poverty traps and natural disasters in ethiopia and honduras

World Development Vol. 35, No. 5, pp. 835–856, 2007� 2007 Elsevier Ltd. All rights reserved

0305-750X/$ - see front matter

doi:10.1016/j.worlddev.2006.09.010www.elsevier.com/locate/worlddev

Poverty Traps and Natural Disasters in Ethiopia

and Honduras

MICHAEL R. CARTERUniversity of Wisconsin, Madison, WI, USA

PETER D. LITTLEUniversity of Kentucky, Lexington, KY, USA

TEWODAJ MOGUESInternational Food Policy Research Institute, Washington, DC, USA

and

WORKNEH NEGATU *

Addis Ababa University, Ethiopia

Summary. — Droughts, hurricanes, and other environmental shocks punctuate the lives of poorand vulnerable populations in many parts of the world. The direct impacts can be horrific, but whatare the longer-term effects of such shocks on households and their livelihoods? Under what circum-stances will shocks push households into poverty traps from which recovery may not be possiblewithout external assistance? In an effort to answer these questions, this paper analyzes the assetdynamics of Ethiopian and Honduran households in the wake of severe environmental shocks.While the patterns are different across countries, both reveal worlds in which the poorest house-holds struggle most with shocks, adopting coping strategies which are costly in terms of both shortterm and long-term well being.
� 2007 Elsevier Ltd. All rights reserved.
Key words — poverty traps, natural disasters, Honduras, Ethiopia

* This publication was made possible by support pro-

vided in part by the World Bank and by the US Agency

for International Development (USAID) Agreement

No. LAG-A-00-96-90016-00 through BASIS Collabora-

tive Research Support Program. Rob White and Fran-

cisco Galarza provided valuable research assistance.

Helpful comments were received from anonymous

referees of this journal, as well as by seminar partici-

pants at the 2005 ASSA meetings, the World Bank, the

University of Manchester’s Social Protection Confer-

ence, and the University of Wisconsin. All views, inter-

pretations, recommendations, and conclusions expressed

in this paper are those of the authors and not necessarily

those of the supporting or collaborating institutions.Final revision accepted: September 11, 2006.

1. INTRODUCTION

Ato Mohammed, 55 and illiterate, resides in the Batidistrict of South Wollo Zone (Ethiopia) and heads ahousehold of nine. He has been chronically foodinsecure for more than 10 years when he lost his onlyoxen due to drought. He sold the animal to buy foodat the time and has not been able to acquire another.Currently, Mohammed holds one hectare of farmland and he has no grazing land. Since he owns nooxen, he has been leasing out the land for share-crop-ping on a 50/50 sharing arrangement. Mohammedand his family members are engaged in various typesof daily labor activities for cash and food, and thehousehold is a regular recipient of food aid.

Mohammed asserts ‘‘oxen are the crucial productiveasset that would liberate me from this insecurity

835

836 WORLD DEVELOPMENT

trap.’’ On the other hand, however, he does not wantto take credit from a regional credit organization tobuy an ox as he does not want to be indebted andfears that the debt may be passed on to his childrenif he fails to repay. He fears that the ox may die dueto lack of adequate feed or animal diseases for whichthere is no dependable animal health service in thecommunity. He also fears that he may not be ableto pay back since crop failure is frequent due to in-sects and droughts.

The direct impacts of the droughts, hurricanes,and other environmental shocks can be horrific,resulting in immediate increases in poverty anddeprivation. But what are the longer-term effectsof shocks on households and their livelihoods?Are households able to quickly reestablish theirlivelihoods and the assets needed to supportthem, or is recovery a slow, drawn-out process,especially for poorer households who may beless able to leverage the resources needed torebuild? Indeed, is there a ‘‘poverty trap’’ fromwhich households can rarely recover, as AtoMohammed’s 1 story suggests? And, if there issuch a trap—understood as a minimum assetthreshold (e.g., one ox), below which accumula-tion and livelihood growth are not feasible—doforward-looking households adopt asset protec-tion strategies designed to avoid the trap butwhich come at the very high cost of immediatelyreduced consumption, with perhaps irreversiblelosses in child health and education? Finally, towhat extent does the existence of deep marketsand, or social networks offset these longer-termconsequences of disaster?

To explore these issues, this paper examinesdata from two macabre, naturally occurringexperiments. The first is Hurricane Mitch whichstruck Honduras and other parts of CentralAmerica in 1998. Through the vagaries ofMitch, some households lost nearly all their pro-ductive assets, while others were left unscathed.These changes in the asset distribution 2 permitus to explore questions of resilience and thespeed of longer-term recovery, and to exam-ine whether there is any evidence of povertytraps.

The prolonged Ethiopian drought of 1998–2000 presents a second kind of disaster experi-ment. Direct destruction of assets was modest,but the income losses of repeated crop failuresin some locations forced households to choosebetween preserving assets, or selling them tomaintain current consumption and health.Examination of household asset holdingsacross the drought cycle provides insight intothe longer-term effects of droughts and into

particular wealth-differentiated asset manage-ment strategies.

This work is part of a comparative projectthat addresses the interrelationships betweenclimatic shocks, markets, and asset recoverystrategies among households in developingcountries (see Little, Ahmed, Carter, & Negatu,2002). In Ethiopia, markets are relatively weak(especially for land, labor, and capital), andnon-market mechanisms are important. Factormarkets are better developed in Honduras, butits inegalitarian agrarian structure may limitthe effectiveness and extent of the social assetsthat may aid recovery in Ethiopia. 3 Data ona sample of 416 rural Ethiopian householdstrack household assets over a seven-year periodof pre-drought (1996–97), drought (1998–2000), and recovery (2001–03). Data on a sam-ple of 850 rural Honduran households capturethe immediate impact of Hurricane Mitch in1998 on assets and income, as well as thesehouseholds’ economic position in 2001, twoand half years after Mitch.

The remainder of this paper is organized asfollows. Section 2 proposes an anatomy of anenvironmental shock, tracing the evolution ofassets through time in the face of a shock,and presents an empirical model of asset accu-mulation to investigate households’ sensitivityto, and resilience from, shocks. The factors thatinfluence rural Honduran households’ exposureto and recovery from the 1998 hurricane areexamined in Section 3. Section 4 describes theincome losses households in northeastern Ethi-opia suffered due to the droughts of 1999–2000and estimates the determinants of long-termasset recovery in the wake of the droughts. Con-cluding remarks are offered in the final section.

2. SHOCKS, POVERTY TRAPS, ANDRESILIENCE

This section puts forward a conceptualframework for thinking about the longer-termeconomic impacts of environmental shocks. Intheir analysis of economic shocks and vulnera-bility, Calvo and Dercon (2005, p. 3) make theimportant observation that economic vulnera-bility is a ‘‘sense of insecurity, of potential harmpeople must feel wary of—something bad canhappen and ‘spell ruin’’’ (emphasis added). Thenotion that a random event (a flood, a drought,an illness, an unemployment spell) can havepermanent effects, spelling ruin for a family,

POVERTY TRAPS AND NATURAL DISASTERS IN ETHIOPIA AND HONDURAS 837

suggests that vulnerability (and perhaps pov-erty) can best be understood through the lensof poverty traps as described by the quote atthe beginning of this paper. Here we followCarter and Barrett (2006) and understand apoverty trap as a critical minimum asset thresh-old, below which families are unable to success-fully educate their children, build up theirproductive assets, and move ahead economi-cally over time. Below the threshold lie thosewho are ruined, who can do no better thanhang on and who are offered no viable pros-pects for economic advance over time. Thoseabove the threshold can be expected to produc-tively invest, accumulate, and advance.

The bifurcation of behavior reflects the bifur-cation of longer-term prospects and thus of eco-nomic wellbeing. In the presence of povertytraps, temporary shocks have permanent adverseconsequences for those knocked beneath thethreshold, while others—those who do not crossbetween regimes—would be expected to recoverfully and quickly from an objectively similarshock (Dercon, 2004). Anticipation of povertytraps would also be expected to shift copingbehavior, inducing those near the trap to protector smooth their assets lest they fall below the crit-ical minimum asset level and face ruin. In theremainder of this section, we explore furtherthe empirical implication of poverty traps.

(a) Asset shocks, sensitivity, and poverty traps

Figure 1 presents a stylized anatomy of a nat-ural disaster that destroys assets and reduces in-

Ass

ets

Tim

Inco

me

shoc

k

Shock Coping

Abp

Awp

Asp

Awp

ε

θ p

Figure 1. Asset shock

comes. The x-axis measures time and the y-axismeasures asset stocks and income shocks.Hypothetical asset trajectories are illustratedfor an initially wealthier household, w (whichbegins with an initial asset level Abw), and aninitially poorer household, p (with pre-shockassets Abp). The trajectories represented by thedashed lines illustrate the case of convergent as-set dynamics in which the poorer householdwould eventually catch up to the wealthierhousehold if it experienced no shocks andstayed on the dashed asset accumulation path.The shock itself is displayed as a short durationevent (e.g., a hurricane) that kills livestock, andwashes away land and plantations, reducing as-sets by Hw(Hp) for the wealthy (poor) house-hold. In the wake of the shock, the rich andpoor households are respectively left with assetstocks Asw and Asp. The shock may also reducecurrent income by an amount e.

We are now in a position to consider theforces that shape the longer-run implicationsof an asset shock such as that illustrated in Fig-ure 1. Two broad sets of forces will shape theasset level we would be able to observe at somefuture time when assets have recovered to theirlonger-term trajectory. The first set concernscoping strategies utilized in the immediateaftermath of the shock. The second set con-cerns the longer-run patterns of asset dynamics,namely the desired and achievable level ofassets and the speed with which a householdcan approach this longer-term level.

Households’ reactions to the direct incomeand asset losses during the coping phase are

Poverty Trap Threshold, A

e

Recovery

Deviation from Expected Income

Arp

s and poverty traps.


structured by the markets and other institutionsto which they have access. 4 Households withfinancial market access might borrow againstfuture earnings to sustain their consumptionstandard without further asset depletion. Infor-mal finance and insurance arrangements canplay the same role, as can receipt of disasterassistance. Another coping strategy is to redi-rect or increase work time (reduce leisure).The effectiveness of this strategy will dependon access to and depth of labor markets.

Households without access to these marketsmay sustain their consumption by furtherdrawing down on their assets (householdsreluctant to draw down on their assets may alsocope by reducing consumption, as the next sec-tion will discuss). While asset sales will help tosmooth household consumption over time, itimplies that assets will exhibit excess sensitivityto shocks (declining by more than the direct as-set shock, H). Note that the severity of this sec-ondary asset decline will also be shaped bychanges in the prices of assets relative to theprice of food and other necessities. Unfavor-able asset price swings (as would be expectedto happen if all households in an area respondto the shock by selling cattle) would threatenfurther decapitalize households in the wake ofa shock.

The market and social mechanisms that bro-ker access to employment and financial serviceswill also shape a household’s resilience and thespeed of its post-shock asset accumulationtrajectory. A household with good access tocapital (via markets, or via informal socialarrangements) can borrow against future earn-ings to immediately rebuild asset stocks. Such ahousehold might be expected to recoverquickly, perhaps returning to the sort of con-vergent, long-term trajectory illustrated inFigure 1. A household without this accessmay face a doubly slow recovery process.

While there are thus multiple reasons whyless well-positioned families may exhibit excesssensitivity to shocks and recover slowly fromthem, the story of Ato Mohammed quotedabove suggests even deeper effects than thosethat would be signaled by slow recovery. Hisstory suggests the existence of a poverty trap,understood as a minimum asset threshold be-low which it is not possible to engineer success-ful asset accumulation, barring positive shocksor interventions that would lift the household’swealth level above the threshold. The dashedhorizontal line in Figure 1, drawn at a povertytrap threshold of A, demonstrates the idea of

such a threshold. As illustrated in that figure,a household falling below that threshold wouldbe unable to accumulate assets and would beobserved over time to follow a path from Asp

to Arp, rather than rejoining its convergentpre-shock trajectory. 5

While households pushed below a criticalthreshold and into a poverty trap would be ex-pected to exhibit slower post-shock growth,they would also be expected to permanently re-main at a lower level than their more fortunateneighbors. Put differently, asset accumulationtrajectories should bifurcate above and belowa critical asset threshold if a poverty trap exists.Later sections of this paper will test for the exis-tence of such a threshold and bifurcated assetgrowth patterns.

(b) Income shocks, asset smoothing andpoverty traps

The sort of poverty trap described by AtoMohammed has a second testable implication.Figure 2 illustrates hypothetical asset trajecto-ries for households buffeted by income shocksthat play out over an extended time (e.g., a se-quence of drought seasons). For simplicity’ssake, we will assume that there are no directasset losses associated with these incomeshocks. An initially wealthier household thatbegins with assets Abw might be expected todraw down assets in the face of income shocksin order to protect (or smooth) its consumptionlevel. Following the drought period, the house-hold might be expected to rebuild its assetstocks, returning toward its initial trajectory.

But will an initially poorer household, whichbegins with assets Abp, pursue the same con-sumption smoothing strategy? If there is indeeda poverty trap, denoted A in Figure 2, then thehousehold might instead choose to reduce con-sumption and protect (or smooth) its assets inorder to avoid the fate of Ato Mohammed.While this strategy may be a last resort forhouseholds that lack other options, it may alsobe pursued by households reluctant to increasetheir future vulnerability by further depletingtheir stock of assets. 6

While the economics literature sometimesdiscusses consumption smoothing as if it wereitself a primary behavioral objective, a fewauthors have noted the existence of constrainedcircumstances that may lead individuals toasset smoothing, despite the fact that assetsmoothing means that consumption becomesunstable and dips to painfully low levels. Dreze

Poverty Trap Threshold

Ass

ets

Time

Inco

me

shoc

k

Drought Recovery

Abw

Abp

Deviation from Expected Income

Figure 2. Income shocks and asset smoothing.


and Sen (1989), for example, note that verypoor people are observed to ‘‘asset-smooth’’in a way that necessarily destabilizes their con-sumption. At a theoretical level, Zimmermanand Carter (2003) show that optimal inter-tem-poral consumption and saving can result in as-set smoothing behavior. Analyzing data from adrought in Zimbabwe, Hoddinott (2006) findsthat poorer households (defined as those withless than two oxen) tend to asset smooth inthe face of drought-induced income losses,while wealthier households above that thresh-old sell assets and smooth consumption. 7 Test-ing for asymmetric responses to randomincome shocks thus provides a second way totest for the existence of poverty traps. 8 In addi-tion, as in the case of asset shocks, post-shocktrajectories can be examined for direct evidenceof poverty traps in the case of individuals whofall below the critical asset threshold.

(c) Empirical strategy

This section puts forward an econometric ap-proach for exploring the longer-run economicimpacts of environmental shocks. With smallmodifications, this approach will be used toinvestigate the impact of environmental shocksin Ethiopia and Honduras. The data availablefor both countries includes measures of pre-shock (Abi), post-shock (Asi), and post-recoveryassets stocks (Ari), measures of shocks received,and indicators of market access and social net-works.

Consider first the following adaptation of astandard, single equilibrium growth model:

gbi ¼ AbibA þHibHðAbi; Li;KiÞþ eibeðAbi; Li;KiÞ þ bzZi þ ti; ð1Þ

where gbi is the asset growth for household iover the time stretching from the pre-shock per-iod to the recovery period (several years afterthe shock). The household’s initial asset level,Abi, is included in the growth regression to cap-ture the idea—common to neoclassical growththeory—that there is a single equilibrium assetlevel toward which households grow. 9 An esti-mate of bA < 0 would signal a convergent accu-mulation process with lower-wealth householdsgrowing rapidly toward the equilibrium level,while the asset growth of wealthier householdswould slow down and approach zero as theequilibrium level is reached. The point at whichthe growth rate equals zero would be a long-term equilibrium or steady state asset position.The hypothetical no-shock trajectories in Fig-ure 1 illustrate such a convergent process.

The variables Hi and ei are, respectively, mea-sures of asset and income shocks. Reflecting thepreceding conceptual discussion, the coeffi-cients that capture the impacts of asset (bH)and income (be) shocks on growth are writtenconditional on initial asset levels (Abi) and onsocial and market access conditions, Li and Ki

(where Li measures household i’s access to off-farm labor markets, and Ki represents thehousehold’s access to financial and/or socialcapital). Finally, regression specification (1)contains other control variables (geographiclocation, lifecycle age of the households, etc.)that are represented by the vector Zi. The termti measures latent, random factors that impact


asset growth. Estimates of the parameters in (1)would permit investigation of many of the keyquestions put forward in this paper. 10 Do poorhouseholds exhibit greater sensitivity to assetshocks and recover more slowly in their wake?Does factor market access mediate such excesssensitivity? Do poor households exhibit mutedsensitivity to income shocks as they defend vul-nerable asset bases?

While estimates of Eqn. (1) can thus permitus to see whether poor and rich households re-spond asymmetrically to randomly distributeddamage, it does not permit us to directly iden-tify the existence of a poverty trap. For exam-ple, a finding several years after a shock thatpoor households are more deeply affected thanrich households may mean that poor house-holds fell into a permanent poverty trap, orsimply that they recover more slowly than richhouseholds and will eventually return to theirpre-shock trajectory and long-run equilibriumdestination. While the difference between thesetwo scenarios is something of an academicpoint, a long enough time period of observation(without further shocks) would in principle suf-fice to straightforwardly resolve this issue.

Absent this kind of very long-term data, themost compelling way to explore this issue withthe available data is to examine post-shockasset accumulation to see if it is possible toidentify long-run equilibrium positions ofhouseholds of different wealth levels. Note thatin the case of an asset shock like a hurricane,the post-shock changes in a household’s assetposition is exogenous. In the case of prolongedincome shock, such as a drought, the change inasset position is endogenous to the household’scoping strategy. Later analysis will attend tothis important difference.

Consider the following model of post-shockasset growth that is general enough to admitdistinct long-run asset equilibrium for house-holds of different wealth levels:

gsi ¼b‘AAsi þ b‘ZZi þ t‘i ; if Asi < c;

buAAsi þ bu

ZZi þ tui ; otherwise;

(ð2Þ

where gsi is household i’s asset growth from thepost-shock to the recovery period, Asi is thehousehold’s stock of assets in the immediatepost-shock period, c is the critical asset levelaround which bifurcation occurs, and the ‘superscripts indicate the parameters whichshape growth in the lower regime, while the usuperscripts designate parameters that lead tothe upper regime equilibrium. A poverty trap

would exist if households below the thresholdlevel (Asi < c) moved toward a lower equilib-rium asset level at which growth became zero.Such a low level equilibrium would occur ifthe parameter b‘A was sharply negative, signal-ing that asset growth collapses quickly towardzero as assets increase. A low level equilibriumcould also occur if the estimate of b‘ZZ werelow.

While (2) captures the basic notion of a pov-erty trap, a fundamental difficulty from a statis-tical perspective is that the threshold level isunknown, if it exists at all. One way to ap-proach this problem has been to employ flexiblenon-parametric methods to explore assetdynamics. As utilized by Lybbert, Barrett,Desta, and Coppock (2004) and later Adato,Carter, and May (2006) and Barrett et al.(2006), this method has been used to identifya critical asset threshold and correspondinghigh- and low-level equilibria. As used by theseearlier studies, these methods have been limitedto bivariate analysis (that is, asset dynamics arestudied without controlling for shocks or othervariables that may temper the accumulationrelationship).

An alternative approach is to employ Han-sen’s (2000) methods and directly estimate thecritical asset threshold around which assetgrowth dynamics bifurcate. In the empiricalanalysis to follow, we will follow Hansen’s pro-posed method and explicitly test for the exis-tence of a critical threshold and examine thedegree to which the parameters conform tothe hypotheses explained earlier. A limitationof the Hansen estimator is that it assumes thatthe threshold is the same for all units (as well asthe more conventional assumption that theerror term, ti, is orthogonal to the explanatoryvariables). While this latter assumption issomewhat troubling for the usual reasons, 11

recent theoretical work (Barrett & Carter,2006; Buera, 2005) indicates that the thresholditself should be a function of individual charac-teristics, including skill, entrepreneurial apti-tude, etc.

While the econometric difficulties behindendogenizing the threshold have yet to be re-solved, we here will employ the standardHansen estimator in order to obtain initial, ten-tative evidence on the possibility of assetthresholds. Our threshold results thus need tobe treated with a degree of caution as it is un-clear whether an estimate of multiple stable as-set positions reflects a true multiple equilibrium(in the sense that any agent would move to the


low position if pushed below the threshold), orwhether it simply identifies the distinctive assetpositions toward which high and low typeagents uniquely gravitate because of theirintrinsic characteristics.

3. GAUGING THE LONGER-TERMIMPACTS OF ASSET SHOCKS

IN HONDURAS

Hurricane Mitch carved a path of destructionacross Honduras in 1998, and the direct impactof the hurricane was almost instantaneous.Drawing on data collected shortly after the hur-ricane, Morris et al. (2001) report that poorrural households lost 30–40% of their crop in-come and measured poverty immediately in-creased 5.5 percentage points, rising from69.2% of households to 74.6%. They also reportthat lower-wealth households lost 15–20% oftheir productive assets (land, livestock, andplantations), compromising their capacity togenerate earnings and livelihood. Unclear,however, from this early study is whetherhouseholds were able to recover from lossesof this magnitude and rebuild their assets andlivelihoods.

The data available for this study provide awindow into these longer-term questions. Some30 months after Mitch, a sample of 850 ruralHonduran households (clustered in 30 munici-palities spread across 6 provinces) was surveyedas part of a study on the impact of land marketliberalization and asset accumulation. Includedin the questionnaire were a number of retro-spective questions that probed into the directimpacts of Mitch on household assets and in-come. The study also collected data on house-hold assets in 2000–01, giving a window intothe longer-term patterns of asset cycles andpoverty traps.

Surveyed households include those located inmountainous and indigenous areas (in theDepartments of Intibuca and Ocotepeque), aswell as households in more commercial farmingareas (including Comayagua and coffee-grow-ing regions of the Department of Santa Barbar-a) as well as the households located along theCaribbean coast which is home to Afro-Carib-bean Garifuna people (in the Department ofColon). 12 In addition to this cultural diversity,the sample also exhibits religious diversity, withsome communities overwhelmingly RomanCatholic, others largely comprised of Evangeli-cal Christians, and yet others some combina-

tion of the two. While these factors surelymatter for how families and communities man-aged to cope with the impacts of Mitch, analy-sis of them is beyond the scope of this paper.

(a) The impact of Hurricane Mitch on assetstocks and growth

Table 1 presents descriptive statistical indica-tors of the impact of Hurricane Mitch. Infor-mation is provided both for the overallsample and for households broken up by assetquartiles. Quartiles were defined based onhouseholds’ pre-Mitch holdings of productiveassets (Ab), where productive assets are definedas the value of land, plantations, machinery,and livestock. Note that we do not have infor-mation on social capital or migration assets.Access to these assets may of course be veryimportant for household coping strategies. Allassets were valued using 2000 price informationand were converted to $US using the marketexchange rate. As can be seen, wealth variessubstantially, averaging $650 for the poorestquartile and just over $75,000 for the wealthiestquartile. Annual household income in 2000–01was six times higher for wealthier householdsthan it was for poorer households ($996 versus$5,967). These low figures are consistent withthe high poverty rates typical of rural Hondu-ras. The variation in them is also a reflectionof the sharp levels of inequality found acrossmuch of rural Honduras.

As can be seen, 44% of households suffered aloss of productive assets from Hurricane Mitch.The percentage of affected households increaseswith household wealth (rising from 22% to 68%from the first to the fourth wealth quartile).This finding contradicts the notion that poorerhouseholds are more vulnerable to shocks,though it may be an artifact of the fact thatpoorer households had relatively little to lose.Indeed, as can be seen, among those house-holds suffering asset losses, poorer householdslost a greater percentage of their productivewealth (31%) than did wealthier households(8%). 13 Across all wealth quartiles, losses wereprimarily comprised of lost plantations andland (note that lost land literally means landrendered unusable because of soil loss, or insome cases, submersion by rivers that hadchanged course).

The second two panels of the table presentadditional descriptive data on households basedon whether or not they experienced a loss ofproductive assets from Mitch. Not surprisingly,

Table 1. Losses due to Hurricane Mitch—Honduras (mean values unless otherwise noted)

All households Pre-Mitch asset quartiles

I II III IV

Pre-Mitch productive assets (US$) 23,769 653 3,998 13,718 76,821Annual household income 2000–01 (US$) 2,440 996 1,127 1,716 5,927

Loss of productive assets

% Households with losses 44.3 21.8 31.7 55.6 68.3% of Pre-Mitch assets losta 12.0 31.1 13.9 12.2 7.5

Structure of asset lossa (% of total assets lost)

Land 29.6 22.6 16.5 25.1 31.4Plantations loss 60.5 62.2 75.4 65.3 58.6Livestock 8.6 13.8 8.1 9.6 8.3Machine 1.3 1.5 0.0 0.1 1.7

Households that lost productive assets

Income shock (US$) 428 144 164 328 722Housing loss (US$) 442 58 310 481 596Aid received (US$) 232 154 330 98 320Median asset growth (%) (pre-Mitch to 2001) �2.6 �5.0 �4.9 �2.1 �2.1

Households without loss of productive assets

Income shock (US$) 93 101 70 95 121Housing loss (US$) 119 187 96 53 89Aid received (US$) 141 88 134 161 261Median asset growth (%) (pre-Mitch to 2001) 5.4 8.8 5.4 4.6 3.0

a Figures calculated only for those households that suffered asset losses.


households that suffered asset losses also expe-rienced greater income losses (primarily cropincome). For the lowest wealth quartile, theselosses averaged 10–15% of annual householdincome. 14 Loss of housing stock was also sub-stantial for many households. Aid (typically inthe form of food and building materials) aver-aged between $50 and $600 across the quartilegroups, but in no case averaged more than10% of the value of lost productive assets.

Finally, Table 1 reports asset growth ratesfrom mid-1998 (pre-Mitch) to early 2001. 15

Across all pre-Mitch wealth quartiles, house-holds without assets losses show substantiallyhigher growth than those that suffered losses.The gap is 13.8% for the lowest quartile wherepoor households with losses had showed �5%net growth (loss) over the post-Mitch period,while poor households without losses had8.8% growth. The gap is a smaller 5.1% forthe wealthiest quartile (�2.1% versus 3% post-Mitch growth).

These growth differences seem to signal thatpoor households are more sensitive to shocks(and less able to recover from them). At thesame time, of those households that did notsuffer any asset losses, poor households tended

to grow faster (8.8%) than did wealthier house-holds (3.1%), signaling a convergent accumula-tion process. We now turn to more thoroughlyexplore these patterns of vulnerability and resil-ience.

(b) Regression analysis of asset loss and recovery

The standard growth model given by (1)above permits us to explore the underlying pat-tern of asset accumulation as well as the impactof shocks. Table 2 displays the results of esti-mating two alternative specifications of (1).The first, or basic, model includes initial assets(as a way to test for convergence), the shockvariables, and basic demographic and regionalcontrol variables. The expanded model includesadditional variables to see if the impacts ofshocks are mediated by labor and capital mar-ket access. Asset shocks, income shocks, andaid received are all normalized by the house-hold’s pre-Mitch asset level. Given this normal-ization, a coefficient of �1 on the asset shockvariable would mean that the household hadnot recovered at all from the shock (e.g., a10% loss of assets would reduce the growth rateby 10 percentage points). A coefficient less than

Table 2. OLS estimates of asset recovery and growth—Honduras

Explanatory variables Basic model Expanded model

Core growth & convergence

Initial assets, Ab (log) �0.55** �0.57**

(Ab)2 0.03** 0.03**

Sensitivity to asset losses, H (by initial asset quartile)

H� Q1b �1.78** �2.88**

H� Q2b �2.21** �2.85**

H� Q3b �1.17** �1.62**

H� Q4b �0.86* �1.22**

Income & other shocks

Income shock, e 0.05 0.04Housing loss (equals 1 if housing loss) �0.14** �0.10**

Mediating factors

Labor market access, L · H – 1.99**

L · H · Ab – �0.19**

Capital market access, K · H – 0.64**

Aid received – �0.04**

Demographic & other controls

Age of household head 0.02** 0.02**

Squared age of head �0.00** �0.00**

Post-Mitch inheritance (equals 1 if received) 0.61** 0.63**

Constant 2.48** 2.55**

Provincial dummies Included Included

R2 0.32 0.34* Significant at the 10% level.** Significant at the 5% level.


�1 would signal that households had furtherdrawn down on their assets in the wake of thehurricane, while a coefficient greater than �1would indicate at least partial recovery fromthe shock.

Two strong patterns emerge in both the basicand expanded models. First, the data signal astrong pattern of convergence as expected: assetgrowth diminishes significantly with the level ofinitial assets. Second, households who were inthe lower pre-hurricane wealth quartiles showmuch greater sensitivity to asset shocks (notethat Q1

b is the lowest or poorest pre-Mitch quar-tile). Both models permit the coefficient of theasset shock to vary with the pre-hurricanewealth quartile. As can be seen, the pattern ofimpact is non-linear across wealth levels. AWald test for equality of the 4 coefficients iseasily rejected for both models. For the basicmodel, the calculated F-statistic value of 3.31,while the 5% critical value F(1, 779) is 2.61. Ef-forts to capture this non-linearity with a parsi-monious specification of the variables (e.g., theasset shock interacted with pre-Mitch wealth)

were unsuccessful. The results for the basicmodel indicate that a 10% asset loss would beexpected to reduce growth for quartile 1 house-holds by about 18 percentage points, 22percentage points for a second quartile house-hold, but only 9 percentage points for a house-hold in the top wealth quartile. 16 Theseasymmetric effects suggest that shocks will off-set the tendency toward convergence, a findingconsistent with the story told by the descriptivestatistics in Table 1.

In addition to these core findings, both mod-els give similar results with respect to incomeshocks and housing loss. Income shocks arecounter-intuitively estimated to have a positivebut small and statistically insignificant impact

on subsequent asset growth. 17 As a controlfor other losses suffered in the hurricane, adummy variable was included in the regression,taking a value of one for households experienc-ing loss of housing. More intuitively, whenhouseholds did experience housing loss, recov-ery of productive assets was significantly slo-wed by approximately 10% percentage points.


Given these average patterns of sensitivity toasset shocks, the expanded model in Table 2permits us to see whether better access to factormarkets mitigate the impacts of shocks. Thelabor market indicator for a household is de-fined as the average off-farm labor market earn-ings within its community (there are a total of31 separate communities within the sample).The variable was scaled to lie between zeroand one by dividing it by the highest commu-nity average earnings level within the sample.The data used to construct this variable wascollected some two and half years after the hur-ricane. This lag in the collection of this infor-mation should make it a more reliableindicator of labor market access rather than ameasure of the endogenous exercise of laborsupply in the wake of immediate needs createdby the hurricane.

The capital access measure was also derivedfrom a survey at the later time period and isbased on a set of questions designed to probewhether or not a household was on its demandcurve for credit (and hence price rationed inthat market), or whether it had excess demandfor credit and hence was subject to some formof quantity rationing. 18 The capital access var-iable was defined as a binary variable taking thevalue of one when the household was pricerationed.

As can be seen in Table 2, labor market ac-cess has a large mitigating effect on shocks,especially for low-wealth households (note thatthe mitigating effect of labor market access dis-sipates quickly as assets grow). Capital marketaccess also strongly mitigates the impacts ofshocks. While this mitigation effect is in princi-ple wealth-neutral, Boucher et al. (2005) showthat access to capital is highly skewed with re-spect to wealth (only 40% of households inthe lowest wealth quintile are price rationedand do not have excess, unmet demand for for-mal sector loans, while almost 90% of house-holds in the top quintile are able to satisfy

Table 3. Sensitivity to and resili

Lowest wealth quar

No shock 31% Asse

Poor marketaccess

G

Pre-shock assets ($) 650 650Post-recovery assets () 902 32130 Month growth rate (%) 39 �50

their demand for capital in the formal creditmarket). Not surprisingly, when these mitigat-ing factors are included in the regression, theaverage effects of asset shocks increase, indicat-ing that households without access to labor andcapital markets are more severely affected bythe shocks than is indicated by the basic model.

In order to draw out the implications of theregression coefficients more clearly, we calcu-late predicted asset levels for a variety of styl-ized low- and high-wealth households thatexperienced different shocks in different marketenvironments. Table 3 presents the results ofthese calculations. Initial asset levels are takento be the mean for each quartile. For each assetlevel, the table contrasts the experience of ahousehold that had no asset shock with theexperience of a household that suffered a 31%asset loss (which was the mean loss level forthe lower-wealth quartile households that expe-rienced losses). Other shocks were set to zeroand all other household characteristics are setto mean levels for the sample.

For the no shock case, the low-wealth house-hold shows a higher growth (39% over 30months) than the high-wealth household (9%),reflecting the convergence property discussedearlier. However, in the high shock scenario,the excess sensitivity of poor households to as-set shocks completely overturns this modestconvergent process. Absent good market ac-cess, a low-wealth household that experiencedan immediate 31% asset loss is estimated toexperience further declines and a net assetgrowth rate of �50% from its pre-Mitch posi-tion to the time of the study 30 months later.A wealthier household that experienced anidentical 31% loss is estimated to have recov-ered partially from the loss and exhibit a netgrowth rate of �28%. Were we to further takeinto account that wealthier households on aver-age only lost 7.5% of their assets (not 31%),then the unequalizing effect of the shock wouldbe further magnified. When poor households

ence from shocks—Honduras

tile Highest wealth quartile

t loss No shock 31% Asset loss

ood marketaccess

Poor marketaccess

Good marketaccess

650 76,821 76,821 76,821686 83,905 54,951 64,5375.5 9 �28 �15


are compared to where they counterfactuallywould have been without a shock, the impactsof the shocks stand out even more sharply.

Finally, Table 3 shows that more buoyantlabor and capital market access serves to offsetthe unequalizing effect of asset shocks. Underthese circumstances, lower-wealth householdswould be able to offset the entire 31% asset loss(climbing back to a net asset change of positive5%). Wealthier households are also estimatedto benefit slightly from better market access,though they still show a lingering effect of theasset shock on their asset accumulation.

These results are especially interesting in thecontext of a related study by Carter and Castil-lo (2005). In that study, Carter and Castillo findthat recovery from Mitch was more rapid incommunities characterized by high levels ofpro-social norms of trust and altruism. Interest-ingly, further analysis by Carter and Castillosuggests that only a subset of households seemsto actually benefit from the pro-social environ-ment, suggesting that there may be processes ofexclusion that prevent all households from ben-efiting from socially mediated access to insur-ance and capital. If correct, when mergedwith the results presented here, we seem to seea situation in which local social mechanismsleave poor Honduran households quite vulner-able to asset shocks. In this environment, accessto supporting capital and especially labor mar-kets seem to be especially important.

Table 4. Threshold poverty

Explanatory variables L

Core growth & convergence

Post-shock initial assets, As (log)

Asset loss

Housing loss

Mediating factors

Labor market access, L

Capital market access, K

Demographic & other controls

Age of household headSquared age of headPost-Mitch inheritance (equals 1 if received)ConstantProvincial dummies

ObservationsR2

Threshold interval estimate (95% interval)

* Significant at the 10% level.** Significant at the 5% level.

(c) Testing for poverty traps

The analysis up to this point has shown thatasset shocks in Honduras have the potential forderailing an otherwise convergent accumula-tion process in which lower-wealth householdsaccumulate more quickly than their richerneighbors. However, it is unclear whether ornot these results signal an irreversible setbackfor low-wealth households who become miredin a poverty trap, or whether their recoveryis simply slower. As discussed in Section 2,threshold estimation potentially offers addi-tional insight into this question by permittingus to estimate the longer-run equilibrium to-ward which lower-wealth household are head-ing.

Table 4 displays the results of this analysis,based on the model in Eqn. (2). The dependentvariable is asset growth over the 30 month per-iod immediately following Mitch. Initial assetsare defined as household’s post-shock assetlevel, As. Note that this definition is differentthan that used in the prior section’s analysis.Growth is in this section defined off the baseof the post-shock assets in order to test the no-tion of an asset poverty trap which suggeststhat households that fall below a critical thresh-old level will become mired in a low level equi-librium trap. Because this post-shock growthmeasure already nets out the direct effect ofthe shock, the regression specification does

trap estimates—Honduras

ow regime As < 244 High regime As > 244

�0.65** �0.04**

�1.46** �0.10

0.27 0.01�1.19** 0.11**

�0.01 0.020.00 �0.00

1.95** 0.486.41** 0.23

Included Included

43 7270.62 0.22

235–249**


not include the shock itself (in contrast to themodel in the prior section).

Using Hansen’s (2000) threshold estimator,we can now estimate whether the data can besplit into two groups—above and below a crit-ical initial asset value—which exhibit signifi-cantly different growth dynamics. Note thatany observations in a low group will includehouseholds that both randomly fell into thisgroup based on the vagaries of HurricaneMitch, as well as households that were alreadyat low asset levels prior to the hurricane. 19

As shown in Table 4, the data indeed suggestthe presence of a threshold at an asset levelof $244. This estimate is quite precise, as the95% confidence interval estimate of the thresh-old ranges from only $235–$249. 20 The explan-atory variables are largely the same as thoseincluded in the Table 2 regression. 21

Table 4 regression results can perhaps bemost easily appreciated by exploring themgraphically. Figure 3 graphs the two regimeswhere all variables except initial assets are heldat their mean levels for the full sample. The lowregime displays a pattern in which growthdrops off very quickly as assets increase, andsignal the existence of a low level equilibriumat about $225 (where expected asset growthequals zero). The upper group displays a pat-tern of only very slowly diminishing growthand an upper-level equilibrium that is orders

0 200 40Post-shock

-0.1

0.4

0.9

1.4

1.9

30 M

onth

Ass

et G

row

th R

ate

E

Figure 3. Post-Mitch po

of magnitude greater than that for the low levelgroup.

While these results should be treated withcaution (see the discussion in Section 2), theyare consistent with the sort of poverty trap de-scribed by Ato Mohammed in the quotation atthe beginning of this paper. While full explora-tion of the poverty trap hypothesis will requirefurther methodological developments (seeCarter & Barrett, 2006), the Hurricane Mitchquasi-experiment does unambiguously indicatethat asset shocks have long-lived, unequalizingeffects.

4. ASSET SMOOTHING AND DROUGHTRECOVERY IN NORTH-EASTERN

ETHIOPIA

Unlike the swift destruction brought aboutby Mitch in Honduras, the drought of the late1990s that afflicted our study area in North-eastern Ethiopia (the South Wollo andOromiya Zones of the Amhara Region) was aprolonged event, with uneven consequencesand a gradual onset. Indeed, the first signs ofdisaster can be traced to the poor short rains(January–April, called the Belg season) of1998 where it is estimated that harvests wereonly 60% of normal yields in the main Belggrowing areas (Government of Ethiopia, 1997,

0 600 800 Assets, As ($'s)

stimated Threshold

As< 244As >244

verty trap—Honduras.


1998a, 1998b). That year the long rainy season(June–September, called the Meher) was nearnormal for all areas except in the Belg growingzones where there is also some dependence onthe Meher season. Because the Meher rains of1998 were near normal in some locations,drought and relief agencies in Ethiopia failedto see the looming disaster until the Belg seasonof 1999 emerged as a massive failure, resultingin 90% loss of crops (see Castro, Amare, Adal,& Tolossa, 1999).

The 1999 Meher season was only somewhatbetter, yielding about 40% of normal harvestsin six of the eight kebeles 22 in which data forthis study are available. Food aid distributionstarted in the region in June 1999, but wasnot widespread until August 1999. To makematters worse, the Belg season of 2000 was verypoor (75% reduction of normal yields). Withmassive imports of food aid and the recoveryof the long rains in 2000, the nutritional statusof the area’s population had returned to nearnormal by early 2001. Thus, the drought ofthe late 1990s was keyed by the failure or nearfailure of three successive short rainy seasons.The first of the crop failures was only 40%,but with such widespread poverty it was en-ough to initiate the downward spiral of exten-sive food insecurity and distress sales of assets(mainly livestock) that characterized the regionfor the better part of 30 months.

The long-term possibilities of asset recoveryfrom this series of shocks would be expectedto be conditioned by several community andhousehold characteristics. For example, in ourstudy area 65% of households participate infuneral clubs (called iddir or kire); 67% in farmwork groups (debo or wonfel) where memberswork on each other’s farms during peak peri-ods; and 18% in religious organizations. Theprevalence of certain social institutions variesby ethnic and religious affiliations. Individualsin the study area either are members of theOromo (25%) or Amhara (75%) ethnic groups,and 86% of the region are Muslims while 14%are Coptic Christians. The study region (as in-deed large parts of Ethiopia) is characterizedby relatively weak labor markets and nearly ab-sent credit markets. Land markets are severelyrestricted in that private ownership is prohib-ited, and legal constraints on land rentals wereonly recently relaxed. In part because insuranceagainst crop loss is practically absent and mar-ket-based coping mechanisms are limited, foodaid makes up a relatively large portion of foodconsumption, as indicated above. Data are

from a seven-round household survey con-ducted over three-and-half years in eight kebe-les in South Wollo and Oromiya zones. Thedataset also includes recall questions on live-stock holdings during 1996–99, which assesshow households fared in terms of their assetsprior to the onset of the drought.

The year 1999 also witnessed the heaviestlivestock asset reductions, both due to animaldeaths as well as sales. Based on qualitative sur-veys in the study area (see Amare, Tolossa,Castro, & Little, 2000; Castro et al., 1999; Lit-tle, Stone, Mogues, Castro, & Negatu, 2006), itis estimated that 25% of livestock reductions in1999 were distress sales at which the seller re-ceived less than 50% of the normal price ofthe animal sold (cattle prices, e.g., droppedfrom an average of 625 birr [US$74] in thepre-drought period to 291 birr [US$35]). Priceswings of this magnitude constituted a hugecapital loss for those forced to sell livestockduring this period. Natural causes clearly pre-cipitated the drought disaster of 1999, which re-sulted in a massive humanitarian effort, but thepopulation’s vulnerability to relatively smallperturbations in climatic events is ‘‘unnatural’’and highlights the extreme poverty in the area.

(a) Livestock losses and their recovery

Figure 4 gives a first indication of how theweather shocks discussed above impactedhouseholds. The top panel shows the evolutionof mean livestock by households in the four,pre-drought wealth quartiles. Livestock assetsare here aggregated in Tropical Livestock Units(TLUs). 23 Following the onset of the droughtin 1998, the top two wealth quartiles appearto exhibit typical consumption smoothingbehavior, as livestock assets begin to dip shar-ply. In contrast, the two lowest quartiles appearto more stubbornly hold on to their livestock,showing on average only small decreases inlivestock near the end of the drought period.This apparent asset smoothing behavior is whatwe would predict if these lower quartile house-holds were in the vicinity of a critical povertytrap threshold (see Hoddinott, 2006 for ananalysis of Zimbabwe).

Analogous to Honduras, one way to furtherprobe the existence of poverty traps is to exam-ine post-shock growth trajectories to see ifthose households that were worst off at theend of the disaster show evidence of being un-able to recover and move toward higher assetlevels. The bottom panel of Figure 4 displays

1996

1997

1998

1999

Jun '00

Dec '00

Jun '01

Nov '01

March '02

Jul '02

Jul 03

1996

1997

1998

1999

Jun '00

Dec '00

Jun '01

Nov '01

March '02

Jul '02

Jul 03

0

2

4

6

8

10

Tro

pica

l Liv

esto

ck U

nits

Pre-Shock Quartiles

Drought Recovery

Wealthiest Quartile

0

2

4

6

8

10

Tro

pica

l Liv

esto

ck U

nits

Post Shock Quartiles

Drought Recovery

Poorest Quartile

Wealthiest Quartile

Poorest Quartile

Figure 4. Evolution of mean livestock holdings by wealth quartiles—Ethiopia.


the livestock trajectories based on wealth quar-tiles defined according to animal holdings at theend of the drought. As can be seen, averageholdings of the poorest quartile were essentiallyzero at that time. Interestingly, however, thisgroup on average managed to add substantiallyto their livestock holdings over the span of thefollowing three years, especially in contrast tothose households better positioned in the wakeof the shocks. The latter also recovered, but ata slower rate. At this descriptive level, thesepatterns are not consistent with what we wouldexpect to see if the poorest households hadindeed fallen into a poverty trap a la AtoMohammed.

This pattern may be partially explained by theinclination of better-off households to continueto market animals after the drought, in order tobenefit from the post-shock boom in livestockprices (Little et al., 2006). It likely also reflectsthe fact that households’ post-shock asset posi-tions were endogenous to their own coping deci-sions over the drought period. Households thathad completely stocked out by mid-2000 mayhave been precisely those households that haveseen it necessary to thoroughly invest in socialcapital for their survival, and thus may have en-joyed relatively stronger endowments of socialcapital and perhaps other assets. Another studyusing the same data has shown that local social


mechanisms assume considerably more impor-tance for the poor during recovery periods (Lit-tle et al., 2006, p. 220). Those households thatapparently defended their livestock could beisolated households who—like Ato Moham-med—would have been trapped by low post-shock resilience and growth had they let theirstocks deplete to almost nothing. Subsequenteconometric analysis will take into consider-ation the endogeneity of drought coping strate-gies discussed above when examining recoveryperiod livestock growth.

(b) Econometric estimates of asset andconsumption smoothing

Employing the asset growth model put for-ward before, we now turn to a more thoroughanalysis of the data using methods that permitus to control for these multiple factors thatinfluence observed outcomes. In doing so, wewill estimate growth over the two distinct timeranges. This section will examine coping strate-gies over the two year drought period, using theasset growth model to study asset draw downor negative growth. The following section willthen turn to a separate examination of post-shock (positive) growth.

To explore coping and negative asset growth,we use a measure of the magnitude of the

Table 5. Estimates of asset changes

Explanatory variables

Shocks

Income shock, eIncome shock * initial assets

Mediating factors

Community social capital, SKLabor market access, L

Access to food aid, F

Demographic factors

Household sizeAge of household headSquared age of headGender of household headLand assets (in timad = 1/4 hectares)

ConstantNumber of censored observationsR2

LR v2ð11Þ

p-Value

Sample size: n = 416; number of clusters: 8.* Significant at the 10% level.** Significant at the 5% level.

shock, ei, directly in the specification as wellas letting the coefficient for ei depend on thepre-shock asset level of the household, Abi.While the available data lack a measure of theseverity of the shock received by each house-hold, we approximate it with the share ofhouseholds in the community of household ithat suffered crop losses in the worst season,the 2000 belg season.

The first column of Table 5 gives the resultsof the estimation of asset changes during thesuccessive poor seasons. We see in the OLS esti-mation result that the shock coefficient is quitesubstantial in size (but not significant) across allestimations: for example, the model associatesa 10 percentage point decrease in the share ofhouseholds of the kebele that experiencedweather-induced crop loss with a 105.7 percent-age point increase in the asset growth rate overthe two year time period of the drought. Whilethis appears to be an enormous effect, it is notquite as huge when compared against the distri-bution of rates of livestock asset changes overthis period. For example, for households whohad a positive growth (about 40% of all house-holds), the median two year growth rates were58%, the third quartile 141%, and the 90th per-centile was 1025%. The large effects of shockscan be seen as taking growth rates back downfrom very high levels.

during drought years—Ethiopia

OLS Tobit

�11.80 �22.59�0.31* �0.21*

0.78 1.384.219** 3.17�5.98 �9.11

�0.15 0.09�0.44 �0.48**

0.004 0.004**

1.64** 3.01**

0.41 0.63

19.62 25.7960

0.0726.010.004


The estimates also show that having higherassets magnifies the detrimental impact ofshocks on growth. This effect is significant atthe 10% level, though small in magnitude, andsuggests that poor households seek to defendtheir assets in the face of successive droughtsrather than liquidate them and perhaps limittheir subsequent chances of recovery. That is,it is consistent with asset smoothing behaviorby the lower-wealth households.

The estimated model also explores the degreeto which social mechanisms, food aid, and la-bor markets bolster asset growth. Access to so-cial institutions is measured as communityaverage membership in social organizations. 24

Food aid is measured as the percentage of acommunity participating in food for work pro-grams. This community level variable bypassesindividual-level endogeneity problems createdwhen more severely shocked households chooseto participate in the program. It may, however,still suffer from the fact that programs areplaced in communities where the drought wasmost severe. The labor market indicator is anaverage of off-farm earnings for all householdswithin a kebele. This number was then normal-ized to vary between zero and one by dividing itby the highest level of average earnings amongthe eight kebeles in the sample. 25

Community membership in social organiza-tions (social capital) increases the rate ofgrowth (or limits the rate of loss) of livestockover this period. There is even stronger evi-dence that labor market access positively affectsthe rate of growth of livestock. Finally, avail-ability of food aid in the community does notappear to protect households’ future assets,and in fact seems to have the opposite effect. 26

As is indicated in Figure 4, after those in thelowest strata bottom out, they do not have theoption to draw down assets further. This moti-vates investigating whether the results inter-preted as evidence of asset smoothing by thepoorer households, as suggested by the esti-mates in the first column of Table 5, are insteaddriven primarily by the inability of householdsin the lowest strata to liquidate assets, for lackof the latter. To check for this, Table 5 also in-cludes a tobit estimation with the asset growthrates censored from below at �100%. 27 In-deed, the share of households that had someholdings at the beginning of the series of weath-er shocks but stocked out by the end of this per-iod is about 6%, and 14% of households had nolivestock at the end of the shocks (irrespectiveof their initial assets). 28 Generally, in Table 5

the censored and ordinary least squares estima-tion do not differ from each other by much, andmore specifically, the effects of asset smoothingare still present, suggesting that this was notmerely driven by the potential growth censor-ing problem. 29

(c) Post-drought asset growth trajectories

This section focuses on the recovery period toexplore the existence of long-run effects andpoverty traps induced by the drought. Follow-ing the strategy outline in Section 2 and model(2), we employ Hansen’s (2000) threshold esti-mator to examine whether there exists a criticalasset threshold around which asset growthdynamics bifurcate. In assessing the determi-nants of the recovery period growth, includinghow households’ starting position affects subse-quent growth, it is important to account forhow households’ post-shock position is endog-enously determined by the coping decisionsthey made during the drought period. Unlikein Honduras, where the change in households’asset stocks immediately after Mitch was exog-enously determined by the destruction due tothe hurricane, in Ethiopia the livestock assetspeople possessed post-drought depended onthe coping strategies they employed. In the fol-lowing analysis of recovery period growth, weaccount for this endogeneity by basing ourmodel on the predicted levels of post-shock as-sets, bAs (estimated using the results in Table 5),rather than the observed levels of post-shockassets, As.

The results employing Hansen’s thresholdestimator are given in Table 6. The dependentvariable is the simple three livestock growthrate (not annualized). The data indicate thepresence of two distinct, statistically differentregression regimes. Households with initial(predicted) assets in excess of 0.59 TLU fallinto the upper regime and those below intothe lower regimes. Approximately, half of theobservations fall into each regime. The thresh-old dividing the two regimes is predicted quiteprecisely (the 95% interval estimate of thethreshold is 0.57–0.61). The Hansen joint-R2,which indicates explanatory fit of the modelgiven the estimated threshold, is 80.9%, higherthan the equivalent single equilibrium model(not reported here), which has an R2 of 75.7%.

Figure 5 graphs the two estimated regressionfunctions, holding all variables except initial as-sets at their mean sample values. Within boththe upper and the lower regimes livestock

Table 6. Threshold poverty trap estimates—Ethiopia

Explanatory variables Low regime As 6 0.65 TLUs High regime As > 0.65 TLUs

Convergence

Initial assets, ln(As) �1.83** �0.80**

Mediating factors

Community social capital, SK 0.23** 0.05Labor market access, L 1.11** 0.57Access to food aid, F �1.14* �0.79

Demographic factors

Household size 0.20** 0.14**

Age of household head 0.04 0.01Squared age of head �0.0004 �0.0001Gender of household head 0.72** 0.67*

Land assets (in timad = 1/4 hectares) 0.10 �0.01

Constant �3.84** �0.45R2 0.72 0.51Number of observations 214 202

* Significant at the 10% level.** Significant at the 5% level.


growth decelerates as initial assets increase. Ac-cess to off-farm opportunities is not only impor-tant for protecting assets during the poorseasons as seen earlier, but also for the assetrecovery period. However, there is stronger evi-dence that this is the case for the low-wealth re-gime. Similarly, the role of access to socialinstitutions appears also to be larger and signif-icant for low-wealth households. Households in

0 1 2Post-shock Assets, As

-2

0

2

4

6

Liv

esto

ck G

row

th o

ver

Rec

over

y Pe

riod

Estimate

Figure 5. Growth equilibria of

the lower regime show rapid initial growth thatquickly dissipates to zero at about 0.5 TLU.Note that this is still less than one oxen. The esti-mated livestock equilibrium for the upper groupis at about five-times higher, at nearly 3 TLU.

The data do thus signal the existence of twodistinctive equilibria as expected from a pov-erty traps perspective. Interestingly, we do seethat poorer households are able to rather

3 4 5 (Tropical Livestock Units)

d Threshold

As< 0.6 TLUAs > 0.6 TLU

two asset regimes—Ethiopia.


quickly rebuild their livestock, at least up to thelevel of this lower equilibrium point. One rea-son why poorer households have been able torebuild their modest animal holdings fasterthan others may relate to the land tenure sys-tem in South Wollo (see Little et al., 2006).As explained earlier, land is controlled by thestate and because of this there is an ‘‘upper lim-it’’ of land (2.5 hectares) that a household isallocated for farming and grazing purposes. 30

In a land-constrained economy like South Wol-lo, where communal grazing areas are few,there are real challenges (in terms of manage-ment and the availability and costs of fodderand pastures) for better-off households to re-build large herds quickly. This is especially soin a post-drought recovery period when landrental prices increase, which is an importantway to increase land holdings under currentconditions. Poor households, in turn, are likelyto have adequate access to lands and pastures,relative to their livestock asset endowments,to rebuild their small herds without the needto purchase fodder or rent additional lands.

Finally, it is worth remarking that one fea-ture of the estimated equilibria seems odd. Sim-ilar to the results obtained for Honduras, theasset equilibrium of the lower regime (i.e., theinitial asset level that is associated with zerogrowth) is just slightly below the threshold va-lue of 0.59 TLUs. While there is nothing con-ceptually wrong with this proximity of theequilibrium to the threshold, we probably needadditional experience with this class of estima-tor to better understand these kinds of results.

5. CONCLUDING REMARKS

In the fictive world of full and complete mar-kets, poor households could draw on loan andinsurance contracts to cope with the oftendisastrous asset and income losses brought bysevere environmental shocks. Drawing on fu-ture earnings, households in this world couldrebuild lost assets and sustain their level of cur-rent consumption without further depletion oftheir productive assets and future possibilities.The story of one Ethiopian household told atthe beginning of this paper is an example ofhow far the actually existing world can be fromthat fictive world. In the real world of AtoMohammed, environmental shocks can decapi-talize the poor, and trap them in impoverishedposition from which they cannot escape. Whenthis happens, a humanitarian problem of disas-

ter relief becomes a long-term developmentproblem.

In an effort to better understand the impactsof natural disaster, this paper has employedlongitudinal data on assets to understand thelonger-term impacts of two shocks, the three-year drought of the late 1990s in Ethiopia andthe 1998 Hurricane Mitch in Honduras. Analy-sis of the Honduran data reveals that the med-ium-term effects of the shock differ by initialhousehold wealth. Relatively wealthy house-holds were able to at least partially rebuild theirlost assets in the three years following theshock. In contrast, for the lowest wealthgroups, the effects of the hurricane on assetswere of longer duration and felt much moreacutely. This differential impact of the hurri-cane is especially striking because it seems tohave derailed what would otherwise have beena convergent growth path in which poorerhouseholds accumulate productive assets morequickly than their richer neighbors. In an effortto determine whether these patterns signal thepresence of a poverty trap, such that thosepoorer households who lost assets to the hurri-cane will never recover, we also estimated athreshold model of poverty traps. While thereliability of the estimates depend on severalstrong assumptions, the estimates do indicatethe presence of a poverty trap such that house-holds that begin beneath (or fall below) an assetthreshold of $250 are expected to gravitate to alow-level equilibrium, orders of magnitude be-neath that of their better off, or more fortunateneighbors. However, more work is clearlyneeded to generalize the estimation methodsand improve the reliability of these estimates.

For Ethiopia we had the opportunity toexamine asset changes over two distinct timeranges, a period of drought and coping, and aperiod of recovery. Analysis of this data weaklyreveals a pattern of asset smoothing among thelowest wealth households, meaning that thehouseholds at the bottom try to hold on to theirfew assets even as income and consumptionpossibilities dwindle during the period of severelosses in agricultural production. Such behavioris consistent with what would be expected in theface of a poverty trap. Following the drought,low-wealth households are estimated to accu-mulate assets faster than non-poor households,controlling for other variables. But similar tothe Honduran case, the threshold estimatesagain signal the existence of a lower equilib-rium at which these poor households settledown and stop growing. The possibility that


shocks have different and more durable effectson the less well-off seems well established,though additional work is needed to increasethe reliability of the threshold estimates whichunderlie the identification of a lower equilib-rium for poor households.

In addition to these econometric issues, theanalysis here has in other dimensions fallenshort of fully resolving all the puzzles and com-plexities of the two disasters studied here. Forexample, an evaluation of the full welfare im-pacts of asset poverty traps would require amore complete set of assets and better informa-tion about asset values (in the Ethiopian case),as well as the consumption consequences of as-set smoothing behavior. Further research mayalso reveal other important facets of wealth-differentiated asset recovery experiences byexploring the asset composition of variedwealth categories of households. For example,do the poor in Ethiopia hold most of the ani-mal assets in small stock and chickens, whilethe better-off own more cattle and plow oxen?These types of assets differ in their ‘‘lumpiness,’’their abilities to breed (hence recover) rapidly,their functions in income and wealth genera-tion, and in the extent to which they areprotected in the face of shocks. Future studiesshould address the likelihood that differentasset portfolios of the rich and poor mayconstrain or facilitate post-disaster recoverypaths.

While future research to solve remainingquestions is always desirable, given the impor-tance of natural disasters it seems worth specu-lating on the policy and developmentimplications of our findings, especially sincemost disasters and their impacts are treated ashumanitarian and not development problems.Building ‘‘productive’’ safety nets that keepvulnerable households from losing their assetsis one important policy intervention in disas-ter-prone areas. As Barrett and Carter (2006)analyze in detail, keeping households abovecritical asset thresholds can generate large so-cial (and private) returns if poverty traps in factexist.

In addition to its implications for safety nets,the analysis here has shown that access to socialand financial capital, and to off-farm earningopportunities, can be especially important forlow-wealth assets as they cope with naturaldisasters. Policies that improve non-farmemployment opportunities, rural market infra-structure, and accessibility of credit—especiallyin the post-disaster period—are important waysthat governments and development agenciescan help limit long-term asset depletion. Giventhe importance of social networks especially forthe poor, any form of development policy needsto be cognizant of the way in which such socialnetworks operate so as to minimize any poten-tial negative impact of programs on the func-tional elements of social institutions.

NOTES

1. Ato (which means Mr.) Mohammed is the head ofone of the households in one of the two country casestudies of this paper.

2. While the destruction wrought by the hurricanewould seem to be random and unrelated to unobservedindividual characteristics, we do not formally establishthis point. We do show in Section 3 that wealthierindividuals more likely suffered some losses, whileamong those affected, poorer households lost a largerpercentage of their assets.

3. Mogues and Carter (2005) theoretically explore theidea that poor households will be less able to accumulateeffective social capital in more polarized and inegalitar-ian economies.

4. While the discussion which follows lists copingstrategies in a rough order of decreasing desirability

for discursive purposes, any household’s true preferenceswill depend on a number of factors.

5. The notion that some households might remainmired in a trap of persistent poverty is surprising fromthe perspective of some dynamic economic theory thatsuggests that less well-off households would be expectedto have every incentive to try to save, accumulate, andcatch up economically with their better off fellowcitizens. Carter and Barrett (2006) discuss the forcesthat could offset convergence, identifying lack of accessto market or socially mediated access as the key force.

6. Another possible motivation for retaining assets in atime of drought are adverse terms of trade for livestockvis-a-vis grain, as often occurs in drought periods (e.g.,see the discussion in Fafchamps, Udry, & Czukas, 1998).The depreciation of the (relative) value of assets,however, would likely—all else equal, including


asset-smoothing motivations—exist for all householdsirrespective of wealth. Hence it is not expected that wewould see asset retention for terms of trade reasons onlyamong the poor. A similar comment applies toMcPeak’s (2004) observation that households may beexpected to hold on to livestock if income losses werecorrelated with asset losses.

7. As Hoddinott stresses, the cost of an asset smooth-ing strategy is not only immediate hunger. In the case ofZimbabwe, asset smoothing permanently reduced thegrowth and, presumably future capacity, of young,nutritionally vulnerable children.

8. Fafchamps et al. (1998) find seemingly puzzlingregression evidence that at least some households in theirWest African sample do not manage their livestock so asto smooth consumption over time in the face of shocks.One explanation of their finding is that a subset of theirhouseholds is at a threshold where asset smoothingbecomes a rational response to shocks, and hence theirdata is a mix of asset and consumption smoothers.

9. Note that this model is a microeconomic adaptationof the standard neoclassical growth model used in theliterature studying convergence in living standardsbetween countries. As in the case of this literature, onecan raise the question about the potential endogeneity ofinitial wealth or income levels. While these initial levelsare temporally pre-determined, they may be correlatedwith household- (or country-) specific characteristics notcaptured by the model. While we control for manyhousehold and community characteristics in our sub-sequent analysis of the model in (1), we are in the endsubject to the same criticisms that have been applied tothe macroeconomic convergence literature.

10. The damaging events meted out by the naturaldisasters (e.g., high winds, rain shortfalls) should berandomly distributed across households. These natural‘‘experiments’’ thus gave us a random subsample of low-wealth households that experienced damage and anothersubsample of low-wealth households that did not (thesame of course holds for better off households).Regression estimates of (1) should thus permit us toreliably recover the impact of the shocks on those typesof household found to be poor and those found to berich.

11. Consistent estimation of (2) will only result if theerror term, ti, is orthogonal to the explanatory variables.While it has been common to make this assumption inthe macro growth literature which regularly regressesgrowth rates on initial conditions (see, e.g., Hansen,2000), it is a questionable assumption.

12. This sample is comprised of two distinct sub-samples: panel and cross-section. The panel households(500) originate from a study conducted in 1994 (Lopez &Valdes, 2000) in which 450 farm households were inter-viewed to analyze the impacts of the initial land titlingprograms. The 2001 survey attempted to follow boththese baseline households and the land they cultivated. Ofthe original baseline households, 362 were resurveyed. Inaddition, 138 ‘‘new’’ panel households were added. In2000, these households were cultivating land that hadbeen worked by the original panel households in 1994.The remaining 350 cross-sectional households were addedin regions that were not covered in the 1994 study.Households that were not operating their own farm in1998 were eliminated from the sample for purposes of thisstudy, reducing the total number of households to 821.

13. This pattern is consistent with what would beexpected if damages were random in the sense that everyhectare had an equal and independent probability ofloss. In this case, fewer poorer households wouldexperience loss (as each had fewer potential hectareson which a loss could potentially occur), but those whoexperience loss would lose a larger fraction of their totalassets.

14. The percentages are approximate as householdincome is measured only for the year 2000–01 and notfor the year of the hurricane.

15. Median growth rates are reported in the table. Meangrowth rates are higher in all cases, but follow the samequalitative pattern. Many of the high growth rates appearto be the result of inheritances received post-Mitch.

16. These differences are statistically different from a10% growth loss for the lowest two quartiles as we canreject the hypothesis at the 1% level that the coefficientof the asset shock is �1. We cannot reject this hypothesisfor the top two quartiles.

17. The insignificance of the income shock variable maybe an artifact of a pattern in the data whereby highestincome losses occurred for households involved in highvalued commercial activities that also gave them betteraccess to market-based coping mechanisms. Interactingincome shocks with the market access variables did not,however, uncover any systematic patterns.

18. Being price rationed is not the same thing as havinga loan. Price rationed includes both households thatborrow as well as those that did not need or want toborrow given the price of credit and their access to othersources of funds (see Boucher, Barham, & Carter, 2005,for further details).


19. We may worry that this latter group has intrinsiccharacteristics which lock them into a low equilibrium.It is thus unclear whether the equilibrium for this lowgroup simply represents a low level convergence club for‘‘low type’’ individuals, or whether it represents agenuine multiple equilibrium traps that would absorbhigh type individuals were they to fall below thethreshold. Efforts to estimate initial asset holding usingdata on parental characteristics were not successful.

20. The joint R-squared for the threshold model is 0.30,in contrast to an R-squared of only 0.17 for the pooledmodel that assumes that there is only a single growthregression regime.

21. The squared initial asset term was tiny and insig-nificant and was not included in the reported regression.Note that the splitting the sample into two groupsshould reduce the need for additional terms to capturehighly non-linear growth relationships.

22. A kebele is an administrative unit comprised of fouror five villages.

23. As used here, a TLU (Tropical Livestock Unit) is: 1TLU = 1 head of cattle (oxen, bull, cow, calf, heifer);0.5 TLU = 1 Horse/Donkey/Mule; 1.4 TLU = 1 camel;0.1 TLU = 1 sheep/goat; 0.05 TLU = 1 chicken. TheTLU factors approximate weight, subsistence (food),and market value of different animals.

24. These include, for example, burial societies (kire),informal credit associations (iqqub), and religiousgroups (mehaber and senbete).

25. With ‘‘labor market’’ in this context we refer to thebroader array of off-farm work opportunities includingbusiness income, rather than narrowly wage labormarket.

26. This result is somewhat surprising. The direction ofthe impact cannot arise from endogeneity as would beexpected if household-level aid receipts were used(Quisumbing, 2003) since the measure used here is acommunity aggregate rather than individual participa-tion in food-for-work. A study focusing on the impacton food aid on welfare in three regions in Ethiopiaincluding South Wollo (Mathys, 1999) points in asimilar direction (though less sharply), finding that whilein the short term asset sales are somewhat reduced with

food aid, months later households tend to resort back toelevated sales. Other work on South Wollo points to thelimitations of food aid in enabling recovery fromdisasters in the long run (Little et al., 2006). Further-more, similar results are obtained in a study (Mogues,2005) using the same dataset and dynamic panel datatechniques to assess the determinants of the evolution oflivestock assets in Ethiopia. Here, the endogenity isinstrumented using GMM techniques. The negativecoefficients in multiple studies and using multiple tech-niques remain a bit of a puzzle and may require furtherinvestigation in the future.

27. Households that entered the post-shock periodwith zero assets were censored to have growth no smallerthan 0% as they had nothing further to lose. To insurethat the restricted choices of these households wereproperly treated by the program, each was assumed tohave one chicken (0.05 TLU) at the end of the shockperiod. If the household was registered as having zeroTLU at the end of the recovery period, their growth ratethus became �100% such that the tobit regressionproperly reflected their censored behavior.

28. Given that 16% of the households enter thebeginning of the drought period with no assets at all,we made an adjustment that would permit obtaininggrowth rates, namely, to increase all livestock assets bythe same small increment.

29. We also examined a specification including initialassets, in order to also see how these may change theresults. Introducing a convergence effect indeed seems toobscure the asset smoothing effect identified in Table 5.This is perhaps not surprising: both the concept ofconvergence as well as asset smoothing have the sameempirical implication in this context, at least for the timerange prior to the recovery period: they both suggestthat the assets of the initially less well off households willgrow at a faster pace (or decline at slower pace) thanbetter off households’ assets. However, in examining therate of change in assets over the drought period, we arenot inherently interested in the question of convergence(discussed more in the next subsection), but in copingstrategies implied by asset changes.

30. However, the legal bound on land size can besignificantly smaller than that for any given household,as the maximum size is determined by the number ofhousehold members and other household characteristics.

REFERENCES

Adato, M., Carter, M. R., & May, J. (2006). Exploringpoverty traps and social exclusion in South Africa

using qualitative and quantitative data. Journal ofDevelopment Studies, 42(2), 226–247.


Amare, Y., Adal Y., Tolossa, D., Castro, A. P., & Little,P. D. (2000). Food security and resource access: afinal report on the community assessments in SouthWello and Oromiya Zones of Amhara Region,Ethiopia. Working Paper, BASIS CRSP, Universityof Wisconsin.

Barrett, C., & Carter, M. R. (2006). Social protectionpolicy to overcome poverty traps and aid traps.Paper presented to Brookings conference on assetbased approaches to poverty reduction in a globalizedcontext, Washington, DC, 27–28 June 2006.

Barrett, C. B., Marenya, P. P., McPeak, J. P., Minten,B., Murithi, F., Oluoch-Kosura, W., et al. (2006).Welfare dynamics in rural Kenya and Madagascar.Journal of Development Studies, 42(2), 248–277.

Boucher, S., Barham, B., & Carter, M. R. (2005). Theimpact of market friendly reforms on credit and landmarkets in Honduras and Nicaragua. World Devel-opment, 33(1), 107–128.

Buera, F. (2005). A dynamic model of entrepreneurshipwith borrowing constraints. Working paper, North-western University.

Calvo, C., & Dercon, S. (2005). Measuring individualvulnerability. Working paper, Oxford University.

Carter, M. R., & Barrett, C. B. (2006). The economics ofpoverty traps and persistent poverty: an asset-basedapproach. Journal of Development Studies, 42(2),178–199.

Carter, M. R., & Castillo, M. (2005). Morals, marketsand mutual insurance: using economic experimentsto study recovery from Hurricane Mitch. In C. B.Barrett (Ed.), Exploring the moral dimensions ofeconomic behavior (pp. 268–287). Oxon, UK: Routl-edge.

Castro, A. P., Amare, Y., Adal, Y., & Tolossa, D.(1999). BASIS/IDR community assessments: Kebeleprofiles. Working Paper, BASIS CRSP, University ofWisconsin.

Dercon, S. (2004). Growth and shocks: evidence fromrural Ethiopia. Journal of Development Economics,74(2), 309–329.

Dreze, J., & Sen, A. K. (1989). Famines and socialresponse. In J. Dreze, & A. Sen (Eds.), Hunger andpublic action (1st ed., pp. 65–81). Oxford, UK:Clarendon Press.

Fafchamps, M., Udry, C., & Czukas, K. (1998).Drought and saving in West Africa: are livestock abuffer stock? Journal of Development Economics,55(2), 273–305.

Government of Ethiopia (1997). Food supply prospect1998: early warning system report. Addis Ababa,Ethiopia: Disaster Prevention and PreparednessCommission.

Government of Ethiopia (1998a). Food supply prospect1999: early warning system report. Addis Ababa,Ethiopia: Disaster Prevention and PreparednessCommission.

Government of Ethiopia (1998b). Food situation in 1998and assistance requirement. Addis Ababa, Ethiopia:Disaster Prevention and Preparedness Commission.

Hansen, B. (2000). Sample splitting and thresholdestimation. Econometrica, 68(3), 575–604.

Hoddinott, J. (2006). Shocks and their consequencesacross and within households in Rural Zimbabwe.Journal of Development Studies, 42(2), 301–321.

Little, P.D., Ahmed, A. G. M., Carter, M. R., & Negatu,W. (2002). Building assets for sustainable recovery andfood security. BASIS Brief, No. 5. Madison, WI:BASIS Research Program, Department of Agricul-tural and Applied Economics, University of Wiscon-sin-Madison.

Little, P. D., Stone, M. P., Mogues, T., Castro, P., &Negatu, W. (2006). ‘Moving in place’: drought andpoverty dynamics in South Wollo, Ethiopia. Journalof Development Studies, 42(2), 200–225.

Lopez, R., & Valdes, A. (2000). Fighting rural poverty inLatin America: new evidence of the effects ofeducation, demographics, and access to land. Eco-nomic Development and Cultural Change, 49(1),197–211.

Lybbert, T., Barrett, C., Desta, S., & Coppock, L.(2004). Stochastic wealth dynamics and risk man-agement among a poor population. Economic Jour-nal, 114(498), 750–777.

Mathys, E. (1999). Assessment of the impact of food aidon household economies of North Wollo, South Wolloand East Hararghe, Ethiopia. Food Security Unit,Save the Children Fund, UK.

McPeak, J. (2004). Contrasting income shocks with assetshocks: livestock sales in Northern Kenya. OxfordEconomic Papers, 56(2), 263–284.

Mogues, T. (2005). Shocks, livestock asset dynamics,and social capital in Ethiopia. Working paper.

Mogues, T., & Carter, M. R. (2005). Social capital andthe reproduction of inequality in polarized societies.Journal of Economic Inequality, 3, 193–219.

Morris, S., Neidecker-Gonzales, O., Carletto, C., Mun-guia, M., Medina, J. M., & Wodon, Q. (2001).Hurricane Mitch and livelihoods of the rural poor inHonduras. World Development, 30(1), 39–60.

Quisumbing, A. (2003). Food aid and child nutrition inrural Ethiopia. World Development, 31(7), 1309–1324.

Zimmerman, F., & Carter, M. R. (2003). Asset smooth-ing, consumption smoothing and the reproduction ofinequality under risk and subsistence constraints.Journal of Development Economics, 71, 233–260.

poverty traps and natural disasters in ethiopia and honduras

Documents