HOUSEHOLD STRUCTURE AND LABOR ALLOCATION: THE CASE OF RISK INSURANCE IN
MALI AND TECHNOLOGY ADOPTION IN BURKINA FASO
By
Aissatou Ouedraogo

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirement
for the degree of
Agricultural, Food and Resource Economics – Doctor of Philosophy
2017

ABSTRACT
HOUSEHOLD STRUCTURE AND LABOR ALLOCATION: THE CASE OF RISK INSURANCE IN
MALI AND TECHNOLOGY ADOPTION IN BURKINA FASO
By
Aissatou Ouedraogo
The low level of adoption of labor-saving technologies makes human labor a central productive factor
among small-scale farmers in sub-Saharan Africa. Yet, many of these farmers face missing labor markets
such that household composition becomes an important source of labor endowment. This dissertation
investigates family structure and labor allocation using detailed household agricultural production data
under two different contexts: (i) when households are subject to production risk and (ii) in the wake of the
introduction of a new agricultural technology.
The first study uses a two-wave plot level dataset collected in 2009 and 2012 and applies multiple
levels of fixed effect and alternative functional forms to investigate labor allocation across collective plots
–collectively owned by extended family members and managed by the heads of extended family
household– and individual plots managed by the other adult male members. The empirical findings
provide strong evidence that collective plots receive much more labor and achieve higher crop yield per
hectare relative to individual plots.
These findings, which are unanticipated with respect to what the theory on collective production
predicts, are rationalized by the claim that collective plots serve as a mechanism for an ex-ante risk
mitigation strategy against production uncertainty. Such a claim is tested using the historical coefficient
of variation of rainfall to show that households living in villages with higher rainfall variability present a
significantly higher labor allocation gap in favor of their collective plots relative to individual plots.
Furthermore, household level data analysis indicates that nuclear households in places with higher
historical rainfall variability are more likely to be engaged in collective farming.
The second study uses randomized control trial (RCT) data combined with the difference-indifference estimation method to investigate household labor response to the introduction of the

microdosing technology (a fertilizer application technique) among extended family households –formed
by multiple nuclear households– and purely nuclear family households in Burkina Faso. The findings
indicate that the intervention significantly increased labor allocation among nuclear households, which is
as expected, given that the microdosing technology is very labor-intensive. In contrast, we find that the
program significantly reduced labor allocation among extended households. Two explanations are
provided for this differential impact according to family structure.
The first relates to differences in control over productive resources across the two types of
households. Specifically, within the extended households, labor is shared across the collective and
individual plots. This labor arrangement is likely to constrain the amount of labor that can be allocated to
the plots receiving the microdosing technology. The second explanation is related to differences in
incentives. Nuclear households are sole claimants of any incremental output, which may give them
greater incentive to exert more effort to meet the labor requirement of the microdosing technology. In
contrast, extended family household members are rewarded based on their need rather than their efforts,
providing them less incentive relative to those in purely nuclear family households.
While the findings from the two studies seem at first glance contrasting, a close consideration of
the settings of the two studies shows that these findings are specific to their context. This calls attention to
how factors governing productive resource allocations among households with missing and incomplete
markets are numerous, multifaceted, and context-specific to the household structure. This highlights the
need for more microeconomic household analysis for better targeting aiming at increasing agricultural
productivity, and consequently, food security.

This dissertation is dedicated to my late brother, Saidou Ouedraogo

iv

ACKNOWLEDGEMENTS

I would like to earnestly thank my major professor and dissertation advisor, Andrew Dillon, for his great
research mentorship and moral support. I will forever be indebted to him and to all my other Dissertation
Guidance Committee members –Leah Lakdawala, Eduardo Nakasone, Maria Porter, and Melinda Smale –
for their invaluable advice and feedback. I am thankful to Dean Karlan, Lori Beaman, and Bram
Thuysbaert, for allowing the use of the Saving for Change data. I am also grateful to USAID and BASISAMA for generous funding for my PhD program. Many thanks to participants of the Development
Brown Bag, my Research Lunch Group (Ashesh Prasann, Edouard Mensah, Godwin Debrah, Joey Goeb,
Joshua Gill, and Sarah Kopper), for their useful comments.
Words are not enough to express my gratefulness to my friends Marcia Benfica, Assa DembelĂŠ,
Marie Steele-Adjognon and Awa Sanou, for generous and warm-heartedness childcare. I am all gratitude
to my friends Hamza Haider, Kazianga Harounan, Nouhoum Traore, and Ayala Wineman for always
being there to answer my numerous questions.
This dissertation would not have been possible without the support of my loving and caring
husband, Yacouba Sore, who helped me keep my sanity throughout the completion of my degree. My
heartfelt appreciation to my kids Kameala and Wendyam for their love and sacrifice. I remain indebted to
my sisters (Adjaratou, Aminata, Awa, and Sarata), brothers (Adama, Idrissa, Youssouf), and my
stepmother (Salimata Sawadogo) for their support, lover, and for believing in me.
Lastly but not least, my highest gratitude to the Almighty for blessing my live with three
exceptional people –my late father (Mahamoud), late mother (Habibah Rabo) and late brother (Saidou) –
who sadly are no longer with us, but whose unconditional love, affection, and faith, will always shape my
life. May they all rest in peace.

v

TABLE OF CONTENTS

LIST OF TABLES ..................................................................................................................................... viii
LIST OF FIGURES ..................................................................................................................................... xi
KEY TO ABBREVIATIONS ..................................................................................................................... xii
INTRODUCTION ........................................................................................................................................ 1
Motivation ................................................................................................................................................. 1
Research contribution ............................................................................................................................... 2
REFERENCES ............................................................................................................................................. 5
1 INTRA-HOUSEHOLD LABOR ALLOCATION AND EVIDENCE OF RISK INSURANCE
AMONG EXTENDED FAMILY MEMBERS IN MALI ............................................................................ 9
1.1
Introduction ................................................................................................................................... 9
1.2
Background ................................................................................................................................. 13
1.3
Conceptual framework ................................................................................................................ 15
1.4
Data source and sample restrictions ............................................................................................ 20
1.4.1
Data source.......................................................................................................................... 20
1.4.2
Sample restrictions .............................................................................................................. 21
1.5
Descriptive statistics ................................................................................................................... 23
1.6
Empirical model .......................................................................................................................... 29
1.6.1
Specifications with household-year-crop fixed effect......................................................... 29
1.6.2
Specifications with plot-manager-year fixed effect ............................................................ 34
1.6.3
Adding weather variability to the household-crop-year fixed effect .................................. 36
1.7
Results: Labor allocation and crop yield..................................................................................... 39
1.7.1
Results with household-crop-year fixed effects .................................................................. 39
1.7.2
Labor intensity and crop yield with plot-manager-year fixed effects ................................. 46
1.8
Evidence for motive of risk insurance ........................................................................................ 51
1.8.1
Rainfall variability and labor allocation .............................................................................. 51
1.8.2
Rainfall variability and participation in collective farming ................................................ 57
1.9
Conclusion and policy implications ............................................................................................ 60
APPENDICES ............................................................................................................................................ 64
APPENDIX A ADDITIONAL DESCRIPTIVE STATISTICS .............................................................. 65
APPENDIX B ROBUSTNESS CHECK BY DROPPING ALL FERTILIZED PLOTS ....................... 69
APPENDIX C GENDER BIAS IN LABOR ALLOCATION AND CROP YIELD.............................. 71
APPENDIX D LABOR HETEROGENEITY BY PLOT MANAGEMENT STRUCTURE ................. 73
APPENDIX E ROBUSTNESS CHECK USING DECILES OF PLOT SIZE ....................................... 76
APPENDIX F ROBUSTNESS CHECK WITH OVERLAPPING PLOT SIZE .................................... 78
APPENDIX G ROBUSTNESS CHECK BY DROPPING PLOTS WITH COMPANION CROPS ..... 80
APPENDIX H ESTIMATION OF THE LOG OF CROP YIELD ACROSS COLLECTIVE AND
HEADS’ INDIVIDUAL PLOTS ............................................................................................................ 81
APPENDIX I YIELD DISPERSION ACROSS COLLECTIVE AND INDIVIDUAL PLOTS ............ 82
REFERENCES ........................................................................................................................................... 84
2 HOUSEHOLD STRUCTURE INFLUENCES LABOR ALLOCATION DECISIONS DUE TO
AGRICULTURAL TECHNOLOGY ADOPTION: EVIDENCE FROM BURKINA FASO ................... 89
2.1
Introduction ................................................................................................................................. 89
vi

2.2
Conceptual framework ................................................................................................................ 92
2.2.1
Experimental Design ........................................................................................................... 92
2.2.2
Household Structure ............................................................................................................ 95
2.3
Identification Strategy ................................................................................................................. 97
2.4
Data ............................................................................................................................................. 99
2.4.1
Sample Restrictions........................................................................................................... 100
2.4.2
Demographic and Production Characteristics by Household Structure ............................ 101
2.4.3
Pre-Treatment Household Demographic and Production Characteristics ......................... 105
2.5
Results ....................................................................................................................................... 108
2.5.1
Results with the full sample .............................................................................................. 108
2.5.2
Results with the restricted sample ..................................................................................... 110
2.6
Understanding the sources of the heterogeneous impact of the program ................................. 116
2.6.1
Labor fixity within the extended households and differences in incentive? ..................... 116
2.6.2
Differences in productive resource control within the extended households? .................. 118
2.7
Conclusion and policy implications .......................................................................................... 119
APPENDICES .......................................................................................................................................... 122
APPENDIX A MICRODOSING TAKE-UP BY HOUSEHOLD STRUCTURE................................ 123
APPENDIX B CHANGES IN LABOR ALLOCATION FOR NON SORGHUM CROPS ................ 124
APPENDIX C CHANGES IN SORGHUM LAND SHARE AND NUMBER OF PLOTS ................ 125
APPENDIX D CHANGES IN LABOR ALLOCATION BY HEADSHIP AT THE PLOT-MANAGER
LEVEL .................................................................................................................................................. 126
APPENDIX E CHANGES IN LABOR ALLOCATION BY HEADSHIP AT PLOT-MANAGER
LEVEL (EXTENDED HOUSEHOLDS ONLY) ................................................................................. 127
REFERENCES ......................................................................................................................................... 128
SUMMARY AND CONCLUDING REMARKS ..................................................................................... 133
Summary ............................................................................................................................................... 133
Concluding remarks .............................................................................................................................. 135
REFERENCES ......................................................................................................................................... 139

vii

LIST OF TABLES

Table 1.1 Empirical specifications and their sub-sample............................................................................ 22
Table 1.2 Plot characteristics, crop choice, and inputs across collective and male individual plots .......... 24
Table 1.3 Labor allocation across collective and male individual plots ..................................................... 27
Table 1.4 Estimation of log of labor allocation (person-days/ha) across collective and male individual
plots .................................................................................................................................................... 41
Table 1.5 Estimation of log of yield (FCFA/ha) across collective and male individual plots .................... 44
Table 1.6 Estimation of log of labor allocation (person-days/ha) across collective and heads’ individual
plots using PMYFE ............................................................................................................................ 48
Table 1.7 Estimation of log of labor allocation (person-days/ha) across collective and male individual
plots with rainfall variability (without plot covariates) ...................................................................... 53
Table 1.8 Estimation of log of labor allocation (person-days/ha) across collective and male individual
plots controlling for rainfall (controlling for plot covariates) ............................................................ 54
Table 1.9 Estimation of log of total labor allocation (person-days/ha) across collective and male
individual plots controlling for rainfall by grain and non-grain production....................................... 56
Table 1.10 Estimation of the log of yield for all crops (FCFA/ha) across collective and male individual
plots controlling for rainfall variability .............................................................................................. 57
Table 1.11 Test for collective farming as ex-ante risk mitigation strategy. Dependent variable is
participation of nuclear household in collective farming ................................................................... 58
Table 1A. 1 Plot characteristics and crop choice across collective and individual plots (Full sample) ...... 65
Table 1A. 2 Plot characteristics and crop choice across all collective and male individual plots .............. 65
Table 1A. 3 Plot characteristics by crop choice across collective and male individual plots ..................... 66
Table 1A.4. Coefficients of correlation between rainfall variability and village covariates....................... 68
Table 1B.1 Log of labor allocation estimation after dropping all plots with inorganic and organic fertilizer
............................................................................................................................................................ 69
Table 1B.2 Log of crop yield (FCFA/ha) estimation after dropping all plots with inorganic and organic
fertilizer .............................................................................................................................................. 69
Table 1C. 1 Log of labor allocation estimation across collective and all individual plots (including female
plots) ................................................................................................................................................... 71

viii

Table 1C. 2 Log of crop yield (FCFA/ha) estimation across collective plots and all individual plots
(including female plots)...................................................................................................................... 72
Table 1D. 1 Linear probability results of labor heterogeneity across collective and individual plots
(without plot covariates) ..................................................................................................................... 74
Table 1D. 2 Linear probability results of labor heterogeneity across collective and individual plots
(controlling for plot covariates) .......................................................................................................... 74
Table 1E. 1 Log of labor allocation estimation across collective and individual plots using plot deciles .. 76
Table 1E. 2 Log of crop yield estimation across collective and individual plots using plot deciles........... 77
Table 1F. 1 Log of labor allocation estimation across plots with overlapping plot size ............................. 78
Table 1F. 2 Log of crop yield (FCFA/ha) estimation across plots with overlapping plot size ................... 78
Table 1G. Log of labor allocation estimation after dropping all plots with companion crops ................... 80
Table 1H. Estimation of the log of crop yield (FCFA/ha) across collective and heads’ individual plot
using PMYFE ..................................................................................................................................... 81
Table 1I. Yield variability across collective and male individual plots ...................................................... 82
Table 2.1 Sample Restrictions and Household Structure ............................................................................ 96
Table 2.2 Demographic Characteristics Across Extended and Nuclear Households (Full sample) ......... 102
Table 2.3 Production Characteristics Across Extended and Nuclear Households (Full sample).............. 102
Table 2.4 Demographic Characteristics Across Extended and Nuclear Households (Restricted Sample)104
Table 2.5 Production Characteristics Across Extended and Nuclear Households (Restricted Sample) ... 104
Table 2.6 Balancing Test Results Across Assigned and Control Households by Household (Full sample)
.......................................................................................................................................................... 106
Table 2.7 Balancing Test Results Across Assigned and Control Households by Household (Restricted
sample) ............................................................................................................................................. 107
Table 2.8 Changes in log of labor allocation across baseline and follow-up (Full Sample) ..................... 108
Table 2.9 Changes in log of labor allocation across baseline and follow-up (Restricted Sample) ........... 111
Table 2.10 Labor allocation results by activity using sample restricted to the lower three quintiles of total
sorghum land .................................................................................................................................... 113
Table 2.11 Changes in log of labor allocation across baseline and follow-up by labor source (Restricted
sample) ............................................................................................................................................. 115
Table 2A. Take-up of microdosing by household structure after program implementation ..................... 123

ix

Table 2B. Changes in labor allocation across baseline and follow-up for crops other than sorghum ...... 124
Table 2C. Changes in sorghum land share and number of plots across baseline and follow-up .............. 125
Table 2D. Changes labor allocation across heads and non-heads members between baseline and follow-up
(All male plot managers) .................................................................................................................. 126
Table 2E. Changes labor allocation across heads and non-heads members between baseline and follow-up
(All male plot managers within extended households) .................................................................... 127

x

LIST OF FIGURES

Figure 1.1 Map of Mali with the Segou region ........................................................................................... 13
Figure 1.2 The “Cercles” included in the study zone.................................................................................. 13
Figure 1.3 Plot size (ha) distribution across collective and individual plots (all crops) ............................. 25
Figure 1.4 Plot size (ha) distribution across collective and individual plots by grain and non-grain crops 26
Figure 1.5 Labor heterogeneity across collective and individual plots (All crops) .................................... 27
Figure 1.6 Labor heterogeneity across collective and individual plots (By cereal and non-cereal crops) .. 28

xi

KEY TO ABBREVIATIONS

AGRODIA

Association des Grossistes et Détaillants d’Intrants Agricoles

CFA

CommunautÊ Financière Africaine

CV

Coefficient of variation

DID

Difference-in-difference

FAO

Food and Agriculture Organization

FCFA

Francs CommunautÊ Financière Africaine.

FE

Fixed effects

Ha

Hectare

HH

Households

HHCYFE

Household-crop-year fixed effect

ICRISAT

International Crops Research Institute for the Semi-Arid Tropics

INERA

Institut de l'Environnement et de Recherches Agricoles

IPA

Innovations for Poverty Action

ITT

Intend to treat

Kg

Kilograms

PMYFE

plot-manager-year fixed effects

RCT

Randomized control trial

SD

Standard Deviation

SfC

Saving for Change

USAID

United States Agency for International Development

xii

INTRODUCTION
Motivation
Most of the world’s food insecure live in Sub-Saharan Africa (SSA) where more than a quarter of the
population is
undernourished (FAO, 2014). A high proportion of these people rely on subsistence farming as their main
source of livelihood. At the same time, agricultural production in the region is plagued by low
productivity due to multiple factors including low and decreasing soil fertility (Drechsel et al., 2001;
Nachtergaele, F. O. et al., 2011; Jones et al., 2013; Josephson et al., 2014) combined with low agricultural
technology adoption rates, and limited access to productive resources, such as labor (Croppenstedt and
Goldstein, 2013; Moser and Barrett, 2003; Lee et al., 2006; Barrett et al., 2008; Smale et al., 2013), inter
alia.
Access to labor has important implications for agricultural productivity among
households in this region, especially in West Africa, where a substantial empirical literature finds that
differences in labor access are strongly reflected in crop yield performance (Udry, 1996; Goldstein and
Udry, 2008; Kazianga and Wahhaj, 2013; Guirkinger et al., 2015; Guirkinger and Platteau, 2017). These
finding mostly stem from the fact that labor is an important productive factor for households in this
region, especially among small-scale subsistence farmers. This is not only because it compensates for the
lack of agricultural technology adoption (de Ridder et al., 2004; West, 2010), but also because most
agricultural technologies available in this region are labor intensive (Moser and Barrett, 2003; Barrett et
al., 2004, Kijima et al., 2011). Trends in agricultural technology adoption show that an overwhelmingly
high proportion of subsistence farmers still heavily rely on hand tools in SSA, due to a decreasing trend in
agricultural mechanization (Ashburner and Kienzle; 2011; Baudron et at., 2015; Sheahan and Barrett,
2014). This makes the availability of human power crucial for successful agricultural intensification.
Despite the central role of human labor in agricultural production among small-scale subsistence
households, they usually face missing or incomplete labor markets and must rely on family members to

1

satisfy their labor needs. In this context, household composition becomes an important source of labor
endowment. This makes an understanding of the relationship between household structure and labor
allocation decision-making relevant for achieving successful implementation of agricultural development
programs.
However, accurately identifying such factors is challenging, especially among rural subsistence
households in SSA, where factors governing the allocation of productive resources are numerous,
multifaceted, and often context-specific to the household structure (Kazianga and Wahhaj, 2017;
Theriault et al., 2017). This underscores the need for further research related to household structure and
productive factor allocation. This has motivated the current dissertation on linkages between household
structure and labor allocation under distinct production organizations in Burkina Faso and Mali.

Research contribution
The dissertation is organized in two parts. In the first part, I investigate how labor allocation
varies across plots within the same household, depending on plot management structure– that is, plots that
are collectively owned by all extended family members in the households or individual plots owned by
nuclear family households. This part of the dissertation contributes to the literature in several ways. First,
to my knowledge, no other study applies plot-manager-year fixed effects to investigate labor allocation
among extended households, where the same plot manager simultaneously manages collective plots and
his own individual plots. Such an approach provides a powerful way to establish that any differences
observed in labor allocation across the collective and individual plots can be solely attributed to the role
that collective plots play within the household, rather than the characteristics of the individual who
manages them.
Second, the study sheds light on a scantily understood subject, which is the role of collective
farming as an ex-ante production risk mitigation strategy. To capture historical exposure to production
uncertainty, I follow the approach of Dillon et al. (2011) by computing the historical coefficient of
variation of rainfall over a period of twenty-seven years (1981-2007). I then use this coefficient of
2

variation to test my claim about collective farming being an ex-ante risk mitigation strategy. Specifically,
I investigate differences in labor allocation across collective and individual plots among extended family
households across villages with different levels of rainfall variability.
The notion that collective farming can be favored when production uncertainty prevails has been
mentioned in the theoretical literature (Putterman, 1983; Carter, 1987; Deininger, 1995; West, 2010) but
with very limited empirical evidence. Among the few studies is the paper by Rosenzweig and Wolpin
(1985), where the authors exploit weather variability to investigate variation in household-level farm
profits by family structure in India. To my knowledge, the current study is the first to use detailed plotlevel data, combined with historical rainfall variability, to empirically investigate collective farming
among extended family household members as an ex-ante risk mitigation strategy.
The second part of the dissertation examines how nuclear and extended family households
allocate labor differently in response to the introduction of a new agricultural technology in Burkina
Faso.1 The technology in question is microdosing, which is a specific fertilization technique. The
technology presents important distinctive features. It has the potential to significantly increase crop yield
and requires much less fertilizer, making it less financially constraining and more environmentally
friendly relative to the broadcasting technique (ICRISAT; 2012; Resnick et al., 2012). These features lead
some analysts to perceive this technology as a promising way to achieve sustainable agricultural
intensification (Juma et al., 2013, Aune and Coulibaly, 2015). However, it is very labor-intensive,
especially at the time of fertilizer application and planting. The benefits of microdosing on one hand, and
its intensive labor requirement on the other hand, make it an ideal subject to examine labor responses to
its introduction.
The study uses data from a randomized control trial (RCT) in which the microdosing technology
is exogenously introduced to the targeted households. The exogenous variation in the availability of the
technology provides a strong way to establish the direction of causality of the differential impact of the

1

This study is co-authored by Andrew Dillon and Maria Porter.

3

technology based on household structure. Moreover, to increase precision in our estimates, we combine
the RCT data with the difference-in-difference estimation method. To our knowledge, the current work is
the first to apply this approach to investigate the potential household structure differential effects of a new
agricultural technology on labor allocation.

4

REFERENCES

5

REFERENCES

Ashburner, J. E., & Kienzle, J. (2011). Investment in agricultural mechanization in Africa: Conclusions
and recommendations of a round table meeting of experts (3-5 June 2009. Arusha, Tanzania). FAO,
Roma (Italia).
Aune, J.B. and Coulibaly, A. (2015). Microdosing of Mineral Fertilizer and Conservation Agriculture for
Sustainable Agricultural Intensification in Sub-Saharan Africa. In Sustainable Intensification to
Advance Food Security and Enhance Climate Resilience in Africa (pp. 223-234). Springer
International Publishing.
Barrett, C. B., Moser, C. M., McHugh, O. V., & Barison, J. (2004). Better technology, better plots, or
better farmers? Identifying changes in productivity and risk among Malagasy rice farmers. American
Journal of Agricultural Economics, 86(4), 869-888.
Barrett, C. B., Sherlund, S. M., & Adesina, A. A. (2008). Shadow wages, allocative inefficiency, and
labor supply in smallholder agriculture. Agricultural Economics, 38(1), 21-34.
Baudron, F., Sims, B., Justice, S., Kahan, D. G., Rose, R., Mkomwa, S., & GĂŠrard, B. (2015). Reexamining appropriate mechanization in Eastern and Southern Africa: two-wheel tractors,
conservation agriculture, and private sector involvement. Food Security, 7(4), 889-904.
Carter, M. R. (1987). Risk sharing and incentives in the decollectivization of agriculture. Oxford
Economic Papers, 39(3), 577-595.
Clarke, L., & Bishop, C. (2002). Farm power-present and future availability in developing
countries. Agricultural Engineering International: CIGR Journal.
Croppenstedt, A., Goldstein, M., & Rosas, N. (2013). Gender and agriculture: inefficiencies, segregation,
and low productivity traps. The World Bank Research Observer, 28(1), 79-109.
de Ridder, N., Breman, H., van Keulen, H., & Stomph, T. J. (2004). Revisiting a ‘cure against land
hunger’: soil fertility management and farming systems dynamics in the West African
Sahel. Agricultural Systems, 80(2), 109-131.
Deininger, K. (1995). Collective agricultural production: A solution for transition economies? World
Development, 23(8), 1317-1334.
Dillon, A., Mueller, V., & Salau, S. (2011). Migratory responses to agricultural risk in northern Nigeria.
American Journal of Agricultural Economics, 93(4), 1048-1061.
Drechsel, P., Gyiele, L., Kunze, D., & Cofie, O. (2001). Population density, soil nutrient depletion, and
economic growth in sub-Saharan Africa. Ecological Economics, 38(2), 251-258.
FAO, IFAD. "WFP (2014) The state of food insecurity in the world 2014: strengthening the enabling
environment for food security and nutrition." FAO, Rome.Goldstein, M., & Udry, C. (2008). The
profits of power: Land rights and agricultural investment in Ghana. Journal of Political
Economy, 116(6), 981-1022.

6

Gollin, D., & Rogerson, R. (2014). Productivity, transport costs and subsistence agriculture. Journal of
Development Economics, 107, 38-48.
Guirkinger, C., Platteau, J. P., & Goetghebuer, T. (2015). Productive inefficiency in extended agricultural
households: Evidence from Mali. Journal of Development Economics, 116, 17-27.
Guirkinger, C., & Platteau, J. P. (2017). Transformation of African Farm Households: A Short Survey of
Economic Contributions. Journal of Demographic Economics, 83(1), 41-50.
ICRISAT. 2012. Fertilizer Microdosing, July 2012.
Jones, A., Breuning-Madsen, H., Brossard, M., Dampha, A., Deckers, J., Dewitte, O., & Le Roux, P.
(2013). Soil atlas of Africa.
Josephson, A. L., Ricker-Gilbert, J., & Florax, R. J. (2014). How does population density influence
agricultural intensification and productivity? Evidence from Ethiopia. Food Policy, 48, 142-152.
Juma, C., Tabo, R., Wilson, K., & Conway, G. (2013). Innovation for sustainable Intensification in
Africa. London: The Montpellier Panel, Agriculture for Impact.Kazianga, H., & Wahhaj, Z. (2013).
Gender, Social Norms, and Household Production in Burkina Faso. Economic Development and
Cultural Change, 61(3), 539-576.
Kazianga, H., & Wahhaj, Z. (2017). Intra-household resource allocation and familial ties. Journal of
Development Economics, 127, 109-132.
Kijima, Y., Otsuka, K. and Sserunkuuma, D. (2011). An inquiry into constraints on a green revolution in
Sub-Saharan Africa: the case of NERICA rice in Uganda. World Development, 39(1), 77-86.
Lee, D. R., Barrett, C. B., & McPeak, J. G. (2006). Policy, technology, and management strategies for
achieving sustainable agricultural intensification. Agricultural economics, 34(2), 123-127.
Moser, C. M., & Barrett, C. B. (2003). The disappointing adoption dynamics of a yield-increasing, low
external-input technology: the case of SRI in Madagascar. Agricultural Systems, 76(3), 1085-1100.
Nachtergaele, F. O., Petri, M., Biancalani, R., Van Lynden, G., Van Velthuizen, H., & Bloise, M. (2011).
An Information database for land degradation assessment at global level. Global land degradation
information system (GLADIS).
Putterman, L. (1983). A modified collective agriculture in rural growth-with-equity: Reconsidering the
private, unimodal solution. World Development, 11(2), 77-100.
Resnick, D., Tarp, F., & Thurlow, J. (2012). The political economy of green growth: Illustrations from
Southern Africa (No. 2012/11). WIDER Working Paper.Rosenzweig, M. R., & Wolpin, K. I. (1985).
Specific experience, household structure, and intergenerational transfers: farm family land and labor
arrangements in developing countries. The Quarterly Journal of Economics, 961-987.
Sheahan, M., & Barrett, C. B. (2014). Understanding the agricultural input landscape in sub-Saharan
Africa: Recent plot, household, and community-level evidence.
Smale, M., Byerlee, D., & Jayne, T. (2013). Maize revolutions in sub-Saharan Africa. In an African green
revolution (pp. 165-195). Springer Netherlands.

7

Theriault, V., Smale, M., & Haider, H. (2017). How Does Gender Affect Sustainable Intensification of
Cereal Production in the West African Sahel? Evidence from Burkina Faso. World Development, 92,
177-191.
Udry, C. 1996. “Gender, Agricultural Production, and the Theory of the Household.” Journal of Political
Economy, 104(5), 1010-1046.
West, C. T. (2010). Household Extension and Fragmentation: Investigating the Socio-Environmental
Dynamics of Mossi Domestic Transitions. Human Ecology 38.3, pp.363-376.

8

1

1.1

INTRA-HOUSEHOLD LABOR ALLOCATION AND EVIDENCE OF RISK
INSURANCE AMONG EXTENDED FAMILY MEMBERS IN MALI
Introduction

Labor plays a crucial role in agriculture among many farmers in Sub-Saharan Africa due to limited access
to labor-saving technologies. Its availability for crop cultivation is important such that plot-level crop
yield performance is contingent on how much labor is allocated to the plot. Indeed, numerous studies on
households in West Africa find differences in crop yield across plots with similar characteristics within
the same household due to differences in labor allocation across these plots (Udry, 1996; Kazianga and
Wahhaj, 2013; Guirkinger and Platteau, 2014; Guirkinger et al., 2015; Kazianga and Wahhaj, 2017;
Guirkinger and Platteau, 2017; Theriault et al., 2017). Given that labor remains a central productive
resource among these farmers (Moser and Barrett, 2003; Barrett et al., 2004; de Ridder et al., 2004; Lee et
al., 2006; West, 2010; Kijima et al., 2011), it is important to understand the underlying factors governing
the household labor allocation decisions to better promote agricultural productivity. This constitutes an
important path for achieving agricultural intensification and, therefore, food security, as most households
in this region mainly rely on agriculture for their livelihood.
However, the primary factors leading to differences in labor allocation are not only multifaceted,
ranging from gender differential impact to social norms, but can also be location specific. Such a
complexity of the household labor allocation decision-making highlights the challenges faced by
development practitioners in this region to improve aggregate measures of household welfare in contexts
where intra-household dynamics can be location-specific. This emphasizes the need for continuous
research efforts in understanding household production decision making in these regions for successful
development policy implementation.2
The current study aims to contribute to this small but growing literature using plot-level data to

2

Using household plot level data from Burkina Faso, Kazianga and Wahhaj (2013) show that plots controlled by the
head of household benefit from higher levels of labor input provision and achieve higher crop yields relative to plots
controlled by other household members. In a similar study carried out by Guirkinger and Platteau (2014) in Mali, the
authors find the opposite to be the case, with farmers devoting greater energy to their own individually managed
plots compared to the collective plots

9

provide an alternative explanation of plausible sources of differences in labor allocation among extended
family households in the Segou region in Mali. Households in this region jointly farm collective plots,
which are controlled by household heads, with output and labor shared by all family members, while
managing their own individual plots, with output and labor shared by immediate family members only.
The theory on collective farming predicts that this type of production organization is likely to
lead to moral hazard on the collective plots (Israelsen, 1980; Carter, 1984; Meyer, 1989). Common
recounted issues are coordination and supervision problems which lead to low input provision and,
consequently, to low levels of output in the production of collective goods (Lin, 1987; Mcmillan, 1987).
Indeed, this theory is supported by existing empirical evidence among rural households in another region
in Mali (Guikinger and Platteau, 2014).
Preliminary empirical results in the current paper, however, provide evidence that depart from
this theory. Specifically, using a two-wave household panel data collected in 2009 and 2012 with a unique
feature that clearly distinguishes collective plots from individual, the study shows with compelling
evidence that collective plots receive significantly much more labor per unit of land and present higher
crop yield performance relative to individual plots.3
The paper makes the claim that the higher labor allocation on the collective plots can be attributed
to motives for ex-ante risk mitigation for these households, to insure themselves against production risk.
Following the same strategy as in Dillon et al. (2011), rainfall variability is used as a proxy for production
risk. Specifically, rainfall variability is computed as the coefficient of variation of rainfall over a period of
twenty-seven years (1981-2007) for each village in the sample.
The view of collective farming among extended family members as risk mitigation strategy has
been explored theoretically but with sparse empirical application. One of the very few studies is by

3

The results remain robust through alternative econometric specifications and sample restrictions, even after
limiting the sample to the same individuals simultaneously managing collective plots and their own individual plots.

10

Rosenzweig and Wolpin (1985) with a focus on variation in farm profits by family structure.4 The current
work focuses on the importance of the role of labor, as a productive factor, by taking a more
microeconomic analysis approach –that is, plot-level labor allocation –when production uncertainty due
to weather variability exists. To my knowledge, no existing empirical work has looked at the links
between ex ante exposure to production risk and intra-household plot-level labor allocation.5 Collective
farming among extended households is also prevalent in other West African countries, such as Burkina
Faso (Kazianga and Wahhaj, 2016). Many of these households share similar production characteristics
and face the same insurance markets failure for consumption smoothing (Townsend, 1994; Dercon, 2002;
Kazianga and Udry, 2006). Consequently, results from the current work would provide insight, not only
in the context of Mali, but also could be generalized beyond the case of Mali.
The argument that collective farming serves as insurance against production risk due to weather
variability may seem at first glance counter-intuitive, since rainfall variability is perceived as a covariate
shock and, consequently, assumed to affect all plots in a given region in a similar manner. Nevertheless,
the setting of the study area makes this arrangement mechanism suitable for ex ante risk mitigation.
Specifically, rainfall, which usually occurs from May to September, constitutes the main source of water
for plots in the region, as households have very limited access to irrigation. Timing of agricultural
activities during this period is critical for crop yield performance (Sultan et al., 2005; Mishra et al., 2008;
Cooper et al., 2008; West, 2010) and, consequently, the household’s ability to secure food availability.
For these households, labor remains an important production factor (de Ridder et al., 2004; Moser
and Barrett, 2003; Barrett et al., 2004; Lee et al., 2006; Kijima et al., 2011), making its availability crucial
to avoid production shortfall. Securing labor availability becomes more critical when weather instabilities

4

Rosensweig and Wolpin (1985) investigate differences in farm profit variation across intergenerational family
households relative to families without elderly members. Kazianga and Wahhaj (2017) tackle the question of risk
sharing among extended family members in Burkina Faso using consumption data. Also, the authors focus on ex
post shocks, while the current works looks at ex exposure to risk.
5
A close existing work looking at plot-level labor allocation is by Akresh (2008) but with a fundamental difference
with the current work. Akresh looks at labor allocation across male and female plots following weather shocks (i.e.,
ex post to weather shocks), while the current study looks at labor arrangement, ex ante to weather variability.

11

are such that households must accelerate the execution of some specific agricultural tasks. However,
households in these regions, as in most Sub-Saharan Africa, face labor constraints (Moser and Barrett,
2003; Lee et al., 2006; Barrett et al., 2008; Smale et al., 2013) because they are not able to hire labor
either due to liquidity constraints or missing labor markets and must rely on their own members to satisfy
their labor requirement. Additionally, high risk of moral hazard in hired labor makes family labor more
favorable (Foster and Rosenweig, 1994; Bharadwaj, 2015)
Given that households composed of extended family members are usually large and
likely to be heterogeneous, by farming collectively, it is possible for them to overcome the adverse
impacts of weather instabilities. Particularly, they are likely to have a larger pool of labor, enabling them
to apply “brute strength” labor when weather fluctuations leave them with a constricted period to execute
some major agricultural activities, (Putterman, 1983; West 2010). In addition, heterogeneity of the
extended household members allows them to take advantage of “age-specific labor force”. Specifically,
they are able to mobilize workers with different endowments with younger members being more
physically productive in the field while elder members have experience in responding to weather
instabilities (Rosenzweig and Wolpin, 1985). These factors make the ground for the central claim
advanced in the current paper –that is, collective farming being perceived as an ex-ante risk mitigation
strategy.
The remainder of the study is organized as follows: Section 1.2 presents the background of the
study zone with a description of the study site and household structure as well as their agricultural
production arrangement. Section 1.3 provides the conceptual framework regarding collective farming as
an ex ante risk mitigation strategy. Description of the data source and sample restrictions are given in
section 1.4 while descriptive statistics are presented in section 1.5. Section 1.6 provides the empirical
approach. Regression results of labor allocation and crop yield performance across collective and
individual plots are presented in section 1.7. Section 1.8 provides empirical evidence of motives for
insurance against risk while section 1.9 concludes the paper with some policy implications.

12

1.2

Background

Mali is one of the world’s ten poorest countries with about 60 percent of its population living in rural
areas (World Bank, 2016) where agriculture is the main source of livelihood. The country is divided in
eight regions among which the Segou region, the study zone of the current work. The Segou region (see
red colored in Figure 1.1) is the second largest region of Mali (in terms of population). The agroecological zone of the region is the Sudano-Sahelian zone. The study covers four cercles in the Segou
region: Bla, San, Segou, and Tominian (see Figure 1.2). While the region is well known for its fishery and
pottery industries, the main economic activity in the rural area remains small-scale farming, where the
family constitutes the main sources for agricultural labor due to missing or incomplete labor markets in
these zones.

Figure 1.1 Map of Mali with the Segou region

Figure 1.2 The “Cercles” included in the study zone

13

The ‘traditional’ farming household structure, as most rural households in Mali, is commonly
composed of extended family members though there has been recently a rise in the formation of nuclear
family households and, according to Guirkinger and Platteau (2014), a more individualization of farm
management. The extended household is comprised of multiple nuclear households and is controlled by a
senior male (the patriarch) who also has his own nuclear household. The production system within the
extended household consists of collective plots controlled by the patriarch and individual plots controlled
by adult males and married females. The patriarch allocates land to the latter.
Collective plots, named “foroba” (which means “big field” in Bambara), are generally larger than
individual plots.6 The decision maker on those plots is usually called the “chef des travaux” (work team
leader). The work team leader tends to be the patriarch or one of the brothers or eldest sons of the
patriarch (Becker, 1996). Individual plot managers are required to share their labor between the collective
plots and their own plots. Part of the production from the collective plots may be sold for liquidity
availability for agricultural input purchase for the next agricultural season and for other expenses within
the extended household. The specific form of the distribution of the production to the nuclear households
for their efforts provided on the collective plots depends on the extended household configuration.
The extended households may fall in two of the four household classifications by Beaman and
Dillon (2012): (i) common agriculture, dwelling and authority or (ii) common agriculture, food, dwelling,
and authority. In the first case, the production is distributed to the nuclear households in the form of
grains. As for the second case, the production is kept in common granaries and extended household
members share collective meals (Beaman and Dillon, 2012).7 The sharing rule of the production in either
case is based on need, rather than performance (Guirkinger and Platteau, 2014), which is very salient
from the second case, where meals are taken collectively.
The existence of collective plots dates to the period prior to colonization, where members of the

The “big fields” are formally referred to as exploitation familiale in Mali
The other two classification by Beaman and Dillon (2012) are (i) common dwelling, authority, and common food
and (ii) common dwelling, authority only.
6
7

14

same lineage collectively farmed the same land allocated by the “land chief”, who was a respected senior
from the first family to settle in the area and had control over all the land in the village. The linear chief
oversaw farming activities on the collective fields and was usually in charge of the redistribution of the
production to individual households based on their needs (World Culture Encyclopedia, Mali). Men and
women also had access to individual plots to maintain their financial autonomy. A land reform in 1986
allowed access to private land, which was largely adopted in the urban areas. However, land tenure in the
rural area mostly remained collective (FAO, 2006). An important point to bear in mind is that formal
land titles are almost inexistent in the rural areas of Mali. Land is the property of the State, and land users
only have usufruct rights (USAID, 2010).

1.3

Conceptual framework

The study zone of the current work is characterized by drylands with limited access to irrigation, making
rain-fed crop cultivation the main source of livelihood for households in this region. Yet, rain-fed crop
production is highly sensitive to rainfall variability (Mishra et al., 2008; Cooper et al., 2008; and Sultan et
al., 2005), making these households vulnerable to production uncertainty and risk. It may therefore be
preferable for extended household members to secure the availability of sufficient labor on the collective
plots to ensure higher levels of production to protect amongst themselves against food shortfalls.
Apparently, the argument that collective farming serves as insurance against production
uncertainty due to weather variability by itself is uncertain. Given that variabilities in rainfall are
perceived as covariate shocks and, consequently, assumed to affect all households in a given region in a
same way, why would households dwelling in the same location use collective farming as ex ante risk
mitigation? While this question cannot be answered definitively, there are nevertheless several factors that
support the notion that collective farming can be viewed as an ex ante risk management strategy.
The risk of unpredictable labor bottlenecks during the agricultural season among nuclear
households may be an important factor. Labor remains a major productive resource among households in
sub-Saharan Africa as farmers tend to compensate for low (or even absent) non-labor input use by

15

engaging in labor intensive practices such as land preparation, building yield enhancing or yield
preserving technologies (manure, zai, stone-bunds etc.) and intensive weeding (de Ridder et al., 2004;
West, 2010), inter alia. Because these practices are labor-intensive, their execution requires households to
have access to a large pool of labor. Yet, it has been widely documented that labor availability remains a
binding constraint among these households, due to their limited ability to hire labor either because of
liquidity constraints or labor shortage at the pick of the agricultural season (Moser and Barrett, 2003;
Barrett et al., 2008; Jack, 2013).
Labor constraints become exacerbated, with dire implications for the household’s ability to secure
food availability, when the environment is characterized by high weather uncertainty. Sultan et al. (2005)
and Mishra et al. (2008) note that farmers in dryland areas tend to sow during early rains for richer soil
nutrients and to reduce weeding labor requirements. However, early sowing may be risky because resowing may be required when it rains fewer than 10 out of 20 days following sowing (Sultan et al.
(2005)). This weather fluctuation induces a large uncertainty in labor requirement of the nuclear
household over the course of the growing season (Adams et al., 1997). Conversely, its impact on labor
availability and the ability to mitigate its adverse impact among nuclear households may be
heterogeneous, given that these households are likely to differ within the same extended household in
terms of their demographic characteristics and farming experience.
To illustrate, consider an extended household where a father and his son are the heads of two
nuclear households.8 Obviously, the father is older and relatively less physically strong compared to the
son.9 Conversely, he (i.e., the father) due to his long-standing farming experience, is more likely to be
well versed in strategic planning in the wake of weather instabilities (Rosenzweig and Wolpin, 1985). As
a result, due to their heterogeneity in terms of their endowments, these two nuclear households present
different levels of exposure even though they share the same weather attributes and may complement one

8

Extended family households may also be formed by married brothers and their respective nuclear families
This difference in age is supported in the current sample, as the average age of the heads of the nuclear households
is about 42 while the corresponding figure for the heads of extended households is about 59 years
9

16

other to mitigate their vulnerability to risk .10 For example, Rosenzweig and Wolpin (1985) use district
level historical rainfall data in India to show that districts with higher rainfall variability tend to have a
higher concentration of intergenerational extended families.11 The authors find that in districts with
higher rainfall variability, these intergenerational extended families achieve higher profitability and
steadier gross returns compared to households with younger members because the former (i.e.,
intergenerational extended families) have elderly members who have specific knowledge allowing them
to deal with adverse weather conditions.
Clearly, information in terms of agricultural practices can be shared across extended family
members, which may undermine the claim that collective farming is favored for risk mitigation through
gains from the experience of the elderly. One of the important aspects to consider here though, is the
knowledge of farm-specificity, which is gradually instilled over the course of several years of farming the
same farm plot (Rosenzweig and Wolpin, 1985).12 For example, some specific areas on the same plot
may need more yield enhancing or water management techniques relative to others, depending on weather
conditions.13 Also farmers are more likely to learn from those of similar age group (Conley and Udry,
2010). An implication for this is that agricultural knowledge transfer occurs mostly across rather than
within households (Beaman and Dillon, 2014)14. This, combined with the farm-specificity knowledge

10

Clearly, nuclear households may use other ways of ex ante risk mitigation such as seasonal migration,
participation in non-agricultural income generating activities or off-farm labor supply. However, these strategies are
unlikely to override the practice of collective farming as ex ante risk mitigation because they are not mutually
exclusive with the later. Indeed, anecdotal evidence shows that many migrant household members in West Africa,
especially in Mali and Burkina Faso, are required to return in their villages at the start of the agricultural campaign
to supply labor on the family farm.
11
The authors find that in districts with higher rainfall variability, these intergenerational extended families achieve
higher profitability and steadier gross returns relative to households with younger members because the former (i.e.,
intergenerational extended families) have elderly members who have specific knowledge allowing them to deal with
adverse weather conditions.
12
Unfortunately, this claim of increasing farm-specificity knowledge due to several years of consecutive years of
farming cannot be directly tested here because of non-availability of data on how long the plot has been exploited.
However, I indirectly test this argument by looking at yield dispersions, over time, across collective plots –which are
likely to have been farmed over several years –and individual plots
13
For example, some area may be more prone to flooding in the wake of too much rain while other areas may be
more severely affected following insufficient rainfall.
14
The authors find that composting knowledge spillovers to women were more likely to occur across women from
different households than within households with opposite gender.

17

assumption provide grant for my claim that the benefit from the elderly’s experience for effectively
responding to weather variability is more likely to occur through collective farming.
Another advantage of collective farming is that, owing to the larger size of the extended
family households, they can access a large pool of labor (Putterman, 1983; West, 2010), allowing them to
apply ‘brute force’ labor when weather fluctuations leave them with only a small window of time to carry
out some specific activities. Agricultural activities must be carried out within a specific timeframe
depending on rainfall conditions, and failing do so can have significant negative consequence for crop
yield performance (Sultan et al., 2005; Mishra et al., 2008; Cooper et al., 2008; West, 2010) and,
consequently, threaten food security. As a result, this guaranteed access to a large pool of labor that
facilitates timely execution of the series of agricultural activities is an important feature that favors
collective farming, especially when labor bottleneck is likely to occur at the peak of the agricultural
season when no external labor is available for hire.
Also, collective farming allows extended family members to mobilize heterogeneous labor (male,
female, and child labor), which increases the effectiveness of the timely execution of farm activities. To
be more precise, access to heterogeneous labor permits delegation of different groups of laborers to
specific agricultural activities, where they are the most efficient, further allowing them to accelerate the
executions of agricultural tasks when needed. For example, while men are more efficient in activities that
require physical strength such as plowing, women and children tend to be more specialized in
transplanting and weeding (Boserup, 1970; Burton and White, 1984; Stone et al., 1995). As a result, such
a specialization facilitated by collective farming increases the ability of extended households living in
high rainfall variability areas to reduce exposure to potential crop failure.
These beneficial features of collective farming serve the basis for the central hypothesis in the
current essay: households who are subject to weather variability will favor collective farming as an ex
ante risk mitigation strategy. The ex-ante risk mitigation argument reflects the precaution of the nuclear
households against production uncertainty by being engaged in collective farming, prior to the realization
of the adverse event. This is different from the household’s responses to weather shocks, which would be

18

classified as ex poste risk copping strategies. For this reason, weather risk in the current dissertation is
defined following the same approach in Dillon et al. (2011) as the variability of rainfall (rather than
rainfall shocks), which captures the historical exposure of the household to production uncertainty.
For the first part of the empirical work, labor demand (expressed in person-days per hectare) on a
given plot in the household will be derived from the following reduced form:
𝐿 = 𝑓(𝑋, 𝐻, 𝑍, 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒, 𝜇),

(1)

where X, H, Z, are vectors of plot, household, and village characteristics, and Îź is a stochastic
error term. Collective indexes whether the plot is collectively farmed. Differences in labor allocation
across collective and individual plot will be captured by

𝜕𝐿
.
𝜕(𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒)

If labor is allocated equally across

collective and individual plots after controlling for all relevant plot characteristics, then this term should
be zero (Udry, 1996; Kazianga and Wahhaj, 2013). Labor allocation (in person-days per hectare) will be
disaggregated by male, female, and child labor, as labor productivity is likely to differ by type of labor.
For instance, with the same number of workers, amount of time, and land size, the productivity obtained
from children is likely to be lower relative to the productivity obtained from men. The same reduced form
will be used to investigate labor heterogeneity, in which case, 𝐿 is a binary outcome taking 1 if mixed
labor is applied on the plot labor (i.e., if men, women, and children work on the plot) and 0 otherwise.
To test the main research contribution of this paper–that is, the claim that collective farming
constitutes an ex-ante risk mitigation mechanism, an interaction term between rainfall variability and the
binary variable Collective is introduced in equation (1), leading to the following reduced form:
𝐿 = 𝑓(𝑋, 𝐻, 𝑍, 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒, [𝐶𝑉_𝑟𝑎𝑖𝑛] ∗ [𝐶𝑜𝑙𝑙e𝑐𝑡𝑖𝑣𝑒], 𝜇),

(2)

where CV rain, measures rainfall variability. Rainfall variability is captured by the coefficient of variation
of rainfall, measured as the historical standard deviation divided by the historical rainfall mean for each
village over a period of twenty-seven year (1981 to 2007). Rainfall variability in village v is defined as
𝐶𝑉𝑟𝑎𝑖𝑛 𝑣 =

𝑆𝐷𝑣
,
𝑥̅𝑣

1

where 𝑆𝐷𝑣 = √(𝑇 ∑𝑇𝑡=1(𝑥𝑡𝑣 − 𝑥̅𝑣 )2 ) is the historical standard deviation and

19

1
𝑇

𝑥̅𝑣 =( ∑𝑇𝑡=1 𝑥𝑡𝑣 ) is the historical rainfall mean, both computed over the period of 1981 to 2007 in village
v. For each of these years, rainfall variability is computed for crop cultivation season (May through
September).
The main proposition tested is that labor allocation (number of person-days per unit of land) is
higher on collective plots among households in sites with greater rainfall variability. Formally,
𝜕𝐿
𝜕([𝐶𝑉_𝑟𝑎𝑖𝑛]∗[𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒])

> 0. This would be consistent with the assumption that collective farming

constitutes an ex ante risk mitigation strategy, which makes households allocate more labor on these
plots. It is important to reiterate that the purpose of this paper is not to show that extended households
devote more labor on the collective plots following weather shocks. What this paper does is to show that
households living in places where weather fluctuations prevail, which constitute a major risk for
production shortfall, will favor this type of production organization. As a result, ex-ante, they put in place
a labor organization system that will systematically lead to higher labor allocation on the collective plots.

1.4

Data source and sample restrictions

1.4.1

Data source

The study uses the Saving for Change (SfC) project data15. The sampling was carried out in four
“cercles” (Bla, San, Segou, and Tominian) in the Segou region. While parts of the region benefit from
high access to irrigation –especially villages in the office du Niger and the cercle of Macina –the
agricultural production in the study zone is dominated by the cultivation of rainfed subsistence crops. The
main staple crops are millet and sorghum.
The data collection consisted of a first and second round surveys implemented in 2009 and 2012,
respectively. Each round covered about 6,000 households in 500 villages. The survey used two household

15

The primary objective of the SfC project was to evaluate a microfinance credit among rural women in the Segou
region. The sampling consisted of randomly selecting a female adult in the age range 18-60. A question was asked
whether the nuclear household of the selected woman shares agricultural production activities or consumption with
another household. If the answer was yes, the nuclear household of the selected woman was identified as being part
of a large household. Otherwise, the household was considered as a nuclear household.

20

classifications. The first is a nuclear household, defined as a person with his direct dependents. The
second classification is an extended household formed by extended family members where multiple
nuclear households live and carry out economic activities and food consumption together.
The survey used two questionnaires: one questionnaire for the nuclear household and another for
the extended household. Only the questionnaire for the nuclear household was administered to households
that are exclusively nuclear while both questionnaires were administered to those who cohabitate and
carry out economic activities with extended family members. In the current work, plots belonging to the
nuclear households are referred to as individual plots while those belonging to the extended household are
classified as collective plots.

1.4.2

Sample restrictions

The SfC project data cover three types of nuclear households. The first category is the group of nuclear
households who do not participate in any collective farming. These households constitute about 25
percent of the sample. The second type, which constitutes most the sample (about 60 percent), is formed
by nuclear households exclusively engaged in collective farming with extended family members. By
default, this group of households provides labor to the collective plots, as the region is characterized by
missing agricultural labor markets. The third group, making 15 percent of the sample consists of
households who participate in collective farming in addition to managing their individual plots.16
As the interest is to investigate differences in labor allocation and crop yields across
collective and individual plots for plots planted to the same crop and in the same year within the same
household, I restrict the sample to the third group of households –that is, those within which collective
and individual plots coexist. Moreover, to eliminate any potential gender specific differential effects, I
also exclude all female individual plots from the sample. Preliminary analysis of the full sample (Table

16

The corresponding figure in Guirkinger and Platteau (2014) is about 26 percent. This is certainly because their
study region (mainly, the Sikasso region) offers more irrigated agriculture such as the production of vegetable crops,
which are more predominant on individual plots.

21

1A.1 in APPENDIX A) indicates that up to 71 percent of individual plots are controlled by female plot
managers.17 However, collective plots tend to be controlled by the head of household (who is mostly a
male household member). Consequently, comparing labor allocation and crop yield across the two types
of plot management structure without eliminating female individual plots will likely lead to outcomes
being driven by gender-specific effects rather than by differences in plot management structure. Finally,
the sample was also restricted to keep crops that are planted across collective and individual plots.
After these sample restrictions, which have the advantage to increase the internal validity of the
study, the final sample retained for the study is about 15 percent of the full sample both at the plot and
household level.18 While this represents a small portion of the full sample, the comparison of the
descriptive statistics with this restricted sample (see section 1.5) to the ones in Table 1A.2 (APPENDIX
A) with the full sample of male plots reveals that the patterns of plot characteristics, crop choice, as well
as labor allocation across collective and individual plots are preserved in the final restricted sample. Table
1.1 summarizes which subsample is used in each of the empirical specifications developed in section 1.6.

Table 1.1 Empirical specifications and their sub-sample
Subsample

Empirical specification

Households within which collective and
individual plots coexist
Household heads who simultaneously manage
collective plots and their own individual plots

Labor allocation across collective and individual plots using
household-crop-year fixed effect
Labor allocation across collective and individual plots using
plot-manager fixed effect
Labor allocation across collective and individual plots using
household-crop-year fixed effect controlling for rainfall
variability
Labor allocation across collective and individual plots using
plot-manager fixed effect controlling for rainfall variability
Likelihood of nuclear households' participation in collective
farming
Yield dispersion across collective and individual plots
between the two survey rounds

Households within which collective and
individual plots coexist
Household heads who simultaneously manage
collective plots and their own individual plots
All categories of households
Households within which collective and
individual plots coexist

17

A robustness check of the main results with the full sample, with an indicator variable for plots controlled by
women, provide consistent coefficient estimates of the key variables. Results are available upon request.
18
Internal validity is increased by not only dropping female plots but also by dropping nuclear households without
individual plots. By default, nuclear households without male individual plots are likely to allocate all their labor to
the collective plots and failing to drop such households will overstate the labor gap across collective and individual
plots.

22

1.5

Descriptive statistics

The descriptive statistics on plot characteristics and crop choice at the plot level across individual and
collective plots are presented in Table 1.2. The results show that collective plots are significantly larger
than individual plots. Also, land tenure and crop choice significantly vary across the two types of plot
management structure. A higher proportion of collective plots tend to be grown to grain crops while the
opposite is observed for vegetable crops. The most predominant grain crops are millet and sorghum and
this is consistent with nationally representative data (USAID-Mali, 2008, World Bank, 2014). The
proportion of plots grown to peanut is also high for both collective and individual plots.
The statistics show that the use rates of both organic and inorganic fertilizer as well as
herbicide/pesticide are significantly high for collective plots relative to individual plots. Fertilizer (either
organic or inorganic) application is labor-intensive in terms of the time needed for its application,
increased weeding efforts, and more time needed for harvest (assuming fertilizer use increases yields).
Conversely, herbicide use significantly reduces weeding efforts. These relationships between technologies
and their differential labor requirements can have important implications for labor allocation across these
plots, which is further discussed in section 1. Note that despite recent findings on a sharp increase in the
use of herbicide in another region in Mali (Haggblade et al., 2017), only about 4 percent and 6 percent of
individual and collective plots, respectively, receive herbicide application in the current sample.
There are also significant mean differences in soil types across collective and individual plots.
However, as shown in Table 1A.3, these differences in soil types by plot management structure disappear
once statistics are computed conditional on crop choice, especially for grain crops. This suggests that the
observed differences in crop choice are driven by differences in soil quality. If soil characteristics are
already embodied in crop choice, the concern of omitted variable bias due to unobserved differences in
soil quality across collective and individual plots should be lessened once crop fixed effects are controlled
for.19 Furthermore, the proportion of plots with companion crops is significantly higher on the collective

19

Plot-level regressions (results not presented here) of soil type on the indicator variable collective controlling for
plot size and plot tenure show no impact of plot management structure on soil quality.

23

plots. Given that labor allocation is recorded at the plot-level for the main crop on the plot, this may
inflate the amount of labor spent on the production of the main crop. To ensure that the main results are
free from bias driven by this potential source of confounding factor, I check the robustness of the results
after excluding these plots from the sample. The results remain consistent with the initial findings.

Table 1.2 Plot characteristics, crop choice, and inputs across collective and male individual plots
Individual
Collective
plots (N=1,703)

plots (N=5,194)

Variables

Mean

SD

Mean

SD

P-value
of diff.

Plot size

0.89

0.94

2.26

2.68

0.00

Clay soil

0.31

0.46

0.35

0.48

0.03

Clay & sandy mix

0.38

0.49

0.34

0.47

0.01

Sandy soil

0.28

0.45

0.28

0.45

0.75

Gravel

0.03

0.18

0.03

0.18

0.83

Household owns plot

0.96

0.19

0.98

0.13

0.00

Main source of water is rain

0.98

0.12

0.98

0.15

0.11

Millet

0.15

0.35

0.28

0.45

0.00

Sorghum

0.1

0.31

0.2

0.4

0.00

Maize

0.06

0.24

0.1

0.31

0.00

Rice

0.05

0.22

0.09

0.28

0.00

Peanut

0.47

0.5

0.11

0.32

0.00

Beans

0.03

0.17

0.03

0.16

0.73

Chickpeas

0.02

0.15

0.04

0.21

0.00

Fonio

0.04

0.19

0.09

0.29

0.00

Sesame

0.09

0.28

0.05

0.22

0.00

Plot has companion crop(s)

0.22

0.41

0.31

0.46

0.00

Inorganic fertilizer (1=yes)

0.11

0.32

0.29

0.45

0.00

Organic fertilizer (1=yes)

0.26

0.44

0.31

0.46

0.00

Soil type

Crop choice

Herbicide/pesticide (1=yes)
0.04
0.189
0.06
0.23
0.00
Plot level observations. Sample restricted to households within which collective and male individual plots coexist.
Top and bottom 1 percent of plot size and outcome variables dropped from sample

The significant differences in means in plot size across collective and individual plots raises the
concern of whether there is any overlapping between the two types of plot management structure. Having

24

an overlapping is important for linear controls to appropriately account for differences in plot size across
collective and individual plots.
Figure 1.3 presents plot size distribution by plot management structure. As expected, there is high
density of individual plots with small plot area, while the opposite is observed for collective plots. The
distributions presented in Figure 1.4 reveal that collective and individual plots tend to overlap mostly for
plot size smaller than 9 and 4 hectares, respectively for grain and non-grain crops. The figure also reveals
that only a very few number of observations do not overlap across collective and individual. Yet, there is
still a concern that differences in yield and labor allocation by plot management structure could be driven
by the larger plot size of collective plots that do not overlap with individual plots. To ease this concern, I
ran additional regressions by restricting the sample to plot size not greater than 9 and 4 for grain and nongrain crop plots, respectively.

Figure 1.3 Plot size (ha) distribution across collective and individual plots (all crops)

Figure 1.3 presents plot size distribution by plot management structure. As expected, there is high
density of individual plots with small plot area, while the opposite is observed for collective plots. The
distributions presented in Figure 1.4 reveal that collective and individual plots tend to overlap mostly for
plot size smaller than 9 and 4 hectares, respectively, for grain and non-grain crops. The figure also reveals
that only a very few number of observations do not overlap across collective and individual. Yet, there is

25

still a concern that differences in yield and labor allocation by plot management structure could be driven
by the larger plot size of collective plots that do not overlap with individual plots. To ease this concern, I
ran additional regressions by restricting the sample to plot size not greater than 9 and 4 for grain and nongrain crop plots, respectively.

Figure 1.4 Plot size (ha) distribution across collective and individual plots by grain and non-grain crops

26

The dependent variables are presented in Table 1.3. Labor variables in terms of heterogeneity of
laborers (male, female, and child) are binary outcomes. Labor intensity (total labor, male, female, and
child labor) are all expressed in number of person-days per hectare while yield is expressed in monetary
value (in 1,000 CFA per hectare).20

Table 1.3 Labor allocation across collective and male individual plots
Individual

Collective

plots (N=1,703)

plots (N=5,194)

Variables

Mean

SD

Mean

SD

P-value of
diff.

Labor heterogeneity
Plot receive male and female labor=1

0.81

0.39

0.87

0.33

0.00

Plot receive male and child labor=1

0.67

0.47

0.88

0.33

0.00

Plot receive all 3 sources of labor=1

0.57

0.50

0.78

0.41

0.00

Labor intensity
Total labor (person-days/ha)

107.62

97.77

105.79

84.49

0.59

Male labor (person-days/ha)

55.65

54.91

51.74

42.69

0.03

Female labor (person-days/ha)

27.25

35.38

25.53

28.43

0.13

Child labor (person-days/ha)

24.87

36.15

28.55

32.22

0.01

Yield (1,000 FCFA per ha)
265.90
246.57 189.90
182.50 0.00
Plot level observations. Sample restricted to households within which collective and male individual plots coexist.
Top and bottom 1 percent of continuous variables dropped from sample.

Figure 1.5 Labor heterogeneity across collective and individual plots (All crops)

20

CFA: CommunautÊ Financière Africaine. The exchange rate at the time of the surveys was approximately
$1US=500 CFA francs. Child labor is labor allocation by members who are between 5 and 17-year-old.

27

The statistics on labor heterogeneity, which are graphically illustrated in Figure 1.5, show that
collective plots tend to receive more heterogeneous labor (i.e., receive more labor from men, women, and
children) compared to individual plots. Note that all plots receive male labor since the sample has been
restricted to male controlled plots only (see sample restrictions in section 1.4). The illustrations by grain
and non-grain crops in Figure 1.6 (statistics not included in Table 1.3) show that the differential
heterogeneity of labor across collective and individual plots is much higher for grain crops, which
constitute the main staple food for these households. This hints that collective farming primarily favors
the production of these crops, which would be consistent with the claim that households engage in this
practice to guarantee their food security.
Conversely, individual plots tend to be more intensively farmed with male and child labor relative
to collective plots, though there is no statistical significant difference in total labor allocation. Also,
individual plots achieve higher yields than collective plots. This indication of higher labor allocation and
crop yields on the individual plots is certainly because these statistics are mean values and do not take
important factors such as crop choice and plot characteristics into account.

Figure 1.6 Labor heterogeneity across collective and individual plots (By cereal and non-cereal crops)

28

1.6

Empirical model

In this section, I specify multiple econometric models to test the two main reduced forms developed in
section 1.3. The first set of specifications, which corresponds to the reduced form in equation (1) is to
investigate labor allocation across collective plots and individual plots. Crop yield performance is also
estimated by plot management structure in this set of specifications. Second, to show that collective
farming serves as an ex-ante risk mitigation strategy, I introduce historical rainfall variability in the
specifications, which corresponds to the reduced form in equation (2) to investigate how the gap in both
labor allocation and crop yield between collective and individual plots varies across sites with different
levels of rainfall variability. Finally, a linear probability model is developed to test the likelihood of
participation in collective farming by nuclear households depending on the historical rainfall variability.
The main identification strategy applied in the specifications for labor allocation and crop yield
performance is the household-crop-year fixed effect approach (HHCYFE), which is constructed as a
unique identifier for each household by combining the household identification number, crop code, and
the year indicator. The motivation for the use of the HHCYFE is that several unobservable and
unobserved factors other than weather variability may influence the household's labor allocation decision
across these plots. The use of the Household-Year-Crop Fixed Effects (which keeps much of these factors
constant within households) presents a great potential to free the estimates of the key variables from
plausible omitted bias (Udry, 1996).

1.6.1

Specifications with household-year-crop fixed effect

The empirical specification for estimating labor allocation across collective and individual plots follows
the same strategy originally applied by Udry (1996), which was motivated by the collective model of the
household developed by Browning and Chiappori (1998). Udry (1996) argues that if the same production

29

technology is used on plots within the same household, then conditional on plot characteristics and crop
choice, the amount of labor allocated across these plots should not be different.21
Labor demand across collective and individual plots within the same household in the same year,
and for the same crop is defined as:
𝐷ℎ𝑘𝑗𝑡 = 𝑋ℎ𝑘𝑗𝑡 𝛽 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 + 𝛿ℎ𝑘𝑡 + 𝜖ℎ𝑘𝑗𝑡

(3)

where 𝐷ℎ𝑘𝑗𝑡 is specified as either a binary indicator for labor heterogeneity or a continuous variable for
the intensity of labor expressed in person-days per hectare on plot j of household h planted to crop k in
year t. There are three separate specifications when 𝐷ℎ𝑘𝑗𝑡 is a binary variable and are defined as follows:
𝐷ℎ𝑘𝑗𝑡 =1 if plot receives a mix of male, female, and child labor and 0 otherwise
𝐷ℎ𝑘𝑗𝑡 =1 if plot receives a mix of male and female labor and 0 otherwise
𝐷ℎ𝑘𝑗𝑡 =1 if plot receives a mix of male, and child labor and 0 otherwise.22
Recall that all plots receive male labor because the sample has been restricted to male plots (see section
1.4 for sample restrictions).
Four estimations are carried out for labor intensity (total, male, female, and child labor) all
expressed in logs of person-days per hectare.23 The disaggregation of labor by labor source is important in
the sense that labor quality differs between men, women, and children. For example, if there are several
children working on a plot, the intensity of labor allocation for that plot would be high, though it may not
be as productive as a corresponding smaller figure provided by men on another plot of similar
characteristics.
𝑋ℎ𝑘𝑗𝑡 is a set of the observable characteristics of plot j, 𝛿ℎ𝑘𝑡 is a household-crop-year fixed
effect; and 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 is an indicator variable for collective plots (1 if a collective plot and 0

21

See Udry (1996) for a formal description of the theoretical model
As described later in this section, the sample is restricted to male plots. Consequently, all plots receive male labor.
So, the case where plots receives a mix of female and child labor only does not apply.
23
An alternative specification to the logarithmic function is the Inverse Hyperbolic Sine transformation (ISH). The
advantage of this functional form is that it does not involve arbitrarily attributing non-zero positive values to
observations with missing values. Therefore, as a further robustness check of the main findings, I apply the IHS
transformation to the outcome variables. The results, available upon request, are similar to the main findings.
22

30

otherwise). Plot characteristics include plot size to control for economies of scales. I also control for soil
type, as labor productivity may be correlated with soil quality and, therefore, may affect the household
labor allocation decision (Fafchamps, 1993; Lamb, 2003; Dillon and Barrett, 2017). Plot tenure (owned
by household or rented) is also controlled for given the large literature on the correlation between plot
tenure and input investment (Hayes, 1997; Gebremedhin and Swinton, 2003; Goldstein and Udry, 2008;
Quisumbing, and Pandolfelli, 2010; Teklewold et al., 2013).24 Input use (mainly organic and inorganic
fertilizer and herbicides) are also included as controls.25 A positive sign of 𝛾̂ would be evidence that
collective plots receive intensively more labor (or are more likely to receive a mix of labor for the case
when 𝐷ℎ𝑘𝑗𝑡 is a binary indicator) than individual plots.
In principle, labor demand would include household characteristics such as the manpower of the
household proxied by the number of adults in the household (Reardon, 1997; Barrett et al., 2004,
Masakure and Henson, 2005; Peterman et at., 2014), access to markets, land prices and wages (Benjamin,
1992, Key et al., 2000, Taylor and Adelman, 2003; Barrett, 2008, Dillon and Barrett, 2017) in order to
account for differences in shadow prices of production factors across plots within the same household as
well as across households. These factors are assumed constant within households for a given year and
given that the empirical strategy considers plots within the same household and controls for householdcrop-year fixed effects, such factors are indirectly accounted for (Udry, 1996).
Note that all plots controlled by female-plot managers are dropped from the sample to eliminate
any gender-differential bias. Furthermore, since the empirical approach investigates labor use on crops
simultaneously grown across collective and individual plots within the same household and in the same
year, the sample is restricted to households that participate in collective farming in addition to managing

24

Plots managed under sharecropping are dropped from the sample as they only represent about 0.3 percent of the
sample.
25
The inclusion of input use controls for differences in the use rates of these inputs across collective and individual
plots. Yet, their inclusion also raises a concern of a reverse causality between the decision to apply these inputs and
labor endowment. To alleviate this concern of potential endogeneity, I run the regressions after dropping the sample
of plots where these inputs are applied. The results, presented in Table 1B.1 and Table 1B.2, remain consistent with
the main findings, both for labor allocation and crop yield.

31

their own male individual plots (see section 1.3 for more details on the categories of households in the
sample).
In addition to investigating labor allocation, I also estimate yield differences to examine whether
gaps in labor allocation will be reflected on yields. Assuming the same production technology across
collective and individual plots, the specification for crop yield is defined as
𝑄ℎ𝑘𝑗𝑡 = 𝑋ℎ𝑘𝑗𝑡 𝛽 + 𝐼ℎ𝑘𝑗𝑡 𝜑 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 + 𝛿ℎ𝑘𝑡 + 𝜖ℎ𝑘𝑗𝑡

(4)

𝑄ℎ𝑘𝑗𝑡 is the log of yield on plot j planted to crop k in year t by household h.
𝐼ℎ𝑘𝑗𝑡 𝜑 is a vector of inputs, mainly organic and inorganic fertilizer and herbicide.26
If crop yield does not vary across collective and individual plots within the household, then Îł = 0;
that is, yield on a plot should not depend on plot management structure (i.e., whether collectively or
individually managed) after controlling for relevant factors that may influence yield.
A shortcoming in the current empirical specification is potential selection bias in plot
management structure that may influence the household’s decision in its productive resource allocation
and crop yield performance across collective and individual plots.27 The same issue of omitted variable
bias is also present in previous work striving to investigate intrahousehold resource allocation across
different plots within the same household (Kazianga and Wahhaj, 2013; Udry, 1996; Goldstein and Udry,
2008 among others).
This concern of potential selection bias of plot management structure is intensified in the current
study because of lack of information on distance to plot location and plot topography. Plots located closer
to the dwelling of the household may receive more care due to easy access. Conversely, households may
work more diligently on plots located at a greater distance to compensate for travel time (Kazianga and
Wahhaj, 2013). With respect to plot topography, it has been shown that plots with uppermost

26

The common technology assumption is motivated by the fact that production technology may vary across crops,
but not across individual and collective plots as long as the same crop is planted across these plots within the same
household and in the same agricultural season (Udry, 1996).
27
Ideally, one would control for plot fixed effects. Unfortunately, such approach is not feasible because plots are not
tracked over time.

32

toposequence are likely to be less productive relative to plots with nearest bottom toposequence (Steiner,
1998; Shorr, 2000; Guirkinger et al., 2015). If both plot location and plot topography are systematically
correlated with plot management structure, failing to include them in the regressions will lead to biased
estimates of 𝛾.
However, existing empirical evidence suggests that farmers tend to make their crop choice
contingent on plot characteristics such as soil type, plot location, and plot topography (Udry, 1996;
Steiner, 1998; Shorr, 2000; Guirkinger et al., 2015). Though I lack data on plot location and plot
topography, I show in Table 1A.3 (APPENDIX A) that conditional on crop choice, all available soil
characteristics are fairly balanced across collective and individual plots, especially for grain crops. And,
since I control for household-crop-year fixed effect in my regressions, the estimates of 𝛾 will condition on
the main factors that are highly related to unobserved plot characteristics and therefore I see the estimates
of 𝛾 as being largely free of these omitted variables bias.
There is still a concern of systematic differences in plot characteristics across collective and
individual plots, especially for non-grain crops, based on the statistics in Table 1A.3 (in APPENDIX A).
To account for these significant differences, separate regressions are run for all crops, grain crops, and
non-grain crops. More importantly, three alternative econometric models (Models 1, 2, and 3) are
specified by gradually adding plot characteristics for some sensitivity analysis. Specifically, Model 1
includes the indicator Collective and plot size only (in addition to controlling for household-year-crop
fixed effects), Model 2 adds input use and plot ownership, while Model 3 adds soil type. The motivation
behind these different specifications is to observe how sensitive yield differences across collective and
individual plots are to the inclusion and omission of available plot characteristics. This allows to
indirectly assess how serious the issue of the omitted unobservable plot characteristics is likely to be.28
Fortunately, the results of the sensitivity analysis do not suggest any correlation between unobservable
plot characteristics and crop yield. Nevertheless, this concern of lacking comprehensive plot physical

28

The same strategy is used by Udry (1996) and Kazianga and Wahhaj (2013).

33

characteristics must remain a caveat in the current study.

1.6.2

Specifications with plot-manager-year fixed effect

The empirical specifications presented so far do not account for plot manager fixed effects because not all
individuals concurrently manage individual and collective plots. Yet, plot manager characteristics may
influence access to inputs and, consequently, crop yield performance (Ondersteijn et al., 2003; Foltz,
2004; Qaim et al., 2006; Barrett et al., 2008), as well as plot management structure (collective versus
individual) such that the coefficient estimates on Collective may suffer from bias due to omitted plot
manager effects. To address this issue, I impose another sample restriction and run additional econometric
specifications.
Specifically, I restrict the sample to individuals (namely the heads of extended households) who
control their own individual plots and collective plots at the same time. Using this restricted sample, I
develop another specification where the household-year-crop fixed effect is replaced with an individualyear fixed effect. Such a specification with a sample restricted to individuals simultaneously controlling
their own individual plots and collective plots constitutes a powerful way to demonstrate the robustness of
the main findings. Given that collective plots tend to be exclusively managed by the head of household,
who is inherently different from the other household members, being able to control for the plot
manager’s unobserved characteristics removes concerns related to possible correlation between plot
manager characteristics and access to labor as well as crop yield performance.
Note that, including individual-year fixed effects means that I also control for individual
characteristics that vary year to year (i.e. are time variant in addition to time invariant). This controls for
anything that is unobserved and specific to an individual in a given year (e.g. changes in access to credit
and input markets or productivity that might change from year to year within an individual). As a result,
the specifications using individual-year fixed effects identifies the gap between collective and individual
plots of off comparisons across these types of plots within an individual in each year.
Limiting the sample to the heads that simultaneously control their own individual plots and
34

collective plots has another advantage beyond relieving the concern of omitted variable bias due to
differences in the identity of the plot manager. As I highlight above, one of the concerns about potential
selection bias of plot assignment as individual versus collective plots is that the head of household may
systematically allocate less productive plots to junior members for individual farming. This is especially
of a concern in the current work as plot characteristics display significant differences for non-grain crops
across collective and individual plots. However, given that the head controls land allocation, it is unlikely
that the later will allocate land of poor quality to himself for his own individual farming and, therefore, it
is reasonable to expect that the head will choose a portion of land that is at least as productive as the
portion of land used for collective farming. As a result, estimates of the coefficients on collective
indicator are less likely to suffer from upward omitted variable bias in the restricted sample.
Lastly, investigating labor allocation and crop yield performance across collective and individual
plots controlled by the same person provides a powerful way to explore whether it is plot management
structure (i.e., collective versus individual) or plot manager identity that matters the most. To be more
precise, if the identity of the plot managers across collective and individual plots matters, then there
should be no differences in labor allocation and crop yields after restricting the sample to individual
simultaneously controlling the two types of plots.
The empirical strategy using the sample of heads is specified in equations (5) and (6) for input
allocation and crop yield, respectively.
𝐷ℎ𝑗𝑡 = 𝑋ℎ𝑗𝑡 𝛽 + 𝐾𝑗 𝜔 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑗 + 𝛿𝑖ℎ𝑡 +𝜖𝑖ℎ𝑗𝑡

(5)

𝑄ℎ𝑗𝑡 = 𝑋ℎ𝑗𝑡 𝛽 + 𝐾𝑗 𝜔 + 𝐼ℎ𝑘𝑗𝑡 𝜑 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑗 + 𝛿𝑖ℎ𝑡 +𝜖𝑖ℎ𝑗𝑡

(6)

Precisely, 𝐷ℎ𝑗𝑡 is the intensity of labor (total, male, female, and child labor) expressed in log person-days
per hectare on plot j managed by the head of household h in year t. 𝑄ℎ𝑗𝑡 is crop yield performance
expressed in log of production in kilogram per hectare. 𝛿𝑖ℎ𝑡 is the plot manager year-fixed effect. Because
the specifications in equations (5) and (6) include plot manager-year fixed effects rather than householdyear-crop fixed effect, the additional term 𝐾𝑗 , which the crop fixed effect, is included to control for the

35

household crop choice portfolio.29
In this subsection, I used sample restrictions and alternative econometric specifications to check
the robustness of the estimated differences in labor allocation and crop yield performance across
collective and individual plots. As a result, even though there may be a selection bias on plot
unobservable characteristics across these two types of plot management structure, it is unlikely that taking
such factors into account would substantially alter the main results presented in this paper.

1.6.3

Adding weather variability to the household-crop-year fixed effect

In section 1.3, I claim that collective farming may constitute an ex ante risk insurance mechanism among
extended family households. To empirically investigate this claim, I introduce rainfall variability
represented by 𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 in equations 7 and 8. Rainfall variability is captured by the coefficient of
variation (CV) of rainfall, measured following Dillon et al., (2011) by taking the historical standard
deviation divided by the historical rainfall mean for each village over a period of twenty-seven years.30
Formally, I estimate the following equations
𝐷ℎ𝑘𝑗𝑡 = 𝑋ℎ𝑘𝑗𝑡 𝛽 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 + 𝜅(𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 ∗ 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 )
+𝛿ℎ𝑘𝑡 + 𝜖ℎ𝑘𝑗𝑡

(7)

and
𝑄ℎ𝑘𝑗𝑡 = 𝑋ℎ𝑘𝑗𝑡 𝛽 + 𝐼ℎ𝑘𝑗𝑡 𝜑 + 𝛾𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 + 𝜅(𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 ∗ 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒ℎ𝑘𝑗 ) + 𝛿ℎ𝑘𝑡 + 𝜖ℎ𝑘𝑗𝑡 (8)
for labor demand and crop yield, respectively, where standard errors are clustered at the village level.
If collective farming is driven by motives for insurance against production risk, then households

29

Ideally, equations (5) and (6) would control plot-manager-year-crop fixed effects rather than plot-manager-year
fixed effects, which would provide a stronger identification, as they would allow estimate the gap between collective
and individual plots for crops grown simultaneously across these plots by the same individual in the same year.
However, such identification strategy would impose a strong restriction on the data, leading to very low variation
across households.
30

𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 =

𝑆𝐷𝑣
𝑥̅𝑣

1

1

𝑇

𝑇

, 𝑤ℎ𝑒𝑟𝑒 𝑆𝐷𝑣 = √( ∑𝑇𝑡=1(𝑥𝑡𝑣 − 𝑥̅𝑣 )2 ) is the historical standard deviation and 𝑥̅𝑣 =( ∑𝑇𝑡=1 𝑥𝑡𝑣 )

is the historical rainfall mean, both computed over the period of 1981 to 2007 in village v.

36

living in sites with high weather variability will pool more labor on collective plots to secure food
availability. As a result, the prediction is that 𝜅 > 0. Furthermore, if collective farming is used as an ex
ante risk management strategy, households in sites with higher weather variability should be more likely
to be engaged in collective farming. This assertion is tested through equation 9, defined as follows:
𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒_𝑓𝑎𝑟𝑚𝑖𝑛𝑔ℎ𝑣 = 𝜏𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 + 𝐺𝑣 𝜉 + 𝑋ℎ𝑣 𝜔 + 𝜖ℎ𝑣

(9)

where 𝐶𝑜𝑙𝑙𝑒𝑐𝑡𝑖𝑣𝑒_𝑓𝑎𝑟𝑚𝑖𝑛𝑔ℎ𝑣 is a binary variable that takes 1 if household h is engaged in
collective farming with extended household members in village v. The sample used here includes all the
three types of households described in section 1.4. The full sample of household is used since the
objective here is to estimate the probability of a household being engaged in collective farming. Because
this estimation is carried out at the household level, the sample of plots is collapsed at the household level
such that households where at least one collective plot exists will be considered as participating in
collective farming.
Though the full sample is used, all female plots are dropped from the sample before putting the
data at the household level. This is mainly due to two reasons. First, female plots are not included in any
of the estimations in the previous sections due to concerns related to gender differential bias. Second and,
most importantly, land distribution to female household members follows different norms than land
allocation to male household members. Land is distributed to male through bloodline whereas women
mostly obtain land through marriage. A nuclear household not participating in collective farming
primarily obtains most its livelihood from the production from male plots, rather than female plots.
Proceeds from women’s plots are for their financial independence and for sometimes cooking small meals
for their own children (Toulmin, 1986).
𝐶𝑉_𝑟𝑎𝑖𝑛𝑣 in equation (9) is defined as above, 𝑋ℎ𝑣 and 𝐺𝑣 are vectors of household and village
characteristics likely to be correlated with the household decision to participate in collective farming, and
𝜖ℎ𝑣 captures other exogenous shocks that affect the probability of household h being engaged in
collective farming in village v. Household level characteristics are classified in terms of household
demographics and asset holding. Demographics are represented by household size and the age of the head
37

of the nuclear household. The underpinning reason for the inclusion of the head’s age is to indirectly
assess the assumption that participation in collective farming allows younger plot managers to benefit
from the experience of the elderly members. As a result, since experience is likely to increase with age,
the age of the nuclear household head is expected to be negatively correlated with the likelihood of
participation in collective farming.
Constraints to land access partly due to missing land markets have been found to be positively
correlated with collective farming among extended family members (Rosenzweig and Wolpin, 1985;
Kazianga and Wahhaj, 2017). This is explored by including total land holding in equation 9. Furthermore,
motives for economies of scale, especially for investment in indivisible agricultural equipment (Carter,
1987; Mathijs and Swinnen, 2001), may also favor collective farming. Consequently, draft animal
ownership by the nuclear family household, which requires high financial capacity, is also included in
equation (9) as asset holding to assess the robustness of the coefficient estimates of 𝜏.
Village level covariates include the total number of households in the village, availability of
formal credit in the village and whether the village experienced outmigration over the last ten years. Since
these covariates are likely to be correlated with rainfall variability, estimations will be carried out with
and without these variables to verify the consistency of the estimates of the key coefficient, which is 𝜏.
The prediction is that 𝜏 is expected to have a positive sign (i.e., 𝜏 > 0). In other words, households in
villages with high rainfall variability will be more likely to be engaged in collective farming. 31 Note that
the full sample of household is used in this specification since the objective here is to estimate the
probability of a household being engaged in collective farming.
Despite controlling for important village covariates that are likely to be both correlated with
rainfall variability and the household decision to participate in collective farming, I must emphasize that
this specification is only descriptive and thus does not establish a framework to determine a causal

31

Recall that the type of risk referred to here is ex ante risk, rather than ex post risk. As a result, even though
households in the same area share the same weather characteristics, collective farming may serve as an ex ante
insurance mechanism among households who anticipate to face production risk due to weather uncertainty.

38

correlation between rainfall variability and the household decision to participate in collective farming.
Nevertheless, the concern of endogeneity between the nuclear household’s decision to engage in
collective farming and village level characteristics is reduced by the inclusion of important village
covariates in equation (9). Additionally, the statistics presented in Table 1A.4 in APPENDIX A show
very little correlation between rainfall variability and the available village covariates. This hints that the
results from the specification in equation (9) are unlikely to be strongly altered by other sources of village
level confounding factors. Therefore, though only descriptive, findings from the current work would
provide suggestive evidence that higher rainfall variability could lead households to engage more in
collective farming as an insurance instrument.

1.7

Results: Labor allocation and crop yield

1.7.1

Results with household-crop-year fixed effects

The summary statistics in Table 1.2 show a large difference in plot size, crop choice, and plot tenure
across collective and individual plots. In addition, the rates of organic and inorganic fertilizer as well as
herbicide/insecticide, whose application has different implication for labor requirement, are significantly
higher for collective plots. These significant differences in means may constitute possible sources of
variation in labor allocation across the two types of plot management structure. Thus, these variables are
included in the labor allocation estimation. While including these inputs controls for differences in their
use rates across collective and individual plots, their inclusion also raises a concern of a reverse causality
between labor endowment and the decision to use them. To check the robustness of the results to this
potential endogeneity, I run the regressions after dropping the sample of plots on which these inputs are
applied. The results, presented in Table 1B.1 and Table 1B.2 (see APPENDIX B) obtained with the
corresponding sample remain consistent with the main findings, both for labor allocation and crop yield.
Other significant sources of variations in the household’s labor allocation decision are market
conditions such as input prices, wages, and access to credit. While these factors are not explicitly included
in the regressions, they are taken into account through the household-year-crop-fixed effects, the same

39

approach applied in existing similar empirical work (e.g., Udry, 1996; Kazianga and Wahhaj, 2013).
Given the identification strategy, the sample is restricted to nuclear households that manage their own
male individual plots concurrently with participating in collective farming. Furthermore, due to the
gender differential bias in both input access and crop yield performance (see review by Quisumbing,
1996), which is corroborated in this study (see Tables 1C.1 and 1C.2 in APPENDIX C), all female plots
are excluded from the sample to eliminate any gender differential bias.32
The results with respect to labor heterogeneity, where the outcome variables are expressed in
binary terms, are presented in Appendix D, Tables 1D.1 and 1D.2. The estimation results for the intensity
of labor allocation, expressed in the log of the total labor on the plot, are presented in Table 1.4. Since
labor quality differs across men, women, and children, the results are provided for total labor as well as
for male, female and child labor. Also, results are presented with the key variable only –that is Collective
–before including the other covariates.

32

A robustness check of the findings with the full sample (i.e., including female plots) presented in APPENDIX C
provides similar with stronger coefficient estimates of the key variable (i.e., the indicator variable Collective)

40

Table 1.4 Estimation of log of labor allocation (person-days/ha) across collective and male individual plots
Model 1: Without plot covariates
Model 2: With plot covariates
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Log(Total
Total(Male
Log(Female
Log(Child
Log(Total
Total(Male
Log(Female
VARIABLES
Labor/ha)
labor/ha)
labor/ha)
labor/ha)
Labor/ha)
labor/ha)
labor/ha)
Collective=1

0.163***
(0.034)

0.384***
(0.038)

0.082**
(0.057)

0.303***
(0.069)

Log of plot size
Organic fertilizer=1
Inorganic fertilizer=1
Herbicide / insecticide=1
Household owns plot=1
Clay & sandy soil (clay is
omitted)

0.384***
(0.039)
-0.248***
(0.023)
0.095**
(0.037)
0.02
(0.043)
-0.183**
(0.093)
0.126
(0.141)

0.303***
(0.043)
-0.246***
(0.026)
0.117***
(0.040)
0.009
(0.046)
-0.169
(0.104)
0.08
(0.141)

0.480***
(0.063)
-0.171***
(0.041)
0.015
(0.059)
0.113
(0.074)
-0.20
(0.182)
-0.148
(0.225)

(8)
Log(Child
labor/ha)
0.688***
(0.081)
-0.127***
(0.045)
0.052
(0.063)
0.037
(0.075)
-0.358**
(0.169)
0.489**
(0.224)

0.08
0.052
-0.001
0.204**
(0.053)
(0.053)
(0.098)
(0.104)
Sandy
-0.062
-0.021
-0.119
-0.04
(0.050)
(0.054)
(0.082)
(0.088)
Gravel
0.034
-0.029
-0.062
0.298
(0.116)
(0.109)
(0.178)
(0.187)
HHCYFE
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
6,897
6,897
6,897
6,897
6,897
6,897
6,897
6,897
R-squared
0.0189
0.0036
0.0301
0.0983
0.1046
0.0773
0.0459
0.1126
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to households
within which collective and male individual plots coexist. All specifications include household-year-crop fixed effects. Top and bottom 1 percent of plot size and
outcome variables dropped from sample

41

There is a strong indication that collective plots are more intensively farmed than individual plots.
This is shown by the positive and significant coefficient estimate of 𝛾 for not only total labor but also for
any type of labor. Specifically, 𝛾̂ = 0.38 in column 8 indicates that total labor allocation per hectare on
collective plots is about 47 percent higher than the corresponding amount allocated to individual plots.
Disaggregating labor allocation by source reveals that labor gap is much higher for female and child
labor.
Other noteworthy findings in Table 1.4 are the opposite signs of the coefficient estimates on
fertilizer and herbicide use. Specifically, the results indicate that fertilizer use (mainly organic fertilizer) is
associated with more labor requirement while herbicide use, which reduces weeding time, significantly
lowers the amount of labor allocated to the plot. These findings are strongly supported by the literature on
the adoption of these technologies (Gianessi and Williams, 2011; Gianess, 2013).
The analysis by labor source shows that the higher labor requirement associated with organic
fertilizer is significant for male labor. This support the view that adoption of organic fertilizer requires
physical strength (production and transportation), making men more suitable for the execution of these
tasks (Feng et al., 2010). In contrast to organic fertilizer, the major source of labor affected by herbicide
use is child labor. This finding corroborates the claim that weeding activities are mostly performed by
children.33
The higher labor amount applied on collective plots does not necessarily indicate that more effort
is being applied on these plots relative to individual plots. As reported in the work of Guirkinger and
Platteau (2014), individuals may be spending more time on the collective plots for only token
contributions in order to fulfill collective labor participation requirements, and save their energy to work
later on their individual plots. If the observed higher labor applied on the collective plots is real effort
provision, then collective plots should achieve higher yields relative to individual plots.
Table 1.5 presents the estimation results of the monetary value of yield (from equation 4)

33

Children are household members who are 5 to 17-year-old.

42

expressed in log terms.34 As in the labor allocation regressions, yield is estimated controlling for
household-year-crop-fixed effects. Note that Table 1.5 contains fewer numbers of observations relative to
Table 1.4 as observations for yield variables are at the crop-level. Two important points must be
highlighted here. The first is the presentation of the results for different samples of crops (Panel A for all
crops, Panel B for grain crops, and Panel C for non-grain crops). The second point to be noted is the use
of three different econometric models (Models 1, 2, and 3) in each of the three panels. Model 1 presents
the results where I only control for the indicator Collective and plot size expressed in log terms. Given
that yield increases with the use of both organic and inorganic fertilizer as well as herbicide use, and
because descriptive statistics show significant differences in the application rates of these inputs, I control
for the use of such inputs in Model 2 besides controlling for plot tenure.35 Model 3 includes soil type,
mainly whether the soil type on the plot is sandy, clay, clay-sandy mixed, or gravel (the omitted soil type
being sandy soil).

34

Village level prices, recorded through community questionnaires at the survey times, were used to convert
production into monetary value
35
As noted earlier, the inclusion of these inputs raises a concern of endogeneity. The robustness of the results to this
potential reverse causality is checked by running the regressions after dropping the sample of plots on which these
inputs are applied. The results are presented in Table 1B.1 and 1B.2 for labor allocation and crop yield (see
APPENDIX B), respectively, are consistent with the main findings.

43

Table 1.5 Estimation of log of yield (FCFA/ha) across collective and male individual plots
Panel A: all crops

Collective=1
Log of plot size

Model 1
0.288***
(0.06)
-0.190***
(0.04)

Organic fertilizer=1
Inorganic fertilizer=1
Herbicide/insecticide=1
Household owns plot=1

Panel B: grain crops
Model 2
0.294***
(0.05)
-0.203***
(0.04)
0.231**
(0.11)
0.083
(0.11)
-0.213
(0.355)
0.141
(0.19)

Model 3
0.293***
(0.05)
-0.198***
(0.04)
0.231**
(0.11)
0.049
(0.11)
-0.188
(0.347)
0.134
(0.19)

Model 1
0.371***
(0.10)
-0.146**
(0.06)

Model 2
0.359***
(0.10)
-0.147**
(0.06)
0.264**
(0.13)
0.012
(0.13)
-0.41
(0.626)
0.113
(0.16)

Panel C: non-grain crops
Model 3
0.377***
(0.10)
-0.151**
(0.06)
0.313**
(0.13)
-0.026
(0.12)
-0.313
(0.565)
0.059
(0.19)

Model 1
0.299***
(0.07)
-0.324***
(0.07)

Model 2
0.309***
(0.07)
-0.337***
(0.07)
0.192
(0.18)
0.047
(0.20)
0.011
(0.282)
0.183
(0.26)

Model 3
0.305***
(0.07)
-0.327***
(0.07)
0.157
(0.18)
-0.011
(0.20)
0.045
(0.259)
0.21
(0.25)

Clay and sandy
soil (clay is
omitted)

Sandy
Gravel

-0.091
(0.09)
-0.254***
(0.09)
-0.038
(0.17)
Yes
5,062
0.061

0.258
(0.17)
-0.227
(0.18)
-0.255
(0.31)
Yes
2,824
0.097

-0.303***
(0.11)
-0.329***
(0.11)
-0.024
(0.25)
Yes
2,238
0.092

HHCYFE
Yes
Yes
Yes
Yes
Yes
Yes
Observations
5,062
5,062
2,824
2,824
2,238
2,238
R-squared
0.043
0.051
0.055
0.067
0.067
0.072
Sample Mean
(SD)
(Mean=11.692; SD=1.091)
Mean=11.595; SD=.941)
(Mean=11.815; SD=1.245
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample
restricted to households within which collective and male individual plots coexist. All specifications include household-year-crop fixed effects.
Top and bottom 1 percent of plot size and outcome variables dropped from sample. Village level prices recorded through community questionnaires were used to
convert production into monetary value. FCFA: Franc CommunautÊ Financière Africaine. The exchange rate at the survey times was approximately $1US=500
FCFA. Fewer number of observations relative to Table 1.4 because yield variables are at the crop-level.

44

The estimates of 𝛾 in Table 1.5 are positive and statistically different from zero at the 1 percent
significance level across all the three models. Both the sign and the significance level of the coefficient
remain robust even when crops are split into grain and non-grain crops (Panels B and C). The estimates of
𝛾 in the most exhaustive specification (i.e., Model 3) in Panels B and C in Table 1.5 (0.377 and 0.305,
respectively) imply that grain and non-grain crop yields are about 46 percent and 36 percent, respectively,
higher on collective plots relative to individual plots. The results also show that crop yield is decreasing
in plot size, suggesting diminishing returns to land.36
Recall that the difference between Models 2 and 3 is the inclusion of soil type in Model 3. The
reason for doing so is to observe how crop yield performance changes across individual and collective
plots when soil observable qualities are controlled for. If soil quality is higher on collective plots, then
failing to control for them is likely to bias upward the coefficient estimates in Model 2. Consequently, the
magnitude of γ̂ in Model 3, where such factors are controlled for, should be smaller relative to the ones in
Model 2. Yet, comparing γ̂ across Models 2 and 3 reveals almost no difference (Panel A) and if any, it
(i.e., γ̂) is slightly higher in Model 3 for grain crops (Panels B), suggesting that individual plots are at
least as productive as collective plots with respect to soil observable characteristics.
Though these findings do not offer a definite ground to infer that collective plots are not
inherently of better quality than individual plots, it seems less probable that the higher crop yield
performance on collective plots is driven by better unobservable or unobserved soil quality on these plots
relative to individual plots. These same alternative econometric specifications have been applied in
previous empirical work to establish that female (or individual) plots are at least as productive as male (or
collective) plots (Udry, 1996; Kazianga and Wahhaj, 2013).
Plot size, in Tables 1.4 and 1.5, is expressed in logarithm terms. Following Udry (1996) and
Kazianga and Wahhaj (2013), regressions are run with plot size expressed in deciles. The results,
presented in Table 1E.1 in Appendix E (see APPENDIX E), are consistent with those presented in Tables

36

Plot size is expressed in log. Similar results are obtained using plot deciles.

45

1.4 and 1.5. Applying the Inverse Hyperbolic Sine transformation (ISH) to the outcome variables and plot
size also provides similar results to the main findings.37
Furthermore, there is a concern that differences in yield and labor allocation across collective and
individual plots could be driven by no overlapping between some collective and individual plots due to
differences in plot size (see figure 1.4). To ease this concern, additional regressions are run by restricting
the sample to plot size where collective and individual plots overlap. The results, presented in Tables 1F.1
and 1F.2 (in APPENDIX F) for crop yield and labor allocation, respectively, remain consistent with the
main results.
Moreover, to test the robustness of the main results to a possible reverse causality between inputs
that are potentially complement (i.e., organic and inorganic fertilizers) or substitute (i.e., herbicide) to
labor, additional regressions are run after excluding all plots receiving such inputs. The results, presented
in Tables 1G.1 and 1G.2, respectively for labor allocation and crop yield performance, are consistent with
the main findings.
Finally, as shown in the descriptive statistics, the proportion of plots with companion crops is
significantly higher for collective plots (see Table 1.2). This combined with the fact that labor is recorded
at the plot-level, for the main crop on the plot, may lead to an overestimation of the intensity of labor on
the collective plots. However, the results presented in APPENDIX G for plots with no companion crops,
remain as the main findings, suggesting no major concerns about this potential source of a confounding
factor.
1.7.2

Labor intensity and crop yield with plot-manager-year fixed effects

A unique feature of the data used in the current work is that it allows to distinguish the head’s own
individual plots from plots that are purely collective. To my knowledge, such a distinction has not yet
been made in the existing literature dealing with similar questions (Kazianga and Wahhaj, 2013);
Guirkinger and Platteau, 2014); and Guirkinger et al., 2015). The advantage of being able to observe and

37

The advantage of the Inverse Hyperbolic Sine transformation is that, contrary to the logarithmic transformation, it
does not involve arbitrarily attributing non-zero positive values to observations with missing values.

46

distinguish between the head’s own individual plots and the collective plots he manages is threefold. The
first is the ability to control for plot-manager-fixed effects, which provides a unique way to address the
concern of omitted variable bias (such as that due to educational level, farming experience, and access to
extension services) as well as his access to productive resources.
The second advantage is that by focusing only on plots managed by the household head, I can
reduce potential bias due to endogenous plot allocation since the head is usually responsible for allocating
plots to household members. For example, among plots managed by the household head, soil quality is
likely to be similar on collective plots relative to individual plots because it is unlikely that the head will
choose a plot of poor soil quality for himself. Finally, this feature of the data provides a unique way to
study whether differences in labor allocation and crop yield across collective and individual plots are
solely because collective plots are almost always managed by the head of household, who is likely to have
more control over productive resources relative to the other household members. If this is the case, then
we should not observe any statistically significant differences in input allocation and in yield after
restricting the sample to collective plots and individual plots of household heads.
The results related to labor allocation with the restricted sample with specification without and
with plot covariates are presented in Table 1.6. Because crop effects are not incorporated in the
construction of the plot-manager-year fixed effects (PMYFE), in addition to controlling for plot
characteristics and organic and inorganic fertilizer use, I also control for crop fixed effects.

47

Table 1.6 Estimation of log of labor allocation (person-days/ha) across collective and heads’ individual plots using PMYFE
(1)
(2)
(3)
(4)
(5)
(6)
Log(Total
Total(Male
Log(Female
Log(Child
Log(Total
Total(Male
VARIABLES
Labor/ha)
labor/ha)
labor/ha)
labor/ha)
Labor/ha)
labor/ha)
Collective=1

0.148***
(0.047)

0.086
(0.053)

0.495***
(0.092)

0.455***
(0.107)

Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Herbicide / insecticide=1
Household owns plot=1
Clay & sandy soil (clay is omitted)
Sandy
Gravel

0.373***
(0.047)
-0.329***
(0.028)
0.119***
(0.043)
0.078
(0.054)
-0.072
(0.118)
0.135
(0.196)
-0.05
(0.056)
-0.109**
(0.050)
0.039
(0.107)

0.315***
(0.052)
-0.334***
(0.032)
0.163***
(0.046)
0.076
(0.056)
-0.057
(0.129)
0.15
(0.217)
-0.104*
(0.059)
-0.08
(0.055)
0.019
(0.114)

(7)
Log(Female
labor/ha)
0.625***
(0.095)
-0.207***
(0.049)
-0.003
(0.082)
0.117
(0.098)
-0.038
(0.192)
0.081
(0.261)
-0.117
(0.106)
-0.142
(0.090)
0.018
(0.165)

(8)
Log(Child
labor/ha)
0.610***
(0.112)
-0.221***
(0.048)
0.095
(0.079)
0.078
(0.098)
-0.094
(0.232)
-0.132
(0.216)
0.254*
(0.136)
0.064
(0.112)
0.328**
(0.158)

PMYFE
Observations
1,743
1,743
1,743
1,743
1,743
1,743
1,743
1,743
R-squared
0.075
0.045
0.114
0.052
0.209
0.175
0.134
0.079
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to heads that
control simultaneously their own individual plots and collective plots. Top and bottom 1 percent of continuous variables dropped from sample

48

The results show that collective plots are significantly more intensively farmed than the head’s
own individual plots, regardless of labor type and the model specification, except for male labor in Table
1.6, column 2. To be more specific, even when collective plots and individual plots are controlled by the
same person, they (i.e., the collective plots) still receive significantly more labor relative to individual
plots. The estimate of 𝛾 for total labor when all plot covariates are included is about 0.37 (Table 1.6,
column 5). Since the dependent variables are expressed in log, this indicates a differential impact for
collective plots on total labor allocation of approximately 45 percent. Disaggregating labor allocation by
labor type reveals that labor allocation gaps are higher for female and child labor, which is consistent with
the results obtained with the household-crop-year fixed effects (Table 1.4).
Consistent with the labor allocation results, crop yields with the restricted sample reported in
Appendix H (see Table 1H in the appendix) are higher for collective plots. Specifically, I find that crop
yields are approximately 46 percent and 34 percent higher on collective plots relative to individual plots
for grain (Model 3, Panel B, Table 1.9) and non-grain crops (Model 3, Panel C, Table 1.9), respectively.
The order of the differences is similar to the finding with the results presented in Table 1.6. The
comparison of 𝛾̂ between Models 2 and 3 shows that the magnitude of the coefficient estimate remains
almost of the exact size regardless of whether the type of soil on the plot is included or not. These results
further support the argument that the large observed differences in yield are unlikely to be driven by
omitted variables bias in soil quality between collective and individual plots.
The consistently positive and significant coefficient estimates of the indicator variable collective
across the alternative econometric specifications and sample restrictions provides strong evidence that
collective plots receive more inputs and present higher yield than individual plots. These findings are not
in line with what it is generally predicted by the economic theory on collective action, which is low
efforts provision in the production of collective goods. Indeed, with one exception (Kazianga and
Wahhaj, 2013), an overwhelming majority of the empirical literature finds evidence that collectively
managed farms suffer from low effort provision leading to lower overall output level (Putterman and
DiGiorgio, 1985; Lin, 1987; Carter, 1987; McMillan et al., 1987; Guirkinger and Platteau. 2014).

49

As a result, a natural question arises from the current paper: Why do we find that collective plots
are more intensively farmed? Particularly, given that the proceeds from the collective plots are for shared
consumption within the household, what deters individuals from shirking on the collective plots and
devoting more effort to their individual plots where they alone claim the farm’s production? While it is
not possible to offer a definite answer to these questions, I posit that motives for insurance against
production risk partly explain these observed labor and yield gaps across the two types of plots
management. This idea of collective farming being favored when production risk prevails has indeed
been theoretically explored, but its empirical application, especially in West Africa, remains very
limited.38
The underpinning assumption for the risk insurance claim is that collective farming can enable
nuclear households to reduce production variability across years. I provide suggestive evidence for this
claim in APPENDIX I, where a Chow test across the two groups of plots, on the effect of the time
indicator variable on yield residuals, shows that changes in yield dispersions across the two survey rounds
are significantly lower for collective plots relative to individual plots. This lower variation on yield
dispersion across years on the collective plots provides somehow an insured agricultural output for
extended household members. Collective farming may therefore be favored by households living in
places with high weather unpredictability, which is a major source for production uncertainty. As a result,
these households are likely to be inclined to allocate more labor on the collective plots relative to their
own plots, where production uncertainty is higher. This assertion is empirically explored in the following
section.
Certainly, there are other potential explanations for why collective plots are more intensively
farmed than individual plots. Such justifications may include more tenure security on collective plots,
though this mechanism is less plausible in the current context, given that the type of crops included in this
study are all short-term rather than perennial crops.39 In addition, because collective plots are more likely

38
39

One of the few existing studies is the work by Rosenzweig and Wolpin (1985) in India.
Also note that there are no formal land titles in rural Mali. Households only have usufruct right to exploit the land.

50

to be farmed by a group of workers, one may argue that this would lead to peer effects, inducing longer
hours of work on the collective plots and, therefore, higher yield. This is not likely to be an important
factor because it is also possible to have more than one worker on the individual plots. As a result, peer
effects may rise on these plots as well. Economies of scale in terms of land exploitation can also be ruled
out as plot size is included as a control in all specifications. However, economies of scale in terms of
indivisible assets such as draft animal may incite participation in collective farming. This is explored in
section 1.8. Finally, another plausible mechanism is social norms that may compel members from the
nuclear households to provide labor on collective plots. This latter mechanism is more plausible but
unfortunately, it cannot be tested with the available data.

1.8

Evidence for motive of risk insurance

Recall that the central assumption in this study is that the higher labor allocation on the collective plots
can be partly attributed to motives for insurance against production uncertainty. Production uncertainty
cannot be directly observed. Nevertheless, given that the main production system in the study area is rainfed crop cultivation, which strongly depends on rainfall, I use rainfall variability as a proxy for production
uncertainty.
This part of the study provides the estimation results of plot-level labor allocation and crop yield
performance across collective and individual plots taking into account rainfall variability represented by
the coefficient of variation of rainfall from 1981 to 2007. The likelihood of the participation of the nuclear
household in collective farming with extended family members depending on rainfall variability is also
presented here.
1.8.1

Rainfall variability and labor allocation

The estimation results related to the hypothesis that higher labor allocation on the collective plots are
partly driven by motives for risk insurance are presented in Table 1.7 (for total labor) and Table 1.8 (for
male, female, and child labor). Three specifications are presented in each table: (1) only controls for the
binary variable Collective and its interaction with the coefficient of variation of rainfall (2) adds plot

51

characteristics, while (3) adds village covariates.40 These results are obtained using the sample where
collective and male individual plots coexist and correspond to the estimation of equation 7. All standard
errors in these regressions are clustered at the village level. Rainfall variability has been standardized to
ease the interpretation of the results –that is, such that the coefficient estimates on the interaction term are
interpreted as an approximate percent increase in the amount of labor following an increase in rainfall
variability by one standard deviation. The coefficient estimate of the interaction term is positive and
statistically different from zero at the 5 percent significance level for total labor when all available
covariates are included (Table 1.7, column 3). When labor is disaggregated, the magnitude of the
coefficient is much larger for female labor with a statistical significance level of 5 percent, followed by
male labor while there is no indication of any significant differences with respect to child labor.

40

Rainfall variability cannot be included in the estimations without interacting it with the collective indicator, as this
variable is constant within villages and between the two rounds of the panel. Rainfall variability is computed for
crop cultivation season (May through September).

52

Table 1.7 Estimation of log of labor allocation (person-days/ha) across collective and male individual plots with
rainfall variability (without plot covariates)
Log(total labor/ha)
(1)
(2)
(3)
Collective
0.177***
0.391***
0.335***
(0.036)
(0.041)
(0.055)
CV_rain*Collective
0.056
0.069*
0.079**
(0.038)
(0.039)
(0.039)
Plot covariates?

No

Yes

Village covariates?

Yes

No
No
Yes
HHCYFE?
Yes
Yes
Yes
Observations
6,391
6,391
6,391
R-squared
0.028
0.104
0.107
Plot level observations. Robust clustered standard error at the village level in parentheses. * p<0.1; ** p<0.05; ***
p<0.01. Sample restricted to households within which collective and male individual plots coexist. All specifications
include household-year-crop fixed effects. CV_rain=Rainfall variability and is computed for crop cultivation season
(May through September) from 1981 to 2007. For ease of interpretation, CV_rain has been standardized. Fewer
observations here relative to Table 1.4 because of missing village-level data for some villages. For ease of
interpretation, CV_rain has been standardized. Top and bottom 1 percent of plot size and outcome variables dropped
from sample. Plot covariates include log of plot size, use of organic fertilizer (1=yes), inorganic fertilizer (1=yes),
plot tenure (=1 if owned by household), and soil type (clay, sandy, clay and sandy mix, or gravel). Village level
covariates include outmigration (i.e., whether village experienced households moving out of village over the last ten
years), access to formal credit in village, and village size (in terms of number of households in village)

53

Table 1.8 Estimation of log of labor allocation (person-days/ha) across collective and male individual plots controlling for rainfall (controlling for plot
covariates)
Panel A: Log(male labor/ha)
Panel B: Log(Female labor/ha)
Panel C: Log(Child labor/ha)
(1)
(2)
(3)
(1)
(2)
(3)
(1)
(2)
(3)
Collective
0.092**
0.305***
0.256*** 0.341***
0.487***
0.501***
0.720***
0.817***
0.617***
(0.040)
(0.049)
(0.064)
(0.061)
(0.068)
(0.099)
(0.070)
(0.084)
(0.124)
CV_rain*Collective 0.062
0.077*
0.087*
0.121**
0.129**
0.133**
-0.04
-0.041
-0.008
(0.045)
(0.046)
(0.046)
(0.058)
(0.059)
(0.061)
(0.063)
(0.064)
(0.065)
Plot covariates?
No
Yes
Yes
No
Yes
Yes
No
Yes
Yes
Village covariates? No
No
Yes
No
No
Yes
No
No
Yes
HHCYFE?
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
6,391
6,391
6,391
6,391
6,391
6,391
6,391
6,391
6,391
R-squared
0.009
0.074
0.077
0.038
0.053
0.055
0.101
0.113
0.121
Plot level observations. Robust clustered standard error at the village level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to households
within which collective and male individual plots coexist. All specifications include household-year-crop fixed effects. CV_rain=Rainfall variability and is
computed for crop cultivation season (May through September) from 1981 to 2007. For ease of interpretation, CV_rain has been standardized. Fewer
observations here relative to Table 1.4 because of missing village-level data for some villages. Top and bottom 1 percent of plot size and outcome variables
dropped from sample. Plot covariates include log of plot size, use of organic fertilizer (1=yes), inorganic fertilizer (1=yes), plot tenure (=1 if owned by
household), and soil type (clay, sandy, clay and sandy mix, or gravel). Village level covariates include outmigration (i.e., whether village experienced households
moving out of village over the last ten years), access to formal credit in village, and village size (in terms of number of households in village)

54

With the inclusion of the interaction term, the differential effect of a plot being collectively
managed on labor allocation is now captured by the sum of the coefficient estimates of the interaction
term and the Collective indicator –that is, 𝛾̂ + 𝜅̂ . Note that the standard deviation of the coefficient of
variation (CV) is 0.005 (with a corresponding mean of 0.216), which is relatively low. Despite this low
cross-sectional variation in rainfall variability across villages, I find significant differences in labor
allocation by different levels of rainfall variability. For example, according to the results presented in
Table 1.7 in column 2, an increase of one standard deviation in the CV of rainfall leads to an additional
labor allocation gap across collective and individual plots by about 7 percent and this is statistically
different than zero at the 10 percent significance level.
The corresponding figures for male and female labor are approximately 8 percent and 13 percent,
respectively (Table 1.8, columns 2). These labor allocation gaps remain robust to the inclusion of village
level covariates (Tables 1.7 and 1.8, columns 3), with larger coefficients and stronger statistical
significance level, suggesting that estimates of 𝜅 are unlikely to be substantially biased by omitted village
level characteristics.
Labor allocation is further disaggregated by grain and non-grain production in Table 1.9, Panel A
and Panel B, respectively. There is a strong indication of different levels of correlation between rainfall
variability and labor allocation across collective and individual plots by type of crop cultivation.
Specifically, while the estimates of coefficients of the interaction term CV_rain*Collective are positive
and significant for grain production, the corresponding figures for non-grain are not statistically different
from zero. This holds regardless of whether plot or village covariates are included. Since grain production
is the main source of livelihood in the study area, these findings provide strong supporting evidence of
my claim that the higher labor allocation on the collective plots can be partly attributed to motives for
securing food availability.
The differences in labor allocation are substantiated in the yield results presented in Table 1.10,
where yield gap across collective and individual plots are higher among households living in areas with
high rainfall variability. For example, considering the specification where all control variables are

55

included (Table 1.10, column 3), households living in villages where the CV of rainfall variability is one
standard deviation higher, achieve approximately 10 percent more crop yields on their collective plots
relative to their individual plots.41

Table 1.9 Estimation of log of total labor allocation (person-days/ha) across collective and male individual plots
controlling for rainfall by grain and non-grain production
VARIABLES

Panel A: Grain only

Panel B: Non-grain only

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

0.135**

0.373***

0.311***

0.185***

0.403***

0.374***

(0.060)

(0.061)

(0.095)

(0.039)

(0.057)

(0.070)

0.119**

0.123**

0.137**

0.01

0.008

0.017

(0.058)

(0.056)

(0.056)

(0.039)

(0.042)

(0.042)

Plot covariates

No

Yes

Yes

No

Yes

Yes

Village covariates

No

No

Yes

No

No

Yes

HHYCFE

Yes

Yes

Yes

Yes

Yes

Yes

Observations

3,794

3,794

3,794

2,598

2,598

2,598

Collective=1
CV_rain*Collective

R-squared
0.029
0.102
0.109
0.029
0.120
0.126
Plot level observations. Robust clustered standard error at the village level in parentheses. * p<0.1; ** p<0.05; ***
p<0.01. Sample restricted to households within which collective and male individual plots coexist. All specifications
include household-year-crop fixed effects. CV_rain=Rainfall variability and is computed for crop cultivation season
(May through September) from 1981 to 2007. For ease of interpretation, CV_rain has been standardized. Top and
bottom 1 percent of plot size and outcome variables dropped from sample. Plot covariates include log of plot size,
use of organic fertilizer (1=yes), inorganic fertilizer (1=yes), plot tenure (=1 if owned by household), and soil type
(clay, sandy, clay and sandy mix, or gravel). Village level covariates include outmigration (i.e., whether village
experienced households moving out of village over the last ten years), access to formal credit in village, and village
size (in terms of number of households in village)

Udry (1996) demonstrates that differences in crop yield resulting from differences in labor
allocation across plots (male and female plots) of similar physical characteristics are evidence for
efficiency loss due to decreasing marginal productivity of labor. Accordingly, while the current work does
not test for efficiency, the fact that the labor allocation gap in Tables 1.7 and 1.8 is translated into a higher
yield gap in Table 1.10 suggests efficiency loss among households in the study area.

41

The results with the data disaggregated by grain and non-grain, not reported here, though positive, did not reveal
any statistical significant difference.

56

Table 1.10 Estimation of the log of yield for all crops (FCFA/ha) across collective and male individual plots
controlling for rainfall variability
Log(yield (FCFA/ha))
(1)
(2)
(3)
Collective
0.080*
0.283***
0.188**
(0.043)
(0.058)
(0.081)
CV_rain*Collective 0.074*
0.088**
0.097**
(0.042)
(0.040)
(0.040)
Plot covariates?
No
Yes
Yes
Village covariates? No
No
Yes
HHCYFE?
Yes
Yes
Yes
Observations
4,728
4,728
4,728
R-squared
0.013
0.061
0.066
Crop level observations. Robust clustered standard error at the village level in parentheses. * p<0.1; ** p<0.05; ***
p<0.01. Sample restricted to households within which collective and male individual plots coexist. All specifications
include household-year-crop fixed effects. CV_rain=Rainfall variability and is computed for crop cultivation season
(May through September) from 1981 to 2007. For ease of interpretation, CV_rain has been standardized. Fewer
observations in column (2) because of missing village-level data for some villages. Top and bottom 1 percent of plot
size and outcome variables dropped from sample. Plot covariates include log of plot size, use of organic fertilizer
(1=yes), inorganic fertilizer (1=yes), plot tenure (=1 if owned by household), and soil type (clay, sandy, clay and
sandy mix, or gravel). Village level covariates include outmigration (whether village experienced households
moving out of village over the last ten years), access to formal credit in village, and village size (in terms of number
of households in village)

1.8.2

Rainfall variability and participation in collective farming

Based on the labor allocation results presented in Tables 1.7 and 1.8, it is sensible to hypothesize that
collective farming may be viewed as an ex ante risk management strategy against production risk among
households in the study zone, which I test by estimating equation 9. Specifically, I estimate a linear
probability model where the outcome variable is a binary variable taking one if the household is engaged
in collective farming, and zero otherwise. The results obtained from the estimation of equation 9 at the
household level are presented in Table 1.11. Recall that the sample used here includes all the three types
of households described in section 4. However, because of differences in norms that govern land
distribution to female and male plot managers (see section 1.6 for more details), all female plots have
been excluded from the sample before putting the data at the household level.42

42

The results (available upon request) remain robust to the inclusion of female plots.

57

Table 1.11 Test for collective farming as ex-ante risk mitigation strategy. Dependent variable is participation of
nuclear household in collective farming
VARIABLES
Model 1 Model 2
Model 3
Model 4
Model 5
Rainfall variability (standardized
CV)
Head age 30-40 (<=30 omitted)
Head age 40-50
Head age 50-60
Head age>60
Household size

0.051***
(0.007)

0.055***
(0.008)
-0.088***
(0.016)
-0.158***
(0.020)
-0.084***
(0.019)
-0.044**
(0.022)
-0.024***
(0.003)

Land holding

0.025***
(0.005)

0.051***
(0.008)

-0.10***
(0.004)
-0.091***
(0.014)

Draft animal=1
Access to formal credit in village

-0.001
(0.026)
-0.000**
0.000
-0.007
(0.016)

Number of HHs in village
Village experienced outmigration

0.023***
(0.005)
-0.036***
(0.012)
-0.073***
(0.015)
-0.031***
(0.015)
-0.025
(0.014)*
0.013***
(0.002)
-0.10***
(0.004)
-0.091***
(0.015)
0.003
(0.016)
0
0.000
-0.02**
(0.011)

Observations
5,325
5,325
5,325
5,325
5,325
R-squared
0.015
0.061
0.580
0.017
0.583
* p<0.1; ** p<0.05; *** p<0.01. Observations are at the household level. Clustered standard errors at the village
level are in parentheses. Sample includes all restricted households (after dropping all female plots). Household
covariates include size of nuclear household and its land holding. Village level covariates include outmigration
(whether village experienced households moving out of village over the last ten years), access to formal credit in
village, and village size (in terms of number of households in village). Rainfall variability is computed for crop
cultivation season (May through September) from 1981 to 2007. For ease of interpretation, rainfall variability has
been standardized.

The results, obtained with the first wave of the data are presented for five different specifications
(Models 1-5).43 The motivation for these alternative specifications is to check the consistency of the
coefficient estimates of the key variable –that is rainfall variability. Model 1 represents the specification
with rainfall variability only. Models 2 and 3 include controls for household demographics (represented

43

I use only the first wave data because participation in collective farming does not vary much across the two
waves. In addition, Historical rainfall variability is the same across the two waves, such that panel regressions with
household fixed effect cannot be used. Results (not shown here) using the second wave of observations are
consistent with the current findings.

58

by the size of the household and the head’s age) and household asset holdings (captured by total land and
possession of a draft animal), respectively. Model 4 controls for village covariates (availability of formal
credit in the village, whether the village experienced outmigration of households over the last ten years,
and the total number of households in the village) while Model 5 includes all covariates.
The results in Model 1 indicate that an increase of one standard deviation in the CV of rainfall is
associated with a nuclear household being more likely to be engaged in collective farming by about 5
percentage point. Given that the standard deviation of the sample is only 0.005, this finding implies that
the differential impact of rainfall variability would have been much higher, had there been more crosssectional variation in rainfall variability.
One of the central arguments for collective farming is the benefit from the experience of the
elderly household members in dealing with adverse weather conditions, which allows extended family
members to mitigate their exposure to production variability (Rosenzweig and Wolpin, 1985). This
argument is supported by the negative and highly statistically significance level of the coefficient
estimates on the nuclear household head’s age range. Particularly, with the omitted age category being
30-year-old or younger, the statistics in Model 2 show that nuclear households where the head’s age is
between 30 and 40 are about 9 percentage point less likely to be engaged in collective farming relative to
their counterparts where the head’s age is less than 30-year old. This statistic goes up to almost 16
percentage points when the head’s age is between 40 and 50. The differential effect of the head’s age
range remains robust to the inclusion of all covariates (Model 5) but the coefficient estimates when the
head’s age is greater than 60, is only marginally statistically different from zero.
The results in Model 2 also show that larger nuclear households are less likely to be engaged in
collective farming. This result is anticipated given that a larger number of individuals in the household
represents more manpower and, hence, no need to rely on collective farming to satisfy labor requirement.
The sign of the coefficient on household size is flipped in Model 5, when all available covariates are
controlled for, which is puzzling, but the magnitude becomes very low (less than one percentage point).
With respect to assets holding, in Model 3, I find that nuclear household with larger land holdings

59

are less likely to engage in collective farming. This evidence is consistent with finding from previous
work in Burkina Faso (Kazianga and Wahhaj, 2017) where the authors evoke limited access to land
markets as one of the driving factors for collective farming among extended family members. A similar
case was observed in India (Rosenzweig and Wolpin, 1985). Model 3 also controls for draft animal
ownership, which represents an important agricultural technology among subsistence farmers for land
preparation. The size of the corresponding coefficient is almost 10 percent and statistically different from
zero at the 1 percent significance level, implying that nuclear households who own a draft animal are
much less likely to be engaged in collective farming. Since draft animal is an indivisible investment, the
current finding strongly corroborates the view of economies of scale, in terms of technology adoption, as
one of the driving factors for collective farming (Carter, 1987; Mathijs, E., & Swinnen, 2001).
Note how the size and the significance level of the coefficient on rainfall variability remains
strongly consistent across the different specifications, except when asset holdings are included as controls
where the size of the coefficient is reduced by about half.44 Therefore, although the current results do not
provide grant for causal inferences, they provide strong suggestive evidence that collective farming
constitutes an ex-ante risk mitigation mechanism for households in the study area of the current work.

1.9

Conclusion and policy implications

This paper investigates household labor allocation and crop yield across collective and individual plots
among rural households in the Segou region of Mali. There is strong empirical evidence that collective
plots are more intensively farmed with labor, which translates into higher crop yield performance on
collective plots by up to 47 and 46 percent, respectively.
I claim that the higher labor allocation and crop yields on collective plots are driven partly by the
way collective farming may constitute an ex ante risk mitigation strategy among households in the study
zone. This assertion is tested by using village level historical rainfall data to show that differences in labor

44

The coefficient on rainfall variability is 0.054 when land holding and draft animal ownership are omitted in Model
5 (results not reported here).

60

allocation across collective and individual plots are higher among households in sites with greater rainfall
variability.
The argument behind the view of collective farming as an ex ante risk mitigation strategy is that it
enables nuclear households to lower their production variability. The study provides evidence in support
of this claim by showing that yield dispersion across time is significantly lower on collective plots
relative to individual plots. Furthermore, I show with strong suggestive evidence that places with higher
rainfall variability tend to have a higher concentration of households engaged in collective farming with
extended household members.
While the evidence of risk insurance presented here is compelling, it is important to be
mindful that this analysis does not provide a basis for causal inference. Ex ante risk insurance motives are
surely not the only potential driving factors behind the observed labor allocation gap. Other plausible
factors include social and cultural norms regarding moral obligations to contribute to joint family welfare,
which may compel nuclear household members to provide more labor on collective plots.45
Despite these limitations, the findings provide novel and noteworthy implications for the study of
the complexity of household economic decision-making in sub-Saharan African countries. A striking
feature is that the results from the current study, while in line with empirical evidence in Burkina Faso by
Kazianga and Wahhaj (2013) –that is, higher intensity of labor use on collective plots relative to
individual plots – differ from the findings of Guirkinger and Platteau (2014) among rural households in
Mali.
Rather than perceiving the current results as conflicting with those obtained by Guirkinger and
Platteau (2014), I consider the two studies to be complementary due to the regional disparities between

45

This is particularly important in the context of Malian culture, where there is social pressure to contribute to the
joint welfare of the whole family, and failing to do so may result in being perceived as selfish (personal
communications and observations). Individuals who detract from following such social norms are usually the subject
of gossip, which can tarnish their reputation and social status. Nevertheless, it is less plausible that social norms
constitute an important factor in explaining the higher labor allocation on the collective plots. In fact, Guirkinger and
Platteau (2014), whose work is in another region in the same country, report that some heads of households
complain about shrinking behaviors exhibited by individual plot managers on the collective plots.

61

the two study zones. To be specific, the study area of the current work is mainly characterized by
drylands, with very limited access to irrigation, where the main source of livelihood is rain-fed crop
cultivation. Untimely execution of farming activities can have dire impacts on crop yield performance.
This may provide incentives for nuclear households in the region to scrupulously work on the collective
plots, where it is possible for them to have “brute strength” and “experienced” labor, as well as a larger
pool of labor supply, allowing them to insure one another against production risk. In contrast, individuals
in the study zone of Guirkinger and Platteau (2014), especially the Sikasso region, have more
opportunities to engage in the cultivation of cash crops on their individual plots, as well as to access nonagricultural income generating opportunities.46
The finding that differences in plot-level intra-household labor allocation are reflected in crop
yield performance across plots of similar characteristics in the current work is consistent with previous
empirical evidence in West Africa. This strongly supports the premise that labor remains a major
productive factor among households in this region. As a result, if motives for ex ante risk insurance are
plausible driving factors in the observed labor allocation gaps among households in the current work,
increasing farmers’ access to productive safety nets and promoting labor saving technologies presents a
strong potential to alleviate labor bottleneck at the peak of the farming season and improve overall
agricultural productivity. For example, herbicide use, which significantly reduces weeding efforts while
increasing crop yield performance, is very low among the studied population despite recent findings of a
steep rise in herbicides use in other regions in Mali.47
The differences in the empirical literature regarding rural household economies highlighted here,

Indeed, the authors find that “mixed farming” –that is, the coexistence of collective and individual plots within the
same household –is more predominant among extended households with a higher proportion of bottomland relative
to extended households without individual plots. This supports the argument of the opportunity for more cash crop
production for “individualized farming” in that study region. In effect, bottomland is not only suited for crops
production such as vegetables, but also can be exploited beyond the main cropping season (Guirkinger et al., 2015)
when there is less competition in labor demand across individual and collective plots, as the latter tend to be mainly
grown to rain-fed crops.
47
Haggblade et al. (2017) find that up to 60 percent of plots in the Sudanian Savana zone in Mali receive herbicide
while the corresponding figure in the current work is only about 5 percent . This sharp difference is likely to be due
to regional disparities in terms of agro-ecological zones and herbicide markets.
46

62

underline the challenge faced by development practitioners in West Africa to improve aggregate measures
of household welfare in contexts where intra-household dynamics can be location-specific. It is therefore
often difficult to tailor policies to specific groups of people or locations. This highlights the need for
continuous research efforts in understanding household production decision making in these regions.

63

APPENDICES

64

APPENDIX A ADDITIONAL DESCRIPTIVE STATISTICS
Table 1A. 1 Plot characteristics and crop choice across collective and individual plots (Full sample)
Individual
Collective
Plots (N=9,889)
Plots (N=41,059)
Mean
SD
Mean
SD
P-value
0.000
Female=1
0.710
0.450
0.000
Plot size
0.600
1.020
2.090
2.700
0.029
Clay soil
0.370
0.480
0.350
0.480
0.022
Clay & sandy mix
0.350
0.480
0.370
0.480
0.903
Sandy Soil
0.260
0.440
0.260
0.440
0.355
Gravel
0.020
0.150
0.030
0.160
0.000
Household owns plot
0.940
0.240
0.980
0.140
0.000
Household rents plots
0.060
0.240
0.020
0.140
Millet
0.000
0.160
0.370
0.270
0.450
Sorghum
0.000
0.090
0.290
0.190
0.390
Maize
0.000
0.030
0.180
0.100
0.300
Rice
0.000
0.040
0.200
0.090
0.290
Peanut
0.000
0.530
0.500
0.140
0.340
Beans
0.062
0.030
0.180
0.030
0.170
Chickpeas
0.002
0.040
0.190
0.050
0.220
Fonio
0.000
0.010
0.100
0.080
0.270
Sesame
0.000
0.050
0.230
0.040
0.200
0.000
Inorganic fertilizer=1
0.110
0.320
0.300
0.460
Organic fertilizer=1
0.230
0.420
0.320
0.470
0.001
Total labor (person-days/ha)
158.090
473.740
132.800
189.390
0.000
Male labor (person-days/ha)
37.760
136.920
62.040
88.350
0.000
Female labor (person-days/ha)
88.210
316.530
35.270
70.230
0.106
Child labor (person-days/ha)
32.540
122.180
35.650
66.280
Plot level observations. Sample includes all households as well as all plots (including female individual plots)

Table 1A. 2 Plot characteristics and crop choice across all collective and male individual plots
Individual Plots
Collective Plots
(N=2,595)
(N=35,802)
VARIABLES
Mean
SD
Mean
SD
Plot size
0.86
0.92
2.15
2.68

65

P-value of Diff.
0.00

Table 1A.2 (cont’d)
Soil type
Clay soil
Clay & sandy mix
Sandy soil
Gravel
Household owns plot
Main source of water is rain
Crop choice
Millet
Sorghum
Maize
Rice
Peanut
Beans
Chickpeas
Fonio
Sesame
Inorganic fertilizer (1=yes)
Organic fertilizer (1=yes)
Inorganic fertilizer (kg/ha)
Total labor (person-days/ha)
Male labor (person-days/ha)

0.32
0.37
0.27
0.03
0.96
0.98

0.47
0.48
0.45
0.17
0.2
0.13

0.35
0.36
0.26
0.03
0.98
0.97

0.48
0.48
0.44
0.16
0.15
0.17

0.03
0.01
0.75
0.83
0.00
0.11

0.14
0.10
0.06
0.06
0.47
0.04
0.02
0.03
0.09
0.11
0.26
7.06
106.86
56.1

0.35
0.30
0.23
0.23
0.5
0.2
0.15
0.17
0.28
0.32
0.44
29.31
97.14
55.79

0.28
0.19
0.1
0.09
0.14
0.03
0.05
0.08
0.04
0.31
0.33
15.65
109.83
52.95

0.45
0.4
0.3
0.29
0.34
0.16
0.22
0.27
0.19
0.46
0.47
53.49
94.67
47.22

0.00
0.00
0.00
0.00
0.00
0.73
0.00
0.00
0.00
0.00
0.00
0.00
0.59
0.03

0.13
Female labor (person-days/ha)
26.77
35.91
28.48
32.39
0.01
Child labor (person-days/ha)
24.14
35.34
28.45
33.73
Plot level observations. Sample includes all households participating in collective farming after dropping all female plots

Table 1A. 3 Plot characteristics by crop choice across collective and male individual plots
Individual plots
Collective plots
N
Mean
SD
N
Mean
Clay_rice
Clay_millet
Clay_sorghum
Clay_maize

83
249
178
104

0.88
0.22
0.47
0.38

0.33
0.42
0.5
0.49

66

450
1,471
1,024
545

0.89
0.23
0.43
0.34

SD

P-value of diff.

0.31
0.42
0.5
0.47

0.83
0.75
0.29
0.39

Table 1A.3 (cont’d)
Clay_fonio
Clay_sesame
Clay_peanut
Clay_beans
Clay_chickpeas
Clay_sandy_rice
Clay_sandy_millet
Clay_sandy_sorghum
Clay_sandy_maize
Clay_sandy_fonio
Clay_sandy_sesame
Clay_sandy_peanut
Clay_sandy_beans
Clay_sandy_chickpeas
Sandy_rice
Sandy_millet
Sandy_sorghum
Sandy_maize
Sandy_fonio
Sandy_sesame
Sandy_peanut
Sandy_beans
Sandy_chickpeas
Gravel_rice
Gravel_millet
Gravel_sorghum
Gravel_maize
Gravel_fonio
Gravel_sesame
Gravel_peanut
Gravel_beans
Gravel_chickpeas

61
145
793
50
40
83
249
178
104
61
145
793
50
40
83
249
178
104
61
145
793
50
40
83
249
178
104
61
145
793
50
40

0.13
0.29
0.27
0.18
0.1
0.05
0.42
0.39
0.36
0.49
0.45
0.38
0.38
0.42
0.05
0.33
0.13
0.18
0.33
0.2
0.32
0.42
0.42
0.02
0.02
0.01
0.08
0.05
0.06
0.03
0.02
0.05

0.34
0.46
0.45
0.39
0.3
0.22
0.49
0.49
0.48
0.5
0.5
0.48
0.49
0.5
0.22
0.47
0.34
0.39
0.47
0.4
0.47
0.5
0.5
0.15
0.15
0.11
0.27
0.22
0.24
0.17
0.14
0.22

Plot level observations. Sample includes all households after dropping all female plots

67

479
257
591
145
232
450
1,471
1,024
545
479
257
591
145
232
450
1,471
1,024
545
479
257
591
145
232
450
1,471
1,024
545
479
257
591
145
232

0.18
0.31
0.28
0.39
0.19
0.07
0.41
0.35
0.38
0.32
0.34
0.33
0.32
0.31
0.02
0.34
0.18
0.25
0.45
0.26
0.35
0.26
0.48
0.02
0.02
0.04
0.03
0.05
0.1
0.04
0.03
0.02

0.38
0.46
0.45
0.49
0.39
0.25
0.49
0.48
0.49
0.47
0.47
0.47
0.47
0.47
0.15
0.47
0.38
0.43
0.5
0.44
0.48
0.44
0.5
0.13
0.15
0.19
0.16
0.22
0.3
0.19
0.16
0.15

0.28
0.72
0.76
0.00
0.11
0.42
0.71
0.40
0.59
0.00
0.04
0.03
0.46
0.19
0.35
0.85
0.08
0.13
0.02
0.21
0.23
0.04
0.54
0.75
0.81
0.01
0.10
0.92
0.15
0.29
0.76
0.43

Table 1A.4. Coefficients of correlation between rainfall variability and village covariates
Correlation coefficient with
Village covariates
rainfall variability
Households in village have access to formal credit in
0.081
village
Village size (number of households in village)
-0.03
Village experienced outmigration of households over last
0.004
10 years

68

APPENDIX B ROBUSTNESS CHECK BY DROPPING ALL FERTILIZED PLOTS
Table 1B.1 Log of labor allocation estimation after dropping all plots with inorganic and organic fertilizer
VARIABLES

Total labor Male labor Female labor

Child labor

Collective=1

0.415***

0.319***

0.474***

0.829***

(0.049)

(0.054)

(0.079)

(0.099)

-0.296***

-0.300***

-0.204***

-0.149**

(0.033)

(0.037)

(0.057)

(0.063)

0.188

-0.015

0.322

0.952***

(0.227)

(0.227)

(0.327)

(0.327)

-0.032

-0.04

-0.22

0.111

(0.086)

(0.084)

(0.143)

(0.166)

-0.147**

-0.088

-0.257**

-0.059

(0.071)

(0.075)

(0.118)

(0.136)

0.131

-0.063

0.245

0.439

(0.145)

(0.153)

(0.278)

(0.328)

4,141

4,141

4,141

4,141

Log of plot size
Household owns plot=1

Clay and sandy soil (clay is omitted)
Sandy
Gravel

Observations

R-squared
0.1338
0.0965
0.061
0.1325
Plot level observations. Robust clustered standard error at the household level in parentheses.
* p<0.1; ** p<0.05; *** p<0.01. Sample restricted to households within which collective and male individual plots
coexist. All plots on which organic and or inorganic fertilizer is used are dropped. All specifications include
household-year-crop fixed effects. Top and bottom 1 percent of plot size and outcome variables dropped from
sample

Table 1B.2 Log of crop yield (FCFA/ha) estimation after dropping all plots with inorganic and organic fertilizer
VARIABLES

All crops

Cereal crops

Non-cereal crops

Collective=1

0.255***

0.383**

0.273***

(0.065)

(0.164)

(0.071)

-0.213***

-0.109

-0.310***

(0.066)

(0.116)

(0.082)

0.064

0.383**

0.103

(0.298)

(0.164)

(0.313)

-0.082

0.732***

-0.266**

(0.126)

(0.219)

(0.126)

-0.349***

0.054

-0.415***

(0.110)

(0.333)

(0.117)

0.011

0.148

-0.085

Log of plot size
Household owns plot=1

Clay and sandy soil (clay is omitted)

Sandy
Gravel

69

Table 1B.2 (cont’d)

Observations

(0.237)

(0.354)

(0.268)

3,009

1,062

1,947

R-squared
0.0641
0.1623
0.093
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist. All plots on
which organic and or inorganic fertilizer is used are dropped. All specifications include household-year-crop fixed
effects. Top and bottom 1 percent of plot size and outcome variables dropped from sample. FCFA: Franc
CommunautÊ Financière Africaine. The exchange rate at the survey times was approximately $1US=500 FCFA

70

APPENDIX C GENDER BIAS IN LABOR ALLOCATION AND CROP YIELD
Table 1C. 1 Log of labor allocation estimation across collective and all individual plots (including female plots)
Total
Male
Child
VARIABLES
labor
labor
Female labor labor
Collective=1
Female plot manager=1
Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

Clay and sandy soil (clay is omitted)
Sandy
Gravel
Observations

0.521***

0.410***

0.698***

0.951***

(0.031)

(0.040)

(0.054)

(0.058)

-0.03

-1.917***

1.303***

-0.280***

(0.032)

(0.052)

(0.060)

(0.062)

-0.384***

-0.257***

-0.389***

-0.232***

(0.012)

(0.020)

(0.017)

(0.022)

0.159***

0.147***

0.101***

0.167***

(0.018)

(0.026)

(0.028)

(0.032)

0.041**

0.063**

0.069**

0.044

(0.021)

(0.031)

(0.032)

(0.039)

0.055

0.177**

0.025

0.15

(0.058)

(0.081)

(0.084)

(0.103)

0.129***

0.071*

0.115***

0.200***

(0.028)

(0.042)

(0.040)

(0.051)

-0.001

-0.044

-0.016

0.021

(0.024)

(0.037)

(0.037)

(0.049)

-0.007

-0.128*

0.04

0.084

(0.057)

(0.068)

(0.086)

(0.102)

46,358

46,358

46,358

46,358

R-squared
0.1718
0.4467
0.2581
0.1229
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective, male, and female plots coexist. All
specifications include household-year-crop fixed effects. Top and bottom 1 percent of plot size and outcome
variables dropped from sample

71

Table 1C. 2 Log of crop yield (FCFA/ha) estimation across collective plots and all individual plots (including
female plots)
VARIABLES

All crops

Cereal crops

Non-cereal crops

Collective=1

0.341***

0.549***

0.338***

(0.047)

(0.090)

(0.057)

-0.391***

-0.684***

-0.317***

(0.049)

(0.087)

(0.059)

-0.158***

-0.231***

-0.283***

(0.027)

(0.044)

(0.043)

0.308***

0.371***

0.155

(0.060)

(0.068)

(0.100)

0.169***

0.145*

0.047

(0.063)

(0.075)

(0.105)

-0.049

-0.007

-0.009

(0.085)

(0.169)

(0.093)

0.013

0.002

0.007

(0.057)

(0.080)

(0.074)

-0.012

0.015

-0.038

(0.054)

(0.097)

(0.063)

-0.059

-0.02

-0.136

(0.165)

(0.233)

(0.208)

31,410

18,349

13,061

Female plot manager=1
Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

Clay and sandy soil (clay is omitted)

Sandy
Gravel
Observations

R-squared
0.1199
0.2662
0.0861
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective, male and female individual plots coexist. All
specifications include household-year-crop fixed effects. Top and bottom 1 percent of plot size and outcome
variables dropped from sample. FCFA: Franc CommunautÊ Financière Africaine. The exchange rate at the survey
times was approximately $1US=500 FCFA

72

APPENDIX D LABOR HETEROGENEITY BY PLOT MANAGEMENT STRUCTURE
Labor heterogeneity estimation results (from equation 3) using the sample restricted to households within
which collective and male individual plots coexist are presented in Tables 1D.1 and 1D.2. Since the key
variable is Collective, results are presented when only that variable is included (Table 1D.1) before
adding the other plot covariates (Table 1D.2). In addition, results are presented separately for all crops,
and grain and non-grain crops.
There is strong evidence that collective plots are more likely to receive more heterogeneous labor
relative to individual plots regardless of whether plot covariates are omitted or included, though the
magnitude of the coefficients are relatively smaller when more controls are added. The results
disaggregated by crop type reveal that grain crops are more likely to receive all combination of labor
compared to non-grain crops, which is consistent with the descriptive statistics in section 1.5. This
suggests that collective farming is practiced to secure the availability of these crops. Since grain crops
constitute the major source of livelihood for these households, the results hint that the latter engage in
collective farming to guarantee food security among themselves.48

48

These results can be extended to the perception that labor heterogeneity allows to achieve higher crop yield
performance. However, the empirical examination of this perception is beyond the scope of this paper.

73

Table 1D. 1 Linear probability results of labor heterogeneity across collective and individual plots (without plot covariates)
All crops
Grain crops
Non-grain crops
All
Male and
Male and
All
Male and
Male and
All
Male and
Male and
VARIABLES sources
female
child
sources
female
child
sources
female
child
x
Collective=1
0.221*** 0.075***
0.206***
0.234*** 0.126***
0.201***
0.211*** 0.037**
0.209***
-0.021
-0.015
-0.02
-0.03
-0.023
-0.029
-0.026
-0.018
-0.024
HHCYFE
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations 6,907
6,907
6,907
4,106
4,106
4,106
2,801
2,801
2,801
R-squared
0.111
0.023
0.110
0.125
0.055
0.111
0.101
0.006
0.110
Plot level observations. Robust clustered standard error at the plot level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to households within
which collective and male individual plots coexist. All plots receive male labor because these plots are all managed by male farmers. Top and bottom 1 percent
of plot size dropped from sample

Table 1D. 2 Linear probability results of labor heterogeneity across collective and individual plots (controlling for plot covariates)
All crops
Grain crops
Non-grain crops
Male and
Male and
Male and
Male and
All
Male and Male and
VARIABLES
All sources female
child
All sources
female
child
sources
female
child
Collective=1
Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1

Household owns
plot=1
Clay & sandy soil
(clay is omitted)
Sandy

0.167***
(0.024)
0.053***
(0.014)
-0.029
(0.018)
0.017
(0.023)

0.042**
(0.018)
0.035***
(0.012)
-0.004
(0.015)
0.019
(0.020)

0.167***
(0.023)
0.037***
(0.013)
-0.01
(0.017)
0.019
(0.022)

0.182***
(0.034)
0.042***
(0.015)
-0.011
(0.017)
0.011
(0.021)

0.106***
(0.026)
0.016
(0.014)
-0.007
(0.015)
0.009
(0.019)

0.159***
(0.033)
0.033**
(0.014)
0.009
(0.017)
0.009
(0.020)

0.135***
(0.034)
0.087***
(0.032)
-0.102*
(0.061)
0.047
(0.077)

-0.011
(0.026)
0.068***
(0.024)
0.013
(0.050)
0.029
(0.061)

0.156***
(0.032)
0.055*
(0.029)
-0.096*
(0.053)
0.083
(0.072)

0.069
(0.073)

-0.044
(0.060)

0.111
(0.071)

-0.014
(0.077)

-0.034
(0.078)

0.056
(0.078)

0.164
(0.123)

-0.049
(0.081)

0.177
(0.114)

0.018
(0.031)
0.018

-0.002
(0.024)
0.014

0.014
(0.028)
-0.007

-0.013
(0.031)
0.015

0.004
(0.027)
0.005

-0.026
(0.025)
-0.017

0.061
(0.061)
0.025

-0.008
(0.045)
0.03

0.069
(0.057)
0.01

74

Table 1D.2 (cont’d)
(0.026)
(0.022)
(0.025)
(0.026)
(0.024)
(0.024)
(0.053)
(0.041)
(0.051)
0.019
0.023
-0.009
-0.029
-0.04
0.018
0.083
0.125
-0.072
(0.052)
(0.047)
(0.059)
(0.052)
(0.049)
(0.042)
(0.093)
(0.083)
(0.122)
HHCYFE
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
6,907
6,907
6,907
4,106
4,106
4,106
2,801
2,801
2,801
R-squared
0.124
0.034
0.120
0.138
0.060
0.122
0.123
0.030
0.129
Plot level observations. Robust clustered standard error at the plot level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to households within
which collective and male individual plots coexist. All plots receive male labor because these plots are all managed by male farmers. Top and bottom 1 percent
of plot size dropped from sample
Gravel

75

APPENDIX E ROBUSTNESS CHECK USING DECILES OF PLOT SIZE
Table 1E. 1 Log of labor allocation estimation across collective and individual plots using plot deciles
VARIABLES

Total labor

Male labor

Female labor

Child labor

collective

0.401***

0.323***

0.498***

0.798***

(0.039)

(0.043)

(0.063)

(0.080)

-0.213***

-0.239***

0.02

0.007

(0.066)

(0.075)

(0.129)

(0.128)

-0.465***

-0.510***

-0.188

-0.213

(0.070)

(0.080)

(0.132)

(0.136)

-0.627***

-0.722***

-0.270*

-0.304*

(0.086)

(0.096)

(0.160)

(0.167)

-0.679***

-0.706***

-0.429***

-0.251*

(0.078)

(0.089)

(0.148)

(0.151)

-0.753***

-0.789***

-0.430***

-0.323**

(0.079)

(0.092)

(0.149)

(0.158)

-0.746***

-0.784***

-0.406**

-0.281

(0.089)

(0.100)

(0.157)

(0.178)

-0.835***

-0.826***

-0.514***

-0.404**

(0.091)

(0.106)

(0.166)

(0.175)

0.094***

0.116***

0.015

0.052

(0.036)

(0.040)

(0.058)

(0.062)

0.007

-0.009

0.122*

0.039

(0.041)

(0.045)

(0.073)

(0.075)

0.124

0.076

-0.157

0.477**

(0.143)

(0.142)

(0.228)

(0.225)

0.068

0.037

-0.011

0.202**

(0.054)

(0.053)

(0.098)

(0.102)

-0.068

-0.027

-0.127

-0.041

(0.050)

(0.053)

(0.082)

(0.087)

0.035

-0.026

-0.063

0.299

(0.115)

(0.109)

(0.177)

(0.185)

6,897

6,897

6,897

6,897

2.surface_decile
4.surface_decile

6.surface_decile
7.surface_decile
8.surface_decile
9.surface_decile
10.surface_decile
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1
Clay and sandy soil (clay is
omitted)
Sandy
Gravel
Observations

R-squared
0.116
0.092
0.053
0.115
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist. All
specifications include household-year-crop fixed effects. Top and bottom 1 percent of plot size and outcome
variables dropped from sample

76

Table 1E. 2 Log of crop yield estimation across collective and individual plots using plot deciles
VARIABLES

All crops

Cereal crops

Non-cereal crops

collective

0.308***

0.355***

0.326***

(0.055)

(0.093)

(0.068)

-0.148

-0.541***

-0.05

(0.097)

(0.204)

(0.104)

-0.325***

-0.433**

-0.348***

(0.108)

(0.207)

(0.124)

-0.351*

-0.753***

-0.196

(0.205)

(0.245)

(0.295)

-0.552***

-0.394

-0.746***

(0.156)

(0.288)

(0.171)

-0.516***

-0.513***

-0.667***

(0.166)

(0.192)

(0.247)

-0.631***

-0.627***

-1.160***

(0.169)

(0.229)

(0.317)

-0.687***

-0.800***

-1.081***

(0.176)

(0.240)

(0.373)

0.238**

0.357***

0.165

(0.108)

(0.134)

(0.166)

0.034

-0.042

-0.08

(0.107)

(0.129)

(0.189)

0.136

0.044

0.166

(0.188)

(0.208)

(0.268)

-0.099

0.299*

-0.306***

(0.094)

(0.160)

(0.112)

-0.264***

-0.211

-0.321***

(0.091)

(0.173)

(0.107)

-0.055

-0.426

-0.12

(0.173)

(0.282)

(0.255)

5,062

2,824

2,238

2.surface_decile
4.surface_decile
6.surface_decile
7.surface_decile
8.surface_decile
9.surface_decile
10.surface_decile
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

Clay and sandy soil (clay is omitted)

Sandy
Gravel
Observations

R-squared
0.0651
0.1326
0.1173
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist. All
specifications include household-year-crop fixed effects. Top and bottom 1 percent of plot size and outcome
variables dropped from sample. FCFA: Franc CommunautÊ Financière Africaine. The exchange rate at the survey
times was approximately $1US=500 FCFA. Fewer number of observations relative to Table 1E.1 because yield
variables are at the crop-level

77

APPENDIX F ROBUSTNESS CHECK WITH OVERLAPPING PLOT SIZE
Table 1F. 1 Log of labor allocation estimation across plots with overlapping plot size
VARIABLES

Total labor

Male labor

Female labor

Child labor

Collective=1

0.380***

0.293***

0.485***

0.791***

(0.039)

(0.044)

(0.063)

(0.081)

-0.256***

-0.260***

-0.175***

-0.133***

(0.024)

(0.027)

(0.042)

(0.046)

0.117***

0.138***

0.044

0.055

(0.037)

(0.040)

(0.056)

(0.063)

-0.008

-0.02

0.098

0.017

(0.042)

(0.044)

(0.075)

(0.078)

0.096

0.052

-0.188

0.457**

(0.139)

(0.138)

(0.226)

(0.228)

0.088

0.063

-0.012

0.206*

(0.055)

(0.055)

(0.099)

(0.107)

-0.082

-0.034

-0.180**

-0.041

(0.052)

(0.056)

(0.084)

(0.090)

0.045

-0.014

-0.072

0.31

(0.120)

(0.113)

(0.184)

(0.195)

6,699

6,699

6,699

6,699

Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

Clay and sandy soil
(clay is omitted)

Sandy
Gravel

Observations

R-squared
0.1064
0.0796
0.0507
0.1138
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist and plots
with overlapping plot size. All specifications include household-year-crop fixed effects. Top and bottom 1 percent of
plot size and outcome variables dropped from sample.

Table 1F. 2 Log of crop yield (FCFA/ha) estimation across plots with overlapping plot size
VARIABLES

All crops

Cereal crops

Non-cereal crops

Collective=1

0.309***

0.387***

0.310***

(0.057)

(0.101)

(0.071)

-0.220***

-0.181**

-0.309***

(0.052)

(0.071)

(0.079)

0.232**

0.282**

0.175

(0.115)

(0.128)

(0.185)

-0.008

-0.08

-0.05

(0.109)

(0.125)

(0.206)

0.102

0.018

0.164

(0.197)

(0.196)

(0.263)

Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

78

Table 1F.2 (cont’d)
Clay and sandy soil
(clay is omitted)

Sandy
Gravel

Observations

-0.112

0.287*

-0.303***

(0.097)

(0.167)

(0.115)

-0.251***

-0.257

-0.293***

(0.094)

(0.203)

(0.108)

0.024

-0.247

0.111

(0.166)

(0.314)

(0.232)

4,890

2,688

2,202

R-squared
0.0644
0.1021
0.0842
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist and plots
with overlapping plot size. All specifications include household-year-crop fixed effects. Top and bottom 1 percent of
plot size and outcome variables dropped from sample. FCFA: Franc CommunautÊ Financière Africaine. The
exchange rate at the survey times was approximately $1US=500 FCFA. Fewer number of observations relative to
Table 1F.1 because yield variables are at the crop-level

79

APPENDIX G ROBUSTNESS CHECK BY DROPPING PLOTS WITH COMPANION
CROPS
Table 1G. Log of labor allocation estimation after dropping all plots with companion crops
VARIABLES

Total labor

Male labor

Female labor

Child labor

Collective=1

0.343***

0.277***

0.388***

0.733***

(0.046)

(0.052)

(0.079)

(0.097)

-0.281***

-0.301***

-0.166***

-0.115*

(0.034)

(0.037)

(0.054)

(0.059)

0.147**

0.192***

-0.007

0.07

(0.059)

(0.062)

(0.095)

(0.104)

0.036

-0.021

0.239**

0.05

(0.065)

(0.068)

(0.107)

(0.126)

0.117

0.075

-0.292

0.600**

(0.172)

(0.175)

(0.267)

(0.268)

0.121

0.072

0.002

0.308**

(0.077)

(0.077)

(0.141)

(0.149)

-0.052

-0.04

-0.104

-0.002

(0.067)

(0.072)

(0.120)

(0.117)

0.082

0.01

0.188

0.22

(0.166)

(0.163)

(0.283)

(0.251)

4,906

4,906

4,906

4,906

Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1
Clay and sandy soil (clay is
omitted)
Sandy
Gravel

Observations

R-squared
0.1087
0.1022
0.0397
0.1107
Plot level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05;
*** p<0.01. Sample restricted to households within which collective and male individual plots coexist. All plots
with companion crops are dropped. All specifications include household-year-crop fixed effects. Top and bottom 1
percent of plot size and outcome variables dropped from sample.

80

APPENDIX H ESTIMATION OF THE LOG OF CROP YIELD ACROSS COLLECTIVE AND HEADS’ INDIVIDUAL PLOTS
Table 1H. Estimation of the log of crop yield (FCFA/ha) across collective and heads’ individual plot using PMYFE
Panel A: All crops
Panel B: Grain crops

Collective=1
Log of plot size
Organic fertilizer=1
Inorganic fertilizer==1
Household owns plot=1

Clay and sandy soil
(clay is omitted)
Sandy
Gravel

Panel C: Non-grain crops

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

0.258***
(0.08)
-0.212***
(0.05)

0.249***
(0.07)
-0.228***
(0.05)
0.116
(0.08)
0.157*
(0.09)
-0.244
(0.29)

0.253***
(0.07)
-0.231***
(0.05)
0.112
(0.08)
0.155*
(0.09)
-0.238
(0.30)

0.357***
(0.12)
-0.197***
(0.06)

0.331***
(0.12)
-0.211***
(0.06)
0.097
(0.08)
0.234**
(0.11)
0.252**
(0.12)

0.329***
(0.12)
-0.207***
(0.06)
0.095
(0.08)
0.234**
(0.11)
0.195
(0.17)

0.253**
(0.10)
-0.374***
(0.10)

0.268***
(0.10)
-0.380***
(0.10)
0.245
(0.20)
-0.014
(0.23)
-0.451
(0.39)

0.268***
(0.10)
-0.384***
(0.10)
0.255
(0.21)
0.001
(0.23)
-0.412
(0.37)

0.096
(0.08)
0.095
(0.10)
0.046
(0.16)
Yes
Yes
1352
0.247

0.098
(0.11)
0.02
(0.15)
-0.257
(0.23)
Yes
Yes
766
0.154

-0.02
(0.18)
0.164
(0.17)
0.192
(0.28)
Yes
Yes
586
0.432

PMFE?
Yes
Yes
Yes
Yes
Yes
Yes
Crop FE?
Yes
Yes
Yes
Yes
Yes
Yes
Observations
1352
1352
766
766
586
586
R-squared
0.239
0.246
0.133
0.150
0.425
0.429
Sample Mean (SD)
(Mean=11.735; SD=1.080)
(Mean=11.605; SD=.926)
(Mean=11.904; SD=1.233)
Crop level observations. Robust clustered standard error at the household level in parentheses. * p<0.1; ** p<0.05; *** p<0.01. Sample restricted to heads that
control simultaneously their own individual plots and collective plots. Top and bottom 1 percent of plot size and outcome variables dropped from sample. Village
level prices recorded through community questionnaires were used to convert production into monetary value FCFA: Franc CommunautÊ Financière Africaine.
The exchange rate at the survey times was approximately $1US=500 FCFA.

81

APPENDIX I YIELD DISPERSION ACROSS COLLECTIVE AND INDIVIDUAL PLOTS
A major underpinning argument in favor of collective farming as an ex ante risk mitigation strategy is that
it reduces the nuclear household’s exposure to production uncertainty. As a result, crop yield performance
should be less variable on collective plots relative to individual plots. Since the data used in this study are
from two rounds, to empirically test this assertion, I regress the residuals obtained from the householdcrop-year fixed effects estimation –that is, the crop yield performance across collective and individual
plots presented in Table 1I –on the time indicator. The central idea is that, if production is less variable
on collective plots, then the yield dispersion across the two survey rounds for group of the collective plots
should be smaller relative to that of the individual plots.
Table 1I. Yield variability across collective and male individual plots
(1)
(2)
Group 1:
Group 2:
VARIABLES
Collective=0
Collective=1
Year=2012

-0.479***

-0.181***

(0.056)

(0.032)

Plot covariates

Yes

Yes

Crop FE

Yes

Yes

Household FE

No

No

Observations

1,304

3,758

R-squared

0.3074

0.2085

chi2(1) = 26.10

26.10

26.10

Probability > chi2 = 0.0000
Sample restricted to households within which collective and male individual plots coexist. Residuals are obtained
from yield estimation across collective and individual plots (originally presented in Table 1.XX).

The results presented in Table 1I, show that the size of the coefficient of the year indicator for
individual plots is more than 2.5 times larger relative to the coefficient for the group of collective plots
(Table 1I, Column 2).
A Chow test, which tests whether the coefficients vary between the two groups indicates that this
difference is different from zero at the 1 percent statistical significance level with a corresponding chi2 of
one degree. This finding provides strong suggestive evidence that, collective farming does indeed allow
households to mitigate production variability across time. These findings are consistent with results from

82

previous work that find less variability in agricultural profit across times among intergenerational
households relative to households without elderly members (Rosenzweig and Wolpin, 1985).

83

REFERENCES

84

REFERENCES

Adams, W., & Mortimore, M. (1997). Agricultural Intensification and Flexibility in the Nigerian Sahel.
The Geographical Journal, 163(2), 150-160.
Beaman, L. and Dillon, A. (2012). Do household definitions matter in survey design? Results from a
randomized survey experiment in Mali. Journal of Development Economics, 98(1), pp.124-135.
Bharadwaj, P. (2015). Fertility and rural labor market inefficiencies: Evidence from India. Journal of
development Economics, 115, 217-232.
Barrett, C. B., Moser, C. M., McHugh, O. V., & Barison, J. (2004). Better technology, better plots, or
better farmers? Identifying changes in productivity and risk among Malagasy rice farmers. American
Journal of Agricultural Economics, 86(4), 869-888.
Barrett, C. B., Sherlund, S. M., & Adesina, A. A. (2008). Shadow wages, allocative inefficiency, and
labor supply in smallholder agriculture. Agricultural Economics, 38(1), 21-34.
Becker, L. (1996). Access to labor in rural Mali. Human Organization, 55(3), 279-288.
Benjamin, D. (1992). Household composition, labor markets, and labor demand: testing for separation in
agricultural household models. Econometrica: Journal of the Econometric Society, 287-322.
Binswanger, H. P., & Deininger, K. (1997). Explaining agricultural and agrarian policies in developing
countries. Journal of Economic Literature, 35(4), 1958-2005.
Boserup, E. (1970). Women's role in economic development.
Browning, M., & Chiappori, P. A. (1998). Efficient intra-household allocations: A general
characterization and empirical tests. Econometrica, 1241-1278.
Burton, M. L., & White, D. R. (1984). Sexual division of labor in agriculture. American Anthropologist,
86(3), 568-583.
Carter, M. R. (1984). Resource allocation and use under collective rights and labour management in
Peruvian coastal agriculture. The Economic Journal, 94(376), 826-846.
Carter, M. R. (1987). Risk sharing and incentives in the decollectivization of agriculture. Oxford
Economic Papers, 39(3), 577-595.
Cooper, P. J. M., Dimes, J., Rao, K. P. C., Shapiro, B., Shiferaw, B., & Twomlow, S. (2008). Coping
better with current climatic variability in the rain-fed farming systems of sub-Saharan Africa: An
essential first step in adapting to future climate change? Agriculture, Ecosystems & Environment,
126(1), 24-35.
Deininger, K. (1995). Collective agricultural production: A solution for transition economies? World
Development, 23(8), 1317-1334.
Dillon, B., & Barrett, C. B. (2017). Agricultural factor markets in Sub-Saharan Africa: an updated view
with formal tests for market failure. Food Policy, 67, 64-77.

85

Dillon, A., Mueller, V., & Salau, S. (2011). Migratory responses to agricultural risk in northern Nigeria.
American Journal of Agricultural Economics, 93(4), 1048-1061.
Fafchamps, M. (1993). Sequential labor decisions under uncertainty: An estimable household model of
West-African farmers. Econometrica: Journal of the Econometric Society, 1173-1197.
Feng, S., Heerink, N., Ruben, R., & Qu, F. (2010). Land rental market, off-farm employment and
agricultural production in Southeast China: A plot-level case study. China Economic Review, 21(4),
598-606.
Foltz, J. D. (2004). Credit market access and profitability in Tunisian agriculture. Agricultural
Economics, 30(3), 229-240.
Food and Agriculture Organization (2006). Country Pasture/Forage Resource Profiles
http://www.fao.org/ag/agp/agpc/doc/counprof/Mali/mal.
Foster, A. D., & Rosenzweig, M. R. (1994). A test for moral hazard in the labor market: Contractual
arrangements, effort, and health. The Review of Economics and Statistics, 213-227.
Gebremedhin, Berhanu, and Scott M. Swinton. "Investment in soil conservation in northern Ethiopia: the
role of land tenure security and public programs." Agricultural Economics 29, no. 1 (2003): 69-84.
Gianessi, L. P. (2013). The increasing importance of herbicides in worldwide crop production. Pest
management science, 69(10), 1099-1105.
Goldstein, M., & Udry, C. (2008). The profits of power: Land rights and agricultural investment in
Ghana. Journal of Political Economy, 116(6), 981-1022.
Guirkinger, C., & Platteau, J. P. (2014). The effect of land scarcity on farm structure: empirical evidence
from Mali. Economic Development and Cultural Change, 62(2), 195-238.
Guirkinger, C., Platteau, J. P., & Goetghebuer, T. (2015). Productive inefficiency in extended agricultural
households: Evidence from Mali. Journal of Development Economics, 116, 17-27.
Haggblade, S., Minten, B., Pray, C., Reardon, T., & Zilberman, D. (2017). The Herbicide Revolution in
Developing Countries: Patterns, Causes, and Implications. The European Journal of Development
Research, 1-27.
Haggblade, S., Smale, M., Theriault, V., & Assima, A. (2017). Causes and consequences of increasing
herbicide use in Mali. The European Journal of Development Research, 1-27.
Hayes, J., Roth, M., & Zepeda, L. (1997). Tenure security, investment and productivity in Gambian
agriculture: A generalized probit analysis. American Journal of Agricultural Economics, 79(2), 369382.
Israelsen, L. D. (1980). Collectives, communes, and incentives. Journal of Comparative Economics, 4(2),
99-124.
Jack, B. K. (2013). Market inefficiencies and the adoption of agricultural technologies in developing
countries.

86

Kazianga, H., & Wahhaj, Z. (2013). Gender, Social Norms, and Household Production in Burkina Faso.
Economic Development and Cultural Change, 61(3), 539-576.
Kazianga, H., Wahhaj, Z., 2016. Intra-household Resource Allocation and Familial Ties. Oxford CSAE
Working Paper WPS 2016-06.
Kazianga, H., & Wahhaj, Z. (2017). Intra-household resource allocation and familial ties. Journal of
Development Economics, 127, 109-132.
Lamb, R. L. (2003). Inverse productivity: Land quality, labor markets, and measurement error. Journal of
Development Economics, 71(1), 71-95.
Lee, D. R., Barrett, C. B., & McPeak, J. G. (2006). Policy, technology, and management strategies for
achieving sustainable agricultural intensification. Agricultural economics, 34(2), 123-127.
Lin, J. Y. (1987). Household Farm, Cooperative Farm, and Efficiency: Evidence from Rural Decollectivization in China. Economic Growth Center, Yale University.
McMillan, J., Whalley, J., & Jing, Z. L. (1987). Incentive effects of price rises and payment-system
changes on Chinese agricultural productivity growth.
Meyer, C. A. (1989). Agrarian reform in the Dominican Republic: An associative solution to the
collective/individual dilemma. World Development, 17(8), 1255-1267.
Mathijs, E., & Swinnen, J. F. (2001). Production organization and efficiency during transition: An
empirical analysis of East German agriculture. The Review of Economics and Statistics, 83(1), 100107.
Mishra, A., Hansen, J. W., Dingkuhn, M., Baron, C., TraorĂŠ, S. B., Ndiaye, O., & Ward, M. N. (2008).
Sorghum yield prediction from seasonal rainfall forecasts in Burkina Faso. Agricultural and forest
meteorology, 148(11), 1798-1814.
Moser, C. M., & Barrett, C. B. (2003). The disappointing adoption dynamics of a yield-increasing, low
external-input technology: the case of SRI in Madagascar. Agricultural Systems, 76(3), 1085-1100.
Ondersteijn, C. J. M., Giesen, G. W. J., & Huirne, R. B. M. (2003). Identification of farmer
characteristics and farm strategies explaining changes in environmental management and
environmental and economic performance of dairy farms. Agricultural Systems, 78(1), 31-55.
Putterman, L., & DiGiorgio, M. (1985). Choice and efficiency in a model of democratic semi-collective
agriculture. Oxford Economic Papers, 37(1), 1-21.
Qaim, M., Subramanian, A., Naik, G., & Zilberman, D. (2006). Adoption of Bt cotton and impact
variability: Insights from India. Applied Economic Perspectives and Policy, 28(1), 48-58.
Quisumbing, A. R. (1996). Male-female differences in agricultural productivity: Methodological issues
and empirical evidence. World Development, 24(10), 1579-1595.
Quisumbing, A. R., & Pandolfelli, L. (2010). Promising approaches to address the needs of poor female
farmers: Resources, constraints, and interventions. World Development, 38(4), 581-592.

87

Ricker-Gilbert, J., Jumbe, C., & Chamberlin, J. (2014). How does population density influence
agricultural intensification and productivity? Evidence from Malawi. Food Policy, 48, 114-128.
Rosenzweig, M. R., & Wolpin, K. I. (1985). Specific experience, household structure, and
intergenerational transfers: farm family land and labor arrangements in developing countries. The
Quarterly Journal of Economics, 961-987.
Shorr, N. (2000). Early utilization of flood-recession soils as a response to the intensification of fishing
and upland agriculture: Resource-use dynamics in a large Tikuna community. Human Ecology, 28(1),
73-107.
Smale, M., Byerlee, D., & Jayne, T. (2013). Maize revolutions in sub-Saharan Africa. In an African green
revolution (pp. 165-195). Springer Netherlands.
Steiner, K. G. (1998). Using farmers' knowledge of soils in making research results more relevant to field
practice: experiences from Rwanda. Agriculture, ecosystems & environment, 69(3), 191-200.
Stone, M., Stone, G., & Robert McC. Netting. (1995). The Sexual Division of Labor in Kofyar
Agriculture. American Ethnologist, 22(1), 165-186. Retrieved from
http://www.jstor.org/stable/646051.
Sultan, B., Baron, C., Dingkuhn, M., Sarr, B., & Janicot, S. (2005). Agricultural impacts of large-scale
variability of the West African monsoon. Agricultural and forest meteorology, 128(1), 93-110.
Teklewold, H., Kassie, M., & Shiferaw, B. (2013). Adoption of multiple sustainable agricultural practices
in rural Ethiopia. Journal of Agricultural Economics, 64(3), 597-623.
Theriault, V., Smale, M., & Haider, H. (2017). How Does Gender Affect Sustainable Intensification of
Cereal Production in the West African Sahel? Evidence from Burkina Faso. World Development, 92,
177-191.
Udry, C. 1996. “Gender, Agricultural Production, and the Theory of the Household.” Journal of Political
Economy, 104(5), 1010-1046.
Udry, C., Hoddinott, J., Alderman, H., & Haddad, L. (1995). Gender differentials in farm productivity:
implications for household efficiency and agricultural policy. Food Policy, 20(5), 407-423.
USAID. (2010). Mali Land Tenure Assessment Report.
http://fsg.afre.msu.edu/promisam_2/references/ARD_2010_Mali_Land_Tenure_Assessment.pdf.
USAID (2008). Plan de SĂŠcuritĂŠ Alimentaire Commune Urbaine de SĂŠgou 2008-2012, Commissariat Ă  la
SĂŠcuritĂŠ Alimentaire, RĂŠpublique du Mali, USAID-Mali, 2008.
World Bank (2016). Country Overview, Mali. http://www.worldbank.org/en/country/mali/overview.
World Bank (2014). LSMS-ISA-Mali.
World Culture Encyclopedia. http://www.everyculture.com/Ja-Ma/Mali.html.

88

2

2.1

HOUSEHOLD STRUCTURE INFLUENCES LABOR ALLOCATION DECISIONS DUE
TO AGRICULTURAL TECHNOLOGY ADOPTION: EVIDENCE FROM BURKINA
FASO
Introduction

Food security has become more tenuous over recent decades because of a combination of worsening soil
fertility and reduced arable land per capita from rising population. Because of these pressures, farmers
must rely on agricultural intensification to ensure sustainable food security (Lele and Stone, 1989; Garnett
et al., 2013). Several programs have been implemented to promote the adoption of agricultural
technologies among farmers (e.g., fertilizer and improved seed subsidies). However, though some
progress has been observed in some regions, especially with the “early green revolution” in Latin
America and Asia (Evenson and Gollin, 2003), low adoption rates of such technologies tend to persist
among subsistence farmers, especially in sub-Saharan Africa.
This has fueled numerous studies to investigate the determinants of agricultural technology
adoption among farmers (Just and Zilberman, 1983; Foster and Rosenzweig, 1995; Moser and Barrett,
2003; Croppenstedt et al., 2003; Barrett et al., 2004; Moreno and Sunding, 2005; Langyintuo and
Mungoma, 2008; Gine and Young, 2009; Duflo and Kremer, 2011; Kijima et al., 2011; Noltze et al.,
2012; Jack, 2013; Khonje et al., 2015; among others). Much of the empirical research has found strong
evidence that the household endowment in labor is a major contributing factor (Croppenstedt et al., 2003;
Moser and Barrett, 2003; Barrett et al., 2004), especially for labor intensive technologies (e.g., cite some
of the technologies). This highlights the importance of investigating how households respond in terms of
their labor allocation to the introduction of a new labor-intensive agricultural technology, especially in a
setting where incomplete or missing labor markets are prevalent.
In this article, we investigate household labor response to the introduction of a fertilizer
application technique called microdosing among rural subsistence farmers in Burkina Faso. The
microdosing technique directly targets the seed-hole and requires much less fertilizer compared to
broadcasting methods. This makes it less financially constraining and a climate-smart technology

89

(Resnick et al., 2012; Aune and Coulibaly, 2015). When applied properly, it has the potential to increase
yield by more than 20 percent relative to the broadcasting technique. But the downside is its intensive
labor requirements (ICRISAT, 2012). These features of microdosing, especially its labor-intensive nature,
make it suitable for exploring how household labor allocation is affected by programs encouraging
agricultural intensification.
We first investigate the household labor response by the household’s endowment in land.
Household land holding, which is a proxy for wealth, has been proven to be an important determinant of
agricultural technology adoption (Just and Zilberman, 1983; Foster and Rosenzweig, 1985; Moreno and
Sunding, 2005; Langyintuo and Mungoma, 2008; Khonje et al., 2015). Larger land holders are likely to
test a new technology using a portion of their land, while small land holders may be skeptical to adopt an
unproven technology, especially, if it is to be applied to the production of a crop that constitutes their
main source of livelihood (Langyintuo et al., 2008).
In ruling out differential treatment effects of microdosing by landholdings, we focus on how labor
allocation responds to the introduction of the technology depending on the structure of the household
(nuclear vs. extended family households).
Why is it important to investigate the differential impact of the microdosing program by
household structure? First, with increasing fragmentation of households to smaller units in the developing
world, mixed household structures – that is, the coexistence of nuclear family households and extended
family households – are predominant in in many sub-Saharan African countries (Bongaarts 2001 and
West, 2010).
Second, the majority of the households in these regions face missing or imperfect labor markets,
(De Janvry et al., 1991; Barrett et al., 2004) and therefore heavily rely on family members as the main
source of their agricultural labor (Udry, 1996; Doss, 2006; Marenya and Barrett, 2007; Morton, 2007;
Kazianga and Wahhaj, 2013).
Third, we expect differences in production organization systems between nuclear and extended
family households would lead to differences in labor response to the introduction of microdosing.

90

Extended households, comprised of multiple nuclear families, differ from nuclear households in that they
have more adult labor available to them. Yet, given their larger size, labor allocation decisions among
extended households could inhibit adoption if coordination problems create high transaction costs
compared to nuclear households (Poteete and Ostrom, 2004). Furthermore, extended family members
simultaneously manage private plots owned by nuclear family members and collective plots, which are
owned by all extended family members. Private plot managers are required to provide labor on the
collective plots before working on their own plots (Becker, 1996; Kazianga and Wahhaj, 2013,
Guirkinger and Platteau, 2014, Guirkinger et al., 2015).
We address potential endogeneity in the decision to adopt microdosing by exogenously varying
the availability of this technology.49 To do so, we implemented a Randomized Control Trial (RCT) where
the treatment included training in the microdosing application technique, along with a starter kit
comprised of certified sorghum seed and fertilizer. We then estimate the Intention To Treat (ITT) effects
with Difference-in-Difference (DID) applied to pre- and post- program intervention data.
Our findings provide evidence of a differential impact of the microdosing program by household
structure, and no indication of any heterogeneous impact by land holding. Specifically, with the sample
restricted to the bottom three quintiles of baseline total landholding, we find that total labor allocated to
sorghum production increased by approximately 16 percent among nuclear households assigned to
treatment (about 15 person-days per hectare), whereas the corresponding figure for extended households
was a reduction of about 29 percent (approximately 33 person-days per hectare).
We attribute the heterogeneous impact of the microdosing by household structure to three factors.
First, we show that the increased labor requirement for fertilizer application induced by the microdosing
intervention led extended households to substitute labor away from other sorghum production activities,
while there is no indication of any substitutions effects of labor among the nuclear households. Second,

49

For example, household characteristics such as wealth and household composition may determine technology
adoption (Croppenstedt et al., 2003; Teklewold et al., 2013). Other endogenous determinants of technology adoption
include productivity differences across households (McMillan, 1987; Doss, 2001, Mendola, 2008; Foster and
Rosenzweig, 2010).

91

we argue that the program provided higher incentives for nuclear households to exert more effort in their
sorghum production because they are the sole claimants of the potential accrued production. Finally, we
present suggestive evidence that differences in resource control is a likely contributing source to the
differential impact of the program across nuclear and extended households.
These findings have important implications for agricultural development policy interventions in
developing countries with heterogeneity in household structures –where some households are formed by
nuclear family members while others are composed by extended family members. Our study also
highlights the need for a better understanding of how households may differ in their agricultural
production organization, depending on household composition. Doing so would enable policy makers to
design more inclusive agricultural development intervention programs and, consequently, for increased
overall agricultural productivity.
The remainder of the paper is structured as follows. We present the conceptual framework where
we describe our experimental design, the microdosing technology as well as how we classify households
according to nuclear or extended categories. In Section 3, we describe the difference-in-difference
methods used to identify the intent to treat effect of the microdosing technology intervention. Section 4
provides a description of the data and sample restrictions. We also summarize differences in demographic
and production characteristics across nuclear and extended households, as well as balancing tests by
treatment assignment within each household category. Section 6 presents concludes the paper.

2.2
2.2.1

Conceptual framework
Experimental Design

The current work is part of a larger program implemented in Burkina Faso, with the objective to increase
food security among rural households. For these households as well as many rural households in the West
African Sahel, sorghum is the main food staple and most widely cultivated dryland crop. Yet, as in much
of this region, average sorghum yields in Burkina Faso are estimated at 0.8 tons per hectare, despite the
potential to attain over 2 tons per hectare (Ministry of Agriculture, Burkina Faso 2010). The specific

92

objective of the program was therefore to improve sorghum yields through the promotion of a particular
application technique of mineral fertilizer, called microdosing. The microdosing technique directly targets
the plant by applying small quantities (about 2 grams) of fertilizer in or near the seed hole during sowing
time, or as top dressing 3 to 4 weeks after the seeded crop begins to emerge. This technique lowers the
required amount of fertilizer per hectare by more than 50 percent (ICRISAT, Fertilizer Microdosing,
January 2012), making it less financially constraining relative to standard fertilizer application techniques.
In the case of hard soil, farmers dig small holes and fill them with manure before the rain begins. Once
the rain starts, the fertilizer and seeds are placed into the moist soil. This technique captures the water, so
that it does not run off the hard-crusted soil, thereby encouraging root growth (ICRISAT, January 2009).
When applied to the seeding of improved sorghum varieties, microdosing raises yields
considerably. In Burkina Faso, INERA (Institut de l'Environnement et de Recherches Agricoles) has
reported grain yields of nearly 2000 kg/ha for improved sorghum varieties. Because the microdosing
technique is climate-smart while lowering adoption barriers due to financial constraints, it has been
perceived as one of the promising ways to sustainably increase agricultural productivity (Montpellier
Panel Report, 2013, Aune and Coulibaly, 2015). Conversely, the primary drawbacks to microdosing are
that it is time-consuming, laborious, and it is difficult to ensure that the correct amount of fertilizer is used
for each dose.
Our experimental design used an encouragement scheme to randomly assign microdosing training
and a starter kit, which included fertilizer and certified sorghum seed sufficient for one half hectare of
sorghum production. We worked with AGRODIA, an organization of local agro-dealers in Burkina Faso,
to supply these micro packets. The seed was certified by INERA, the public agricultural research
institution in the country. INERA also provided training in microdosing to all households in treated
villages.
We selected three provinces for the study: Bam and Sanmatenga from the Center-North region
and PassorĂŠ from the North region. These provinces were selected because there is a high prevalence of
sorghum cultivation and little cultivation of cash crops such as cotton. We collected a complete list of all

93

925 villages in these three provinces. Because this study was part of a larger project which involved
collecting social network censuses in many of the villages, we restricted our potential pool of villages to
those that were not too large to conduct such censuses. We therefore kept villages where the population
size was between 190 and 800, provided that the number of households was no greater than 120
households. The number of households per village varied between 70 and 120 households. From the
remaining 226 villages, 80 villages were randomly selected for the treatment group studied here, and 20
villages were randomly selected for the control group. In all these villages, a village enumeration included
questions about plot information, sorghum production, and adoption of improved seeds.
Using this village enumeration, in each village we then randomly sampled approximately 30
households growing sorghum. For these households, we conducted baseline and follow-up household
surveys, where we collected detailed production and socio-economic information. Thus, detailed
household surveys were conducted for 2400 households in all 80 villages in the treatment group, and for
600 households in all 20 villages in the control group.
In each of 80 treated villages, approximately 15 of the 30 surveyed households were randomly
selected to receive free micro-packets of certified sorghum seed and fertilizer.50 In the current analysis,
we compare these treated households to the households surveyed in control villages which did not receive
any micro-packets or any training in micro-dosing. We exclude all households in treatment villages who
did not receive free micro-packets but may have received training in microdosing. We do so to eliminate
any potential spillover effects due to possible technological transfer, which in turn may lead to
underestimation of the program’s impact on treated households.
Finally, we do not impose compliance in our analysis as we are not able to capture actual take-up
rates of the practices of the microdosing technique. This is because many households consider other
existing fertilization techniques as microdosing. More importantly, self-declared take-up rates is more
likely to be inflated in the control villages relative to the treatment villages, given that the treatment

50

INERA reported that one village did not receive the intervention due difficulties in accessing the village. That
village has been dropped from the sample.

94

certainly increased knowledge of microdosing practice in the treatment villages. Though we are not able
to ensure clear differences in microdosing adoption among control and treatment groups, we are able to
observe whether the pattern of self-declared take-up rates are similar across treatment and control groups
by household structure as shown in Table 2A (see APPENDIX A)

2.2.2

Household Structure

We define a household as “a socio-economic unit within which one or more members, related or not, live
in the same house or concession, pool their resources, and jointly meet the bulk of their food and other
basic needs under the authority of one of them, called the head of household” (Beaman and Dillon, 2012).
This definition of the household may include multiple “nuclear” households cohabiting in one “extended”
household. Such cohabitation of multiple nuclear households may arise through two possible channels:
vertical extension of the household, where married sons and their nuclear families have common
residency with their fathers; and horizontal extension, where married brothers and their respective nuclear
families cohabitate (Laslett and Wall 1982; West 2010; Kazianga and Wahhaj 2017).
Therefore, we define an “extended” household as a household that meets at least one of the
following criteria: households having one married male who is not the household head; or household
heads having married sons, married brothers, daughters-in-law, or sisters-in-law living in the household
(Table 1, Panel A). Households that do not meet these criteria are classified as “nuclear” households.
According to this definition, 354 of the 1,059 households in our final sample (about 33 percent) are
extended households, and the remaining households are classified as nuclear households.
The leading criterion for being qualified as an extended household is the household having more
than one married male. This condition is met by about 29 percent of the final sample. When classified by
relationship to the head, approximately 16 percent of the heads have at least one married son and nearly 8
percent of them live with one or multiple married brothers. The father and married son relationship is
mirrored by approximately 16 percent of heads having at least one daughter-in-law. However, while only
about 9 percent of the heads cohabitate with married brothers, up to 13 percent of them have at least one

95

sister-in-law (Table 2.1, Panel A).

Table 2.1 Sample Restrictions and Household Structure
Sample restrictions

Number excluded

Number remaining

All households in baseline

0

1,522

All households in follow-up

0

1,428

Households in both surveys

102

1,420

Remaining male-headed HHs
Remaining male-headed HHs with at least one plot managed
by male

133

1,287

32

1,255

Remaining male-headed HHs producing sorghum at baseline

73

1,182

94

1,098

39

1,059

Remaining male-headed HHs producing sorghum both at
baseline and follow-up
Final sample after dropping outliers and missing values
Household structure

HH has at least one additional married male besides the head

Number of Households
Panel A: Final full
Panel B: Final
sample
Restricted sample
311
157

HH head has at least one married son

170

84

HH head has at least one married brother

89

43

HH head has at least one daughter-in-law

162

73

HH head has at least one sister-in-law
Extended HHs (HHs that meet at least one of the criteria
above)
Nuclear HHs (HHs that meet none of the criteria above)

133

64

354

181

705

496

Total number of HHs
1,059
677
Note: Observations are at the household level
Outliers identified as observations with values more than three times the standard deviation above the mean
Sorghum plot is identified as a plot on which sorghum is the main crop
Restricted sample is for households in the first three quintiles of land holding

Based on these criteria for the household structure classifications, extended household are
certainly different from nuclear households in terms of their demographic characteristics and their
landholding, as they are more likely to have larger landholding relative to the latter. In addition, given
their larger size, they are more likely to have a larger pool of labor but may also face some coordination
costs in productive factors allocation.
Finally, the microdosing intervention may generate different levels of incentives in terms of effort
allocated to sorghum production depending on household structure since output produced in the extended

96

households is shared amongst all extended family members. Precisely, nuclear households will be more
motivated to exert higher efforts in their sorghum production following the reception of the microdosing
kits because they get to keep all the marginal production that will originate from their marginal effort.
Conversely in the extended households, the sharing rule is based on need, rather than on effort.51 Based
on these factors, our main research question is: How does labor allocation respond differently across
households with differing family structure? To answer this question, we use the following general
equation:
𝑌𝑖 = 𝛼 + 𝛽1 𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙_𝑠𝑡𝑟𝑢𝑐𝑡𝑢𝑟𝑒 + 𝜷𝟐 𝑿𝒊 + 𝜖i

(2.1)

where 𝑌𝑖 is the labor allocation response to the introduction of the microdosing technology,
Household_structure is an indicator variable that takes 1 if an extended household and 0 otherwise, and
𝑿𝒊 is a vector of household and production characteristics. The key coefficient in equation (1) is 𝛽1 . Our
main prediction is that 𝛽1 is different from zero. In other words, the microdosing intervention will have
heterogeneous effects across nuclear and extended households.

2.3

Identification Strategy

Our randomized study design enables us to evaluate the impact of the program by applying Difference-inDifference (DID) methods to identify the Intent to Treat (ITT) effect of being given training in
microdosing and free packets of seed and fertilizer. Because technology adoption and household
characteristics may be simultaneously determined (McMillan, 1987; Doss, 2001, Mendola, 2008; Foster
and Rosenzweig, 2010), the randomization design permits identification of the direction of causality. The
ITT approach compares outcomes between households assigned to treatment to those in the control
villages without taking compliance into account. We use this approach instead of investigating the impact
of the actual application of the microdosing technique because some households, especially those in the

51

This claim stems from the assumption that the microdosing kit will be applied on collective plots in some of the
extended households. Production from the collective plots is considered as a public good within these households
(Kazianga and Wahhaj, 2013).

97

control villages, may inaccurately consider prior existing fertilization techniques, such as zai as
microdosing.
As we are interested in the differential effects of the intervention depending on household
endowment in land and on whether households are comprised of either extended or nuclear family
structures (as outlined above), we carry out two distinct identification strategies. In the first, we interact
treatment assignment (Tj) with baseline total land holding (Lhj0), with households indexed by h and
villages indexed by j. Therefore, we estimate OLS for the following set of regressions:
(Yhj1 - Yhj0) =
β1Tj + β2Lhj0 + β3(Tj x Lhj0) + β4Ehj + β5LSharehj0 + µk + (ɛhj1 - ɛhj0),

(2.2)

where we difference outcome measures across pre- and post- treatment surveys. Superscripts 1 and 0 refer
to follow-up and baseline data respectively. Equation (2) enables us to obtain DID estimators for β1 and β3
because it estimates changes in outcomes between baseline and follow-up values across assigned and
control households. All outcome variables are measured per unit of land to account for potential
economies of scale across different landholdings. Our primary outcomes are labor allocation in persondays per hectare in total and disaggregated by tasks, including the following: plowing, planting, inorganic
and organic fertilizer application, weeding, and harvesting. Since treatment assignment is allocated by
village, we cluster all standard errors by village. In this specification, we also control for household
structure (Ehj), a binary variable that takes 1 if an extended household and 0 otherwise.
To identify household structure differential impacts, we carry out a second identification strategy
where baseline total land holding is replaced with household structure. We interact treatment assignment
(Tj) with the indicator variable for household structure (Ehj) and apply OLS estimation to the set of
regressions defined as
(Yhj1 - Yhj0) =
β1Tj + β2Ehj + β3(Tj x Ehj) + β4Lhj0 + β5LSharehj0 + µk + (ɛhj1 - ɛhj0),
where we also control for baseline total land holding (Lhj0).

98

(2.3)

In both specifications (2) and (3), we control for the share of total cultivated land devoted to sorghum as
the main crop at baseline (LSharehj0) and commune fixed effects (Âľk).52
Finally, we remain mindful of the fact that nuclear and extended households differ from one
another in a variety of ways. Not only are extended households comprised of more family members, they
also have significantly larger land holdings. As a result, we rule out that differential estimates in ITT in
specification (3) are due to differences in land holding at baseline by restricting our sample to households
in the lower three quintiles of land holdings at baseline. We do so because a disproportionately higher
number of extended households compared to nuclear households have land holdings in the top two
quintiles. For this restricted sample, we find that average land holdings at baseline are similar for both
nuclear and extended households (see Table 3.2). Sample means tests for this restricted sample also
indicate that extended and nuclear households are relatively similar in terms of baseline fertilizer use
rates, whether sorghum or maize is the main crop at baseline, and total sorghum production at baseline.
Note that for the wider sample, extended and nuclear households differ along all of these dimensions (see
Table 2.2).

2.4

Data

We collected detailed household survey data within our study villages, both before and after the
intervention. We surveyed all those who received starter kits in treatment villages. In addition, we
surveyed approximately 30 households at random from each of the control villages. The baseline survey
took place in late 2013 and early 2014. Households received microdosing training and starter kits at the
starting of the 2014-2015 agricultural campaign, and the follow-up survey was conducted one year later –
in late 2014 and early 2015. Our multi-topic household survey collected information on various
demographic, socio-economic, and agricultural production characteristics. We collected data at the

52

Ideally, we would include village fixed effect but this is not feasible given that treatment assignment does not vary
within village. This is because we dropped all households that are not assigned to treatment in treatment the villages
to avoid spillovers, which may lead to underestimating the impact of the intervention.

99

household level as well as at the individual plot manager level. In total, we collected 1,529 household
surveys at baseline. With a household attrition rate of 6.7 percent, we collected household surveys both at
baseline and follow-up for 1,426 households.

2.4.1

Sample Restrictions

As our primary interest is to determine the differential impact of the microdosing intervention by
household structure, and since male and female headed households are considerably different from one
another, we first restrict our sample to male-headed households. Extended households are mostly headed
by men (Colson 1962; Kazianga and Wahhaj 2015), and most female headed households are nuclear
households. This restriction ensures that in comparing across different household structures, we are not
conflating these differences with differences in male or female headed households. This drops about 10
percent of the households.
In addition, to eliminate any potentially differentiated effects specific to gender differences, we
also exclude all female-controlled plots from the sample. Preliminary analysis of the sample (results not
presented here) indicated that the proportion of female-plot managers among extended households is
much higher relative to nuclear households. Consequently, comparing labor allocation across the two
types of households without eliminating female-controlled plots will likely lead to outcomes being driven
by gender-specific effects rather than by differences in household structure.
Given that our identification strategy consists of investigating changes in labor allocation to
sorghum production across the pre- and post- treatment periods, the sample is further restricted to maleheaded households that had at least one male-controlled sorghum plot in each of the two rounds, resulting
in 1,098 remaining households.
Moreover, we identified outliers in labor allocation measured in person-days per hectare as those
observations with values greater than three times the standard deviation. This restriction eliminates 39
households from our sample, resulting in a final full sample size of 1,059 households.
Finally, since the treatment targeted small sorghum growers and given that extended households

100

tend to have larger land holdings, there is a concern that any findings on differential impact by household
structure would be driven by differences in land holding. We address this concern by carrying out a final
sample restriction where land holding is balanced across household structures. Specifically, we restrict the
sample to the lower three quintiles of land holding, leading to a final sample size of 677 households of
which about 27 percent are extended family households. Table 2.1 summarizes these sequential sample
restrictions.

2.4.2

Demographic and Production Characteristics by Household Structure

Tables 2.2 and 2.3 present summary statistics of demographic and production characteristics for extended
and nuclear households prior to our intervention. The two types of households are very different in terms
of their demographic and production characteristics. First, the heads of extended households are on
average about six years older than nuclear heads. As expected, the numbers of adult males and females
are significantly higher for extended households.
These differences in household composition will likely impact their labor response to the
introduction of the microdosing technology. In particular, the larger number of adults in the extended
households implies the need for more coordination efforts in their labor allocation decisions. As labor
markets are almost nonexistent in the study zone, households rely heavily on family labor for their
agricultural production activities (Udry 1996; Kazianga and Wahhaj 2013). Therefore, since the
microdosing technology is potentially labor-intensive, households may have to readjust their labor
allocation if they were already facing labor constraints before the intervention. This possible need for
labor allocation readjustment is likely to be higher for extended households relative to nuclear
households.

101

Table 2.2 Demographic Characteristics Across Extended and Nuclear Households (Full sample)
Extended HHs
Nuclear HHs (N=705)
(N=354)
VARIABLES
Mean
SD
Mean
SD
Household is Assigned Treatment
0.66
0.48
0.65
0.48
Head Age
46.24
14.08
52.10
15.41
Number of Adult Males
1.90
1.13
3.61
1.79
Number of Adult Females
2.18
1.27
4.15
2.23
Household size (members older than 15 years)
4.51
1.88
8.12
3.45

P-Value
0.94
0.00
0.00
0.00
0.00

Note: Observations at the household level
Table 2.3 Production Characteristics Across Extended and Nuclear Households (Full sample)
Nuclear HHs (N=705)
VARIABLES
Mean
SD
Household Uses Fertilizer
0.39
0.49
Intensity of fertilizer (kg/ha)
8.95
20.40
Total Land Holding (ha)
3.85
2.50
Total land where sorghum is main crop (ha)
2.97
2.14
Total land where millet is main crop (ha)
0.71
1.24
Total land where maize is main crop (ha)
0.05
0.22
Total land where rice is main crop (ha)
0.01
0.11
Total land where peanut is main crop (ha)
0.04
0.21
Labor Allocation to Plowing (person-days/ha)
9.779
9.924
Labor Allocation to Planting (person-days/ha)
Labor Allocation to Fertilizer Application (person-days/ha)
Labor Allocation to Manure Application (person-days/ha)
Labor Allocation to Weeding (person-days/ha)
Labor Allocation to Harvest (person-days/ha)
Total Labor Allocation (person-days/ha)
Total sorghum production
Total sorghum yield (kg/ha)

13.62
2.390
4.651
37.44
14.85
82.80
1,513
630.3

Note: Observations are at the household level

102

10.55
4.451
5.963
37.66
13.88
60.20
1,288
546.2

Extended HHs (N=354)
Mean
SD
0.48
0.50
9.86
15.97
5.66
4.10
4.40
3.61
0.94
1.41
0.09
0.32
0.09
1.10
0.03
0.19
11.34
11.91
15.64
3.259
5.714
40.16
17.72
93.88
2,144
583.8

10.75
5.169
7.206
31.67
14.15
57.73
1,877
477.3

P-Value
0.01
0.42
0.00
0.00
0.00
0.02
0.32
0.72
0.05
0.00
0.01
0.02
0.21
0.00
0.00
0.000
0.11

The larger household size of the extended household (in terms of adult members who are more
than 14 years old) is further reflected in larger total land holdings devoted to agricultural production,
especially for sorghum. Moreover, the proportion of households that apply fertilizer among extended
households exceeds that of nuclear households by 9 percent, and the difference is statistically significant
at 5 percent. However, there is no statistically significant difference in the intensity of fertilizer
application. Relative to nuclear households, labor allocation expressed in number of person-days per
hectare is higher among extended households for all activities. However, there is no significant statistical
difference in sorghum yield across the two types of household structure. Finally, Tables 2.4 and 2.5
present descriptive statistics for household demographic and production characteristics after restricting
the sample to households in the lower three quintiles of land holdings. As we highlight above, the
motivation for such a restriction is to ensure that total land holding is not the main source of differences in
the ITT estimates across extended and nuclear households.
Overall, much of the mean differences previously observed in Table 2.2 subsides or no longer
exists. For example, not only is total land holding balanced by household structure, but so too are total
sorghum land, fertilizer application rates, and total sorghum production.

103

Table 2.4 Demographic Characteristics Across Extended and Nuclear Households (Restricted Sample)
Nuclear HHs (N=496)
VARIABLES
Mean
SD
Household is Assigned Treatment
0.66
0.48
Head Age
45.09
14.32
Number of Adult Males
1.776
1.096
Number of Adult Females
1.976
1.121
Household size (members older than 15 years)
4.141
1.712

Extended HHs (N=181)
Mean
SD
0.65
0.48
52.01
15.93
3.243
1.357
3.624
1.681
7.215
2.360

P-Value
0.997
0.000
0.000
0.000
0.000

Note: Observations are at the household level. Sample is restricted to the lower three quintiles of land holding
Table 2.5 Production Characteristics Across Extended and Nuclear Households (Restricted Sample)
Nuclear HHs (N=496)
Extended HHs (N=181)
VARIABLES
Mean
SD
Mean
SD
Household Uses Fertilizer
0.379
0.486
0.464
0.500
Intensity of fertilizer (kg/ha)
9.898
22.40
11.81
18.63
Total Land Holding (ha)
1.839
0.779
1.940
0.775
Total land where sorghum is main crop (ha)
1.882
0.905
2.109
1.601
Total land where millet is main crop (ha)
0.749
1.277
1.006
1.410
Total land where maize is main crop (ha)
0.0462
0.211
0.0762
0.279
Total land where rice is main crop (ha)
0.00479
0.0436
0.122
1.489
Total land where peanut is main crop (ha)
0.0286
0.165
0.0138
0.0902
Labor Allocation to Plowing (person-days/ha)
10.16
10.83
14.15
14.57
Labor Allocation to Planting (person-days/ha)
14.52
11.60
18.74
12.76
Labor Allocation to Fertilizer Application (person-days/ha)
2.636
4.855
3.968
6.046
Labor Allocation to Manure Application (person-days/ha)
4.899
6.367
7.064
8.882
Labor Allocation to Weeding (person-days/ha)
41.28
42.21
48.32
37.83
Labor Allocation to Harvest (person-days/ha)
16.19
15.66
21.95
17.32
Total Labor Allocation (person-days/ha)
89.75
66.29
114.2
67.08
Total sorghum production
1,188
966.0
1,352
1,053
Total sorghum yield (kg/ha)
705.0
605.4
689.7
574.5
Note: Observations are at the household level. Sample is restricted to the lower three quintiles of land holding

104

P-Value
0.080
0.245
0.094
0.049
0.033
0.162
0.313
0.161
0.003
0.000
0.008
0.004
0.027
0.001
0.000
0.099
0.751

2.4.3

Pre-Treatment Household Demographic and Production Characteristics

In order to attribute any changes in labor allocation among assigned households to the microdosing
treatment, it is important to ensure that households assigned to treatment and control groups share similar
pre-intervention characteristics. Since we are interested in the heterogeneous effects of the microdosing
treatment by household structure, we conduct balancing tests across treated and control households within
each category of household structure.
Table 2.6 presents summary statistics of pre-intervention household demographic and production
characteristics by assignment status for extended and nuclear households (see Panels A and B). All pvalues corresponding to the statistical significance level of the mean differences across the assigned and
control households greatly exceed the 10 percent significance level, indicating that the randomization
produced balanced assignment and control groups of households.
Finally, Table 2.7 presents balancing tests by treatment assignment within household structure
after restricting the sample to households in the lower three quintiles of total land holdings. Overall, the
balance observed in Table 4 is preserved in the restricted sample. As a result, any differences in the ITT
observed across nuclear and extended family households can be attributed to the heterogeneous impact of
the treatment assignment by household structure.

105

Table 2.6 Balancing Test Results Across Assigned and Control Households by Household (Full sample)
Panel A: Extended Households
Control
Assigned
(N=120)
(N=234)
Mean
SD
Mean
SD
P-Value
VARIABLES
Head Age
52.08
15.66 52.1
15.3
0.994

Panel B: Nuclear Households
Control
Assigned
(N=465)
(N=240)
Mean
SD
Mean
SD
46.27
14.38
45.95
14.02

P-Value
0.785

Number of Adult Males

3.525

1.773

3.632

1.678

0.674

1.858

1.009

1.927

1.187

0.498

Number of Adult Females
Total number of adults
Household more than one married male

4.092
8.008
0.85

1.879
2.932
0.359

4.184
8.124
0.893

2.39
3.63
0.31

0.742
0.795
0.318

2.192
4.529
0

1.296
1.785
0

2.178
4.443
0

1.218
1.868
0

0.9
0.615
.

Household head has at least one married son

0.483

0.502

0.483

0.501

0.994

0

0

0

0

.

Household head has at least one married brother

0.25

0.435

0.252

0.435

0.964

0

0

0

0

.

Household head has at least one daughter-in-law

0.483

0.502

0.444

0.498

0.518

0

0

0

0

.

Household head has at least one sister-in-law
Household Uses Fertilizer
Intensity of fertilizer (kg/ha)

0.358
0.392
8.124

0.482
0.49
16.75

0.385
0.521
10.74

0.488
0.501
15.51

0.67
0.107
0.218

0
0.338
7.368

0
0.474
19.41

0
0.417
9.768

0
0.494
20.87

.
0.157
0.286

Total Land Holding (ha)

4.324

3.029

3.974

3.046

0.46

2.977

2.066

2.782

1.952

0.401

Total land where sorghum is main crop (ha)

4.672

3.652

4.234

3.523

0.449

3.009

2.11

2.869

2.154

0.551

Labor Allocation to Plowing (person-days/ha)

11.62

14.32

11.2

10.5

0.817

9.319

9.517

10.02

10.13

0.486

Labor Allocation to Planting (person-days/ha)
Labor Allocation to Fertilizer Application
(person-days/ha)
Labor Allocation to Manure Application
(person-days/ha)
Labor Allocation to Weeding (person-days/ha)

15.25

10.5

15.84

10.9

0.688

13.54

9.754

13.66

10.94

0.896

2.957

5.046

3.413

5.236

0.472

2.077

4.167

2.551

4.587

0.357

4.761

6.29

6.203

7.6

0.108

5.074

6.917

4.432

5.401

0.34

38.1

28.42

41.21

33.22

0.385

38.6

35.14

36.85

38.92

0.562

Labor Allocation to Harvest (person-days/ha)

16.35

13.36

18.41

14.52

0.261

14.65

11.62

14.95

14.93

0.786

89.11
537.4

57.31
544.3

96.33
607.6

57.91
438.4

0.34
0.259

83.37
671

59.13
640.1

82.5
609.3

60.81
490.1

0.859
0.225

Labor Allocation (person-days/ha)
Total sorghum yield (kg/ha)
Note: Observations are at the household level

106

Table 2.7 Balancing Test Results Across Assigned and Control Households by Household (Restricted sample)
Panel A: Extended Households
Assigned
Control (N=58)
(N=123)
VARIABLES
Mean
SD
Mean
SD
P-value
Head Age
Number of Adult Males
Number of Adult Females
Total number of adults
HH more than one married male
HH head has at least one married son
HH head has at least one married brother
HH head has at least one daughter-in-law
HH head has at least one sister-in-law
HH Uses Fertilizer
Intensity of fertilizer (kg/ha)
Total Land Holding (ha)
Total land where sorghum is main crop (ha)
Labor Allocation to Plowing (person-days/ha)
Labor Allocation to Planting (person-days/ha)
Labor Allocation to Fertilizer Application
(person-days/ha)
Labor Allocation to Manure Application
(person-days/ha)
Labor Allocation to Weeding (person-days/ha)
Labor Allocation to Harvest (person-days/ha)
Labor Allocation (person-days/ha)

Panel B: Nuclear Households
Control
Assigned(N=337
(N=159)
)
Mean
SD
Mean
SD

P-value

53.17
3.10
3.72
7.28
0.85
0.26
0.52
0.52
0.31
0.36
8.99
2.01
2.07
13.91
18.35

16.92
1.25
1.67
2.24
0.37
0.44
0.50
0.50
0.47
0.49
19.57
0.83
0.83
18.90
12.85

51.45
3.31
3.58
7.19
0.88
0.23
0.44
0.35
0.37
0.51
13.14
1.91
2.13
14.27
18.92

15.47
1.40
1.69
2.42
0.33
0.42
0.50
0.48
0.49
0.50
18.10
0.75
1.86
12.10
12.76

0.55
0.38
0.57
0.80
0.50
0.61
0.30
0.03
0.37
0.30
0.77
0.54
0.12
0.91
0.81

45.24
1.78
1.96
4.20
0.33
9.10
1.85
1.86
10.12
14.94

14.34
1.00
1.05
1.60
0.47
22.72
0.76
0.76
10.73
11.01

45.03
1.77
1.98
4.12
0.40
10.27
1.84
1.89
10.17
14.33

14.33
1.14
1.16
1.76
0.49
22.27
0.79
0.97
10.90
11.88

0.88
0.96
0.84
0.70
.
.
.
.
.
0.56
0.19
0.68
0.26
0.97
0.60

3.02

5.53

4.42

6.25

0.14

2.34

4.42

2.77

5.05

0.46

5.42
46.84
19.78

7.64
34.99
17.64

7.84
49.02
22.97

9.34
39.22
17.14

0.71
0.34

5.53
44.94
16.53

7.57
39.66
13.26

4.60
39.56
16.02

5.70
43.31
16.69

107.30

71.71

117.50

64.83

0.42

94.56

65.96

87.48

66.42

0.27

775.10

732.0

671.90

533.50

0.13

Total sorghum yield (kg/ha)

0.10

663.60
711.00
702.00 500.40 0.70
Note: Observations are at the household level. Sample is restricted to the lower three quintiles of land holding

107

0.26
0.16
0.74

2.5

Results

2.5.1

Results with the full sample

The primary focus of this paper is the heterogeneity of the treatment effect by household structure.
However, since the treatment provided enough fertilizer and seed for a small parcel (one half hectare),
one might expect that smaller landholders may be more greatly affected by the intervention in comparison
to larger landholders. Thus, baseline land holdings may be a potentially important factor in determining
the effectiveness of the treatment on labor allocation. Consequently, we present two sets of results in
Panels A and B in Table 2.8 corresponding to equations (2.2) and (2.3).
Three models are presented in each panel. Model A1 in Panel A includes the treatment
assignment and the log of baseline land holdings as well as their interaction terms. Model A2 adds
commune fixed effects to account for the potential effects of location-specific unobservables. These
include factors such as access to inputs, labor and output markets, as well as differences in social and
cultural norms that are likely to affect self-selection into a particular household structure. In Model A3,
we add a binary indicator of household structure (extended households, with nuclear households being the
comparison). We also add the share of total land devoted to sorghum cultivation in Model A3 because the
prevalence of sorghum production in the household could affect labor allocation to sorghum production
activities such as microdosing, with potential labor constraints to adoption decisions. The specifications
in Panel B are analogous to those in Panel A, where the interaction term is now between household
structure and treatment status.

Table 2.8 Changes in log of labor allocation across baseline and follow-up (Full Sample)
Panel A: Interaction with log of land
VARIABLES

Model A1

Model A2

Model A3

Assigned

-0.013

0.013

0.014

(0.14)

(0.15)

(0.15)

0.383***

0.432***

0.442***

(0.08)

(0.08)

(0.08)

0.036

-0.004

-0.004

Logland
Assigned*Logland

108

Table 2.8 (cont’d)
(0.10)

(0.10)

Extended

(0.10)
0.006
(0.06)

Sorghum land share

-0.077
(0.16)

Commune FE included

No

Yes

Yes

Observations

1,059

1,059

1,059

R-squared

0.060

0.117

0.117

VARIABLES

Model B1

Model B2

Model B3

Assigned

0.081

0.074

0.079

(0.079)

(0.073)

(0.072)

0.264***

0.263***

0.142

(0.095)

(0.094)

(0.100)

-0.205*

-0.223**

-0.206*

(0.114)

(0.110)

(0.111)

Panel B: Interaction with household structure

Extended
Assigned*Extended
Logland

0.437***
(0.056)

Sorghum land share

-0.073
(0.14)

(β1+β3)

-0.125

-0.150

-0.126

P-value of F-test (β1+β3)=0

0.166

0.088

0.144

Commune FE included

No

Yes

Yes

Observations

1,059

1,059

1,059

R-squared
0.008
0.120
0.120
Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 ***
p<0.01. Only households with at least one male-controlled plot where sorghum is main crop across baseline and
follow-up are included in all regressions. For extended household definition, see section 3. Logland and sorghum
land share are baseline observations

For the specification with respect to household structure in Panel B, Table 2.8, the ITT effect of
the intervention for nuclear households is captured by the coefficient on the variable “Assigned” (i.e., β1).
Here, even though the program seems to not have induced significant changes in labor allocation among
nuclear households in treatment villages, the magnitude of the coefficients is much larger relative to those
in Panel A in Table 2.8 with relatively smaller standard errors. This indicates that the program may have
induced a slight increase in labor allocation among nuclear households. As the overall effect on extended

109

households is the sum of the coefficient estimates on the treatment variable and the interaction term (i.e.,
β1+β3), the net effect for these households is -0.125, -0.150, and -0.126, across Model B1, B2, and B3,
respectively. These negative coefficients imply that the microdosing intervention induced a decrease in
labor allocation among extended households. The magnitude of each of these sums is relatively high,
providing suggestive evidence of heterogeneous impact of the treatment assignment by household
structure (though the sums of these coefficients are only marginally statistically different from zero; see
Table 2.8, Panel B).

2.5.2

Results with the restricted sample

As shown in Table 2.2, there is a large disparity between pre-program landholdings across nuclear and
extended family households. Sorghum land holdings among extended family households is almost 50
percent larger. It is therefore possible that the results presented thus far could be a mere artifact of
differences in baseline land holdings across nuclear and extended households. To rule out this possibility,
we carry out additional regressions with the sample restricted to the lower three quintiles of land holdings,
where we have relatively balanced land holdings by household structure as well as other important
production characteristics such as fertilizer use rates, total land devoted to sorghum cultivation, and total
sorghum production (see Table 3.2). One important point in the results with the restricted sample in Table
2.4 is that while production characteristics tend to be relatively balanced across extended and nuclear
households, household demographics remain significantly different. Therefore, differences in treatment
impact across extended and nuclear households in the restricted sample can be attributed to the
heterogeneous impact of the program by household structure.
The results obtained with the restricted sample are presented in Table 7. Similar to the results
with the full sample, there is no evidence of heterogeneous treatment impact by baseline landholding in
the specification where treatment assignment is interacted with log of landholdings (see Models A1, A2,
and A3 in Panel A, Table 2.9).
When we interact treatment assignment and household structure, we find a significant difference

110

in ITT estimates by household structure regardless of the model specification (Panel B, Table 2.9).
Particularly, in Model B3, where we include both baseline landholdings and sorghum land share, as well
as commune fixed effects, we find that assignment to treatment induced nuclear households to increase
labor allocation to sorghum production by approximately 17 percent and this is statistically different from
zero at the 5 percent significance level. With the baseline mean value of total labor allocation being about
90 person-days per hectare, this corresponds to an increase of nearly 15 person-days per hectare among
nuclear households. This finding is consistent with the fact that the microdosing technique is labor
intensive.

Table 2.9 Changes in log of labor allocation across baseline and follow-up (Restricted Sample)
Panel A: Interaction with log of land
VARIABLES

Model A1

Model A2

Model A3

Assigned

-0.196

-0.223

-0.215

(0.24)

(0.25)

(0.25)

0.217

0.21

0.247

(0.20)

(0.20)

(0.20)

0.24

0.255

0.25

(0.24)

(0.24)

(0.24)

Logland
Assigned*Logland
Extended

-0.198
(0.16)

Sorghum land share

0.026
(0.07)

Commune FE included

No

Yes

Yes

Observations

677

677

677

R-squared

0.019

0.097

0.100

VARIABLES

Model B1

Model B2

Model B3

Assigned

0.159*

0.165**

0.163**

(0.090)

(0.079)

(0.081)

0.373***

0.368***

0.328***

(0.116)

(0.112)

(0.112)

Panel B: Interaction with household structure

Extended

111

Table 2.9 (cont’d)
Assigned*Extended

-0.431***

-0.464***

-0.451***

(0.145)

(0.142)

(0.139)

Logland

0.411***
(0.114)

Sorghum land share

-0.196
(0.16)

(β1+β3)

-0.272

-0.299

-0.288

P-value of F-test (β1+β3) = 0

0.032

0.014

0.013

Commune FE included

No

Yes

Yes

Observations

677

677

677

R-squared
0.060
0.117
0.117
Note: Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 ***
p<0.01. Sample restricted to the first three quintiles of total sorghum land holding. Only households with at least one
male-controlled plot where sorghum is main crop across baseline and follow-up are included in all regressions. For
extended household definition, see section 3. Logland and sorghum land share are baseline observations

While the program induced an increase in labor allocation among nuclear households, the results
suggest that it had the opposite effect on extended households. The differential impact by household
structure is shown by the negative coefficient on the interaction term. Assigned*Extended (i.e., β3),
which is -0.45 and is statistically significant at the 1 percent level (Model B3, Panel B, Table 2.9). This
implies that overall, the impact of the program on total labor allocated to sorghum production among
extended households is approximately 45 percent lower relative to its impact on nuclear households.
The net impact on the extended households, which is the sum of the coefficients (β1+β3), is
negative 0.29. The descriptive statistics presented in Table 2.5 show that baseline total labor allocated to
sorghum production among the extended households in the restricted sample is about 114 person-days per
hectare. Consequently, the negative 0.29 indicates that the microdosing intervention induced a reduction
in labor allocation for sorghum production by about 33 person-days per hectare among the extended
households. These results provide evidence that the prior estimates in Table 2.8 may have underestimated
differences in ITT effects of the intervention by household structure.
To understand how the intervention increased total labor allocation among the assigned nuclear
households, we present results for specific activities in Table 2.10. Here, we focus on Model B3, where

112

treatment is interacted with household structure using the most exhaustive specification –that is, where we
control for baseline land holdings, sorghum land share and commune fixed effects. The outcome variables
are changes (across pre- and post-intervention periods) in time spent on: plowing, planting, manure
application, inorganic fertilizer application, weeding, and harvest, and are all expressed in log.
Labor allocation results with respect to fertilizer application indicate that the program had a
significant impact for all treated households, by up to 42 percent and 57 percent , respectively, for nuclear
and extended households. Based on the baseline average values presented in Table 2.5, these results
indicate that the treatment induced an increase of about 1 and 2 person-days per hectare on time devoted
to inorganic fertilizer application, respectively, among nuclear and extended households. This increase in
fertilizer labor suggests that the beneficiary households did use the fertilizer in their field.

Table 2.10 Labor allocation results by activity using sample restricted to the lower three quintiles of total sorghum
land
Plowing
Planting
Manure
Fertilizer
Weeding
Harvest
VARIABLES
Assigned (β1)
Extended

-0.003
(0.13)
-0.143
(0.22)

0.169**
(0.08)
0.266**
(0.13)

0.138
(0.11)
0.518**
(0.20)

0.421***
(0.12)
-0.123
(0.21)

0.141
(0.12)
0.427*
(0.22)

0.214**
(0.10)
0.352***
(0.13)

Assigned*Extended
(β3)

0.002
-0.324**
-0.568**
0.153
-0.480*
-0.397**
(0.26)
(0.16)
(0.24)
(0.28)
(0.28)
(0.18)
Logland
-0.316
-0.22
-0.421*
0.092
0.192
-0.18
(0.22)
(0.17)
(0.22)
(0.23)
(0.24)
(0.19)
Sorghum land share
0.348
0.405***
0.139
-0.045
0.057
0.491***
(0.22)
(0.13)
(0.19)
(0.19)
(0.22)
(0.16)
(β1+β3)
-0.00
-0.155
-0.430*
0.574**
-0.339
-0.182
P-value of (β1+β3) = 0
0.997
0.267
0.061
0.028
0.161
0.239
Commune FE included
Yes
Yes
Yes
Yes
Yes
Yes
Observations
677
677
677
677
677
677
R-squared
0.101
0.055
0.090
0.063
0.142
0.081
Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 ***
p<0.01. Sample restricted to the first three quintiles of total sorghum land holding. Only households with at least one
male-controlled plot where sorghum is main crop across baseline and follow-up are included in all regressions. For
extended household definition, see section 3. Logland and sorghum land share are baseline observations

While the program seems to have similarly affected nuclear and extended households in terms of
their labor devoted to inorganic fertilizer application, ITT estimates with respect to the other labor

113

activities indicate that the program positively affected nuclear households only, and in some instances had
the opposite effect on extended households. This is shown by the positive and significant coefficient β1
for labor allocated to planting and harvest, and the negative interaction term between Assigned and
Extended, which is also statistically significant (i.e. β3). Given that the overall effect on extended
household is (β1+ β3), the net effect on these households is negative, although not statistically different
from zero. With respect to time spent on manure application, the program induced a significant decrease
among extended households, while no significant changes are detected among nuclear households.
The estimation by labor (i.e., male, female, or child labor), presented in Table 2.11, reveals that
the heterogeneous impact of the treatment mostly comes from male labor allocation followed by female
labor. As for child labor, there are no discernible differences across the two types of household structure.

114

Table 2.11 Changes in log of labor allocation across baseline and follow-up by labor source (Restricted sample)
Male labor
Female labor
Child labor
VARIABLES
Model 1 Model 2
Model 3
Model 1
Model 2
Model 3
Model 1
Model 2

Model 3

Assigned (β1)

0.144

0.190**

0.189**

0.222**

0.186*

0.182*

-0.031

0.07

0.072

Extended

(0.097)
0.470***

(0.084)
0.468***

(0.087)
0.424**

(0.103)
0.367**

(0.107)
0.336**

(0.109)
0.314**

(0.213)
0.321

(0.174)
0.405

(0.179)
0.378

Assigned*Extended (β3)

(0.158)
-0.503**

(0.170)
-0.547***

(0.168)
-0.533***

(0.146)
-0.384**

(0.147)
-0.369**

(0.155)
-0.359*

(0.356)
-0.239

(0.321)
-0.371

(0.328)
-0.365

(0.192)

(0.202)

(0.196)
-0.294*

(0.190)

(0.184)

(0.189)
0.008

(0.414)

(0.390)

(0.392)
-0.362

Logland

(0.174)
(0.189)
(0.363)
0.377***
0.385**
0.009
(0.142)
(0.164)
(0.326)
(β1+β3)
-0.360
-0.356
-0.344
-0.162
-0.183
-0.177
-0.270
-0.301
-0.293
P-value of (β1+β3) =0
0.018
0.034
0.031
0.385
0.284
0.304
0.490
0.345
0.357
Commune FE included
No
Yes
Yes
No
Yes
Yes
No
Yes
Yes
Observations
677
677
677
677
677
677
677
677
677
R-squared
0.014
0.077
0.090
0.010
0.063
0.074
0.002
0.071
0.073
Note: Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 *** p<0.01. Sample restricted to the first three
quintiles of total sorghum land holding. Only households with at least one male-controlled plot where sorghum is main crop across baseline and follow-up are
included in all regressions. For extended household definition, see section 3. Logland and sorghum land share are baseline observations.
Sorghum land share

115

2.6

Understanding the sources of the heterogeneous impact of the program

The heterogeneous impact by household structure may be attributed to several reasons. One important
feature of the microdosing application technique is that it is labor intensive, especially at planting and
fertilizer application times. Thus, the lack of increase in the overall labor allocated to sorghum production
among the extended households, despite suggestive evidence that these households applied the fertilizer
distributed to them on their sorghum field, is puzzling. This is especially startling, given that these
households are more likely to have a larger pool of labor relative to the nuclear households given their
larger size. While the differential impact of the intervention by household structure is counter-intuitive,
we offer a few factors that may rationalize the current findings.

2.6.1

Labor fixity within the extended households and differences in incentive?

One plausible factor associated with the heterogeneous effect of the intervention is that extended
household members may face fixed labor or some coordination costs in the reallocation of their labor. As
a result, the increased labor requirement in fertilizer application may have substituted labor away from
other activities. This speculation is supported by the significant reduction in the time spent on manure
application among extended households.
To investigate whether the microdosing substituted labor away from the production of the other
crops, especially for the nuclear households who experienced an increase in labor allocated to sorghum
production, we estimate changes in labor allocation with respect to the other crops. Two noteworthy
findings must be highlighted in the corresponding results presented in Table 2B (APPENDIX B). First,
we find that the treatment did not induce any changes in the amount of labor allocated to the production
of the other crops. Second, there is no indication of heterogeneous impact of the program by household
structure. One of the immediate implication of these results is that the intervention did not substitute labor
away from the production of the other crops. This corroborates our assumption that the increase in the
time spent on fertilizer application for sorghum production among the extended households resulted from
a decrease in manure application time for the same crop, indicating labor fixity among these households.

116

A pressing question is: Where did the increased time spent on sorghum production among the
nuclear households come from? One may stipulate that the program led nuclear households to adjust the
share of total land devoted to sorghum cultivation to adequately respond to the labor requirement of the
microdosing technology. However, this claim is not sustained by the data, as results presented in Table
2C (APPENDIX C) do not show any evidence of changes in the share of sorghum land, neither by
treatment status, nor by household structure. The results with respect to changes in the number of plots
presented in the same table yielded no noticeable differences either. Based on these results, our
assumption is that the microdosing intervention induced nuclear households to exert more effort by
spending more time on their sorghum production.
How might have the microdosing induced labor increase among the nuclear households? Recall
that one important distinction between nuclear and extended households in terms of their production
organization is the coexistence of collective and individual plots within the extended households (West,
2010; Kazianga and Wahhaj, 2013; Guirkinger and Platteau, 2014). Labor allocation across the two types
of plots requires managers of the individual plots, who are junior members relative to the head, to provide
labor on the collective plots, which are managed by the head, before working on their individual plots.
Production from the collective plots is shared among all extended family members. The sharing rule is
mostly based on need instead of effort provision (Guirkinger and Platteau, 2014) either by taking common
meals or distributing output to each nuclear household (Beaman and Dillon, 2012).53
With this sharing scheme within the extended households, an increase in effort provided by an
individual does not necessarily translate into an increase in her / his reward. In contrast, within the pure
nuclear households, production is for the head, his spouse(s), unmarried children, and other immediate
dependents. Consequently, for the nuclear households, any accrued production from the microdosing

53

One may argue that the higher labor allocation among the nuclear households relative to the extended households
is due to moral hazard on the collective plots within the extended households. While the available data do not allow
us to rule out this possibility, we claim that the issue of moral hazard is unlikely to be a driving factor based on
compelling evidence in the empirical literature that, because of social norms, collective plots receive much more
labor relative to individual plots among households in the same region (Kazianga and Wahhaj, 2013).

117

technology is the sole ownership of these households, which may provide them with greater incentives to
exert more effort to their sorghum production activities following the reception of the kits.
Furthermore, labor endowment is more likely to be centralized at the head’s level within the pure
nuclear households, enabling the later to make household members spend longer times on sorghum plots
that received the microdosing kit to respond to the increased labor requirement. Conversely, in the
extended households where multiple nuclear households exist, labor division is such that members can
only work for a certain amount of time on the collective plots, after which they switch to their own
individual plots. The potential constraint that this labor division imposes on the time spent on the
collective plots may also constitute a plausible reason for the heterogeneous impact of the intervention by
household structure.

2.6.2

Differences in productive resource control within the extended households?

The claims of fixed labor and differences in incentives by household structure provided above stem from
the assumption that the microdosing kit is applied on the collective plots within the extended households.
However, any sorghum grower in the household was eligible to be randomly selected to receive the kit.
Given that junior plot managers are required to work on the collective plots before working on their
individual plots , another plausible source of the differential impact of the microdosing program by
household structure could be the following: if the recipient of the microdosing kit in the extended
household is a junior plot manager with less control over productive resources within the household
compared to the household head, he may not be able to mobilize the required labor for the technology on
his plot.
Plot-manager level regression results are presented in Table 2D (APPENDIX D). Overall, we do
not find any significant differences in changes in total labor allocated to sorghum production by
headship.54 However, the analysis by specific activity reveals that recipients of the microdosing kits who

54

Note from the description of our intervention that there is only one recipient by household.

118

are heads of households experienced a significant increase in time spent on fertilizer application relative
to non-head members.
We also run regressions where we control the headship of the recipient across nuclear and
extended households (results not reported here). We did not detect any significant statistical difference
due to imprecise estimates, certainly because of low variation in the headship of the recipient, especially
among nuclear households. Indeed, up to 96 percent of the recipients in the nuclear households are heads
of households, while the corresponding figure for extended households is about 81 percent.
Because almost all recipients in the nuclear households are heads, we ran additional regressions
where we keep a sample of extended households only to investigate differences across head and non-head
recipients. Overall. The results (see APPENDIX E) remain robust, confirming that recipients who are
heads experienced a larger increase in their labor allocation compared to non-head recipients.

2.7

Conclusion and policy implications

We have shown that there are significant heterogeneous impacts of the microdosing program by
household structure. Specifically, we find that the intervention increased labor allocation to sorghum
production among nuclear households, which is consistent with the fact that the microdosing technique is
labor intensive. We find no indication that the increase in labor allocated to sorghum production among
these households resulted from labor being substituted away from the production of the other crops.
In contrast, for the extended households, we find that the increased labor requirement for fertilizer
application substituted labor from time spent on manure application, resulting into an overall decrease in
the total amount of labor allocation to sorghum production. These findings are startling, given the
presumably bigger pool of labor available to the extended households due to their larger size. We explain
this unanticipated heterogeneous impact of the intervention by a couple of factors stemming from
differences in the production system across the two types of households.
First, we attribute the increase in labor allocation among the nuclear households to higher
incentives for the latter to exert more effort in their sorghum production following the reception of the

119

kits. Nuclear households are sole claimants of their production and, consequently, can keep any potential
increase in sorghum production following the introduction of the microdosing technology. In contrast, in
the extended households, where multiple nuclear families coexist, members jointly work on collective
plots and share production among themselves based on the need of each nuclear family, instead of the
level of their efforts. Consequently, if the microdosing is applied on a collective plot within the extended
households, the incitation to meet the increased required labor will be lower among these households
relative to the nuclear households.
Furthermore, extended family members can only allocate a certain amount of time on the
collective plots as they must also allocate part of their labor on their own individual plots. This cap on the
amount of time that can be spent on the collective plots limits the ability of these households to meet the
labor increase induced by the adoption of the microdosing. This is backed by the suggestive evidence that
extended households reduced the amount of time spent on manure application to meet the additional labor
required for fertilizer application.
Labor constraint on sorghum plots receiving the microdosing technology also holds when the
technology is applied on individual plots, especially, if the recipient is a non-head member, who has
relatively less control over productive resources relative to the head of household. We support this claim
with suggestive evidence indicating that recipients of the microdosing kits who are heads of households
experienced a greater increase in labor allocated to fertilizer application relative to those who are nonhead members.
Despite some limitations in the interpretation of our results, the current findings have important
implications for agricultural development policy interventions in developing countries, where increasing
fragmentation of households to smaller units has led to heterogeneity in household structures. Most
agricultural technologies in the developing world, especially among small scale subsistence farmers, tend
to be labor intensive. At the same time, these households face missing labor markets and must resort to
their own members to satisfy their labor need such that household composition becomes an important
determinant of labor endowment. As a result, the differential impact of the microdosing intervention by

120

household structure highlights the importance of understanding households labor allocation decision
making depending on their structure for successful implementation of agricultural development
interventions. This is important for the promotion of labor-intensive technologies among farmers who
face labor constraints due to limited control over productive resources and missing labor markets.
Another implication for the increased labor among the nuclear households is that they would be
more likely to adopt labor intensive agricultural technologies. This, combined with the importance of
technology adoption for increased agricultural productivity and the increasing fragmentation of larger
households into smaller units, the current findings suggest high potential for future agricultural
intensification. These implications reconcile with findings among similar population in Burkina Faso,
where the authors find more productive efficiency among nuclear households relative to extended
households.

121

APPENDICES

122

APPENDIX A MICRODOSING TAKE-UP BY HOUSEHOLD STRUCTURE
Table 2A. Take-up of microdosing by household structure after program implementation
Nuclear treated HHs

Extended treated HHs

Mean

SD

Mean

SD

P-value of
diff

Full sample

0.652

0.477

0.676

0.469

0.435

Restricted sample

0.668

0.471

0.629

0.484

0.333

123

APPENDIX B CHANGES IN LABOR ALLOCATION FOR NON SORGHUM CROPS
Table 2B. Changes in labor allocation across baseline and follow-up for crops other than sorghum
Plowing
Planting
Fertilizer Manure
Weeding
VARIABLES
Total

Assigned

Harvest

0.01

0.035

-0.047

0.049

0.089

-0.037

0.057

(0.09)

(0.16)

(0.11)

(0.14)

(0.16)

(0.14)

(0.12)

0.093

0.28

-0.009

0.144

0.139

0.208

0.117

(0.13)

(0.24)

(0.13)

(0.20)

(0.27)

(0.20)

(0.21)

-0.23

-0.467

-0.13

-0.058

-0.139

-0.404

0.022

(0.17)

(0.30)

(0.17)

(0.28)

(0.29)

(0.28)

(0.26)

-0.2

0.244

-0.385

0.341

-0.171

0.099

-0.619*

(0.28)

(0.41)

(0.29)

(0.41)

(0.54)

(0.40)

(0.36)

0.318**

0.202

0.353**

0.165

0.095

0.169

0.259*

(0.15)

(0.19)

(0.14)

(0.19)

(0.25)

(0.20)

(0.15)

Commune FE included

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Observations

482

482

482

482

482

482

482

Extended
Assigned*Extended
Sorghum land share
Logland

R-squared
0.089
0.098
0.105
0.077
0.071
0.191
0.104
Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 ***
p<0.01. Only households with at least one male-controlled plot where sorghum is main crop across baseline and
follow-up are included in all regressions. For extended household definition, see section 3. Logland and sorghum
land share are baseline observations. Smaller number of observations as only few households have male-controlled
plots where sorghum is not the main crop

124

APPENDIX C CHANGES IN SORGHUM LAND SHARE AND NUMBER OF PLOTS
Table 2C. Changes in sorghum land share and number of plots across baseline and follow-up
Changes in sorghum land share

Changes in number of plots

VARIABLES

Model 1

Model 2

Model 3

Model 1

Model 2

Model 3

Assigned

0.021

-0.004

-0.006

0.168

0.075

0.065

(0.025)

(0.023)

(0.022)

(0.103)

(0.089)

(0.082)

0.001

-0.003

0.031

-0.187

-0.175

0.066

(0.025)

(0.025)

(0.025)

(0.151)

(0.156)

(0.125)

-0.028

-0.026

-0.031

-0.055

-0.063

-0.098

(0.033)

(0.034)

(0.034)

(0.174)

(0.180)

(0.152)

Extended
Assigned*Extended
Logland

-0.129***

-0.889***

(0.017)

(0.070)

Commune FE
included

No

No

Yes

No

No

Yes

Observations

1,059

1,059

1,059

1,059

1,059

1,059

R-squared
0.002
0.045
0.096
0.0136
0.0629
0.2095
Note: Household level regressions. Standard errors clustered by village in parentheses. * p<0.1 ** p<0.05 ***
p<0.01. Only households with at least one male-controlled plot where sorghum is main crop across baseline and
follow-up are included in all regressions. For extended household definition, see section 3. Logland and sorghum
land share are baseline observations.

125

APPENDIX D CHANGES IN LABOR ALLOCATION BY HEADSHIP AT THE PLOT-MANAGER LEVEL
Table 2D. Changes labor allocation across heads and non-heads members between baseline and follow-up (All male plot managers)
VARIABLES
Total labor
Plowing
Planting
Fertilizer
Manure
Weeding
Harvest
Recipient is head

0.072
0.06
0.061
0.275***
0.093
0.041
0.054
(0.05)
(0.08)
(0.05)
(0.07)
(0.09)
(0.08)
(0.06)
Sorghum land share
-0.254**
-0.248
-0.149
0.119
-0.453**
-0.211
-0.351**
(0.12)
(0.19)
(0.13)
(0.16)
(0.18)
(0.16)
(0.15)
Logland
0.230***
0.213***
0.260***
0.085
0.114*
0.133*
0.257***
(0.04)
(0.07)
(0.05)
(0.06)
(0.07)
(0.07)
(0.05)
Commune FE included
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
1,150
1,150
1,150
1,150
1,150
1,150
1,150
R-squared
0.0633
0.0482
0.056
0.054
0.0553
0.0884
0.0617
Note: Plot-manager level regressions. Standard errors clustered at the household level in parentheses. * p<0.1 ** p<0.05 *** p<0.01. Only male plots managers
with at least one plot where sorghum is main crop across baseline and follow-up are included in all regressions. Logland and sorghum land share are baseline
observations.

126

APPENDIX E CHANGES IN LABOR ALLOCATION BY HEADSHIP AT PLOT-MANAGER LEVEL (EXTENDED
HOUSEHOLDS ONLY)
Table 2E. Changes labor allocation across heads and non-heads members between baseline and follow-up (All male plot managers within extended households)
VARIABLES
Total labor
Plowing
Planting
Fertilizer
Manure
Weeding
Harvest
Recipient is head

0.09
0.206
0.132
0.270**
0.054
0.071
0.015
(0.091)
(0.138)
(0.096)
(0.134)
(0.159)
(0.137)
(0.109)
Sorghum land share
-0.347*
-0.14
-0.188
0.339
-0.836***
-0.446*
-0.298
(0.201)
(0.323)
(0.247)
(0.245)
(0.315)
(0.233)
(0.250)
Logland
0.192***
0.300***
0.185**
-0.098
0.005
0.197*
0.142*
(0.067)
(0.109)
(0.072)
(0.093)
(0.105)
(0.112)
(0.085)
Commune FE included
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
415
415
415
415
415
415
415
R-squared
0.107
0.119
0.078
0.111
0.130
0.109
0.068
Note: Plot-manager level regressions. Standard errors clustered at the household level in parentheses. * p<0.1 ** p<0.05 *** p<0.01. Only male plots managers
with at least one plot where sorghum is main crop across baseline and follow-up are included in all regressions. Logland and sorghum land share are baseline
observations.

127

REFERENCES

128

REFERENCES

Aune, J.B. and Coulibaly, A. (2015). Microdosing of Mineral Fertilizer and Conservation Agriculture for
Sustainable Agricultural Intensification in Sub-Saharan Africa. In Sustainable Intensification to
Advance Food Security and Enhance Climate Resilience in Africa (pp. 223-234). Springer
International Publishing.
Barrett, C. B., Moser, C. M., McHugh, O. V., & Barison, J. (2004). Better technology, better plots, or
better farmers? Identifying changes in productivity and risk among Malagasy rice farmers. American
Journal of Agricultural Economics, 86(4), 869-888.
Beaman, L. and Dillon, A. (2012). Do household definitions matter in survey design? Results from a
randomized survey experiment in Mali. Journal of Development Economics, 98(1), pp.124-135.
Becker, G. (1981). A Treatise on the Family. Cambridge, Mass.: Harvard University Press.
Bongaarts, J., 2001. Household size and composition in the developing world in the 1990s. Population
studies, 55(3), pp.263-279.
Bielders, C.L. and GĂŠrard, B. (2015). Millet response to microdose fertilization in southwestern Niger:
effect of antecedent fertility management and environmental factors. Field Crops Research, 171,
pp.165-175.
Colson, E. (1962). Family change in contemporary Africa. Annals of the New York Academy of
Sciences, 96(2), pp.641-652.
Croppenstedt, A., Demeke, M., & Meschi, M. M. (2003). Technology adoption in the presence of
constraints: the case of fertilizer demand in Ethiopia. Review of Development Economics, 7(1), 5870.
De Janvry, A., Fafchamps, M., & Sadoulet, E. (1991). Peasant household behaviour with missing
markets: some paradoxes explained. The Economic Journal, 101(409), 1400-1417.
Doss, C.R. (2001). Designing agricultural technology for African women farmers: Lessons from 25 years
of experience. World development, 29(12), pp.2075-2092.
Doss, C. R. (2006). Analyzing technology adoption using microstudies: limitations, challenges, and
opportunities for improvement. Agricultural Economics, 34(3), 207-219.
Duflo, E., Kremer, M., & Robinson, J. (2011). Nudging farmers to use fertilizer: Theory and experimental
evidence from Kenya. The American Economic Review, 101(6), 2350-2390.
Evenson, R. E., & Gollin, D. (2003). Assessing the impact of the Green Revolution, 1960 to
2000. Science, 300(5620), 758-762.
Foster, A. D., & Rosenzweig, M. R. (1995). Learning by doing and learning from others: Human capital
and technical change in agriculture. Journal of political Economy, 103(6), 1176-1209.

129

Foster, A.D. and Rosenzweig, M.R. (2010). Microeconomics of technology adoption. Annu. Rev. Econ.,
2(1), pp.395-424.
Galor, O. (2005). The demographic transition and the emergence of sustained economic growth. Journal
of the European Economic Association, 3(2-3), pp.494-504.
Galor, O. and Weil, D.N. (2000). Population, technology, and growth: From Malthusian stagnation to the
demographic transition and beyond. American economic review, pp.806-828.
Garnett, T., Appleby, M.C., Balmford, A., Bateman, I.J., Benton, T.G., Bloomer, P., Burlingame, B.,
Dawkins, M., Dolan, L., Fraser, D. and Herrero, M. (2013). Sustainable intensification in agriculture:
premises and policies. Science, 341(6141), pp.33-34.
GinĂŠ, X., & Yang, D. (2009). Insurance, credit, and technology adoption: Field experimental evidence
from Malawi. Journal of development Economics, 89(1), 1-11.
Guirkinger, C., & Platteau, J. P. (2014). The effect of land scarcity on farm structure: empirical evidence
from Mali. Economic Development and Cultural Change, 62(2), 195-238.
Guirkinger, C., Platteau, J. P., & Goetghebuer, T. (2015). Productive inefficiency in extended agricultural
households: Evidence from Mali. Journal of Development Economics, 116, 17-27.
ICRISAT. 2012. Fertilizer Microdosing, July 2012.
Jack, B. K. (2013). Market inefficiencies and the adoption of agricultural technologies in developing
countries.
Jayne, T. S., Yamano, T., Weber, M. T., Tschirley, D., Benfica, R., Chapoto, A., & Zulu, B. (2003).
Smallholder income and land distribution in Africa: implications for poverty reduction
strategies. Food policy, 28(3), 253-275.
Just, R. E., & Zilberman, D. (1983). Stochastic structure, farm size and technology adoption in
developing agriculture. Oxford Economic Papers, 35(2), 307-328.
Karlan, D., Osei, R., Osei-Akoto, I., & Udry, C. (2014). Agricultural decisions after relaxing credit and
risk constraints. The Quarterly Journal of Economics, 129(2), 597-652.
Kazianga, H., & Wahhaj, Z. (2013). Gender, Social Norms, and Household Production in Burkina
Faso. Economic Development and Cultural Change, 61(3), 539-576.
Kazianga, H., & Wahhaj, Z. (2017). Intra-household resource allocation and familial ties. Journal of
Development Economics, 127, 109-132.
Kijima, Y., Otsuka, K. and Sserunkuuma, D., 2011. An inquiry into constraints on a green revolution in
Sub-Saharan Africa: the case of NERICA rice in Uganda. World Development, 39(1), pp.77-86.
Khonje, M., Manda, J., Alene, A. D., & Kassie, M. (2015). Analysis of adoption and impacts of improved
maize varieties in eastern Zambia. World Development, 66, 695-706.
Langyintuo, A. S., & Mungoma, C. (2008). The effect of household wealth on the adoption of improved
maize varieties in Zambia. Food policy, 33(6), 550-559.

130

Laslett, P., Wall, R., Nakamura, J.I. and Miyamoto, M., 1982. Household and family in past time.
Economic Development and Cultural Change, 30(2), pp.229-269.
Marenya, P. P., & Barrett, C. B. (2007). Household-level determinants of adoption of improved natural
resources management practices among smallholder farmers in western Kenya. Food policy, 32(4),
515-536.
McMillan, D.E., 1987. Monitoring the evolution of household economic systems over time in farming
systems research. Development and Change, 18(2), pp.295-314.
Mendola, M., 2008. Migration and technological change in rural households: Complements or
substitutes? Journal of Development Economics, 85(1), pp.150-175.
Ministère de l’Agriculture, de l’Hydraulique et des Resources Halieutiques. 2010. Diagnostic du secteur
Semencier du Burkina Faso. Document final. Ouagadougou, Burkina Faso, November.
Moreno, G., and Sunding, D. L. (2005). Joint estimation of technology adoption and land allocation with
implications for the design of conservation policy. American Journal of Agricultural
Economics, 87(4), 1009-1019.
Morton, J. F. (2007). The impact of climate change on smallholder and subsistence
agriculture. Proceedings of the national academy of sciences, 104(50), 19680-19685.
Moser, C. M., & Barrett, C. B. (2003). The disappointing adoption dynamics of a yield-increasing, low
external-input technology: the case of SRI in Madagascar. Agricultural Systems, 76(3), 1085-1100.
Noltze, M., Schwarze, S., & Qaim, M. (2012). Understanding the adoption of system technologies in
smallholder agriculture: The system of rice intensification (SRI) in Timor Leste. Agricultural
systems, 108, 64-73.
Poteete, A.R. and Ostrom, E. (2004). Heterogeneity, group size and collective action: The role of
institutions in forest management. Development and change, 35(3), pp.435-461.
Resnick, D., Tarp, F., & Thurlow, J. (2012). The political economy of green growth: Illustrations from
Southern Africa (No. 2012/11). WIDER Working Paper.Rosenzweig, M. R., & Wolpin, K. I. (1985).
Specific experience, household structure, and intergenerational transfers: farm family land and labor
arrangements in developing countries. The Quarterly Journal of Economics, 961-987.
Rosenzweig, M.R. and Wolpin, K.I. (1985). Specific experience, household structure, and
intergenerational transfers: farm family land and labor arrangements in developing countries.
Quarterly Journal of Economics, pp.961-987.
Teklewold, H., Kassie, M., & Shiferaw, B. (2013). Adoption of multiple sustainable agricultural practices
in rural Ethiopia. Journal of Agricultural Economics, 64(3), 597-623.
Toulmin, C. (1986). Access to food, dry season strategies and household size amongst the Bambara of
Central Mali. IDS Bulletin, 17(3), 58-66.
Udry, C. (1996). Gender, agricultural production, and the theory of the household. Journal of Political
Economy, pp.1010-1046.

131

West, C. T. (2010). Household Extension and Fragmentation: Investigating the Socio-Environmental
Dynamics of Mossi Domestic Transitions. Human Ecology 38.3, pp.363-376.

132

SUMMARY AND CONCLUDING REMARKS
Summary
This dissertation, organized in two studies, explores household structure and labor allocation in two
different settings. The first part investigates labor allocation across plots within the same household
among extended family members in rural Mali, where households are subject to production risk due to
weather uncertainties. The second part examines labor allocation response to agricultural technology
adoption, namely fertilizer microdosing, across households with differing family structure –that is,
nuclear households versus extended households– in rural Burkina Faso.
In the first study, I develop a conceptual framework that describes how collective farming can be
favored when households are exposed to production uncertainty due to high weather variability. In the
empirical part, I use a two-wave plot level dataset collected in 2009 and 2012 and apply multiple levels of
fixed effect and alternative functional forms and find strong empirical evidence that collective plots –
managed by the heads of extended family households with output shared amongst all extended family
members– are more intensively farmed with labor. To rule out the possibility that the longer time spent on
the collective plots is not to satisfy nominal labor contribution on these plots, I also estimate crop yield
performance and find that the higher labor allocation on the collective plots translates into higher crop
yield performance, relative to that found on individual plots.55 Findings also reveal that collective plots
are more likely to receive more diverse sources of labor (i.e., labor from men, women, and children)
relative to male individual plots.
I claim that the higher level of labor allocated to the collective plots can partly be explained as
part of an ex-ante risk mitigation strategy against production uncertainty. To empirically test this claim, I
investigate the labor allocation gap across collective and individual plots using the historical coefficient of
variation of rainfall over twenty-seven years (1981-2007) as a proxy for production uncertainty. I find that

55

Recall that individual plot managers are required to work on the collective plots before working on their own
plots. Guirkinger and Platteau (2014) report complaints from the head of households about individuals tending to
shirk on the collective plots to conserve their energy to work later on their own individual plots.

133

households living in villages with higher rainfall variability present a significantly higher labor allocation
gap in favor of their collective plots relative to individual plots. Furthermore, using household level data,
I find that nuclear households in places with higher historical rainfall variability are more likely to be
engaged in collective farming. The results remain robust to the inclusion of household asset holdings and
village-level covariates that may also be correlated with rainfall variability, providing strong suggestive
evidence that collective farming may indeed constitute an ex-ante risk mitigation strategy among these
households.
The second study uses randomized control trial (RCT) data combined with the difference-indifference estimation method to investigate labor allocation response to the introduction of a specific
fertilization technique, namely microdosing The program targeted rural subsistence sorghum growers in
Burkina Faso for whom sorghum constitutes a major staple food, with the objective of improving
sorghum productivity among these households and, consequently, increasing household food security.
The results show that while the intervention increased labor allocation to fertilizer application for
sorghum production among all targeted households, its overall impact by household structure was starkly
heterogeneous. To be precise, while the microdosing significantly increased labor allocation among
nuclear family households, as expected given the labor-intensive nature of the technology, its overall
impact on extended households was negative. Subsequent analyses disaggregated by specific agricultural
tasks reveal that extended households substituted labor away from other sorghum production tasks to
meet the additional labor needed to apply fertilizer, while there is no indication of such substitution
effects among nuclear households. Furthermore, we find no evidence of labor substitution effects that
diverted labor away from the production of other crops to sorghum production. Therefore, the findings
imply that the microdosing technology induced an increase in the overall agricultural labor allocation for
nuclear households.
One of the explanations we provide regarding the differential impact of the intervention relates to
the fact that, like labor arrangements in Mali, individual plot managers in the extended households are
required to work on the collective plots before working on their own individual plots. With this limited

134

access to labor, non-head individual plot managers selected to receive the microdosing kit may not be
able to acquire the additional labor required by the microdosing technique. This explanation appears to
be sustained by the data as we find that recipients of microdosing kits who are household heads
experienced a greater increase in labor allocated to fertilizer application, relative to non-head recipients.
Note that the gender differential aspect of the intervention has not been covered in this study as
the paper focuses on nuclear versus extended family structures. Given the extent to which access to
productive resources is strongly impacted by gender (see review by Quisumbing, 1996) while virtually all
female-headed households are nuclear households, a consideration of gender would constitute a major
confounding factor. Consequently, including gender would be a threat to the internal validity of the study.
The gender question may constitute a separate research question on its own.

Concluding remarks
Taken together, these two studies emphasize the context specificity of labor allocation decision-making
among subsistence farmers in West Africa. With respect to the first study, the theory on the production of
collective goods would predict greater labor allocation on the individual plots due to moral hazard
associated with the collective plots (Guirkinger and Platteau, 2015; Guirkinger and Platteau 2017). While
I do not test for moral hazard in this study, the consistency of the higher labor allocation on the collective
plots across different econometric specifications and sample restrictions strongly suggests that collective
plots do not suffer from the issue of moral hazard.
One may argue that the guaranteed access to the output from collective plots may serve as a
source of moral hazard on the individual plots. To be more precise, given the benefits of collective
farming and knowing that the contribution of labor to collective plots allows nuclear households to share
the production, individuals may be incentivized to minimize the labor allocated to their own plots. If this
argument sustains, it would further support the claim that collective farming constitutes an insurance
mechanism among these farmers. Otherwise, why would individuals be willing to provide more effort to
the production of a collective good to the detriment of the production of a private good for which they are

135

the sole claimants?
Regarding the labor response to the introduction of the microdosing technology in the second
study, one would have expected a higher labor allocation response among extended family households,
owing to their larger size, relative to nuclear family households. The findings from the two studies
therefore seem contradictory at first glance. However, a close consideration of the nuances between labor
allocation arrangements in the two studies shows the contextual specificity of each of them leading to
three important considerations.
First, recall that the finding at the plot manager level in the second study reveals that recipients of
the microdosing kits who are household heads experienced a higher increase in labor allocation devoted
to fertilizer application relative to non-head recipients. If we assume that plots controlled by the head in
the extended households are collective plots (as in Kazianga and Wahhaj, 2013, 2017), this finding would
be in line with the results of the first study where I find that collective plots receive more labor relative to
individual plots.
Second, in contrast to the first study, the nuclear households in the second paper, who
experienced an increase in their labor response to the introduction of the technology, do not participate in
collective farming. Several factors may explain the nonparticipation of these households in collective
farming, including the possibility that they are self-sufficient regarding their labor needs. In addition,
within the extended family households, multiple nuclear households coexist and must also allocate part of
their labor endowment on their own individual plots within their nuclear household. In contrast, within
the pure nuclear family household, there is only one head of household, which makes the latter more
likely to have centralized control over labor endowment within the household. Consequently, he is more
likely to be able to make workers more effectively respond to the labor requirement of the microdosing
technology relative to the extended households where more units are involved in the production system.56

56

Note that this argument does not contradict the claim that collective farming allows extended family members to
apply a larger amount of labor with brute force when weather uncertainty leaves them with a small window of time
to carry out some specific agricultural tasks. Weather uncertainty is historical and extended households can ex-ante
put in place a labor arrangement system that allows them to deal with unpredictable weather outcomes. However,

136

Finally, nuclear households are the sole claimants of any potential increased output induced by
the microdosing technology, in contrast to the extended households, where output is shared amongst
members based on need rather than the contribution of effort (Guirkinger and Platteau, 2017; Kazianga
and Wahhaj, 2013; 2017). As a result, nuclear households may have higher incentives to exert more effort
in their sorghum production following the receipt of the microdosing kit, leading to a higher overall
increase in their labor allocation relative to the extended family households.57
Combined with the demographic transition characterized by the increasing fragmentation of
larger households into nuclear households in West Africa, this finding of positive labor response to the
introduction of the microdosing suggests greater potentials for future agricultural technology diffusion.
Whether these findings (that is, the positive response of labor among the nuclear households) imply
focusing targeting efforts on this type of household structure for labor-intensive agricultural development
interventions remains an open-ended question. The decisive answer to this question will ultimately
depend on the objective and the ethical issues surrounding the motives for such interventions.
It is important to highlight that the current work does not claim that collective farming among
extended family members is ineffective, per se, especially given the beneficial features of collective
farming described in the first paper. Rather, paper points out the central role played by human labor as a
productive resource among small-scale subsistence farmers and how its limited availability may have
implications for agricultural intensification. Though the issue of labor constraints among small scale
subsistence farmers has been extensively explored in the existing empirical literature, the current work
brings important insights. Specifically, it both reinforces the importance of a consistent household

the introduction of the microdosing technology brings a different nature of labor coordination, as it comes with
specific embodied recommendations.
57
Note that this potential difference in incentives does not contradict the claim that nuclear household members
have an incentive to diligently work on the collective plots in the first study in the case of Mali. There, the claim is
that nuclear households engage in collective farming to overcome labor bottlenecks when production uncertainties
prevail due to weather variability. In the case of the microdosing in Burkina Faso, the nuclear households operate in
‘autarky’ and may be less labor constrained. Note that nuclear households operating in ‘autarky’ also exist among
the studied population in Mali. However, they are not covered in that study because the objective of the empirical
investigation there is to look at differences in labor allocation across plots within the same household.

137

definition for successful targeting in social program interventions (Beaman and Dillon, 2012), and it
illustrates how the dynamics in household labor allocation decision making can be complex, contextspecific and highly dependent on the structure of the household. This highlights the challenges faced by
development practitioners to replicate development programs aiming at alleviating these labor constraints
at large scales, which underscores the need for an improved understanding of rural household economies.

138

REFERENCES

139

REFERENCES

Beaman, L., & Dillon, A. (2012). Do household definitions matter in survey design? Results from a
randomized survey experiment in Mali. Journal of Development Economics, 98(1), 124-135.
Guirkinger, C., Platteau, J. P., & Goetghebuer, T. (2015). Productive inefficiency in extended agricultural
households: Evidence from Mali. Journal of Development Economics, 116, 17-27.
Guirkinger, C., & Platteau, J. P. (2017). Transformation of African Farm Households: A Short Survey of
Economic Contributions. Journal of Demographic Economics, 83(1), 41-50.
Kazianga, H., & Wahhaj, Z. (2013). Gender, Social Norms, and Household Production in Burkina Faso.
Economic Development and Cultural Change, 61(3), 539-576.
Kazianga, H., & Wahhaj, Z. (2017). Intra-household resource allocation and familialties. Journal of
Development Economics, 127, 109-132.
Quisumbing, A. R. (1996). Male-female differences in agricultural productivity: Methodological issues
and empirical evidence. World Development, 24(10), 1579-1595.
West, C. T. (2010). Household extension and fragmentation: Investigating the socio-environmental
dynamics of Mossi domestic transitions. Human ecology, 38(3), 363-376.

140