Donors Versus Implementing Agencies

Get Complete Project Material File(s) Now! »

Food aid in Ethiopia

Ethiopia has been one of the world’s major recipients of international food aid for decades. As a result, over the last twenty years, food aid has amounted to one-tenth of domestic pro-duction in Ethiopia (Planel, 2005). For wheat, a major staple in the country, food aid has even reached 40 percent of domestic production. 3
Ethiopia has faced a major shift in food aid policy in the mid 2000s. Before that date, food aid was basically repeated emergency interventions. While those interventions were successful in terms of alleviating starvation, they did not prevent asset depletion and were not integrated in agricultural development activities (Berhane et al. , 2014).
Against this background, a number of policy changes have occurred. First allocation criteria of free food aid were reformed(DRMFSS, 1995, 2003). Before 2003 those who used to be eligible for free food delivery were the elderly, disabled persons, lactating or pregnant women, and hou-sehold members attending to young children. In 2003, the Disaster Risk Management and Food Security Sector revised the official guidelines and introduced the Household Economic Approach. This method is based on a survey that assesses hazard probability and coping strategies at the household level. For instance, it takes into account resources available to the household, such as assets (livestock) or relatives who could give transfers. While the Household Economic Approach is based on sound economic theory, it is hard to apply on the ground, partly because the ins-titutional channels through which aid is actually allocated have hardly changed (Shoham, 2005).
Secondly starting 2004-2005, Ethiopia, with the help of donors, implemented the Productive Safety Net Program (PSNP). This multi-year program seeks to prevent asset depletion at the household level and build assets at the community level ; it also ensures timely and predictable cash and/or food transfers to chronically food-insecure people. 4 The program covers now more than 50 percent of the communities (woredas) in the country. 5 The shift from “annual emergency aid » to an integrated safety net approach is likely to have influenced households’ marketing behavior and is worth studying over time.

Related studies on Ethiopia

As one of the countries most dependent on food aid, Ethiopia has been the focus of nume-rous studies.
A first stream of work focuses on targeting and dependency. Aid allocation in Ethiopia re-sults from a three-step process where the government decides the geographical allocation of aid at the regional level, regional leaders decide the allocation of aid by woreda, and local leaders at the Peasant Association (PA) level select households within each community. All steps are subject to inefficiency and potential political capture. According to Clay et al. (1999), Jayne et al. (2001) and Enten (2008), allocation at the woreda level results from negotiations between the government, the administrative staff and local communities and, as a result, is not (entirely) related to effective needs. These three papers, although written ten years apart, show that the Tigray region has been favored because of its close ties to the government. 6 At the local level, re-cipient households with political connections and involved in village organizations receive more food aid than recipient households without connections (Broussard et al. , 2014). The system perpetuates itself, as PA leaders who are elected are reportedly manipulating the election, by threatening voters that they will be excluded from federal support (Human Rights Watch, 2010).
Two consequences emerge from these papers on allocation process in Ethiopia. First, tar-geting is likely to be imperfect. Only 22 percent of food-insecure people received some aid ; this comes either because their district was not targeted or because their household was not selected (Planel, 2005). As allocation under the new Household Economic Approach relies less than before on easily households’ observable characteristics such as age and gender, it may be subject to political capture.
Second, because of political stability in Ethiopia, with a national coalition staying in power for many years, it is likely that the same politically-connected households have received aid over time. Hence, the part of the selection that is based on unobservable households characteristics such as political connections may be considered as time-invariant.
Political capture is not the only culprit of poor targeting. The fixed costs of setting operations and identifying needs also account for the inertia of food aid allocation. Jayne et al. (2002) show, based on a nationally representative rural dataset of 1996, that the spatial allocation of aid in 1996 is highly correlated with the spatial pattern of vulnerability in 1984 during the famine and is concentrated in areas that are not the poorest. The inertia is particularly prevalent for food-for-work, possibly because the latter is often a multi-year program.
Asfaw et al. (2011) investigate the determinants of participation in food aid programs and the impact of such programs on poverty reduction, based on the ERHS surveys from 1999 and 2004. They show that households’ size and asset endowments determine the extent of poverty alleviation and food aid dependency. Based on quantitative and qualitative data from 1999-2000 and 2002-2003, Little (2008) finds that food aid plays a significant role in households’ recovery strategies, without creating dependency. This is due to the fact that aid deliveries are poorly timed and come with uncertainty. Bevan & Pankhurst (2006) have conducted interviews in 20 villages, including the villages surveyed in the ERHS. Their study gives a sense of attitudes towards aid. Respondents mention that aid in the long-term can make « people lazy ». They also claim that food aid may come too late, is insufficient and distributed in centers that are too far away.
A second stream of literature investigates the impact of food aid on food prices and food production. Levinsohn & McMillan (2007) argue that the impact of aid on poverty depends on its effect on prices and on the household being a net seller or a net buyer. Based on two nationally representative household surveys in 1999-2000, they estimate the welfare impact of a change in prices and infer the impact of food aid on prices using a partial equilibrium model. They find that aid is alleviating poverty in the short term, as net buyers are more numerous than net sellers and poorer. Kirwan & McMillan (2007) extend the time span of the previous analysis to the period 1970-2003 and use indirect evidence based on aggregate data on production and prices. They find no correlation between food aid and producer prices, the latter declining stea-dily after 1984 while food aid, mostly driven by variations in the US price of wheat, has been volatile. As a consequence, food aid might have an impact on long-run production, not through prices but because of uncertainty about shipments that might have deterred investment in the wheat sector. Re-examining the relationship between aid and prices, Assefa Arega & Shively (2014), using monthly data over 2007-2010, do not find an impact of food aid on local producer prices of wheat, teff and maize in Ethiopia.
Using a computable general equilibrium model calibrated to Ethiopia in 2000, Gelan (2006) finds a disincentive impact of food aid on domestic food production. Removing food aid stimu-lates demand and generates an expansion of the food producing sector with a slight increase in producer prices. In general equilibrium, consumers would substitute between grains ; as house-holds receive wheat for free, they would shift away from maize or teff, hurting not only wheat growers but also the producers of other cereals.
Abdulai et al. (2005) re-examines the impact of aid on food production both at the micro level on Ethiopia and at the macro level with a VAR model estimated on 42 Sub-Saharan Afri-can countries. They do not find evidence of a disincentive impact with either method. If any, the macro analysis tends even to find a positive impact of food aid on production one or two years later. The micro analysis is based on two rounds of the ERHS in 1994 and one in 1995. They estimate the impact of receiving food aid on various outcomes: labor supply (of various sorts: on-and off-farm, wage work and own business, male and female), agricultural investment and use of inputs, and informal labor sharing. Some of these outcomes are of a 0/1 type. Others are conti-nuous with zero values, and are estimated with a tobit. A naive estimation finds that aid has a strong disincentive impact. However, once controlled for household characteristics that might explain aid allocation (location, age, gender and education of household head, household size and holdings of land and oxen), only one impact remain significant (and positive), on off-farm female wage work. Then they estimate a model where aid is endogenous and is instrumented by past aid, reflecting an inertia effect as in Jayne et al. (2002). Aid received one year before, in early 1994, has a small disincentive effect on family labor supply for permanent and semi-permanent crops. On the contrary, contemporaneous aid (that received in 1995) has a positive impact on the same type of labor supply. Both past and current aid increase male labor supply of off-farm work. Overall, these findings make a very convincing case on the convergence of macro and micro analysis and the non existence of disincentive effects of aid in Ethiopia, at least in the short run.
Recent papers focus on the PNSP (Gilligan et al. , 2009; Hoddinott et al. , 2012; Berhane et al. , 2014) using propensity score matching and difference-in-difference estimations. The propensity score matching is based on observable households’ characteristics. It also accounts for unobservable characteristics at the village level, as it compares treated and control hou-seholds from the same woreda. The first paper finds a weak impact of PNSP in its first year of implementation in 2006 because of delays and under-payment of transfers. Aid recipients tend to increase their livestock suggesting a positive impact of aid on production. The second paper finds a positive impact of the PSNP on agricultural inputs use, especially when it is coupled with the OFSP extension program. The third paper considers treatment as continuous: it is the number of years of PNSP transfers. The paper compares the outcome between one and five years of PNSP. The propensity score is based on the demographic characteristics of the households before the program. Aid has a positive effect on food security and livestock holdings.
A third direction in the literature compares the different types of aid (Yamano et al. , 2000; Gilligan & Hoddinott, 2007; Bezu & Holden, 2008). For instance, food-for-work (FFW) target household members that are able to work and provide them a job with payments usually in-kind. If work requirements are harsh, not all eligible households enroll in the program, thus, there is self-selection on top of eligibility criteria. By contrast, free distribution is aiming at those that cannot work, children or elderly people. Bezu & Holden (2008) finds that food-for-work has encouraged the adoption of fertilizer in Tigray in 2001. They estimate a Heckman two-step model where first the household decides whether to adopt fertilizers, before deciding the actual quantity, conditional on selection. Gilligan & Hoddinott (2007) compares two programs that were expanded after the 2002 drought, free food distribution (FFD) of the « Gratuitous Relief », and food-for-work (the « Employment Generation Scheme » or EGS). They use the 1999 and 2004 waves of the ERHS and estimate a propensity score. They find that EGS participants had significantly lower growth of livestock holdings ; the effect is partly driven by outliers (some hou-seholds with large livestock in the control group). Households could also have decreased their precautionary saving as they felt protected and insured by aid. On the other hand, free food distribution was better targeted and smaller in size and had no significant impact on livestock.
Yamano et al. (2000) also distinguish between FFW and free distribution and examine their impact on purchases and sales separately. They argue that looking at net sales is not sufficient in order to assess the impact on local markets. They find that FFW decreases the purchase of wheat, while free distribution decreases the level of sales albeit the effect is small and not statistically significant.
To push their argument one step further, one would like to examine other types of market participation, such as households that grow wheat for their own consumption. Moreover, aid might not only influence the quantities sold or bought, but also the 0/1 decision of the type of market participation, for instance, determining producers who were growing wheat for their own consumption, to sell on the local market. Moreover, Yamano et al. (2000) were not controlling for the endogeneity of aid allocation. Last, we would like to take advantage of a panel stretching over 1994 and 2009 and contrast the short-run and the long-run impact of aid as well as look for any change in households’ behavior following the reform of aid policy in Ethiopia in the mid 2000s.
7. Another differentiating characteristic of food aid is whether it is sourced from local or regional procurement (LRP) or shipped from overseas. Lentz et al. (2013) show that LRP aid reduces delay and improves the adequacy between needs and shipments ; thus, it should reduce the risk of disincentive effect. Violette et al. (2013) show that LRP is more culturally accepted. This may reduce the negative impact on markets, as households are more likely to consume LRP aid instead of selling it, a consequence that is not investigated in their paper. Garg et al. (2013) examines the potential price effect of LRP aid and do not find any statistically significant impact. Unfortunately the EHRS does not provide information on the type of procurement. At the country level, one quarter of aid in wheat comes from local purchases (INTERFAIS-WFP).

Data and descriptive statistics

Our data comes from the Ethiopian Rural Household Survey Dataset (EHRS), a longitudi-nal survey which covers some villages between 1989 and 2009. The survey results from a joint project between Addis Ababa University, the CSAE at the University of Oxford and IFPRI. The data are not nationally representative but account for the diversity of non-pastoral farming systems in the country (see Dercon & Hoddinott (2009) for more details). The survey gives information on household characteristics, agriculture and livestock, food consumption, transfers and remittances, health, women’s activities, and information at the village level on electricity and water, health services and education, wages, production and marketing.
Most of the results of this paper are based on a balanced panel of 1215 households in 15 villages, followed over 5 rounds (in 1994, 1995, 1999, 2004 and 2009). 8 In the robustness checks, we also run the estimations on the whole (unbalanced) sample.
Table 1.1 provides descriptive statistics of the sample. The poverty rate was 48.2 percent in 1994, decreased in the late 1990s and early 2000s, but has returned to its previous level in 2009. Households are cultivating 1.5 hectares on average. The worst harvest took place in 1995 with only 533 kgs of wheat produced by the average household, and the best in 2009 with a production three times higher. The size of livestock holdings has increased continually since 1994 and reaches an average of 5 tropical livestock units in 2009 (one tropical livestock unit – TLU – equals 1 cow or 10 goats or 11 sheep or 100 chickens).
The variables of interest are whether a household has received free food aid or food-for-work, and the quantities received. We focus on one crop, wheat, which is one of the major cereals in Ethiopia. From the mid-1990s, wheat consumption has increased steadily in both urban and rural areas and wheat has become one of the top priority crops deemed to solve food security challenges in the country (Tefera, 2012). Thus, a large share of food aid is provided in wheat (74 percent in our sample). 9
The share of recipients is highly variable: only seven percent of households received free food aid in 1995 whereas almost 30 percent did so in 2009. 10 Hence, on average, only one third of beneficiaries receives a transfer again in the next round (Table 1.2).
The share varies between villages as well, from zero to almost 80 percent. Quantities of wheat received per household vary from 30 kilograms in 1995 to 100 kilograms in 1999. The share of household benefiting from food-for-work programs was stable during the 1990s at around 10–11 percent. It has doubled after 2004. 11
Looking at targeting criteria (Table 1.3), recipient households have fewer and older members. They have fewer children on average, though we would have expected the opposite, given the official allocation guidelines before 2004. Food-for-work and free food aid recipients seem to differ in terms of agricultural assets and household composition. Households receiving free food are smaller than those receiving food-for-work but have more old-age members. Food-for-work households, as expected, cultivate less land than other households and have less livestock.
Regarding households market participation, we define four groups. First, households can be wheat buyers or sellers (these categories are defined in gross terms). They can grow wheat for their own consumption, without selling or buying it: these households are called « autarkic ». Finally, they can be « non-involved » (in any wheat-related activity), meaning that they neither produce nor buy wheat. Household are considered as producers if they sow wheat, even if they get no harvest.
All four types of market participation are present in Ethiopia. 12 The share of households cultivating wheat (for their own use or to sell) increases over time, going from 24 percent in 1994 to 32 percent in 2009. 11 percent of households were sellers and 18 percent buyers in 2009 ; 20 percent were in autarky and 55 percent were « non-involved ». As buyers and sellers are defined in gross terms, they might overlap (as some households are doing both) but these are in very small number, making up less than four percent of the sample.
Households’ market participation status is not stable across rounds (Table 1.4). Transition happens mostly between buyers and non-involved households, and to a lesser extent between sellers and autarkic households. In addition, only three-quarter of households that have grown wheat at time t cultivate it again at time t + 1 (not reported in the Table).
In the descriptive statistics, food aid recipients differ from other households in terms of their market participation status. Beneficiaries are more likely to be non-involved in wheat-related activity and less likely to be autarkic households or sellers (Table 1.5). They are as likely to buy wheat. Regarding quantities, aid recipients produce less (the difference being significant at 1 percent level of confidence for autarkic households) ; they also buy more wheat. However, they sell as much as non-recipient households. How much of these differences come from selection and endogenous aid allocation and how much could be triggered by aid itself?
[Table 1.5 here]

READ  The PWA-MIP/HASI/Huygens instrument, revisiting the data collected on the surface of Titan 

Empirical specification

On production

We model simultaneously the production decision and the quantity produced. We allow food aid to affect both stages differently. We run a panel Tobit type II with selection and endogeneity (Semykina & Wooldridge, 2010). The model is defined as:
yit = yit∗ ∗ sit
sit = 1sit∗>0 (1.1)
= xitβ + γ F Ait + ci
y∗ 1 + uit
it 1 1
sit∗ = xitβ2t + z1itδt + γ2tF Ait + ci2 + uit2
where yit is the quantity of wheat produced in year t by household i, and sit is the househol-d’s 0/1 decision to produce. Both are observed if s∗it, the latent variable that drives production decision, is higher than 0. yit∗ is the latent variable that drives the level of production. F Ait is the quantity of wheat aid received in the last 12 months by each household (through free food dis-tribution and/or food-for-work programs). If food aid is well targeted, we should expect that it affects neither the decision to produce nor the quantities. ci1 and ci2 are households fixed effects.
xit are control variables, meant to capture market conditions and transactions costs: whe-ther there is a daily market within the Peasant Association, the distance to the nearest market and regional dummies. We also add consumption and production shifters such as household size (per adult equivalent), the age of the household’s head, whether the household is poor, the amount of non-food consumption, livestock size (in tropical livestock units) and the size of the cultivated plot (in hectares). The household size takes into account the fact that larger families can allocate more labor on their plots. The size of livestock holdings matters in two ways for cereals production: first, part of the harvest is used to feed the livestock ; second, manure is used as a fertilize and may improve harvest. Poor households may invest less in agriculture because they are cash-constrained ; they may also grow less risky crops (hence, often less productive) in order to reduce risk.
We also include observable households characteristics that explain food-aid allocation and affect both productivity and the demand for food, such as the share of women, children and elderly within the household.
We control for climatic shocks by including the monthly average level of rainfall during the planting and growing season, measured in nearby meteorological stations and interpolated in a 50 km2 grid around each village. 13 Last, we take into account health shocks, namely, whether a household member was sick during the previous month. We cannot control for all health shocks that could have occurred during the growing season, but we assume that recent illness is a proxy for previous bad health.
The estimation procedure is as follows:
— For each round, we estimate P(sit = 1|zi) = Φ zitδta + γ2tF Ait + ziξta + F Aiξ2at where zit includes xit and z1it, the excluding variable (see below). zi is the individual mean over time of zit and F Ai the average quantity of food aid received by household i over time.
— Next, we compute the inverse Mills ratio, ˆ . λit
— For = 1 we estimate a pooled two-stage least square with ∗ = + + ˆ + sit yit xitβ ziν γλit eit using , ˆ and as instruments of and the endogenous variable . zit1 λit zi xit F Ait
— Finally, we estimate the variance, applying a panel bootstrap.
Three issues arise with this type of estimation: endogeneity of food aid allocation, reverse causality, and the need to have an exclusion variable that differentiates between the decision to produce and the level of production.
First, on the issue of selection, the distribution of food aid is not random, because of tar-geting. We first control for time-varying observable characteristics that drive aid allocation as stated in the official guidelines (such as poverty status or household composition). Selection is also driven by unobservable characteristics such as political connections. We assume that se-lection on these unobservable characteristics is time-invariant. This is a reasonable assumption because the institutional setting of food aid allocation in Ethiopia has been stable over time, as well as the political affiliation of PA leaders who ultimately allocate aid to a given house-hold within a village. Hence, we assume that PA leaders favor the same households and do not switch aid recipients before and after an election. If the assumption holds, we can exploit the panel nature of the data in order to control for individual time-invariant unobservable heteroge-neity in the distribution of aid. More precisely, we assume that endogeneity is conditional on the unobserved fixed household effect, ci2, only through the time averages of households’ variables. 14
The equations above estimate the impact of food aid received by a household in year t on its production later that year. Could it be the case instead, that food aid reacts to agricultural output of the same year, raising the concern of reverse causality?
Actually, the nation-wide amount of aid in a given year is based on previous year’s total production. The government estimates in the last quarter the number of households by woreda that are likely to be in need during the upcoming year and calls for pledges by international donors. Then, at the local level, aid is mostly distributed to households during the lean season, between April and September. It is also during these months that households decide what to produce during the main harvest (meher), which starts in October and goes through December. Normally, there is no aid distributed during the meher harvest. 15 There is also a minor harvest in May and June (belg), during which wheat is not grown.
Rain in June and July is a good predictor of the meher harvest to come, and food aid could be adjusted accordingly. In practice, however, the amount of aid is not revised before August and due to regular delays (DRMFSS, 2012), actual quantities given to households start to be affected only in September and more often after January (see Figure A1.1 in 1.8 for an illustra-tion of the calendar in 2004). Hence, because of this timing, it seems that the concern of reverse causality can be ruled out.
Moreover, most results will present the impact of current aid on current production, but we will also look at the impact of aid received over all rounds between 1994 and 2009 (a kind of dependency effect), as well as the impact on the outcome in 2009 of aid received in all years but the last (a lagged effect of aid).
For our estimation strategy to hold, we need an excluding variable z1it that explains the decision to produce but not the level of production. We choose for that purpose the share of religious holidays during the planting season as agricultural activity is banned during these days. According to farmers’ interviews it seems the prohibition is actually binding (Bevan & Pankhurst, 1996):
People have the strict belief that if they go to the field on these religious holidays, they will be punished by God. If they see someone working on these days, he will be admonished by elders of the community. If he insists on working on these days, he will be condemned and ostracized by the community. – Debre Birhan PA (p.24).
People go to church on [holy] days to pray and attend religious ceremonies. If anyone is found working (ploughing, harvesting or weeding) on these religious holidays he is criticized by the community and must pay a fine to the church. In some cases this situation may be a fundamental obstacle in the production process. For example this year the belg rain [. . . ] came on April 14th. The following days [. . . ] were the usual holidays [. . . ]. Then followed the week-long holiday of Himamat [. . . ]. Then came [. . . ] a holiday of four consecutive days making the total number of holidays 16 consecutive days. That means that if the rain only lasted two weeks the Shumsheha peasants could not plough, and therefore would not get any belg harvest. Peasants referred to this year’s belg as yeslam belg (Muslim’s belg) indicating that only the Muslims could plough. – Shumsheha PA (p.22).
Muslims fast for a month during Ramadan. Since they stay awake during the nights they do not work effectively on most of the days. – Imbidir PA (p.22).
As not all religious holidays are on a fixed calendar day, their number varies by year 16 ; moreover, the planting season also varies across years and regions. Hence the excluding variable is differentiated across households, years and regions.
We also include household’s religion directly in the estimations, as it could affect production decisions and productivity through difference in preferences, ritual fasting or prohibitions. We assume that once controlled for the direct impact of religion, the share of religious holidays only affects the planting decisions. This is in line with Kijima & Gonzalez (2013) who find that in Madagascar, religion does not affect agricultural productivity but determines the choice of crops. 17 Table 1.6 shows the expected probability of being a wheat producer, depending on the household’s religious group and the share of days-off during the planting season (controlling for other household characteristics). The probability is lower for Muslims than for other religious groups, and decreases with the share of holidays.

Table of contents :

General Introduction 
1 Does Food Aid Disrupt Local Food Market? Evidence from Rural Ethiopia 
1.1 Introduction
1.2 Context
1.2.1 Food aid in Ethiopia
1.2.2 Related studies on Ethiopia
1.3 Data and descriptive statistics
1.4 Empirical specification
1.4.1 On production
1.4.2 On sales and purchases
1.5 Results and analysis
1.5.1 On production
1.5.2 On sales and purchases
1.5.3 Robustness checks
1.6 Conclusion
1.7 Figures and tables
1.8 Appendix
2 Donors Versus Implementing Agencies: Who Fragments Humanitarian Aid? 
2.1 Introduction
2.2 Humanitarian aid: data and descriptive statistics
2.2.1 Data
2.2.2 Descriptive statistics
2.3 Fragmentation of humanitarian aid
2.3.1 Indicators of aid fragmentation
2.3.2 Donor and implementing agency fragmentation
2.4 Delegating aid and its fragmentation: potential consequences
2.4.1 Positive impacts of delegation and fragmentation on aid efficiency
2.4.2 Negative impacts of delegation and fragmentation on aid efficiency
2.5 Three case studies of implementing agency fragmentation
2.5.1 Haiti 2010: the burden of fragmentation
2.5.2 Pakistan 2010: a useful fragmentation
2.5.3 Sudan 2010: the leading role of the UN
2.6 Conclusion
2.7 Figures and tables
3 To Give or Not to Give? How Do Donors React to European Food Aid Allocation? 
3.1 Introduction
3.2 Empirical strategy
3.2.1 Specification
3.2.2 Instrumental strategy
3.2.3 Potential concerns
3.3 Data and descriptive statistics
3.3.1 Food aid statistics
3.3.2 Controls
3.4 Empirical results
3.4.1 Baseline results
3.4.2 Bilateral reactions
3.4.3 Placebo tests and robustness checks
3.5 A donor typology
3.5.1 Setting
3.5.2 Reaction function
3.5.3 Typology
3.6 Conclusion
3.7 Figures and tables
3.8 Appendix
General Conclusion 


Related Posts