the harvest is
known, although we will allow for credit. Purchased inputs, such as fertiliser, are divisible
and can be used in small quantities, with limited startup costs in production. Still,
transactions costs in contacting and travelling to suppliers, and learning, may imply some
sunk costs6. Net returns from agricultural production can be defined as yt = g(xt ) pxxt 
I(xt).m, with m defined as sunk costs incurred from using fertiliser (m 0) and I(xt) is the
indicator function taking on the value one if fertiliser is used, and zero otherwise.
6World Bank (2006b) reports that most fertiliser in Ethiopia is sold in bags of 25 kilograms.
5
Assume that households optimize intertemporal welfare defined in consumption.
Suppose they have an intertemporally additive utility function u defined over lifetime T as:
ut = (1+ )t v(c )
T
(1)
=t
with (.) instantaneous utility derived from consumption c ( 0) and the rate of time
preference, ('(.)>0, ''(.)<0). Define r as the rate of returns of savings between periods and
At+1 as assets at the beginning of period t+1. Assets evolve from one period to the next
according to:
At +1 = (1+ r)(At + g(xt) pxxt  I(xt ).mct) (2)
We assume for simplicity that assets can be liquidated at any point in time. Consumption
prices are used as the numéraire. Consumption decisions are made after income has been
generated from production, while inputs have to be paid beforehand unless credit is available.
Even though (formal) consumption credit is rarely available, input credit is common,
an important distinction to be taken into account. In Ethiopia, inputs such as fertiliser are
provided under regional government guarantees usually offering seasonal credit without
collateral. Repayment is strictly enforced, and default rates are low, albeit nonzero.7
Following Evans and Jovanovic (1989), suppose that that enforcement is not perfect, but that
those caught are punished by losing the equivalent of a proportion of their assets, and that the
net return on their assets and production for those not repaying pxxt will only be
g(xt )At  I(xt ).m , with (>0) determined by factors such as the probability of getting
caught and the share of assets impounded when caught. However, lenders will only offer
7Bonger et al. (2004) for example found that 20 percent of the farmers in their sample did not fully repay the
fertiliser credit, largely due to harvest failure. Those farmers faced severe penalties such as imprisonment, or
had to sell livestock and other property or sell their food items.
6
credit if there are incentives to repay loans, or if net returns when repaying outweigh returns
when cheating, i.e.8:
At  pxxt 0 (3)
This equation is then the seasonal credit constraint, whereby credit is an increasing
function of initial assets levels. When assets are fully liquid, values <1 won't hold since the
household can use cash purchases up to the value of its assets. If =1, a household's
purchases of inputs must be fully collateralised. They are indifferent between borrowing for
the purchase of inputs (to repay later) and selling the assets now to purchase the inputs.
Higher values imply access to notfully collateralised credit, such as borrowing against future
harvests. Reflecting the Ethiopian realityseasonal loans without collateral but with harsh
enforcementwe assume 1, nesting the more conventional case of credit constraints
( At  pxxt 0 ).
Consumption is decided after income has been generated from production and after
seasonal credit has been repaid. Borrowing against future income is not possible9 and
consumption is limited to the sum of the realized income and the value of assets At at the end
of t, or formally:
At = At + g(xt )  pxxt  I(xt )m  ct 0
+1 (4)
This is the "consumption credit constraint" `expost', which together with the (`exante')
seasonal credit constraint (3) and the transversality condition AT+1=0, i.e. there is no savings
beyond the last period T, and ct 0, form the constraints of the optimization problem. The
value function, defined in initial asset levels, can be written as:
8It is assumed that no interest rate is charged on seasonal credit, but this could be easily introduced, without
affecting the general thrust of the results.
9Given that informal insurance is only partially effective at best in insuring households against idiosyncratic
shocks and ineffective in insuring them against covariant shocks (Townsend (1995), Morduch (1995)), we
abstract from effects of informal insurance on consumption smoothing. Lamb (2003) highlights the potential for
ex post consumption smoothing in semiarid India through the labor market. In Ethiopia, opportunities for both
offfarm employment and seasonal migration are very limited.
7
Vt (At ) = maxv(ct ) + 1 Vt+1((1+ r).(At + g(xt )  ct  pxxt  I(xt )m))
ct ,xt 1+ (5)
+ t(At  px xt
) + t (At + g(xt )  pxxt I(xt )m  ct )

Solving this problem backwardly implies that we get the solution for optimal
consumption in each period and then derive the optimal input decision, given the optimal rule
for deciding on consumption. The optimal consumption rule satisfies:
Vt (1+ r)
(6)
ct = v'(ct) (1+ ) .Vt+1'(At+1 )  t = 0
Given (6), when the input decision is being taken (and given that the seasonal credit
constraint has builtin incentives for credit repayment to be the optimal decision expost), the
subsequent intertemporal (or `expost') budget constraint is not relevant and only the seasonal
credit constraint matters. The optimal level of input use (for nonzero input use) can be
obtained from (5) as:
Vt = (1+ r) yt t px = 0 (7)
xt ( 1+ ) .Vt+1'(At+1) + t xt
Substituting (6) into (7) gives the optimal decision rule for the adoption of xt:
Vt = v'(ct)xt 
yt tpx = 0 (8)
xt
Equation (6) is the standard rule for intertemporal consumption with (consumption)
credit constraints: if binding, then the marginal utility of current consumption is higher than
the marginal value of future (appropriately discounted) consumption, or consumption now is
lower than optimal. Equation (8) shows how the seasonal credit constraint may result in
inefficiency: if seasonal budget constraints do not bind (t=0), an efficient allocation in
production is obtained when the marginal value product equals the input price, i.e. when the
marginal net return equals zero. Otherwise, the seasonal credit constraint and decreasing
8
marginal returns imply that suboptimal and lower input levels are obtained, with positive
marginal net returns. Note furthermore that, given positive marginal returns, the larger is the
input use, the larger is the household's income and the less likely it is that t binds. In the
deterministic case, it is thus optimal to choose the input level that maximizes income, as this
also minimizes the risk that the intertemporal liquidity constraint binds.10
To introduce risk in the income process, consider gross returns being governed by
g(xt, t), whereby t is a random, serially uncorrelated shock, realised after input decisions
have been made. It is assumed that min[ g(xt ,t )]= a 0, i.e. the lowest gross returns are
nonnegative. Define yt=g(xt,t)pxxtI(xt)m as the net returns to production and
Et(yt) xt > 0 (with Et the expectation at the beginning of t before yt is known), so that
expected returns with risky inputs are always higher than without inputs, albeit at a
decreasing rate, 2Et(yt ) xt < 0. It is further assumed thatyt t > 0, that yt xt > 0 if
2
t > 0 and that yt xt < 0ift < 0. In good times, choosing more risk increases net
returns, while in bad times, choosing more risk results in losses, or the dispersion of net
returns increases as inputs increase.11 If the household were maximising expected income
(i.e. riskneutrality), it would choose xt such that Et yt g(xt,t)
xt = 0, or Et xt = px, and some
inputs would always be used given the assumptions (including that input use is profitable in
expectation).
10 The sunk costs m do not play a role in deciding the level of input use here, since these decision rules are
derived for nonzero input use. Nonzero input use will only apply if it is profitable to do so, i.e. it yields
positive net returns yt. The sunk costs imply that there is a threshold level of fertiliser below which marginal
return to using more fertiliser is positive but levels used are zero since it would result in negative overall profits
(net returns). Introducing risk does not affect this. Formally, this is equivalent to introducing the condition g(xt)
pxxtI(xt)m 0 into the optimization problem.
11We remain agnostic about whether the increased riskiness stems from higher risk in yields or (given the non
zero costs of inputs) higher risk in returns, even if yields are not `more risky'. Inputs could also be perceived to
be more risky given limited knowledge of the new production technique inducing a degree of subjective
uncertainty which typically declines as producers become more familiar with the technique (Hiebert, 1974). In
the empirical analysis, we do not find evidence that fertiliser results in more risky yields, but the nonzero costs
result in higher, but more risky returns as assumed in the model.
9
Assume that the household maximises the expected flow of utility from consumption,
ut = (1+ )t Et v(c ). All assumptions made before regarding risk aversion are
T
=t
maintained. The period t value function for the household can now be written as:
1
Vt (At ) = max Et v (ct ) +
1+ Vt+1((1+ r).(At + g(xt ,t )  ct  pxxt  I(xt )m) + t(At  px xt)
ct ,xt +t(At + g(xt,t) pxxt  I(xt)mct)
(9)
with t and t defined as before. Backwardly solving this problem, we first derive the optimal
consumption rule (after uncertainty over income has been resolved):
Vt (1+ r) Vt '(At ) + t = 0
(10)
ct = v'(ct)  Et(1+ ) . +1 +1
Given (10), we subsequently take the derivative of (9) with respect to xt at the beginning of t,
i.e. before uncertainty has been resolved and obtain the optimal decision rule for xt (for non
zero input allocations) as:
Vt = Et
'  = 0 (11)
xt ( (1+ r) Vt+1(At+1)
1+ ) +t yt t px
xt
Expanding equation (11) yields:
Vt ' +t Etxt +cov(
yt ' yt t px = 0 (12)
xt = Et( (1+ r)Vt+1(At+1)
1+) (1+ r)Vt+1(At+1)
1+) +t ,xt 
Since t and t+1 are uncorrelated and given (10):
cov ( (1+ r)Vt ' ' t = covv'(ct),xt
yt
1+ ) +1(At ) + t , xt = covEt (
+1 yt (1+ r)Vt
1+ ) +1(At ) + t ,
+1 yxt
From (10) it can furthermore be seen that Et (v'(ct )) = Et (1+ r)
(1+ ) . Vt+1 '(At+1) + t at the
beginning of t. Using these insights, (12) can be rewritten as
10
Vt yt = 0
xt = Et[v'(ct)].Et xt +Covv'(ct),xt  yt t px (13)
which is equivalent to:
Vt = Et v'(ct)
 = 0 (14)
xt yt t px
xt
While there is obviously much similarity between (14) and (8), uncertainty and risk
aversion make it no longer optimal to maximize income by maximizing the amount of inputs
used. Given risk aversion, low levels of consumption (i.e. higher marginal utility) will have a
higher weight in the expected value in (14), so that incentives exist to allocate inputs at lower
levels (i.e. higher marginal returns) than without risk. In other words, risk may result in lower
risky inputs in production.12
These insights follow readily from equation (13). To see this, note that the
covariance in (13) is nonpositive. In moving from a good to a bad state of the world, the
marginal return to input use turns negative, or income is lower for higher levels of input use
in a bad state of the world. If as a result, the consumption credit constraint (t) binds and
current consumption has to be reduced, the marginal utility of consumption will increase
the marginal return to risky inputs and the marginal utility of consumption move in the
opposite direction. In other words, if at least in some states of the world the consumption
constraint (t) is likely to bind, the covariance is negative. As a result, the expected marginal
return to input use increases, making it optimal to reduce the use of risky inputs. If liquidity
constraints never bind, the covariance will be zero.
12This may seem a trivial result, and could be obtained from a static model with income risk and risk aversion.
However, this result is dependent on consumption being lower when poor harvests occur, and in an
intertemporal model this means that intertemporal constraints have to be taken into account. To see this, note
from (10) that the presence of credit constraints in particular states of the world at t (t>0) implies that current
marginal utility will be higher (and thus consumption lower).
11
More broadly, the insights related to equations (13) and (10) can help us to identify
whether the choice of risky inputs is determined just by seasonal `exante' credit constraints
(t) or whether possible riskrelated intertemporal `expost' credit constraints (t) are relevant
as well. Consider the following scenarios. If consumption can be kept smooth over time (t:
t=0 and t0), then the covariance between marginal returns in different states of the world
and marginal utility is zero. The only cause for a deviation from a riskneutral allocation
based on expected marginal returns to inputs equal to zero would be seasonal credit
constraints (t>0) effectively similar to (8). These seasonal credit constraints would be
determined by the levels of assets available and the nature of the credit market constraints at
the time of the input decision.
However, whenever ex post consumption credit constraints are more likely to bind
due to limited ex post coping capacity (t>0 for some t and t0), the choice of risky inputs
is likely affected as it affects the likelihood of ex post credit constraints to bind. In particular,
poorer households (with limited assets At) are more likely to forego risky inputs (such as
fertiliser), not just because they have less access to credit (t0) but also because they are less
able to avoid consumption shortfalls (t>0). Note furthermore, at low levels of consumption,
instantaneous utility is likely to be steep and highly concave as the household is concerned
about very low levels of consumption. As a result, small reductions in income will result in
large increases in marginal utility and consumption credit constraints will be much more
likely to bind.
In sum, factors contributing to more likely binding consumption credit constraints
including more risky production patterns and less smoothing possibilitiesreduce people's
willingness to take risk. This effect goes beyond just risk averse preferences: risk averse
households with appropriate means for risksharing and consumption smoothing, so that their
marginal utilities expost are not affected by particular outcomes, could take decisions on
12
production as if they were riskneutral. Lower risk taking results in lower returns on average,
and perpetuates poverty.
3 Empirical Approach
To test the key prediction from our model fewer risky inputs will be used, ceteris paribus,
when households face higher expost downside consumption riska credible identification of
each household's downside consumption risk is needed. This information is subsequently
used to explore whether households' expectations about ex post consumption downfalls (t)
affect their ex ante decisions on modern input use, in addition to seasonal credit constraints
(related to the need for working capital). If the likelihood of consumption downfall is
properly controlled for, then measures of current liquidity positions (At) and any required
down payments for inputs would allow one to distinguish (exante) working capital
constraints (t) from (expost) consumption smoothing constraints (t).
Equation (14) shows that input decisions are based on expost consumption outcomes
in expectation. Obviously, using information on actual consumption expost to proxy this is
endogenous as it is itself a function of actual adoption of the inputs. Initial assets could be
used as instruments (Morduch, 1990), though they would not allow us to distinguish seasonal
credit constraints from expost consumption risk. Instead, a parametrised version of the ex
post consumption model is first estimated including explicit information on shocks, and the
exante consumption risk faced by the household is then simulated using historical data on
rainfall shocks and conditioned on other current household and community characteristics.13
By applying the envelope theorem to equation (6) and assuming that households can
perfectly smooth consumption, the optimal consumption path is defined by:
13 This means that we assume rational expectations, i.e. households know the underlying consumption model
including the exante distribution of the stochastic variables and the effect of these shocks on consumption. This
procedure is similar to Kazianga and Udry (2006) in their test for precautionary savings in Burkina Faso.
13
v'(ct) = (1+ r)
(1+ ) v '(ct )
+1 (15)
Assuming constant relative risk aversion with direct marginal utility defined at t as ct e , t
with defined as the coefficient of relative risk aversion and t a general taste shifter, taking
logs, and introducing subscript i to denote households, an empirical specification can be
obtained from (15) as14:
ln ccit = ln(1+ r) ln(1+ ) + (it+1 it )
it+1 1 ( ) (16)
Equation (16) suggests that the path of consumption over time is only affected by taste and
preference shifters, as long as there are no binding liquidity constraints over time and
provided the underlying factors determining wealth (or permanent income) are not changing.
Overidentifying this equation to reflect shocks (Sit+1) to income and possible heterogeneity
in households' capacity to cope with these shocks15, this leads to the following linear
specification:
ln cit+1 ` (17)
cit =0 +1Zit +1 +2Sit +1 Bit + eit +1
with Zit+1 taste and other preference shifters (such as, changes in household composition,
price changes, or seasonality shifters), Sit+1, shocks linked to idiosyncratic and common risk
and eit+1 an error term. If consumption credit constraints do not bind, then 2 is zero.
The first regression to be estimated is based on (17), though expressed as a household
fixed effects levels model, rather than a difference model. Define Xit as a set of (exogenous)
household characteristics affecting preferences and `permanent' income (such as changes in
14In (16), r and are assumed constant across households. It is less straightforward to defend that is assumed
constant, i.e. constant relative risk aversion with the same coefficient across households. Given fixed effects,
some of the heterogeneity in risk preferences will nevertheless be controlled for in this regression, as well as in
the adoption regression discussed below.
14
household composition) and Sit are variables that describe different sources of risk (such as
weather, pests, and general health conditions). To capture households' differential ability to
cope with changes in income, Sit is interacted with liquid asset levels Bit, proxied by livestock
at t1 to avoid endogeneity. Livestock is the most important liquid asset in rural Ethiopia.
The indicator variable Git is introduced to reflect the fact that it may be easier to protect
consumption from positive (e.g. good harvests) than from negative shocks. Unobserved
household fixed effects are represented by vi and it is a white noise error term..
lncit = o + 1Xit + 2Sit.Bit.Git + 3Bit +i + it (18)
From equation (18) consumption expectations for different possible values of the shock
variables can be obtained to investigate whether expected values in `bad' years matter for
input adoption.
Equation (14) implied that the demand for risky inputs will be influenced by factors
influencing the marginal utility and the value of the marginal productivity of these inputs.
These include: Lit fixed inputs (such as land endowments) and other household specific
characteristics, some of which are fixed, such as land quality or risk preferences; Vit, input
and output prices and other community16 and agroecological characteristics; Ait, asset
variables capturing (exante) working capitalrelated credit constraints at the time of making
the input use decisions; and g(cit), expectations about (expost) consumption outcomes,
weighted towards the anticipated downside risk in consumption; it reflecting unobserved
community characteristics; and i reflecting unobserved time invariant household
characteristics. uit is a white noise error term. In a linear specification, this could be written
as:
15 Datt and Hoogeveen (2003) and Christiaensen and Subbarao (2005) provide empirical evidence of the
differential ability of households to smooth consumption in the face of shocks in the Philippines and Kenya
respectively.
16Vit may also capture changing availability of fertiliser, the changing presence of extension officers offering
information on fertiliser use, and more general learning and increased familiarity with fertiliser use.
15
xit = 0 + 1Lit + 2Vit + 3Ait + 4g(cit ) + +i + uit (19)
it
Identification of g(cit) is achieved by using counterfactual values based on possible
(not actual) realisations of the shock variables. From equation (18), different households
have a differential ability to bear risk over time. When combined with the historical rainfall
distribution data for each cluster, time variant and household specific ex ante (counterfactual)
consumption distributions can be generated to construct the potential downside risk variable
at the time of the fertiliser use decision g(cit).
This leaves the choice of the relevant counterfactual. If fertiliser adoption results in
changing the distribution of consumption outcomes (higher mean but higher variance), then
estimating (18) without controlling for fertiliser use effectively ignores the differential
consumption risk distribution for users versus nonusers. Inclusion of fertiliser use in model
(18) would introduce an endogenous variablethe theory clearly showed nonseparability of
consumption and production decisions if risk is not fully insured. Instead, the reduced form
specification as in (18) is used, in effect ignoring fertiliser as a variable shifting the
distribution of consumption, to avoid introducing endogeneity. It will be shown below that
fertiliser use results in larger downside income risk, and that households have in general a
limited ability to smooth consumption. The reduced form approach thus offers a lower bound
on the risk faced by households.
16
4 Fertiliser use, agricultural production and households in rural Ethiopia
Our data are taken from the Ethiopia Rural Household Survey (ERHS), which comprises
1477 households in 15 Peasant Associations across the four major regions in Ethiopia.
Households were surveyed 4 times between 1994 and 1999. The sample is broadly
representative for the main farming systems in the country, including the oxplough cereal
producing areas in the northern and central areas (about 63 percent of the sample), the enset17
dominated areas which are typically also suitable for coffee and chat (about 30 percent) and
the much smaller hoebased cereal areas (about 7 percent).18 Attrition is low. For about 88
percent of the sample, we have observations in each year, but many households were
recovered within the sample during the five year period and in 1999 there is information for
94 percent of the households interviewed in 1994. Estimations are done on the unbalanced
panel.
Cereal yields in Ethiopia are currently only about 1,250 kg per hectare, compared
with 2,500 and 4,500 kg per hectare in South and East Asia respectively.19 Yields increased
only marginally over the past decade (by 0.3 percent per year between 1991/92 and 2003/4),
most of it accounted for by increased maize yields and an expansion of the fertilized area
(about 43 percent of the area under cereals fertilized in 2002/3, up from 32.5 percent in
1994/5). Intensity remained constant at about 100kg per hectare and the number of users
remained at less than a quarter of all farmers.20
Consistent with national trends both adoption rates and intensity of fertiliser use in the
sample are low with an overall expansion in the main cereal areas and a decline in the enset
and other permanent crop areas (Table 1). Only 22 percent of all households used fertiliser in
each period and many households switch in and out of fertiliser from year to year. In 1999,
17A permanent food crop, commonly known as false banana.
18See Dercon, Hoddinott, and Woldehanna (2005) for more details on the sample.
19Average yields during 20032005 (World Bank, 2007).
20See World Bank (2006b) for a comprehensive review of the performance of the agricultural sector in Ethiopia
17
14 percent of households did not use fertiliser, despite having used it in one of the three
earlier rounds and 3 percent stopped using fertiliser after having used it every round before.
Three quarters of all purchased quantities are multiples of 25 kilogrammes (a small bag),
suggesting the existence of a small threshold due to fixed costs or other indivisibilities.
Table 1: Fertiliser use in ERHS, 19941999
Incidence of farmers using fertiliser (%) Application rate per hectare1) (kg)
Main Enset and other Total Main Enset and other Total
cereal permanent crop cereal permanent crop
areas areas areas areas
1994 43.0 44.5 43.5 35.0 33.9 34.6
1995 41.6 28.0 37.0 31.1 13.2 25.1
1997 49.9 41.8 47.3 32.1 25.0 29.8
1999 50.0 36.4 45.5 39.0 18.1 32.2
1)Average across users and nonusers.
Source: Data from the ERHS 199499
In 1994, about half the farmers quoted costs as the main reason for not using modern
inputs (including fertiliser), increasing to about 60 percent in 1999 (Table 2).21 Only 15
percent mentioned limited availability as a constraint.22 The percentage of farmers indicating
nonsuitability of the agroclimatic conditions as reasons for nonadoption, declined from 19
percent in 1994 to only 8 percent in 1999. This coincided with the declining importance of
knowledge and skills in adopting modern inputs and suggests better understanding of how
fertiliser works over time. Only 3 percent reported not to have the relevant skills in 1999.
over the past 25 years (including the functioning of the fertiliser market).
21In 1994, the question distinguished profitability from costs and lack of cash, and very few farmers (2 percent)
suggested that the profitability of fertiliser itself was the problem, settling for cost reasons, suggestive of credit
constraints. In 1999, this distinction was not made and the two groups are reported together in table 2, even
though profitability had likely declined following the fertiliser price increase since 1997.
22In 1994, lack of availability was concentrated in two enset growing villages. In 1999, this response was
concentrated in one coffee producing village (Adado, near Dilla in SNNPR), and most likely referred to
pesticides for coffee disease.
18
Table 2: Main two reasons for not using modern inputs (% of farmers reporting)
1994 1999
Too expensive/low profitability/lack of cash 49 60
Fertiliser not available in area 15 15
Soil/crops/climate not suitable 19 8
Don't have skills 10 3
Other 7 15
Source: ERHS
Consistent with the observation that input use is expensive, real fertiliser prices in the
areas under study at the time of planting increased by 28 percent between 1993/4 and
1998/99 (see Table 3). This follows the gradual removal of panterritorial price fixing,
completed by 1997/98. However, cereal prices did not follow this increase and the relative
outputfertiliser price decreased considerably during the latter years of the survey.
Table 3 : Evolution of fertiliser and cereal prices during 19941999.
Average Fertiliser Price per Average cereal price Average cereal/fertiliser
Quintal (per 100 kg in 1993 (Ethiopian birr) per 100 kg price ratio
prices) in 1993 prices
1993/94 141 227 1.61
1994/95 129 242 1.88
1996/97 176 212 1.20
1998/99 180 221 1.23
Fertiliser prices are the average of DAP and UREA prices around the time of planting for the main season
(June). DAP and UREA are the two main types of fertiliser used. Cereal prices are averages based on village
specific prices derived from price surveys in the Ethiopian Rural Household Survey at the time of the harvest
given that loans must repaid upon harvesting. All prices are deflated using the consumer price index.
Source: Development Studies Associates (2001), IMF, International Financial Statistics, and
own calculations using ERHS.
The cost of credit adds to the cost of fertiliser. In 1999, 71 percent of those
purchasing fertiliser used seasonal credit and the implicit median interest rate is calculated at
57 percent per year.23 While the data on the terms of loans are noisy, incurring credit appears
23A down payment of about 0.65 birr per kg or about 30 percent of the purchase price is required. Interest
payments and other costs related to the loan have a median of about 0.34 birr per kg, though the data on the
terms of loans in the survey are noisy and there is a lot of variation around this. Given a median repayment
period of about 7 months, this suggests an interest rate of about 30 percent for the median loan duration (or 57
percent per year). In an economy with consistently very low inflation (in this period, about 4 percent per annum
on average) this is substantial.
19
costly. Most of those using cash to purchase fertiliser mention the perceived high interest
rates (and not lack of fertiliser availability) as main reason for using cash. The latter is not an
important reason for not using fertiliser either.
Understanding the profitability of fertiliser use under uncertainty is critical for our
analysis. However, data on input application and outputs by crop and plot were only collected
in 1999. In other years, the data were collected at the farmlevel, so crops that use modern
inputs cannot be identified. Given the high variability in climatic and other conditions, it is
thus difficult to establish the distribution of returns for crops in different areas, with and
without fertiliser application.24 Nonetheless, comparing the (counterfactual) yields across
plots with and without fertiliser based on (crosssectional) production function analysis, is
still suggestive. About 28 percent of all cereal plots in the sample are fertilized.25
A standard CobbDouglas production function is estimated linking plotlevel cereal
yields to plot size, livestock, the input of different types of labour (male, female and
children), fertiliser input, herbicide, fungicide, insecticide, controls for land quality and slope,
controls for the particular crop grown as well as a series of villagelevel (e.g. rainfall) and
idiosyncratic (e.g. pests) shock variables. A set of interaction terms of the rainfall variables
and fertiliser use are introduced to account for different sensitivity of yields to fertiliser
across the rainfall distribution. Yields are significantly affected by fertiliser use, by rainfall
and other shock variables, as well as their interactions. The full estimation results and
robustness tests using household fixed effects are reported in Appendix 1.
This regression is used to construct (counterfactual) yield and return distributions for
an otherwise average farmer for 1999, by simulating through the rainfall distribution from the
nearest rainfall station to each survey site26 and the distribution of a selfreported quality of
24Lack of agricultural labour input data in the other rounds and inconsistencies in the questionnaire design of
the crop production modules across rounds made production analysis at the household level inappropriate.
25Fertiliser is mainly used on plots with cereals but most farmers using fertiliser do not use it on all their plots.
26For each station, rainfall levels over the past 20 years or more are available. In simulating the counterfactual
20
rainfall index derived from the different survey rounds. For each crop, average fertilizer
application rates among those using fertilizer are used.
Across the rainfall distribution, cereal yields on fertilized plots dominate yields on
nonfertilized plots, despite being more risky (Figure 1). Gains can be up to 24 percent
(given average fertiliser use)27, declining as we move away from the median rainfall and
virtually disappearing in periods of extreme droughts and floods (below the 20th and above
the 80th percentile).28 At the 19945 prices, fertiliser use is also profitable over a wide range
of rainfalls, though not at the extremes. At median rainfall levels, returns were 11 percent
higher when using fertiliser in 199495; beyond the 3060th percentile rainfall range, they
were lower.
While the estimates are approximate at best, these results give credence to our
contention that fertiliser use is a high return, but high risk technology. Given the sharp
increase in fertiliser price across the survey rounds, returns to using fertiliser were lower in
the last few years of our sample, and turning negative more rapidly when moving away from
the median.29 In conclusion, our sample encompasses a period where cereal production under
fertiliser yields is more profitable on average, but more risky, facilitating the identification of
the effect of limited ex post coping capacity on fertiliser adoption. The findings further
highlight the critical need to control for changing price ratios and other factors in analyzing
fertiliser use.
cereal yield distribution, the 10th percentile rainfall is equivalent to the 10th percentile of the rainfall
distribution, and so on.
27Detailed results can be found in Appendix 1.
28As 87 percent of the sample had rainfall levels between the 21st and 80th percentile, caution is warranted in
interpreting the counterfactual results for very low and very high rainfall.
29The profitability of fertiliser use is sensitive to the cerealfertiliser price ratio and at average 199799 prices,
cereal cultivation with fertiliser use may on average no longer be more profitable than cereal cultivation without
fertiliser use. Indeed, most recently, two World Bank studies (2006b; Morris, et al., 2007) have highlighted the
lack of profitability as an important factor in explaining limited fertiliser use in Ethiopia in particular, and Sub
Saharan Africa more widely. Yet, they also emphasize that fertiliser use is the outcome of many other demand
and supply factors, including agronomic practices that strongly affect the physical (and thus also the economic)
returns to fertiliser use.
21
Figure 1: Distribution of yields and returns for cereals in 19941995
2000
1800
1600
s)nruter(rr 1400
1200
bi
d 1000
an
s)dleiy( 800
600
kg
400
200
0
10th 20th 30th 40th 50th 60th 70th 80th 90th
percentile of rainfall distribution
yield, no fertiliser yield, fertiliser return, no fertiliser 9495 return, fertiliser 9495
Note: Simulations based on the estimates reported in table A.1, column (3). Yields based on estimated output
per hectare for an average plot (i.e. with mean characteristics for plot, farmer and village), among 2294 plots,
and the mean application rate for fertiliser users for each crop. For example, the 10th percentile reflects cereal
yields when the rainfall was equivalent to the 10th percentile of the rainfall distribution, i.e. very poor rains.
Returns are the gross returns (yield times output price, evaluated at the mean output price in 19945) minus the
cost of the fertiliser (using the mean fertiliser price of at the time of planting in 19945). Prices are expressed in
1999 prices.
Source: Own calculations using EHRS.
5 Empirical results
Equation (18) is first estimated using a household fixed effects regression applied to the
1994, 1995, 1997 and 1999 ERHS data. "Permanent" income terms are thus tied up in the
fixed household effects. To investigate the sensitivity of consumption to shocks, (village
level) rainfall30, and a set of variables describing idiosyncratic shocks are included. As
downside risk may be harder to handle"good" shocks could presumably be savedthe
rainfall variable is interacted with a dummy with value one when rainfall is below the median
level of the last twenty years. The idiosyncratic shock variables include an index based on
(selfreported) descriptions of the quality of `rainfall' (one if the rainfall distribution was
30Not only do we have four time periods, but there is substantial between and withinvillage variation in this
period in the villagelevel rainfall data. First, 24 percent of the variation is withinvillage variation, i.e. not
explained by village fixedeffects and time dummies. Only 6 percent is explained by the time dummies, i.e.
22
satisfactory in all respect, zero if unsatisfactory in all respectssee appendix), an index of
`nonrain' crop shocks, such as weed damage, plant diseases and insect damage, which is one
if there is no problem, and illness shocks (the number of adults ill in the month before the
survey).
To capture differential risk bearing capacity, all shocks are interacted with the (natural
logarithm) of livestock holdings per adult (measured at t1). Livestock values were scaled by
the median value of livestock in the village to control for the different role livestock may play
across farming systems.31 Despite considerable persistency, livestock possessions vary over
timeabout 25 percent of variation remains in the livestock holdings data after controlling
for household fixed effects and time trends. Through the interaction terms, this variation
allows us to identify the timevarying consumption risk at the household level.
Livestock holdings at t1 are also included separately in the regression to control for
timevarying changes in wealth.32 Finally, in addition to basic timevarying household
characteristics, such as household demographics, which may reflect taste shifters33, a post
harvest period dummy is included, to control for seasonal variation in food prices and
consumption.
The effect of villagelevel rainfall shocks on consumption is substantial, but smaller
for those with relatively high livestock holdings (column (1), Table 4). For example, a 10
percent drop in rainfall reduces consumption by 1.5 percent in a household with 2.7 times
more livestock than the village median, compared to 2 percent for someone with median
livestock holdings.34 Idiosyncratic shocks have no significant impact.35
common patterns over time between the villages, and the rest, 70 percent, is between villagevariation.
31The findings are only marginally affected by this scaling, but offer advantages in terms of interpretation.
32Initially, land was included as well but it was systematically insignificant, possibly because of
multicollinearity (Pearson correlation coefficient between land and livestock possessions is 0.75). As it was only
included as an alternative proxy for wealth, it was dropped.
33Demographic variables also act as an implicit control for incorrect equivalence scales and the lack of allowing
for economies of scale in the nutritional adult equivalent corrections used in the construction of the left hand
side variable, consumption per adult equivalent.
34Note that given our definition, the interaction term is zero for someone with median livestock holdings. The
23
Those with higher levels of livestock (above the median) are less sensitive to poor
rainfall outcomes than to rainfall outcomes above the long term median (column 2). In other
words, `positive' rainfall shocks are not fully saved, but reflected in consumption; a result
consistent with other studies on Ethiopia (Dercon, 2004).36 For simulating counterfactual
consumption distributions for our sampled households, the insignificant rainfall variable
interacted with below the median rainfall dummy is dropped (column (3)).
The two key conclusions for the purposes of this paper are, first, that covariate rainfall
shocks are the most important source of uninsured risk in these communities and second, that
households clearly differ in their ability to cope with shocks depending on their livestock
wealth.
specification controls for village specific rainfall averages through the fixed effects. Finally, the result does not
necessarily imply that livestock is being used to smooth consumption, but rather, that better smoothing is
correlated with having more assets. Wealthier households may have better access to village level risksharing or
to other coping mechanisms, such as offfarm employment or migration opportunities when required.
35The lack of significance of the idiosyncratic shocks may follow from errors in measuring (selfreported)
idiosyncratic shocks. As yields are highly sensitive to some of these shocks (appendix 1), it is also plausible that
households actually manage to insure themselves against idiosyncratic risk..
36This may reflect choices by householdsa preference for feasting when good rains occurrather than their
inability to smooth consumption. This provides support for focusing on the downside risk in consumption in
assessing its relevance for production decisions, as pursued below.
24
Table 4 : Determinants of real consumption per adult. Household fixed effects regression.
Dependent variable: ln of real consumption per adult equivalent (1) (2) (3)
ln livestock holdings in t1 (per adult, relative to village median) 0.445 0.554 0.547
[2.57]** [3.07]*** [3.03]***
ln annual level of rainfall 0.200 0.218 0.190
[4.30]*** [4.03]*** [4.07]***
ln rainfall * ln livestock holdings at t1 (relative to median) 0.049 0.062 0.061
[1.96]* [2.41]** [2.37]**
quality of rain index (1 is best, 0 is worst) 0.091 0.094 0.086
[0.74] [0.77] [0.70]
quality of rain * ln livestock holdings 0.027 0.017 0.018
[0.38] [0.24] [0.25]
quality of rain squared 0.065 0.065 0.059
[0.59] [0.59] [0.53]
quality of rain squared * ln livestock holdings 0.013 0.009 0.008
[0.21] [0.14] [0.14]
Index of pests, trampling, frost, flooding (1 is best, 0 is worst) 0.003 0.015 0.019
[0.02] [0.09] [0.12]
Index of pests etc * ln livestock holdings in t1 0.114 0.111 0.111
[1.40] [1.35] [1.35]
index of pests squared 0.253 0.252 0.247
[1.02] [1.02] [0.99]
index of pests squared*ln livestock holdings in t1 0.226 0.234 0.23
[1.63] [1.68]* [1.66]*
Number of adults ill in last 4 weeks 0.021 0.022 0.023
[1.08] [1.15] [1.21]
adults ill * ln livestock holdings in t1 0.001 0.002 0.001
[0.07] [0.17] [0.15]
ln rainfall * rainfall below median dummy 0.005
[1.00]
ln rainfall * ln livestock * below median 0.005 0.006
[1.93]* [2.02]**
Constant 3.532 3.417 3.614
[10.77]*** [8.88]*** [10.94]***
Observations 4336 4336 4336
Rsquared 0.13 0.13 0.13
Absolute value of t statistics in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%.
Note: The regression includes further controls for whether consumption is measured in post harvest period, and
variables controlling for changes in demographic composition (male adults, female adults, children 05, children
515, all by gender) and sex of the head. All values in 1994 prices, adult equivalent corrections based on
nutritional scales, see Dercon and Krishnan (2000).
In estimating (19), we control for household labour characteristics (adults and
children above 15, and sex of the head) as well as land holdings (and a square), given
25
relatively poorly functioning labour markets and missing land markets.37 Livestock holdings
at t1 are included to capture seasonal (working capital) credit constraints.
To test whether the risk of low consumption outcomes affects fertiliser use, we
include the predicted counterfactual level of consumption, simulated based on column (3),
table 4 and using historical rainfall data from the nearest rainfall station near each survey site.
Three alternatives to represent the presence of downside risk are explored. First, the level of
consumption if the 20th percentile of the rainfall distribution were realised, is used.
Subsequently, the probabilityweighted mean value of the natural logarithm of consumption
across the entire rainfall distribution has been takenwhich is equivalent to using expected
utility given constant relative risk aversion with a coefficient of one. Finally, to focus more
on the downside risk in consumption, the previous measure is truncated at the median level of
rainfall. It is noted again that in calibrating the downside risk on the historical rainfall
distribution, reliance on withinsample variables to supply instruments is avoided, and the
fundamental identification problem in assessing the impact of consumption risk on
production decisions is circumvented.
Since fertiliser use is far from general, limited dependent variable models are
appropriate. But fixed effects in limited dependent variable models based on the normal
distribution (such as the probit and tobit) yield inconsistent estimates, as fixed effects cannot
be treated as incidental parameters without biasing the other model coefficients (as long as N
> T) (Hsiao, 1986). To get around this, we use the conditional fixed effects logit model to
explore adoption of fertiliser (not the amount used) and the Honoré semiparametric fixed
effect tobit estimator (Honoré, 1992) to look at the effect of consumption risk on fertiliser
application rates.38
37All land is stateowned and allocated by the local government for farmers to cultivate.
38This estimator is consistent, even if N is increasing relative to T. We used an adapted version of Honoré's
Pantob programme, using Gauss 6.0. We tried a number of alternative versions. We report the results from a
simple quadratic loss function with zero bandwith, but all models estimated offered very similar estimates for
26
In addition to household fixed effects that allow us to capture all timeinvariant
household characteristics, such as risk preferences, skill and education levels or permanent
income, both approaches also control for time varying village level effects. This allows us to
control for all community wide influences on fertilizer adoption and use, including prices,
availability, general economic trends, increased extension service availability, broad changes
in the delivery systems and general villagewide learning over time. They also force all
identification of variables of interest to be based on within village variation in each period,
making any significant results rather remarkable and robust.
From column (1), table 5, it is clear that the possibility of low consumption ex post
following low rainfall and low harvests reduces a household's likelihood of adopting fertiliser
ex ante (column (1), table 5). This finding is quite striking given that it is identified from
farm households switching in and out of fertilizer, substantially reducing the sample size (to
417 households), and after controlling for all household time invariant and community time
varying effects. Though positively correlated, the effect of livestock, our other variable of
interest as the main measure of liquid wealth, is not statistically significant. Households with
more labor (male and female adults) are more likely to adopt fertiliser. The same holds for
male headed households.39 The villagetime dummies are mostly significant.
The second column shows the estimated coefficients of the determinants of the
intensity of fertiliser use based on the Honoré household fixed effects tobit. As before, there
is a strongly significant positive effect from the counterfactual level of consumption if rains
were to be relatively poor.
the coefficients of the variables of interest.
39We have to be cautious with this effect, as it is identified from households with a change in the head of the
household, mainly involving a change from male to female head, due to the death of the head of the household.
About 11 percent of households experienced a change in the sex of the head of the household, more than 75
27
Table 5: Explaining fertilizer adoption and aplication Rates (log kg per hectare):
(1) (2) (3)
Conditional fixed effects logit Honoré Fixed Effects Standard Tobit Model
model Tobit Model (n=4397) (n=4397)
(n=1540, 417 groups)
Whether using fertiliser (yes=1) Ln fertiliser in kg per ha Ln fertiliser in kg per ha
Coeff. zvalue Coeff. zvalue Coeff. zvalue
Male child <15 0.069 0.59 0.105 2.01 ** 0.037 0.78
Male adults 0.354 3.01 *** 0.071 1.51 0.135 2.96 ***
Female child <15 0.032 0.26 0.118 2.28 ** 0.029 0.59
Female adults 0.269 2.51 ** 0.055 1.25 0.220 4.75 ***
Sex head (male=1) 0.779 2.13 ** 0.443 2.35 ** 0.500 4.05 ***
Ln Livestock/1000 (birr) 0.160 1.45 0.068 1.17 0.693 13.55 ***
at t1
Ln Land per adult 0.114 0.73 0.936 6.44 *** 0.001 0.01
Ln consumption at 20th 0.236 2.02 ** 0.165 2.69 *** 0.431 5.87 ***
percentile for rain
Time (round) dummies, YES YES YES
and interaction between
village and time dummies
included but not reported
Jointsignificance tests 2650.11*** 766.2***
Wald chi2
Notes: Natural logarithm of fertiliser in kg plus 1 per hectare. One is added to allow zero values to be
transformed in logarithms. Land per adult and livestock per adult (in `000) is adjusted by 0.01 to allow
logarithms. * significant at 10%; ** significant at 5%; *** significant at 1%.
The coefficients for livestock, land and downside consumption risk are lower in the
household fixed effects specification than in the standard tobit, presented in column (3) for
comparison Indeed, the level of consumption in `bad years' as well as current livestock
wealth and land holdings (at t1) are likely correlated with fixed characteristics of the
household, such as higher permanent income, that in turn positively affects fertiliser use.
More striking is that the consumption outcome in a bad year remains relevant, even in
the fixed effects regression, while the effect of livestock wealth around the time of the
fertiliser decision disappearsthe household's position at the time of the fertiliser decision
matters largely to extent that it may cause hardship in bad years. This supports our
interpretation that potential expost credit constraints when harvests fail matter in a distinct
manner from the seasonal credit constraint. Furthermore, households with higher landlabor
ratios are found to use less fertilisers (even though their adoption of fertiliser is unaffected).
percent from male to female headed.
28
This is consistent with their comparative advantage, and only emerged when the household's
total (liquid and illiquid) wealth was properly controlled (through the household fixed
effects).
These results are robust to the use of alternative definitions of the downside risk
variable. When the downside risk variable is defined as the expected value of the logarithm
of consumption for rainfall levels up to the median (column (1), table 6), the results are
virtually identical as in table 5, column (2). Very similar results were obtained when using
the expected value of the logarithm of consumption (not truncated), giving a higher weight to
downside consumption risk than upside risk.40
40These and all other robustness checks also hold in the fixed effects conditional logit model (not reported
here).
29
Table 6: Robustness checks
(1) (2) (3) (4) (5)
Honoré Fixed Effects Honoré Fixed Effects Honoré Fixed Effects Conditional fixed Honoré Fixed Effects
Tobit Model (n=4397) Tobit Model Tobit Model effects logit model Tobit Model
(n=4397) (n=4397) (n=1540, 417 groups) (n=3260)
Dependent Variable Ln fertiliser in kg per Ln fertiliser in kg per Ln fertiliser in kg per Whether using Ln fertiliser in kg per
ha ha ha fertiliser (yes=1) ha
coeff zvalue coeff zvalue coeff zvalue Coeff zvalue Coeff zvalue
Male child <15 0.105 2.01 ** 0.087 1.34 0.103 1.97 ** 0.071 0.61 0.064 0.97
Male adults 0.072 1.51 0.041 0.72 0.073 1.54 0.356 3.03 *** 0.039 0.63
Female child <15 0.118 2.28 ** 0.107 1.66 * 0.118 2.28 ** 0.029 0.24 0.113 1.93 *
Female adults 0.055 1.26 0.105 2.01 ** 0.056 1.27 0.267 2.49 ** 0.076 1.32
Sex head (male=1) 0.443 2.35 ** 0.536 2.76 *** 0.446 2.36 ** 0.776 2.12 ** 0.731 3.31 ***
Ln Livestock at t1 0.067 1.16 0.035 0.51 0.113 1.49 0.264 1.06 0.132 1.96 **
Ln Land per adult 0.936 6.44 *** 0.979 6.48 *** 0.936 6.45 *** 0.117 0.75 0.771 4.40 ***
Ln consumption 20th perc 0.166 2.70 *** 0.237 2.03 ** 0.221 2.49 **
Expected ln consumption 0.165 2.69 ***
Below median rainfall
Ln consumption 20th perc 0.478 1.64 *
based on predicted cons
Ln Livestock at t1 0.105 0.88 0.207 0.47
interacted with down
payment fertiliser loan
Ln consumption at t1 0.034 0.35
Time (round) dummies, and YES YES YES YES YES
interaction between village
and time dummies included
but not reported
* significant at 10%; ** significant at 5%; *** significant at 1%.
So far, the counterfactual distribution of consumption for different levels of rainfall
was calculated starting from actual consumption, and then using the estimated coefficients
from table 4. The counterfactual consumption levels in column (2), table 6, start from
predicted values of consumption based on table 4. If actual consumption is measured with
substantial measurement error, then the estimates in column (2) will be superior; however,
predicted consumption ignores any factors causing consumption differences between
households known to the household but not observed by the researcher. The results reported
so far are maintained: insignificant effects from livestock holdings at t1, but counterfactual
consumption if the harvest were to be poor, is still significant. The coefficient on the latter is
higher than in column (2), table 5, consistent with attenuation bias when using our original
estimate for counterfactual consumption.
The next two columns investigate further whether seasonal credit constraints matter at
all, and whether our interpretation is robust. It could be argued that simply using asset
holdings at t1 is a poor proxy of the true cost exante of fertiliser. Even though credit is
available, a down payment is required. As discussed in section 4, a payment of about 30
percent of the purchase price is required, but there is considerable variation across areas. This
variation is exploited to explore the impact of seasonal credit constraints further (columns (3)
and (4)). In particular, even though the village timevarying fixed effects contain this
variation in down payment, it could be expected that in areas with higher downpayment, the
same asset holdings would result in lower fertiliser application rates in other words, an
interaction effect of asset holdings and downpayments should be negative if seasonal
constraints matter. The results do not confirm this: this variable, just as the asset levels at t1,
is insignificant.
Finally, in column (5), a further robustness test is shown, focusing on the
interpretation of our counterfactual consumption term. It could be argued that as consumption
levels move only slowly, and as our prediction model is based on the evolution of actual
consumption, our counterfactual consumption level is likely to be highly correlated with
consumption and welfare levels at t1. In fact, our measure may then be more a reflection of
current conditions, not really capturing a counterfactual outcome. Similarly, it could be
argued that livestock holdings are a poor proxy of current conditions, and our predicted
consumption levels are in fact a better measure of current circumstances. One way of
examining the robustness of our interpretation is by including a stronger measure of living
conditions exante, via consumption levels at t1, before the fertiliser decision, as well as our
counterfactual prediction. Even though we lose a round of data now41, the results regarding
the relevance of counterfactual consumption for fertiliser application are confirmed, while the
consumption level at t1 is simply insignificant. Taken together, these results highlight the
impact of possibly low consumption outcomes expost on fertiliser use, controlling for the
level of assets at the time of fertiliser purchase, and household fixed characteristics and time
varying village level conditions.
41Contrary to livestock levels, which were easily collected at t1 via a set of simple recall questions, this is not
possible for consumption.
32
6 Conclusions
This paper has investigated the impact of the risk of poor consumption outcomes on
the adoption and use of fertiliser in Ethiopia. Fertiliser results in higher yields and substantial
returns on average. However, as a costly input, when harvests are poor, for example due to
poor weather conditions, returns tend to be low given the sunk cost of fertiliser, making it a
high risk activity with moderately higher returns compared to not using fertiliser. We
developed a simple framework to assess whether the possibility of poor consumption
outcomes affects modern input use, controlling for a simple asset effect, that could also be a
reflection of seasonal working capital constraints.
Using data from Ethiopia, we find evidence that after controlling for these seasonal
working capital constraints, as well as household fixed effects (including factors such as risk
preferences and permanent income) and timevarying community fixed effects (including
factors such as input output price ratios and extension programs) fertiliser application rates
are significantly lower due to downside risk in consumption. Consequently, measures to
remove the downside risk of agricultural innovations via insurance systems would have
beneficial impacts on stimulating their spread.
The presence of a link between downside consumption risk and modern input
adoption also suggests that risk is a cause of perpetuating poverty: those (poorer) households
unable to protect themselves against downside risk are forced to avoid some downside risk by
reducing their use of profitable modern inputs. As such, risk induces the persistence of
poverty for some, as if trapped in low return, lower risk agriculture.
33
References
Abrar, S., O., Morrissey, and T., Rayner, 2004, Crop Level Supply Response by Agro
Climatic Region in Ethiopia, Journal of Agricultural Economics, 552: pp. 289311.
Asfaw, A., and A., Admassie, 2004, The Role of Education on the Adoption of Chemical
Fertiliser under Different SocioEconomic Environments in Ethiopia, Agricultural
Economics, 30: pp. 215228.
Bonger, T., G., Ayele, and T., Kuma, 2004, Agricultural Extension, Adoption and Diffusion
in Ethiopia, Ethiopian Development Research Institute, Research Report 1, Addis Ababa:
Ethiopia.
Carlsson, F., G., Kohlin, A., Mekonnen and M., Yesuf, 2005, Are Agricultural Extension
Packages What Ethiopian Farmers Want? A Stated Preference Analysis, Working Papers in
Economics, No. 172, Department of Economics, Goteborg University.
Christiaensen, L., and K., Subbarao, 2005, Towards an Understanding of Household
Vulnerability in Rural Kenya, Journal of African Economies, 144: pp. 520558.
Christiaensen, L., and L., Demery, 2007, Down to Earth: Agriculture and Poverty Reduction
in Africa, Directions in Development, World Bank: Washington D.C.
Croppenstedt, A., Demeke, M., and M., Meschi, 2003, Technology Adoption in the Presence
of Constraints: The Case of Fertiliser Demand in Ethiopia, Review of Development
Economics, 71: pp. 5870.
Dadi, L., M., Burton, and A., Ozanne, 2004, Duration Analysis of Technology Adoption in
Ethiopian Agriculture, Journal of Agricultural Economics, 553: pp. 613631.
Datt, G., and H., Hoogeveen, 2003, El Nino or El Peso? Crisis, Poverty and Income
Distribution in the Philippines, World Development, 317: pp. 11031124.
Deaton, A., 1992, Understanding Consumption, Clarendon Press, Oxford.
Dercon, S., 1996, Risk, Crop Choice, and Savings: Evidence from Tanzania, Economic
Development and Cultural Change, 443: pp. 485513.
Dercon, S., 2002, "Income Risk, Coping Strategies and Safety Nets", (2002), World Bank
Research Observer, December, Vol. 17 (2), pp. 14166.
Dercon, S. And P.Krishnan, 2000, "Vulnerability, seasonality and poverty in Ethiopia",
Journal of Development Studies, 366, 2pp. 2553.
Dercon, S., 2004, Growth and Shocks: Evidence from Rural Ethiopia, Journal of
Development Economics, 742: 309329.
34
Dercon, S., J.,Hoddinott, and T.,Woldehanna, 2005, "Shocks and consumption in 15
Ethiopian Villages, 19992004", Special Issue on Risk, Poverty and Vulnerability in Africa,
Journal of African Economies, 144, pp. 559585.
Duflo, E., M.Kremer and J. Robinson, (2006), "Why Don't Farmers Use Fertiliser? Evidence
from Field Experiments in Western Kenya" MIT and Harvard, mimeo.
Eswaran, M. and A. Kotwal, 1990, "Implications of Credit Constraints for Risk Behaviour in
Less Developed Economies", Oxford Economic Papers, 422: pp. 47382.
Evans, D.S. and B. Jovanivic (1989), "An Estimated Model of Enterpreneurial Choice under
Liquidity Constraints", Journal of Political Economy, 974: pp. 80827.
Feder, G., R. Just and D. Zilberman (1985), "Adoption of agricultural innovations in
developing countries: a survey", Economic Development and Cultural Change, 332: pp.
255298.
Just, R., and R., Pope, 1979, Production Function Estimation and Related Risk
Considerations, American Journal of Agricultural Economics, 612: pp. 27684.
Hiebert, L. Dean, 1974, Risk, Learning, and the Adoption of Fertilizer Responsive Seed
Varieties, American Journal of Agricultural Economics, 564: pp. 76468.
Honoré, B.E., 1992, Trimmed Lad and Least Squares Estimation of Truncated and Censored
Regression Models with Fixed Effects, Econometrica, 603:, pp. 533565.
Hsiao, C., 1986, Analysis of Panel Data, Econometric Society monographs, Cambridy
University Press.
Kazianga, Harounan, and Christopher, Udry, 2006, "Consumption Smoothing? Livestock,
Insurance and Drought in Rural Burkina Faso", Journal of Development Economics, 792:
pp. 41346.
Kurosaki, T., and M., Fafchamps, 2002, Insurance Market Efficiency and Crop Choices in
Pakistan, Journal of Development Economics, 672: pp. 49453.
Lamb, R., 2003, Fertiliser Use, Risk and OffFarm Labor Markets in the SemiArid Tropics
of India, American Journal of Agricultural Economics, 852: pp. 359371.
Morduch, J., 1990, "Risk, Production and Savings", Harvard University, mimeo.
Morduch, J., 1995, Income Smoothing and Consumption Smoothing, Journal of Economic
Perspectives, 93: pp. 103114.
Morris, M., et al., 2007, Fertiliser Use in African Agriculture Lessons Learned and Good
Practice Guidelines, Directions in Development, Agriculture and Rural Development, World
Bank: Washington D.C..
35
Narayan, Deepa with Raj Patel, Kai Schafft, Anne Rademacher and Sarah KochSchulte,
2000. Voices of the Poor: Can Anyone Hear Us? New York, N.Y.: Published for the World
Bank, Oxford University Press.
Rosenzweig, M., and H., Binswanger, 1993, Wealth, Weather Risk and the Composition and
Profitability of Agricultural Investments, Economic Journal, 103416: pp. 5678.
Sandmo, A., 1971, "On the Theory of the Competitive Firm under Price Uncertainty",
American Economic Review, 61: pp. 6573.
Townsend, R.M. 1995, "Consumption Insurance: An Evaluation of RiskBearing Systems in
LowIncome Economies." 93): pp. 83102.
World Bank, 2006a, Well Being and Poverty in Ethiopia: The Role of Agriculture and
Agency, Report No. 29468ET, World Bank: Washington D.C.
World Bank, 2006b, Ethiopia  Policies for ProPoor Agricultural Growth, Africa Region,
World Bank: Washington D.C., mimeographed.
World Bank, 2007, World Development Indicators 2006, World Bank: Washington D.C..
Zimmerman, F.,J., and M.R., Carter, 2003, Asset Smoothing, Consumption Smoothing, and
the Reproduction of Inequality Under Risk and Subsistence Constraints, Journal of
Development Economics, 712: pp. 233260.
36
Appendix 1: The distribution of cereal yields and returns
Table A.1 explores the determinants of output per hectare using plotlevel data from 1999.
Yields are regressed on plot level inputs including the plot size in hectares, hours worked by
male and female adults, and children (up to the age of 15 years), and purchased inputs, such
as fertiliser, herbicides, fungicides, insecticides, and other pest controls (in kg). All variables
are expressed in natural logarithms (plus 0.01 to allow for zero values to be defined when
taking logarithms). Table A.1 adds controls for the different crops (whereby white teff is the
basegroup), whether the plot is intercropped, local standard descriptions for land quality (lem,
lemteuf, teuf, with lem the best and teuf the base group) and the slope of the plot (flat,
sloping and `geddel', the base group, meaning a steep slope). The regression also controls at
the household level for the number of livestock in tropical livestock units (as a proxy for both
wealth and for access to oxen for ploughing).
Plotspecific, selfreported `shocks' related to flooding, weed damage, animals damaging
crops, etc. are introduced as dummies with one suggesting a `bad' event. A number of
variables describing the rainfall distribution and experience in this particular year are further
introduced. First, we have access to historical rainfall data in most locations going back for
about 20 years, based on rainfall in the nearest station. We use village level variables to
describe whether the rainfall in the 1999 season was in a particular quintile of the historical
distribution (from 019, 2039, etc.) To reflect differences in `normal' conditions, we control
for the median level of rainfall in each area. We also use a selfreported index of how good
the rain distribution was in this particular year, based on plotlevel data from the 1999 survey:
was the rain on time, did it rain too long, into the harvest period, was there enough at critical
points, etc. The answers to these questions were added into a simple score, normalised to one
if the rain was as good as it could get, and zero if it was bad on all counts. The squared value
of this index was used.
Finally, market access was proxied by a variable for `road access and quality' based on
community level surveys. It is an index on a scale of six, with the value one a road with
access for all possible vehicles in all seasons, and six only a track not accessible in the rainy
season. The value of three is for example a road accessible to trucks most of the time.
Application of OLS shows that here are clear signs of lower yields on larger plots, with
returns to male labour, fertiliser and insecticide strongly significant (column (1), table A1).
Livestock and good land quality appears to matter as well. Most of the idiosyncratic shocks
(with the exception of frost damage, a rare event in the data) have the expected signs and
several cause substantial and significant yield losses. Areas with good roads appear to have
higher yields on average, consistent with increased incentives for producing crops. The effect
is large, though dropping this variable did not affect the signs or size of the rainfall variables.
Unsurprisingly, rainfall matters significantly, with median levels increasing yields (but only
marginally in magnitude). The realised quintile of the rainfall in 1999 based on the historical
distribution also displays the expected signs and magnitude. As there were no observations in
the 1999 data of the highest quintile, the lowest quintile is the basegroup, and all other three
quintiles showed significantly higher yields (by at least 20 percent). Having rainfall around
the 40 to 60 percentile compared to historical distribution (ie around the median) had the
highest yield gain. Finally, the selfreported `quality' of rainfall index showed higher yields
with a better index but at a decreasing rate.
Table A.1 PlotLevel Production Function (ln output in kg per hectare) 1999
37
(1) (2) (3)
Ln yield Ln yield Ln yield
Ln Plot size (in hectares) 0.618 0.582 0.616
[25.42]*** [18.11]*** [25.22]***
Ln Hours Male Labour 0.267 0.409 0.274
[11.40]*** [11.31]*** [11.55]***
Ln Hours Female Labour 0.014 0.065 0.017
[0.96] [2.51]** [1.12]
Ln Hours Child Labour (<15 years) 0.003 0.056 0.003
[0.25] [1.84]* [0.23]
Ln Fertiliser in kg 0.034 0.028 0.003
[8.34]*** [5.39]*** [0.10]
Ln Herbicide in kg 0.205 0.261 0.175
[1.80]* [1.46] [1.52]
Ln Fungicide in kg 0.015 0.913 0.011
[0.05] [1.16] [0.04]
Ln Insecticide in kg 0.144 0.191 0.138
[2.06]** [2.00]** [1.98]**
Ln Other Pest Control in kg 0.224 0.18 0.213
[0.85] [0.36] [0.81]
Ln Livestock (livestock units) 0.072 0.072
[5.33]*** [5.37]***
Is the plot intercropped? 0.12 0.072 0.119
[2.41]** [1.07] [2.41]**
Black Teff 0.143 0.114 0.162
[2.61]*** [2.14]** [2.96]***
Barley 0.33 0.565 0.345
[6.35]*** [10.36]*** [6.59]***
Wheat 0.267 0.407 0.272
[5.53]*** [8.45]*** [5.60]***
Maize 0.461 0.581 0.474
[9.69]*** [12.35]*** [9.97]***
Sorghum 0.416 0.434 0.409
[6.73]*** [6.83]*** [6.60]***
Land quality: very good (lem) 0.101 0.192 0.095
[2.24]** [3.38]*** [2.11]**
Land quality: OK (lemteuf) 0.037 0.077 0.037
[0.83] [1.34] [0.81]
Land Slope: flat (Medda) 0.011 0.089 0.017
[0.11] [0.60] [0.16]
Land Slope: sloping (Dagathama) 0.037 0.171 0.036
[0.35] [1.10] [0.34]
Wind damage? 0.07 0.017 0.074
[1.42] [0.14] [1.51]
Hail damage? 0.029 0.139 0.05
[0.46] [1.06] [0.79]
Frost damage? 0.145 0.047 0.152
[3.43]*** [0.56] [3.60]***
Flood damage? 0.151 0.139 0.146
[3.09]*** [1.58] [2.99]***
Plant disease damage? 0.094 0.182 0.108
[1.02] [1.05] [1.17]
Insect damage? 0.066 0.08 0.058
38
(1) (2) (3)
Ln yield Ln yield Ln yield
[0.98] [0.89] [0.87]
Weed damage? 0.218 0.145 0.225
[2.79]*** [1.35] [2.88]***
Bird damage? 0.207 0.209 0.21
[2.13]** [1.44] [2.17]**
Wild animals damage? 0.049 0.066 0.052
[0.76] [0.62] [0.81]
Livestock trampling damage? 0.025 0.069 0.012
[0.18] [0.32] [0.09]
Road quality index 0.081 0.081
(1=best, 6=worst) [7.94]*** [7.91]***
Median rainfall in area 0.001 0.001
[2.52]** [2.72]***
Is rainfall 2040 percentile? 0.284 0.472
[2.37]** [2.93]***
Is rainfall 4060 percentile? 0.38 0.475
[3.00]*** [2.76]***
Is rainfall 6080 percentile? 0.218 0.403
[1.87]* [2.54]**
Rainfall distribution index 0.752 0.749
(1=best, 0=worst) [4.18]*** [4.16]***
Rainfall distribution index squared 0.501 0.503
[2.75]*** [2.77]***
Rainfall index*ln fertiliser 0.042
[3.33]***
2040% * ln fertiliser 0.049
[1.49]
4060% * ln fertiliser 0.084
[2.30]**
6080% * ln fertiliser 0.057
[1.74]*
Constant 4.04 4.005 3.816
[20.11]*** [19.67]*** [16.69]***
Observations 2502 2502 2502
Rsquared 0.4 0.31 0.41
Absolute value of t statistics in brackets
* significant at 10%; ** significant at 5%; *** significant at 1%
Column (2) looks at the robustness of the results by showing the household fixed effects
regression results. By definition, all the household and community level variables are
absorbed in the fixed effects, so that no results on rainfall can be reported. Importantly, the
coefficient on fertiliser retains its significance, and is similar in value. While there are some
signs of correlations of some of the plotlevel variables with the household fixed effect (i.e.
bias in column (1)), it may not affect our inference on the impact of fertiliser on yields. This
is a striking result, as identification of the impact has to be done across the plots of each
household, who usually farm about 23 plots each. As there is relatively little variation in
terms of most of the idiosyncratic shocks, it should not come as a surprise that the impacts of
most shocks are measured with considerable error.
39
Finally, to test whether the returns to fertiliser are rainfall dependent, interaction terms of the
level of fertiliser use and the various rainfall variables are introduced in column (3). These
interactions are generally significant. However, the impact of fertiliser without interaction is
now no longer different from zero. This term is now the basegroup for the interaction terms
giving the impact of fertiliser when the rains are really poor (the base group is the 0 to 20th
percentile of the rainfall distribution). Or, during drought fertiliser has no impact.
Simulations are undertaken to obtain clearer inference on the impact of fertiliser on yields
across the rainfall distribution.
For the villagelevel rainfall distribution, the historical data are used to show the distribution
of yields for different levels of rainfall, by comparing otherwise similar plots with `mean'
characteristics in terms of fertiliser use and nonuse. For the selfreported rainfall distribution
index we rely on the distribution implied across all four rounds of the householdlevel self
reported shocks. The frequency distribution of all possible outcomes between 19941999 is
used in the simulations. As was shown in Dercon and Krishnan (2000), this index is largely
(although not perfectly) covariate by village; the simulations assume perfect covariance,
although some sensitivity analysis showed that less than perfect covariance hardly affected
the results reported in the main text.
The equations above use pooled data for all crops. We experimented with specifications by
crop, but the sample sizes become much smaller, especially to identify the impact of rainfall
across the distribution as most villages do not grow all different possible crops in this year.
Broadly speaking, the coefficients are comparable though estimated with less accuracy.
Restricting the cropspecific effects to multiplicative shifters of yields appears to capture the
diversity of yields reasonably well in the data.
The regressions are used to construct counterfactual yield distribution for different levels of
rainfall, based on the historical rainfall distribution, and assuming here that the household
specific quality of rainfall index is covariate with village rainfall. As the results are in size
dominated by the village rainfall index, alternative assumptions on the quality of rainfall did
not make much difference. The results suggest considerably higher yields for fertilised plots
compared to nonfertilised plots (Table A2). Across cereals, median rainfall conditions offer
about 24 percent higher yields when using the average amount of fertiliser. While yields on
fertilised plots dominate those on nonfertilised plots across the rainfall distribution, the
relative benefit systematically declines the further rainfall levels are from the median.
Table A.3 shows calculations of the implied returns per hectare for these different yield
levels. These returns are simply defined as the mean gross return (yields times output prices)
minus (for the simulations for fertilised plots) the cost of the fertiliser, i.e. mean application
rates times the fertiliser price, assuming a cash purchase. To get a sense of the distribution of
returns by fertilised and nonfertilised plots during the survey period, we use the mean output
and fertiliser price for the period 199499. In the main text, we offer in figure 1 the
distribution for all cereals using average 199495 prices and compare it to 199799 prices to
reflect changing profitability.
For all crops, using fertiliser is profitable as long as the rainfall is not deviating too much
from the usual patterns of rainfall. With median rainfall, the returns on fertilised plots are
about 10 percent higher than for nonfertilised plots. Given that using fertiliser offers little
yield gain at low rainfall levels, nonfertilised plots have higher returns per hectare in these
40
circumstances. The data suggest that with abundant rain, returns are also better for non
fertilised crops.
Table A.2 Fertiliser use and the distribution of yields
Yields in kg per ha All Teff Barley Wheat Maize Sorghum
across the rainfall cereals
distribution
No fert No fert No fert No fert No Fert No fert
fert fert fert fert fert fert
10th percentile 440 443 319 321 394 397 424 427 574 578 584 588
20th percentile 568 618 412 450 508 555 547 597 741 803 754 806
30th percentile 790 912 573 668 707 822 761 885 1030 1184 1049 1175
40th percentile 850 1018 616 748 761 920 819 990 1108 1320 1129 1301
50th percentile 889 1106 644 815 796 1002 856 1079 1159 1434 1180 1403
60th percentile 879 1027 637 753 787 927 846 998 1146 1334 1167 1320
70th percentile 853 939 618 686 764 845 821 909 1112 1221 1133 1222
80th percentile 847 917 614 669 758 824 815 887 1104 1193 1124 1197
90th percentile 838 899 607 655 750 808 807 869 1092 1170 1112 1176
Note: Simulations based on the estimates reported in table A.1, column (3). Yields based on estimated output
per hectare in 1999 for an average plot (i.e. with mean characteristics for plot, farmer and village), among 2294
plots. Yields on fertilised plots based on approximately the mean application rate for fertiliser users for each
crop. The table offers the counterfactual yield distribution for different cereals across the rainfall distribution.
For example, the 10th percentile gives yields when the rainfall was equivalent to the 10th percentile of the
rainfall distribution, i.e. very poor rains, the 50th percentile when rains were to be at the median level of rainfall
in all villages, and the 90th percentile is very abundant rainfall, i.e. rain at the 90th percentile of the historical
distribution. The cropspecific yield estimates are based on calculating counterfactual yields only for those plots
currently growing the crop.
Table A.3 Fertiliser use and the distribution of returns 199499
Returns per All cereals Teff Barley Wheat Maize Sorghum
hectare across the
rainfall
distribution
No fert No fert No fert No fert No Fert No fert
fert fert fert fert fert fert
10th percentile 793 544 688 478 772 535 867 622 742 459 1180 861
20th percentile 1023 858 888 757 997 846 1119 970 957 751 1523 1301
30th percentile 1423 1388 1235 1226 1387 1369 1556 1559 1332 1243 2119 2046
40th percentile 1531 1579 1329 1398 1492 1560 1675 1773 1433 1419 2280 2301
50th percentile 1601 1740 1389 1543 1560 1721 1751 1955 1499 1567 2384 2507
60th percentile 1582 1596 1373 1410 1542 1575 1731 1790 1481 1437 2356 2339
70th percentile 1536 1437 1333 1264 1497 1413 1680 1608 1438 1291 2287 2141
80th percentile 1525 1397 1323 1227 1486 1373 1668 1563 1427 1255 2270 2092
90th percentile 1509 1365 1309 1198 1470 1341 1650 1527 1412 1225 2246 2050
Note: Simulations based on the estimates reported in table A.1, column (3).. Returns are the gross returns (yield
times output price, evaluated at the mean output price in 199499) minus the cost of the fertiliser (using the
mean fertiliser price at the time of planting in 199499).
41