DETERMINANTS OF LAND ALLOCATION IN A MULTI-CROP FARMING SYSTEM:
AN APPLICATION OF THE FRACTIONAL MULTINOMIAL LOGIT MODEL
TO AGRICULTURAL HOUSEHOLDS IN MALI
By
James Edward Allen IV

A THESIS
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
MASTER OF SCIENCE
Agricultural, Food and Resource Economics
2012

ABSTRACT
DETERMINANTS OF LAND ALLOCATION IN A MULTI-CROP FARMING SYSTEM:
AN APPLICATION OF THE FRACTIONAL MULTINOMIAL LOGIT MODEL
TO AGRICULTURAL HOUSEHOLDS IN MALI
By
James Edward Allen IV
Effective food security work in developing countries, such as Mali, relies on a thorough
understanding of the rural farming system. A common approach is to study land allocation
decisions to specific crops. In accomplishing this, one challenge is to model all production
outcomes in a multi-crop system. This thesis attempts to overcome this challenge in order to
study the determinants of how much of a household’s cultivated land it allocates to cotton,
maize, sorghum, millet, and secondary crops. First, incorporating insights gained by the author
while serving as Peace Corps Volunteer in Mali, the agricultural household model helps to
identify factors that explain land allocation to various crops. This framework is applied to survey
data from seven villages in Mali’s Koutiala production zone. A fractional multinomial logit
econometric model is used to estimate the effect of household and production attributes on shares
of cotton, maize, sorghum, millet, and secondary crops simultaneously, the results of which are
presented as average marginal effects. Among other results, the analysis shows that ethnic
groups not native to the Koutiala area are associated with significantly smaller shares of maize,
and that villages with better market access are correlated with much higher shares of millet and
smaller shares of maize, sorghum, and especially cotton. These results, along with personal
experience, inform recommendations for policymakers, such as the need to reduce transaction
costs for coarse grain markets, promote maize and secondary crops as nutritious and marketable
goods, and remain mindful of remote villages during the restructuring of Mali’s cotton industry.

This thesis is dedicated to the people of Simona.
Ala k’a suman tige nyuman di aw ma, sisan ani waati be.

iii

ACKNOWLEDGMENTS

I have learned that a thesis is not something you write, but something you do; so while I
may be its only author, I owe it to those who have helped me during this process. First and
foremost, I thank the people of Mali, and specifically those in Simona and Yorosso, who worked
with me while I served as a Peace Corps Volunteer. Not only did you care for my wife Joye and
me with shelter, food, and generous hospitality, but you also taught me so much about life in
general and agricultural economics in particular. Specifically, thank you to families of Dorro
Guita, Rube Dao, Tie Sanou, Kalifa Guita, and Sekouba Traore. Thank you also to the dedicated
staff of Peace Corps Mali, including Mike Simsik, Jolie Dennis, Karim Sissoko, Yacouba Kone,
and many others, as well as to the former members of Mali’s Food Security Task Force.
I also want to thank my thesis committee for their gracious efforts, especially when I
have worked on most of my thesis from afar. All have been patient as I narrowed down my thesis
topic and research questions from a long list of ideas. Specifically, Dr. Eric Crawford, my
committee chair, has delivered honest and helpful advice every step of the way and has granted
me the independence to work at my own pace from remote locations. Similarly, Dr. John Staatz
led me in an independent study on food systems organization while I was in Mali to help fulfill
my credit requirements and even treated me to a steak dinner in Bamako. Both were also
instrumental in our placement in Mali and allowing me to participate in the important food
security work that Michigan State is doing there. For his part, Dr. Soren Anderson worked with
me patiently over coffee, scones, and burritos on at least five different occasions as I developed
the models for this thesis and ultimately led me toward the fractional multinomial logit. For all
these contributions, I truly thank my thesis committee and their steadfast support.

iv

There are many other professors and researchers who have helped me along the way.
Thank you to my mentors at Alma College, including Dr. Derick Hulme, who taught me how to
write and fostered my passion for doing something about world hunger. Also thanks to Drs.
Robert Cunningham and Scott Loveridge, who both encouraged me to pursue agricultural
economics at Michigan State. Furthermore, Drs. Duncan Boughton, Valerie Kelly and Abdoul
Murekezi all delivered helpful insights as I developed my thesis topic and studied the survey data
from the Koutiala cercle. Dr. Nango Dembele, who works with Michigan State’s PROMISAM,
provided support as well as a valuable contact at CMDT’s regional office in Koutiala. Finally,
Dr. Veronique Theriault has graciously kept an open door to let me share my experiences and put
them in the context of applicable research. I am grateful for all of their collective guidance.
Friends have also been a tremendous resource. For the encouraging words and support
from my fellow students at Michigan State, I thank Megan Sheahan, Ty Lawson, Miriam
Sohlberg, Guilherme Signorini, Rachel Ford, and all those who sent that card to Mali back in
November of 2010. There are also friends from Peace Corps, namely Koutiala-kaw, who had to
hear about this thesis far too much for their liking, but were enduring listeners nonetheless.
Finally, I want to thank my family, starting with my two parents James and Debra Allen,
who have been mentoring me now for over a quarter century. The opportunities and
encouragement you have given me from day one have undoubtedly led me here. However, I save
the most heartfelt thank you for my wife Joye, who has remained supportive of me even while
carrying an intestinal parasite, working all day in unbearable heat, being evacuated from
unexpected civil conflict, moving to Kentucky and starting pharmacy school. My love for you is
longer than this thesis, and I promise to make up for the hundreds of meals you have prepared
that allowed me to “just keep working.” I could not have done this without you.

v

TABLE OF CONTENTS

LIST OF TABLES .......................................................................................................................viii
LIST OF FIGURES ....................................................................................................................ix
CHAPTER 1: INTRODUCTION ...............................................................................................1
1.1
Problem Statement ..........................................................................................................1
1.2
Literature Review ............................................................................................................3
1.3
Research Methods ...........................................................................................................8
1.4
Thesis Organization ........................................................................................................9
CHAPTER 2: BACKGROUND .................................................................................................10
2.1
Geographic Scope ...........................................................................................................10
2.2
The Farming System in the Koutiala Production Zone ..................................................13
2.2.1
Cotton ............................................................................................................14
2.2.2
Coarse Grains ...............................................................................................20
2.2.3
Secondary Crops ...........................................................................................23
2.2.4
Conclusion ....................................................................................................25
CHAPTER 3: CONCEPTUAL MODEL ....................................................................................27
3.1
Malian Farmers as Rational Actors .................................................................................27
3.2
Malian Household Maximization Problem .....................................................................29
3.3
Determinants of Crop Choice and Land Allocation .......................................................33
CHATPER 4: EMPIRICIAL APPLICATION ...........................................................................37
4.1
Data Description .............................................................................................................37
4.2
Dependent Variables ........................................................................................................41
4.3
Independent Variables ....................................................................................................43
4.3.1
Labor and Land: Household Members per Hectare and Inactivity ..............45
4.3.2
Capital: Farming Equipment ........................................................................48
4.3.3
Transport .......................................................................................................49
4.3.4
Household Characteristics: Literacy and Ethnic Identity ............................51
4.3.5
Expected Price ..............................................................................................55
4.3.6
Determinants Not Included ...........................................................................56
CHAPTER 5: ECONOMETRIC MODEL AND RESULTS .....................................................60
5.1
Fractional Multinomial Logit Model ..............................................................................60
5.2
Presentation of Average Marginal Effects ......................................................................63
5.2.1
Labor and Land: Household Members per Hectare and Inactivity ..............66
5.2.2
Household Characteristics: Literacy and Ethnic Identity ............................69
5.2.3
Capital: Farming Equipment ........................................................................72
5.2.4
Transport .......................................................................................................73
5.2.5
Time and Location ........................................................................................74
vi

5.3

Application of Results to Example Households .............................................................78

CHAPTER 6: CONCLUSIONS .................................................................................................81
6.1
Summary of Methodology ..............................................................................................81
6.2
Summary of Findings ......................................................................................................82
6.3
Policy Recommendations ................................................................................................85
6.4
Limitations ......................................................................................................................89
6.5
Future Research ..............................................................................................................91
APPENDICES ............................................................................................................................94
Appendix 1: The Household in the Koutiala Production Zone ...................................................95
Appendix 2: Fertilizer and Credit for Staple Crop Production ...................................................99
Appendix 3: Development of Agricultural Household Model ...................................................103
A.3.1
Generalized Agricultural Household Model .................................................103
A.3.2
Market Failures and the Agricultural Household Model .............................107
Appendix 4: Description of Logit, Fractional Logit, and Multinomial Logit Models ................112
A.4.1
The Logit Model for Binary Response ..........................................................112
A.4.2
The Logit Model for Fractional Response .....................................................113
A.4.3
The Logit Model for Multinomial Response ..................................................115
Appendix 5: Specification of the Fractional Multinomial Logit Model in STATA ...................117
Appendix 6: STATA Code for Calculation of Average Marginal Effects .................................119
Appendix 7: Fractional Multinomial Logit Results off of Cotton Share ....................................121
REFERENCES ...........................................................................................................................124

vii

LIST OF TABLES

Table 1: Adoption Rates (%) of Coarse Grains by Village .........................................................22
Table 2: Adoption Rates (%) of Secondary Crops by Village .....................................................23
Table 3: Description of Six Villages in Sample from Koutiala Cercle .......................................39
Table 4: Summary Statistics for Dependent Variables ................................................................42
Table 5: Summary of Independent Variables ..............................................................................44
Table 6: Average Marginal Effects of the Independent Variables .............................................64
Table A.1: Fractional Multinomial Logit Results Pivoted off of Cotton Share..........................122

viii

LIST OF FIGURES

Figure 1: Map Highlighting Mali's Sikasso Region ....................................................................11
Figure 2: Cercles of the Sikasso Region ......................................................................................11
Figure 3: Total Hectares Alloted to Cotton in Mali, 1960/61 to 2007/08....................................18
Figure 4: Cultivated Hectares of Cotton in the Koutiala Zone, 1990/91 to 2010/11 ...................19
Figure 5: Hectares of Cotton and Coarse Grains in the Koutiala Zone, 1999/00 to 2010/11 ......25

ix

CHAPTER 1: INTRODUCTION
1.1

Problem Statement
Comprehending farmers’ land allocation decisions is a significant challenge within the

field of agricultural economics and has thus been the subject of many studies. This is because
various factors, including transactions costs, risk aversion, and household preferences are
incorporated into the decision-making process. It is further complicated in developing countries,
such as Mali, a land-locked West African country where households are largely subsistent, face
greater market obstacles, and often grow three or more crops. Yet, serving as a Peace Corps
Volunteer in Mali from July 2010 to April 2012, I was to see first-hand, living in the rural village
of Simona, how important it is to understand the complexities of farming systems in order to
engage in effective agricultural development work. Certainly, in a country where 70 percent of
the population relies heavily on rain-fed agriculture for their livelihoods, it is critical that
policymakers understand how and why farmers devote certain shares of their land to growing
cotton, maize, sorghum, millet, or other field crops.
Therefore, this thesis aims to develop a relevant model and test it using empirical data
from the historical cotton basin of Mali’s southern-most Sikasso region, which can benefit
policymakers as they attempt to improve household food security. Indeed, the region has long
struggled with food insecurity despite being “the breadbasket of Mali,” with poverty rates 30
1

percent higher than any other region (USAID 2010). With a poor growing season in 2011
largely due to drought, a still unresolved civil unrest, and current threats of locust swarms, it is
almost certain that both short- and long-term interventions will be needed to diminish household
1

The phenomenon is known as the paradox of Sikasso and is discussed at length in Delarue et
al. (2009).
1

food insecurity in the future. Understanding how certain factors affect crop allocation of coarse
grains and other crops can lead to more effective, evidence-based strategies.
Moreover, many households in Sikasso are dependent on Mali’s cotton industry, which is
managed by a 98-percent government-owned monopsony known as the Malienne Compagnie de
Developpement des Textiles (CMDT) (Mali Agricultural Sector Assessment 2011, henceforth,
Mali Assessment). In Simona, the village where I served as a Peace Corps Volunteer, cotton
production served as the main income-generating activity for 96 percent of households. Also,
CMDT provides improved access to credit and coarse-grain inputs for many villages, especially
in remote areas. However, due to debt worth approximately two hundred million US dollars that
accumulated between the 2004/05 and 2008/09 seasons, negotiations have been underway since
early 2010 to privatize CMDT into four companies, though recent bumper crops and current civil
unrest makes this deal uncertain. Still, given the significant effect that CMDT and its policies
have on its suppliers, both CMDT and Malian policymakers can benefit from understanding the
relationship between cotton and coarse grain production, whether CMDT restructures or not.
In order to address these concerns, this thesis poses the following research question:
Which market and household characteristics significantly affect land allocation of cotton, maize,
sorghum, millet, and secondary field crops? In other words, what determines how much of a
household’s land it devotes to specific crops? Not only will answers to this question give insights
to Malian policymakers as they try to tackle complex problems in the Sikasso region, but perhaps
more importantly, the development of an econometric method that does so will equip economists
to evaluate land allocation decisions in multi-crop farming systems under other circumstances
where it could assist in addressing similar issues.

2

1.2

Literature Review
There is a rich literature dedicated to studying determinants of supply response, crop

selection, and land allocation in developing countries, which were reviewed for this study. The
first set stemmed from the Nerlovian supply model, which originates from Nerlove (1956) but
has been expanded and modified upon since. Askari and Cummings (1977) state that, in basic
terms, the model estimates a household’s agricultural production in terms of price expectations
as a function of past prices as well as partial area adjustments, though it does include a vector of
non-price factors that may affect supply response. In a sense, it attempts to model present supply
as a function of past supply, which can be effective for forecasting.
Indeed, it has already been adopted to study the cotton industry in Mali; for example,
Theriault (forthcoming) examines the effect of CMDT institutional variables on cotton supply
over a decade. She finds that timely payments by CMDT and higher credit recovery rates had a
significant positive effect on future cotton supply. Yet, because the study focused on this one
crop, it does not explore how these institutional variables affected cotton’s relationship with
coarse grains. Considering four Malian crops, Vitale, Djourra, and Sidibe (2009) estimate supply
response over 14 years while controlling for crop rotation, finding that Malian producers’
responses to output prices are nearly twice as inelastic as those of producers in developed
countries. Still, because their dependent variables are expressed as acreages and not shares, their
work struggles to show changes in one crop relative to another. Understanding this relationship
is essential as land and other production constraints limit the amount of resources that can be
devoted to multiple crops.
While it has its strengths in certain applications, the Nerlovian model will not work to
answer the research question put forth by this thesis. From a practical viewpoint, the Nerlovian

3

model requires time-series data that are not available for this study. Also, its theory of partial
adjustment may not be fitting for Mali; as will be discussed, many farmers have dropped cotton
production altogether in certain years—a change that is hardly partial. Furthermore, the
assumption of the Nerlovian model that farmers will tweak their past production based on
present circumstances does not fully model the foundations of household decision-making. It is
better to attempt a model that rationally explains how farmers might annually re-evaluate crop
choice and land allocation based on developments in market and household characteristics in
order to understand decisions made in the past, present, and future.
The need for such a model brings into play a second set of literature, which seeks to
model household maximizing behavior and apply various methods of empirical analysis. Its
primary working hypothesis is that semi-subsistent households are rational, but do not
necessarily aim to maximize profit. This is because these farms are not traditional firms, but as
partial consumers of what they produce, they seek to maximize household utility or have
multiple objective decision-making problems. To test this theory, studies have employed CobbDouglas production functions (De Boer and Chandra 1978; Barnum and Squire 1978), goal
programming software and simulations (Lee, Tipton, and Leung 1994), and linear programming
(Ahn, Singh, and Squire 1981), finding that these other models are often more effective for
explaining rural household behavior than one that assumes profit-maximization. In particular, the
agricultural household model (discussed in Chapter 3) has been adopted and adapted to study a
wide-scope of issues, including transactions costs and market participation (Omamo 1998;
Barrett 2008; Goetz 1992), missing markets (de Janvry, Fafchamps, and Sadoulet 1991; Van
Dusen and Taylor 2003), risk aversion (Fafchamps 1992; Hazell 1982), labor availability
(Benjamin 1992), and credit access (Dorward 2011) to list a few. These papers and others make

4

important contributions to economists’ understanding of the obstacles facing semi-subsistent
households and how such barriers may influence their supply response.
However, because they are set in various countries and often focus on a particular topic
relevant within the agricultural household model, the literature as a whole remains incomplete.
Additionally, the details and assumptions of these studies have not been discussed in the context
of Mali’s cotton-growing region; for example, many studies assume the existence of an active
labor market, though one is missing in the rural villages studied here. Moreover, most studies
take data from a single or dual-crop production system, or simplify crop diversity into two
categories—typically cash and staple crops (De Boer and Chandra 2001). Thus, they are often
not able to predict crop-specific supply response in a multi-crop system, though production of
three of more crops is the norm in Mali and other developing countries. Therefore, while there is
wealth of literature on which to draw in designing a conceptual model for Malian households,
there is not an analytical technique that satisfies the scope of the research question given that I
want to estimate the effect of multiple determinants on a multi-crop system.
A final bit of research should be discussed here as it most closely relates to this thesis’s
research objectives: current work carried out by the Project to Mobilize Food Security Initiatives
in Mali (PROMISAM), a joint USAID/Mali, Gates foundation, and Michigan State University
program aimed at assisting Mali’s Food Security Commission. One research activity under
PROMISAM aims to analyze the relationship between Mali’s cotton and coarse grain subsectors
and particularly how participation in the cotton value chain affects coarse grain productivity and
food security. The hypothesis that the cotton industry may affect coarse grain yields arises from
the fact that farmers have improved access to credit and fertilizer through CMDT. Yet, recent
PROMISAM field work by Boughton and Dembelé (2010) has noted that Mali is in “a process of

5

transition from a cotton-cereal production system, where the cotton subsector facilitates access to
fertilizer for cereals as well as cotton, to a cereal-based production system.” This is a
considerable development, and where true, may greatly reduce the efficacy of interventions
through CMDT designed to improve farm productivity.
Following this report, scholars associated with PROMISAM analyzed survey data from
Mali’s cotton-growing area (the same data used for this thesis, described in Chapter 4). In
another thesis, Sako (forthcoming) is exploring the determinants of participation in the cotton
industry—whether or not a household grew cotton—across the three survey rounds.
Additionally, Murekezi et al. (forthcoming) is updating evidence on the interactions between
cotton and coarse grain production, last analyzed quantitatively by Dioné (1989). Estimating the
effect of cotton area on the quantity of fertilizer applied to each coarse grain, and then the effect
of that fertilizer on productivity, Murekezi et al. find that participation in the cotton industry is
associated with higher levels of fertilizer use, which significantly increases maize yields. This is
an important verification of the perceived relationship between cotton and maize production in
Mali.
Since both these upcoming studies focus on distinct, albeit important, aspects of Mali’s
multi-crop system in the cotton basin, it would be beneficial to analyze the farming system as a
whole. Not only may doing so challenge or reaffirm their findings, but it may also detect
unintended consequences of cotton production or policy interventions. For example, participating
in the cotton industry may improve fertilizer use and maize productivity, but does it improve
land allocation to maize or other coarse grains as well? After all, dropping cotton frees up a
certain set of constrained household resources—land, labor, and capital—allowing farmers to
plant more hectares of course grains and spend more time cultivating coarse grains (or other food

6

crops), leading to increased production levels. On the other hand, as many think, dropping cotton
production potentially limits access to other constrained resources that can be secured with credit
through CMDT, such as fertilizers and insecticides, preventing farmers from maintaining coarse
grain production levels, particularly for maize, which relies more on chemical inputs. Or finally,
these factors could have counteracting effects that lead to insignificant changes in household
coarse grain production as a result of growing cotton, depending on a household’s wealth and
access to markets. Thus, given the importance of such findings for policymakers, there is room
for a similar study on these topics using a different approach. After all, the dynamics of farming
systems are too complex to be fully explained by any one study, including this thesis.
This example not only stresses the complexity of resource allocation in a multi-crop
system, but also brings up another critical aspect of farmer’s decision-making for production:
though decisions are made continuously throughout each growing season, the rational farmer
makes decisions after considering all crops simultaneously. Accordingly, one cannot assume that
farmers decide to grow cotton first, notice improved access to fertilizer, and then decide to grow
maize, because this assumes a sequential decision-making process that may instead be
simultaneous. Perhaps some farmers grow cotton because that makes it more affordable to grow
maize, and so on. While Murekezi et al. (forthcoming) address this simultaneity bias using an
adapted control function approach with instrumental variables, another alternative would be to
develop a method that explores the relationship between production decisions without treating a
single crop’s planting decision as an independent variable that explains another crop’s planting
decision. Such a technique could potentially enrich and augment PROMISAM’s current findings
as well as be applied to other contexts if it proves to be an effective estimator.

7

1.3

Research Methods
A review of the existing literature reveals unsolved puzzles, and a few gaps where this

thesis hopes to make some contributions. The first is an agricultural household conceptual model
discussed and fitted to the context of farmers in Mali’s Cotton Basin. Rather than assuming each
household farm acts like a profit-maximizing firm, the agricultural household model sees each
farm as a family with a utility-maximization problem. The primary implication contrasting these
two maximization assumptions is that as a utility-maximizing unit, the tastes and preferences of
the household are incorporated into its own production decisions. The utility-maximization
problem is then subject to resource constraints for income, time, and production. The
constraining effect of high transaction costs that occur because of missing markets is also a key
factor in the model.
The framework proposed from the agricultural household model will then inform the
selection of variables from available sources. These data come from two survey rounds of about
150 households—25 from six villages—from the Koutiala cercle in 2008/09 and 2009/10. Fifty
households are also added to the sample from one round of surveys in 2010/11. These data
originate from one village (i.e., Simona in the Yorosso cercle) where I served as a Peace Corps
Volunteer. In total, this gives our dataset 350 observations on which to draw for the analysis.
Next, these data will be used to estimate a fractional multinomial logit econometric
model, which will simultaneously estimate changes in shares of total farm land allocated to
staple crops as a function of market and household variables. Using an approach formalized by
Papke and Wooldridge (1996), a standard logit model, which is typically used to generate
predicted probabilities for binary outcomes, will be used instead to fit predicted crop shares to
actual crop shares. Thus, by modeling shares of all crops as outcome variables, I avoid the risk of

8

a simultaneity bias associated with including one crop’s production decision as an explanatory
variable for another crop’s production decision. Using the coefficients generated from the
fractional multinomial logit through a quasi-maximum likelihood estimator, I will calculate the
average marginal effects of the explanatory variables on crop shares.
To my knowledge, this thesis will be the first use of this econometric technique to study
multi-crop system land allocation decisions. It better recognizes the complexity of the decision
and also is able to predict shares for all crops simultaneously. Doing so will contribute to related
literature where the household decision-making model is oversimplified or where the analytical
focus on a single crop in a multi-crop system neglects to account for trade-offs or synergies
between crops.
1.4

Thesis Organization
The remainder of this thesis will proceed as follows. Chapter 2 provides context for the

issues in this paper by describing the farming system in Mali’s Koutiala production zone based
on CMDT data and observations attained as a Peace Corps Volunteer in Mali. Chapter 3 then
presents the agricultural household model, a conceptual framework for determining which
factors affect crop choice and resource allocation. Chapter 4 describes the data and identifies the
variables for the empirical application of the model. Chapter 5 defines the fractional multinomial
logit and then presents and discusses its results. Finally, Chapter 7 concludes.

9

CHAPTER 2: BACKGROUND
This chapter increasingly narrows its discussion of Mali in the context of this study. It
describes the Koutiala production zone where the research was conducted and provides an
overview of Koutiala’s staple crops, their subsectors, and possible synergies between them.
Informed by PROMISAM research, CMDT data, and personal observation, the chapter’s aim is
to provide background for those unfamiliar with the Koutiala production zone in Mali. A detailed
description of the typical household found in the Koutiala area is also provided in Appendix 1.
2.1

Geographic Scope
Mali is a land-locked country in West Africa, nearly twice the size of Texas, which lies

east of Senegal and north of Cote d’Ivoire (CIA 2012). It has a population of about 16.5 million
people, though 90 percent of the population lives in the southern half of the country. This is
because the north-eastern regions are located near or in the Sahara desert with a few major
settlements along the Niger River, though at the time of this writing, this area is occupied by
rebel forces. Both rainfall and population density increase as one travels farther south, so it is no
surprise that in the southern-most region of Sikasso, which is highlighted in Figure 1, nearly all
rural households are largely reliant on rain-fed agriculture (USAID 2010).
The government of Mali has a four-tiered administrative system. Its eight regions are
overseen by the capital district of Bamako, and each region is divided into cercles. Figure 2
shows that the Sikasso region is split into seven cercles, each of which is subdivided into smaller
communes. Finally, communes are comprised of several rural villages that are traditionally run
by village chiefs, not part of the formal administration. For example, I served as a Peace Corps
Volunteer in Simona, which was one of twelve villages in its commune in the Yorosso cercle.

10

Figure 1: Map Highlighting Mali's Sikasso Region (Profoss 2011)
For interpretation of the references to color in this and all other figures, the reader is referred to
the electronic version of this thesis.

Figure 2: Cercles of the Sikasso Region (Rarelibra 2006)

Koutiala

Bougouni

Yanfolila
s

Sikasso

Kolondieba
Kadiolo

11

Yorosso

Given its size and economic importance, the cotton company CMDT has its own
administration consisting of six production zones, four of which cover the Sikasso region. While
this thesis’s results may be applicable to all cotton-producing regions in Mali, its data only come
from one of the CMDT production zones. The Koutiala production zone covers Sikasso’s
Koutiala and Yorosso cercles, located to the upper right of Figure 2. The Koutiala cercle consists
of 35 communes, while neighboring Yorosso cercle only manages 9. Geographically, the two
cercles are very similar, with the Yorosso cercle being higher elevated and generally more
isolated from economic activity. The city of Koutiala itself, despite not currently serving as a
regional capital, is Mali’s third most populous city. Located in the “Old Cotton Basin” and at
crossroads between Sikasso, Burkina Faso, and roads to the Niger River, it is a regional trade
hub with a strong industry in cotton manufacturing. Since this production zone is historically the
most significant producer of cotton in the country, it makes sense to focus here given the
relevance of cotton to Mali’s economy and to the research question.
This chapter will place greater emphasis on Simona, the village in which I served for two
years as a Peace Corps Volunteer, as a case study for a rural community in the Koutiala
production zone. This means that while the information presented is applicable to most Malian
households in this area, Simona can serve as a concrete example and reference point. Qualitative
information about Simona comes from personal observation or informal interviews with various
local farmers, while any quantitative data were collected by a survey of the 2010/11 production
cycle conducted in May of 2011 in which 50 out of Simona’s 140 households were randomly
selected to participate. Since this dataset is also used in our econometric analysis, Chapter 4
provides a more detailed description for how the data were collected.

12

Some aggregate data for the Koutiala production zone are also presented in this chapter to
describe average crop adoption and shares. While the econometric analysis will incorporate
PROMISAM survey data from six villages in the Koutiala cercle, this information will not be
presented here. Rather, to get more complete data that also include the Yorosso cercle, I will
examine information brought together by CMDT’s regional Koutiala office. Statistics and counts
of producers are compiled on a yearly basis through CMDT thanks to their network of producer
organizations, extension agents, and district offices. Much of these data are collected not only for
statistical purposes, but also to order inputs and set the cotton producer price.
2.2

The Farming System in the Koutiala Production Zone
Farming systems research is defined by its focus on the interactions between the

production of different crops at the farm-level (Boughton 1994a). In Mali, the portfolio of crops
that characterize a farming system vary across regions. For example, in CMDT’s Koutiala
production zone the standard portfolio includes cotton, maize, sorghum, millet, and some
secondary crops, such as peanuts, cowpeas, sweet potatoes, or vegetables. However, just north of
the Yorosso cercle, where CMDT is not present, the standard crop portfolio includes fonio, but
not cotton or much maize. Understanding a village requires knowledge of its farming system and
the agricultural subsectors in which it is involved.
For this study, we are particularly interested in the interactions during the planting phase
of the production cycle. The start of this cycle normally occurs in late April, when planting
potentially begins, and goes to late May and June, when households face a final decision for crop
choice and land allocation. This time is important as it defines the break between agricultural
cycles; so, when it is stated that data are collected for the 2008/09 agricultural cycle, this means
that these crops were planted, grown, and harvested in 2008, but represent the household’s
13

supply of crop (whether it is consumed or sold) throughout 2009. The following subsections
intend to explain the characteristics of each crop subsector and how these may influence planting
decisions.
2.2.1

Cotton

Today, Mali’s cotton subsector is controlled by CMDT (Compagnie Malienne de
Développement des Textiles), a cotton monopsony that manages most aspects of input
procurement, cotton production, processing, and marketing. It is highly centralized and has been
since its development from a French cotton parastatal in 1974 (Tefft 2003). Yet CMDT also
retained its economic and social objectives: cotton production as well as building capacity
through farmers’ organizations, agriculture extension, and functional literacy programs.
Furthermore, CMDT has provided farmers with oxen, donkeys, plows, and carts, which are all
heavily utilized in crop production today, and moreover, introduced many farmers to maize
during the food crisis in the 1970s. For these reasons, many farmers in the Koutiala area
appreciate CMDT, though the cotton industry has struggled in the past decade.
Volatile year-to-year world cotton prices and a fluctuating exchange rate of the West
African franc (FCFA) to the US dollar have forced the cotton producer price-setting board to
balance between farmer livelihoods and CMDT’s profitability, with little success. Lower prices
have led to decreases in the number of cotton producers and hectares cultivated, which have
resulted in cotton production levels in 2007/08 that were only 40 percent of those a decade
earlier in 1997/98. The decline also created debt for CMDT over the five years that began in the
2004/05 season and ended in the 2008/09 season; CMDT’s total deficit—even after payments
from a support fund—was well over a hundred billion FCFA or about two hundred million US
dollars (Mali Assessment, 2011) To make matters worse, late payments to farmers and evidence
14

of corruption have generated distrust among farmers, causing a major cotton strike in 2001,
which led to reforms, and a smaller strike in 2008. Because of these struggles, the government of
Mali has been working over the past five years to privatize CMDT and splitting it into four
separate companies, each of which will have state-protected monopolies in different parts of the
country (Theriault and Sterns 2012).
Despite these troubles, the rise in cotton prices over the past few years has the potential to
reignite enthusiasm for the cash crop’s production among rural farmers. After the strike in
2007/08, the cotton producer price rose to 200 FCFA per kilo, but then dropped to 170 FCFA
and 185 FCFA per kilo in 2009/10 and 2010/11, respectively (Mali Assessment 2011); these are
the years examined in this study. Since then, the producer price of cotton rose 38 percent to 255
FCFA per kilo in 2011/12, and this price was maintained for the 2012/13 season (ACP Cotton
2011; ACP Cotton 2012). Despite civil political turmoil, these higher producer prices are likely
to boost land allotted to cotton throughout the Koutiala production zone.
Aside from the potentially volatile producer price for cotton, participating in the cotton
industry has many pros and cons. Cotton is a labor- and input-intensive crop that offers its
producers unique incentives through CMDT. Living in the Koutiala production zone means that
evaluating the trade-offs between the costs and benefits of cotton production is a decision that
every rural households must make. While some decide that it is too risky or not profitable
enough, other households are dependent on cotton production, especially in remote areas. In
Simona, where 96 percent of households grow cotton, one farmer told me: "Ni kouri te, fen te"—
without cotton, we have nothing.
The perks to growing cotton are tied to the institutional design of CMDT and are aimed at
reducing direct and indirect costs associated with production, marketing, and risk management.

15

To start, a legal mechanism established in 1994 requires that CMDT set the farm-price of cotton
before each agricultural cycle, and this has been guaranteed thus far. Even though farmers must
order their cotton seed, fertilizer, and insecticide around October of the previous year—before
they have even harvested their cotton—knowing the cotton price before planting allows
households to adjust production decisions, and it reduces their vulnerability to price shocks.
Then, once the harvest is completed, CMDT is capable of paying farmers shortly after the raw
cotton is collected and delivered to its processing facilities. This gives farmers the cash needed to
pay back loans and their children’s school fees without having to sell their coarse grains early,
when supply of coarse grains is high and prices are low. This possibly explains the correlation
between cotton producer strikes in 2001 and 2008 and complaints of late payments from years
prior.
Another important advantage of growing cotton is that it can help farmers procure inputs
on credit backed by future cotton revenues. As of 2010, farmers could receive credit annually for
up to 74,125 FCFA (roughly 150 US dollars) per hectare of cotton to procure their cotton
variable inputs (e.g., fertilizer, seed, and pesticide) as well as fertilizers for coarse grains. This
amount is then deducted from the payment for the next cotton harvest along with a small
premium. With a premium of only seven percent for purchasing sacks of fertilizer on cotton
credit, this input procurement method is preferred by many households in Simona, as described
in Appendix 2. Furthermore, CMDT occasionally partners with Mali’s agricultural development
bank, the Banque Nationale de Dévelopment Agriocole (BNDA), to help farmers purchase
livestock on cotton credit over a longer period of time.
Finally, CMDT covers transportation expenses for collecting the raw cotton by sending
trucks into the rural villages with the expectation that producer organizations will load their own

16

cotton into each truck. CMDT will also send trucks to deliver purchased seed and fertilizer. This
is noteworthy considering that one Simona farmer recalled his grandfather walking 100
kilometers to Koutiala to sell his cotton to CMDT forty years ago. This collection and delivery
service decreases transportation and transaction costs for households, especially those in remote
areas, which is often a barrier to relying on other cash crops for income generation.
However, cotton has higher investment costs than the other field crops. Seed, fertilizer,
and insecticide and must be applied close to their specified quantities provided by CMDT
extension agents in order to attain the highest quality. For example, extension agents say that a
household should apply three 50 kilogram sacks of Complexe Coton and one 50 kilogram sack
of Urea per hectare of cotton as well as spray insecticide on their cotton fields three times.

2

Cotton also requires a lot of family labor, especially at the harvest, when it is not uncommon to
see entire households, including women and children, working all day in the stifling heat,
picking cotton with their bare hands. There are also factors that cannot be controlled, such as
rainfall, which must come early enough to give the cotton time to grow but then stop so that the
budding cotton flowers will remain dry until plucked. In 2010, the rains in Simona overstayed
their welcome and a lot of healthy cotton began to mold before it could be harvested. Finally,
cotton is tough on the land and drains nutrients from the soil, which can lead to fallow fields over
time. Thus, when households find an alternative income-generating activity that is less risky and
more sustainable, they may be likely to turn away from cotton.
Below, Figure 3 uses CMDT data to chart total hectares devoted to cotton in the Koutiala
production zone and all zones between 1960/61 and 2007/08.

For the first two decades

following Mali’s independence in 1960, the Koutiala zone represented roughly half of cotton
2

A detailed description of fertilizer requirements for cotton farmers is provided in Appendix 2.
17

hectares planted in Mali. However, this percentage fell as CMDT expanded to other areas. Figure
3 also shows the effect of the protests in 2000/01 and then 2007/08, especially in other
production zones.
Figure 3: Total Hectares Alloted to Cotton in Mali, 1960/61 to 2007/08

Source: Author's manipulation of data from CMDT, Koutiala Regional Office
Next, Figure 4 zooms in on cotton hectares planted in Koutiala since 1990/91 and
extending to 2010/11. This figure shows the volatility that has defined the past two decades of
Mali’s cotton industry in Koutiala, whereas before it appears to have been characterized by slow
but reliable growth. The figure shows a strong increase in cotton hectares beginning in 1994,
after CMDT was mandated to establish a guaranteed pan-territorial producer price and the FCFA
currency was devalued. Production then peaked in 1997/98. Additionally, the figure shows that
the protests in 2000/01 were not as influential in Koutiala, but did cause a significant decline in
2007/08. The data used for this thesis come from the three years after this strike, as cotton
producers have slowly returned to the cotton industry.

18

Figure 4: Cultivated Hectares of Cotton in the Koutiala Zone, 1990/91 to 2010/11

Source: Author's manipulation of data from CMDT, Koutiala Regional Office
It was mentioned previously that 96 percent of Simona households grew cotton in
2010/11, but is this percentage the same across the entire production zone? In the six villages
that PROMISAM surveyed from the Koutiala cercle—the details of which are described in
Chapter 5—the fraction of households that grew cotton was not so high. In 2008/09 and 2009/10,
the average rate of adoption was 57.3 percent, with a large standard deviation across villages. In
the village with the highest adoption rate, 85.2 percent of households grew cotton, while only
18.8 percent grew cotton in the village with the lowest adoption rate. However, according to a
CMDT representative for the Yorosso cercle, the three communes near Simona (specifically,
Kifosso, Koury, and Yorosso) all boast villages in which over 95 percent of households grow
cotton. The three exceptions are the commune headquarters themselves, where the mayors
reside, since these communes have higher populations and more diversified economies. The
representative suggested that many households adopt cotton in these cercles because they are
19

more isolated from the economic activity surrounding Koutiala, so that households have fewer
alternative sources of income and poorer access to active coarse grain markets, accessible credit,
and affordable fertilizer. Therefore, they are more reliant on the services provided by CMDT.
2.2.2

Coarse Grains

Unlike the highly centralized CMDT, the structure of the course grain subsector consists
of a wide range of private actors since the liberalization of the cereal market in the early 1980s.
Today, the general coarse grain value chain involves various input dealers, creditors, farmers,
collectors, wholesalers, semi-wholesalers, retailers, processors, and transportation, and storage
service providers (Mali Assessment 2011). Furthermore, whether or not a household decides to
sell its coarse grain, there are food processing requirements before consumption. While this work
can be done manually within the household, enterprises have sprung up to grind coarse grains
into flour, even in rural villages. Maize’s value chain differs from millet and sorghum in that it
can also be consumed as a fresh crop, can be processed for livestock feed, and has higher input
requirements; still, the value chains of maize, millet, and sorghum are similar.

3

Whereas the cotton industry has struggled, the coarse grain subsector, and especially
maize production, has been characterized by steady growth. Indeed, both millet and sorghum
production have increased, but maize has recorded the fastest growth of any rain-fed coarse grain
cereal in Mali for the past two decades. The share of maize within total cereal production has
increased from around 10 percent in 1991 to nearly 18 percent in the late 2000s (Mali
Assessment 2011). Increasingly, many subsistence farmers are turning to maize due to its

3

Both Boughton (1994b) and Diallo (2011) have studied Mali’s maize subsector in great detail
and have diagrams illustrating its value chain.

20

growing market demand and higher yields—each stalk is capable of producing between two and
three healthy ears, whereas millet and sorghum only produce one tassel.
However, maize is the riskiest coarse grain to produce with higher input requirements for
labor, fertilizer, and rainfall. By comparison, millet and sorghum are better at resisting drought
and more suited to arid climates and soil of poor quality. While both millet and sorghum yields
increase with fertilizer use, most Malian farmers choose not to apply fertilizer to millet or
sorghum, presumably because the perceived marginal return of doing so is less than the marginal
cost. However, all three coarse grains are susceptible to pests, such as crop-specific disease or
locusts. Understanding coarse grain input requirements and the nutrient-draining nature of cotton
also explain crop rotation as taught by CMDT and practiced by many households in Simona and
elsewhere. After growing cotton, a typical household would grow maize, which benefits from
residual fertilizer applied to the cotton, followed by sorghum and millet. CMDT also teaches that
crop rotation can help with weed and pest control.
Whether it is because of risk-adverse behavior or household demand for a diverse diet, it
is common for households to produce all three coarse grains. While similar, a few defining traits
determine southern Malians’ preference for both production and consumption decisions. To start,
all three coarse grains are staples to the southern Malian diet as the main ingredient in ‘toh’, a
simple food consumed with sauce for most lunches and dinners. In general, maize toh is
considered to taste the best (based on informal polls), although sorghum toh is also tasty and is
rich in iron, and the sourer taste of millet is preferred for breakfast porridge. As discussed in
Chapter 3, perhaps certain taste preferences also affect production choices.
Table 2.1 shows the adoption rate of these coarse grains by the seven villages in this
thesis’s dataset, later described in Chapter 5. Data for the first six villages comes from the

21

Koutiala cercle and was combined between 2008/09 and 2009/10, while the final row displays
statistics from Simona in 2010/11. These statistics vary less than the adoption rate for cotton,
discussed previously. Of the coarse grains, adoption rates for maize vary the most between
villages, falling as low as 65 percent in a village that is very loyal to sorghum and millet.
Sorghum is by far the most widely adopted of these coarse grains, with all but four households in
the entire sample choosing to grow it. However, all coarse grains are adopted by the majority of
households within a village.
Cercle
Koutiala
Koutiala
Koutiala
Koutiala
Koutiala
Koutiala
Yorosso

Table 1: Adoption Rates (%) of Coarse Grains by Village
Commune
Village
Maize
Sorghum
Fagui
Nampala II
78
100
Sinkolo
Tonon
83
100
Sincina
Kaniko
84
100
Sincina
Try I
65
100
Koutialaa
Signe
86
96
Mpessoba
Gantiesso
94
96
Yorosso
Simona
96
100
84
99
Average Adoption Rate:
Average Hectares per
1.3
3.2
Household (if Adopted):
Share of Total Hectares
16%
46%
Allocated to Coarse Grains:

Millet
94
60
100
100
92
83
92
89
2.9
38%

Encouraged by recent growth, the Malian government has set ambitious goals to increase
its country’s coarse grain production significantly. Although the goals would have production for
all coarse grains increasing in absolute terms, maize’s share would increase, moving from 15
percent to 39 percent of total cereal production, while corresponding shares of millet and
sorghum would fall from roughly 50 percent to just 33 percent by 2012/13 (Mali Assessment
2011). Thus, with maize emerging as a priority food crop, it will be especially important to
examine its relationship to the share of land devoted to cotton and the other coarse grains.

22

2.2.3

Secondary Crops

In addition to cotton and the coarse grains, Malian households may choose to grow a
variety of other crops during the rainy season, including peanuts, cowpeas, sweet potato, or other
vegetables. While some crops can be grown in cold season when there is less of a labor
constraint, the rainy season provides regular watering, an activity that requires much effort in the
cold season after the rains stop. Hence, outside of a small vegetable garden, many households in
the Koutiala area grow other crops in rainy season. Since this study is mostly concerned with
these crops when they compete with cotton and coarse grain production for shares of cultivated
land, the discussion will be focused entirely on their production in rainy season.
Table 2: Adoption Rates (%) of Secondary Crops by Village
Cercle Commune Village Secondary Peanut Sweet Cowpea Rice Other
Crops
Potato
Koutiala
Fagui
Nampala
90
68
0
56
32
34
II
Koutiala
Sinkolo
Tonon
52
27
0
29
13
33
Koutiala
Sincina
Kaniko
84
68
26
62
8
28
Koutiala
Sincina
Try I
85
71
40
56
0
35
Koutiala Koutialaa
Signe
90
74
22
72
6
10
Koutiala Mpessoba Gantiesso
98
91
13
57
31
39
Yorosso
Yorosso
Simona
66
44
8
24
10
48
81
64
15
51
15
33
Average Adoption Rate:
Average Hectares per Household
2.1
0.8
0.5
1.3
0.9
0.8
(if Adopted):
Share of Total Hectares
37%
5%
35%
8%
15%
Allocated to Secondary Crops:
This study will refer to these other crops as the “secondary crops,” not as a judgment of
their economic or nutritional value, but due to their “secondary” role in the Koutiala farming
system and this study’s focus on cotton and coarse grains. Since all secondary crops will be
aggregated for the econometric analysis, Table 2.2 summarizes key information about common
secondary crops in the seven villages surveyed in the Koutiala production zone, including
adoption rates by village, the average hectares planted per household (if adopted), and the
23

average share of total secondary crop hectares made up by each individual crop. Data for the first
six villages was combined between 2008/09 and 2009/10 as are described later in Chapter 5. The
final row displays statistics from Simona in 2010/11.
The adoption rates of various secondary crops vary by village. For example, cowpeas and
peanuts are grown by the majority of households in some villages, and the former is typically
cultivated on more than one hectare. Rice also has a higher adoption rate in Gantiesso and
Nampala II and, where adopted, is grown on almost one hectare. Compared with the Koutiala
villages, Simona’s adoption rates average a bit lower. Only Simona’s “Other” category has
higher measures with a 48 percent adoption rate, likely due to increased jatropha production on
marginal lands, which has been promoted there by entrepreneurs and non-profits alike.
Clearly, the types of crops being clustered together as secondary are diverse and, in some
cases, seem to be common household staples, but they also have much in common. At least in
Simona, there were a couple ways to identify secondary crops. First, they were often planted on
marginal lands and only after the successful planting of the coarse grains. Moreover, these plots
were usually planted and maintained by the household’s women using manual tools like the hoe
or ‘daba’. Then, after the harvest, the yields were mainly used for household consumption, as
snacks or to supplement meals, though households with excess could sell at market. Rice
production is the exception here, as it is labor- and capital-intensive crop like cotton, but its
adoption rates are too low to include it as its own outcome category.
Thus, one may expect that households growing secondary crops share some telling
characteristics. Their ability to diversify may be the result of excess labor and human capital
(e.g., knowledge of sweet potato farming) as well as aversion to risk. These households may also
have an increased preference for these food products, which are often considered luxury goods.

24

On the other hand, if a farmer devotes a significant share of his land to a secondary crop, he may
have recognized his comparative advantage in producing the crop as an income-generating
activity. These are patterns that may appear in the analysis of the model.
2.2.4

Conclusion

In this chapter, I have described each subsector and presented some basic descriptive
information on the adoption of these crops in the Koutiala production zone. Using the CMDT
data, I will finish by comparing aggregate hectares devoted to cotton and coarse grains.
Unfortunately, CMDT does not collect data on any secondary crops, but they will be included in
the econometric analysis described in later chapters.
Figure 5: Hectares of Cotton and Coarse Grains in the Koutiala Zone, 1999/00 to 2010/11

Source: Author's manipulation of data from CMDT, Koutiala Regional Office
Figure 5 presents the total cultivated hectares for cotton, maize, sorghum, and millet over
the past decade in the Koutiala production zone. Of course, cotton is the most uneven; in two
years, it drops from occupying the most land area to aggregate hectares that are below sorghum
25

and millet, and just above maize. All three coarse grains saw increases in hectares planted,
though small declines for millet and maize occurred at the same time as the cotton strike.
Ultimately, it appears that millet had the largest growth over the decade in absolute terms, adding
over 30,000 hectares. Surprisingly, it does not appear that the decline in cotton was compensated
with large increases in hectares allocated to coarse grains, as one might expect due to the freeing
up of inputs such as land and labor that would were previously dedicated to cotton production.
But again, this graph does not account for secondary crops and is only a measure of land area,
not of yield or production.
However, while these aggregate figures illustrate the effects of the 2007 cotton strike on
crop shares, they do not explain what is happening at the household level, which is critical for
understanding the farming system in greater detail. The structure and conduct of the subsectors
described in this chapter hint at what might influence farmers’ planting decisions, but there are
many other factors to consider. To determine why households make certain planting decisions,
Chapter 3 presents an agricultural household model, a theoretical framework for household
decision-making that is the foundation for our analysis.

26

CHAPTER 3: CONCEPTUAL MODEL
This chapter presents and justifies the theoretical framework for this thesis by discussing
its assumptions of human decision-making in the context of Mali’s Koutiala production zone,
summarized in Chapter 2 and Appendix 1. It begins by discussing the Malian farmer as a rational
actor and his household’s utility maximization problem. Next, I present an agricultural household
model and use it to justify a reduced-form land allocation function that will inform the selection
of explanatory variables in Chapter 4.
3.1

Malian Farmers as Rational Actors
In perhaps his most influential work, Theodore Schultz (1964) put forth his “efficient but

poor” hypothesis, which rejected that poor farmers were inefficient because of cultural
characteristics (e.g., laziness), but rather, asserted that poor farmers make efficient decisions with
the few resources to which they have access. Thus, the problem is not the decision-making, but
the lack of resources and technology. His theory had significant implications for development
policy as it implies that the most effective way to assist these farmers is to introduce and
encourage the adoption of new factors of production that increase a farm’s productive potential
(Ball and Pounder 1996) and the human capital needed to exploit it.
Today, there is still debate as to whether poor farmers are efficient—that is, do they
equate marginal returns with marginal costs—but Schultz’s work has brought about an implicit
paradigm in modern development economics of the “rational but poor” farmer (Duflo 2003).
This paradigm implies that the poor farmer should be treated as a rational agent as assumed
under the neo-classical economic model, recognizing that both efficiency and rationality may
vary given different social and cultural values. Thus, as long as these social and cultural values

27

are understood along with a given set of economic incentives, the bottom line is that the
decisions of poor farmers are logical, and perhaps predictable.
Yet for many years economists did not presume that the rural poor were rational, and it is
even a question discussed among Peace Corps Volunteers in Mali. After all, is it logical to buy
and brew tea and sugar daily, but not use soap to wash your hands before meals? In Mali, maybe
so. Consider that in Mali, household decisions are made by its older men, who derive their value
from physical wealth as well as their social status within the community. Especially in small
villages, respect is a highly valued commodity that affects the individual and his household. As it
is customary to drink three rounds of tea after lunch during the day’s hottest hours, brewing tea
to share with neighbors and friends is a way to earn and maintain respect in the village, and
conversely, refusing even to sit and drink tea is considered rude. However, washing with soap,
while cheap, is seen as a hassle, and those Malians who have been educated about its benefits
have yet to be convinced of its value in saved health expenditures. An equivalent comparison
may be flossing teeth in United States—a healthy habit that most are taught but that far fewer
practice regularly, despite what they may tell their dentists. Therefore, if the perceived value of
soap is low while the perceived social value of tea is high, and both cost approximately the same,
the rational decision is to buy tea and sugar.
Another question commonly asked by Peace Corps Volunteers: how is it rational to buy
your family its third motorcycle when all of your children are chronically malnourished? Again,
it comes down to the defining the specific value system employed by Malian households. So
what gives a third motorcycle value? Economically, it may reduce costs for market transactions
requiring transport, especially for larger households of twenty people or more. It may also be a
gift or a promised benefit to a hard-working adult male, working either on the farm or away,

28

which earns him respect and ensures his further contributions to the household. Additionally, it
generates respect for the household as a status symbol and can be used to give favors to
neighbors and friends in need of transport. Similarly, it further reduces the risk of being stranded
in a health-related emergency when quick transport to a regional hospital is dire. Finally, a
motorcycle is a method of turning cash into capital, and less cash means reduced social pressure
to give aid to struggling neighbors and friends come the next hungry season. By contrast, the
perceived value of having nourished children is lower since most families do not even
understand that their children are malnourished. The bloated belly on many children due to
kwashiorkor, a protein-deficiency, is often seen as a positive sign of a full stomach. Therefore,
until Malians understand that their children are malnourished and how this may stunt physical
and mental growth, and thus future household productivity, it is rational to purchase a third
motorcycle even when a Peace Corps Volunteer thinks that same money should be spent on more
vitamin- and protein-filled foodstuffs.
Like many studies in the field of development economics before it, this thesis also
assumes that its population, the Malian household, is a rational agent while understanding that
they face an entirely different set of market and social incentives that may easily be
misunderstood by the researcher. The goal of the researcher then is to properly understand the
factors that do influence this rational decision-making. To quote Schultz, “Most people in the
world are poor. If we knew the economy of being poor, we would know much of the economics
that really matter” (1980).
3.2

Malian Household Maximization Problem
The next matter to settle is what the rational Malian household is attempting to maximize.

This question is complicated by the semi-subsistent nature of the Malian household in the
29

Koutiala production zone since it exhibits the characteristics of both a producer and a consumer.
It is heavily engaged in agricultural production and must make decisions concerning the
allocation of inputs and outputs, yet it consumes a significant proportion of that production to
meet its own household’s needs and desires. Whereas the goal of competing producers or firms is
to maximize profit by producing to the point where marginal cost is equal to marginal revenue,
the goal of an individual consumer is to maximize utility (i.e., the preference for a certain set of
goods over another) given a budget constraint (Nicholson and Snyder 2008). As a semicommercial farm, do Malian households seek to maximize profit, utility, or both?
In the presence of perfect input and output markets (i.e., those free of transaction costs
and fully capable of hedging risk), it should be safe to assume that households seek to maximize
profit. In this scenario, households would choose the crop or set of crops in which they held
particular comparative advantage. Then, they would choose to produce efficiently and maximize
profit knowing that they can consume more, and thus attain a higher level of utility, with more
profit than with less profit. In other words, if every utility maximization problem is subject to a
budget constraint, then the primary objective of the household would be to loosen the budget
constrain by increasing their income through profit maximization. This profit level directly
affects consumption, but consumption does not affect production decisions. The household
model literature labels this concept as separability (Udry 1996).
On the other hand, non-separability occurs when a household’s decisions regarding
production are affected by its consumer characteristics, such as individual taste preferences and
household demographics. This is often true when market failures limit the household’s ability to
utilize additional profit earned through specialization (de Janvry and Sadoulet 2006).
Specifically, a market failure occurs when the cost of a market transaction generates a disutility

30

greater than its potential utility gain, so that even if the market exists it is still not used for the
transaction (de Janvry, Fafchamps and Sadoulet 1991). For example, if a household wants both
sorghum and millet, but the cost of finding a vendor and transporting a ton of millet is greater
than the additional profit earned by exclusively growing sorghum, then the rational decision is to
simply grow both at the desired quantity. In addition to high transaction costs, other market
failures concerning risk, credit, labor, and land can also lead to non-separability in the household
model. In fact, Udry (1996) uses separability as a benchmark for measuring the extent of market
failures in Africa. When combined, these various potential market failures weaken the argument
that Malian households seek to maximize profits, because it becomes logical to take household
consumption preferences into account.
This makes way for the second solution to the household maximization problem, which
supports non-separability. It assumes that semi-commercial farms aim to maximize expected
household utility, which is some function of expected utility of each individual in the household.
Under this maximization problem, agricultural production can then be incorporated as part of the
household’s resource constraints. For instance, agricultural production is a large determinant of a
household’s budget constraint, contributing to consumption or cash generation depending on
whether or not crops are sold at market. Since markets are unreliable, most farmers make
production decisions based on consumption preferences. These assumptions form the basis for
the agricultural household model: a utility-maximizing model subject to budget and other
resource constraints.
Other methods have been proposed for explaining the rural household’s maximization
problem, most of which try to mix profit and utility maximization into a multi-objective
approach. Fisk (1962) promoted a subsistence affluence model in which households first devote

31

resources to meet their own demand and then employ any excess resources toward income
generation. In a sense, this splits the agricultural household model into a two-step process, a
framework that has proven useful for some studies (Stent and Webb 1975; De Boer and Chandra
1978). Similarly, another study used goal-programming software to maximize five objective
functions, four to maximize crop yield for consumption during each season and another to
maximize annual net revenue (Lee, Tipton and Leung 1994). Some researchers have even
experimented with using responses from goal preference surveys to weigh the importance of
various household objective functions, though without finding significant differences between
those results and others found assuming profit-maximization (Barnett, Blake and McCarl 1982).
These alternative approaches have merit as they represent efforts to better explain farmer
decision-making in developing countries, but they may overcomplicate what can be simplified
by the agricultural household model.
For this reason, this thesis will develop an extension of the agricultural household model.
It elegantly captures the effects of a household’s consumption and production decisions and has a
more substantial literature that expands upon it, which has been summarized in Appendix 3.
Most importantly, its utility-maximizing approach seems more congruent with how Malians
seem to talk about agricultural production as a part of their lives. For example, when asked why
they hoped to increase production of a particular food crop for next year, farmers from Simona
most often responded that it tasted good and their households enjoyed it—very few mentioned
that they expected its price to go up. Perhaps because many Malians do not truly believe that
they can become rich through hard work on the farm, it is rational that agricultural production is
more a means to utility maximization.

32

3.3

Determinants of Crop Choice and Land Allocation
So far, this chapter has argued that Malian farmers are rational and their households’

primary objective is to maximize expected utility subject to a set of resource constraints. These
assumptions are best represented in the agricultural household model which was popularized by
4

Singh, Squire, and Strauss (1986), but has been greatly expanded upon since. According to
Dorward (2011), the main contributions of this model are representing the interactions between
consumption and production decisions (i.e., non-separability, as defined above) present in
household decision-making among the rural poor and, as a result of later work, identifying the
effect of market failures for labor, variable inputs, credit, and staple crops.
By adapting the agricultural household model for farming households in Koutiala, I hope
to specify a theoretical model that helps support assumptions made about rural household
decision-making in the empirical analysis. Thus, this framework identifies potential determinants
of crop shares that will soon establish which variables are included in the econometric analysis.
A summary of this agricultural household model used for this process is presented below:
Max U = f (X, Y, H)

utility maximization, s.t.

PxX + TCbx + PyY + TCby ≤ PxQ – TCsx – PvV – TCby + πy

budget constraint

Q = f (F, V, A, K, σ)

production constraint

H+F+O=T

time constraint

X, Y, Q ≥ 0
As rational agents, poor farmers strive to maximize their expectation of future household
utility, which is a function of leisure H, consumption of agricultural goods X, and consumption
4

Again, see Appendix 3 for a summary of the agricultural household model and some of its
extensions.
33

of other goods Y. Utility is maximized subject to a budget constraint in which the cost of
consumption, determined by prices Px and Py and transaction costs for buying TCb, is less than
or equal to profits from other activities πy and profits from crop production, which is equal to
this value of production (output Q times Px) minus the cost of inputs (V times Pv) and
transaction costs for selling TCsx. Also present is a production constraint, for which output is a
function of farm labor F, other variable inputs V, land area A, capital K, and risk σ. Household
choices are further restricted by a time constraint, which limits the sum of farm labor F, leisure
H, and off-farm labor O to be less than the household’s total time endowment T. Note that all
labor is assumed to come exclusively from the household, and so does not factor into the budget
constraint. Finally, we assume that X, Y, and Q are non-negative values.
The research question asks which factors affect land allocation of standard Malian crops
in the country’s cotton-growing region. Land allocation of crops is the main determinant of
agricultural production, represented by the variable Q. While other household choices (e.g.,
fertilizer usage) affect total production, most supply response models use land area as a proxy for
total production, since the two are highly correlated and because agricultural production is
difficult to measure (Askari and Cummings 1977). Thus, here, Q will represent a vector of the
portfolio of crops chosen and the share of land devoted to those crops.
Within the model, four different types of variables influence the choice of Q. First,
variables that may lead households to specialize in the production of a particular crop include the
availability of farm labor F, other variable inputs V, land area A, capital K, and depending on the
crop, price P. Second, variables that may discourage specialization and lead to diversification
include risk σ, transaction costs TC, and seasonal food constraints. These factors force a
household to contemplate the reality of missing markets, and how this might change their

34

expected utility. These are the competing influences that emerge from the modified budget
constraint.
Third, as a result of missing markets, the agricultural household model also predicts that
household tastes and preferences will directly affect the crop planting decisions and production.
Thus, household-level variables that may affect preferences for food, other consumption goods,
and leisure—such as ethnic group and family structure—should also be included. In addition, a
household’s education may affect its preferences for consuming various goods, while also
equipping it to make more efficient production choices. Thus, understanding the individual traits
of household members is essential.
Finally, a fourth set of variables exists that serve as alternative ways for the household to
earn profit and seek more utility. This group consists of the profits from other activities π y,
though such profits may represent other outcome variables in a model seeking to maximize
utility. That is, the decision to engage in other income-generating activities, be it raising
livestock or sending a family member to the city, requires an assessment of household inputs and
future consumption needs, much like agricultural production. However, if the decision to
participate in other income-generating activities is made in conjunction with the choice to grow
agricultural products, then they are, by definition, simultaneous outcomes. Recognizing a
possible simultaneity bias, I will not represent other income-generating activities in the empirical
model. Nevertheless, I recognize it here in the conceptual model as a possible influence on land
allocation decisions.
To conclude, according to the agricultural household model, crop choice and land
allocation at the time of planting is a function of production factors, the characteristics of local

35

missing markets, household members, and alternative income-generating activities. Together,
these make the following expression:
Q = f (P, F, V, A, K, σ, TC, Hh, πy)

reduced-form land allocation function

(1)

In the next chapter, I identify specific variables within the dataset that represent most of the
factors in this reduced-form land allocation function.

36

CHAPTER 4: EMPIRICAL APPLICATION
Having discussed the agricultural household model, this chapter presents the application
of the model to household-level data on agricultural production in Mali. This includes describing
the source and characteristics of the data, identifying the dependent and independent variables,
and justifying their inclusion in the model.
4.1

Data Description
To analyze crop shares at the household level, I will draw on data from a couple of

sources. Three hundred observations come from two rounds of surveys in the Koutiala cercle,
which gathered information on the same households in both 2008/09 and 2009/10. These 150
households were pulled from 6 villages with about 25 households surveyed per village. Data
from the Yorosso cercle, which is similarly (if not more) dependent on cotton, consist of an
additional 50 observations from Simona in the 2010/11 farming season.
However, the data from the original study that motivated these three survey rounds will
not be used in this thesis. Known as the RuralStruc project, the 2006/07 survey was conducted as
part of a World Bank study evaluating the effect of structural adjustment programs on rural
economies and production systems from seven countries, four of which—Kenya, Madagascar,
Mali, and Senegal—are in Africa. Within Mali, four different cercles were surveyed, each
5

representing a unique production system in Mali. Whereas other cercles were of interest
because of their rice or exclusive coarse grain production, Koutiala was selected because it is in
5

The original four cercles were Koutiala (cotton zone, Sikasso region), Tominian (coarse grain
zone, Segou region), Macina (irrigated rice zone, Segou region), and Diema (Kayes region).
However, due to limited resources, the Diema cercle was not included in later surveys helmed by
IER and PROMISAM.
37

the heart of CMDT’s cotton zone (Samake 2008). The results from different cercles were than
analyzed, compared with each other, and then compared with development indicators from other
countries.
Not surprisingly, the RuralStruc survey had various implementation partners, including
Mali’s Institut d’Economie Rurale (IER), France’s Centre de Coopération Internationale en
Recherche Agronomique pour le Développement (CIRAD), and Michigan State University with
its Project to Mobilize Food Security Initiatives in Mali (PROMSIAM). Capitalizing on their
involvement in the RuralStruc survey, IER and PROMISAM also conducted a second and third
round of surveys in 2008/09 and 2009/10, respectively, with funding from USAID and the Bill
and Melinda Gates foundation. Where families had not moved or been reorganized, the survey
sample consists of those households originally surveyed from the RuralStruc project (Murekezi
forthcoming). These are the two surveys whose data is used for our analysis, but again, the
RuralStruc data are not used in this study due to some inconsistencies between it and the design
and methodology employed by IER and PROMISAM.
The surveys led by IER and PROMISAM were very comprehensive. They asked for
detailed information on household demographics and acquired assets, including farming
equipment and livestock. Further, the dataset describes the household’s agricultural production in
full, including land allotment and some other measures not needed for this study, such as
production levels and input costs. While the RuralStruc and original Koutiala dataset actually
contain households from three cercles, this thesis does not use data for all the households
surveyed, but only those from Koutiala.
So how were the villages and households selected for the survey? Again, each cercle was
strategically selected to best represent the different agricultural systems in Mali. Similarly, the

38

selection of the six villages was purposive in order to ensure representation of intra-cercle
diversity in a region known for its economic, social, and ecological homogeneity. The criteria for
village selection were primarily based on two factors: population and access to markets. The
latter criterion simply distinguished whether market access was easy or difficult; three villages
are described by each category. Table 3 below, adapted from Samake (2008) Table 81, lists the
six villages, their population, ease of market access, average cultivated hectares per household,
and the average number of members present in the household.
Table 3: Description of Six Villages in Sample from Koutiala Cercle
Commune
Pop.
Market
Avg. Cultivated
Avg. Size of
1998
Access
Hectares per Household
Household
Nampala II
Fagui
982
Difficult
10.6
14.9
Tonon
Sinkolo
286
Difficult
7.6
14.1
Kaniko
Sincina
1735
Easy
10.9
20.2
Try I
Sincina
864
Easy
8.5
16.1
Signe
Koutialaa
1005
Easy
9.9
18.3
Gantiesso
Mpessoba
3219
Difficult
10.7
18.5
Village

The RuralStruc report states that often households differ more within villages than
between villages, on average. For this reason, households within the six villages were chosen at
random to best capture this variation. Twenty-five households were surveyed in five villages,
and twenty-eight in the sixth, but after data cleaning, this portion of the sample consists of 153
households (Murekezi 2011). However, because we have two observations for each household,
this doubles our sample size and means that each village is represented by nearly 50
observations. Because outcomes are likely to be highly correlated over time within the same
household, I am careful to control statistically for potential correlation in the econometric model
presented in Chapter 5.
Having lived east of the Koutiala cercle for two years as a Peace Corps Volunteer, I
thought it was important to include at least one village from the Yorosso cercle in the dataset; the

39

area relies heavily on cotton, is similar climatically, but is generally more isolated. Therefore, I
implemented a survey in Simona for the production year 2010/11. Based on the PROMISAM
survey used in the Koutitala cercle, households were asked a range of questions concerning their
family’s demographics, livestock and capital assets, field allocation, crop production, input
procurement, market participation, and plans for the following year. Out of Simona’s 140
households, 50 were randomly selected to participate; only one abstained, which I replaced with
a randomly-selected substitute household. This allows Simona to be fairly represented as a
village in the dataset, bringing the total number of observations up to 350. Compared to the other
villages in Table 4.1, Simona has an approximate population of 2,500, difficult market access, an
average of 11.6 cultivated hectares per household, and an average household size of 18.9 people.
Due to the amount of work required and the village’s preference for speaking Minianka,
literate locals were trained as enumerators to carry out the survey under my supervision.
Admittedly, given their lack of experience, this was a possible source of human error. Thus,
upon receiving each survey, I reviewed it privately with the enumerator, discussed any errors or
inconsistencies, and required him to return to the households if necessary. Additionally, the
survey was re-designed specifically for Simona in order to best address relevant research
questions and to assist the enumerators. For example, surveys had to be translated into Bambara
and some terminology had to be revised, such as concepts of “units” and “individual property.”
This reworking was only possible after attaining a cultural understanding of the village thanks to
many informal interviews with Simona farmers.
One main limitation of the Koutiala and Simona data is their susceptibility to human
error. All estimates of hectares planted as well as total production are based on household
memory. Thus, efforts were made by IER to include multiple household members, men and

40

women, during the survey interview to encourage collaboration and, hopefully, more reliable
estimates. For Simona, efforts were also taken to survey groups of farmers within each
household, instead of just the household head, who is sometimes removed from current farming
activity due to old age. Additionally, all survey rounds were conducted between February and
May of the production cycle, at least three to four months after harvest and nine months since
planting. While any delay in survey questioning raises the probability of human error, surveys
were implemented during these months because they are Mali’s inactive, hot season. Since
planting typically begins in May and the harvest usually ends in November, this was the optimal
time to interview a large number of households.
4.2

Dependent Variables
For any empirical application, the selection of variables must be informed by a

conceptual model. At the end of Chapter 3, it was argued that the agricultural household model is
a fair representation of how households in the Koutiala production zone make decisions
regarding crop choice and land allocation. Thus, equation (1) is repeated here as equation (2):
Q = f (P, F, V, A, K, σ, TC, Hh, πy)

reduced-form land allocation function

(2)

Again, it states that land allocation Q is a function of expected prices P, the availability of family
labor F, other variable inputs V, land area A, owned equipment (or capital) K, risk σ, transaction
costs TC, household characteristics Hh, and finally other sources of household income πy. The
challenge is now selecting data measured in the Koutiala and Simona survey rounds that best
represent these variables.
Starting with the left-side of equation (2), Q is the dependent variable, representing a
vector of crop shares for the portfolio of crops chosen by a household. As discussed in Chapter 2,
the portfolio of crops for the Koutiala production zone consists primarily of cotton, maize,
41

sorghum, and millet. Households may also cultivate peanuts, beans, or other crops, all of which
are grouped together under the category of secondary crops. The shares of a household’s total
cultivated hectares devoted to each of these crops are represented by CottonShare, MaizeShare,
SorghumShare, MilletShare, and SecondaryShare, respectively. This gives us five dependent
variables—the percentage of total cultivated land devoted to cotton, maize, sorghum, millet, and
secondary crops—the sum of which represents all cultivated land on a farm.
Table 4: Summary Statistics for Dependent Variables
Mean
Std. Dev. % 0 Share Minimum Maximum
(after 0)
CottonShare
.1354
.1243
.3714
.0561
.4615
Variable

MaizeShare

.1046

.0791

.1600

.0217

.6000

SorghumShare

.3488

.1774

.0114

.0313

1.000

MilletShare

.2401

.1607

.1143

.0435

.9231

SecondaryShare

.1712

.1512

.1886

.0014

.6957

Table 4 summarizes some basic descriptive information about these five dependent
variables, including mean, standard deviation, percentage of observations with zero share
(meaning they did not grow the crop at all), minimum shares (after zero), and maximum shares.
The standard deviations in Table 4 are revealing as they represent the variance in their respected
crop shares across households, a sufficient amount of which is necessary to produce significant
results in the econometric analysis. The standard deviations also show that crop shares definitely
vary between households.
Another important detail revealed in Table 4 is that while every crop has some
observations report a zero share, a significant portion of households are not growing cotton. In
fact, only about 63% of observations grew any cotton at all. Therefore, more so than with the

42

other crops, results that suggest a decrease in the expected share of cotton due to a change in the
explanatory variable may explain either a marginal decrease in the share of land devoted to
cotton or, perhaps for some households, dropping cotton altogether. This will also be an
important detail for interpreting the results.
4.3

Independent Variables
Continuing with the right side of equation (2), variables present in the dataset were

selected to represent the explanatory factors for crop shares as informed by the reduced-form
land allocation function and the agricultural household model. On the next page, Table 5 displays
these variables with a brief description. After, further explanation is also provided for each factor
represented in equation (2), including whether or not it is represented in the model and why.

43

Variable
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

Table 5: Summary of Independent Variables
Description

Mean

LABOR AND LAND: Household Members per Hectare and Inactivity
MenPerHct
# of adult males (age ≥15) per hectare
0.42
WomenPerHct
# of adult females (age ≥15) per hectare
0.50
YouthPerHct
# of children (6 ≤ age ≤ 14) per hectare
0.57
InfantsPerHct
# of children (age ≤ 5) per hectare
0.33
WomenPerHct*
Interaction term between (2) and (4)
0.22
InfantsPerHct
%MenInactive
% of adult males who are inactive
0.05
MenPerHct*
Interaction term between (1) and (10)
0.02
%MenInactive
%WomenInactive % of adult females who are inactive
0.12
WomenPerHct*
Interaction term between (2) and (12)
0.06
%WomenInactive
CAPITAL: Farming Equipment
Sprayers
# of sprayers owned
1.17
Plows
# of plows owned
1.75
Oxen
# of draft oxen owned
2.95
Plows*Oxen
Interaction term between (15) and (16)
7.35
TRANSACTION COSTS: Modes of Transportation
Motorcycles
# of motorcycles owned
0.63
Bicycles
# of bikes owned
2.11
Carts
# of draft animal carts owned
0.99
HOUSEHOLD CHARACTERISTICS: Literacy and Ethnic Identity
%MenLiterate
% of adult males who are literate
0.64
MenPerHct*
Interaction term between (1) and (6)
0.27
%MenLiterate
%WomenLiterate % of adult females who are literate
0.50
WomenPerHct*
Interaction term between (2) and (8)
0.26
%WomenLiterate
Bambara
Dummy: 1 if Bambara ethnicity, 0 if not
0.13
Senoufo
Dummy: 1 if Senoufo ethnicity, 0 if not
0.09
Peulh
Dummy: 1 if Peulh ethnicity, 0 if not
0.06
OtherEthnic
Dummy: 1 if other ethnicity, 0 if not
0.03
TIME AND LOCATION: Year and Village
Year_2010
Dummy: 1 if from year 2009/10, 0 if not
0.43
Village_Tonon
Dummy: 1 if from Tonon, 0 if not
0.14
Village_Kaniko
Dummy: 1 if from Kaniko, 0 if not
0.14
Village_TryI
Dummy: 1 if from Try I, 0 if not
0.14
Village_Signe
Dummy: 1 if from Signe, 0 if not
0.14
Village_Gantiesso Dummy: 1 if from Gantiesso, 0 if not
0.15
Village_Simona
Dummy: 1 if from Simona and 2010/11, 0 if not
0.14

44

Std Dev
0.23
0.30
0.59
0.32
0.57
0.13
0.06
0.17
0.09
0.87
1.86
2.72
17.87
0.79
1.49
0.54
0.40
0.25
0.45
0.32
0.34
0.28
0.23
0.18
0.50
0.34
0.35
0.34
0.35
0.36
0.35

4.3.1

Labor and Land: Household Members per Hectare and Inactivity

Labor for crop production and food preparation is a key production factor that may also
explain land allocation decisions. In the Koutiala and Simona surveys, family labor was
measured for each individual by asking whether they typically farmed in the mornings,
afternoon, or all day. However, this information cannot represent a household’s labor input in a
land allocation model. Firstly, while this method measures labor input on a given day, it ignores
that labor demands change seasonally, as planting and harvest times often require the most labor.
It is not uncommon to see the entire household—at least ages eight and up—helping to pick
cotton all day for the few dry days when the cotton boll is mature. However, even if this is the
only time a woman works all day, the household head may indicate that she is a full-time farmer,
because she is when she needs to be. Accurate measures of degrees of household labor must
include a seasonal component to capture the labor demands of planting, weeding, applying
variable inputs, and the harvest. Yet even if this more detailed data were available, it would be
best for a model trying to predict crop production, but might have still generated skewed results
since labor allocation decisions can be endogenous to decisions regarding land allocation and
crop choice. For these reasons, these data are not used to estimate family labor in this study.
Rather, we will represent family labor with household size, which in Mali’s Koutiala
production zone, is assumed to be the household’s available supply of labor for farming or other
activities. Additionally, a household’s size largely defines its consumption needs. However,
since our focus is on production potential and consumption demands, a simple headcount to
calculate household population is not an accurate representation, as it measures infants on par
with adult males. One option then is to calculate an adult equivalence measure that assigns
weights to the headcounts of different age and gender categories based on each category’s

45

expected labor input and food consumption. However, one weight could not accurately adjust for
both—for example, an infant eats something, but contributes nothing to production—and an
estimation’s dependence on weights is never ideal when a better option is available. Thus, the
model starts by including multiple variables for each relevant age and gender categories: adult
men, adult women, youth (ages 6-14), and infants (ages 0-5).
It is also important to consider the supply of land available to the household to cultivate.
In this study, the household’s number of total cultivated hectares is the measure for household
land area. It is the sum of the hectares devoted to each crop that is used to calculate the shares for
the dependent variable. However, for over two-thirds of households, the sum of cultivated
hectares is not the total number of hectares that the household possesses. The difference is
comprised of land rented to other households, newly created farm land not yet ready for
cultivation, fallow land that is set aside to regain fertility, and unused arable land.
By choosing total cultivated hectares as the measure of available farm land, I assume that
this final category—unused arable land—is non-existent or at most trivial. It is my opinion that if
a household possesses land that is suitable for farming, it will cultivate something there, even
though a common environmental concern is that households continue to cultivate land that needs
to rest every few years. This is the same assumption made by the survey instrument, which
collected data on hectares of fallow, rented, and newly created land, but left no room to record
hectares of unused arable land. Thus, the sum of cultivated hectares, and not total hectares, is the
measure used to represent a household’s supply of farm land.
Now, one might expect that if crop production is largely used for household
consumption, then land area must be looked at in terms of how many mouths it can feed. In other
words, the effect of land area on crop share decisions depends on how many people are working

46

and eating off that land. For example, I expect households with a higher rate of people per
cultivated hectares to produce more maize, which yields more calories per hectare than sorghum
or millet. Therefore, the four variables representing headcounts of adult men, adult women,
youth, and infants are calculated on a per-hectare basis. MenPerHct and WomenPerHct measure
males and females, respectively, ages 15 and up per hectare, YouthPerHct measures children
between ages 6 and 14 per hectare, and InfantsPerHct measure children between ages 0 and 5
per hectare. Together, these four variables represent a household’s labor supply and consumption
demand relative to the household’s available supply of land. I have also inserted an interaction
term between WomenPerHct and InfantsPerHct to control for an expected negative effect of
children ages five and under on their mother’s (or other caregiver’s) labor productivity. While
Malian women are fond of slinging their babies to their back while working, a high number of
young children relative to potential female adult caregivers sharply changes the composition and
potential labor supply of the household, and thus may affect crop choice and land allocation.
Another related factor that may affect crop choice is what percentage of adult men and
women are inactive. An adult may be inactive because he is elderly, severely or chronically ill,
disabled, or for some women, in the final stages of pregnancy. In these situations, the individual
often has similar consumption needs as other adults in the household, but is limited in their
ability to contribute to farm labor. To consider these cases, the variables %MenInactive and
%WomenInactive are included to measure the percentage of household adult men and women,
respectively, which were considered by the household to be “inactive” during the farming
season. Interaction terms between these variables and MenPerHct and WomenPerHct,
respectively, are also added to the model to capture how the relationship changes across
households of varying size. It is important to note that while these individuals may contribute

47

little or none to farm labor, this rarely means that they do not work in any capacity. Often
inactive women will care for young children and assist with food preparation, while inactive men
may take up some other income-generating activity (e.g., in Simona, elderly men seemed to
prefer rope-making while socializing with friends). Yet these contributions are normally small,
and so land allocation decisions must consider inactive family members.
Finally, as there can be many types of farm labor, the characteristics of the land
possessed by the household—such as its elevation, soil type, position in the water table, and
nutrient availability—can vary greatly as well. Unfortunately, data were not collected on these
attributes. However, these land characteristics are somewhat controlled for with the village
dummy variables discussed further below, though admittedly, such characteristics can often vary
even within the same village. This may be an important consideration in future studies.
4.3.2

Capital: Farming Equipment

For rural farmers in Mali’s Koutiala production zone, capital accumulation is basically
the attainment of farming equipment that can improve agricultural production. Examples range
from motorized tractors to ‘dabas’, a hand-held tool that breaks up earth like a hoe. However, no
farmer in the sample has a tractor and every household (I assume) has a daba. The key is to
identify capital that is used by enough households to generate reliable estimates, but still varies
between households.
The variables Sprayers and Plows represent the number of sprayers and draft plows,
respectively, owned by the household. Similarly, the variable Oxen represents the number of
bulls or oxen the family owns to pull the draft plow or other draft-powered equipment. The
alternative to selecting a few choice types of capital is to calculate an all-inclusive equipment
index, but this would require subjective weighting that may ignore how the possession of crop48

specific technology encourages its production. For example, sprayers are used for pesticide
application, which is mostly practiced on cotton production. Thus, one might expect that owning
high number of sprayers would help convince a household that they can handle growing a larger
share of cotton. Finally, an interaction term is included between Plows and Oxen, because I
expect to see a high correlation between the number of draft animals and draft-powered plows
owned by the household. After all, each is only functional if it has the other.
By including capital in the model, there is a risk of simultaneity bias; that is, maybe a
household owns many sprayers because of other factors that lead the household to prefer cotton
production. Sprayers do not lead to more cotton, but cotton leads to more sprayers. I justify this
possible inconsistency by looking at the time-frame in which crop decisions are made. As part of
a long-term plan to increase cotton shares, a household may plan on investing in additional
sprayers, plows, and draft animals over many years. However, at the start of planting, the share
of land devoted to cotton is not decided by the household’s long-term goals, but by their current
capital constraints (among other things), especially since it is unlikely that a household has
excess cash or credit during this season to purchase more equipment. Since this study does not
examine crop shares over a long stretch of time, we will assume that a household’s existing
capital is a determining factor of its yearly crop share decisions, and not vice versa.
4.3.3

Transaction Costs: Modes of Transportation

As discussed in Chapter 3, transactions costs—those associated with the transport and
exchange of market goods—have been modeled as a main contributor to market failures.
Transportation costs incurred by a market vendor, or the long walk to market for a household
buyer, may raise the price of a product enough to prevent an exchange from being mutually
beneficial. However, estimating transaction costs per household is very difficult. Not only is it
49

about distance to and food availability at nearby markets, which can greatly vary even within one
spread-out village, but also consists of relationships between household members and market
vendors and the ability to transport crops or travel long distances. I will incorporate some
variables here to represent transaction costs as best as possible with the available data.
First, specific village-level dummy variables represent ease of market access, as
described in the RuralStruc classification of the sampled villages, shown above in Table 4.1. It
states that observations identified by the variables Village_Kaniko, Village_TryI, or
Village_Signe have “easy” access to a market. While the RuralStruc report does not explicitly
reveal how these labels were assigned, a rough spatial mapping of the villages reveals that
relative distance to the city of Koutiala was probably the test. The three villages defined as
“easy” are approximately twenty kilometers or less from Koutiala, which could be traveled back
and forth by donkey cart on market day if need be. However, while Koutiala is the primary
market town for both Koutiala and Yorosso cercles, this does not mean that it is the only
available market; in fact, even Simona had its own market. Thus, while all villages likely have
access to a market, the villages of Kaniko, Try I, and Signe are within walking distance of a large
market town that exchanges a far greater range of goods and services. Since most goods in the
smaller markets probably went through Koutiala or a similar large town (e.g., Sikasso), access to
trade in Koutiala likely reduces transactions costs for market vendors.
Access to private transportation can also reduce transactions costs incurred by a
household. To start, the variables Motorcycles and Bicycles are equal to all of the motorcycles
and pedal bicycles, respectively, that are owned by a household (no household in the sample
owned a car or truck). Though capable of carrying a small load, these vehicles are primarily for
transportation of people. However, both can reduce transportation costs between locations, and

50

greater mobility can improve other transaction costs as it allows a household to better maintain
relationships across distances. For example, many farmers in Simona made a point to travel to a
larger market in the cercle at least once a month in order to greet family, friends, and vendors
that they knew. Among other reasons, these farmers understood that failure to greet on a regular
basis could quickly lead to a breakdown in the relationship. Therefore, one might expect that
households with greater transportation capabilities travel to regional markets more often, and
thus will face fewer other transaction costs when trading crops at market.
Next, the variable Carts, which counts the number of carts owned by a household, assists
in transporting large amounts of goods into or away from the city or nearest market.
Additionally, carts can be a production factor, as they also help households transport cotton and
coarse grains to storage after the harvest. An assumption made here is that each cart is
accompanied by a draft animal that is probably a donkey, but possibly a horse or mule. Between
the village dummies and variables for household ownership of motorcycles, bicycles, and carts,
the model controls for potential differences in transaction costs in the best way the data allow.
4.3.4

Household Characteristics: Literacy and Ethnic Identity

The first step in characterizing a household is to define its size and composition; this was
done in the series of variables that describe a household’s labor supply and consumption demand
per hectare. After that, other important household traits to consider are the education levels of its
adult men and women as well as the ethnic group to which the family identifies. As discussed in
Chapter 3, the rationale behind the agricultural household model is that because many
households consume their own production, one should expect that variables representing
differences in preferences have a significant impact on crop share decision-making.

51

Starting with education, the variables %MenLiterate and %WomenLiterate are defined by
the percentage of the household’s adult men and women, respectively, who are described as
literate by the household. These variables are also included as interactions with MenPerHct and
WomenPerHct, respectively, to account for changes in the impact that adults can have on
farming decisions depending on the size of household. This set of four variables incorporates
literacy as an explanatory factor in the model. Directly, literacy can simplify trade and reduce
transaction costs, which may cause a highly literate household to grow more of a marketable
crop. Literacy also serves as a proxy for other factors that are difficult to measure, such as
intelligence and personal motivation—characteristics that increase the likelihood of completing
literacy education and may affect crop share decisions over the long-term. Moreover, literacy
may affect individual preferences if the literacy class was accompanied with nutrition education,
as is often the case. It will be important to recall what explanatory factors these variables are
capturing when interpreting the results.
Notice that in this empirical model, the educational level of the household head is not
considered by itself, as is common in many development studies. The reasoning for this is twofold. First, as discussed in Appendix 1, I assert that while household decisions are authorized or
approved by the household head, most adult males participate in the decision-making process
depending on the respect and authority granted to them by the family. The eldest brother or
elderly father will consider the input of other married men and may grant special weight to their
opinions if they are the hardest working (motivation) or wisest (intelligence). Secondly, the
intelligence and motivation of a household head is often seen in the percentage of his household
members who are educated, even if he is not formally educated himself. One case comes to mind
of a respected elder with nearly a dozen sons, all of whom had received some formal education

52

and most of whom were literate. It is likely that some of the motivation behind the sons’ studies
was the result of their father’s motivation and intelligence, though the elder himself was illiterate
and too old (and proud) to really benefit from taking a class. If a variable were to measure
household intelligence and drive by only acknowledging his personal educational background,
and not that of his family’s, it would have failed to capture his own wisdom.
Additionally, variables that measure adult female literacy, and by extension her
intelligence and motivation, are also included in the model. While women are not very influential
in the patriarchal decision-making process, household leadership will often allow one or more
women to produce crops on their own, especially some of the secondary crops which require less
equipment to cultivate. Opportunities to do so may be through a garden managed by the village
women’s cooperative or a smaller parcel of land devoted to the production of beans or peanuts.
While such instances may be the result of very deliberate plans by household leadership to
maximize labor supply, it often is accompanied—if not entirely motivated—by the household
women, either by a desire to generate income or provide a more diverse or nutritious diet for
their family. However, these efforts are often perceived as inferior farming and must be in
addition to household activities and fieldwork on primary crops during peak labor demands.
Specifically, literacy training can be expected to encourage women’s agricultural
production in a few ways. First, being literate may assist with market trade or cooperative
participation, making it more profitable to cultivate other crops. Secondly, the confidence gained
by a woman after having acquired a skill that many men lack may also motivate her to pursue
interests in farming or gardening. Furthermore, training received on nutrition and gender
empowerment, which usually accompanies literacy programs and classroom studies, may
encourage crop production and change her preferences toward more nutritious foods. And as

53

stated before, literacy also serves as a proxy for inherent intelligence and drive; one may expect
that gifted and motivated women are more likely to seek and complete literacy training as well as
make additional contributions to household agricultural production.
Next, a series of binary dummy variables representing ethnic groups are incorporated into
the model. Since the vast majority of households in the Koutiala production zone identify
themselves as Minianka, this ethnic group will be the omitted category to which other ethnic
groups are compared. The variables Bambara, Senoufo, Peulh, and OtherEthnic represent
alternative ethnic identities, and the estimated coefficients on these variables will represent the
changes in crop shares associated with each ethnic group relative to the Minianka group. Since
each are a binary variable (=1 if part of the ethnic group; =0 if not), the means presented in Table
5, when multiplied by 100, are the percentage of the sample that identify with that ethnic group.
The variable OtherEthnic is comprised of five ethnic groups, the Soninke and Malinke people
from the west and the Bozo, Bobo, and Dogon from north-eastern Mali. Together, they total only
twelve households, which is why they are grouped together into one variable, despite the fact
that they represent very diverse cultures. For example, most Bozo reside along the Niger River
and rely heavily on fishing for their food and livelihood, which is a unique identity for an ethnic
group in a landlocked country.
The purpose of dummy variables for ethnic identity is that it may hint at each
household’s tastes and preferences. If the theory behind the agricultural household model is
correct, then in the presence of transaction costs or other barriers to market transactions, a
household’s tastes and preferences may factor directly into its production choices, including for
crop shares. While an ideal analysis might gather information on household food preferences
directly or weight each individual’s food preferences according to household seniority, this

54

subjective information would be difficult to collect and interpret. In its place, ethnic identity can
explain some of the differences in tastes and preferences across households as certain groups
prefer traditional crops and meals, particularly those households migrating from distant regions.
For example, those from the north may be less familiar with how to grow and prepare maize,
preferring to focus on millet, sorghum, or fonio instead. Of course, this assumes that recent
immigrants to new areas still retain many of the traditions and preferences of their home regions.
Cotton shares may even be associated with ethnic group as the Minianka and Senoufo peoples,
originating from Koutiala’s Old Cotton Basin, may have retained cultural knowledge regarding
its production as well as perceive the cash crop as linked to their ethnic identity.
One concern is that some of the ethnic identities are highly correlated with certain
villages. For example, 60.0 percent of the Peulh observations are in one village, 73.9 percent of
the Bambara are in another, and finally, 93.3 percent of Senoufo are in yet another. For this and
other reasons, village-level dummies are also included in the model to control for differences in
geographic location across the sample.
4.3.5

Expected Price

If supply response models have taught anything, it is that farmers respond to expected
price changes. However, spanning merely three years, only two of which cover the same
households, the data do not support empirical estimates of how price could affect crop shares,
even if the data were used as a two-year panel dataset. Rather, the differences in expected price
will be captured by dummy variables representing the time and location from where and when
the observations originate. This approach assumes that households from the same village and in
the same year have very similar price expectations. Given that price expectations and land

55

allocation strategies are a common topic of discussion among adult men in the dry seasons
before planting, this assumption—while not true in every case—is practical.
Specifically, the time dummy variables will represent differences between the 2008/09,
2009/10, and 2010/11 growing seasons and expected changes in price that may have occurred
due to unique weather patterns for each of those years. Additionally, the presence of village
dummy variables will control for geographic differences in the expected price of coarse grains,
including those that may result from foreseen transaction and transportation costs, which were
already discussed. However, there is no need to control for geographic differences in the
expected price of cotton, since, as stated in Chapter 2, CMDT is mandated to set a pan-territorial
producer price for cotton before each planting season.
Finally, the variable Village_Simona identifies the fifty observations from the village in
which I served as a Peace Corps Volunteer, which lies east of the other six villages in the
Yorosso cercle. These data also represent land allocation at least a year later than the others,
during the 2010/11 farming season. Therefore, the Village_Simona dummy variable represents
an expected price change through both a time and location difference.
4.3.6

Determinants Not Included

Mostly due to a lack of sufficient data and possible simultaneity biases, some of the
determinants shown in equation (2), the reduced-form land allocation function, are not
represented in the model. These are profits generated by off-farm employment (πy), variable
inputs (V), and a household’s risk preference (σ). This final subsection of Chapter 4 explains the
justification for the exclusion of these variables.
The first determinant category not included in the model or Table 5 is profits generated
by off-farm employment. Originally, this was partially represented in the model by several
56

variables equal to a household’s number of cattle, sheep, goats, hogs, and poultry. However, it
was determined that animal-raising is too closely interrelated to crop production as part of the
farming system, thus leading to a simultaneity bias that also does little to explain the effects of
off-farm work on crop shares. Ideally, off-farm employment would also include profits gained
(or lost) from migratory household members, many of whom go to larger cities such as Koutiala
or the capital Bamako to find work or study. Additionally, off-farm income can come from work
as a service provider in a village, such as a mechanic, carpenter, shopkeeper, or blacksmith.
Unfortunately, in either case, adequate data were not collected in the surveys to provide a
reliable estimate of profits earned (or lost) from these activities. Yet, based on my observation,
in-village service providers are the least available during rainy seasons since farming becomes
their top priority. In Simona, every single household participated in agricultural production,
including the top blacksmith, village chief, imam, and pastor, and this is not unique for small
villages. Yet while agricultural production may be important for all households in remote
communities, it will remain unclear in the analysis how crop shares may be affected by these
alternative income sources.
Next, variable inputs are the second determinant in equation (2) that is not represented in
Table 5. Originally, it was included in the model since most households order their fertilizer
from CMDT long before planting and because knowledge of planned fertilizer application can
affect crop shares. For example, a farmer can produce the same maize on his field as his
neighbors do if he plants less but uses more fertilizer. However, even though data pertaining to
fertilizer type, purchase method, and quantity applied is available, a variable to represent inputs
is ultimately excluded due to issues of simultaneity. First of all, it is unrealistic to assume that
fertilizer procurement always comes before planting decisions for coarse grains. Households can

57

decide after planting to apply more or less fertilizer than anticipated due to unexpected changes
in weather. Other households may even wait and purchase fertilizer from a market vendor.
Secondly, because fertilizer is a requirement for successful cotton production, it must be
considered a simultaneous decision. In other words, if you grow cotton, you will use fertilizer;
the two decisions go together. This high correlation between fertilizer and cotton hectares also
makes fertilizer an unfair estimator of coarse grain shares; households without fertilizer grow no
cotton and have higher shares of coarse grains, while those with fertilizer also grow cotton and
thus have lower shares of coarse grains. Other options might include finding proxies to represent
the price or availability of fertilizer, but government subsidies and the ability of households to
procure through CMDT (discussed in Chapter 2 and Appendix 2) mean that these measures are
identical across most, if not all, households in the sample. Therefore, representation of variable
inputs such as chemical fertilizers is left out of the model.
Finally, a household’s risk preference is the other category present in equation (2) not
represented in Table 5. Here, risk does not refer to the vulnerability of a household to certain
shocks, but rather the household’s preference for accepting or averting risk and how this attitude
affect their land allocation decisions. However, while understanding a household’s risk
preference may explain much of their decision-making, it is not included in the model for two
reasons. First, risk aversion preference is terribly difficult to measure or proxy at the household
level, especially because it is perceived differently by each individual—their estimated
probabilities that given events will occur (e.g., weather) as well as aversion behavior. Then of
course, this must be weighted to calculate risk aversion at the household level. Secondly, risk
preference is tricky to represent in an econometric model because it is difficult to truly separate
from every other independent variable. Risk preference are an engrained part of the decision-

58

making process, and therefore are in embedded in other variables. For example, choosing to own
a motorcycle or expand a field require an assessment of risk. Since these variables are among the
explanatory variables, they can also represent a household’s risk preference indirectly. Similarly,
there is no explicit measure of household wealth, but measures of the total cultivated hectares
and ownership of farming equipment reflect household wealth indirectly. As with wealth, a
general representation or proxy for risk aversion preference is another variable that ought to be
included, but is not due to a lack of such a proxy in the data.

59

CHAPTER 5: ECONOMETRIC MODEL AND RESULTS
This chapter begins with a description of the fractional multinomial logit model—the
econometric technique necessary to estimate the effect of the explanatory variables on the shares
of multiple crops simultaneously. It then presents the results in the form of average marginal
effects. Later sections discuss the results for specified categories of related explanatory variables,
consistent with classification used in Chapter 4. The chapter ends with an application of the
estimated average marginal effects to two example scenarios that characterize a smaller,
disadvantaged household and a larger, better-equipped household.
5.1

Fractional Multinomial Logit Model
As its name suggests, the estimation technique used in this paper combines two variations

on the standard logit model: the fractional logit and the multinomial logit, all three of which are
summarized in Appendix 4. The outcome (i.e., the fractional multinomial logit) is that the
explained variable y is able to represent the different shares of various types of y, all of which
sum to one, much like the various categories in a pie chart. For this reason, the model is in the
family of multivariate fractional logit models (e.g., Mullahy 2011; Murteira and Ramalho 2012),
because it is measuring the changes in shares of multiple variables simultaneously as a result of
some explanatory variables. In other words, it allows one to ask how the slices of a pie chart
change between observations as a result of differences in a certain set of related factors. In this
case, the whole pie chart is a household’s total number of cultivated hectares, meaning that the
fractional multinomial logit model can help to see how changes in market and household
characteristics affect the share of land devoted to cotton, maize, sorghum, millet, or secondary
crops.

60

Combining some main elements of the fractional logit and the multinomial logit models
6

to come up with the fractional multinomial logit model is fairly straightforward. The fractional
logit model differs from the standard logit model as it treats the dependent variable as an
expected value defined by an interval rather than a response probability (Papke and Wooldridge
1996). Similarly, the fractional multinomial logit model must ensure that the expected share of
any outcome j lies between parameters A and B and that the sum of shares for all outcomes sums
to unity. Mathematically,
A ≤ E(Sj | x) ≤ B, j = 0,…, J, where A=0 and B=1,

(3)

J

∑ j=0 E(Sj | x) = 1

(4)

This technique permits the evaluation of shares of total farm land instead of the probability of
whether or not a crop was cultivated.
The multinomial logit describes a technique for comparing the response probabilities for
several categorical variables through use of a pivot outcome, which is the difference between one
and the sum of expected shares for all other outcomes. Likewise, the fractional multinomial logit
model defines a pivot outcome as well, but again, its dependent variables are fractional outcomes
(i.e., crop shares), not response probabilities. Defining j = 0 as the pivot outcome, the fractional
multinomial model also must establish expressions for every outcome within the logit
framework.
E(Sj | x) = G(β0 + βkx k) = G(z) = exp(z)/[1 + ∑exp(z)],

j = 1, 2, …, J.

(5)

E(S0 | x) = G(β0 + βkx k) = G(z) = 1 / [1 + ∑exp(z)],

j=0

(6)

6

Again, see Appendix 4 for a summary of the logit, fractional logit, and multinomial logit
estimation methods.
61

Use of the pivot outcome equation (6) to estimate multiple outcomes makes it possible to
evaluate the effect of explanatory variables on several crops simultaneously. Therefore, when
joined together, the fractional multinomial logit model estimates coefficients which predict the
expected share of several categorical outcomes within a defined interval, such as the share of
cultivated land that a Malian household devotes to various crops.
By embedding the fractional logit function into the multinomial logit quasi-likelihood
function, the econometric model can measure shares of outcomes—not probabilities—in what is
a simplified form of the log likelihood function (Mullahy 2011). This new function, as a member
of the linear exponential family, uses a quasi-maximum likelihood estimator (QMLE) and is
efficient and consistently normally distributed provided the fractional logit function holds true
7

(Ye and Pendyala 2005). The QMLE approach will maximize this new function and, with the
assistance of a fractional multinomial logit STATA package (Buis 2008, updated 2012), run until
it has converged and is able to predict crop shares.
However, because the multinomial logit estimator requires some normalization, these
QMLE estimates will correspond to the coefficients in the multinomial shares model. Thus, it
produces coefficients that may be difficult to interpret (Mullahy 2011). For this reason, using the
coefficients predicted from an estimation of the fractional multinomial logit model, I calculate
average marginal effects for every explanatory variable on each crop outcome, taking into

7

For a more detailed description of fractional logit models and the quasi-maximum likelihood
technique, see Papke and Wooldridge (1996) or Ye and Pendyala (2005). For specifics on the
fractional multinomial logit model as well as information related to specification testing or other
multivariate fractional models, see Koch (2010), Mullahy (2011), and Murteira and Ramalho
(2012).

62

8

account the coefficients for interaction terms when applicable. Section 5.2 will present the
estimated average marginal effects, calculated from coefficients estimated with a fractional
multinomial logit model, and then discuss their economic meaning.
5.2

Presentation of Average Marginal Effects
Using the data described in Chapter 2, the variables informed by the agricultural

household model in Chapter 3, the data described in Chapter 4, and the fractional multinomial
logit model explained above, this thesis will now present the mode’s results and average
marginal effects on crop shares due to changes in various explanatory variables.
Drawing from 334 observations, the fractional multinomial model converged on a log
pseudo-likelihood of -480.03 with a Wald chi-squared of 2004.73. To control for potential
correlation over time within the same household, observations were “clustered” by a household
identification number to ensure that standard errors were estimated robustly. The results of the
fractional multinomial logit pivoted off of CottonShare are provided in the Appendix 7.
Now, this chapter focuses its attention on the average marginal effects of the independent
variables on crop shares, as presented in Table 6. Average marginal effects that are statistically
different from zero at the 10%, 5%, and 1% levels are indicated with one, two, or three asterisks,
respectively; coefficients that are not statistically different from zero at the 10% level or below
receive no asterisk. Of the model’s 125 coefficients for average marginal effects, only 30 are
significant at the 10% level, though some are close.

8

See Appendix 5 for a specification of the fractional multinomial logit model in STATA and
Appendix 6 for the code used to calculate average marginal effects.
63

Table 6: Average Marginal Effects of the Independent Variables (Derived from Results of Fractional Multinomial Logit)
Log pseudolikelihood = -480.02629
Obs: 334
Wald Chi^2: 2004.7
Prob > Chi^2: 0.0000
Cotton Share
Maize Share
Sorghum Share
Millet Share
Secondary Share
Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig
MenPerHct
0.0032 0.0400
0.0422 0.0267
-0.0714 0.0465
0.0080 0.0496
0.0180 0.0452
WomenPerHct
-0.0434 0.0464
-0.0014 0.0323
0.0148 0.0542
0.1189 0.0597 ** -0.0889 0.0557
YouthPerHct
0.0074 0.0225
0.0150 0.0246
0.0424 0.0298
-0.0962 0.0358 *** 0.0314 0.0348
InfantsPerHct
0.0580 0.0292 ** 0.0024 0.0315
0.0428 0.0512
-0.0347 0.0541
-0.0684 0.0528
%MenInactive
-0.0817 0.0653
0.0377 0.0304
0.1324 0.1031
-0.0841 0.0757
-0.0044 0.0557
%WomenInactive -0.0223 0.0577
0.0143 0.0343
0.1054 0.0745
-0.0703 0.0472
-0.0271 0.0501
Sprayers
0.0202 0.0066 *** -0.0013 0.0052
-0.0210 0.0102 ** -0.0275 0.0128 ** 0.0296 0.0141 **
Plows
0.0053 0.0071
0.0063 0.0059
-0.0001 0.0102
0.0011 0.0108
-0.0126 0.0124
Oxen
0.0071 0.0050
0.0049 0.0027 * 0.0030 0.0062
-0.0064 0.0051
-0.0086 0.0065
Motorcycles
-0.0009 0.0133
0.0054 0.0088
-0.0270 0.0161 * 0.0197 0.0122
0.0027 0.0146
Bicycles
0.0012 0.0050
-0.0014 0.0038
0.0057 0.0088
0.0038 0.0084
-0.0093 0.0082
Carts
0.0311 0.0154 ** -0.0037 0.0116
-0.0287 0.0167 * 0.0015 0.0202
-0.0003 0.0215
%MenLiterate
0.0041 0.0171
-0.0075 0.0192
-0.0292 0.0412
-0.0120 0.0313
0.0446 0.0326
%WomenLiterate 0.0250 0.0293
0.0182 0.0231
-0.0379 0.0481
0.0013 0.0493
-0.0066 0.0607
Bambara
0.0176 0.0388
-0.0484 0.0119 *** 0.0393 0.0517
-0.0132 0.0332
0.0048 0.0474
Senoufo
-0.0316 0.0244
0.0009 0.0226
0.0590 0.0466
-0.0271 0.0351
-0.0013 0.0442
Peulh
-0.0076 0.0301
-0.0361 0.0197 * 0.0677 0.0468
0.0516 0.0908
-0.0756 0.0664
OtherEthnic
-0.0441 0.0305
-0.0523 0.0169 *** 0.0106 0.0622
0.1167 0.0713
-0.0310 0.0356
Year_2010
0.0635 0.0213 *** -0.0096 0.0176
0.0098 0.0417
-0.0235 0.0398
-0.0401 0.0542
Village_Tonon
-0.0328 0.0294
0.1010 0.0277 *** 0.1386 0.0517 *** -0.1595 0.0298 *** -0.0473 0.0420
Village_Kaniko
-0.0306 0.0247
0.0404 0.0244 * -0.1181 0.0418 *** 0.0211 0.0334
0.0871 0.0494 *
Village_TryI
-0.1290 0.0137 *** -0.0161 0.0235
0.0103 0.0536
0.0014 0.0338
0.1335 0.0593 **
Village_Signe
-0.0664 0.0223 *** 0.0286 0.0267
-0.0449 0.0489
0.0428 0.0497
0.0399 0.0434
Village_Gantiesso 0.0069 0.0336
0.0223 0.0189
-0.0742 0.0452
-0.0952 0.0278 *** 0.1402 0.0482 ***
Village_Simona
0.0887 0.0382 ** 0.0453 0.0208 ** 0.0156 0.0362
-0.0735 0.0273 *** -0.0761 0.0310 **
Significance Legend:
*** = P<.01
** = P < .05
* = P < .10

64

A few other points must be made about the interpretation of the coefficients in Table 6.
For non-binary variables, the coefficients represent the mean of the change in crop shares as a
result of a marginal change in the explanatory variables for all observations. So for example, the
first coefficient for MenPerHct under the outcome CottonShare is 0.0032, which suggests that a
one-unit increase in MenPerHct, all else equal, is associated with an average increase of 0.32%
for land allocated to cotton across all households, though in statistical and economic terms, this
is no different than zero. For binary variables, the coefficients represent the average change in
crop shares resulting from a shift in the variables’ minimum to its maximum, across all
households. Thus, the coefficient for Bambara under CottonShare is 0.0176, which suggests
that—relative to a Minianka household—a Bambara household has an average of 1.76% more
land devoted to cotton. The upcoming discussion will highlight coefficients deemed to have
economic and statistical relevance to the research questions.
Furthermore, because crop shares must always sum to one—as they are defined by a
finite amount of total cultivated hectares—the sum of the average marginal effects for any one
variable is zero; in other words, what an independent variable takes away from some crop shares,
it has to give to others. Additionally, the average marginal effects of interaction terms are not
presented as such a measure does not exist; under the assumption of “all else equal,” an
interaction term has no marginal effect as it is only the product of two other explanatory
variables. However, the coefficients estimated for the interaction terms by the fractional
multinomial logit model were incorporated into the calculation of the average marginal effects
9

for those explanatory variables involved in the interaction term. In short, while the interaction

9

Again, see Appendix 6 for the STATA code used to calculate average marginal effects.
65

terms have no measure of average marginal effect for themselves, the consequences of the
interactions are present in the average marginal effects in Table 6.
5.2.1

Labor and Land: Household Members per Hectare and Inactivity

To start, I evaluate the effect of labor and land supply on crop choice and land allocation.
Here, variables are expressed in terms of household members per hectares, but for ease of
interpretation, the results will be analyzed specifically for households with 6.25, 9, and 12.25
hectares of total cultivated land. These values represent the 25th, 50th, and 75th percentile of total
cultivated land in the sample, which will proxy for a small, medium, and large farm in the
following discussion.
The results indicate that every additional adult male per hectare is associated, on average,
with a 7.1% decrease in land allocated to sorghum, more than of half of which is transferred into
maize crop. For every additional male on a small, medium, and large farm, this translates into
1.1%, 0.8%, and 0.6% decrease in the household’s share of sorghum, respectively. Thus, this
effect seems small considering that the median household has four adult males present, and the
effect is surprisingly inconclusive for labor-intensive cotton. However, the draw away from
sorghum, which has relatively low input costs, to more marketable food crops—particularly
labor-intensive maize—is logical given that more men per hectare is usually equated to more
labor per unit of land.
Moreover, every additional adult female per hectare is associated, on average, with a
11.9% increase in land allocated to millet, more than of half of which is taken from land
allocated to secondary crops. This effect translates into a 1.9%, 1.3%, and 1.0% decrease in a
household’s share of millet for every additional woman present on a small, medium, and large
farm, respectively. Although the median household has four adult females present—as with
66

men—these effects are greater than those described for adult males per hectare. An average
marginal effect linked with an increase in millet share may suggest that women are influential in
millet production, though the converse effect on share of secondary crops, which are more often
cultivated by women, casts some confusion. Another possibility is that, all else equal, more
women per hectare may be the result of more wives among the household’s men, and thus a sign
of increased wealth. This may be associated with a slight increase in millet share, which is
considered more marketable than sorghum.
Another important consideration in a discussion about the effect of labor on crop choice
and land allocation is the percentage of adult men and women in a household who are inactive or
otherwise unable to assist with farm labor. As is, the results described in the paragraph above
assume that all household members are healthy and able to contribute to fieldwork. However, the
results find that if an additional 25% of a household’s adult men are inactive (i.e., one of the four
men in the median household), this is associated with nearly 2.0% decreases in the share of
cotton and millet and an opposing 3.3% increase in sorghum share. Further, a 25% increase in
the number of adult women who are inactive is correlated with a 2.6% increase in land allocated
to sorghum and a 1.8% decrease in millet share. While none of these coefficients for inactivity
are statistically significant at the 10% level, an interesting pattern emerges. The slight increases
in sorghum at the expense of millet and cotton support the role of sorghum as the “safe” crop,
grown by families who are vulnerable to risk and have more mouths to feed with less labor input.
Additionally, secondary crops seem to be the least affected, perhaps because inactive household
members are most able to contribute to the production of these crops; for example, elderly
women are able—and quite skilled at—cracking peanut shells, though this task could be
delegated elsewhere if no one was inactive in the household.

67

Despite providing less (if any) to the labor supply, youth and infants per hectare also
affect land allocation of crops, though probably as consumers of crop production. The results
suggest that a one-unit increase in the number of children per hectare, ages 6 to 15, is correlated
with a 9.6% decrease in land allocated to millet—nearly the opposite effect of adult females per
hectare—and small compensating increases in sorghum and secondary crop shares. In a small,
medium, and large household, this means that every additional child, ages 6 to 15, decreases
millet share by an average of 1.5%, 1.1%, and 0.8%, respectively. This effect seems to
counteract that of adult women per hectare even more if one considers that the median household
has four youths as well. It may be that these effects will cancel each other out in most cases, but
in households where there are many more adult women than youths, these women will often
focus on millet production. So conversely, when there are many more youths than adult women,
the additional work of caring and feeding these children distracts the women from millet
production and lowers its share of total cultivated land.
Finally, a one-unit increase in the number of infants per hectare, ages 5 and under, is
linked to a 5.8% increase in a household’s share of cotton, which is largely compensated by a
6.8% decrease in secondary crop share, among others. However, this is only a 0.9%, 0.6%, or
0.5% increase for every additional infant on a small, medium, or large farm, respectively, and the
median household in the sample had only two infants in the household. Still, the significance
behind this result is unexpected, as infants have nothing to contribute to cotton production
(though even children as young as eight can help during the cotton harvest), and the cash crop
will not directly provide food for them. Possible explanations include that households may be
just a bit more likely to grow cotton if they have infants, as these young children may require
more cash expenditures related to post-natal care. However, that additional infants take away

68

from land allocated to secondary crops is expected if women, in fact, are responsible for most
secondary crop production, because additional infants requires additional resources (e.g., time)
devoted to caregiving.
5.2.2

Household Characteristics: Literacy and Ethnic Identity

Next, I discuss the results related to the effect of household characteristics on land
allocation of crops. Derived from variables that represent household traits and not market or
resource constraints, these results may provide evidence for or against the agricultural household
model. As discussed in Chapter 3, this model states that because smallholder farms in Mali’s
Koutiala production zone are both producers and consumers of their field crops, due in large part
to market failures, they will take into consideration their own tastes and preferences when
making crop production decisions. The variables included in the model to represent household
characteristics relate to adult literacy and ethnic identity.
Starting with adult literacy, which serves as a measure of education and a possible proxy
for motivation and intelligence, the average marginal effects on crop shares in Table 6 resulting
from the fraction of adult men and women who are literate are statistically insignificant and thus
inconclusive. Still, the results find that a 25% increase in the percentage of household men who
are literate (an additional one person out of four) is correlated with a 1.1% increase in secondary
crop share and a nearly 0.7% decrease in sorghum share, with all the magnitude of other effects
being even closer to zero. The same 25% increase in household percentage of women who are
literate is associated with a 0.6% increase in cotton share and 0.9% decrease in sorghum share,
with negligible effects for all other shares.
That both measures of household literacy are linked to decreases (albeit small) in
sorghum share is not surprising, as one expects that education and the motivated individuals who
69

seek it likely make a household less vulnerable, possibly because they are able to support the
household in other ways. However, the positive effect of male literacy on secondary crops shares
and that of female literacy on cotton shares was unexpected, as I had predicted the opposite.
First, literate men are able to attain leadership positions within cotton producer associations, so it
seemed reasonable that more literate households could become more invested in cotton
production through participation in their association. Conversely, I had thought that the largest
positive effect for women’s literacy would be on secondary crops, since many women’s
associations tend to simultaneously promote literacy classes and secondary crop cultivation (e.g.,
gardening), though perhaps this effect is not as influential as expected. Yet, even if women’s
gardens are in fact a significant factor, it is possible that the household men did not relay this to
survey enumerators or that this production occurred in cold season, which is not included in the
data. Two other possible explanations remain. First, men often seek training to learn production
techniques for secondary crops, meaning that there well could be a relationship between adult
male literacy and land allocation to secondary crops. Secondly, the low magnitude and high
standard errors for all coefficients in this category do not provide solid evidence to validate any
significant inquiry.
Another set of variables in Table 6 contain binary variables representing ethnic identity.
The results suggest that, compared to the Minianka ethnic group, many ethnic identities in the
Koutiala production zone are associated with a smaller share of cultivated land allotted to maize.
All else equal, Bambara, Peulh, and “other ethnicity” households are correlated with a 4.8%,
3.6%, and 5.2% decrease in maize share, respectively, relative to Minianka and Senoufo
ethnicities. Both of these latter ethnic groups have inhabited the Koutiala area longer and were
more likely present when maize was first introduced and promoted by CMDT, which may help

70

to explain this difference. Yet, the difference in maize share is not directly proportional to
differences in cotton shares between ethnic groups, though the two have overlapping value
chains.
Other results are noteworthy though not statistically significant. First, all ethnic groups
present in the model are also linked to an increase sorghum share relative to the Minianka group,
ranging between 1.1% and 6.7%. Also, all are associated with a decrease in cotton share relative
to Minianka households, between 0.8% and 4.4%, except for the Bambara people. This trade-off
between cotton and sorghum could be related to cultural identity or generational knowledge
sharing, as the Koutiala production zone has long been known for its cotton production and is
mostly inhabited by the Minianka people. The trade-off also makes sense if other ethnicities are
often more vulnerable to risk, and thus preferred a safer crop like sorghum to one with high
investment costs. For example, it is possible that minority groups, having immigrated to the area,
may not be as established as the Minianka people, which may result in financial insecurity and
having to farm on second-choice land; such conditions make it difficult to consider anything but
staple crop production. Finally, Peulh and “other” ethnicity households are found with 5.1% and
11.6% greater shares of millet, along with 7.6% and 3.1% smaller shares of secondary crops,
respectively. This more closely aligns to what one might find further north in Mali—from which
many of these people originate—where an arid climate is more suitable to millet than production
of secondary crops. Ethnic identities only serve as one measure of a household’s tastes and
preferences, but evidence that these variables actually affect crop shares justifies, to some
degree, the agricultural household model.

71

5.2.3

Capital: Farming Equipment

The correlation between capital (i.e., farming equipment) and crop shares are also
estimated in the model. The results suggest that each additional pesticide sprayer owned by a
household is correlated with a 2.0% increase in cotton share and a 3.0% increase in the share of
secondary crops, which comes at the expense of a 2.1% and 2.8% decrease in sorghum and
millet share, respectively. On average, these coefficients can explain the difference in the
composition of 1/10 of total cultivated hectares between a household with zero sprayers (21% of
the sample) and another with two (26% of the sample); the latter will devote it to cotton and
secondary crops in place of millet and sorghum. This outcome is logical considering that the
sprayer is normally used for pesticide application on cotton and some types of secondary crops
and rarely on coarse grains. Input-intensive maize may be the exception, though the effect of
additional household sprayers on its share of farm land is inconclusive.
Next, results for the number of draft plows and oxen owned by the household, which
incorporates an interaction term between them, imply that they have very little average marginal
effect on any outcome, especially considering that the median household has only one plow and
two oxen to pull it. Compared to a household with no plows or oxen, the median household in
the sample is linked to 2.0% and 1.6% increase household share of cotton and maize,
respectively, made up for by a 1.2% decrease in millet share and a 3.0% decrease in secondary
crops, though only one of the coefficients is significant at the 10% level. Still, the results do not
contradict expectations, as many farmers do not risk investing in the production of cotton and
maize if they cannot properly prepare their fields, though some are able to borrow or rent; this
would lead to an average increase in the share millet and especially secondary crops, the latter of
which are usually cultivated by hand tools and benefit the least from oxen-drawn equipment.

72

Overall, these results provide evidence that owning capital that makes cotton and, to a
lesser extent, maize less labor intensive—as it enhances labor productivity—affects a farmer’s
planting decisions. Still, given that these tools are required for cotton and maize production, the
magnitudes are smaller than anticipated. While ownership of tools may be linked to higher
shares of crops that use those tools, it is certainly far from the dominant determining factor.
5.2.4

Transport

It was theorized that access to modes of transportation that may reduce transportation and
even transaction costs might affect crop shares. While over half of households in the sample do
not own a motorcycle, for those that do, each additional motorcycle is associated with an average
2.7% decrease in total cultivated land allocated to sorghum. This is largely made up for by a
2.0% increase in millet share and an even smaller increase in maize, though neither increase was
statistically significant. Ownership of bicycles had a smaller effect, even when considering that
the median household owned two, and the results here are inconclusive. At best, they suggest
that an additional bicycle may be linked with an almost 1% decrease in secondary crop share,
which is unexpected given the occasional use of bicycles to transport smaller loads of secondary
crops to and from local markets. Overall, both modes of transportation seem to have limited
explanatory power.
The effect of owning a cart to transport goods to market—or also from the fields to the
household—was also examined. Table 6 shows that an additional cart was estimated to be
correlated with a 3.1% increase a household’s share of cotton, most of which is taken from a
2.9% decrease in sorghum share, with minimal changes to shares of maize, millet, and secondary
crops. These results make sense if the cart is primarily seen as a piece of farming equipment that
assists with production rather than a way that reduces transport costs to market. Carts are very
73

useful for transporting organic and inorganic fertilizer to cotton fields and during the harvest,
when the plucked cotton needs to be kept off the ground and sent to a dry location. However,
because CMDT sends semi-trucks to collect the cotton harvest from each producer association,
the cart does not significantly reduce transportation or transaction costs to market. So while
possession of carts may influence crop choice and land allocation, it does not necessarily do so
as predicted in Chapter 3.
5.2.5

Time and Location

Variables representing time and location were mainly included in the model as controls,
but their results offer valuable insights as well. First, the results indicate that in the 2009/10
farming season, households located in the Koutiala cercle increased the share of their land
allocated to cotton by 6.4% compared to 2008/09, an effect that reflects the increase of cotton
hectares planted between that time throughout the Koutiala production zone, as seen in Figure
2.5. Even though the price of cotton fell 30 FCFA per kilo between these two years (roughly 60
US dollars per metric ton), this upward trend represents a recovery from a major producer strike
in the 2007/08 growing season (Mali Assessment 2011). However, this increase in cotton share
was done mostly at the expense of secondary crops and millet, with minimal effects on maize
and sorghum, though an average 0.4 hectare increase of total cultivated land between these years
may also explain some change in crop shares without necessarily assuming that land was
redistributed to different crops.
Continuing, the collective set of village dummy variables showcases some interesting
results, especially when one considers a primary difference between them—ease of market
access—which applies to the villages Kaniko, Try I, and Signe, as described in Table 3. Again,
ease of market access refers to the fact that these villages are approximately twenty kilometers or
74

less from Koutiala, the region’s central market, whereas others are farther away. The findings
suggest that, relative to the first village of Nampala II (the base village), which has difficult
access to Koutiala’s markets, these three villages devote less land to cotton and sorghum;
specifically, Try I and Signe have smaller shares of cotton by 12.9% and 6.6%, respectively, and
Kaniko and Signe have smaller shares of sorghum by 11.8% and 4.5%, respectively. To
compensate for these changes, Kaniko, Try I, and Signe all grow higher shares of millet,
especially relative to the other villages with poor market access, Tonon and Gantiesso.
Furthermore, Kaniko and Try I have higher shares of secondary crops (relative to Nampala II):
8.7% and 13.4%, respectively.
The patterns for the villages farther from markets are a bit more varied. Relative to
Nampala II, households in the second village of Tonon are associated with a 10.1% increase in
maize share and a 13.9% increase in sorghum share, on average, which is balanced by a 16.0%
decrease in millet share and small predicted decreases of cotton and secondary crops as well.
While the results for Gantiesso have in common with Tonon a high decrease in millet share
(9.5%), its households are correlated with 14.0% increase in secondary crop shares and a 7.4%
decrease in sorghum share.
Collectively, these results suggest that the villages with better market access to Koutiala
have more in common than those that do not, which is possible for a couple reasons. First,
because Kaniko, Try I, and Signe are all within a relatively small radius of Koutiala, it is more
likely that the villages are similar in terms of climate and soil quality. Conversely, the locations
of Nampala II, Tonon, and Gantiesso are more spread out and characteristically diverse, which
may help to explain differences between them, such as the average household’s share of
sorghum, which may differ between Tonon and Gantiesso by as much as 20%.

75

The second reason is that access to Koutiala’s markets may provide certain incentives
that motivate a specific crop choice and land allocation in its surrounding villages. Consider the
decline in cotton share for Kaniko, Try I, and Signe. It may be that access to Koutiala and its
markets gives households a greater opportunity to earn income through coarse grain or secondary
crop production, which may eliminate the need to produce cotton in order to earn cash. Also, a
key benefit of participation in the cotton industry is that CMDT helps to reduce transportation
and transaction costs for crop sales and input procurement by driving semi-trucks out to its
village—no matter how remote. In Simona, where one farmer remembers his grandfather
walking 100 kilometers to sell cotton in Koutiala, this service continues to be an influential
incentive because they have difficult access to large and functioning markets otherwise.
However, this service is less valuable when a trip to a substantial market can be made daily by
donkey cart. Thus, the correlation between villages with better market access and smaller shares
of cotton is logical, yet revealing for those working at CMDT.
Furthermore, households in the villages of Kaniko, Try I, and Signe have generally
higher shares of millet and secondary crops. For millet, this may be because a strong export
demand from Cote d’Ivoire makes it a viable and marketable cash crop, especially for those near
the markets of Koutiala, a regional trade hub. But unlike millet, sorghum does not currently have
a similar export demand. Additionally, if households are already selling their millet in Koutiala,
they face fewer transaction costs to buying other cereals to diversify their diet, providing another
incentive for them to specialize in millet production. Better access to Koutiala’s market also
makes production of secondary crops more promising since fruits and vegetables are often in the
highest demand in the city and, if not, its markets are full of traders willing to take products to
sell in remote markets. For example, farmers who produce watermelons would likely sell them

76

and living close to the city would reduce the high transportation and transaction costs involved in
moving the large and perishable fruit, making its production more profitable.
Altogether, the magnitudes and significance of the coefficients for the village dummy
variables suggest that village location seems to explain much more about the effect of transaction
costs on crop choice and land allocation than ownership of motorcycles, bicycles, and carts.
However, better market access had a relatively inconclusive effect on maize. This may be
because of the overlap between the cotton and maize value chains; many farmers use credit
through CMDT to purchase fertilizer for maize production. So perhaps while better market
access may make maize more appealing as a cash crop in its own right, the negative effect of
market access on cotton production, and by extension input procurement for maize production,
balances out its effect on maize shares. However, clearly maize is grown more in some villages,
such as Tonon and Kaniko, perhaps due to favorable geographic or climatic conditions.
Finally, results under the location variable Village_Simona represent a seventh village in
the Yorosso cercle that I surveyed in the following 2010/11 growing season. So, compared to the
base data, it is both a different time and location, and one that is more economically isolated.
Relative to Nampala II in 2008/09, the results suggest that Simona households have increased
shares of cotton and maize by 8.9% and 4.5%, respectively, which is offset by decreases in
shares of millet and secondary crops of 7.4% and 7.6%, respectively. Not unlike some of the
villages in the Koutiala cercle described above, the average household in Simona differs in its
land allocation decisions for about 15% of its land, relative to the average household in Nampala
II, all else equal. Additionally, these results reinforce patterns discussed above given how
Simona’s economic isolation and distance from Koutiala give it incentives to produce cotton
along with maize instead of larger shares of potentially marketable millet and secondary crops.

77

Not surprisingly, only 4% percent of households formally sold millet that year and only 8% of
households formally sold any coarse grain.
5.3

Application of Results to Example Households
To conclude this chapter, I will apply the average marginal effects from Table 6 to a

couple of example scenarios, now that the coefficients have been discussed individually. This
will demonstrate the usefulness of the average marginal effects when trying to predict total crop
shares for specific cases. In particular, I want to examine differences between a wealthier, larger
household and a small, disadvantaged household. For simplicity, both cases will be from the
same year, village, and ethnic group, dismissing the need to consider these coefficients in the
calculation. The predicted crop shares for both households, and the differences between them,
highlight many of the results discussed above.
The first household has six adult males, eight adult females, eight young boys and girls
over the age five, and four children under five. The household farms on 16 hectares and owns
two sprayers, three plows, six draft animals, two carts, two motorcycles and three bikes.
Furthermore, three adult men are literate along with two of the adult women, and only the elderly
grandmother is considered inactive. These figures are slightly better than the average for a
household of this size. The second household falls below the average 2008/09 Minianka
household in the dataset. It is comprised of two adult males, two adult females, three boys and
girls over five, and three children under five. The household farms on six hectares with one
plow, one ox, one bicycle, but does not own a sprayer, motorcycle, or a cart. One adult male is
literate and none are inactive. Again, these numbers are slightly lower or more disadvantaged
than a typical household of this size.

78

Assuming that both households belong to the same year, village, and ethnic group, the
larger first household is estimated to allocate 13% more land to cotton and 4% more land to
maize, in exchange for 15% less share of sorghum, relative to the smaller second household. The
differences between shares of millet and secondary crops are relatively small at 1% less for the
first household, evening out all crop share differences to sum at zero. These differences between
the aggregate average marginal effects for two households emphasize the role of cotton and, to
some extent, maize as a cash crop for households with the proper farming equipment and labor
supply.
For the sake of another example, consider if the smaller and disadvantaged household
now has one inactive male—an unfortunate but possible scenario. This one change greatly
affects the differences in crop shares between the two households. Now, the larger household is
estimated to allocate 17% more to cotton share and 22% less to sorghum share, relative to the
smaller second household, with small differences in maize, millet, and secondary crop shares that
even out all differences to sum at zero. The strengthening of the divide between cotton and
sorghum shares highlights sorghum’s role as a crop for vulnerable households with many mouths
to feed relative to the land, labor, and other inputs available.
Finally, while there are some very significant average marginal effects on millet and
secondary crop shares, they did not emerge in this application. This may be because the values
chosen here for specific variables canceled out any variable-specific positive or negative
impacts. For example, coefficients for the effect of the number of women and youth per hectare
on millet share are both statistically significant and impressive in magnitude, but in opposite
directions; thus, these effects mostly balance out here since the households in each scenario had
similar headcounts of women and youth. Furthermore, determinants of millet and secondary crop

79

shares—along with maize shares—may have a lot to do with village or ethnic group. For
example, as a marketable crop, millet sees much higher shares in villages closer to Koutiala. Yet
because both scenarios were equal by year and location, these differences in millet share did not
emerge.
The primary purpose of this exercise was to demonstrate that the average marginal effects
on crop shares add up, though some clearly have more effect than others. It presented two
scenarios of realistic, yet different, households in the Koutiala production zone and predicted
how their crop shares may differ relative to each other. In the second example, the model was
able to predict land allocation differences for almost a quarter of the household’s total cultivated
land, even though these families could have been neighbors. Using the coefficients, it is also
possible to predict the expected crop shares for different years and ethnic groups represented in
the data.

80

CHAPTER 6: CONCLUSIONS
To conclude this study, I will present a summary of its methodology and then make
policy recommendations based on relevant findings. Also, limitations will be discussed,
including aspects from survey design to data manipulation. Lastly, I will suggest opportunities
for future research that build on this thesis and previous works.
6.1

Summary of Methodology
This thesis began with a research question concerning land allocation of field crops in

Mali, the understanding of which is important for improving household food security and
preparing for challenges facing Mali’s cotton industry. Specifically, policymakers need to know
how certain market and household characteristics affect planting decisions of cotton, coarse
grains, and secondary crops. Whereas many studies examine this issue one crop at a time—
perhaps building on a supply response model or estimating the probability of crop adoption with
a logit or probit model—I undertook a different approach.
First, I applied the agricultural household model to my understanding of Mali’s Koutiala
production zone, which established a framework that assumed that household farms face a utility
maximization problem subject to a budget, production, and time constraint. Thus, their land
allocation function is determined by expected prices, family labor, inputs, land area, access to
capital, risk, transaction costs, household characteristics, and alternative sources of income.
Variables were selected to represent these factors, employing data from two survey rounds of
150 households in Mali’s Koutiala cercle and an additional round of 50 households from the
neighboring Yorosso cercle a year later. In all, 350 observations represented seven villages.

81

Variables consisted of household members per hectare, inactivity of adults, available farming
equipment, literacy, ethnic identity, village location, and available modes of transport.
To estimate the relationship between these factors and land allocation decisions in the
data, Chapter 5 employed a fractional multinomial logit. This technique builds on a standard
logit by allowing for categorical, non-binary, dependent variables whose values are fractions
which sum to one. For my purposes, the share of land allocated to cotton, maize, sorghum, millet
and other “secondary” crops served as the five dependent variables, all of which, when
combined, equaled the total number of hectares cultivated by a household. The fractional
multinomial logit results were estimated through quasi-maximum likelihood using the fractional
multinomial logit package in STATA. The resulting average marginal effects of the explanatory
variables on each crop share category were presented and discussed.
6.2

Summary of Findings
This thesis’s research question asked which determinants were responsible for land

allocation of cotton, maize, sorghum, millet, and secondary field crops. Overall, the most
influential sets of variables were those representing ethnic identity and village location, the latter
of which may be due to proximity to markets in Koutiala. Villages closer to Koutiala were
closely associated with much higher shares of millet and lower shares of maize, sorghum, and
especially cotton. Finally, the effects of ethnic identity varied by group, but compared to the
Minianka majority, most grew a higher share of millet and smaller shares of maize and, to a
lesser degree, cotton and secondary crops. Meanwhile, variables sets representing family and
farm size, literacy, farming equipment, and modes of transportation had significant results for
some variables on particular outcomes, but were not as revealing overall.

82

Specifically, the model estimated that the key determinants of increased land allocation to
cotton were the number of infants per hectare and the number of pesticide sprayers and carts
owned by the household. It was discussed that while infants cannot assist with cotton production
nor consume it directly, this link may be because families with young children expect many outof-pocket expenses—from medicine to future school fees—that require cash, which can be
earned through cotton production. Also, since cotton growing has high start-up costs, the
ownership of capital needed for successful cotton production, such as carts and sprayers, helps
farmers decide to grow more of it. Village location was also influential, suggesting that those
households living in a village closer to Koutiala planted smaller shares of cotton, perhaps
because better access to Koutiala’s markets makes it easier to trade and earn income with other
goods. Although CMDT transports all inputs and output to villages where cotton is grown,
proximity to Koutiala also makes these efforts to reduce farmer’s transaction costs less of an
incentive.
Above all else, the key determinants of land allocation to maize were ethnic identity and
the number of adult males and youth, between the ages of 6 and 15, per hectare. Households
belonging to the Bambara, Peulh, or another “other” ethnic group devoted less land to maize
relative to Minianka and Senoufo households, perhaps reflecting a cultural or historical
preference for maize. Additionally, that more household members per hectare was correlated
with higher maize shares is expected as its high-yielding potential is good for families with little
land and many mouths to feed, as long as they possess the labor supply to support its production.
Finally, owning capital (e.g., plows or oxen) also helped to increase maize shares—as with
cotton—as farming equipment significantly increases labor productivity for tasks that are
necessary for successful maize production.

83

Earlier discussions in this thesis of sorghum and millet reviewed how these coarse grains
have very similar subsectors, but the results imply that the determinants for how much land is
devoted to each differ greatly. Increased sorghum share is associated with more infants per
hectare, higher percentages of inactive adults in the household, and it often serves as the tradeoff for increased shares of cotton and maize for additional capital. Further, its share greatly
decreases as the number of adult men per hectare goes up. It seems to be a crop for vulnerable
households with many mouths to feed, but not enough capital or labor to consider planting more
maize; certainly, this is the picture that emerged from the example scenarios in Section 5.3.
While I would have expected millet to fulfill a similar role, the results suggest that it is not so
much a safety crop as a marketable one. In fact, millet shares are inversely related to the
percentage of household adults who are inactive, and increased shares of millet are highly
correlated with villages in close proximity to Koutiala, where well-functioning coarse grain
markets exist to export millet to Cote d’Ivoire or elsewhere. So while their subsectors may be
similar, the determinants of land allocation to sorghum and millet have distinct differences.
Finally, key determinants of land allocation to secondary crops vary a lot by location as
well. This may be due to increased access to thriving markets—all three villages near Koutiala
had higher shares of secondary crops—or due to geographic differences; the village of Gantiesso,
which also had high shares of secondary crops, may have been located near a river that made rice
production possible. On the other hand, smaller shares of secondary crops were linked to
increased number of women and infants per hectare and some ethnic groups, especially the Peulh
people. Since it was thought that women are mostly responsible for secondary crop production,
the negative effect of adult women per hectare on secondary crop shares was unexpected, though
this same negative effect attached to infants per hectare is logical after considering that infants

84

require time and energy from women that then could not be used in crop production. Perhaps
households with fewer women, and thus fewer infants, are more likely to devote land and other
resources to secondary crops, which have higher investment start-up costs.
6.3

Policy Recommendations
These results, along with some personal observation and evidence from other sources,

can help inform recommendations for Malian agricultural policymakers. Since determining the
shares of various crops is a zero-sum game—that is, a situation in which all gains are someone
else’s losses—the recommendations for improving cotton and coarse grain production oppose
the other’s. Yet, if households responded to new incentives by clearing new fields and planting
additional hectares of crops or increasing adoption of fertilizer to boost yields, it is possible that
both cotton and coarse grain production can increase overall. Of course, such changes cannot be
analyzed using this thesis’s methodology, as it only examines crops shares, but should be kept in
mind.
For Mali’s cotton company CMDT, the primary recommendation is stressing the
importance of not losing their institutional advantage, which has been helping households attain
farming equipment, offering fertilizer procurement though cheap cotton-backed credit, covering
transport of inputs and output, and paying farmers shortly after harvest. First, CMDT has
historically assisted farmers in attaining cotton-farming equipment, such as sprayers, plows, draft
animals, and donkey carts, all of which help explain higher shares of cotton production
associated with ownership of such equipment. However, as older tools deteriorate and new
households are established in the Koutiala production zone, there must be ways to help farmers
overcome the steep start-up costs attached to cotton production. Offering fertilizer procurement
and covering transport also help reduce transaction costs for farmers, a relevant factor for
85

isolated villages without easy access to Koutiala’s markets. Finally, as stressed by Theriault
(2010), the cotton industry must make predictable and timely payments to its suppliers who rely
on this income soon after the harvest. Failure to do so is what led to a cotton strike in 2007/08
and likely explains why, in the results, cotton shares increased in 2009/10 relative to 2008/09
despite the fact that the producer price of cotton dropped by 30 FCFA between those years (Mali
Assessment 2011). These perks of participating in the cotton industry are considered part of its
expected incentives package and thus need to be met or expanded in order to maintain or expand
cotton production.
For those looking to reform the cotton industry, a pertinent question is whether or not this
expected incentives package is sustainable. If so, what would be the result of cutting certain
services; which has the highest cost-to-benefit ratio? Unfortunately, answering these tough
questions is beyond the scope of this thesis. However, it is important to reiterate that while some
villages in the Koutiala production zone seem to be parting ways with CMDT, other villages are
still dependent on it. I assert this from experience, having lived in Simona for almost two years,
and this claim is support by the significantly positive correlation between cotton share and the
variable Village_Simona. Additionally, interviews conducted for this thesis affirm that many
villages in the Yorosso cercle have a similar dependency on the cotton industry. Moreover, many
farmers there either fail to understand or underestimate CMDT’s financial woes and their
potential implications. For them, talk of cotton reform is frightening, as it threatens the primary
livelihood they have been raised to understand. Any transition or change must not only consider
sustainability and profitability, but perhaps more importantly, dependent farmers’ well-being and
ability to adapt. Transition can also be helped through communication (by radio, extension agent,

86

or otherwise) and clear and predictable timetables that give farmers the ability to revisit their
household’s utility maximization problem.
For agricultural policymakers hoping to boost coarse grain production and sales, the
results have some usefulness as well. With regards to maize, it remains cultivated on the least
amount of hectares, especially by ethnic groups not native to the area—despite tremendous
growth overall throughout Mali in the last decade. This may because of varying taste and
cooking preferences between ethnic groups or because CMDT extension, which introduced and
promoted maize in this area years ago, has dwindled and may not continue to encourage maize
production as much. Either way, if Mali hopes for Koutiala to help in increasing the country’s
share of maize within total cereal production, it may help to increase promotion of maize as a
delicious and marketable crop, along with providing resources for farmers to learn more about its
cultivation. From experience, I believe that radio programs may be a successful starting point as
they are non-confrontational, inexpensive to broadcast, and capable of reaching a wide audience.
Moreover, the results suggest that while sorghum and millet exhibit similar value chains,
each has a different role in the Koutiala production zone. Sorghum is the “safe” crop— grown
more by vulnerable households—and millet is the marketable coarse grain. This result is telling
given that much of the policy discussion now is about the growing demand and marketability of
maize as opposed to millet. These differences in the role that sorghum and millet play should be
studied more, either through qualitative or quantitative analysis, to determine if these differences
exist only in the dataset or truly represent trends across the production zone, or even the entire
county. Additionally, policymakers should continue work that reduces transaction costs for
coarse grain markets as the results suggest that better accessibility to markets enable households
to cultivate greater shares of coarse grains (particularly millet) relative to cotton. For sorghum,

87

reduced transaction costs would also make it easier for vulnerable households to purchase this
“safe” crop, which in turn may encourage greater market participation. Overall, any action taken
now to strengthen the sustainability of these coarse grains markets helps to reduce household
vulnerability to potential changes resulting from the restructuring of the cotton industry.
One question that came to mind frequently while serving as a Peace Corps Volunteer in
this area was: if coarse grains were put on the same playing field as cotton—that is, if they all
had the same incentives package—what would crop shares and production look like? To do this,
one approach is to reform the cotton industry, as is being discussed anyway, so that it either
offers fewer services or that those services can apply to more aspects of coarse grain production.
Another is to reform agricultural support markets so that they become viable alternatives to
services offered by CMDT. For example, is it possible to reform, support, promote, or raise
awareness of Mali’s microfinance institutions, so that they offer more farm-friendly financial
products that are competitive with CMDT’s loans for fertilizer? This could be done if payments
on loans could be delayed until cereal prices were higher (around April or May), giving farmers
an opportunity to maximize their profit from coarse grain production before paying off their
debt. Moreover, many people still do not understand the financial mechanisms offered by
microfinance institutions, and actions aimed at educating farmers about their benefits and
appropriate use of a loan or bank account could make these options more accessible to rural
households.
Furthermore, the results identify some helpful patterns for secondary crops that may be of
use to policymakers. Again, secondary crops can be beneficial in the farming system because,
even though they likely limit specialization in cotton or a coarse grain, their successful
production can diversify a household’s income portfolio or diet, which can help to reduce risk

88

and can improve household nutritional intake. Yet, the results indicate certain types of
households that are associated with lower shares of secondary crops: those from economically
isolated areas like Simona, minority ethnic groups, and those with more women per hectare and
children under the age of 6 per hectare. Results to variables representing village location
indicated that secondary crops shares seem to vary much by geographic location, regardless of
market access, and are primarily grown when there is extra family labor to dedicate to it or
suitable environmental conditions (e.g., a river to grow rice and other water-intensive crops).
Continued education to rural audiences on the nutritional importance, potential use,
marketability, and different production techniques of certain secondary crops may boost their
production over time while at least exposing households to the idea that these are essential for a
well-balanced diet, especially for their children.
Lastly, the focus on the village of Simona in this study also highlights a need to conduct
more research in the Yorosso cercle, which has consistently higher cotton adoption rates among
households than those in the Koutiala cercle. Also, since the relationship between the cotton and
coarse grain subsectors is stronger in this area, its population will be more vulnerable to sudden
changes as a result of the CMDT restructuring. Therefore, special consideration must be given to
the coarse grain subsector, and particularly maize, in the Yorosso cercle throughout cotton
industry reform as the food security of Simona, and its neighboring villages, may depend on it.
6.4

Limitations
Since every economic model seeks to simplify a complex reality while maintaining

explanatory power, it is bound to have limitations from the data it employs or the model itself.
This thesis is no exception, and in fact, most of the fractional multinomial logit’s limitations
have already been discussed. Theoretically, its measurement of crop shares disregards the effect
89

of fertilizer and careful maintenance on production, giving the impression that hectares planted
to specific crops is always a zero-sum game. Secondly, it tells us which variables higher or lower
shares of particular crops but not necessarily why this is the case, whereas other models may
facilitate the use of crop-specific variables to help explain these correlations. However, these
limitations were discussed and kept in mind throughout the interpretation.
Secondly, other possible limitations of this study result from human errors made during
the collection of the data. This could be the fault of the respondent, who may have provided an
estimate without knowing for sure. Another possibility is that the respondent lied in an attempt to
impress or evoke pity from the enumerator, who could have been perceived as a government
official or non-profit worker. Sometimes enumerators make errors whether by mistake or
intention. Especially if a question is embarrassing or one outcome is typically given for over 90
percent of respondents, enumerators may just fill it in without asking. These were some of the
issues that emerged for the 50 surveys I conducted in Simona, where I trained local enumerators
who spoke the native Minianka, although as discussed in Chapter 4, steps were taken to
minimize collection errors. For example, I encouraged enumerators to meet with a group of the
household’s men over a round of tea to ensure that their farmers were giving input, since some
household heads are no longer active in agricultural production. Similar strategies were adapted
for the Koutiala survey rounds too. Still, the possibility of some error exists.
Along the same lines, in hindsight it would have been beneficial to include or revise
additional variables to use in the present analysis, but because they were not collected, they
could not be included here. This includes a host of village-level variables such as distance to
paved road, nearest weekly market, or access to a microfinance institution, all of which would
have enriched the analysis of the results for village location dummy variables. Also, it would

90

have been valuable to know how much households earned from alternative sources of income
and whether this was earned outside the village or not. Another excluded variable was some
representation of last year’s food stock at the time of planting, which may affect crop shares as
households aim to offset expected food shortages. Lastly, the survey asked whether each family
member put in a morning, afternoon, or full-day of farm labor to draw up its labor supply
estimates, but it did not disaggregate this request for different times of the year (e.g., planting,
weeding, or harvest). In the future, more accurate estimates of family labor, perhaps lagged one
year, may have a place in this model.
Finally, as in many studies of agriculture in the developing world, these results would be
strengthened with additional data. Fifty observations per village is close to, but not quite enough,
for us to be truly confident that it fully represents the village, and certainly seven villages out of
45 communes leaves room for misrepresentation. The timeline in which these surveys were
conducted also creates room for additional error, though these were controlled for as best as
possible in the model.
6.5

Future Research
This thesis began with two primary motives. One was to provide insights to Malian

policymakers as they attempt to address food security and cotton industry reform in the Koutiala
production zone. In this regard, there is certainly more research to do. While examining crop
shares and land allocation has been a useful exercise, there are various other approaches that
have been and should be done in order to understand changing trends of crop preference, input
adoption, market failures, and child nutrition.
Two emergent transitions in Mali make future research there more relevant than ever.
The first is the much-discussed cotton industry reform, the future of which is uncertain, but the
91

impact of which could be significant as the cotton industry has been a monopsony there since
Mali’s independence. Thus, continued research on Mali’s cotton subsector, along with its social
and economic impacts, may be useful to policymakers and should address at least three topics.
First, understanding which of CMDT’s services are needed and valued the most by rural farmers
could assist with reform efforts, especially if the industry becomes privatized. Secondly, it is
important to understand which existing markets have developed which compete with CMDT, or
its services, especially those that may have enabled some households to not rely so heavily on
the cotton industry in recent years (e.g., fertilizer merchants, alternative cash crops). In other
words, what alternatives have helped bring about, or emerged in response to, a decline in
participation in the cotton industry, and how might these alternatives be encouraged or improved
within the broader reform effort? Finally, such research should maintain a special focus on
cotton-dependent households, like many of those in the central part of the Yorosso cercle, which
can inform strategies to assist communities in adapting to potential change in the cotton industry.
Research that exclusively focuses on villages closer to the city of Koutiala for the sake of
convenience are at risk of overlooking vulnerable cotton-dependent communities, which—as this
thesis has argued—face a different set of economic circumstances.
The second and more troubling transition is that of the country’s political instability and
civil unrest after its coup d’état in late March of 2012 that was quickly followed by a rebel
takeover of Mali’s three northern-most regions. This has resulted in uncertainty in the South,
made worse by the threat of economic sanctions, and horrific killings and crimes against
civilians in the North. Thankfully for the villages in the Koutiala production zone, their lives and
land are safe, and the rains this year are good, though some villages have relied on food aid.
However, given that the final outcome of this struggle remains uncertain, there will certainly be a

92

need to reexamine food security to see if progress has regressed or continues as it has over the
last decade.
The other motive of this study was to develop a new method of modeling household land
allocation for various crops in developing countries. To do this, we adapted the agricultural
household model for use in Mali’s Koutiala production zone and applied the relatively new
fractional multinomial logit framework. While neither is without its limitations, some described
just above, the overall model had definite explanatory power and was useful in discussing
determinants of crop choice and planting at the household level. What is needed now is
additional work applying this model to different circumstances. Within Mali, it can be applied to
other regions and their alternative crop portfolios. Another idea is to open up the dependent
variable representing shares of secondary crops to see how the explanatory factors affect peanuts,
sweet potatoes, sesame, and vegetables differently. In fact, use of the model on a generous
dataset in any developing country can help to provide evidence for theoretical discussions of the
agricultural household model, such as the extent of the effect of transaction costs or inter-annual
credit restraints. As it can compare all crops simultaneously, it can serve as an additional tool to
study the farming system on the household farm, which continues to be the most fundamental
economic unit in the majority of the developing world.

93

APPENDICES

94

Appendix 1: The Household in the Koutiala Production Zone
The term “household” is used extensively throughout this paper, though the standard
definition of this term varies greatly across cultures. In Mali, a household is typically a large
patriarchal social structure including an average of three generations in which relationships are
relational or marital. However, depending on family divisions or migrations, household size can
vary greatly. For example, during the 2009/10 growing season, households in Simona had an
average of 18.9 members, though the smallest household was roughly 4 members and the largest
(by far) was 84. Each household normally has one established location close to a village center
and another settlement near its fields. The latter is inhabited during the farming season in order
to increase labor efficiency on the farm.
Authority and privileges within the household are granted with age and are genderspecific. The head of the household, the eldest male, is traditionally granted the final say for all
household decisions. One Malian friend confided that he believed the fiancé of his daughter to be
a bad man, but could not break the engagement or protest because his elder brother, the
household head, had already negotiated the marriage. Thus, many household-level development
studies consider characteristics of the household head as explanatory variables for a given
dependent variable, though many household heads work with the other adult males in the family
on complicated matters. For example, an elderly man will make his decisions along with his
sons, who currently manage the household’s fields, or an elder brother consults with his younger
siblings to maintain family peace. While conferring with a woman is not unheard of, it is by and
large the exception to the rule.
The respected men within a household also make most decisions regarding family
consumption. They are responsible for managing the household’s finances and food stock in

95

general, especially the share for their wives and children. In a traditional village such as Simona,
a woman will not take grain out of storage without receiving her husband’s permission first, and
similarly, she must ask her husband for money before going to market and be ready to explain
her expected purchases. As a result of this weighted decision-making, respected men often
consume better goods than the rest of the family. Since meats and vegetables are still considered
luxury goods in many communities, these are mostly consumed by the men when they eat first,
leaving less for women and children. This is not done out of spite, but out of tradition and
unawareness (the nutritional value of these foods is not widely recognized). When these men,
and particularly the head, make household decisions, every member of the household is
considered, but the hypothetical weights given to each individual is determined by the man’s
moral and social values.
In terms of labor, the role of men in the household is dictated by the seasonal calendar. In
the rainy season, which goes from June to September, rainfall is adequate for the planting and
maintaining of crops. These few months are busy for an entire family, but particularly for men,
who often wake at dawn and work until dusk with only a two-hour break at lunch. For this
reason, these few months are alternatively called the farming season. Moreover, this time leading
up to the harvest is also known as the hungry season, when many households run low on last
year’s supply, and any available excess grains can sell for a high price.
Then, October to January, known as the cold season, becomes increasingly dry and cool,
which are good conditions for crop harvest and gardening. Normally, there is enough food for
everyone during this time, but prices are also lower since grain supply is abundant. So during this
time, men have to take stock of the harvest and carefully strategize how they will manage their
food and finances for the year. Additionally, cold season is an opportunity to make mud bricks,

96

and then construct or repair structures, because the next six months without rain will allow them
to dry and harden.
Finally, from February to May, the hot season is the time when villagers (but mostly
men) take leisure as high temperatures make people lethargic. Many rest and spend their
evenings at weddings until the early rains in May begin to alert people of the start of the next
crop cycle. It is important to note that men rest to various degrees during hot season; some seem
to spend an entire two months in the shade, while others continue to build up their own
compound, volunteer for a community project, or work in village as a service provider (e.g.,
carpenter, blacksmith, or mechanic). Of course, this latter group can work all year, but get most
of their business during the dry seasons from December to May.
The role of women within a southern Malian household is mostly restricted to food
preparation, child-rearing, and some agricultural activity. With permission from their husbands,
women may also participate in associations or income-generating activities, such as gardening or
shea butter production. While this work is seen as inferior or even degrading for a man, it was
accepted by nearly all Peace Corps Volunteers that Malian women work far longer and harder
than men, whose farm labor is seasonal. Development efforts have focused on Malian women
not only to promote gender equality and empowerment, but also because of evidence that income
earned by women often has a greater positive effect on children’s health than men’s income
(Tefft, Kelly, Wise and Staatz 2003). Yet, effective partnering with women requires an
appreciation of their time and decision-making constraints.
Similar to the men, respect among women is a function of their age and marital status, the
latter of which is complicated by practiced polygamy that allows (not requires) men to marry up
to four wives according to Islamic law. Keeping this in mind, it is easy to see why these

97

households can grow to be so large. In Simona, the respected patriarch of an established family
had four wives, over a dozen children, and could not recall the number of his grandchildren.
Such size makes it more difficult for the adult men to agree on household decisions, but also can
provide a safety net, especially for widows, their children, and the elderly.
A support network and safety net is particularly important in remote villages where
residents have limited access to basic goods and services. Theoretically, every village has its own
water pump, maternity, and elementary school, but misunderstandings between villagers, the
government, and non-profits about who is responsible for maintaining these facilities can lead to
situations in which they are not functional. Next, there is a series of other public goods, such as
middle schools, microfinance banks, and water towers that are given to villages able to convince
a benefactor of their need. However, larger institutions such as hospitals, high schools, and radio
towers are only found in important cities, such as Yorosso or Koutiala.
Weekly markets are not held in every village, and although Simona has one, it offers
significantly less than some larger villages in the Yorosso cercle. For most households, getting to
a larger market requires some form of transportation, such as a donkey cart (if buying or selling a
lot), bicycle, or motorcycle. Finally, public transportation to Koutiala and beyond is only
accessible on paved roads. This means if the household’s village is located off the roads,
privately owned transport is required to access it as well. These obstacles, which vary by village,
are some of the challenges that increase risk and transportation costs for rural households.

98

Appendix 2: Fertilizer and Credit for Staple Crop Production
In CMDT’s Koutiala production zone, there are three types of fertilizer most commonly
used for staple crop production. Complexe Coton is intended for use on cotton, Complexe
Céréale is intended for use on coarse grains, and urea can be applied on both. But while all of
these crops can benefit from the appropriate fertilizer, cotton and maize are especially dependent
on fertilizers in order to be profitable. However, even after a large fertilizer subsidy was
instituted in 2008, these fertilizers can still be expensive for a Malian household; a bag of
subsidized fertilizer costs a minimum of 12,500 FCFA (roughly 25 US dollars), although before
the subsidy it cost anywhere between 18,000 and 20,000 FCFA (roughly 36 and 40 US dollars,
respectively). Part of the problem is that nearly every household faces a cash constraint around
July or August (during hungry season) when fertilizer is needed. In order to purchase fertilizer,
households must choose between two suppliers and two methods of payment. The two suppliers
are CMDT and market vendors and payment can be made in cash or on credit. Since the CMDT
mechanism is complex, we will begin by discussing procurement through a market vendor.
If a household has cash and is willing to spend it, it may purchase the fertilizer from a
market vendor directly. However, since these fertilizer vendors tend to be closer to paved roads
and growing towns, the remote household may incur high transport costs. Even in Simona’s rural
weekly market, established shops (boutiques) and vendors sold various input products including
weed-killer, fungicide, and insecticides. However, these vendors did not sell Complexe Coton,
Complexe Céréale, or urea fertilizers. Still, they are noteworthy because they represent a
developing infrastructure that may be able to handle all input distribution in the future, even in
remote rural areas. Purchasing on cash through a market vendor was an attractive option for only
one farmer in Simona, where historically CMDT was the sole supplier of fertilizer and other

99

inputs. When asked why they refrained from selling fertilizers, these market vendors responded
that no one was willing to buy from them. The price was too high relative to household
disposable incomes in the period before the harvest when cash was low and needed to purchase
food and maintain farming equipment.
Thus, an alternative is that a household purchases fertilizer from a vendor after applying
for a loan through a microfinance institution, but this is rarely done. First, at an interest rate
around 20 percent, this is a far more expensive line of credit than going through CMDT. Further,
the perceived punishment for failing to follow through on payments is as severe as
imprisonment. Because of this risk, use of a microfinance loan for fertilizer procurement may not
be the most attractive option for households growing cotton, though it may be reasonable if
cotton production is deemed unprofitable and cash is unavailable.
On the other hand, the CMDT mechanism allows farmers to purchase their inputs on
credit. Sacks of fertilizer have a premium rate that is just over 7 percent, raising the total cost of
a fertilizer sack to 13,415 FCFA (roughly 27 US dollars). Essentially the mechanism stipulates
that each year, farmers can receive credit for up to 74,125 FCFA per hectare of cotton (as of
2010) in order to procure coarse grain fertilizer and cotton inputs, which is then deducted from
their cotton payment after the harvest for a small premium. Otherwise, households can still
purchase at the base rate if they are willing to pay in cash. Either way, CMDT covers the input’s
transport cost to the village. Additionally, current policy allows households to purchase inputs
for coarse grains through the same system. Particularly for households interested in expanding
maize production, this mechanism simplifies fertilizer procurement by reducing transaction costs
and the need for cash or microfinance credit in June or July, as the hungry season approaches.
Furthermore, due to CMDT structural reforms made in 1974 (Theriault and Sterns 2012),

100

all financial exchanges between the producer and CMDT are made through a producers’ village
association, including debts. Therefore, if an elephant crushes your cotton crop and you do not
produce enough to repay, this debt will not put you in prison, but makes you responsible to your
association. Then according to local custom on debt repayment, you will have repay as the
association sees fit. Thus, the association serves as a producer network and safety net, which
does not exist for most other crops.
For Simona farmers, evidence suggests that purchase of fertilizer through CMDT was a
very attractive option. Households that cultivated cotton in 2010/11 acquired all three types of
fertilizer exclusively through CMDT. Only one farmer, who was unable to rent the livestock,
equipment, and labor required to plant cotton that year, purchased fertilizer for maize at the
weekly market after making a special order in advance. It is also important to highlight that 97.7
percent of all fertilizer purchased for coarse grain production in Simona was placed on maize
fields during the 2010/11 farming season.
One option that was avoided by Simona farmers was relying on credit from Kafo Jiginiw,
a microfinance institution with a branch in Simona that was open during weekly market. The low
demand for its lending and bank services was a frustration that Kafo Jiginiw officials expressed
at a public meeting (Kafo Jiginiw 2011). It was pointed out that of the eight villages in the
Yorosso cercle that have a bank branch, Simona’s had the lowest amount of money in active
accounts as well as the lowest number and value of loans taken out. Indeed, over half of the
accounts had less than 2000 FCFA in them (roughly 4 US dollars). According to villagers, credit
from Kafo Jiginiw was not desirable because of the need to make regular early payments rarely
made financial sense as good prices for coarse grains—another possible source of income—were
not available until April. As discussed, such loans also had a high interest rate and stiff

101

repercussions for defaulting, though over 98 percent of Kafo Jiginw loans taken out in Simona
were paid back on time. Given that all Simona farmers who grew cotton in 2010/11 used their
CMDT credit to purchase fertilizer for both their cotton and coarse grains, it is clear that the
Kafo Jiginiw line of credit was not preferred.
While only one person in Simona purchased fertilizer from a market vendor using cash,
the figure was slightly higher in the six villages surveyed in the Koutiala cercle. Of the 849
market transactions made by households to procure fertilizer in 2009/10 and 2010/11, 10.8
percent were made in cash and to a market vendor. Still, purchasing fertilizer on credit through
CMDT was the most common method, making up 84.5 percent of transactions. The fact that this
rate is higher than the average adoption rate of cotton suggests that those who grew cotton were
more likely to purchase fertilizer and/or purchase it for more than just cotton. Most of the
remaining transactions were made with “other” fertilizer suppliers, and those made with market
vendors using credit were fewer than one percent. These statistics provide some evidence on the
PROMISAM’s hypothesis that the cotton industry may positively affect coarse grain
productivity by selling coarse grain fertilizer on cheap, cotton-based credit.

102

Appendix 3: Development of Agricultural Household Model
This appendix provides a summary of some of the background research that brought
about the development of the agricultural household model adopted by this thesis. The first
section presents the basic model, as popularized by Singh, Squire, and Straus (1986), which is
void of considerations for risk behavior, market failures, and transactions costs. The second
section examines more recent literature that attempts to incorporate these concepts, namely
transaction costs, relying heavily on Omamo (1998) and de Janvry, Fafchamps, and Sadoulet
(1991).
A.3.1 Generalized Agricultural Household Model
In designing a model specific to the Malian context, it is necessary to discuss the
literature and determine which resource constraints and market failures are relevant factors for
crop choice and land allocation for Malian households. The following section presents the
foundational model and then discusses market failures. Since almost every new author in the
literature has unique labels for their variables, I will be consistent in my representation while still
paraphrasing the contribution of their work.
Building on the early work of Barnum and Squire (1979) and later Singh, Squire, and
Strauss (1986), I assume a static agricultural household model in which the primary objective of
the household, as discussed, is to maximize expected utility during a given production cycle as a
function of consumption of farm-grown consumption products X (e.g., millet or milk), a vector
of other consumption goods Y (e.g., clothes or radios), and leisure H:
Max U = f (X, Y, H)

utility maximization function

(A.1)

However, households maximize expected utility subject to constraints on income, time,
and production technology. First, the cash income constraint dictates that the value of goods X
103

and Y consumed by the household cannot exceed household cash income. Simply put, assuming
the absence of credit or savings, the household cannot spend more than it earns in a given
production cycle. Income from agricultural production is the value of agricultural output P xQ
minus the cost of inputs PvV and total labor inputs net of labor contributed by the family L – F.
Formally,
PxX + PyY ≤ PxQ – PL(L – F) – PvV + πy

cash income constraint

(A.2)

where Px and Py are a vector of prices for farm-grown goods X and other goods Y, respectively;
Q is a vector of the household’s agricultural output; PL is a market wage; L is total labor input
and F is family labor input (if positive, L – F indicates hired labor); Pv is a vector of prices
associated with a vector of non-labor inputs V; and πy represents profits from other activities.
Since agricultural production is the principal income-generating activity for most Malian
households, this equation is key for describing the relationship between household consumption
(on its left) and production (on its right).
Equation (A.2) also predicts how trade affects a household’s budget constraint in the
Singh, Squire, and Strauss model. Applying basic algebra to equation (A.2) produces the
following:
PyY ≤ Px(Q – X) – PL(L – F) – PvV + πy

alternative cash income constraint

(A.3)

In equation (A.3), the difference between Q and X, that is household production and
consumption of a particular farm-grown crop, is its market surplus. When positive, it indicates
that a household’s production exceeded its own demand, allowing additional income to be earned
through its sale. When negative, a household’s production fell short of its own demand, leading
to purchases from the market. In the case of cotton, households consume very little if any of their
crop, so their market surplus is positive and revenue is generated. However, if a household

104

consumes maize but does not grow any, then its market surplus is negative, though there is also
no labor or input costs. In the case that a particular crop is both sold and purchased at market—a
common outcome resulting from cash flow constraints or financial mismanagement—the market
surplus indicates whether the household earned a net gain or loss from market transactions.
Overall, this equation suggests that households should focus on their comparative advantage in
order to make profit, and if their comparative advantage lies in agricultural production of a
particular crop, then they should produce and sell this crop in order to avoid being a net buyer of
agricultural products overall.
Next, the Singh, Squire, and Strauss model hypothesizes that each household faces a
time constraint, which is straight-forward. Very simply, households must divide their total time T
between leisure H and family labor F, which includes on-farm and non-farm labor that may
generate profit πy from other activities. However, later authors (Benjamin 1992) define family
labor F in two separate variables to distinguish between labor that goes toward agricultural
production X as well as off-farm labor Y. Though it is not in the original model, the same is done
to equation (A.4):
H+F+O=T

time constraint

(A.4)

where O represents off-farm household labor.
Not only does this better represent the necessary trade-off between work and leisure,
which in the case of tea drinking may also increase one’s social capital in the community, it also
illustrates the challenge of labor allocation between various activities. For example, if one
household member can find profitable employment in a city, this can significantly reduce a
household’s production potential and alter their crop choice and land allocation decisions. Thus,

105

a decision has to be made by the household head as to whether it is best for the individual to take
the city job, though this may affect crop production.
By aggregating all labor through an entire production cycle, what the constraint is still
unable to capture fully is the intra-seasonal labor requirements. For instance, more family labor
is required for the task of harvesting cotton relative to that needed for applying pesticides to
cotton fields. However, most research prior to Schultz (1964) assumed there was always excess
labor on the farm since they often observed households when labor was not in the highest
demand. Because of intra-seasonal labor demands, household develop strategies to increase this
labor supply around this time, including calling back a son who works off-farm or pulling
children out of school. This can make aggregate family labor input, on or off-farm, difficult to
estimate.
Finally, in addition to cash income and time constraints, each household is limited by
agricultural output Q as a function of various inputs that make up the production technology:
Q = f (L, V, A, K)

production constraint

(A.5)

where the new variables A and K represent a household’s total land and available capital (e.g.,
farming equipment), respectively.
Note that all of the variables included on the right-side of equation (A.5) are capable of
being changed by the household, but over different time frames. Even within a given production
cycle, a household is able to adjust crop inputs, such as fertilizer, and perhaps even labor supply
during peak times. On the other hand, total land and available capital are more likely changed in
the long term. A household is typically only able to clear a few new hectares of land each year,
assuming that it has already taken steps to acquire the land through the village chief, and new

106

equipment or repairs are considered larger expenditures. Still, all of these factors are likely to
affect crop choice and land allocation.
If the time and production constraints are substituted into the cash income constraint,
utility can be maximized with only a single constraint. This process yields a set of first-order
conditions that equates the marginal revenue products for labor and fertilizer to their respective
market prices. It also predicts a “profit effect” that occurs when an increase in food prices also
increases household income. This effect is normally ignored in traditional demand models, yet it
certainly can reduce or even outweigh the negative income and substitution effects that result
from a price increase.
Thus, whether or not a price change helps a household depends on whether it is a net
buyer or seller of food products; net sellers can attain additional income through higher prices by
the “profit effect,” whereas net buyers may suffer through overall reduction in income. Since
more often than not, the poorest of households fall into the latter category, many governments
attempt to maintain low food prices through policy interventions. Mali is no exception, but since
cotton is not consumed by its farmers, it is often viewed as the income-generating commodity.
Either way, this discussion, which stems from the Singh, Squire, and Strauss model, highlights
its potential usefulness in discussing market accessibility and participation. It is no surprise then
that the extensions of the agricultural household model are mostly related to the issue of missing
markets and other market failures.
A.3.2 Market Failures and the Agricultural Household Model
The most severe flaw with the generalized model presented above is that it assumes
access to perfect markets (i.e., those free of transaction costs and fully capable of hedging risk)—
an assumption that does not even truly hold in developed countries, and certainly not in Mali.
107

Even for farmers participating in the cotton industry, which simplifies market transactions for
inputs, outputs, and credit through CMDT, their production incentives for cotton can still be
affected by failures in other related markets, such as those for coarse grains. Market failures also
provide a rational explanation for non-separability, as high transaction costs create a gap between
prices observed in markets and the prices at which trade is profitable for particular households.
These reasons call for a reassessment of the Singh, Squire, and Strauss model and its predictions
on marketed goods.
The most extreme form of a market failure is when a market does not exist. In most of the
Koutiala production zone, this is essentially the case of the labor market for agricultural
production. In Simona, the village in which I served as a Peace Corps Volunteer, I recall only
one household that had hired one man to maintain a large garden year-round—and this
household was headed by one of the most respected men in the region. Most households draw
mainly from their own members’ labor supply during times of peak labor demand, as discussed
above. Less often, friends and neighbors may work together on each other’s fields—exchanging
a nearly equivalent amount of labor for both parties—or may assist in special circumstances,
such as when church-goers help the pastor plant his crop. However, given the absence of any real
formal or informal labor market, this study will assume that all labor input needs are met by the
household; that is, in equation (A.2), L – F always equals 0. This assumption does not eliminate
the need for the time constraint (A.4), as it continues to define L in the production constraint
(A.5) and present possible off-farm labor needs O that bring in household income through other
activities. However, in upcoming discussions concerning how missing markets affect the Singh,
Squire, and Strauss model, labor will be removed as a cost of production in variations of the cash
income constraint, though household farm labor F will be included as a factor of production.

108

Even where markets exist, high transaction costs and other barriers may prevent
profitable trades that would otherwise occur from happening. For example, consider a Peace
Corps Volunteer who hears of one market vendor selling watermelon for a reasonable price, but
at a stall across town—a 45 minute walk. That is a 1.5 hour commitment, half of which is spent
carrying a watermelon down a hot and crowded street, and the other option, a taxi, will cost five
times more than the watermelon itself. Similarly, for a rural household, the ease of transaction
often has to do with the distance between the farm and market as transport in much of the
developing world is expensive and unreliable. From Simona, the nearest paved road was seven
kilometers and the nearest large market town was twenty kilometers, only open on Mondays. In
Simona, transaction costs were a key factor in household decision-making.
Recognizing the importance of such costs in understanding the rationale of the rural poor,
Omamo (1998) integrated transactions costs into the Singh, Squire, and Strauss model, citing
studies which show that they discourage trade and can influence household-level marketing
incentives and decisions (de Janvry, Fafchamps, and Sadoulet 1991). His change is simple: his
budget constraint states that the cost of all products consumed plus the transaction costs of
buying them must be less than or equal to the value of all products produced minus the
transaction costs of selling them; for either a purchase or a sale, the transaction costs decreases
the household income level. Hence, modifying equation (A.2) and retaining the cost of inputs
(not accounted for in Omamo’s model):
PxX + TCbx + PyY + TCby ≤ PxQ - TCsx – PvV – TCbv + πy transaction costs

(A.6)

where TC represents a transaction cost for purchases (subscript b) and sales (subscript s). I also
included a variable for any potential transaction cost associated with purchasing inputs.
Admittedly, the variable TC covers a broad spectrum. In cases for which the price of goods X

109

and Y are estimated from anywhere other than the household’s market, TC may include price
changes incurred by the market’s middlemen. For example, a merchant purchasing coarse grain
will reduce his farm-gate price beyond his profit margin to account for the cost of transporting
the grain to its place of sale. Further, de Janvry and Sadoulet (2006) improve this by modeling
both fixed and variable transaction costs, the latter being added to a good’s market price, but
Omamo’s simpler model is adequate for my purposes.
A primary contribution of equation (A.6), according to its author, is that it summarizes
the trade-off between the gains from specialization and increased transaction costs that affect a
rural household’s decision to diversify its crops—a tension first identified by North (1981).
Rearranging equation (A.6) reveals that:
TCbx + TCby + TCbv + TCsx ≤ Px(Q – X) – PvV + πy – PyY

diversification

(A.7)

In words, the profits earned from all household activities X and Y, minus whatever is spent on
other market goods, must be greater than all the transaction costs of buying market goods
(including inputs) and selling off household production. If true, then rational households will try
to avoid market transactions that are particularly expensive, such as the buying or selling of
coarse grain, which includes finding a trustworthy merchant, bargaining over the price,
researching the actual price, and still accepting a somewhat adjusted price to account for the
merchant’s transportation fees. Therefore, in villages where transaction costs are high, the model
predicts that farmers are more likely to produce their consumption demand themselves (i.e.,
production is partially determined by consumption preferences). Where this is true, a
household’s preference for variety and nutrition requirements often lead them to diversify their
crops—as long as the difference in profit from producing one crop over the other is less than the
transaction costs required to sell the more profitable one and subsequently purchase the latter.

110

Since this thesis is attempting to predict crop selection and land allocation, the potential effect of
transaction costs on crop diversification is an important factor to include in the econometric
model.
Finally, two specialized extensions of the agricultural household model discussed by de
Janvry and Sadoulet (2006) attempt to capture how market failures for seasonal credit and risk
affect production decisions. Seasonal credit constraints are modeled by adding a slight premium
to all sales and purchases representing the shadow price of liquidity, and risk requires intertemporal modeling to analyze household preference for credit and savings. Additionally,
Dorward (2011) recently modeled seasonal food rationing in rural households by dividing up key
consumption and production factors into separate variables representing before and after the
harvest. Arguing that a flaw of previous models is the assumption that households only aim to
produce future consumption, he predicts that current survival needs—possibly present in the preharvest “hunger” season—may compromise production goals. Such circumstances may lead a
household to harvest its coarse grains early before the crop has fully matured. Or if the family is
predicting a household food crisis prior to the planting season, a family may allot more land to a
faster-growing crop, and perhaps plant earlier, in order to meet pre-harvest consumption needs.
These factors, while difficult to model, are important to consider when selecting choice
dependent variables that may determine how farmers allocate land to various crops.

111

Appendix 4: Description of Logit, Fractional Logit, and Multinomial Logit Models
This appendix has three sections, the first of which reviews the standard logit model. The
second and third sections cover the fractional logit and the multinomial logit model, respectively.
Please see the body of the text in Section 5.1 for a description of the fractional multinomial logit
method employed by thesis.
A.4.1 The Logit Model for Binary Response
The econometric model used in this paper is more complicated than the basic logit model,
so a brief review is necessary before proceeding. To begin, a logit model can be used to estimate
response probabilities for binary independent variables. Its main feature (along with the probit
model) is that it defines the independent variable, a probability, through a function whose values
are between zero and one. This makes it an improvement over employing Ordinary Least
Squares (OLS) to regress a response probability with a Linear Probability Model (LPM), which
runs the risk of predicting probabilities that are negative or greater than one due to its linear and
unbounded nature. While the LPM states
P(y = 1|x) = β0 + βkxk, and βk = ∂P(y = 1|x) / ∂xk,

(A.8)

the logit model restrains the outcome of the explanatory variables using the function G, where
P(y = 1|x) = G(β0 + βkxk) = G(z), and G(z) = exp(z)/[1 + exp(z)].

(A.9)

Under this function, the response probability is always between zero and one. (Note that in a
probit model, which will not be discussed further here, G is the standard normal cumulative
distribution function).
Next, to estimate a logit model, I use maximum likelihood estimation (MLE) technique
instead of OLS, as it automatically accounts for heteroskedasticity. To obtain my estimator,

112

which is conditional on the explanatory variables, I need to find the density of y given x, which
can be written as
y

f(y|x;β) = f(y|z) = [G(z)] [1 – G(z)]

1-y

, y = 0,1.

(A.10)

Thus, when y = 1, I get G(z) and when y = 0, we get 1 – G(z). Finally, to get the estimator, I take
the log of the above equation in order to obtain the log-likelihood function:
log(β) = (y)log[G(x)] + (1 – y)log[1 – G(z)].

(A.11)

By summing equation (A.11) for all observations—under the assumption that all observations
are independent and identically distributed random variables—I calculate a log-likelihood
equation for the sample. Once maximized, this sample log-likelihood equation produces the logit
estimator β (Wooldridge 2006). This overview provides a foundation for explaining the
technicalities of the following variations and how they differ from the original logit model.
A.4.2 The Logit Model for Fractional Response
In their seminal work, Papke and Wooldridge (1996) adapted the logit model for
fractional response variables, which differ from binary variables in that instead of equaling either
outcome A or outcome B, they are equal to or fall into the interval between A and B. Therefore,
the model does not attempt to determine the probability of y given x, but the expected value of y,
bounded by two values A and B, given x. For example, proportion of weekly hours spent
working, participation rates in a voluntary pension plan, or share of farm land devoted to cotton
could all be fractional dependent variables, because they are bound between zero and one
hundred percent, hereafter simplified as one.
Because fractional response variables are bounded, determining their expected value
using OLS has similar consequences as applying the LPM to binary dependent variables; as
discussed above, the linear nature of the regression may potentially predict values that lay
113

outside the established bounds. For example, through OLS, one may find that given a particular
set of household characteristics, a farmer may devote -6 percent of his land to cotton, which is an
awkward outcome to explain. As with the standard logit model for binary outcomes, it is
important to ensure that the predictions for fractional dependent variables remain bounded.
As Papke and Wooldridge (1996) discuss, one alternative is to transform y into a logodds ratio within the OLS linear function, so that
E[f(y)|x] = E(log[y/(1-y)] | x) = β0 + βkxk,

(A.12)

but this presents two problems. First, even a well-defined model in this form makes it difficult to
recover the expected fractional value, since estimation of this equation yields the expected value
of the logged odds ratio. Secondly, while this transformation does bind the dependent variable
between zero and one, the equation does not hold true if y is equal to these bounds (e.g., if one of
the shares is zero or one). Therefore, since many households choose not to grow a particular
crop—meaning the share of that crop is zero—this alternative seems far from optimal.
This is where I can apply the logit framework from above, applied to the circumstances
of fractional dependent variables:
E(y | x) = G(β0 + βkxk) = G(z), and G(z) = exp(z)/[1 + exp(z)].

(A.13)

In addition to binding the expected value by zero and one, without excluding the possibility of it
equaling these values, the advantages to this transformation are that it makes no assumptions
about how the dependent variable is obtained or the data’s sample size. Furthermore, the logit
framework allows us to develop an estimator for β through the Bernoulli log-likelihood function,
which again, is
log(β) = (y)log[G(z)] + (1 – y)log[1 – G(z)].

114

(A.14)

Once summed for all observations and subsequently maximized, this provides a quasi-maximum
likelihood estimator (QMLE) of β that is both consistent and easy to calculate. In applying this
technique to data evaluating participation rates in 401(k) pension plans, Papke and Wooldridge
(2009) found their method to be robust and a better estimator than the linear model alternatives.
A.4.3 The Logit Model for Multinomial Response
The logit model can be extended to instances where the response y has more than two
categorical outcomes. For this study, y will represent five different categories of crop types.
While these alternatives may be assigned labels, such as the integers zero through four, these
assignments are arbitrary and are in no particular order. The purpose of the model is to determine
the probability that the dependent variable will be a certain categorical outcome given the
independent variables. Thus, in mathematical terms
P(y = j | x), j = 0, 1, 2, …, J.

(A.15)

In this case, the probability of growing a particular crop, whether it is cotton, maize or another,
for a given plot of land fits this description and is likely to vary as a result of changes in market
and household characteristics present in the explanatory variables.
A key facet of this expression is that it assumes that the explanatory variables do not
change for each alternative J. Rather, the set of x variables is specific to the household but not to
a given outcome. Therefore, it assumes that factors affecting crop choice are limited to market
and household attributes that could affect all possible crops, not to variables that could only
influence a particular crop without any effect on another. If crop-specific variables were
included, the model would have to be further generalized and become a conditional logit model.
However, the conceptual model in the previous chapter presents variables that could potentially

115

affect all possible outcomes in this study. Hence, maintaining the same explanatory variables for
each alternative does not pose a problem.
Continuing, as the dependent variable is bounded between zero and one as a probability,
it is necessary to apply the logit function from above, so that
P(y = j|x) = G(β0 + βkxk) = G(z) = exp(z)/[1 + ∑exp(z)],

j = 1, 2, …, J.

(A.16)

Notice that the denominator of the logit function now requires summing for all possible j
outcomes, and that if J = 1, we would simply return to the binary logit model. Also note that the
summing excludes the outcome j = 0. This is because it is necessary to set up j = 0 as the pivot
outcome, which will allow for estimating the effect of the explanatory variables on the
probability response of outcome j in relation to the probability response of outcome j = 0.
Otherwise it would be impossible to determine the coefficients within our system of equations.
Since it is assumed that a categorical outcome must be chosen, it follows that the
response probabilities must sum to one. Thus, it is known that the probability that y = 0 is the
difference between one and the sum of all other response probabilities. Therefore using the
identity that defines all other response probabilities above in (A.16), this means
P(y = 0|x) = 1 – ∑G(z) = 1 – ∑exp(z)/[1 + ∑exp(z)] = 1 / [1 + ∑exp(z)].

(A.17)

Given (A.17), I can now use the probability response for outcome 0 as a pivot for the other
outcomes. If the above identities (A.16) and (A.17) hold true, then
P(y = j|x) / P(y = 0|x) = {exp(z)/[1 + ∑exp(z)]} / {1 / [1 + ∑exp(z)]} = exp(z),

(A.18)

or with a log transformation
log [P(y = j|x) / P(y = 0|x)] = z = β0 + βkxk.

(A.19)

As with the standard logit model, these coefficients are best estimated by maximum likelihood
after summing the equation for all observations in the sample (Wooldridge 2001).
116

Appendix 5: Specification of the Fractional Multinomial Logit Model in STATA
To apply our econometric model to the dataset, I will employ the fractional multinomial
logit STATA package authored by Maarten Buis (2008, updated 2012). The program fits the data
to a fractional multinomial logit through quasi-maximum likelihood and, for this reason, implies
the robust option. The package also includes some post-estimation tools that can display
marginal effects and discrete changes, which may be utilized here.
Furthermore, it was described in Chapter 4 that the dataset consists of 150 households
from the Koutiala cercle surveyed in 2008/09 and 2009/10, creating a total of 300 observations.
In addition, 50 households from Simona in the Yorosso cercle were surveyed in 2010/11, making
350 observations in all, 50 from each of the seven villages. However, because the same 150
households represent 300 observations, this needs to be controlled for when calculating the
robust standard error. Even though there is a dummy variable for the 2009/10 season to account
for changes over time, there may still be unexplained traits that are unique to a particular
household that make it partial to dividing its crop shares in a specific way. In other words, these
observations are not independent from each other because they come from the same household.
Thankfully, STATA’s “cluster” option calculates the standard errors by group—in this case by
household identification number—so that the standard errors are robust to arbitrary
heteroskedasticity and serial correlation within households.

117

The notation used by the fractional multinomial logit package for a simple estimation is
“fmlogit (dependent variables), eta(independent variables).” Thus, if I incorporate the variables
chosen from Chapter 4 that follow from the agricultural household model, I get the following:
fmlogit CottonShare MaizeShare SorghumShare MilletShare SecondaryShare,
eta(MenPerHct WomenPerHct YouthPerHct InfantsPerHct Interaction1 %MenInactive
Interaction2 %WomenInactive Interaction3 Sprayers Plows Oxen Interaction4
Motorcycles Bicycles Carts %MenLiterate Interaction5 %WomenLiterate Interaction6
Bambara Senoufo Peulh OtherEthnic Year_2010 Village_Tonon Village_Kaniko
Village_TryI Village_Signe Village_Gantiesso Village_Simona)
cluster(Household_ID)

10

Note the use of the cluster option on each household to control for endogenous correlation within
observations from the same household.

10

Where Interaction1 = WomenPerHct* InfantsPerHct,
Interaction 2 = MenPerHct*%MenInactive, Interaction 3 = WomenPerHct*%WomenInactive,
Interaction 4 = Plows*Oxen, Interaction 5 = MenPerHct*%MenLiterate, and
Interaction 6 = WomenPerHct*%WomenLiterate.
118

Appendix 6: STATA Code for Calculation of Average Marginal Effects
To calculate average marginal effects, code was written in STATA to predict the changes
in specified outcome given a marginal change in each explanatory variable (and applicable
interaction terms). This was then bootstrapped for 50 iterations to estimate a meaningful standard
error. The following code was used to calculate one average marginal effect for one outcome—in
this case, the average marginal effect of MenPerHct on CottonShare. Note the incorporation of
relevant interaction terms when calculating the overall average marginal effect.
program drop _all
program my_ape_MenPerHct, rclass
quietly fmlogit CottonShare MaizeShare SorghumShare MilletShare SecondaryShare,
eta(MenPerHct WomenPerHct YouthPerHct InfantsPerHct Interaction1 %MenInactive
Interaction2 %WomenInactive Interaction3 Sprayers Plows Oxen Interaction4
Motorcycles Bicycles Carts %MenLiterate Interaction5 %WomenLiterate Interaction6
Bambara Senoufo Peulh OtherEthnic Year_2010 Village_Tonon Village_Kaniko
Village_TryI Village_Signe Village_Gantiesso Village_Simona) cluster(Household_ID)
predict share_hat1, pr outcome(CottonShare)
gen temp1 = MenPerHct
gen temp2 = Interaction2
gen temp3 = Interaction5
replace MenPerHct = MenPerHct + 0.0001
replace Interaction2 = MenPerHct*%MenInactive
replace Interaction5 = MenPerHct*%MenLiterate
predict share_hat2, pr outcome(CottonShare)
replace MenPerHct = temp1
replace Interaction2 = temp2
replace Interaction5 = temp3

119

gen mfxMenPerHct = (share_hat2 - share_hat1)/0.0001
summarize mfxMenPerHct
return scalar ape = r(mean)
drop share_hat1 temp1-temp3 share_hat2 mfxMenPerHct
end
bootstrap ape_MenPerHct = r(ape), rep(50) seed(123) cluster(No_expl)
idcluster(record_No_expl) nowarn:my_ape_MenPerHct
clear all

120

Appendix 7: Fractional Multinomial Logit Results Pivoted off of Cotton Share
Table A.1 displays the results from the fractional multinomial logit model using
CottonShare as its pivot outcome. In Chapter 4, it was discussed that in order for the fractional
multinomial logit to work—or any multinomial logit—it needs for one category to be the pivot
outcome from which all coefficients are based. Changing the pivot outcome does not change the
effect of one variable on a particular crop share, but only the perspective from which this effect
is observed. Thus, the difference in value between coefficients will not change if the pivot
changes, but the numerical values will as the each new pivot response shift what is defined as
zero. For this reason, if the coefficient on Y2, pivoted on Y1, was .05, the coefficient on Y1, if
pivoted on Y2, would be -.05. However, while all of the coefficients describing the effect on one
crop share relative to another can be derived from one table, the robust standard errors do vary
for every pivot outcome.
While not strictly consistent with the quasi-maximum likelihood estimation approach,
which returns expected crop shares in levels, one can interpret the coefficients in this Table A.1
directly in terms of a logged share ratio. Given the multinomial logit functional form for
expected crop shares, the logged expected share of a crop relative to that of the pivot crop—in
this case, cotton—is given by:
log [E(Sj | x) / E(S0 | x)] = z = β0 + βkxk.

(A.20)

Thus, 100*βk is the percent change in a crop’s expected share relative to that of cotton for a
marginal one-unit change in xk. If this value is negative, the crop’s share decreases relative to
that of cotton, while if this value is positive, the crop’s share increases relative to that of cotton.

121

Table A.1: Fractional Multinomial Logit Results Pivoted off of CottonShare
Log pseudolikelihood = -480.02629
Obs: 334
Wald Chi^2: 2004.7
Prob > Chi^2: 0.0000
Maize Share
Sorghum Share
Millet Share
Secondary Share
Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig
1 MenPerHct
0.9667 0.5167 * 0.3652 0.5894
0.5566 0.7122
0.3656 0.8698
2 WomenPerHct
-0.4804 0.8017
-1.0650 0.7762
-0.4102 0.8846
-1.4508 0.8711 *
3 YouthPerHct
0.0839 0.2296
0.0636 0.1916
-0.4840 0.2318 ** 0.1174 0.2593
4 InfantsPerHct
-0.7767 0.6522
-1.0736 0.5945 * -1.8323 0.6846 *** -1.8809 0.7757 **
5 Interaction: (2) x (5) 0.9365 1.4106
1.7754 1.1853
2.7099 1.2861 ** 2.3026 1.3273 *
6 %MenInactive
2.2720 1.2488 * 1.9097 1.6170
0.3501 1.1397
1.8023 1.4719
7 Interaction: (1) x (6) -3.1859 2.3939
-2.2233 3.0794
-0.3254 2.7253
-3.0942 3.4149
8 %WomenInactive
-0.1169 0.7208
-0.1586 0.6526
0.0007 0.7115
-0.5220 0.7602
9 Interaction: (2) x (8) 0.9700 1.9428
1.4236 2.0401
-0.1700 2.0930
1.1676 2.1969
10 Sprayers
-0.1665 0.0661 ** -0.2215 0.0652 *** -0.2762 0.0751 *** 0.0210 0.0872
11 Plows
-0.0309 0.0675
-0.1492 0.0539 *** -0.1304 0.0656 ** -0.2530 0.0871 ***
12 Oxen
-0.0357 0.0390
-0.1090 0.0447 ** -0.1386 0.0432 *** -0.1869 0.0598 ***
13 Interaction: (11) x (12) 0.0115 0.0081
0.0299 0.0064 *** 0.0246 0.0076 *** 0.0399 0.0107 ***
14 Motorcycles
0.0571 0.0937
-0.0740 0.0940
0.0938 0.1062
0.0266 0.1244
15 Bicycles
-0.0223 0.0481
0.0081 0.0502
0.0067 0.0597
-0.0661 0.0616
16 Carts
-0.2722 0.1683
-0.3285 0.1586 ** -0.2371 0.1891
-0.2439 0.2265
17 %MenLiterate
0.1524 0.3219
0.1795 0.3668
0.2369 0.4364
0.3077 0.4592
18 Interaction: (1) x (17) -0.6420 0.6777
-0.7557 0.7784
-0.7894 0.9265
-0.1619 1.0336
19 %WomenLiterate
-0.3121 0.3472
-0.9498 0.3581 *** -0.5441 0.4350
-0.5557 0.4616
20 Interaction: (2) x (19) 0.7095 0.6832
1.4396 0.6411 ** 0.8478 0.7622
0.7630 0.7118
21 Bambara
-0.6831 0.2450 *** -0.0120 0.2436
-0.1826 0.2440
-0.0971 0.4048
22 Senoufo
0.2739 0.2653
0.4354 0.2653
0.1455 0.3154
0.2580 0.3837
23 Peulh
-0.3360 0.2920
0.2558 0.2876
0.2696 0.4165
-0.5179 0.5090
24 OtherEthnic
-0.2578 0.3019
0.4450 0.3425
0.8419 0.5640
0.2214 0.5364
Significance Legend:
*** = P<.01
** = P < .05
* = P < .10
122

Table A.1 (cont'd): Fractional Multinomial Logit Results Pivoted off of CottonShare
Log pseudolikelihood = -480.02629
Obs: 334
Wald Chi^2: 2004.7
Prob > Chi^2: 0.0000
Maize Share
Sorghum Share
Millet Share
Secondary Share
Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig Coef Rbst SE Sig
25 Year_2010
-0.5653 0.1978 *** -0.4542 0.2465 * -0.5979 0.3342 * -0.7405 0.3814 *
26 Village_Tonon
0.9747 0.2725 *** 0.6106 0.2563 ** -0.7074 0.3841 * -0.0783 0.4351
27 Village_Kaniko
0.5786 0.2932 ** -0.1542 0.3035
0.3480 0.3453
0.7167 0.3618 **
28 Village_TryI
1.7042 0.4927 *** 1.9078 0.4567 *** 1.9018 0.4741 *** 2.5380 0.4777 ***
29 Village_Signe
0.8547 0.2521 *** 0.4771 0.2593 * 0.8036 0.2853 *** 0.8558 0.3642 **
30 Village_Gantiesso
0.1488 0.2419
-0.2929 0.2454
-0.5200 0.2878 * 0.6354 0.2998 **
31 Village_Simona
-0.2027 0.2302
-0.5461 0.2639 ** -0.9598 0.2995 *** -1.1583 0.3760 ***
Constant
0.1494 0.4564
2.4887 0.4740 *** 2.2872 0.4574 *** 1.8691 0.5614 ***
Significance Legend:
*** = P<.01
** = P < .05
* = P < .10

123

REFERENCES

124

REFERENCES

ACP Cotton. (2011). Mali - CMDT : 500.000 tonnes de coton-graine pour 2011 / Mali - CMDT:
500,000 tons of seed cotton for 2011. Retrieved September 3, 2012, from http://www.cotonacp.org/en/news/mali-CMDT-500000-tonnes-de-coton-graine-pour-2011-mali-CMDT-500000tons-seed-cotton-2011
ACP Cotton. (2012). Mali - La CMDT maintient le prix au producteur de coton pour 2012/13 /
CMDT keeps cotton producer price for 2012-2013. Retrieved September 3, 2012, from
http://www.coton-acp.org/en/news/mali-la-CMDT-maintient-le-prix-au-producteur-de-cotonpour-201213-CMDT-keeps-cotton-producer-pr
Ahn, C. Y., Singh, I., and Squire, L. (1981). A Model of an Agricultural Household in a MultiCrop Economy: The Case of Korea. The Review of Economics and Statistics, 63(4), 520-525.
Askari, H., and Cummings, J. T. (1977). Estimating Agricultural Supply Response with the
Nerlove Model: A Survey. International Economic Review, 18(2), 257-292.
Ball, R., and Pounder, L. (1996). `Efficient but poor' revisited. Economic Development and
Cultural Change, 44(4), 735.
Barnett, D., Blake, B., and McCarl, B. A. (1982). Goal Programming via Multidimensional
Scaling Applied to Senegalese Subsistence Farms. American Journal of Agricultural Economics,
64(4), 720-727.
Barnum, H., and Squire, L. (1979). A model of an agricultural household: theory and evidence.
Baltimore, MD: Published for the World Bank by Johns Hopkins University Press.
Barnum, H. N., and Squire, L. (1979). An econometric application of the theory of the farm
household. Journal of Development Economics, 6(1), 79-102.
Barrett, C. B. (2008). Smallholder market participation: Concepts and evidence from eastern and
southern Africa. Food Policy, 33(4), 299-317.
Benjamin, D. (1992). Household Composition, Labor Markets, and Labor Demand: Testing for
Separation in Agricultural Household Models. Econometrica, 60(2), 287-322.
Boer, A. J. D., and Chandra, S. (1978). A Model of Crop Selection in Semi-Subsistence
Agriculture and an Application to Mixed Agriculture in Fiji. American Journal of Agricultural
Economics, 60(3), 436-444.
Boughton, D. (1994). A Commodity Subsector Approach to the Design of Agricultural Research:
The Case of Maize in Mali (Doctoral dissertation, Michigan State University, 1994).

125

Boughton, D., and de Frahan, B. H. (1994). Agricultural Research Impact Assessment: The Case
of Maize Technology Adoption in Southern Mali (International Development Working Paper 41).
East Lansing, MI: Agricultural, Food and Resource Economics, Michigan State University.
Boughton, D., and Dembele, N. N. (2010). Rapid Reconnaissance of Coarse Grain Production
and Marketing in the CMDT Zone of Southern Mali: field work report of the IER-CSAPROMISAM team December 13 – 19, 2009 (Working paper 2010-1) East Lansing, MI: Food
Security III Cooperative Agreement, Michigan State University.
Buis, M. L. (2008). FRACTIONAL MULTINOMIAL LOGIT: Stata module fitting a fractional
multinomial logit model by quasi maximum likelihood [computer software]. Boston College
Department of Economics. Retrieved from http://ideas.repec.org/c/boc/bocode/s456976.html
Buis, M. L. (2012). fractional multinomial logit: module fitting a fractional multinomial logit
model by quasi-maximum likelihood. Retrieved September 5, 2012, from
http://maartenbuis.nl/software/fractional multinomial logit.html
CIA: Central Intelligence Agency. (2012). Mali. The World Factbook. Retrieved September 5,
2012, from https://www.cia.gov/library/publications/the-world-factbook/geos/ml.html
Delarue, J., Mesplé-Somps, S., Naudet, J. D., and Robilliard, A. S. (2009). Le paradoxe de
Sikasso: coton et pauvreté au Mali. (Développement, Institutions and Analyses de Long terme
working paper DT/2009-09). Retrieved September 9, 2012, from
http://www.dial.prd.fr/dial_publications/PDF/Doc_travail/2009-09GB.pdf
Diallo, A. (2011). An Analysis of the Recent Evolution of Mali’s Maize Subsector. (Master of
Science Plan B Paper, Michigan State University, 2011).
Dioné, J. (1989). Informing food security policy in Mali: interactions between technology,
institutions and market reforms. (Doctoral dissertation, Michigan State University, 1989).
Dorward, A. (2011). Conceptualising seasonal financial market failures and credit rationing in
applied rural household models. Retrieved from University of London, School of Oriental and
African Studies, Centre for Development, Environment and Policy at:
http://eprints.soas.ac.uk/10806/
Duflo, E. (2006). Poor but Rational? In A. V. Banerjee, R. Benabou and D. Mookherjee (Eds.),
Understanding poverty (pp. 367-378). New York, NY: Oxford University Press.
Fafchamps, M. (1992). Cash Crop Production, Food Price Volatility, and Rural Market
Integration in the Third World. American Journal of Agricultural Economics, 74(1), 90-99.
Fisk, E. K. (1962). Planning in a Primitive Economy: Special Problems of Papua New Guinea.
Economic Record, 38(84), 462-478.

126

Goetz, S. J. (1992). A Selectivity Model of Household Food Marketing Behavior in Sub-Saharan
Africa. American Journal of Agricultural Economics, 74(2), 444-452.
Hazell, P. B. R. (1982). Application of Risk Preference Estimates in Firm-Household and
Agricultural Sector Models. American Journal of Agricultural Economics, 64(2), 384-390.
Janvry, A. d., Fafchamps, M., and Sadoulet, E. (1991). Peasant Household Behaviour with
Missing Markets: Some Paradoxes Explained. The Economic Journal, 101(409), 1400-1417.
Janvry, A. d., and Sadoulet, E. (2006). Progress in the Modeling of Rural Households’ Behavior
under Market Failures. In A. Janvry and R. Kanbur (Eds.), Poverty, Inequality and Development
(Vol. 1, pp. 155-181): Springer US.
Kafo Jiginw. (2011). Annual Report to Simona. Simona, Mali.
Koch, S. F. (2010). Fractional Multinomial Response Models with an Application to Expenditure
Shares (Working paper 2010-21) Pretoria, South Africa: Economics, University of Pretoria.
Lee, D. J., Tipton, T., and Leung, P. (1995). Modeling cropping decisions in a rural developing
country: a multiple-objective programming approach. Agricultural Systems, 49(2), 101-111.
Mullahy, J. (2010). Multivariate Fractional Regression Estimation of Econometric Share Models
(NBER Working Papers 16354). Cambridge, MA: National Bureau of Economic Research.
Retrieved September 9, 2012, from http://www.nber.org/papers/w16354.
Murekezi, A., Mather, D., Kelly, V., and Boughton, D. (forthcoming). Effects of Household
Participation in Cotton Production on Coarse Grain Productivity in Mali: Evidence from the
Cotton Zone of Koutiala (International Development Working Paper). East Lansing, MI:
Agricultural, Food and Resource Economics. Michigan State University
Murteira, J. M. R., and Ramalho, J. J. S. (2012). Regression Analysis of Multivariate Fractional
Data (Working Paper). Retrieved on September 9, 2012, from http://www.eeaesem.com/files/papers/eea-esem/2012/2277/multivfracdata.pdf.
Nerlove, M. (1956). Estimates of the Elasticities of Supply of Selected Agricultural
Commodities. Journal of Farm Economics, 38(2), 496-509.
Nicholson, W., and Snyder, C. (2008). Microeconomic theory: basic principles and extensions
(10th ed.). Belmont, CA: Thomson Business and Economics.
North, D. C. (1981). Structure and Change in Economic History. New York: W.W. Norton.
Omamo, S. W. (1998). Farm-to-Market Transaction Costs and Specialisation in Small-Scale
Agriculture: Explorations with a Non-separable Household Model. Journal of Development
Studies, 35(2), 152-163.

127

Papke, L. E., and Wooldridge, J. M. (1996). Econometric Methods for Fractional Response
Variables With an Application to 401 (K) Plan Participation Rates. Journal of Applied
Econometrics, 11(6), 619-632.
Profoss, and NordNordWest (Cartographer). (2011). Mali - Sikasso. Retrieved from
http://en.wikipedia.org/wiki/File:Mali_-_Sikasso.svg#filehistory
Rarelibra (Cartographer). (2006). Map of the Cercles of Sikasso Region in Mali. Retrieved from
http://en.wikipedia.org/wiki/File:Sikasso_cercles.png
Sako, M. (forthcoming). (Master of Science Thesis, Michigan State University, 2011).
Samake, A., et al. (2008). Changements structurels des économies rurales dans la
mondialisation : Programme RuralStruc Mali - Phase II (World Bank Report). Bamako, Mali:
Michigan State University, IInstitut d’Economie Rurale du Mali, Centre de Coopération
Internationale en Recherche Agronomique pour le Développement.
Schultz, T. W. (1964). Transforming Traditional Agriculture. New Haven, Conn: Yale
University Press.
Schultz, T. W. (1980). Nobel Lecture: The Economics of Being Poor. Journal of Political
Economy, 88(4), 639-651.
Singh, I., Squire, L., and Strauss, J. (1986). A Survey of Agricultural Household Models: Recent
Findings and Policy Implications. The World Bank Economic Review, 1(1), 149-179.
Staatz, J., et al. (2011). Mali Agricultural Sector Assessment (Report prepared for United States
Agency for International Development). East Lansing, MI: Food Security III Cooperative
Agreement, Michigan State University.
STATA Annotated Output: Multinomial Logistic Regression. (2012). Retrieved August 26,
2012, from http://www.ats.ucla.edu/stat/stata/output/stata_mlogit_output.htm
Stent, W. R., and Webb, L. R. (1975). Subsistence Affluence and Market Economy in Papua
New Guinea. Economic Record, 51(4), 522-538.
Tefft, J., and Benoit-Cattin, M. (presenter). (2003). Mali's White Revolution: Smallholder Cotton
from 1960 to 2003. Presented at the International Conference on Successes in African
Agriculture Building for the Future, Pretoria, South Africa.
Tefft, J., Kelly, V., Wise, V., and Staatz, J. (2003). Linkages between Child Nutrition and
Agricultural Growth in Mali: A Summary of Preliminary Findings (Food Security International
Development Policy Syntheses Vol. 64) East Lansing, MI: Agricultural, Food and Resource
Economics, Michigan State University.

128

Theriault, V., Serra, R., and Sterns, J. A. (2010). Institutions in the Malian Cotton Sector:
Determinants of Supply. Retrieved from University of Florida, Food and Resource Economics,
on September 9, 2012, at: http://www.fred.ifas.ufl.edu/pdf/workingpaper2.pdf
Theriault, V., and Sterns, J. A. (2012). John R. Commons and the Evolution of Institutions: The
Case of the Malian Cotton Sector. Paper presented at the Agricultural and Applied Economics
Association's 2012 AAEA Annual Meeting, Seattle, Washington.
Udry, C. (1996). Efficiency and market structure: testing for profit maximization in African
agriculture. Department of Economics. Northwestern University. Evanston, Ill.
USAID: United States Agency for International Development. (2011). Country Overview: Mali.
Bamako, Mali.
Van Dusen, M. E., and Taylor, J. E. (2003). Missing Markets and Crop Diversity: Evidence from
Mexico. Environment and Development Economics, 10, 513-531.
Vitale, J. D., Djourra, H., and Sidibé, A. (2009). Estimating the supply response of cotton and
cereal crops in smallholder production systems: recent evidence from Mali. Agricultural
Economics, 40(5), 519-533.
Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data. Cambridge,
Mass.: MIT Press.
Wooldridge, J. M. (2006). Introductory Econometrics: A Modern Approach (3rd ed.). Mason,
OH: Thomson Higher Education.
Ye, X., and Pendyala, R. M. (2005). A Model of Daily Time Use Allocation Using Fractional
Logit Methodology. Paper presented at the 16th International Symposium on Transportation and
Traffic Theory, College Park Maryland, United States.

129