Академический Документы
Профессиональный Документы
Культура Документы
Contract farming has often been associated with an increase in the income of participating house-
holds. It is unclear, however, whether contract farming increases other aspects of household welfare.
We use data from six regions of Madagascar and a selection-on-observables design in which we con-
trol for a household’s marginal utility of participating in contract farming, which we elicited via a
contingent valuation experiment, to show that participating in contract farming reduces the duration
of a household’s hungry season by about eight days on average. Moreover, participation in contract
Key words: Contract farming, outgrower schemes, grower-processor contracts, agricultural value
chains, food security.
JEL codes: L24, O13, O14, Q12.
Although the benefits of economic specializa- two crops for sale—is the emergence of an in-
tion have been widely understood since the termediate sector between the agricultural
publication of Adam Smith’s 1776 Wealth of and manufacturing sectors. The institution
Nations (1976), if not earlier, a persistent lack that perhaps best represents the emergence
of specialization is one of the prime factors en- of such an agro-industrial sector is contract
abling economic underdevelopment in most of farming, or the economic institution wherein
the world’s poorest countries. In such coun- a processing firm contracts the production of
tries, whose economies remain largely agrar- commercial crops out to smallholder farmers;
ian, the structural transformation, or transition this institution is the cornerstone of agricul-
from subsistence to commercial agriculture, tural value chains. In one of the earliest stud-
has so far proven elusive. ies of contract farming in economics, Grosh
One of the first steps in the transition from (1994) noted that this institution can resolve
subsistence to commercial agriculture—that several market failures that result from risk
is, the transition from many smallholder and uncertainty, imperfect factor markets,
farmers producing small quantities of several and reluctance to adopt new technology.
crops for home consumption to fewer large Since then, contract farming has been studied
farms producing large quantities of one or in many countries and across many crops, and
policy makers have often hailed the institu-
tion as a tool for rural poverty alleviation.
Bellemare is an associate professor in the Department of
But does participation in agricultural value
Applied Economics and Director, Center for International Food chains make people better off? Although an
and Agricultural Policy, University of Minnesota. Novak is a important body of literature explores the ef-
PhD candidate in the Department of Applied Economics,
University of Minnesota. The authors are grateful to the
fects of participation in contract farming on
National Institute of Food and Agriculture for funding this work household income or some variant thereof
through grant MIN-14-061, “Smallholder Participation in (Porter and Phillips-Howard 1997; Singh
Agricultural Value Chains: Evidence from Madagascar.” The au-
thors thank Teevrat Garg for invaluable help getting the data or- 2002; Warning and Key 2002; Simmons 2005;
ganized. We also thank Benjamin Wood, Ali Hill, Travis Maertens and Swinnen 2009; Minten,
Lybbert, Matthieu Stigler, two anonymous reviewers, seminar Randrianarison, and Swinnen 2009; Miyata,
participants at Cornell, as well as participants at the 2014 AAEA
annual meetings, the 2014 POLICOFA conference in Dar es Minot, and Hu 2009; Rao and Qaim 2011;
Salaam, and the 2015 CSAE conference at Oxford for comments Barrett et al. 2012; Bellemare 2012;
and suggestions that made for a much-improved manuscript. All
remaining errors are ours. Correspondence to be sent to:
Michelson 2013; Narayanan 2014), we study
mbellema@umn.edu. whether participation in contract farming
improves food security, defined here as the Discussions have incorporated think-
reported duration of the hungry season expe- ing around value chain frameworks,
rienced by a household, that is, the length of which emerged in the late 1990s to
time during which one member of the house- help development actors design in-
hold or more goes without three meals a terventions that responded to the
day.1,2,3 This question is important for three needs of the private sector and con-
reasons. First, because the hungry season co- tributed to development outcomes.
incides with those months before households Value chain approaches can provide
obtain cash for their crops (both contracted useful frameworks to examine the
and not contracted) at harvest, it is not imme- food system and the potential to
diately obvious that the households involved achieve improved nutritional out-
variable in a two-stage least squares design to Ordinary Least Squares and Duration
obtain a local average treatment effect. Models
Our core regression results suggest that par- The core equation we estimate is
ticipation in contract farming decreases the re-
ported duration of the hungry season by ð1Þ yi ¼ a1 þ b1 xi þ c1 Di þ ei
approximately eight days (i.e., 0.28 months)
for the average household in our data; propen- where yi 0 is the reported duration of the
sity score matching results are largely consis- hungry season experienced by household i
tent with those regression results. Our hazard measured in months, xi is a vector of control
and duration model results suggest that partic- variables (which, in a slight abuse of notation,
ipation in contract farming increases the likeli- also includes district dummies), Di is a vari-
months a household’s most recent hungry consider two things. The first is the number
season lasted—we use three distinct estima- of non-participating households to match to
tors to estimate equation (3). The first is the participating households. When matching
ordinary least squares (OLS) estimator, with replacement, matching only one house-
wherein c tells us how much shorter the hun- hold raises the likelihood that the two
gry season is, on average, for households that matched households are very similar.
participate in contract farming. The next two Increasing the number of matched house-
estimators are the Cox proportional hazards holds can decrease the similarity between
model and the survival time regression matched households but increase the pool of
model, two workhorse estimators used in the households upon which we draw inferences.
study of duration data (Lancaster 1992).7 In The second important consideration is the
The contingent valuation experiment used respondent is asked whether he would partic-
in this article is the same as that used in ipate in a contract farming agreement costing
Bellemare (2012). Each respondent was $62.50, we have coded all other amounts,
asked whether he would participate in a con- $12.50, $25.00, $37.50, $50.00, and $75.00, as
tract farming agreement that would raise his “No.” In order to remedy this shortcoming,
income by 10% in exchange for a one-time in an additional set of estimations we enforce
monetary investment. The amount of the monotonic switching on the part of our re-
monetary investment was randomly selected spondents. That is, if a respondent answers
from six investment amounts of $12.50, “Yes” to participating in the hypothetical
$25.00, $37.50, $50.00, $62.50, or $75.00.8 contract farming agreement for a given ran-
The size of investment was determined at domly selected investment value, we code all
farming. The next section explains why this be more likely to participate in contract farm-
constitutes a selection-on-observables re- ing. This would compromise the identification
search design in the context of regression or, of the ATE, and it could definitely be a con-
alternatively, why it satisfies the conditional cern in our application given that households
independence assumption in the context of with better access to food may be more will-
matching. ing to enter into contract farming agree-
ments. It should be the case, however, that a
Identification respondent who is more willing to enter into
a contract farming agreement because he is
How does a set of proxies for a respon- more food secure will have a higher marginal
dent’s marginal utility from participating in utility of participating in contract farming.
of contract farming commissioned by the and though the piece rate paid to growers by
World Bank. The data cover six regions and processors tends to be fixed, there was one
two communes per region. Three of these re- processor in particular that paid a floating (i.
gions were chosen because they exhibited a e., market conditions-driven) piece rate.
relatively high prevalence of contract farm- Most processors are Malagasy companies,
ing; the other three were chosen because the with the notable exception of Sodexo, a
government of Madagascar viewed them as French company that is perhaps best known
high-priority areas for economic develop- for operating cafeterias in universities, hospi-
ment. In all regions, the two communes with tals, and other workplaces.
the highest density of contract farming were We present descriptive statistics for our
surveyed. Commune-level data were ob- sample in table 1, broken down by whether
11
USD 1 2,000 Ariary at the time the data were collected.
12
Such idiosyncrasies are conceivable given the random as-
0.25
Distribution of Hungry Season Duration those variables. Before discussing the results
in table 2, however, we discuss the results of
.2
Table 2a. Estimation Results for OLS, Cox Proportional Hazard, and Survival-time
Regressions Omitting WTP Variables
Variables OLS Cox Survival Time
Dependent Variable: Duration of Hungry Season
Contract farming participant 0.294** 0.150** 0.171**
(0.142) (0.062) (0.070)
Household size 0.050 0.011 0.013
(0.036) (0.015) (0.017)
Dependency ratio 0.576 0.254 0.286
(0.365) (0.157) (0.181)
as well as the value of the assets owned by his In an effort to identify the mechanism
household are all associated with a shorter through which participation in contract farm-
hungry season and a greater likelihood of ing might reduce the length of the hungry
exiting the hungry season at any given time. season (in the OLS specification) or increase
According to the OLS results, an increase in the likelihood of exiting the hungry season
household income in associated with a (in the Cox proportional hazards and survival
shorter hungry season, while income is statis- time regression specifications), we estimated
tically insignificant in the Cox proportional additional specifications (not shown) similar
hazards and survival time regression results. to those in table 2 but in which we include (a)
This lends credence to our claim that it is in- an additional control variable that measures
sufficient to infer nutritional outcomes based the proportion of all crops under contract
solely on income changes. Lastly, though the that are rice crops, or (b) an additional con-
contingent-valuation dummies are not indi- trol variable that measures the proportion of
vidually significant in any of the models pre- all crops under contract that are food crops.
sented in table 2, they are jointly significant Including those additional controls allows us
at less than the 10% level for the OLS model. to determine whether participation in
368 March 2017 Amer. J. Agr. Econ.
Table 2b. Estimation Results for OLS, Cox Proportional Hazard, and Survival-time
Regressions Including WTP Variables
Variables OLS Cox Survival Time
Dependent Variable: Duration of Hungry Season
Contract farming participant 0.277* 0.166*** 0.188***
(0.145) (0.063) (0.071)
Household size 0.052 0.013 0.015
(0.036) (0.015) (0.017)
Dependency ratio 0.517 0.226 0.247
(0.366) (0.158) (0.181)
contract farming improves food security via leakage (whereby contract farming partici-
some sort of positive spillover (whereby con- pants might steal some of the crops they grow
tract farming participants become more pro- under contract in order to eat them). In no
ductive on their own food crops because they case was the coefficient on those interaction
grow food crops under contract) or via terms significant at less than the 10% level,
Bellemare and Novak Contract Farming and Food Security 369
which allows us to rule out those female children? Though our data do not al-
mechanisms. low us to determine the precise mechanism
behind these findings, it is not unlikely that
Treatment Heterogeneity because children—specifically girls—require
fewer calories, we may see that the addition
We now turn to treatment heterogeneity by of calories in the household creates a larger
number of children and by number of chil- reduction in the number of skipped meals for
dren of each gender in the household. Table households with more children. In other
3 shows estimation results for OLS, Cox pro- words, the marginal welfare impacts of partic-
portional hazards, and survival time regres- ipating in contract farming will be highest for
sion models in which the treatment variable children, specifically girls, and so it is not sur-
Table 3. Estimation Results for OLS, Cox Proportional Hazard, and Survival-time
Regressions Exploring Treatment Heterogeneity I
Variables OLS Cox Survival Time
Dependent Variable: Duration of Hungry Season
Contract Farming Participant 0.210 0.009 0.004
(0.253) (0.109) (0.125)
Contract farming participant x number of kids 0.191** 0.060* 0.070*
(0.082) (0.034) (0.039)
Number of kids in household 0.172 0.053 0.060
(0.121) (0.050) (0.057)
Table 4. Estimation Results for OLS, Cox Proportional Hazard, and Survival-Time
Regressions Exploring Treatment Heterogeneity II
Variables OLS Cox Survival Time
Dependent Variable: Duration of Hungry Season
Contract farming participant 0.206 0.005 0.013
(0.254) (0.109) (0.125)
Contract farming participant x girls 0.215* 0.118** 0.137**
(0.120) (0.054) (0.061)
Contract farming participant x boys 0.163 0.015 0.018
(0.120) (0.048) (0.054)
(ATE). As expected, the results for all three detailed in the subsection Identification
matching routines show that the largest effect Strategy. Because imputations yield gener-
(in absolute value) is for households that did ated regressors, we bootstrap the standard er-
participate, followed by the effect for all rors but omit sampling weights in
households in the sample, and lastly the ef- Supplementary online appendix tables A2,
fect for those households that did not partici- A3, and A4.13 Given the results in
pate. Recall that the estimated ATE is most Supplementary online appendix tables A2,
comparable to the estimated effects for the A3, and A4, our core results appear to be ro-
OLS results because OLS reports the ATE. bust to a change in how we proxy for respon-
The ATE ranges from -0.127 to -0.272. dent marginal utility to participate in a
This represents a reduction in the length of hypothetical contract farming agreement that
the hungry season by between four and eight would increase household income by 10%.
days—an effect that is very close to what we Similarly, when we include sampling weights
find using our OLS specification in the first but do not bootstrap the standard errors in
column of table 2b. The effect for participat- Supplementary online appendix tables A5,
ing households is larger, that is, the ATT A6, and A7, our core results appear once
ranges from -0.194 to -0.305. This is a reduc- again robust to a change in how we proxy for
tion in the length of the hungry season by six respondent marginal utility to participate in a
to nine days. hypothetical contract farming agreement that
would increase household income by 10%.
Supplementary online appendix table A8
Robustness Checks presents the results of treatment regressions
In order to ensure that our results are ro- wherein responses to the contingent-valuation
bust, we estimate a number of alternative questions are used as instrumental variables
specifications. Table 5 presents the results of for participation in contract farming, as in
two estimators that aim to minimize the ef- Bellemare (2012). Taking both the OLS re-
fects of outliers. The first specification is a sults in table 2b and the treatment regression
median regression. Intuitively, a median re- results in Supplementary online appendix ta
gression is similar to an OLS regression, ex- ble A8 at face value—that is, assuming that
cept that it focuses on the conditional median they both identify causal impacts—would sug-
rather than the conditional mean. The second gest that the local average treatment effect
specification is a robust regression (LATE; i.e., the estimated coefficient on par-
(Rousseeuw and Yohai 1987). In both cases, ticipation in contract farming in either column
results are very similar to the core OLS result of Supplementary online appendix table A8)
in the first column of table 2b. is much larger than the ATE (i.e., the esti-
In Supplementary online appendix tables mated coefficient on participation in contract
A2 to A7, we present estimation results simi- farming in the first column of table 2b). In
lar to our core results in tables 2, 3, and 4, re-
spectively, but with one important difference.
In tables A2, A3, and A4, the responses to 13
We do not show results in which we use sampling weights
the contingent-valuation questions that were and bootstrapped standard errors, because the use of the latter
not posed to the respondent are imputed, as precludes incorporating the former.
374 March 2017 Amer. J. Agr. Econ.
other words, if one is willing to believe that the reported duration of the hungry season for
both specifications are well-specified and iden- those households that were induced to partici-
tify causal relationships, one would conclude pate because they would derive a higher mar-
that participating in contract farming is associ- ginal utility from participating in the
ated with an almost two-month decrease in hypothetical contract farming arrangement.
But taking into account the potential effect of order to believe that our estimates are causal,
participating in contract farming for one must trust that our proxies for respon-
everyone—including nonparticipants—the ef- dent marginal utility of participation in con-
fect is severely moderated. In other words, the tract farming derived from our contingent
fact that the LATE exceeds the ATE implies valuation experiment fully account for the se-
that compliers (i.e., those households that par- lection process whereby households choose
ticipate in contract farming because they derive to participate in contract farming. This is an
higher marginal utility from doing so and those assumption that is untestable. Moreover,
households that do not participate because comparing the OLS specification in the first
they would not derive higher marginal utility columns of tables 2a and 2b shows that the
from doing so) derive higher benefits than de- coefficients are not statistically different
be measured much more precisely by measur- These are important results given that chil-
ing each household member’s consumption of dren, particularly girls, bear the largest bur-
calories, macronutrients (e.g., carbohydrates, den of food insecurity, the consequences of
fat, and protein), or micronutrients (i.e., spe- which include stunting, wasting, listlessness,
cific vitamins and minerals). The data used in and cognitive impairment (Alderman,
this article were not collected for the specific Hoddinott, and Kinsey 2006; Ruel and
purpose of studying food insecurity, and mea- Alderman 2013). Our results suggest that pol-
suring food insecurity accurately would re- icies that lower barriers to entering contract
quire individual-level rather than household- farming agreements for households with chil-
level survey questionnaires. dren, and particularly girls, may lead to big
gains in terms of food security.
14 15
Again, the difference between Bellemare (2012) and this A reviewer commented that households that participate in
study is twofold. First, whereas Bellemare looked at income as contract farming might see their hungry seasons start later than
his outcome variable, we look at the duration of the hungry sea- nonparticipant households. Looking into whether that is the
son as ours. Second, whereas Bellemare used respondent WTP to case, we found that the start date of the hungry season was not
participate in contract farming as an instrumental variable, we significantly different for households that do and do not partici-
use it as a control variable in a selection-on-observables design. pate in contract farming.
Bellemare and Novak Contract Farming and Food Security 377