ASSESSING DEVELOPMENT OUTCOMES WHEN

WEATHER, LAND, AND PEOPLE DIFFER

By

Jarrad Godwin Farris

A DISSERTATION

Submitted to

Michigan State University

in partial fulfillment of the requirements

for the degree of

Agricultural, Food, and Resource Economics - Doctor of Philosophy

2020

ABSTRACT

ASSESSING DEVELOPMENT OUTCOMES WHEN

WEATHER, LAND, AND PEOPLE DIFFER

By

Jarrad Godwin Farris

Agricultural development and policy design rely on careful analysis of the factors that

influence the welfare and decision making of agricultural households. This dissertation lever-

ages diverse data types, cross-disciplinary knowledge, and applied econometrics to assess the

underlying factors that influence child welfare and agricultural production decisions. In doing

so, it reveals the role of observed and unobserved differences in shaping our understanding

of decision making in agricultural production systems.

The first chapter evaluates the impacts of in utero rainfall on child growth in rural

Rwanda and assesses whether estimates based on aggregate in utero rainfall are attenuated

by intra-seasonal in utero rainfall effects. The in utero period of a child’s life is critical

for his or her development. For families relying on rain-fed agricultural production, such

development can be severely impacted by the timing of rainfall shocks. Yet, evidence of

in utero rainfall effects has been mixed. My results suggest that this mixed evidence may

be driven by a focus on aggregate rainfall measures, which ignore cropping period specific

heterogeneity in rainfall effects on human health.

The second chapter assesses the separability hypothesis which posits that agricultural

households make their production decisions separately from their consumption decisions.

This theory relates closely to the completeness of markets and provides an important av-

enue for understanding how agricultural households are likely to respond to new policies and

programs. The current standard identification strategy for testing whether this separabil-

ity hypothesis holds is to estimate reduced form regressions of household labor demand on

household demographic characteristics, using household fixed effects to address unobserved

heterogeneity. Using plot panel data from Rwanda, I apply an alternative test that controls

for unobserved heterogeneity in land quality. Using simulations, I then show that the stan-

dard approach based on household fixed effects is prone to omitted variable bias from the

endogeneity of household demographic characteristics with unobserved land characteristics.

Simulations indicate that this bias is exacerbated as the land market becomes more active.

The third chapter examines the role of farmer personality in the effectiveness of a

community-based extension program for promoting improved bean varieties in Tanzania. I

develop a conceptual framework which shows that the information gained from community-

based extension activities may be heterogeneous by farmer personality. I then examine this

potential heterogeneity empirically using a unique dataset of the Big Six personality traits

measured using the Midlife Development Inventory (MIDI). My findings suggest that per-

sonality characteristics influenced which farmers benefited from the extension program. In

particular, more extraverted farmers appear to have benefited more from residing in villages

that received trial packs of improved bean seed relative to less extraverted farmers. This is

consistent with their increased sociability and has implications for the types of farmers likely

to gain from community-based extension programs.

ACKNOWLEDGMENTS

A huge thank you to my advisor, Mywish Maredia, for always encouraging me to ask ques-

tions, seek out diverse research, and find ways to improve. Thank you for being such a

wonderful role model.

I am also forever grateful for the other members of my committee: Maria Porter, for

instilling a problem solving tenacity; Songqing Jin, for believing in a plot panel and making

sure it happened; and Robby Richardson, for encouraging the integration of cross-disciplinary

ideas. I am very fortunate for all of the support and guidance from my committee throughout

this process.

I would also like to thank Nicky Mason, for her unwavering support and for making

difficult development topics accessible; David Ortega, for introducing me to the idea of

pursuing the intersection of psychology and economics; Roy Black, for teaching me about tart

cherry production; and Trey Malone, for not giving up on a student’s research. A heartfelt

thank you also goes to the many, many faculty, staff, and students in the Department of

Agricultural, Food, and Resource Economics who shared knowledge, feedback, advice, and

kindness throughout this process.

I am also grateful for the fieldwork and data collection support from Incisive Africa, the

Agricultural Research Institute-Uyole, CIP, PIM, Imbaraga, and YWCA. Thanks also go to

all the farmers and their family members who took the time to share their knowledge and

make the surveys possible as well as to the SurveyCTO staff for answering technical CAPI

programming questions at all hours of the night.

To my amazing family, you are the support system that kept me going. A huge hug

to my parents, for showing me how to appreciate the wonders of the backyard and globe

alike; Leigh, for your open ear, kind heart, and helping me de-stress; Fern, Khalil, Gwen,

Andy, and Whitney, for teaching me how to look up with an open mind and balance critical

iv

thinking with deep breaths; and Sarah, for all the emotional support, fresh perspectives,

good vibes, and laughs–thanks for putting up with me and making me go outside.

v

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

x

Chapter 1 Growing Pains: Cropping Period-Specific In Utero Rainfall
Shocks Impact Child Growth in Rural Rwanda . . . . . . . . . .
I.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II. Crop Yield Response to Rainfall . . . . . . . . . . . . . . . . . . . . . . . . .
III. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III.1 Household Survey and Anthropometrics Data . . . . . . . . . . . . .
III.2 Rainfall Data and In Utero Rainfall Shocks . . . . . . . . . . . . . . .
IV. Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V. Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V.1 Main Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V.2
Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1
2
5
7
7
8
9
11
11
14
15
17
26

Chapter 2 Does Unobserved Land Quality Bias Separability Tests? . . . . 30
31
33
36
38
40
42
42
46
48
50
51
60
65

I.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II. Theoretical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III. Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI. Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
VII. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX A: Tables and Figures . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX B: Supplementary Tables
. . . . . . . . . . . . . . . . . . . . .
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

VI.1
VI.2

Simulation Methods
Simulation Results

Chapter 3 Farmer Personality and Community-Based Extension Effective-

I.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II. Background and Experimental Design . . . . . . . . . . . . . . . . . . . . . .
III. Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ness in Tanzania . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
69
71
73
80

vi

V. Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VII. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX A: Tables and Figures . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX B: Supplementary Tables

85
87
91
93
94
. . . . . . . . . . . . . . . . . . . . . 109
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

vii

LIST OF TABLES

Table 1.1: Summary Statistics

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

Table 1.2: Annual In Utero Rainfall Effects on Child Growth (Height-for-Age Z-Score) 19

Table 1.3: Cropping Period-Specific In Utero Rainfall Effects on Child Growth (Height-
for-Age Z-Score) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 1.4: In Utero Rainfall Effects on Child Growth (Height-for-Age Z-Score) by
Demographic and Household Characteristics . . . . . . . . . . . . . . . . .

Table 1.5: In Utero Rainfall Effects on Height-for-Age Z-Score: Incorporating Minor
. . . . . . . . . . . . . . . . . . . .

Season and Rainfall in Other Periods

20

21

22

Table 2.1: Household Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . .

52

Table 2.2: Plot Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53

Table 2.3: Farm and Plot Labor Demand (Log of Person Days per Season) on House-
hold Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Table 2.4: Sum of Labor Demand Across Plot Panel Plots (Log of Person Days per
. . . . . . . . . . . . . . . . . . . .

Season) on Household Characteristics

54

55

Table 2.5: Simulated Second Period Land Changes . . . . . . . . . . . . . . . . . . .

56

Table 2.6: Number of Observed Type I Errors in 1,000 Replications . . . . . . . . . .

57

Table 2.7: Labor Demand (Log of Person Days per Season) on Household Character-
istics: Accounting for Potential Child Productivity Differences . . . . . . .

Table 2.8: Farm Labor Demand (Log of Person Days per Season) on Household Char-
acteristics: Controlling for Land Quality Proxy . . . . . . . . . . . . . . .

Table 2.9: Farm Labor Demand (Log of Person Days per Season) on Household Char-
. . . . . . . . . . . . . . . . . .

acteristics: Incorporating All Households

Table 2.10: Farm and Plot Labor Demand (Log of Person Days per Season) on House-
hold Characteristics: Excluding Migrant Households . . . . . . . . . . . .

61

62

63

64

viii

Table 3.1: VBAA Balance Tests

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

95

Table 3.2: Farmer Survey Balance Tests . . . . . . . . . . . . . . . . . . . . . . . . .

96

Table 3.3: Household Attrition Tests . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

Table 3.4: Key Characteristics by Agricultural Year and Treatment Group . . . . . .

98

Table 3.5: Midlife Development Inventory (MIDI) Personality Traits . . . . . . . . .

99

Table 3.6: Personality Trait Cronbach’s Alpha Values

. . . . . . . . . . . . . . . . . 100

Table 3.7: Farmer-VBAA Personality Traits Comparison . . . . . . . . . . . . . . . . 101

Table 3.8: Physical Distance Summary Statistics . . . . . . . . . . . . . . . . . . . . 102

Table 3.9: Farmer Adoption of Improved Bean Varieties in 2018 Ag. Year

. . . . . . 103

Table 3.10: Heterogeneous Trial Pack Effects on Farmer Adoption of Improved Bean

Varieties in 2018 Ag. Year

. . . . . . . . . . . . . . . . . . . . . . . . . . 104

Table 3.11: Farmer Interactions with VBAA . . . . . . . . . . . . . . . . . . . . . . . 105

Table 3.12: Farmer Interactions with Other Bean Farmers . . . . . . . . . . . . . . . . 106

Table 3.13: Heterogeneous Trial Pack Effects on the Proportion of Bean Farming House-

holds that the Farmer has Sought Farming Advice From . . . . . . . . . . 107

Table 3.14: Heterogeneous Trial Pack Effects on the Proportion of Bean Farming House-

holds that the Farmer has Discussed Bean Farming With . . . . . . . . . 108

Table 3.15: Farmer Adoption of Improved Bean Varieties in 2018 Ag. Year: Controlling

for Pre-Intervention Outcome . . . . . . . . . . . . . . . . . . . . . . . . . 110

Table 3.16: Heterogeneous Trial Pack Effects on Farmer Adoption of Improved Bean

Varieties in 2018 Ag. Year: Controlling for Pre-Intervention Outcome

. . 111

ix

LIST OF FIGURES

Figure 1.1: Rwanda’s Cropping Period Calendar Around Two Rainy Seasons – Includes
Minor Season . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 1.2: Rwanda’s Cropping Period Calendar – Includes Minor Season and Restricts
Major Seasons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

24

Figure 1.3: Histogram of Child Birth Months . . . . . . . . . . . . . . . . . . . . . . .

25

Figure 2.1: Empirical Distribution via Household Fixed Effects . . . . . . . . . . . . .

58

Figure 2.2: Empirical Distribution via Plot Fixed Effects . . . . . . . . . . . . . . . .

59

x

Chapter 1

Growing Pains: Cropping Period-Specific In Utero Rainfall Shocks Impact
Child Growth in Rural Rwanda

1

I.

Introduction

Throughout the developing world, households reliant on rain-fed agricultural production face

increasing risk of food insecurity due to increased variability and uncertainty in rainfall (e.g.

Funk et al., 2019). Climate change has increased both the risk of prolonged drought, as well

as heavy rainfall (Lehmann et al., 2018). The increasing prevalence of such weather shocks

is likely to have very dire long-term consequences for those who are exposed to such shocks

in the most critical early stages of human development.

The in utero period of a child’s life is critical for his or her development, with long-term

implications for human capital accumulation (Almond et al., 2017; Barker, 1998). Exposure

to poor environmental conditions during this period has been linked to a wide array of

negative consequences. For example, negative rainfall shocks in the year before birth result

in lower birth weights and increased infant mortality (Rocha & Soares, 2015). The negative

consequences of in utero shocks need not be limited to infancy, nor do they require extreme

events. In their recent meta-analysis, Almond et al. (2017) emphasize that even moderate

early life shocks can have long-term negative consequences.

This article contributes to previous research studying the effects of in utero rainfall shocks

by differentiating between cropping-period rainfall shocks that are specific to planting, mid-

season, and harvest seasons. My identification strategy captures potential differences in the

marginal effects of rainfall depending on the crop development stage in which the rainfall

shocks occur. Such differences would be subsumed in aggregate measures of rainfall (e.g.

deviations from annual averages) that have been used in many previous studies.

I argue that the direction of the effect of rainfall shocks on child growth differs according

to the cropping period at the time of the rainfall shock, which implies that aggregating

rainfall outcomes into a single measure will attenuate the estimated effects on child growth.

By not taking such seasonal differences into account, rainfall measures that are homogeneous

across different cropping periods in a given year may not be capturing important differences

2

in cropping period-specific shocks to child growth. Such seasonal differences are particularly

important in countries where agricultural production is primarily rain-fed, where irrigation

is largely infeasible, and where there are multiple production seasons, implying relatively

brief windows for planting, growing and harvesting seasons.

Such agricultural production is exemplified in Rwanda, the focus of this study. In rural

Rwanda, as in many rural developing country contexts, agriculture is the principal source

of livelihood. As rural Rwandan households rely primarily on rainfed agriculture, rainfall

outcomes are a key source of exogenous variability in household incomes and hence nutrition

availability.

The literature measuring rainfall impacts on health has led to contradictory conclusions.

For example, two meta-analyses combining DHS data on African countries find both positive

and negative effects of rainfall on mortality (Comfort, 2016; Kim, 2010). In Mexico, rainfall

that was one standard deviation below average during the wet season reduced height-for-age

z-scores (HAZ) for a subsample of children in the north region, but increased HAZ for a

subsample of children in the center and south regions (Skoufias et al., 2011). Alongside

these mixed findings, other studies that have measured the effects of either annual rainfall

shocks (e.g. Abiona, 2017; Burgess et al., 2017; Comfort, 2016; Kim, 2010; Maccini & Yang,

2009; Rocha & Soares, 2015) or rainfall shocks during the growing season (e.g. Kudamatsu

et al., 2012; Shively, 2017; Skoufias et al., 2011) have found a positive relationship between

rainfall and health outcomes.

One response to the mixed findings on the rainfall-health relationship is a strand of the

literature that seeks to assess this relationship using non-rainfall measures. Early work in

Rwanda analyzed a crop failure in 1988-1989 that occurred for a particular cohort of children

in one region of the country (Akresh et al., 2011). Girls born in the region during the famine

had reduced growth rates compared to girls born in the same region at a different time, or

girls born at the same time but in a different region. No effect was found for boys. Similarly,

3

in a panel analysis of children in Zimbabwe, children who were 12-24 months old during

a 1994-1995 food shortage were 1.5 to 2 cm shorter on average than children who reached

that age in a non-food shortage year (Hoddinott & Kinsey, 2001). Among communities

in Ethiopia who were asked to directly report the months when food was relatively scarce,

children exposed to more food-scarce months in utero had significantly lower heights at age

eight (Miller, 2017).

Rainfall-based measures have several advantages over non-rainfall-based indicators. Rain-

fall data can assess the impact of small deviations without extreme events or self-reporting

of food scarcity. The rainfall data used in this study is at the 0.05 degree resolution, allow-

ing daily rainfall to be estimated at the household level. As rainfall data and rainfall-based

measures improve, rainfall-based measures become increasingly useful in teasing out complex

relationships between nutrition availability and health.

This study contributes in several ways to an emerging literature exploring the rainfall-

health relationship using intra-season rainfall measures. First, I allow for cropping period-

specific rainfall shocks to differ in whether they improve or harm children’s health. Prior

studies which have incorporated intra-season heterogeneity have forced all positive or all

negative deviations to have the same-signed effect, regardless of the cropping period during

which the rainfall occurred (Cornwell & Inder, 2015; Tiwari et al., 2017). Grouping all devi-

ations in the same direction into a single category ignores potentially important differences

in the timing of rainfall. Second, I examine rainfall effects on health from three different

major cropping periods, as opposed to examining effects from rainfall in only one cropping

period (Aguilar & Vicarelli, 2011).

I show that aggregate measures of rainfall shocks conceal acute differences in the rela-

tionship between rainfall and child growth. As the intra-year relationship between rainfall

and agricultural productivity need not be constant, the relationship between rainfall expe-

rienced in utero and child growth may also vary according to the cropping period. I find

4

that during planting and in mid-season, increases in rainfall have a positive relationship with

child growth. Maize yields can be particularly sensitive to water deficits during the mid-

season flowering and grain-filling stages. Less rainfall during the harvest period, however,

can improve yields, as this reduces grain water content (Barron et al., 2003). Indeed, I find

that increases in rainfall during harvest reduce child growth.

The remainder of the paper proceeds as follows. Section two discusses the crop yield

response to rainfall as it relates to rural Rwanda and describes the cropping period specific

rainfall measures. The third and fourth sections describe the data sources and the empirical

strategy respectively. The fifth section outlines the empirical results and the final section

concludes.

II. Crop Yield Response to Rainfall

Two seasons with the same total rainfall can have very different yield outcomes depending

on how rainfall was distributed within the season (Brown, 2008; HarvestChoice, 2010). The

crop-yield response to rainfall varies over the course of the growing season (Brown, 2008;

Doorenbos & Kassam, 1979; Steduto et al., 2012). For most crops, yield response to water

requirements is relatively low in early and late growth periods and relatively high during

periods of flowering and crop formation. For example, for one of Rwanda’s most important

food crops – beans – the yield response to water deficit is four to five times larger during

flowering and pod filling phases compared to during its vegetative and ripening periods

(Doorenbos & Kassam, 1979).

The relationship between rainfall and yield is not always positive. Extreme rainfall can

lead to waterlogging and aeration stress. As in crop stress from insufficient rainfall, aeration

stress from excess rainfall can lead to crop yield reductions (Steduto et al., 2012). For

example, maize – a major staple in Rwanda – is relatively sensitive to water stress during

the flowering period (Doorenbos & Kassam, 1979).

5

Even in the absence of extreme rainfall events, there may be periods when less rainfall

is beneficial. A period of little rainfall at harvest can improve maize yield by reducing grain

water content (Barron et al., 2003). Similarly, a period of no rain for 20 to 25 days before

dry bean harvest is ideal (Doorenbos & Kassam, 1979).

To capture key differences in crop yield response, I distinguish between three distinct

periods for each agricultural season. The land preparation and planting period is the period

when most crops are beginning their growth. The mid-season period is the period when

most crops are in their critical growth stages and are relatively sensitive to water stress.

Finally, the harvest period is the period when most crops are mature and may benefit from

a tapering of rainfall.

As shown in Figure 1.1, Rwanda’s first major agricultural season, which extends from

February to July, is split into: land preparation and planting (February and March), mid-

season (April and May), and harvest (June and July). Similarly, Rwanda’s other major

agricultural season, which extends from August to January, is split into land preparation

and planting (August and September), mid-season (October and November), and harvest

(December and January).1

For each of these individual cropping periods, I estimate a distinct marginal effect of in

utero rainfall shocks on child height-for-age z-scores, holding all other potential determinants

constant (i.e. parental characteristics, and other environmental factors).2

I define the in

utero period of a child’s life as the entire year before birth.3 I denote rainfall shocks (specific

to child i in household h) in the land preparation and planting, mid-season, and harvest

periods while in utero (t = 0) by the following: R1

ih0. These rainfall shocks

ih0, R2

ih0, and R3

determine the height (i.e., length) at birth of child i in household h (Hih0), which in turn

1I also estimate robustness specifications incorporating alternative season and cropping period definitions.
2As height is a long-term measure of health, there are natural dynamics in the production process for
child height. Strauss and Thomas (1998) and Maccini and Yang (2009) provide useful simple reduced forms
for dynamic health production functions.

3The additional three months preceding pregnancy are included as a child’s health endowment can be

influenced by the mother’s health pre-conception (Rocha & Soares, 2015).

6

affects subsequent height-for-age z-scores in year t (HAZiht).

III. Data

III.1 Household Survey and Anthropometrics Data

This study uses household survey data collected in 2014 and 2017 from households who had

children under the age of five.4 This survey was a two round panel survey of households

from eight districts across three provinces in Rwanda: Northern, Southern, and Eastern.

In the enumeration areas, there are significant concerns regarding food security and child

malnutrition.

In 2014, the survey used a three-stage cluster sampling method where the primary, sec-

ondary, and tertiary sampling units were the sector, village, and household respectively. A

census was first conducted in each of the 252 selected villages to collect basic information

of all households. A random sample of households with either a pregnant woman or child

under the age of five were then selected (Peters et al., 2015). The survey then followed up

with these same households in 2017 for the second round of the survey.

The survey instrument was comprehensive, covering a wide array of household and indi-

vidual characteristics. It collected detailed child-specific information on nutrition and health,

as well as weight, height, age, and gender.

This study uses the anthropometric information collected for all children less than the

age of five. If a child was under the age of five and was measured in both survey rounds, the

most recent measurement from the second round of surveying is used.

The rate of stunting in the sample is 39%, well above the expected rate of 2.3% in a

healthy population (World Health Organization [WHO], 1997). This is not surprising given

the rural, developing country context. The mean and median HAZ in the sample are -1.57

4The survey was conducted as part of an impact evaluation of the International Potato Center’s Scaling

Up Sweet Potato Through Agriculture and Nutrition (CIP-SUSTAIN) project.

7

and -1.64 respectively, indicating that most of the children are below average height for age

(see Table 1.1). These averages also highlight the relevance of this research for the study

area.

III.2 Rainfall Data and In Utero Rainfall Shocks

The most reliable source of rainfall data is considered to be the Climate Hazards Group

InfraRed Precipitation with Station data (CHIRPS), according to the Famine Early Warning

Systems Network (Famine Early Warning Systems Network [FEWS NET], 2018). This

publicly available dataset combines 0.05 degree satellite data with in-station data to provide

daily rainfall estimates from January 1st 1981 to near present day (Funk et al., 2015).

I estimate average rainfall within five kilometers of each household, with GPS coordinates

for each day from January 1st, 1981 to December 31st, 2017. A five-kilometer radius captures

the majority of household plots, as more than 85% of all plots were within an hour’s walk

from households’ dwellings. For households missing GPS coordinates, I estimate the average

rainfall among all other households in the same village.

Rainfall shocks are commonly measured as the deviation of the log of rainfall in a given

year from the log of historical annual average rainfall (Maccini & Yang, 2009; Rocha & Soares,

2015). However, a focus on annual rainfall shock measures may mask cropping period-specific

heterogeneity in the effects of in utero rainfall shocks. I address this limitation by applying

the same log-deviation, in utero shock measure as used by Rocha and Soares (2015), but

take a cropping period approach. I estimate in utero rainfall shocks (Rk

imh0) as the deviation

of the log of total rainfall in cropping period k that child i born in month m in household

h experienced while in utero, from the log of the historical average for that period. The

equation for Rk

imh0 is:

Rk

imh0 = ln

(cid:32) m−1(cid:88)

t=m−12

(cid:33)

zkt ∗ rht

8

− ln (¯rkh) ∀ k = 1, 2, 3

(1.1)

where m is the birth month of child i in household h;5 zkt is an indicator variable equal to

one if month t is included in cropping period k and zero otherwise; rht is the monthly rainfall

(measured in millimeters) that occurred within a 5 kilometer radius of household h in month

t; and ¯rhj is household h’s long-run yearly average rainfall for cropping period k over the

1981-2017 time period.

I illustrate Rk

imh0 with a particular example of a child born in October 2014. The harvest

period rainfall would be the sum of the rainfall experienced by the child’s household in the

months of December 2013 and January, June, and July 2014. This summation represents

the total rainfall the child experienced while in utero during this cropping period. The in

utero shock measure is then the natural log of the average harvest period rainfall for the

child’s household in all years from 1981-2017, subtracted from the natural log of the in

utero summation. This same process is repeated for the land preparation and planting and

mid-season periods.

The measures of in utero rainfall shocks vary considerably across cropping periods (see

Table 1.1). During land preparation and planting, rainfall shocks are about 6% above his-

torical averages. But there is less overall variability during the mid-season period, as average

in utero rainfall shocks are about 1.6% below historical averages. During the harvest period,

in utero rainfall shocks are 2.7% below historical averages.

IV. Empirical Strategy

My identification strategy is to differentiate any in utero rainfall effects from other spatial

and temporal child growth determinants that may be correlated with rainfall outcomes. By

including a broad set of controls in the econometric specification, I isolate the effects of in

utero rainfall deviations on child growth from the average conditions of children in similar

5Each month in the rainfall dataset is assigned a running integer value (e.g. January 1981 = 1, February

1981 = 2, etc.).

9

circumstances and localities. The main empirical specification is given by the following:

HAZimhvd = Rimh0θ + ageiγ + agei ∗ Rimh0δ + Xiβ + ωi + σm + cv + τdt + uimhvd

(1.2)

The outcome variable is HAZimhvd, the height-for-age z-score (HAZ) for child i born in month

m in household h in village v and district d. The main coefficients of interest, θ, are the

coefficients on Rimh0 - the log-deviations of in utero rainfall for the three cropping periods

(based on each child’s birthday and household GPS location). I interact Rimh0 with child age

in months, to account for the fact that rainfall shocks during the in utero period may impact

more recently born children differently than older children. All regressions also include: two

child-specific controls, gender and age squared (Xi); and a survey round indicator (ωi).

Birth month may influence a child’s development path and would be correlated with

season-specific in utero rainfall. For instance, children born during lean months are likely

to have different child growth outcomes compared to children born in post-harvest months.

To control for such differences, I include birth month fixed effects, σm. The distribution of

birth months is relatively even throughout the year, with slight jumps in January, May, and

December (see Figure 1.3 for a histogram of children’s birth months).

Village location and other time constant village-level unobservables may influence child

growth and be correlated with rainfall deviations. For example, certain villages may be

located in more mountainous terrain with little access to healthcare. To address this potential

issue, I include village fixed effects, cv.

The economies of Rwanda’s districts develop at different rates, and the average growth

path of children will vary according to local economic development (Maccini and Yang 2009).

To control for differences in economic development paths across districts, I include a time

trend t which is allowed to vary by the district-specific coefficient, τd.

To take account of village-level spatial correlation in the idiosyncratic error term, uimhvd,

10

I cluster standard errors at the village level.

A remaining concern with my identification strategy is due to potential serial correlation

in log deviations in rainfall. If rainfall outcomes are serially correlated, then the estimated

marginal effects could be driven by the effect of rainfall outcomes in other periods (Maccini

& Yang, 2009; Rocha & Soares, 2015). I rule out this possibility, as the estimates of in utero

period effects are robust to including rainfall shocks prior to a child’s in utero period.

Finally, I obtain similar estimates when I control for unobserved household-level hetero-

geneity in child growth outcomes, such as family heritage and the household environment.

To do so, I estimate a robustness regression with household fixed effects. As this specification

is restricted to the subsample of households with anthropometric measurements for at least

two children under age five, the sample size is too small to obtain precise estimates.

V. Empirical Results

V.1 Main Findings

In Table 1.2, I present results from estimating equation 1.2, where I use only one annual

rainfall measure. These estimates are based on the assumptions that the in utero rainfall

effect on HAZ is homogenous throughout the year. In column (1), I control for demographic

characteristics only. In column (2), I add birth month fixed effects to control for general

temporal differences in child growth outcomes that may be correlated with rainfall. In column

(3), I add village fixed effects to control for spatial differences in birth outcomes that could

be correlated with rainfall. In column (4), I add further controls for unobserved district-

specific time trends that could influence children’s development paths and be correlated with

in utero rainfall outcomes.

As estimates are not statistically significantly different from zero, I cannot reject the null

hypothesis that the impact is in fact zero. The point estimates suggest that increases in in

utero rainfall may raise HAZ. Estimates are consistent across these four specifications, all

11

with wide confidence intervals.

The wide confidence interval in estimates could be due to one of two reasons. First, the

relatively high fluctuation in annual rainfall may lead to high standard errors. Second, this

aggregate, annual measure of rainfall may be suppressing intra-year, seasonal heterogeneity

in how rainfall impacts yield at various times of the year, and therefore also children’s growth

outcomes.

Estimates of equation 1.2, using cropping-period specific rainfall deviations rather than

a single annual deviation, suggest that the annual measure attenuates impact estimates due

to seasonal differences in impacts. Increases in in utero rainfall in the land preparation and

planting or mid-season periods significantly raise HAZ. In contrast, increases in in utero

rainfall during the harvest period significantly lowers HAZ. These results are summarized

in Table 1.3, where covariates in each specification mirror those in Table 1.2. Estimates

are consistent and statistically significant across all specifications. Results in column (4) of

Table 1.3 imply that a one standard deviation increase in in utero rainfall during the land

preparation and planting or mid-season periods raises the HAZ of a one year old child by 0.28

or 0.26 standard deviations, respectively.6 These effects represent 18% and 16% increases at

the sample mean HAZ of -1.57. In the harvest period, a one standard deviation increase in

in utero rainfall reduces a one year old child’s expected HAZ by 0.24 standard deviations, a

decrease of 15% from the sample mean.

The findings also suggest that households may be able to compensate for some of the

growth effects of in utero rainfall. In utero rainfall effects are largest for infants, declining

as children age. These differences are statistically significant. This declining impact over

6The calculation, using land preparation and planting as an example, is as follows:

0.143 ∗ (2.560 − 0.049 ∗ 12) ≈ 0.28

where 0.143 is a one standard deviation or 14.3% increase in in utero land preparation and planting period
rainfall, 2.560 is the coefficient on the log deviation in in utero land preparation and planting rainfall, -0.049
is the coefficient on the interaction between the rainfall term and child age in months, and 12 is the age in
months of a one year old child.

12

ages is distinct from the general decline in HAZ with age that is typical of the developing

country context (Akresh et al., 2011; Groppo & Kraehnert, 2016).

I find further heterogeneity in rainfall impacts on HAZ, with estimated effects being

concentrated among girls. As the sample is evenly split between boys and girls, I estimate

separate regressions for each of them (see columns (1) and (2) in Table 1.4). There are several

potential drivers for this concentration of in utero rainfall effects among girls. For instance,

this difference could be a result of weaker girls being more likely to survive pregnancy and

early infancy than boys. Boys have a lower in utero survival rate than girls (Bruckner and

Catalano 2018). Boy infants also have the highest death rate of any age-sex group; only

the strongest boys survive and I do not observe the boys who die in utero. Alternatively,

the observed difference could be due to parents’ behaviors in response to rainfall shocks.

Gender-preferences in parenting could result in the sampled boys being less susceptible to

long-term growth effects from in utero rainfall shocks.

Another potential source of heterogeneity in the rainfall impacts on HAZ is households’

degrees of exposure to rainfall shocks. Some households are likely to be better positioned

to react to changes in rainfall than others. As the dataset was collected after each child’s

in utero period, I cannot directly address this potential heterogeneity in this study and it

remains a promising area for future work. I do, however, have data on household wellbeing

at the time of the anthropometric measurements which can serve as an imperfect proxy of

household wellbeing while the child was in utero. I rank the likelihood of each sampled house-

hold being below Rwanda’s national poverty line at the time of the child’s anthropometric

measurements using a Rwanda-specific poverty index developed in Schreiner (2010). This

poverty index uses household characteristics and assets to predict the likelihood of a given

household falling below Rwanda’s national poverty line. Column (3) of Table 1.4 summarizes

estimates of equation 1.2 for children in relatively poor households (bottom 50th percentile

of poverty index). In Column (4), I present the same estimates for relatively wealthy house-

13

holds (top 50th percentile). Overall, this subsample analysis by current household poverty

status suggests that even the relatively wealthy households in the rural sample of agricul-

tural households are unable to substantially mitigate the effects of rainfall shocks on their

children’s growth outcomes.

V.2 Robustness Checks

The overall findings are robust to a number of different specifications. First, results are

nearly identical when I add birth order and sibling controls, as shown in column (5) of Table

1.4. In column (6) of Table 1.4, I estimate a robustness regression with household fixed effects

in lieu of village fixed effects. This specification is restricted to the subsample of households

with anthropometric measurements for at least two children under age five. For comparison,

in column (7) I also estimate the village fixed effects specification from Table 1.3 column

(4) on this subsample of multiple child households. Unlike village fixed effects, household

fixed effects controls for unobserved household-level heterogeneity in child growth outcomes,

such as family heritage and the household environment. The results are qualitatively similar,

with increases in early and mid-season rainfall associated with increases in expected HAZ

and increases in harvest period rainfall associated with decrease in expected HAZ.

The main findings are robust to accounting for Rwanda’s minor agricultural season,

which I have ignored thus far. Figure 1.1 shows that a minor agricultural season occurs

simultaneously with that of the two major seasons. The minor harvest occurs just before the

onset of the lean season in October and November. Approximately 70% of the households

in the sample farm in this season. To test whether the findings may be sensitive to the

incorporation of this minor season, I first estimate regressions for the subsample of households

who farmed in the minor season. The results, presented in column (2) of Table 1.5, show

nearly identical results to those found with the broad sample. I further assess the sensitivity

of the results to the season and cropping period specifications via two alternative season

14

definitions. First, I incorporate the minor season by assigning the minor season months of

July, August, and September to multiple cropping periods (see Figure 1.1). For example,

in utero September rainfall is included simultaneously in both the harvest period (to reflect

minor season activity), and the land preparation and planting period (to reflect second major

season activity). Alternatively, as shown in Figure 1.2, I restrict the first major season harvest

and second major season land preparation and planting to June and September respectively.

The results of these alternative definitions are presented in columns (4) and (5) of Table 1.5.

The findings are robust across all definitions of cropping periods.

Finally, results are not driven by the possibility of serial correlation in rainfall outcomes.

I test for this possible serial correlation by controlling for the rainfall measures and their

interactions for the 13-24 month period before a child’s birth. As shown in column (6) of

Table 1.5, when controlling for the same rainfall shocks in another year, all season-specific

rainfall effects remain very similar to original estimates. In addition, coefficient estimates

are not statistically significant or meaningful in magnitude for the season-specific rainfall

shocks during the 13-24 month period prior to birth, or their interactions with child age.

VI. Conclusion

In this article, I have shown that aggregate rainfall measures often ignore intra-seasonal

heterogeneity in rainfall effects and attenuate estimated impacts of in utero rainfall shocks

on children’s health. My analysis contributes to prior studies analyzing the relationship

between rainfall and child health, which have primarily used annual or growing season rainfall

measures. I identify intra-seasonal heterogeneity in rainfall effects by using high resolution

rainfall data and a unique dataset of at-risk children in rural Rwanda. I have shown that

intra-year or intra-season rainfall effects differ in sign.

The results are consistent with agronomic research finding that low yield can be due to

both insufficient rainfall in the mid-season flowering and yield formation periods, as well as

15

excess rainfall during the harvest period (Barron et al., 2003; Doorenbos & Kassam, 1979).

I find that increases in in utero rainfall in the land preparation and planting or mid-season

periods increases expected child height-for-age z-scores. But during the harvest period,

increases in in utero rainfall decreases expected child HAZ. I have shown these effects to

be robust to alternative definitions of the cropping periods, as well as to estimates where

I have controlled for potential serial correlation in rainfall. The findings also suggest that

households may be compensating for some of the growth effects of in utero rainfall.

These estimates provide a lower bound, because of potential attenuation bias due to

classical measurement error (Wooldridge, 2010). The measurement of in utero rainfall ap-

proximates the actual rainfall experienced by a child when in utero. In addition, to ensure

that cropping period effects are not endogenous to household’s planting decisions, I have

assumed cropping periods to be constant across space and time.

In fact, they vary with

household cropping decisions and with the onset of rains.

The relationship between weather shocks and individual health is an issue that is be-

coming increasingly salient as extreme weather events - particularly drought and flooding

- become more frequent over time. Understanding the complexity of this weather-health

relationship can help inform policies and programs which aim to identify and distribute re-

sources towards at risk pregnant women and their children. I have shown that identifying or

predicting the timing of rainfall such events in the context of local agricultural production

seasons is important in anticipating negative consequences for children’s long-term health.

16

APPENDIX

17

Table 1.1: Summary Statistics

Variables
Height-for-Age Z-Score (HAZ)
Stunted (1=yes)
In utero annual rainfall
In utero land preparation and
planting period rainfall
-0.016
In utero mid-season rainfall
-0.027
In utero harvest period rainfall
33.692
Child’s age in months
0.499
Child is female (1=yes)
0.248
First born (1=yes)
0.277
Second born (1=yes)
2.022
Number of siblings
Poverty likelihood
47.6
Survey round indicator (1=data from 2017) 0.551

0.060

0

1

Mean Std. Dev. Min Max Median
-1.572
-5.980 5.610 -1.640
0.390
0.017

-0.277 0.374 0.004

1.598
0.488
0.106

0.143

-0.346 0.435 0.047

0.171
0.237
15.550
0.500
0.432
0.448
1.646
19.6
0.497

-0.395 0.531 -0.025
-0.738 0.473 0.010

0
0
0
0
0
0
0

36

2

51.8

59
1
1
1
9
100
1

Note: 3,093 child observations. Stunting is defined as a HAZ below -2. Rainfall variables are the
natural log deviations of rainfall for a given cropping period in the 12 months before birth from the
historical average for that period. Cropping period definitions are defined as in Figure 1a. Birth
order and number of siblings are based on siblings currently living in the household; data on these
two variables are missing for 47 children. Poverty likelihood is the approximate percent chance that
a child’s household is below the Rwanda national poverty line based on the Rwanda poverty index
defined in Schreiner (2010).

18

Table 1.2: Annual In Utero Rainfall Effects on Child Growth (Height-for-Age Z-Score)

(4)
Variables
(1)
1.546
In utero annual rainfall
1.220
(1.168)
(1.004)
−0.016
In utero annual rainfall X child age in months −0.010
(0.029)
(0.026)
0.180∗∗∗
0.185∗∗∗
(0.056)
(0.058)
−0.083∗∗∗ −0.083∗∗∗ −0.088∗∗∗ −0.136∗∗∗
(0.041)
(0.009)
0.001∗∗∗
0.001∗∗∗
(0.000)
(0.000)

(2)
0.997
(1.034)
−0.003
(0.026)
0.185∗∗∗
(0.056)

(3)
1.038
(1.098)
−0.006
(0.028)
0.181∗∗∗
(0.058)

(0.009)
0.001∗∗∗
(0.000)

(0.010)
0.001∗∗∗
(0.000)

Child is female (1=yes)

Child’s age in months

Child’s age in months squared

Birth month fixed effects?
Village fixed effects?
District-specific time trends?

Observations
R-squared
Number of villages

No
No
No

3,093
0.070
251

Yes
No
No

3,093
0.077
251

Yes
Yes
No

3,093
0.081
251

Yes
Yes
Yes

3,093
0.086
251

Note: Robust standard errors clustered at the village level in parentheses. The dependent variable is HAZ.
Rainfall variables are the natural log deviations of rainfall 12 months before birth from the historical
annual average. All specifications include a survey round indicator and an overall constant. *** p < 0.01.

19

(0.481)
−0.026
(0.017)

(3)

1.473∗∗
(0.696)
1.611∗∗
(0.660)

(1)
1.234∗
2.560∗∗∗
(0.655)
(0.818)
1.597∗∗∗
2.033∗∗∗
(0.690)
(0.603)
−1.099∗∗ −1.191∗∗∗ −1.322∗∗∗ −1.556∗∗∗
(0.500)
(0.444)
−0.021
−0.049∗∗
(0.020)
(0.017)
−0.035∗∗ −0.031∗∗ −0.036∗∗ −0.044∗∗∗
(0.017)
(0.015)
0.035∗∗∗
0.047∗∗∗
(0.012)
(0.013)
0.185∗∗∗
0.185∗∗∗
(0.055)
(0.058)
−0.072∗∗∗ −0.072∗∗∗ −0.075∗∗∗ −0.128∗∗∗
(0.041)
(0.010)
0.001∗∗∗
0.001∗∗∗
(0.000)
(0.000)

(0.017)
0.040∗∗∗
(0.013)
0.185∗∗∗
(0.057)

(0.015)
0.037∗∗∗
(0.012)
0.186∗∗∗
(0.056)

(0.011)
0.001∗∗∗
(0.000)

(2)

1.405∗∗
(0.671)
1.478∗∗
(0.616)

(0.445)
−0.024
(0.017)

(0.010)
0.001∗∗∗
(0.000)

Variables
In utero land preparation and
planting period rainfall
In utero mid-season rainfall

In utero harvest period rainfall

In utero land prep. and planting
pd. rainfall X child age in months
In utero mid-season rainfall X
child age in months
In utero harvest period rainfall X
child age in months
Child is female (1=yes)

Child’s age in months

Child’s age in months squared

Birth month fixed effects?
Village fixed effects?
District-specific time trends?

Observations
R-squared
Number of villages

Table 1.3: Cropping Period-Specific In Utero Rainfall Effects

on Child Growth (Height-for-Age Z-Score)

(4)

No
No
No

3,093
0.076
251

Yes
No
No

3,093
0.083
251

Yes
Yes
No

3,093
0.088
251

Yes
Yes
Yes

3,093
0.096
251

Note: Robust standard errors clustered at the village level in parentheses. The dependent variable
is HAZ. Rainfall variables are the natural log deviations of rainfall for a given cropping period in
the 12 months before birth from the historical average for that period. All specifications include a
survey round indicator and an overall constant. * p < 0.10; ** p < 0.05; *** p < 0.01.

20

Table 1.4: In Utero Rainfall Effects on Child Growth (Height-for-Age Z-Score)

by Demographic and Household Characteristics

Relatively Relatively Sibling

Multiple Child Subsample
Household

Village

Female

(1)

Poor Wealthy Controls Fixed Effects Fixed Effects
(3)

(7)

(5)

2.092∗∗
(0.997)
1.405
(0.933)
−1.676∗∗∗
(0.586)
−0.028
(0.024)
−0.031
(0.023)
0.057∗∗∗
(0.016)

Male
(2)
1.751
(1.219)
2.307∗∗
(1.046)

3.067∗∗∗
(1.093)
0.901
(1.007)

Variables
(6)
In utero land preparation and planting 3.194∗∗∗
1.505
(1.080)
period rainfall
(1.151)
1.848∗
In utero mid-season rainfall
0.477
(0.977)
(1.053)
−2.070∗∗∗ −0.984 −1.403∗∗ −1.605∗∗ −1.493∗∗∗ −1.792∗∗∗
(0.630)
(0.658)
−0.020
(0.027)
−0.005
(0.027)
0.052∗∗∗
(0.018)

In utero harvest period rainfall
(0.500)
In utero land preparation and planting −0.056∗∗ −0.038 −0.050∗ −0.047∗ −0.049∗∗
(0.019)
period rainfall X child age in months
(0.026)
−0.038 −0.056∗∗ −0.012 −0.064∗∗∗ −0.051∗∗∗
In utero mid-season rainfall
(0.017)
(0.025)
X child age in months
0.056∗∗∗
0.045∗∗∗
In utero harvest period rainfall
(0.013)
(0.018)
X child age in months

(4)
2.080∗
(1.162)
2.680∗∗∗
(0.884)

2.490∗∗∗
(0.811)
2.264∗∗∗
(0.690)

(0.024)
0.048∗∗∗
(0.017)

(0.022)
0.047∗∗∗
(0.018)

(0.029)

(0.026)

(0.028)

(0.760)

(0.655)

(0.647)

(0.026)
0.032
(0.020)

Observations
R-squared
Number of villages

1,539
0.119
245

1,546
0.084
246

1,570
0.132
243

1,503
0.105
237

3,046
0.100
251

1,943
0.162
242

1,943
0.137
242

Note: Robust standard errors clustered at the village level in parentheses. Dependent variable is HAZ. Rainfall variables are the natural log
deviations of rainfall for a given cropping period in the 12 months before birth from the historical average for that period. Cols. (1) and (2)
split the sample by the child’s sex. Cols. (3) and (4) split the sample by households in the bottom and top 50th percentile, respectively, using
the Rwanda poverty index defined in Schreiner (2010). Col. (5) includes controls for first born, second born, and number of siblings. Cols. (6)
and (7) are restricted to the subsample of 883 households with two or more child observations. All specifications include age in months, age in
months squared, birth month fixed effects, district specific time trends, a survey round indicator, and an overall constant. All cols. except (1)
and (2) include a female indicator. All cols. except (6) include village fixed effects. * p < 0.10; ** p < 0.05; *** p < 0.01.

21

Table 1.5: In Utero Rainfall Effects on Height-for-Age Z-Score:

Incorporating Minor Season and Rainfall in Other Periods

Minor

Season Prod.

Subsample

Season Definitions
Include Minor with:

Controlling
for Rainfall
Limited 12-24 Months
Overlap Before Birth

Main
(1)

(4)

(2)

In utero mid-season rainfall

Variables
2.433∗∗∗
In utero land preparation and planting period rainfall 2.560∗∗∗
(0.935)
(0.818)
2.507∗∗∗
2.033∗∗∗
(0.715)
(0.690)
−1.556∗∗∗ −1.931∗∗∗ −2.329∗∗∗ −2.107∗∗∗
In utero harvest period rainfall
(0.500)
(0.740)
(0.577)
In utero land preparation and planting period rainfall −0.049∗∗
−0.050∗∗ −0.062∗∗∗ −0.053∗∗
(0.020)
X child age in months
(0.021)
(0.023)
−0.044∗∗∗ −0.041∗∗ −0.052∗∗∗ −0.056∗∗∗
In utero mid-season rainfall X child age in months
(0.018)
(0.017)
In utero harvest period rainfall X child age in months 0.047∗∗∗
0.064∗∗∗
(0.018)
(0.013)

2.794∗∗∗
(1.020)
1.901∗∗
(0.839)

(0.020)
0.057∗∗∗
(0.015)

(0.017)
0.069∗∗∗
(0.018)

Overlap

(3)

3.056∗∗∗
(0.960)
2.376∗∗∗
(0.685)

(0.772)

(0.022)

(5)
1.808∗
(0.963)
1.900∗∗
(0.744)
−1.122∗
(0.586)
−0.031
(0.023)
−0.041∗∗
(0.018)
0.035∗∗
(0.015)

3,093
0.098
251

Observations
R-squared
Number of villages

3,093
0.096
251

2,150
0.105
249

3,093
0.099
251

3,093
0.097
251

Note: Robust standard errors clustered at the village level in parentheses. Dependent variable is HAZ. Rainfall variables are the natural
log deviations of rainfall for a given cropping period in the 12 months before birth from the historical average for that period. Col. (1)
repeats the preferred specification from Table 3 Col. (4). Col. (2) restricts the sample to children in households that farmed in the
minor season. Col. (3) assigns in utero rainfall in the minor season months of July, August, and September to multiple cropping periods
as shown in Figure 1a. Col. (4) assigns in utero rainfall in September to multiple cropping periods as shown in Figure 1b. Col. (5)
controls for the cropping period rainfall variables for the period 13-24 months before birth (none of which are significant). All
specifications include a female indicator, age in months, age in months squared, birth month fixed effects, village fixed effects, district
specific time trends, a survey round indicator, and an overall constant. * p < 0.10; ** p < 0.05; *** p < 0.01.

22

Figure 1.1: Rwanda’s Cropping Period Calendar Around Two Rainy Seasons –

Includes Minor Season

Note: Figure adapted from Famine Early Warning Systems Network (2017).

23

Figure 1.2: Rwanda’s Cropping Period Calendar –
Includes Minor Season and Restricts Major Seasons

Note: Figure adapted from Famine Early Warning Systems Network (2017).

24

Figure 1.3: Histogram of Child Birth Months

25

REFERENCES

26

REFERENCES

Abiona, O. (2017). Adverse Effects of Early Life Extreme Precipitation Shocks on Short-term
Health and Adulthood Welfare Outcomes. Review of Development Economics, 21 (4),
1229–1254.

Aguilar, A., & Vicarelli, M. (2011). El Nino and Mexican Children: Medium-Term Effects of

Early-Life Weather Shocks on Cognitive and Health Outcomes (Working Paper).

Akresh, R., Verwimp, P., & Bundervoet, T. (2011). Civil War, Crop Failure, and Child
Stunting in Rwanda. Economic Development and Cultural Change, 59 (4), 777–810.

Almond, D., Currie, J., & Duque, V. (2017, January). Childhood Circumstances and Adult
Outcomes: Act II (NBER Working Paper No. 23017). National Bureau of Economic
Research. Cambridge, MA.

Barker, D. J. P. (1998). In Utero Programming of Chronic Disease. Clinical Science, 95 (2),

115–128.

Barron, J., Rockstr¨om, J., Gichuki, F., & Hatibu, N. (2003). Dry Spell Analysis and Maize
Yields for Two Semi-Arid Locations in East Africa. Agricultural and Forest Meteo-
rology, 117 (1-2), 23–37.

Brown, M. E. (2008). Famine Early Warning Systems and Remote Sensing Data. Berlin,

Springer.

Burgess, R., Deschenes, O., Donaldson, D., & Greenstone, M. (2017, April 20). Weather,

Climate Change, and Death in India.

Comfort, A. B. (2016). Long-Term Effect of In Utero Conditions on Maternal Survival Later
in Life: Evidence from Sub-Saharan Africa. Journal of Population Economics, 29 (2),
493–527.

Cornwell, K., & Inder, B. (2015). Child Health and Rainfall in Early Life. The Journal of

Development Studies, 51 (7), 865–880.

Doorenbos, J., & Kassam, A. (1979). Yield Response to Water (Agriculture Organization of

the United Nations (FAO) Irrigation and Drainage Paper No. 33). Rome, Italy.

Famine Early Warning Systems Network. (2018). Building Rainfall Assumptions for Scenario

Development: Guidance Document Number 2. Washington, D.C.

27

Funk, C., Hoell, A., Nicholson, S., Korecha, D., Galu, G., Artan, G., Teshome, F., Hailer-
mariam, K., Segele, Z., Harrison, L., Tadege, A., Atheru, Z., Pomposi, C., & Pedreros,
D. (2019). Examining the Potential Contributions of Extreme “Western V” Sea Sur-
face Temperatures to the March-June East African Drought. Bulletin of the American
Meteorological Society, 100 (1), S55–S60.

Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Row-
land, J., Harrison, L., Hoell, A., & Michaelsen, J. (2015). The Climate Hazards In-
frared Precipitation with Stations—A New Environmental Record for Monitoring
Extremes. Scientific Data, 2, 150066.

Groppo, V., & Kraehnert, K. (2016). Extreme Weather Events and Child Height: Evidence

from Mongolia. World Development, 86, 59–78.

HarvestChoice. (2010). Rainfall Variability and Crop Yield Potential. International Food
Policy Research Institute. Washington, D.C. and University of Minnesota, St. Paul,
MN. Retrieved January 16, 2019, from http://harvestchoice.org/node/2240

Hoddinott, J., & Kinsey, B. (2001). Child Growth in the Time of Drought. Oxford Bulletin

of Economics and Statistics, 63 (4), 409–436.

Kim, Y. S. (2010). The Impact of Rainfall on Early Child Health (Working Paper).

Kudamatsu, M., Persson, T., & Stromberg, D. (2012). Weather and Infant Mortality in

Africa (Working Paper).

Lehmann, J., Mempel, F., & Coumou, D. (2018). Increased Occurrence of Record-Wet and
Record-Dry Months Reflect Changes in Mean Rainfall. Geophysical Research Letters,
45 (24), 13, 468–13, 476.

Maccini, S., & Yang, D. (2009). Under the Weather: Health, Schooling, and Economic Con-

sequences of Early-Life Rainfall. American Economic Review, 99 (3), 1006–1026.

Miller, R. (2017). Childhood Health and Prenatal Exposure to Seasonal Food Scarcity in

Ethiopia. World Development, 99, 350–376.

Peters, C., Farris, J., Porter, M., Maredia, M. K., & Jin, S. (2015). Impact Evaluation of
Scaling up Sweet Potato Through Agriculture and Nutrition (SUSTAIN) Project in
Rwanda: Baseline Report (Unpublished research report to SUSTAIN project, Inter-
national Potato Center, Nairobi, Kenya).

Rocha, R., & Soares, R. R. (2015). Water scarcity and birth outcomes in the Brazilian

semiarid. Journal of Development Economics, 112, 72–91.

28

Schreiner, M. (2010). Simple Poverty Scorecard Poverty-Assessment Tool Rwanda, 118. http:

//www.simplepovertyscorecard.com/RWA 2005 ENG.pdf

Shively, G. E. (2017). Infrastructure mitigates the sensitivity of child growth to local agri-
culture and rainfall in Nepal and Uganda. Proceedings of the National Academy of
Sciences, 114 (5), 903–908.

Skoufias, E., Vinha, K., & Conroy, H. V. (2011, February). The impacts of climate variability
on welfare in rural Mexico (World Bank Working Paper No. 5555). The World Bank.

Steduto, P., Hsiao, T., Fereres, E., & Raes, D. (Eds.). (2012). Crop Yield Response to Water.

Rome, Italy, Food and Agriculture Organization of the United Nations.

Strauss, J., & Thomas, D. (1998). Health, Nutrition, and Economic Development. Journal

of Economic Literature, 36 (2), 766–817.

Tiwari, S., Jacoby, H. G., & Skoufias, E. (2017). Monsoon Babies: Rainfall Shocks and Child
Nutrition in Nepal. Economic Development and Cultural Change, 65 (2), 167–188.
https://doi.org/10.1086/689308

Wooldridge, J. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, MA, MIT Press.

World Health Organization. (1997). WHO Global Database on Child Growth and Malnutri-

tion [Database]. World Health Organization. Geneva, Switzerland.

29

Chapter 2

Does Unobserved Land Quality Bias Separability Tests?

30

I.

Introduction

Modeling agricultural household decision making is integral to the design and evaluation of

development programs and policies. A key breakpoint in such models is whether agricultural

households make their production decisions separately from their consumption characteristics

and preferences. The existence of separability affects households’ production responses to

new opportunities and shocks and provides an indication of the completeness of markets

(Benjamin, 1992; Singh et al., 1986).

Numerous studies implicitly or explicitly assume separability of agricultural production

decisions (e.g. Conley & Udry, 2010; Foster & Rosenzweig, 1995; Sheahan et al., 2013; Suri,

2011). When production decisions are non-separable from consumption decisions, ignoring

this non-separability may vastly misrepresent household production decisions and the impli-

cations of policies (LaFave & Thomas, 2016; Singh et al., 1986). For instance, a separable

model cannot predict the preferential adoption of a new agricultural technology by larger

households with greater availability of family labor. Similarly, such models cannot account

for autarkic decision making based on member consumption preferences. The sensitivity

of models of agricultural household behavior to this hypothesis begets the importance of

accurate separability tests.

This paper addresses a major identification challenge for tests of this separability hypothesis–

the potential endogeneity of household demographic characteristics with unobserved land

quality.1 Using a unique plot-panel dataset, I test the separability of rural Rwandan agri-

cultural households’ production decisions while controlling for the endogeneity of household

demographic characteristics with land quality. I then use simulations to assess the suscepti-

bility of standard tests based on household fixed effects to ignoring unobserved heterogeneity

in land quality.

1This study defines land quality in the broadest sense to include all land characteristics which affect

agricultural productivity (e.g. soil type, nutrients, organic matter content, slope, etc.).

31

The difficulty of controlling for land quality and its likely correlation with household

demographic characteristics has long been a key identification challenge in the separability

literature (Benjamin, 1992; Udry, 1999). Early, seminal work based on cross-sectional data

relies on observable proxies for land quality and tends to fail to reject separability (Benjamin,

1992; Pitt & Rosenzweig, 1986). More recent work relies on household or farm fixed effects

and tends to come to the opposite conclusion (e.g. Dillon et al., 2019; Kopper, 2018; LaFave

& Thomas, 2016).

This paper makes two main contributions to this literature. First, using a recent dataset

from Rwanda, I control for potentially confounding unobserved land characteristics by lever-

aging intra-plot variability in agricultural input demand. Common tests of separability using

household panel data control for factors fixed at the household or farm level, such as the

quality of household decision making. These tests, however, are threatened by the likely

correlation of household characteristics with land quality and other unobserved land char-

acteristics when farmland is not static across survey waves. I find that the non-separability

result in Rwanda is robust to controlling for land quality and other unobserved time invari-

ant plot characteristics. This emphasizes the need to integrate consumption characteristics

into models of production decision making and support programs and policies designed to

alleviate market failures in agricultural settings.

Second, I use simulations to examine a future with well-functioning markets where the

separability hypothesis holds, but consumption traits are correlated with unobserved plot

characteristics. Using these simulated datasets, I show that separability tests based on

household fixed effects are prone to bias, and that ignoring unobserved land quality can lead

to false rejections of separability. Furthermore, this bias is exacerbated as the land market

becomes more active.

This relationship to land market activity is particularly important given the close link be-

tween the separability of agricultural household behavior and the existence of well-functioning

32

markets (Benjamin, 1992; Singh et al., 1986). Separability tests are more useful in contexts

with an active land market as this increases the likelihood that a separable agricultural

household model (AHM) may accurately describe household production responses; the sim-

ulation results, however, suggest that standard tests based on household panel data are likely

to perform worse in these contexts.

These findings highlight the need for future research to incorporate more robust means

of controlling for unobserved land quality, such as plot panel data which enables the use

of plot fixed effects. In areas with functioning land markets where some households change

operated land area between survey waves, inadequate control of land quality in reduced form

separability tests based on household fixed effects could drive biased inference on agricultural

household decision making.

The remainder of the paper is structured as follows. In the first and second sections, I

describe the theoretical framework underlying the AHM and empirical strategy underpin-

ning reduced-form separability tests, with a focus on plot-level characteristics. In the third

and fourth sections, I describe the rural Rwandan plot-panel dataset used in the empirical

application and present the Rwanda results. In the fifth section, I simulate data to illustrate

the potential bias from unobserved plot-characteristics and how it is exacerbated by shifts

in cultivated land between periods. I conclude in the final section with a summary of the

key findings and implications.

II. Theoretical Model

In this section, I illustrate the intuition behind reduced form separability tests and highlight

the role of unobserved land quality. I do so by incorporating unobserved land quality `a la

Udry (1999) into the LaFave and Thomas (2016) and Dillon et al. (2019) dynamic extensions

of the static AHM in Singh et al. (1986).

33

A household’s objective is to maximize expected discounted utility as follows:

(cid:34) ∞(cid:88)

(cid:35)

max E

βt−1U (xmt, xat, xlt; Dt, µt)

(2.1)

t=1

where household utility in time period t is captured by a time separable, concave, strictly
increasing utility function, U (·), over a vector of market goods, xmt, a vector of agricultural

goods, xat, and leisure, xlt. The utility derived from these goods differs according to house-

hold consumption preferences observed by the analyst (e.g. demographic characteristics),

Dt, and a composite of characteristics unobserved by the analyst, µt. Utility derived in
future time periods is discounted at the rate βt−1.

The household’s budget constraint in period t is:

pmt · xmt + pat · xat + wtxlt +

1

1 + τt

Wt+1 = wtEt + πt + Wt

(2.2)

where the prices of the market goods, agricultural goods, and leisure are pmt, pat, and wt

respectively, Wt+1 is wealth in the next period, which is negative if the household is in debt

and positive otherwise, τt is the interest rate for borrowing or lending, Et is the household’s

total time endowment, and πt is total farm profit.

Total farm profit, πt, is determined by the household’s agricultural input choices and is

the sum of profit across all the household’s plots as follows:

Nt(cid:88)

t, (cid:101)Ai, Zi

t − rt(cid:101)Ai

πt =

pqtf (Li

t; vt) − wtLi

t − pzt · Zi

t

(2.3)

i=1

where Nt
is the number of plots the household farms in the given period. The farm-
production technology, f (·), determines agricultural output on plot i and is a function of

t, quality-adjusted plot size, (cid:101)Ai, a vector of other inputs, Zi

t, and an exogenous,

labor input, Li

community-specific shock, vt. The agricultural output price, wage rate, land rental rate, and

34

other input prices are given by pqt, wt, rt, and pzt respectively.2

Quality-adjusted plot size, (cid:101)Ai, reflects that plots have varying qualities which influence

their productivities. Two plots of the same size may produce different outputs depending

on the quality of each plot, ceteris paribus. In determining the productivity of land input,

the size of each plot is adjusted to account for quality differences as follows:

(cid:101)Ai = θiAi ∀ i ∈ {1, 2, . . . , Nt}

(2.4)

where θi and Ai are the quality and size of plot i respectively.

The key characteristic of this AHM problem is that farm input decisions are indepen-

dent of household consumption preferences. For instance, the first order condition for labor

t, (cid:101)Ai, Zi

∂Li
t

∂f (Li

t; vt)

demand on a given plot is:

pqt

= wt

(2.5)

This optimality condition is independent of household demographic characteristics (and other

characteristics which only affect consumption decisions). Thus, in the separable AHM, pro-

duction decisions are based solely on profit maximization. Optimal household labor demand,

as well as other input demands, are invariant to changes in household preferences or demo-

graphic characteristics. Reduced form tests of the separability hypothesis rely on this result

to assess whether production decisions are consistent with the separable AHM.

The optimality condition in equation 2.5 does, however, depend on plot characteristics.

For example, the analyst may observe plot size, Ai

t, but not plot quality, θi

t. This could be

problematic for the reduced form separability test if household demographic characteristics,

Dt, are correlated with these unobserved plot quality characteristics (e.g. if larger households

tend to have better quality land). In the next section, I assess the implications of this problem

for empirical reduced form separability tests.

2For ease of exposition, this model focuses on a single, land-based agricultural output. The separability

result extends to multiple outputs (Singh et al., 1986).

35

III. Empirical Strategy

Popularized in LaFave and Thomas (2016), the current standard reduced form test of the

separability hypothesis applies the household fixed effects estimator to total farm labor

demand as follows:

lnLcht = κ + δDcht + βXcht + ηct + ηh + 
cht

(2.6)

where Lcht is the total person-days of labor used during an agricultural season in year t by

household h in community c, κ is the overall intercept, Dcht is a vector of household demo-

graphic characteristics, Xcht is a vector of other observed characteristics which affect labor

demand and are potentially correlated with Dcht, and 
cht is the idiosyncratic error.3 The

null hypothesis of interest is δ = 0 as this implies that household demographic characteris-

tics do not influence labor input demand and the separability hypothesis cannot be rejected

(Benjamin, 1992; LaFave & Thomas, 2016).

The community-time fixed effects, ηct, exploit variation within a community in a given

year to control for any time varying community-level characteristics which may be correlated

with household demographic characteristics. For example, community-wide shocks and prices

(LaFave & Thomas, 2016).

The household fixed effects, ηh, exploit within-household variability in labor demand

to allow household demographics to be arbitrarily correlated with time invariant household

characteristics (both observed and unobserved). This is an important improvement over early

studies of separability which relied on observed variables to control for farm characteristics

and other correlates to household demographic characteristics (Benjamin, 1992; LaFave &

Thomas, 2016).

The specification in equation 2.6, however, does not control for plot-level unobservables

which may be correlated with household demographics within a particular community. For

3Xcht controls for time varying farm size and characteristics that reflect differences in farmer experience,

such as household head characteristics.

36

example, soil quality, an important factor in input demand, is typically unobserved and

likely correlated with household size and other demographic characteristics (Kopper, 2018;

Udry, 1999). Failure to adequately control for such plot-level characteristics could result in

spurious correlation of Dcht and 
cht, biasing the test of separability.

Given these plot-specific characteristics and the aggregated household-level input demand

in equation 2.6, the idiosyncratic error can be approximated as follows:

Nt(cid:88)

i=1


cht =

ηi + vcht

(2.7)

where ηi are time invariant plot-level unobservables, Nt is the number of plots at time t,

and vcht is the remaining composite idiosyncratic error. If a household does not change its

farmed plots between survey waves, then(cid:80)Nt

i=1 ηi will be subsumed by ηh. Thus, if none of the

sampled households change the composition of their farmed land between survey waves, then

the separability test given in equation 2.6 will control for unobserved plot characteristics.

If some portion of the households in the sample alter their farmed landholdings between

survey waves, whether through newly rented in, rented out, bought, or sold plots, then the

aggregate sum of time invariant plot characteristics is no longer constant between survey

waves and will not be differenced out by household fixed effects. In this case, a correlation

between household demographic characteristics and time invariant plot characteristics may

cause a rejection of the null hypothesis of separability even if separability holds.

This threat to identification is addressed by plot fixed effects:4

ln(cid:103)Lchit = κ + δDcht + βXcht + ηct + ηi + 
chit

(2.8)

4This identification strategy is contingent on collecting repeated labor use data for the same plot operated
If the land market is extremely volatile in a particular context and nearly all
by the same household.
households change all of their farmed plots over the study time period, then the plot fixed effects approach
is infeasible.

37

where (cid:103)Lchit is the total person-days of labor used during an agricultural season of year t on

plot i of household h in community c, ηi are plot fixed effects, and 
chit is the idiosyncratic

error for plot i.5

By using plot panel data with plot fixed effects, any unobserved time invariant plot

characteristics correlated with the household demographic characteristics are differenced

away. Similarly, as time invariant household and community characteristics are fixed over

time for a given plot, these characteristics are also subsumed by the plot fixed effects (Udry,

1999).

Importantly, whether a plot is observed in the plot panel or not can be arbitrarily cor-

related with Dcht, Xcht, ηct, and ηi without affecting the consistency of the test in equation

2.8; for instance, a household’s decision to farm (or not farm) a given plot in a particular

period can be correlated with their demographic characteristics in Dcht, a community-level

shock, or time invariant plot or household characteristics without affecting the consistent

estimation of δ (Wooldridge, 2010, pg. 829).6

IV. Data

I assess separability in Rwanda using a two-round, panel survey conducted in Rwanda in

2014 and 2017. The initial survey wave used a three-stage cluster sampling method within

Rwanda’s Northern, Southern, and Eastern provinces where the sector, village, and house-

hold were the primary, secondary, and tertiary sampling units respectively. The survey is

not representative of Rwanda as a whole; rather, the sampling frame focused on promoting

food security and nutrition in rural areas and targeted households with a pregnant woman

5Separable labor demand on a given plot is a function of plot size which is controlled for via ηi; therefore,

Xcht omits total farm size in this specification.

6Let si = [si1, si2, . . . , siT ] where sit is an indicator variable equal to one if a given plot i was observed
in period t and T is the number of survey waves. Similarly, denote the elements in equation 2.8 as Zit =
[Dcht, Xcht, ηct] and Zi = [Zi1, Zi2,··· , ZiT]. Estimation of δ is consistent if E(
chit | si, Zi, ηi) = 0 ∀ t and
the outer product of the covariate matrix is nonsingular. This requires si to be uncorrelated with 
chit after
controlling for [Zi, ηi], but does not restrict the relationship between si and [Zi, ηi]. Wooldridge (2010, pg.
829) provides a formal proof.

38

or child under five. The survey collected detailed plot-level agricultural information as well

as household demographic information.

This analysis focuses on total plot-specific labor demand during the major February to

June agricultural season. Total labor demand is measured as the sum of the labor days of

family and hired labor for land preparation, planting, and field management after planting.

Harvest labor is excluded as harvest labor requirements are typically proportional to yield

rather than being a production choice variable. Child labor days, defined as labor provided

by household members under 15 years old, are scaled by 0.5 to reflect productivity differences

relative to adult labor (Dillon & Barrett, 2017).

During the 2017 survey round, plots were linked to the 2014 survey round by the main

household member responsible for agricultural decisions. After describing a given 2017 plot,

the respondent was read a list of the household’s unique plot descriptions and plot sizes

reported in the 2014 survey round. The respondent then either linked the given 2017 plot

to a unique plot description provided in the 2014 round, reported the 2017 plot to be a

new plot, or reported that they did not know and could not identify a match. Using this

method, approximately 51% of plots observed in 2017 were successfully linked back to 2014

plot observations.7

Table 2.1 provides household level characteristics for the 1,494 households with a least

one plot in the plot panel subsample relative to the full analytical sample of 1,800 house-

holds. Although observed characteristics are similar between groups and a vast majority of

households have at least one plot in the plot panel sample, I restrict the household level

separability analysis to the subsample of 1,494 households with at least one plot in the plot

panel sample. This reduces concerns that households without a plot in the plot panel may

7Respondent matched plots above the 95th percentile for the absolute value of plot size difference between
survey waves are trimmed from the plot panel subsample. The remaining 95% of matched plots have a mean
absolute plot size difference between survey waves of 0.098 hectares with a 0.998 correlation in plot size
between survey waves. The trimmed plots have an average absolute plot size difference between survey
waves of 16.407 hectares and a -0.057 correlation in plot size between survey waves, suggesting that these
plots were erroneously matched.

39

follow a markedly different decision framework.

Not all plots farmed by these 1,494 households were observed in both survey waves.

Table 2.2 provides summary statistics for the 4,580 plot-wave observations in the plot panel

subsample relative to the full sample of 7,303 plot-wave observations. As the plot panel

plots are not a random subset of each household’s farmed plots, they are unlikely to be

representative of households’ total landholdings. For example, plots in the plot panel sample

are smaller on average. As discussed previously, consistent estimation of the separability

test based on the plot fixed effects estimator is not reliant on balanced plot characteristics

(Wooldridge, 2010, pg. 829).

V. Results

Separability tests based on the household fixed effects specification in equation 2.6 reject

the null hypothesis of separability in both the parsimonious regression of the natural log

of household size and the expanded regression with shares of household members by age

group (Table 2.3 columns 1 and 2). The validity of these findings is reliant on the exogeneity

of household demographic characteristics given controls for the natural log of farmed land

area, community-wide time varying shocks, and unobserved time invariant household or farm

specific heterogeneity. These household fixed effects results are robust to taking into account

potential productivity differences in child household members (Appendix Table 2.7 columns

1 and 2), controlling for a land quality proxy (Appendix Table 2.8 columns 3 and 4), and to

including the households without a plot in the plot panel sample (Appendix Table 2.9).

Separability tests based on the plot fixed effects specification in equation 2.8 suggest that

the non-separability result in this sample of Rwandan households is also robust to controlling

for unobserved land quality and other unobserved time invariant plot characteristics (Table

2.3 columns 3 and 4). While the parsimonious plot fixed effects specification fails to reject

the null of separability, separability is rejected once the specification is expanded to include

40

shares of household members by age group. These plot fixed effects results are robust to

taking into account potential productivity differences in child household members (Appendix

Table 2.7 columns 3 and 4). The findings suggest that this sample of small-scale, agricultural

households integrate demographic characteristics into their production decisions.

Restricting the household fixed effects specification to only plots in the plot panel sample

provides another useful check on these results. Unlike the summation of labor demand over

all plots farmed by a household in a given year, summing household labor demand over only

plots in the plot panel subsample forces land quality and other unobserved time-invariant

land characteristics to remain fixed at the household level; this enables the household fixed

effects specification to control for unobserved land quality in a similar manner to the plot

fixed effect specification. The results, presented in Table 2.4, are consistent with those

based on plot fixed effects; the parsimonious regression of log household size fails to reject

separability, but separability is rejected in the expanded regression with shares of household

members by age.

Restricting the analysis to the subsample of households where no person left or joined

the household apart from children born between survey waves provides a further check on

the main results (Appendix Table 2.10). This subsample analysis reduces the likelihood that

endogeneity of the demographic variables with the remaining idiosyncratic error drives the

non-separability result as changes in the household demographic variables exclude migrants

in or out of the household.8 This restriction has the disadvantage, however, of an inability to

assess separability for households with a non-birth related change to their household rosters

and a large reduction in sample size.

In the parsimonious plot fixed effects specification

on this subsample, separability is rejected (at the 1% level). Furthermore, the estimated

coefficient on the log of household size is negative, indicating that a child birth between

8LaFave and Thomas (2016) further restrict the sample to households with static rosters where changes
in the demographic variables are driven solely by aging. I do not have adequate power to further restrict the
sample in this way as the survey sampling frame targeted Rwandan households with a pregnant woman or
child under five.

41

survey waves reduces labor demand.9 This is consistent with a reduction in a new mother’s

own-farm labor supply following childbirth which is not offset by hired labor.

While the more robust plot fixed effects tests corroborate those based on household

fixed effects for this sample of Rwandan households, the latter should not be relied on as

a primary indicator of whether household decision making is consistent with separability

when land quality is unobserved and farmed land is not fixed. Next, I demonstrate this

in a controlled environment via simulation, assessing the performance of tests based on the

household fixed effects and plot fixed effects estimators as land markets become more active.

VI. Simulation

VI.1 Simulation Methods

In this section, I simulate the models outlined in Section III to analyze the performance of

tests in a future with well-functioning markets where separability holds, but consumption

traits are correlated with unobserved plot characteristics. The findings from this simulation

show that separability tests based on the household fixed effects estimator are prone to bias,

and that ignoring unobserved land quality can lead to false rejections of separability. The

simulation results further show how tests based on the plot fixed effects estimator address

these issues.

What follows is a description of the simulation procedure for a single replication. I repeat

this process 1,000 times and compare the performance of the household fixed effects and plot

fixed effects specifications under different scenarios.

First, I generate three-level panel data (community, household, and plot) over two time

9The other robustness specifications also have negative household size coefficients, but the estimated

coefficients are not significant.

42

periods by the following process:

Ychit = exp(κ + ηc + ηh + ηi + 
chit)

(2.9)

where Ychit is analogous to labor demand in time t on plot i of household h in community

c. This process was chosen to mirror, in a simplistic way, the linear-in-logs specification of

this study’s empirical strategy.10

Data is simulated for 250 communities and 25,000 households, 100 households per com-

munity, using this data generating process. Community and household-level unobservables
are drawn independent and identically distributed (i.i.d.) as ηc ∼ N (0, 4) and ηh ∼ N (0, 4)

respectively. Each household in the same community is assigned a common ηc and each plot-

observation within the same household is assigned a common ηh. Plot level unobservables,

ηi, are drawn as:

where Xc,h,t=1 = (cid:101)Xc,h,t=1 + 1, (cid:101)Xc,h,t=1 ∼ P oisson(4), and Zchi ∼ N (0, 4).

ηi = aXc,h,t=1 + bZchi

(2.10)

Xc,h,t=1 is analogous to a household demographic variable in the first period. This struc-

ture is chosen to simulate correlation between observable household demographics and un-

observed time-invariant plot characteristics. Given this structure and the independence of

Xc,h,t=1 and Zchi, a can be derived as follows:

a =

Cov(Xc,h,t=1, ηi)

V ar(Xc,h,t=1)

=

Cov(Xc,h,t=1, ηi)

4

(2.11)

10This simulation focuses on the correlation between plot unobservables and a household demographic
variable. Thus, other characteristics are not included in the data generating process and community level
unobservables are simulated as time constant and thus absorbed by the household or plot fixed effects.

43

Similarly, b can be derived as:

b =

Cov(Zchi, ηi)

V ar(Zchi)

=

Cov(Zchi, ηi)

4

The population correlation between Xc,h,t=1 and ηi is then given by:

Corr(Xc,h,t=1, ηi) =

Cov(Xc,h,t=1, ηi)

σXση

=

Cov(Xc,h,t=1, ηi)

√
4

a2 + b2

(2.12)

(2.13)

Defining Cov(Xc,h,t=1, ηi) and Cov(Zchi, ηi) as one and four respectively sets a population

correlation between the first period household demographic variable and time invariant plot

unobservables of approximately one quarter (0.2425). This population correlation is larger

than that of plot unobservables and Xcht once temporal variation in Xcht is introduced.

The simulated dependent variable is finalized by defining the overall intercept, κ, at
idiosyncratic errors from 
chit ∼ N (0, 1). Ychit is then computed by

20 and drawing i.i.d.

combining all of its component parts as shown in equation 2.9.

In order to simulate exogenous temporal variation in Xcht, which is analogous to changes

in a household demographic variable between survey waves, each household is assigned an

i.i.d. draw from δh = (cid:101)δh + 1 where (cid:101)δh ∼ P oisson(2). For each household, a draw from

U nif orm[0, 1] determines how δh is allocated to Xc,h,t=2. One third of simulated households

are randomly assigned Xcht increases in period two as given by Xc,h,t=2 = Xc,h,t=1 + δh.

Similarly, one third of simulated households are randomly assigned Xcht decreases in period
two as given by Xc,h,t=2 = Xc,h,t=1−δh.11 The remaining one third of households are assigned

Xc,h,t=2 = Xc,h,t=1.

The number of plot observations, nht, varies by household and time period. In the first

period, each household’s number of farmed plots is determined by i.i.d. draws of Nht = (cid:101)Nht+1
where (cid:101)Nht ∼ P oisson(2).

11Xc,h,t=2 is set to one if the simulated decrease would cause Xc,h,t=2 to fall below one.

44

Several different scenarios, summarized in Table 2.5, simulate varying degrees of farmed

land changes between survey waves. The first scenario mimics the unlikely case of no house-

hold changing farmed plots between survey waves by maintaining the first period plot alloca-

tions. This is chosen to demonstrate a case where separability tests based on the household

fixed effects estimator are unbiased.

The other three scenarios demonstrate an active (and increasingly active) land market.

They do so by simulating more realistic cases where some of the households’ farmed plots vary

between survey waves (whether due to newly rented in, rented out, bought, or sold plots).

For these scenarios, each of a household’s plots is assigned an i.i.d. draw from U nif orm[0, 1].

Similarly, for each household, five potential new plots are simulated and assigned i.i.d. draws

from U nif orm[0, 1]. These draws are used to determine which plots are farmed by each

household in the second period.

Under Scenario A, a given household plot farmed in the first period is maintained in

the second period with probability 0.99. In addition, each household has a small chance of

farming one or more new plots. Each of the five potential new plots are incorporated into a

household’s farmed plots in the second period with probability 0.03. On average across the

1,000 replications, this results in 16% of households experiencing a change in farmed plots

in the second period.

Under Scenario B, a given household plot farmed in the first period is maintained in the

second period with probability 0.95. Similarly, the chance of farming each of five potential

new plots increases to 0.05. On average across the 1,000 replications, this results in 30% of

households experiencing a change in farmed plots in the second period.

Under Scenario C, the probability of a household maintaining a given first period plot

is kept unchanged at 0.95. The chance of farming each of the five new plots, however, is

increased slightly to 0.07. On average across the 1,000 replications, this results in 37% of

45

households experiencing a change in farmed plots in the second period.12

For each of these scenarios, I run the household fixed effects and plot fixed effects specifi-

cations and store the results of the reduced form separability tests. Having stored the results

from the first replication of the simulation, I then repeat this entire process 1,000 times. In

each replication, I take new draws from the respective distributions of ηc, ηh, Xc,h,t=1, Zchi,


chit, δh, Nht, and the Uniform[0,1] variables. I then use these new draws to compute ηi, Ychit,

and Xc,h,t=2. Using the given replication’s simulated dataset, I then estimate the household

fixed effects and plot fixed effects specifications under each of the four scenarios and record

the results of the reduced form separability tests.

VI.2 Simulation Results

The simulation results demonstrate the susceptibility of the separability test based on the

household fixed effects estimator to unobserved heterogeneity in land quality. The bias in

the household fixed effects estimator increases as the land market becomes more active. In

contrast, the plot fixed effects estimator is robust to a correlation of the household demo-

graphic variable with unobserved, time invariant land quality, regardless of the level of land

market activity.

Table 2.6 reports the number of Type I errors under different levels of land market

activity for each estimator over the 1,000 replications. As this is simulated data, the data

generating process is known and the separability hypothesis holds. That is, as Xcht is not a

causal determinant of Ychit, systematic rejections of the null of hypothesis of separability are

indicative of bias in the estimator. Across 1,000 replications, an unbiased estimator would

incorrectly reject the null of separability at the 5% and 10% levels approximately 50 and 100

times respectively.

When none of the simulated households change their farmed land between survey waves,

12These scenarios are conservative relative to the Rwanda sample where slightly more than half of all

households reported a change in farmed plots between the 2014 and 2017 survey waves.

46

the Type I error rates of both the household fixed effects and plot fixed effects estimators are

indicative of their unbiasedness. The plot fixed effects estimator incorrectly rejects the null

hypothesis of separability at the 5% and 10% levels 50 and 94 times respectively. Similarly,

the household fixed effects estimator incorrectly rejects the null hypothesis of separability at

the 5% and 10% levels 44 and 92 times respectively.

The performance of the household fixed effects estimator worsens as the percentage of

households with a land change increases. Under Scenario A, where 84% of simulated house-

holds have static farms and 16% gain or lose a second period plot, tests based on the house-

hold fixed effects estimator reject the null of separability 117 and 190 times at the 5% and

10% levels respectively. When 30% and 37% of simulated households gain or lose a second

period plot, the Type I errors at the 5% level increase to 133 and 245 respectively. These

Type I error rates represent more than 100% increases relative to an unbiased estimator.

The empirical distributions of the coefficient estimate on Xcht presented in Figure 2.1

further illustrate the Type I error rate of tests based on the household fixed effects estimator.

When all households have static farmland, the distribution of the coefficient estimates of

the demographic variable using household fixed effects is correctly centered on zero. In all

other scenarios, however, the distribution is shifted rightward. The magnitude of this shift

increases as the land market becomes more active and a larger percentage of households

change farmed land between survey waves.

In contrast, separability tests based on the plot fixed effects estimator are unaffected by

the percentage of households that gained or lost a plot in the second survey wave. When

16%, 30% or 37% of households change one or more plots between survey waves, the plot

fixed effects estimator incorrectly rejects separability 47, 49, and 49 times respectively. The

empirical distributions of the coefficient estimate on Xcht based on plot fixed effects illustrate

this consistency (Figure 2.2); the empirical distributions are tightly centered around zero

under each scenario.

47

VII. Conclusion

Whether agricultural households make their production decisions separately from their con-

sumption decisions is key to the design and evaluation of development programs and policies.

Non-separability affects households’ production responses to new opportunities and shocks;

its existence also provides an indication of market failures.

This paper confronts an important identification challenge in common tests of this sep-

arability assumption–the endogeneity of household demographic characteristics with unob-

served land quality. Leveraging intra-plot variability in labor demand, I find that the non-

separability result in Rwanda is robust to controlling for land quality and other unobserved

time invariant land characteristics. Furthermore, using simulated data where the separa-

bility hypothesis is known to hold, I find that tests based on intra-household variability in

labor demand are prone to false rejections of separability and that the likelihood of a false

rejection increases as the land market becomes more active.

The Rwanda results are limited by the short, two-wave plot panel data which reduces

observed intra-plot variability. This short timing, however, reduces the likelihood that plot

selection into the plot panel will lead to inconsistent estimation of the separability test

(Wooldridge, 2010, pg. 830). I also cannot control for unobserved land characteristics when

assessing the separability of agricultural households without a plot in the plot panel. The

inclusion or exclusion of households not in the plot panel subsample, however, does not

impact results based on household fixed effects.

The simulation results are also limited by the applicability of the underlying assumptions.

In particular, the consequence of ignoring unobserved land quality in separability tests is

dependent on the correlation with household demographic characteristics. The applicability

of the plot panel based test is also dependent on some stability in households’ operated plots.

If the land market is volatile and few plots are maintained by the same household overtime,

then capturing plot panel survey data may not be feasible.

48

Despite these limitations, the findings in this paper provide important implications for

the evaluation of households’ responses to agricultural policies and programs. The robustness

of the non-separability result in Rwanda begets the importance of agricultural development

programs and policies which reduce market inefficiencies and consider the role of households’

consumption preferences in their production decisions. Furthermore, the simulation results

suggest that if the land market is not active and the vast majority of households have

static farmland across all survey waves, then the appropriateness of studying households’

production responses to policies in isolation of consumption decisions may be accurately

assessed with household panel data; however, household production responses to policies are

more likely to be separable when markets are working well, increasing the value of separability

tests in contexts with relatively active land markets. When the land market is active over

the sampling period, this paper’s findings suggest that a reliance on household panel data

may misinform inference on the separability of agricultural households’ production decisions.

This could lead to a mis-characterization of households’ production responses to agricultural

policies. Future work should incorporate plot panel data in a variety of different contexts

to provide an updated view of agricultural household decision making which is robust to

unobserved heterogeneity in land quality.

49

APPENDICES

50

APPENDIX A:

Tables and Figures

51

Table 2.1: Household Summary Statistics

Plot Panel
Households Households

All

Household size

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

Share of male members 61 years or older

Farmed land area across all plots (ha)

Number of farmed plots

Number of household-wave observations

5.05
(1.78)
0.24
(0.18)
0.24
(0.18)
0.26
(0.12)
0.22
(0.13)
0.02
(0.09)
0.02
(0.06)
1.08

(12.92)

2.44
(1.40)
2,988

5.05
(1.78)
0.24
(0.18)
0.24
(0.18)
0.26
(0.12)
0.22
(0.14)
0.02
(0.09)
0.01
(0.06)
1.27

(13.08)

2.34
(1.40)
3,600

Note: Means with standard deviations in parentheses. Plot panel households refers to
the subsample of households with at least one plot successfully linked across survey
waves. Farmed land area and number of farmed plots correspond to the February to
June agricultural season.

52

Table 2.2: Plot Summary Statistics

Panel Subsample All Plots

Plot size (ha)

Total labor demand (labor-days)

Family labor demand (labor-days)

Hired labor demand (labor-days)

Number of plot-wave observations

0.22
(2.96)
24.81
(36.22)
21.30
(25.23)

3.50

(24.16)
4,580

0.44
(8.09)
24.21
(35.29)
20.76
(25.38)

3.45

(22.17)
7,303

Note: Means with standard deviations in parentheses. These statistics are
restricted to plots farmed by the 1,494 households with at least one plot in the
plot panel. Panel subsample refers to the plots successfully linked across
survey waves. All plots refers to any plot farmed by the 1,494 households in at
least one period. Labor demand consists of labor days used for land
preparation, planting, and field management after planting for the February to
June agricultural season. Labor from children is scaled by 0.5.

53

Table 2.3: Farm and Plot Labor Demand (Log of Person Days per Season) on

Household Characteristics

Household FE
(2)
(1)

Plot FE

(3)

(4)

0.242∗∗
(0.111)

ln(Household size)

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

F-test p-value for demogr. vars. joint signif.
Number of observations
Number of FE groups

0.030
2,988
1,494

0.356∗∗∗
(0.134)
0.222
(0.687)
0.221
(0.675)
0.909
(0.650)
0.683
(0.659)
0.940
(0.672)

0.059
2,988
1,494

-0.012
(0.112)

0.916
4,580
2,290

-0.021
(0.140)
1.842∗∗∗
(0.609)
2.057∗∗∗
(0.621)
2.298∗∗∗
(0.612)
2.221∗∗∗
(0.596)
2.096∗∗∗
(0.599)

0.001
4,580
2,290

Note: Coefficient estimates with village-level cluster-robust standard errors in parentheses.
Dependent variable in (1) and (2) is the natural log of the sum of pre-harvest person days of labor
used by a household across all operated plots during the February to June agricultural season of a
given year. Dependent variable in (3) and (4) is the natural log of pre-harvest person days of labor
used on a plot during the February to June agricultural season of a given year. The omitted
demographic share in (2) and (4) is male members 61 years or older. All models control for the age,
education, and gender of the household head and community-time fixed effects. (1) and (2) include
household fixed effects and the natural log of farmed land area in a given year. (3) and (4) include
plot fixed effects. * p < 0.10; ** p < 0.05; *** p < 0.01.

54

Table 2.4: Sum of Labor Demand Across Plot Panel Plots (Log

of Person Days per Season) on Household Characteristics

(1)

(2)

ln(Household size)

-0.052
(0.110)

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

F-test p-value for demogr. vars. joint signif.
Number of observations
Number of FE groups

0.635
2,988
1,494

-0.069
(0.138)
1.609∗∗∗
(0.607)
1.720∗∗∗
(0.604)
1.938∗∗∗
(0.599)
2.003∗∗∗
(0.581)
1.670∗∗∗
(0.562)

0.025
2,988
1,494

Note: Coefficient estimates with village-level cluster-robust standard errors
in parentheses. Dependent variable is the natural log of the sum of
pre-harvest person days of labor used by a household across all plot-panel
plots during the February to June agricultural season of a given year. Plots
observed in only one survey wave (and thus excluded from the plot panel)
are omitted. The omitted demographic share in (2) is male members 61 years
or older. All models control for the age, education, and gender of the
household head, household fixed effects, and community-time fixed effects.
* p < 0.10; ** p < 0.05; *** p < 0.01.

55

Table 2.5: Simulated Second Period Land Changes

Scenario

Static Farmland
Scenario A
Scenario B
Scenario C

Prob. Keeping
Each old plot Each of 5 new plots that Lost a Plot that Gained a Plot with Land Change

Avg. % of HHs

Prob. Gaining

Avg. % of HHs

Avg. % of HHs

1

0.99
0.95
0.95

0

0.03
0.05
0.07

0%
2%
10%
10%

0%
14%
22%
30%

0%
16%
30%
37%

Note: The probability of keeping each old plot is the probability that a given plot farmed in the first period is maintained in the
second period. The probability of gaining each of five new plots is the probability that a given new plot is incorporated into a
household’s farmed plots in the second period. The average percent of households that lost a plot corresponds to the percentage of
households that stopped farming a first period plot in the second period averaged across the 1,000 replications. The average
percent of households that gained a plot corresponds to the percentage of households that farmed a new plot in the second period
that they had not farmed in the first period averaged across the 1,000 replications. The average percent of households with a land
change is the percentage of simulated households that lost or gained at least one plot between survey waves averaged across the
1,000 replications.

56

Table 2.6: Number of Observed Type I Errors in 1,000 Replications

Percentage of Simulated Households with Land Change

0%

16%

30%

37%

Estimator
Plot Fixed Effects
Household Fixed Effects

Significance Level Significance Level Significance Level Significance Level
0.05
50
44

0.10
90
353

0.10
94
92

0.10
93
190

0.05
47
117

0.05
49
133

0.10
90
205

0.05
49
245

Note: Type I error is for incorrect rejection of the null hypothesis of separability. Percentage of simulated households with
land change is the percentage of simulated households that lost or gained at least one plot between survey waves averaged
across the 1,000 replications. An unbiased estimator would reject the null at the 5% and 10% levels approximately 50 and
100 times respectively across 1,000 replications.

57

Figure 2.1: Empirical Distribution via Household Fixed Effects

58

Figure 2.2: Empirical Distribution via Plot Fixed Effects

59

APPENDIX B:

Supplementary Tables

60

Table 2.7: Labor Demand (Log of Person Days per Season) on Household

Characteristics: Accounting for Potential Child Productivity Differences

Household FE
(1)
(2)

Plot FE

(3)

(4)

(0.123)

Share of male members 0 to 14 years old

Share of female members 0 to 14 years old

ln(Num. of adult equivalent household members) 0.362∗∗∗ 0.355∗∗∗ 0.055 −0.019
(0.133) (0.128) (0.139)
1.825∗∗∗
0.442
(0.582)
(0.665)
2.041∗∗∗
0.442
(0.592)
(0.656)
2.295∗∗∗
0.900
(0.613)
(0.652)
2.219∗∗∗
0.679
(0.596)
(0.660)
2.096∗∗∗
0.937
(0.599)
(0.671)

Share of female members 15 to 60 years old

Share of female members 61 years or older

Share of male members 15 to 60 years old

F-test p-value for demogr. vars. joint signif.
Number of observations
Number of FE groups

0.003
2,988
1,494

0.058
2,988
1,494

0.668
4,580
2,290

0.001
4,580
2,290

Note: Coefficient estimates with village-level cluster-robust standard errors in parentheses.
Dependent variable in (1) and (2) is the natural log of the sum of pre-harvest person days of labor
used by a household across all operated plots during the February to June agricultural season of a
given year. Dependent variable in (3) and (4) is the natural log of pre-harvest person days of labor
used on a plot during the February to June agricultural season of a given year. Number of adult
equivalent household members counts members under 15 as half an adult to account for potential
labor productivity differences. The omitted demographic share in (2) and (4) is male members 61
years or older. All models control for the age, education, and gender of the household head and
community-time fixed effects. (1) and (2) include household fixed effects and the natural log of
farmed land area in a given year. (3) and (4) include plot fixed effects.
* p < 0.10; ** p < 0.05; *** p < 0.01.

61

Table 2.8: Farm Labor Demand (Log of Person Days per Season) on Household

Characteristics: Controlling for Land Quality Proxy

No Land Quality

Controlling for

Control

Land Quality Proxy

ln(Household size)

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

(1)

(2)

(3)

0.242∗∗ 0.356∗∗∗ 0.240∗∗
(0.111)
(0.111)

(0.134)
0.222
(0.687)
0.221
(0.675)
0.909
(0.650)
0.683
(0.659)
0.940
(0.672)

F-test p-value for demogr. vars. joint signif. 0.030
2,988
Number of observations
Number of FE groups
1,494

0.059
2,988
1,494

0.031
2,988
1,494

(4)

0.358∗∗∗
(0.134)
0.211
(0.686)
0.213
(0.675)
0.906
(0.650)
0.664
(0.660)
0.991
(0.667)

0.059
2,988
1,494

Note: Coefficient estimates with village-level cluster-robust standard errors in parentheses.
Dependent variable is the natural log of the sum of pre-harvest person days of labor used by a
household across all operated plots during the February to June agricultural season of a given year.
(1) and (2) reproduce the results from columns (1) and (2) of Table 2.3. (3) and (4) control for
household’s reported high quality farmed land area. The omitted demographic share in (2) and (4)
is male members 61 years or older. All models control for age, education, and gender of the
household head, the natural log of farmed land area in a given year, household fixed effects, and
community-time fixed effects. * p < 0.10; ** p < 0.05; *** p < 0.01.

62

Table 2.9: Farm Labor Demand (Log of Person Days per Season) on Household

Characteristics: Incorporating All Households

Plot Panel Households All Households

(1)

0.242∗∗
(0.111)

ln(Household size)

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

F-test p-value for demogr. vars. joint signif. 0.030
2,988
Number of observations
Number of FE groups
1,494

(2)

0.356∗∗∗
(0.134)
0.222
(0.687)
0.221
(0.675)
0.909
(0.650)
0.683
(0.659)
0.940
(0.672)

0.059
2,988
1,494

(3)

(4)

0.190∗∗ 0.332∗∗∗
(0.092) (0.112)
0.063
(0.632)
0.102
(0.600)
0.875
(0.595)
0.687
(0.576)
0.820
(0.614)

0.040
3,600
1,800

0.013
3,600
1,800

Note: Coefficient estimates with village-level cluster-robust standard errors in parentheses. Columns
(1) and (2) are restricted to the subsample of households with at least one plot in the plot panel,
reproducing the results from columns (1) and (2) of Table 2.3. Columns (3) and (4) include all
households. Dependent variable is the natural log of the sum of pre-harvest person days of labor
used by a household across all operated plots during the February to June agricultural season of a
given year. The omitted demographic share in (2) and (4) is male members 61 years or older. All
models control for age, education, and gender of the household head, the natural log of farmed land
area in a given year, household fixed effects, and community-time fixed effects.
* p < 0.10; ** p < 0.05; *** p < 0.01.

63

Table 2.10: Farm and Plot Labor Demand (Log of Person Days per Season) on Household

Characteristics: Excluding Migrant Households

Household Fixed Effects

Plot Fixed Effects
Excl. Migrant & Excl. Migrant HHs

Excl. Migrant HHs

(1)

-0.492∗
(0.251)

ln(Household size)

Share of female members 0 to 14 years old

Share of male members 0 to 14 years old

Share of female members 15 to 60 years old

Share of male members 15 to 60 years old

Share of female members 61 years or older

F-test p-value for demogr. vars. joint signif. 0.051
2,208
Number of observations
Number of FE groups
1,104

Non-Plot Panel HHs

(2)

(3)

(4)

-0.349
(0.376)
0.312
(1.659)
0.667
(1.642)
0.484
(1.844)
0.873
(1.557)
1.515
(1.897)

0.255
2,208
1,104

-0.389
(0.275)

0.158
1,864
932

-0.051
(0.414)
-0.751
(1.466)
-0.348
(1.454)
-0.364
(1.709)
0.176
(1.378)
0.748
(1.739)

0.345
1,864
932

(5)

-0.814∗∗∗
(0.282)

0.004
2,962
1,481

(6)

-0.566
(0.433)
0.541
(1.501)
0.358
(1.483)
0.692
(1.758)
0.921
(1.465)
1.242
(1.897)

0.073
2,962
1,481

Note: Coefficient estimates with village-level cluster-robust standard errors in parentheses. All columns are restricted to the
subsample of households where no person left or joined the household except for children born between survey waves.
Columns (3) and (4) further restrict the household sample to exclude households without a plot in the plot panel. Dependent
variable in (1)-(4) is the natural log of the sum of pre-harvest person days of labor used by a household across all operated
plots during the February to June agricultural season of a given year. Dependent variable in (5) and (6) is the natural log of
pre-harvest person days of labor used on a plot during the February to June agricultural season of a given year. The omitted
demographic share in (2), (4), and (6) is male members 61 years or older. All models control for the age, education, and
gender of the household head and community-time fixed effects. (1)-(4) include household fixed effects and the natural log of
farmed land area in a given year. (5) and (6) include plot fixed effects. * p < 0.10; ** p < 0.05; *** p < 0.01.

64

REFERENCES

65

REFERENCES

Benjamin, D. (1992). Household Composition Labor Markets and Labor Demand, Testing

for Separation in Ag Household Models. Econometrica, 60 (2), 287–322.

Conley, T. G., & Udry, C. R. (2010). Learning about a New Technology: Pineapple in Ghana.

American Economic Review, 100 (1), 35–69.

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.

Dillon, B., Brummund, P., & Mwabu, G. (2019). Asymmetric non-separation and rural labor

markets. Journal of Development Economics, 139, 78–96.

Foster, A. D., & Rosenzweig, M. R. (1995). Learning by Doing and Learning from Oth-
ers: Human Capital and Technical Change in Agriculture. The Journal of Political
Economy, 103 (6), 1176–1209.

Kopper, S. A. (2018, July 30). Agricultural labor markets and fertilizer demand: Intensifica-

tion is not a single factor problem for non-separable households (Working Paper).

LaFave, D., & Thomas, D. (2016). Farms, Families, and Markets: New Evidence on Com-

pleteness of Markets in Agricultural Settings. Econometrica, 84 (5), 1917–1960.

Pitt, M., & Rosenzweig, M. (1986). Agricultural Prices, Food, Consumption and the Health
and Productivity of Indonesian Farmers. In I. Singh, L. Squire, & J. Strauss (Eds.),
Agricultural Household Models: Extensions, Applications, and Policy. Baltimore, The
Johns Hopkins University Press.

Sheahan, M., Black, R., & Jayne, T. (2013). Are Kenyan farmers under-utilizing fertilizer?
Implications for input intensification strategies and research. Food Policy, 41, 39–52.

Singh, I., Squire, L., & Strauss, J. (1986). Agricultural Household Models: Extensions, Ap-

plications, and Policy. Baltimore, The Johns Hopkins University Press.

Suri, T. (2011). Selection and Comparative Advantage in Technology Adoption. Economet-

rica, 79 (1), 159–209.

Udry, C. R. (1999). Efficiency and Market Structure: Testing for Profit Maximization in
African Agriculture. In G. Ranis & L. Raut (Eds.), Trade, Growth and Development:
Essays in Honor of T.N. Srinivasan. Amsterdam, Elsevier Science.

66

Wooldridge, J. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, MA, MIT Press.

67

Chapter 3

Farmer Personality and Community-Based Extension Effectiveness in Tanzania

68

I.

Introduction

Diffusion of new agricultural technologies is critical to agricultural transformation in de-

veloping countries. The development literature abounds with examples of new agricultural

technologies that have low adoption rates despite experimental studies demonstrating their

superior profitability relative to the prevailing technology. Low adoption of a new agricul-

tural technology may be driven by farmers’ rational assessments of its lack of suitability to

their local context or cultural practices. For example, the relative benefits of a new agricul-

tural technology may require complementary inputs that a farmer cannot access due to the

lack of well-functioning credit or input markets (Foster & Rosenzweig, 2010; Jack, 2013).

In other cases, however, a new agricultural technology may have low adoption despite being

well-suited to the local context. Lack of take-up in these latter cases can be characterized

as an asymmetric information problem whereby the technology ‘buyer’ (e.g. the farmer) is

not aware of or does not trust the credibility of available information signals on the relative

benefits of the new agricultural technology (Macho-Stadler & P´erez-Castrillo, 2001).

One means of providing a credible signal is extension programs which provide a crucial

link between pure research and the general public. Extension programs are particularly

important in the developing country context where agricultural innovations have the poten-

tial to lift millions out of poverty, but only if those innovations are known and adopted by

farmers. Evidence on the effectiveness of extension programs has been mixed, particularly

in Sub-Saharan Africa (Anderson & Feder, 2004; Krishnan & Patnam, 2014; Norton & Al-

wang, 2020). Lead-farmer extension, which relies on local farmers to act as community-based

extensionists in their villages, provides one potential means of increasing the trust in and

effectiveness of extension efforts (BenYishay & Mobarak, 2019). This form of community-

based extension has been shown to be effective in the study area of Tanzania (Nakano et al.,

2018). Even within a community-based extension program, however, there may be hetero-

geneity in a farmer’s likelihood of receiving and trusting the extension ‘signal.’

69

Farmer personality may be an important source of such heterogeneity in community-

based extension effectiveness.1 Extension, and particularly community-based extension, has

historically relied on the ability of extensionists to establish and nurture relationships with

farmers. The psychology literature provides ample evidence for the role of personality in such

interpersonal outcomes (Brandst¨atter et al., 2018; Ozer & Benet-Mart´ınez, 2006; Weidmann

et al., 2017; Wilson et al., 2016). There is also increasing evidence for the role of personality

in a variety of consumer, health, labor, and social outcomes (Bazzani et al., 2017; Lin et

al., 2019; Ozer & Benet-Mart´ınez, 2006; Soto, 2019). A recent meta-replication study of

78 published personality trait-outcome associations found a vast majority to be replicable

(Soto, 2019).

Despite this, few studies have explored the potential role of personality in extension

and technology adoption in the agricultural context (Ali et al., 2017). There is also a

dearth of personality studies in developing country contexts (Van der Linden et al., 2018).

Building on a randomized control trial (RCT), this study addresses these gaps by assessing

whether farmer personality influences the effectiveness of extension approaches for promoting

improved bean varieties in Tanzania. I motivate this analysis with a conceptual framework for

recipient farmers’ responses to extension activities under asymmetric uncertainty and assess

the potential heterogeneity empirically using a unique dataset of the Big Six personality traits

measured using the Midlife Development Inventory (MIDI) (Lachman & Weaver, 1997, 2005).

My analysis provides insights on the types of farmers most likely to benefit from community-

based extension initiatives as reflected in their adoption behavior and social interactions.

The remainder of this paper is organized as follows. First, I describe in more detail

the extension program and RCT which encompass this study’s contextual background and

experimental design. Second, I develop a conceptual model of adoption under asymmetric

information which informs my hypotheses. I describe the data, including the measures of

1Personality traits can be broadly defined as “. . . relatively enduring patterns of thoughts, feelings, and

behaviors” (Roberts, 2009, pg. 140)

70

personality, in the third section. The fourth and fifth sections detail the empirical model

and results respectively. I conclude in the final section with a summary of key findings and

implications.

II. Background and Experimental Design

My empirical analysis relies on a community-based extension program in the Southern High-

lands of Tanzania developed by Farm Input Promotions Africa Ltd. (FIPS-Africa). FIPS-

Africa is a non-governmental organization that aims to improve small farmers’ welfare by

increasing awareness of and access to improved agricultural inputs. As part of the extension

program, one lead-farmer, which FIPS-Africa refers to as a Village-Based Agricultural Ad-

visor (VBAA), is selected for each village in the program (Melkani and Mason 2018). These

VBAAs are selected by the village members themselves based on a variety of factors, includ-

ing farming experience, communication skills, and willingness to coordinate with FIPS-Africa

(Morgan, 2018). The VBAAs are volunteers, but receive training from FIPS-Africa on good

agronomic practices and small business development as well as support for village-based ex-

tension services. FIPS-Africa has traditionally focused on maize and provided funding for

all VBAAs to conduct a demonstration plot comparison of varieties as well as distribute

small trial packs of improved varieties to farmers in the village for personal trials (Melkani

& Mason, 2018).

FIPS-Africa, in conjunction with Agricultural Research Institute Uyole (ARI-Uyole),

the International Center for Tropical Agriculture (CIAT), and Michigan State University,

conducted an RCT in 2017 to assess the marginal value-added of these small trial pack

distributions. As FIPS-Africa had focused on improved maize varieties in its past activities,

the RCT was based around improved bean varieties.

In particular, the RCT focused on

Njano Uyole and Uyole 96, two high-yielding disease tolerant bean varieties developed by

ARI-Uyole and CIAT in the Mbeya Region (Ibid).

71

While the full RCT encompassed all 230 active VBAAs in Tanzania’s Southern High-

lands, this study focuses on a subsample of 32 VBAAs in the Mbeya and Mbozi Districts

for which a detailed farmer level survey was conducted. Each VBAA in the subsample

led a demonstration plot comparison in the 2017 Major Season (March 2017-July 2017) of

Njano Uyole, Uyole 96, Uyole 03, and a prevailing local bean variety.2 Each demonstration

plot was divided into 16 equal-sized sub-plots (four for each variety). This allowed for a

demonstration plot comparison of the four varieties when untreated, treated with a fungi-

cide/insecticide (Apron Star), treated with fertilizer (YaraMila CEREAL), and treated with

both the fungicide/insecticide and fertilizer. While each VBAA conducted a demonstration

plot comparison of bean varieties, half were randomly assigned to provide free trial packs of

Uyole 96 and Njano Uyole seed for farmers to conduct their own personal trials. The VBAAs

that conducted only the demonstration plot are not a “true control.” Rather, the RCT as-

sesses the marginal value of VBAA’s providing trial packs along with a demonstration plot

relative to a demonstration plot alone (Ibid).

To reduce the probability of pre-treatment differences between the demonstration plot

only (demo-only) and demonstration plot plus trial pack (demo-trial) subsamples, the VBAAs

were stratified by district and, within each district, paired according to a Mahalanobis greedy

pairwise matching index.3 One VBAA from each pair was randomly assigned to the demo-

trial group. In addition to the common demonstration plot, the VBAAs randomly selected to

the demo-trial group received 150 trial packs of improved bean seed to give out to the farm-

ers in their village at the start of the 2017 Major Season (March 2017-July 2017). VBAAs

followed FIPS-Africa’s usual practice of distributing trial packs, which is to give them to

participants at the demonstration plot planting, keeping to one trial pack per household.4

2While all 32 villages in this subsample used Uyole 03 as the third improved bean variety in the demon-
stration plot, different improved varieties were used in its place in the other villages in the larger RCT. This
analysis focuses on Njano Uyole and Uyole 96 as these were the two common improved varieties with the
most widespread applicability (Melkani & Mason, 2018).

3The characteristics used are described in Melkani and Mason (2018).
4If any extra trial packs remained following the demonstration plot planting, distribution was up to the

72

Each trial pack contained 400 grams (g) of bean seed equally split between the local bean

variety and one of the three improved bean varieties used on the demonstration plot. Of the

200g of the bean seed for a given variety contained in each pack, 100g was pre-treated with

the insecticide/fungicide while 100g was left untreated. The VBAAs instructed each of the

farmers that received the trial pack on how to setup a demo plot on their own farm split into

four sub-plots. Thus, the farmers that received trial packs had an opportunity to mimic a

portion of the larger village-level demonstration plot on their own farms (Ibid).

The stratified, pairwise matching treatment randomization was designed to ensure bal-

ance on VBAA characteristics. Melkani and Mason (2018) conducted a baseline survey of all

230 VBAAs in early 2017 prior to the start of the RCT. Table 3.1 provides the balance test

results across key characteristics for the subsample of VBAAs in this study.5 The results

indicate that VBAAs in the demo-only and demo-trial groups are balanced across a variety

of demographic characteristics as well as past FIPS-Africa activities.

This previously implemented community-based extension RCT generated experimental

variation in the potential “signals” recipient farmers received about the new, improved bean

varieties from their VBAAs. I exploit this variability in this paper. In the next section I

develop a conceptual model to illustrate the likely value of the added trial pack information

and how it may be heterogeneous according to the personality traits of recipient farmers.

III. Conceptual Framework

In this section, I develop a conceptual model to explore the role of community-based extension

activities and farmer personality traits in addressing the information asymmetries inherent

in the introduction of a new seed variety. I show how community-based extension activities

that prioritize information signals from varied contexts can better enable farmers to update

their beliefs about a new variety. Furthermore, I show that farmer personality traits can

VBAA’s discretion, keeping to the one per household requirement.

5The full set of balance tests across all 230 VBAAs is available in Melkani and Mason (2018).

73

influence the likelihood of receiving these information signals which has implications for the

types of farmers likely to benefit from community-based extension activities.

Prior to its release, a new seed variety k undergoes rigorous performance evaluation under

researcher or farmer-managed trials conducted at experiment stations, central experimental

plots in communities, or to a limited extent on farmers’ fields.

It is released as a new

and improved variety if, on average across these trials, its performance was deemed better

than existing popular varieties, at least in one trait (e.g.

in profitability, yield, nutrition

quality, processing quality, resistance to drought, pests and disease, etc.). In the conceptual

model of varietal development, I assume that despite being considered an improved variety,

under a farmer’s growing conditions and given his or her varietal trait preferences, this

newly developed variety k can be one of two types relative to the prevailing variety, a high

performance type (θH) or a low performance type (θL).6 When a new variety is promoted,

farmers do not know whether a variety is θH or θL on traits they consider important and

must learn about it from the extensionist, agricultural input dealers, other farmers, and their

own experience. The performance of the prevailing variety (θ0) is given by π0 and is assumed

constant.7 If adopted, a new variety can either have a good performance outcome (πG) or

bad performance outcome (πB), where πG > π0 > πB. While a newly developed variety

has a nonzero chance of either outcome regardless of type, the probability of a good (bad)

outcome is higher (lower) for newly developed high type varieties than newly developed low

type varieties:

ρGH = P (πG|θ = θH) > P (πG|θ = θL) = ρGL
ρBk = (1 − ρGk) ∀ k = H, L

(3.1)

This framework reflects the idea that farmers are very familiar with the prevailing local

6Performance includes, but is not limited to, profitability. Farmers may also value non-monetary varietal

characteristics.

7This is not as strong an assumption as it first appears. π0 can be thought of as the long-run certainty
equivalent performance of the prevailing technology which serves as the benchmark for comparing the relative
benefits of new technologies.

74

variety’s long-term performance potential while newly developed varieties must be evaluated

on the basis of shorter term, noisy outcomes. Furthermore, it reflects the idea that some

new varieties are better than others in terms of their performance potential relative to the

prevailing technology.

I assume farmers are risk averse expected utility maximizers and that the relative prob-

ability distributions of good and bad outcomes imply that if a farmer knew that the new

variety was θH with certainty, then he or she would adopt the new variety. In contrast, I

assume that if a farmer knew that a new variety was θL with certainty, then he or she would

not adopt the new variety. These conditions imply the following:

EU (θH) > EU (π0) > EU (θL)

(3.2)

ρGHU (πG) + ρBHU (πB) > U (π0) > ρGLU (πG) + ρBLU (πB)

This setup is developed to illustrate a simple case where, given full information, a farmer

would rationally choose to adopt a new variety that is type θH, while rationally choosing

not to adopt a new variety that is type θL despite both having the potential for higher

performance under “good” conditions. As a farmer does not know the type of a newly

developed variety, he or she must develop a prior of the probability of a new variety’s type.

Without any additional information on the type of a new variety, this prior is given by γH
and γL for θH and θL, respectively, where γL = 1 − γH. Given no means of updating this

prior about the type of a new variety, I assume that a vast majority of farmers will not adopt

a newly developed seed variety. That is:

γHEU (θH) + γLEU (θL) ≤ U (π0)

(3.3)

If this were not the case, then most farmers would adopt a newly developed variety as soon

as it became available without any additional information. Thus, this condition provides

75

the opportunity to explore the role of extension efforts to facilitate technology diffusion by

helping farmers update their beliefs about the relative benefits of a new technology.

As a farmer does not know the performance characteristics of a new variety and extension-

ists may promote varieties of either performance type, this context describes an asymmetric

information problem.8 A farmer must make probabilistic inferences about the type of a new

variety based on the information signals he or she receives from extension activities. The

quantity and quality of the information available will depend on the promotional activities

of the extension program. For example, if an extensionist promotes a new variety by estab-

lishing a demonstration plot that compares the new variety side-by-side with the prevailing

local variety, then this activity provides one outcome for a farmer to assess. If, however, an

extensionist distributes free trial packs within the village for farmers to test a new variety on

their own farms, then this creates many more potential outcomes for a farmer to evaluate.

I denote N as the number of potential outcomes that extension activities create in order for

farmers to assess the underlying type of a new variety.

Suppose N = 1 and a farmer must base his or her belief about the underlying type of

the variety based on a single outcome (e.g. the extensionist provides a demonstration plot,

but does not distribute trial packs). If a farmer observes this outcome (either good or bad),

I assume he or she will update his or her beliefs about the variety’s type via Bayes’ Rule:

P (θH|π1

i ) =

ρiHγH

ρiHγH + ρiLγL

for i ∈ {G, B}

(3.4)

where π1

i is the observed outcome, ρiH and ρiL are the conditional probabilities of observing

that outcome i given that the new variety is of the high type or low type respectively, and γH

and γL are a farmer’s prior beliefs about the probability of a high or low type respectively. As
defined in equation 3.1, ρGH > ρGL, which implies that P (θH|πG) > P (θL|πG). That is, if a
8Extensionists are informed technology ‘sellers’ given the task of promoting varieties (of either type) by

outside institutions (e.g. non-governmental organizations or public research institutes).

76

good outcome is observed, then a farmer will adjust his or her belief about the underlying type

of the new variety towards θH (and vice versa if a poor outcome is observed). Furthermore,

as ρiLγL > 0, a farmer will not know the type of the new variety with certainty even after

observing the outcome. A good outcome might be a “lucky” low type result while a poor

outcome could be an “unlucky” high type result.

This posterior updating, however, assumes that the outcome π1

i will be costlessly ob-

served and incorporated by each farmer. More realistically, whether a farmer observes and

incorporates π1
i

is likely to depend on the quantity and quality of a farmer’s interactions

with the extensionist and other farmers. If a farmer does not hear of π1

i or is distrustful of

the information source, then he or she will not use this outcome to update his or her prior

beliefs (γH and γL).

Let δ(π1

i |X) be the probability that a farmer learns of and incorporates π1

i into his or her

beliefs about the new variety where X are individual characteristics (e.g. farmer personality)

that influence the quantity and quality of a farmer’s interactions. The affect of outcome π1
i

on a farmer’s belief about the type of the underlying variety is then:

[Equation 3.4] w.p. δ(π1

i |X)
w.p. 1 − δ(π1

γH

P (θH|π1

i , X) =

for i ∈ {G, B}

(3.5)

i |X)

That is, a farmer learns of and pays attention to the information inherent in π1

i with proba-
i |X) and disregards it (either due to lack of awareness or distrust) with probability
i |X).

bility δ(π1
1 − δ(π1

If N > 1 and the farmer is aware of and incorporates a second outcome, π2

i , into his or her

decision making, then he or she will further update his or her beliefs through an extension

77

of Bayes’ rule as follows:

P (θH|π1

i , π2

j ) =

P (π2

j|θH, π1

i )P (θH|π1

P (π2

j|θH, π1

i )P (θH|π1
i , X) + P (π2

i , X)
j|θL, π1

i )P (θL|π1

i , X)

for i, j ∈ {G, B}

(3.6)

where P (θH|π1

i , X) is as defined in equation 3.5. This delineation demonstrates that the

additional value of the second outcome is inversely related to the absolute value of the

covariance between the two outcomes. To see this, note that if the two outcomes are perfectly

correlated (i.e. π2

j = π1

i with certainty) then P (π2

i ) = 1 and equation 3.6 collapses

i |θH, π1

first.9

to equation 3.5. That is, the second outcome provides no new information over that of the
j|θH), so knowing the first outcome (π1
i )
i ). Then, the value of

provides no information about the likelihood of the second outcome (π1

In contrast, suppose P (π2

j|θH, π1

i ) = P (π2

the second outcome is “as good as” that of the first outcome in terms of updating a farmer’s

probabilistic beliefs. Thus, while a second outcome helps a farmer update his or her beliefs

about the underlying type of a new variety, the marginal contribution will be larger if the

second outcome comes from a more dissimilar context than that of the first outcome.10

Therefore, receiving information signals from more diverse sources will, on average, bet-

ter enable farmers to correctly pin down the underlying type of the new variety.11 These

results highlight the potential role of the trial pack distribution and farmer personality traits;

trial pack distribution creates a diverse set of information sources to learn from and farmer

personality traits may affect how farmers react to and process available information.

The updating in equation 3.6, however, is contingent on the farmer’s receipt of the

information signal. As in π1

i , the existence of π2

i alone does not guarantee that a farmer will

9In other words, any potential value of a perfectly correlated second outcome is encapsulated in equation
3.5. After observing the first outcome, observing a perfectly correlated second outcome does not help the
farmer update his or her beliefs beyond what he or she already knew.

10Marginal contribution refers to the magnitude of the shift of a farmer’s belief about the type of a variety

in equation 3.6 relative to equation 3.5.

11Although observing outcomes from similar sources reduces the likelihood of a farmer receiving conflicting

information, this also increases the likelihood of honing in on an incorrect assessment.

78

benefit from the additional information it provides. The affect of outcome π2

i on a farmer’s

belief about the type of the underlying variety depends on the farmer’s awareness of the

information source. Thus, the more realistic effect of an additional outcome is:

[Equation 3.6] w.p. δ(π2

i |X)
[Equation 3.5] w.p. 1 − δ(π2

i |X)

P (θH|π1

i , π2

i , X) =

for i ∈ {G, B}

(3.7)

where δ(π2

i |X) is the probability of the farmer observing and incorporating π2

i into his or her

decision making given X, the characteristics (e.g. farmer personality traits) that affect the

quantity and quality of a farmer’s interactions.

These results illustrate the potential role of extension activities and farmer personality

traits that encourage exposure and learning from varied contexts. As extension activities

increase the number of information signals, N , the extensionist increases the likelihood of

a farmer updating his beliefs about the new variety as well as increases the accuracy of a

farmer’s assessment; however, extension activities which encourage experimentation in varied

contexts, like the distribution of trial packs, may be better suited to relieving the new variety

adoption asymmetric information problem than additional trials on a single demonstration

plot. Furthermore, farmer personality characteristics, which are likely to affect the quality

and quantity of a farmer’s interactions, may play a vital role in addressing this asymmetry.

The effect of this additional extension effort is likely to be heterogeneous according to a

farmer’s awareness of and reaction to new information. Therefore, I hypothesize that the

impact of community based extension will differ by farmer personality characteristics. In

the next section, I describe the data used to test this hypothesis.

79

IV. Data

I utilize data from 740 bean farmers and 32 VBAAs (one per village) from 32 villages which

participated in the previously implemented RCT on VBAA extension activities (Melkani &

Mason, 2018). These 32 villages represent a subsample of those in the Mbozi and Mbeya

Districts that participated in the larger RCT. In each of the 32 villages, a random sample of

approximately 25 bean-growing households were selected for two rounds of a detailed survey

conducted immediately following the 2017 and 2018 Major Seasons. As the population of

interest is bean farmers in each village, non-bean growing households were excluded from

the sampling frame.

The 2017 farmer survey collected a range of detailed household and respondent level in-

formation including variety and plot specific bean production and varietal use in the minor

and major bean growing seasons in 2016 (pre-intervention), minor bean growing season in

2017 (pre-intervention), and in major season 2017 (during intervention) as well as household

and respondent sociodemographic characteristics and household GPS location. Table 3.2

provides covariate balance tests based on the farmer sample.12 Across most farmer charac-

teristics, I fail to reject the null hypothesis of equality between the demo-only and demo-trial

VBAA villages. In particular, farmer adoption of Njano Uyole and Uyole 96 before the start

of the intervention was low and not statistically different between demo-only and demo-trial

villages. Households in demo-only villages, however, are slightly larger on average than

those in demo-trial villages (mean household size was 5.67 and 5.19 in the demo-only and

demo-trial villages respectively). Additionally, demo-only households were more likely to

be related to the village chairman than those in demo-trial villages.

I control for these

unbalanced covariates in my analysis.

To assess the impact of the RCT intervention on farmer adoption of promoted bean

12The detailed 2017 farmer survey occurred after the incidence of the treatment in the 2017 Major Season.
Thus, farmer characteristics which may have been influenced by the treatment (e.g. knowledge of improved
bean varieties and VBAA-activities) are excluded from the balancing tests.

80

varieties, a follow up survey of the same households was conducted immediately following the

2018 Major Season. Analysis of this second round of data from post-intervention can provide

evidence of any short-term adoption effects and which subgroups of farmers were most likely

to benefit (or not) from the trial packs treatment after the program’s completion. Of the

791 households surveyed in the 2017 round, 740 households were successfully resurveyed

in the 2018 round (an attrition rate of 6%). To assess whether this household attrition

was correlated with village treatment status, I estimated logit regressions of the household

attrition indicator variable on village treatment status controlling for the VBAA-pair used

in the random treatment assignment. As shown in Table 3.3, I do not find evidence that

household attrition was related to village treatment assignment even after controlling for

2017 household size and relationship to the village chairmen.

Table 3.4 provides a comparison of key characteristics across three agricultural years

by demo-only and demo-trial villages. In the year before the intervention, only 7.8% and

6.6% of sampled households had adopted Uyole 96 or Njano Uyole in demo-only and demo-

trial villages respectively.13 When the RCT was implemented in the 2017 agricultural year,

adoption in demo-only villages remained relatively flat at 8.5%. As expected, the reported

adoption in demo-trial villages was higher (13%) in 2017, reflecting an increased planting of

these varieties due to the distribution of trial packs of improved bean seed. By the 2018 agri-

cultural year (i.e. the post-intervention year), the bean varieties had diffused more widely

in both demo-only and demo-trial villages with 21% of sampled households growing at least

one of the two varieties. This increase in the use of these varieties is suggestive of house-

holds receiving information signals in 2017 that induced take-up. While the adoption rate

was more than 50 percent higher among the subsample of households residing in demo-trial

villages relative to the demo-only subsample at the time of the intervention, the adoption

rates between the two subsamples were similar in 2018 (20-21%), but represent an approxi-

13I define adoption as the respondent reporting that his or her household grew either variety in the given

agricultural year.

81

mately 2.5 and 3 times increases from pre-intervention to post-intervention for the demo-only

subsample and demo-trial subsamples respectively. In my subsequent empirical analysis, I

explore the farmer characteristics which may have led to heterogeneity in the value of this

additional trial pack signal.

In addition to the respondent and household characteristics collected in the 2017 survey

round, farmer social interactions were also measured in the 2018 survey round (although

not specific to that year). Each farmer was asked: “Have you ever met [Name]?” where

[Name] was replaced by the name of the VBAA working in his or her village. If the farmer

replied yes, he or she was asked several related follow-up questions including: (1) Was your

VBAA named [Name]? (2) Have you ever gone to [Name] for advice about farming? and

(3) Have you ever discussed bean farming with [Name]? Along with these VBAA-specific

questions, each farmer was also asked whether he or she had ever met, ever sought farming

advice from, and ever discussed bean farming with each of the approximately 24 other bean

farming households that were sampled in that farmer’s village. Summary statistics for these

social interaction outcomes are presented in Table 3.4.

In addition to social interaction outcomes, personality traits were also measured in the

2018 survey round via the Lachman and Weaver (1997) Midlife Development Inventory

(MIDI). This scale is based on the widely studied Five Factor Model of personality which was

expanded to include a sixth trait (e.g. Bazzani et al., 2017; Grebitus et al., 2013; Lin et al.,

2019; Van der Linden et al., 2018). As described in Table 3.5, MIDI captures personality by

grouping 31 adjectives into six major traits of agency, agreeableness, openness to experience,

neuroticism, extraversion, and conscientiousness. Relative to other personality inventories

based on the Five Factor Model, MIDI has the advantage of being relatively easy to elicit

as well as measuring an additional sixth trait, agency. Respondents rated how well each

adjective described them on a scale of: one (not at all), two (a little), three (some), and four

(a lot). Respondents’ personality traits are measured as the average scores of the associated

82

adjectives.

The MIDI scale was translated into Swahili through multiple rounds of translation and

backward translation. First, two experts were asked to translate each word independently.

One was a Tanzanian professor at a U.S. university who teaches Swahili to English speaking

students, and the other was a collaborator based in Tanzania. Another professor from a

U.S. university who is based in Tanzania (and is Tanzanian) was then asked to backward

translate the Swahili translated words received from the local collaborator.

In the last

round, the survey enumerators and two supervisors, who are all fluent in both Swahili and

English, were presented with the two sets of translation and one set of backward translation

and as a group went through a rigorous consultation process to come up with the final

Swahili translation of each word. This final round with the enumerators and supervisors

also ensured that a common understanding was reached among all on each personality trait

being measured.

This rigorous approach to translating and localizing the MIDI scale is particularly impor-

tant given the rural, developing country sample. Capturing personality traits in developing

country contexts accurately is challenging. Laajaj et al. (2019) compare the validity of

personality trait measurements across a variety of contexts, finding that personality trait

measurements from developing country surveys tended to be less reliable. Although the

data collection for this study preceded the publication of Laajaj et al. (2019), the extensive

translation and enumerator training for the personality module is a main approach recom-

mended by Laajaj et al. (2019) for improving the measurement of personality in developing

country contexts.

Furthermore, the Cronbach’s alpha values of the personality traits in this study’s sam-

ple, the main measure of internal consistency and reliability for personality scales, compare

favorably to the developing country in-person surveys analyzed in Laajaj et al. (2019). Cron-

bach’s alpha increases (i.e. improves) as the within-group correlation between grouped items

83

increases and as the number of items increases (Laajaj et al., 2019). As shown in Table 3.6,

the majority of the Cronbach’s alpha values for the personality traits in this study’s sample

are 0.7 or greater with an average of 0.69. The six developing country surveys analyzed in

Laajaj et al. (2019) with 44 item personality scales (i.e. more than the 31 item MIDI scale

used in this study) had an average Cronbach’s alpha value of 0.62 (ibid).

As discussed previously, each village self-selected their VBAA based on a variety of

farming characteristics and other skills. As village-selected lead-farmers, I would expect

VBAAs to differ from the average farmer on skill-related variables such as farming experience.

Whether these differences include personality, however, is uncertain a priori. Farmers might

prefer to have a VBAA that is of similar personality temperament or one that stands out

from the average. As shown in Table 3.7, in this study’s sample I find the latter. That

is, VBAAs tend to score higher than the average farmer across a variety of personality

dimensions. Of the six personality traits, the three traits in which VBAAs differ the most

from the farmer sample are in openness to experience, extraversion, and agency. Individuals

high in openness to experience are more willing to experiment with new ideas, making them

good candidates for a lead-farmer position. Similarly, being high in extraversion, which is a

measure of sociability and tendency to be outgoing, may make a lead-farmer position more

attractive. High agency, which measures self-confidence, dominance, and outspokenness,

may also help a farmer win the position of VBAA and succeed in marketing new ideas

once doing so. Along with these major traits, VBAAs also stood out from the farmer

sample in terms of agreeableness and conscientiousness. Agreeableness measures kindness

and likeability, characteristics which may make a candidate more likely to be chosen for a

lead farmer position. Conscientiousness, which is a measure of organization, responsibility,

and work ethic, is a trait which may be particularly important for a VBAA’s effectiveness

in his or her role. All of these differences are indicative of lead-farmer personality being

markedly different than the average farmer. The trait where VBAAs are most similar to the

84

farmer sample is neuroticism, which is a negative personality trait associated with anxiety

and worry (Grebitus et al., 2013).

In my empirical analysis, I control for these VBAA

personality traits.

Along with controlling for VBAA personality in my analysis, I also control for a more

common dimension of similarity, physical distance. Farmers that interact more frequently

with their VBAA are more likely to be exposed to extension information and this frequency

of interaction is likely correlated with physical distance.

I measure physical distance by

capturing the GPS location of each VBAA’s and sampled farmer’s homestead as well as that

of the VBAA-led demonstration plot. A farmer’s distances to the VBAA’s homestead and

demonstration plot are measured via the GPS-based linear distances in kilometers. Table

3.8 reports summary statistics for these physical distances.

V. Empirical Strategy

My first outcome of interest is farmer adoption of the improved bean varieties post-intervention

(i.e.

in the 2018 agricultural year). This adoption analysis provides an indication of the

marginal value (in terms of farmer adoption probability) of the additional information sig-

nals from the distribution of trial packs of bean seed relative to the information signals from

a demonstration plot alone. It also enables an assessment of whether farmers with certain

personality traits were more likely to benefit (in terms of adoption) from these additional

information signals.

Along with adoption, I also investigate the relationships between the trial pack distribu-

tion and farmer personality traits on farmers’ social interactions with their VBAA and other

farmers. Analysis of these intermediate outcomes serves two main purposes. First, it eval-

uates the increase in the bilateral exchange of information between farmers and VBAAs–a

key mechanism through which community-based extension is intended to encourage farmer

take-up of the improved bean varieties. Second, it assesses whether farmer personality char-

85

acteristics predict differences in these key social mechanisms for community-based extension

effectiveness.14

I analyze three farmer-VBAA interaction outcomes: whether the farmer (1) has ever met

his or her VBAA and identified him/her as the VBAA, (2) has ever sought farming advice

from his/her VBAA, and (3) has ever discussed bean farming with his/her VBAA. Similarly,

I analyze three other peer interaction outcome variables: the proportion of a random sample
of ∼ 24 bean farming households in the farmer’s village that he/she (1) has ever met, (2)

has ever sought farming advice from, and (3) has ever discussed bean farming with. The

analysis of these six outcomes provides evidence on the extent to which farmer personality

plays a role in the quantity and quality of VBAA and peer farmer interactions.

For each outcome, I specify the following equation:

E(Yijk|T rialjk, Pijk, Xijk) = G(α + βT rialjk + Pijkδ + Xijkζ)

(3.8)

where Yijk is the binary or fractional outcome variable for farmer i in village j of VBAA-pair
k, G(·) is a logistic function, T rialjk is an indicator variable equal to one if village j of

VBAA-pair k was randomly assigned to the demo-trial group and zero otherwise, Pijk is

a vector of farmer personality characteristics, Xijk is a vector of controls including farmer

and VBAA sociodemographics, VBAA personality characteristics, and VBAA-pair indicator

variables. For the binary response and fractional response outcome variables, I report average

marginal effects from the logit regression and fractional logit regression respectively (Papke

& Wooldridge, 1996; Wooldridge, 2010). This specification evaluates the effect of the trial

pack treatment as well as whether farmer personality types are associated with different

adoption and social outcomes.

14With only 32 villages and VBAAs in the sample (one VBAA per village), I am unable to assess the
potential effect of VBAA personality characteristics on community adoption of the improved varieties. This
remains a promising area for future work with an expanded sample.

86

I also estimate the following supplementary linear model:

E(Yijk|T rialjk, Pijk, Xijk) = α + βT rialjk + T rialjk ∗ Pijkγ + Pijkδ + Xijkζ

(3.9)

where farmer personality characteristics, Pijk, are interacted with the trial indicator variable.

This specification enables an assessment of potential heterogeneity in the effect of the trial

pack treatment by farmer personality traits.

In addition to these main specifications, I also specify adoption models for equations 3.8

and 3.9 which control for the pre-intervention outcome (i.e. adoption in the 2016 agricultural

year). The results for these specifications, based on the ANCOVA framework outlined in

McKenzie (2012), are nearly identical to that of the main specifications (see Appendix Tables

3.15 and 3.16). As in the balancing tests, these robustness specifications suggest that residing

in a village that received trial packs of bean seed is exogenous and that my empirical strategy

controls for unobserved, pre-intervention differences between farmers in demo-trial and demo-

only villages.

VI. Results

A key outcome in terms of the diffusion of these improved bean varieties is whether the trial

pack treatment or personality traits are associated with changes in the likelihood of farmer

adoption of the new varieties. The results based on farmer adoption of Uyole 96 or Njano

Uyole in the 2018 agricultural year are presented in Table 3.9. As discussed previously, the

overall adoption rate among the sample of households increased dramatically from 10.7%

in 2017 to 21% in 2018. The adoption gap between demo-only and demo-trial villages,

however, declined. Given this, it is not surprising that I find limited evidence that residing

in a village that received trial packs increased farmer adoption of these improved varieties

on average. Across most specifications, I cannot reject the null that, for the average farmer

87

in the sample, residing in a village with additional information signals from the trial pack

distribution did not increase the likelihood of adopting the improved varieties relative to

residing in a village where only the VBAA-led demonstration plot took place.

I also do

not find evidence of differences in the likelihood of farmer adoption by personality type.

That is, farmer personality traits alone do not predict improved variety adoption in the 2018

agricultural year.

There may, however, be important heterogeneity that this initial specification cannot

characterize. As shown in the conceptual framework, personality traits may influence the

likelihood that a farmer seeks out additional information signals as well as how he or she

reacts to the information received; therefore, the effect of the trial pack distribution may be

heterogenous by farmer personality types. Table 3.10 presents the results for equation 3.9

which allows the effect of residing in a trial pack village on farmer adoption of Njano Uyole

or Uyole 96 to vary by farmer personality traits. While the impact of trial packs on the

diffusion of these improved bean varieties was attenuated for the average farmer, extraverted

farmers residing in trial pack villages were more likely to adopt these improved bean varieties

than their peers. This suggests that farmer personality may play a vital role in the likelihood

of benefiting from community-based extension efforts which aim to increase the exchange of

information between farmers.

I assess this potential social mechanism of the community-based extension program via

outcome variables that measure farmers’ social interactions with their VBAA and other

farmers. Table 3.11 columns 1-3 present the results for equation 3.8 based on whether a

farmer has ever met the VBAA in his or her village and correctly identified him or her as

the VBAA. The findings show that farmers residing in trial pack villages were more likely

to be able to identify their VBAA relative to farmers in demonstration plot only villages.

This suggests that the trial pack distribution may have boosted farmer participation in and

knowledge of extension activities. Furthermore, farmers scoring higher in agreeableness,

88

which is a measure of friendliness, and openness to experience were more likely to correctly

identify their VBAA. Farmers scoring high in neuroticism, which is a measure of anxiety

and worry, however, were less likely to be able to identify their VBAA. These results are

generally robust to additional demographic controls and suggest that farmer personality

traits are associated with differences in farmers’ basic awareness of the existence of a VBAA

in their village.

The results based on differences in a farmer’s likelihood of ever seeking out farming

advice from his or her VBAA, a stronger measure of information exchange, were more mixed

(Table 3.11 columns 4-6). There is some evidence that the trial pack distribution may have

increased the likelihood of farmers seeking advice from their VBAA, but this result is not

robust. Similarly, there is some evidence that farmers scoring high in agency, which is a

measure of self-confidence, were less likely to seek out farming advice. This is consistent

with these farmers being more sure of their own practices and less willing to seek out others’

opinions. Scoring high in openness to experience, however, may offset this effect.

A more middle ground means of information exchange relative to having met a VBAA

or sought farming advice from him or her, is whether farmers indicate that they have ever

discussed bean farming with their VBAA; results based on this outcome, which are presented

in Table 3.11 columns 7-9, show more robust evidence for trial pack and personality differ-

ences. In particular, residing in a village that received trial packs increases the likelihood

that a farmer has discussed bean farming with his or her VBAA by approximately seven

percentage points. Higher agency (i.e. more self-confident) farmers are also much less likely

to have ever discussed bean farming with their VBAAs; the predicted reduction from scoring

a single point higher in agency would more than offset the gain in likelihood from residing

in a village that received trial packs.

Along with a farmer’s interactions with his or her VBAA, other important social outcomes

are his or her interactions with other peer farmers in the village. The results based on

89

the proportion of a sample of ∼ 24 bean farming households in a farmer’s village that

he or she has ever met (Table 3.12 columns 1-3), ever sought farming advice from (Table

3.12 columns 4-6), and ever discussed bean farming with (Table 3.12 columns 7-9), are

similar to those based on farmer-VBAA interaction. The findings suggest that the trial pack

distribution increased the proportion of bean farming households that farmers have met and

have discussed bean farming with. Similar to that of the VBAA outcomes, however, there is

no evidence that it increased the proportion of bean farming households that the farmer has

sought advice from. This suggests that the trial pack distribution increased the bi-lateral

exchange of information between farmers in the village, but not to an extent that farmers

were more willing to seek out farming advice.

The role of personality traits in farmer interactions with their peers are also very similar

to that of the VBAA social outcomes.

In particular, higher agency and neuroticism are

associated with declines in peer social interaction while openness to experience is associated

with an increase in peer social interaction. There is also some evidence that increases in

agreeableness (friendliness) reduced the willingness of farmers to seek advice from their

peers, but this result is not robust to the inclusion of demographic controls.

Finally, a heterogenous effects analysis on the proportion of bean farming households

that a farmer has sought advice from (Table 3.13) and discussed bean farming with (Table

3.14) provides evidence for why extraverted farmers may have benefited more from the trial

pack distribution. The results show that more extroverted farmers residing in trial pack

villages were more likely to have sought out relevant social interactions with peer farmers.

This is consistent with extraversion playing an important role in the quantity and quality

of trial pack information signals a farmer receives and leading to an increase in take-up of

the improved bean varieties. This exploratory social outcome analysis suggests that farmer

personality may raise farmer awareness of extension information and also provides evidence

that the trial pack treatment increased the exchange of information.

90

VII. Conclusions

In this study, I examine the role of community-based extension and farmer personality on

adoption of improved bean varieties and social interactions. I develop a conceptual framework

demonstrating how farmer adoption of new varieties can be modeled as an asymmetric

information problem whereby the farmer seeks to discover the underlying, unobserved relative

benefits of a new bean variety based on extension activities. I show that farmer personality

can influence the likelihood of receiving a benefits signal. I then examine this heterogeneity

empirically using a unique dataset of the Big Six personality traits measured using the

Midlife Development Inventory (MIDI) for bean farmers in the Mbeya Region of Tanzania

(Lachman & Weaver, 1997).

In terms of adoption impacts, although the trial pack distribution increased immediate

planting of the improved varieties in the year of the intervention, diffusion rates of the

improved bean varieties converged between demo-only and demo-trial villages in the one

year post-intervention. Correspondingly, I find limited evidence that providing additional

extension resources to VBAAs in the form of trial packs of bean seed increased the average

farmer’s post-intervention likelihood of adoption of the improved bean varieties relative to

only providing VBAAs resources for a bean demonstration plot. This suggests that, on

average, the demonstration plot alone provided enough of a meaningful signal to farmers to

kickstart diffusion.

Although the benefits of the trial pack treatment were attenuated for the average farmer,

I do find evidence of heterogeneous treatment effects by personality traits. In particular,

extraverted farmers residing in trial pack villages were more likely to adopt Njano Uyole or

Uyole 96. These farmers were also more likely to seek out farming advice and discuss bean

farming with their neighbors, suggesting that the benefits of the trial pack treatment may

have been greater (smaller) for more (less) sociable farmers.

I also find that farmers residing in villages randomly selected to receive trial packs were

91

more likely to be able to identify their VBAA as well as have discussed bean farming with him

or her. Furthermore, farmer personality traits were also associated with differences in social

outcomes related to extension effectiveness. In particular, farmers with higher openness to

experience were more likely to know their VBAA as well as have discussed bean farming with

him or her. Similarly, farmers that scored higher on the openness to experience personality

trait also discussed bean farming with a greater proportion of farmers in their village on

average. These findings are consistent with the trial pack treatment and farmer personality

traits affecting farmer awareness of information on bean farming.

Although these findings should be interpreted as predictors and not necessarily causal,

they have important implications for the design of community-based extension programs

and warrant future research. Community-based extension programs are inherently social,

relying on already established informal institutional connections to facilitate information

flows. Personality influences interpersonal relationships, yet it has received relatively little

attention in the technology adoption and extension literature. My findings suggest that,

much like in social relationships, personality characteristics may influence who benefits from

community-based extension programs.

In this study’s context, the marginal gains from

providing an experience-based information treatment (like the trial pack) appear to have

been higher for extroverted recipient farmers. This suggests that personality influences the

potential beneficiaries of community-based extension programs and that community-based

extension effectiveness may be increased by considering the role of personality in the quality

and quantity of farmers interactions within their communities.

92

APPENDICES

93

APPENDIX A:

Tables and Figures

94

Table 3.1: VBAA Balance Tests

Mean Value

Demo-only Demo-trial P-value

Characteristics (as of baseline survey)
Age (years)
Is female
Level of education is primary or less
Farming experience (years)
Maize farming experience
Bean farming experience
Land area owned in acres (2016 ag. year)
Household size
Indicators of activities
Distributed free maize seed
Distributed free bean seed
Setup maize demonstration plot
Setup bean demonstration plot
Sold commercial maize seed
Sold commercial bean seed
Sold commercial fertilizer
Sold pesticides or seed treatments
No. free maize seed packs allocated by FIPS-Africa
No. of farmers given free maize seed packs
No. free bean seed packs allocated by FIPS-Africa
No. of farmers given free bean seed packs
No. of maize demonstration plots setup
No. of people involved in maize demonstration plot
No. of bean demonstration plots setup
No. of people involved in bean demonstration plot
Personality traits
Agency
Agreeableness
Openness to experience
Neuroticism
Extraversion
Conscientiousness

(n=16)
44.44
0.19
0.75
23.56
23.56
21.44
14.28
8.31

0.94
0.56
1.00
0.56
0.50
0.06
0.06
0.13
198.44
225.38
131.94
131.94
1.19
12.31
0.75
4.31

3.35
3.71
3.65
2.09
3.67
3.67

(n=16)
46.75
0.19
0.88
23.06
23.06
20.00
10.97
6.56

0.94
0.56
0.81
0.38
0.50
0.00
0.06
0.06
214.63
211.06
123.44
104.81
1.00
8.44
0.44
5.31

3.29
3.60
3.55
2.14
3.54
3.67

from t-test

0.508
1.000
0.382
0.889
0.889
0.729
0.476
0.156

1.000
1.000
0.083
0.303
1.000
0.333
1.000
0.560
0.743
0.782
0.888
0.641
0.480
0.235
0.250
0.688

0.590
0.429
0.476
0.820
0.357
1.000

Note: Tests of equality of means. t-test assumes unequal variances. Unless otherwise noted, VBAA
characteristics are based on current status at the time of the baseline VBAA survey in early 2017.
Indicators of VBAA performance are for the 2016 agricultural year. Personality traits were measured in
2018 on a scale of one (not at all), two (a little), three (some), and four (a lot).

95

Table 3.2: Farmer Survey Balance Tests

Characteristics
Grew Njano Uyole or Uyole 96 in 2016 ag. year
Received bean seed from FIPS-Africa VBAA in 2016 ag. year
Bean production area in 2016 ag. year (acres)
Land owned (acres)
Household head is female
Household head’s age (years)
Household’s highest level of education is primary or less
Household size
Respondent or spouse related to the village chairman
Respondent or spouse related to the village extension officer
Physical distance (km) between farmer’s and
VBAA’s homestead
Physical distance (km) between farmer’s
homestead and VBAA demo. plot
Personality traits
Agency
Agreeableness
Openness to experience
Neuroticism
Extraversion
Conscientiousness

Mean Value

Demo-only Demo-trial

P-value

(n=399)

(n=392)

from t-test

0.08
0.01
0.94
3.94
0.18
47.88
0.64
5.67
0.31
0.02

1.33

1.78

2.88
3.31
2.90
1.95
3.12
3.34

0.07
0.01
0.85
4.11
0.22
48.47
0.67
5.19
0.19
0.02

1.49

1.67

2.91
3.36
2.91
1.92
3.17
3.35

0.537
0.688
0.245
0.545
0.205
0.757
0.311
0.016
0.000
0.590

0.087

0.307

0.309
0.103
0.876
0.497
0.257
0.862

Note: Tests of equality of means. t-test assumes unequal variances. Unless noted below, characteristics were collected in
the 2017 farmer survey. Household head gender and age were collected in 2018. Physical distance is based on GPS data
collected in 2018. Personality traits were collected in 2018 on a scale of one (not at all), two (a little), three (some), and
four (a lot).

96

Table 3.3: Household Attrition Tests

Variable
Village treatment status (1=demo-trial; 0=demo only)

Household size in 2017

Respondent or spouse related to village chairman in 2017

Observations
Pseudo-R2

(1)
0.313
(0.294)

791

0.0277

(2)
0.158
(0.302)
−0.260∗∗∗
(0.073)
−0.763∗
(0.460)

791

0.0776

Note: Logit regressions. Dependent variable equals one if the household is only observed in the
first survey round (2017), zero otherwise. All regressions include indicator variables for
VBAA-pair used in the random assignment. Standard errors in parentheses.
* p < 0.10; ** p < 0.05; *** p < 0.01.

97

Table 3.4: Key Characteristics by Agricultural Year and Treatment Group

2016 Ag. Year
Pre-intervention

2017 Ag. Year

Intervention year

2018 Ag. Year

Post-intervention

Demo-only Demo-trial Demo-only Demo-trial Demo-only Demo-trial

(n=399)

(n=392)

(n=399)

(n=392)

(n=377)

(n=363)

Variable
Information specific to the given ag. year:
Farmer’s household grew Njano Uyole or Uyole 96

Farmer’s household has ever received training on
Uyole 96 or Njano Uyole
Farmer’s household received a trial pack of bean seed

Bean production area (acres)

0.08
(0.27)

-

0.01
(0.09)
0.94
(1.11)

0.07
(0.25)

-

0.01
(0.10)
0.85
(1.05)

Information collected in 2018, but not specific to that ag. year:
Farmer has ever met VBAA and identified him/her
-
as the VBAA
Farmer has ever sought farming advice from VBAA

-

-

-

0.09
(0.28)
0.05
(0.22)
0.05
(0.22)
1.35
(1.07)

-

-

0.13
(0.34)
0.05
(0.23)
0.16
(0.37)
1.22
(1.09)

-

-

-

0.20
(0.40)
0.16
(0.37)
0.03
(0.18)
1.13
(1.07)

0.552
(0.498)
0.475
(0.500)
0.491
(0.501)

0.786
(0.225)
0.130
(0.223)
0.151
(0.228)

0.21
(0.41)
0.17
(0.37)
0.08
(0.28)
1.04
(1.08)

0.634
(0.482)
0.474
(0.500)
0.521
(0.500)

0.847
(0.192)
0.122
(0.209)
0.158
(0.235)

Farmer has ever discussed bean farming with VBAA
Proportion of a random sample of ∼ 24 bean farming households in the farmer’s village that he/she...
...has ever met

-

-

-

-

-

-

-

...has ever sought farming advice from

...has ever discussed bean farming with

-

-

-

-

-

-

-

-

Note: Mean values by agricultural year and treatment group. Standard deviations in parentheses. Dash indicates outcome not measured for given
year. Characteristics for 2016 agricultural year were collected retroactively in 2017. The trial pack variable for 2016 is for any bean seed received
from a FIPS-Africa VBAA (more broad than the Njano Uyole or Uyole 96 bean seed definition used for 2017 and 2018).

98

Table 3.5: Midlife Development Inventory (MIDI) Personality Traits

Corresponding adjectives
Self-confident, Forceful, Assertive, Outspoken, Dominant
Helpful, Warm, Caring, Softhearted, Sympathetic

Trait
Agency
Agreeableness
Openness to experience Creative, Imaginative, Intelligent, Curious, Broadminded, Sophisticated, Adventurous
Neuroticism
Extraversion
Conscientiousness

Moody, Worrying, Nervous, Calm(-)
Outgoing, Friendly, Lively, Active, Talkative
Organized, Responsible, Hardworking, Careless(-), Thorough

Note: Adapted from Lachman and Weaver (2005). Respondents rated how well each adjective described them on a scale of: one (not
at all), two (a little), three (some), and four (a lot). Traits are the average scores of the associated adjectives. (-) indicates that the
adjective was reverse coded when scoring.

99

Table 3.6: Personality Trait Cronbach’s

Alpha Values

Trait
Agency
Agreeableness
Openness to Experience
Neuroticism
Extraversion
Conscientiousness

Alpha (N=772)

0.60
0.78
0.74
0.59
0.72
0.70

100

Table 3.7: Farmer-VBAA Personality Traits

Comparison

Trait

Agency

Agreeableness

Openness

Neuroticism

Extraversion

Conscientiousness

Farmers VBAAs
(N=740)
(N=32)

P-value

from t-test

2.89
(0.52)
3.33
(0.48)
2.91
(0.47)
1.94
(0.59)
3.14
(0.5)
3.35
(0.47)

3.32
(0.32)
3.66
(0.39)
3.60
(0.38)
2.12
(0.57)
3.61
(0.41)
3.67
(0.4)

0.000

0.000

0.000

0.088

0.000

0.000

Note: t-tests of equality of means assuming unequal variances.
Personality traits were measured on a scale of one (not at all),
two (a little), three (some), and four (a lot).

101

Table 3.8: Physical Distance Summary Statistics

Variable
Physical distance (km) between farmer’s and
VBAA’s homestead
Physical distance (km) between farmer’s
homestead and VBAA demonstration plot

Mean Std. Dev. Min Max

1.41

1.73

1.34

0.02

7.11

1.48

0.05

7.24

Note: N=740. Physical distances are GPS-based linear distances.

102

Table 3.9: Farmer Adoption of Improved Bean Varieties in 2018 Ag. Year

Variable
Trial

Agency

Agreeableness

Openness to experience

Neuroticism

Extraversion

Conscientiousness

(1)
0.007
(0.032)

(3)

(2)
0.097∗∗
0.047
(0.033)
(0.046)
0.028
0.032
(0.039)
(0.038)
0.019
0.021
(0.038)
(0.040)
−0.042 −0.048
(0.054)
(0.051)
0.012
0.011
(0.019)
(0.021)
0.062
0.062
(0.041)
(0.041)
−0.019 −0.023
(0.044)
(0.042)

VBAA-pair indicators
VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
Pseudo R-squared
P-value of test that farmer personality
average partial effects are jointly zero

Yes
No
No
740
0.171

Yes
Yes
No
740
0.211

Yes
Yes
Yes
740
0.220

0.158

0.104

Note: Logit average partial effects with standard errors clustered at the village level in
parentheses. Dependent variable is equal to one if the farmer’s household adopted
Njano Uyole or Uyole 96 in the 2018 agricultural year. Trial is an indicator variable
equal to one if the farmer resides in a village where the VBAA received trial packs.
Personality traits are measured on a scale of one (not at all) to four (a lot).
* p < 0.10; ** p < 0.05; *** p < 0.01.
\a VBAA personality controls are the same as that of the farmer. Distance controls are
the GPS-based physical distances (km) from the farmer’s homestead to the VBAA’s
homestead and VBAA demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for
farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age;
indicator for farmer household’s and VBAA’s highest education achieved is primary or
less.

103

Table 3.10: Heterogeneous Trial Pack Effects on Farmer

Adoption of Improved Bean Varieties in 2018 Ag. Year

Variable
Trial

Trial*Agency

Trial*Agreeableness

Trial*Openness to experience

Trial*Neuroticism

Trial*Extraversion

Trial*Conscientiousness

(1)
(2)
0.234
0.312
(0.219)
(0.235)
−0.048 −0.049
(0.075)
(0.073)
−0.071 −0.068
(0.087)
(0.082)
−0.002 −0.007
(0.112)
(0.107)
0.008
0.000
(0.042)
(0.041)
0.152∗∗
0.149∗∗
(0.072)
(0.072)
−0.090 −0.092
(0.078)
(0.078)

VBAA-pair indicators
Farmer and VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
R-squared

Yes
Yes
No
740
0.219

Yes
Yes
Yes
740
0.227

Note: Linear probability model coefficients with standard errors clustered at the village
level in parentheses. Dependent variable is equal to one if the farmer’s household adopted
Njano Uyole or Uyole 96 in the 2018 agricultural year. Trial is an indicator variable equal
to one if the farmer resides in a village where the VBAA received trial packs.
* p < 0.10; ** p < 0.05; *** p < 0.01.
\a Farmer and VBAA personality controls are their six personality traits measured on a
scale of one (not at all) to four (a lot). Distance controls are the GPS-based physical
distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA
demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for
farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age;
indicator for farmer household’s and VBAA’s highest education achieved is primary or less.

104

Table 3.11: Farmer Interactions with VBAA

Met VBAA

Farmer has ever...
Sought Farming

(1)
0.081∗
(0.044)

Variable
Trial

Agency

Agreeableness

Openness to experience

Neuroticism

Extraversion

Conscientiousness

(2)

(3)
(5)
0.059 −0.000 0.025
0.104∗∗∗
(0.036)
(0.035) (0.021)
(0.037)
−0.054
−0.056
(0.039)
(0.038)
0.073∗
0.103∗∗
(0.044)
(0.044)
0.114∗∗
0.057
(0.049)
(0.055)
−0.159∗∗∗ −0.164∗∗∗
(0.037)
(0.039)
−0.085
−0.078
(0.055)
(0.052)
−0.000
0.011
(0.051)
(0.053)

Advice from VBAA
(6)
(4)
0.057∗∗
(0.028)
−0.069 −0.111∗∗
(0.051)
(0.053)
−0.012
0.016
(0.059)
(0.063)
0.097∗
0.038
(0.059)
(0.056)
0.018
0.015
(0.044)
(0.041)
−0.094 −0.086
(0.061)
(0.064)
0.067
0.060
(0.067)
(0.069)

Discussed Bean

(8)

Farming with VBAA
(7)
(9)
0.069∗∗
0.031 0.068∗∗∗
(0.029)
(0.038) (0.023)
−0.086∗ −0.134∗∗∗
(0.046)
(0.047)
0.018
0.043
(0.054)
(0.060)
0.123∗∗
0.076
(0.059)
(0.054)
0.028
0.025
(0.045)
(0.043)
−0.047 −0.040
(0.063)
(0.064)
0.067
0.054
(0.057)
(0.060)

VBAA-pair indicators
VBAA personality and distance \a
Farmer and VBAA demographics \b
Observations
Pseudo R-squared
P-value of test that farmer personality
average partial effects are jointly zero

Yes
No
No
740
0.058

Yes
Yes
No
740
0.169

0.000

Yes
Yes
Yes
740
0.193

0.000

Yes
No
No
740
0.038

Yes
Yes
No
740
0.070

0.064

Yes
Yes
Yes
740
0.096

0.002

Yes
No
No
740
0.043

Yes
Yes
No
740
0.084

0.011

Yes
Yes
Yes
740
0.112

0.001

Note: Logit average partial effects with standard errors clustered at the village level in parentheses. Trial is an indicator variable equal to one if
the farmer resides in a village where the VBAA received trial packs. Personality traits are measured on a scale of one (not at all) to four (a lot).
* p < 0.10; ** p < 0.05; *** p < 0.01. \a VBAA personality controls are the same as that of the farmer. Distance controls are the GPS-based
physical distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA demonstration plot. \b Demographic controls (2017
ag. year) are: farmer household’s size; indicators for farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age;
indicator for farmer household’s and VBAA’s highest education achieved is primary or less.

105

Variable
Trial

Agency

Agreeableness

Openness to experience

Neuroticism

Extraversion

Conscientiousness

Table 3.12: Farmer Interactions with Other Bean Farmers

Proportion of Bean Farming Households that the Farmer has ever...

Met

Sought Farming

Advice From

Discussed Bean
Farming With

(1)

(2)

0.417∗∗∗ 0.052∗∗∗ 0.080∗∗∗ −0.086
(0.021) (0.211)
(0.144)

(3)

(4)

(0.014)
0.006 −0.015
(0.024)
(0.024)
0.013
0.008
(0.024)
(0.023)
0.023
0.014
(0.029)
(0.027)
0.039∗∗∗ 0.037∗∗∗
(0.013)
(0.012)
0.034
0.033
(0.028)
(0.025)
−0.034 −0.022
(0.030)
(0.029)

(5)
(6)
0.009
0.026
(0.020)
(0.020)
−0.066∗∗∗ −0.092∗∗∗
(0.017)
(0.019)
−0.043∗∗ −0.026
(0.019)
(0.022)
0.053∗∗∗
0.010
(0.020)
(0.022)
−0.024 −0.024∗
(0.014)
(0.016)
0.000
0.011
(0.030)
(0.032)
0.036∗
0.031
(0.022)
(0.023)

(7)
0.049
(0.181)

(9)

(8)
0.035∗∗
0.013
(0.021)
(0.017)
−0.061∗∗∗ −0.086∗∗∗
(0.020)
(0.021)
−0.026
−0.038
(0.023)
(0.025)
0.079∗∗∗
0.043∗
(0.023)
(0.019)
0.003
0.002
(0.018)
(0.018)
0.033
0.047
(0.032)
(0.033)
0.039∗
0.047∗∗
(0.020)
(0.020)

VBAA-pair indicators
VBAA personality and distance \a
Farmer and VBAA demographics \b
Observations
Pseudo R-squared
P-value of test that farmer personality
average partial effects are jointly zero

Yes
No
No
740
0.054

Yes
Yes
No
740
0.069

Yes
Yes
Yes
740
0.083

Yes
No
No
740
0.056

0.003

0.004

Yes
Yes
No
740
0.097

0.001

Yes
No
No
740
0.063

Yes
Yes
Yes
740
0.135

0.000

Yes
Yes
No
740
0.093

0.000

Yes
Yes
Yes
740
0.118

0.000

Note: Fractional response logistic regression average partial effects with standard errors clustered at the village level in parentheses. Proportions
are based on a random sample of ∼ 25 bean farming households in the farmer’s village. Trial is an indicator variable equal to one if the farmer
resides in a village where the VBAA received trial packs. Personality traits are measured on a scale of one (not at all) to four (a lot).
* p < 0.10; ** p < 0.05; *** p < 0.01. \a VBAA personality controls are the same as that of the farmer. Distance controls are the GPS-based
physical distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA demonstration plot. \b Demographic controls (2017 ag.
year) are: farmer household’s size; indicators for farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age; indicator for
farmer household’s and VBAA’s highest education achieved is primary or less.

106

Table 3.13: Heterogeneous Trial Pack Effects on the Proportion of Bean
Farming Households that the Farmer has Sought Farming Advice From

Variable
Trial

Trial*Agency

Trial*Agreeableness

Trial*Openness to experience

Trial*Neuroticism

Trial*Extraversion

Trial*Conscientiousness

(2)

(1)

−0.131 −0.130
(0.142)
(0.140)
−0.073∗ −0.069∗
(0.038)
(0.039)
−0.004 −0.015
(0.038)
(0.039)
−0.020
0.005
(0.045)
(0.048)
0.060∗∗
0.060∗∗
(0.025)
(0.027)
0.122∗
0.110∗
(0.057)
(0.063)
−0.029 −0.030
(0.044)
(0.044)

VBAA-pair indicators
Farmer and VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
R-squared

Yes
Yes
No
740
0.195

Yes
Yes
Yes
740
0.254

Note: Linear probability model coefficients with standard errors clustered at the village level
in parentheses. Dependent variable is the proportion of a random sample of ∼ 25 bean
farming households in the farmer’s village that he/she has sought farming advice from. Trial
is an indicator variable equal to one if the farmer resides in a village where the VBAA
received trial packs. * p < 0.10; ** p < 0.05; *** p < 0.01.
\a Farmer and VBAA personality controls are their six personality traits measured on a
scale of one (not at all) to four (a lot). Distance controls are the GPS-based physical
distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA
demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for farmer
or spouse related to village chairmen; farmer’s and VBAA’s gender and age; indicator for
farmer household’s and VBAA’s highest education achieved is primary or less.

107

Table 3.14: Heterogeneous Trial Pack Effects on the Proportion of Bean
Farming Households that the Farmer has Discussed Bean Farming With

Variable
Trial

Trial*Agency

Trial*Agreeableness

Trial*Openness to experience

Trial*Neuroticism

Trial*Extraversion

Trial*Conscientiousness

(2)

(1)

−0.206 −0.204
(0.205)
(0.199)
−0.061 −0.055
(0.043)
(0.042)
−0.037 −0.047
(0.046)
(0.044)
−0.021 −0.005
(0.042)
(0.044)
0.054
0.056
(0.036)
(0.036)
0.157∗∗∗
0.145∗∗
(0.054)
(0.057)
−0.011 −0.009
(0.043)
(0.042)

VBAA-pair indicators
Farmer and VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
R-squared

Yes
Yes
No
740
0.213

Yes
Yes
Yes
740
0.257

Note: Linear probability model coefficients with standard errors clustered at the village
level in parentheses. Dependent variable is the proportion of a random sample of ∼ 25
bean farming households in the farmer’s village that he/she has discussed bean farming
with. Trial is an indicator variable equal to one if the farmer resides in a village where the
VBAA received trial packs. * p < 0.10; ** p < 0.05; *** p < 0.01.
\a Farmer and VBAA personality controls are their six personality traits measured on a
scale of one (not at all) to four (a lot). Distance controls are the GPS-based physical
distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA
demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for farmer
or spouse related to village chairmen; farmer’s and VBAA’s gender and age; indicator for
farmer household’s and VBAA’s highest education achieved is primary or less.

108

APPENDIX B:

Supplementary Tables

109

Table 3.15: Farmer Adoption of Improved Bean Varieties in 2018 Ag.

Year: Controlling for Pre-Intervention Outcome

Variable
Trial

Agency

Agreeableness

Openness to experience

Neuroticism

Extraversion

Conscientiousness

(1)
0.007
(0.032)

(3)

(2)
0.098∗∗
0.048
(0.033)
(0.046)
0.029
0.033
(0.038)
(0.038)
0.017
0.019
(0.038)
(0.040)
−0.045 −0.050
(0.054)
(0.051)
0.012
0.010
(0.019)
(0.021)
0.065
0.065
(0.041)
(0.040)
−0.018 −0.021
(0.044)
(0.042)

VBAA-pair indicators
VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
Pseudo R-squared
P-value of test that farmer personality
average partial effects are jointly zero

Yes
No
No
740
0.171

Yes
Yes
No
740
0.212

Yes
Yes
Yes
740
0.221

0.132

0.0797

Note: Logit average partial effects with standard errors clustered at the village level in
parentheses. Dependent variable is equal to one if the farmer’s household adopted
Njano Uyole or Uyole 96 in the 2018 agricultural year. All specifications control for
adoption of Njano Uyole or Uyole 96 in the 2016 agricultural year. Trial is an indicator
variable equal to one if the farmer resides in a village where the VBAA received trial
packs. Personality traits are measured on a scale of one (not at all) to four (a lot). *
p < 0.10; ** p < 0.05; *** p < 0.01.
\a VBAA personality controls are the same as that of the farmer. Distance controls are
the GPS-based physical distances (km) from the farmer’s homestead to the VBAA’s
homestead and VBAA demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for
farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age;
indicator for farmer household’s and VBAA’s highest education achieved is primary or
less.

110

Table 3.16: Heterogeneous Trial Pack Effects on Farmer
Adoption of Improved Bean Varieties in 2018 Ag. Year:

Controlling for Pre-Intervention Outcome

Variable
Trial

Trial*Agency

Trial*Agreeableness

Trial*Openness to experience

Trial*Neuroticism

Trial*Extraversion

Trial*Conscientiousness

(1)
(2)
0.239
0.313
(0.219)
(0.236)
-0.047
-0.048
(0.073)
(0.075)
−0.075 −0.071
(0.085)
(0.080)
−0.002 −0.006
(0.113)
(0.108)
0.008
0.000
(0.042)
(0.041)
0.154∗∗
0.151∗∗
(0.072)
(0.072)
−0.091 −0.093
(0.077)
(0.077)

VBAA-pair indicators
Farmer and VBAA personality and distance controls \a
Farmer and VBAA demographic controls \b
Observations
R-squared

Yes
Yes
No
740
0.220

Yes
Yes
Yes
740
0.228

Note: Linear probability model coefficients with standard errors clustered at the village
level in parentheses. Dependent variable is equal to one if the farmer’s household adopted
Njano Uyole or Uyole 96 in the 2018 agricultural year. All specifications control for
adoption of Njano Uyole or Uyole 96 in the 2016 agricultural year. Trial is an indicator
variable equal to one if the farmer resides in a village where the VBAA received trial
packs. * p < 0.10; ** p < 0.05; *** p < 0.01.
\a Farmer and VBAA personality controls are their six personality traits measured on a
scale of one (not at all) to four (a lot). Distance controls are the GPS-based physical
distances (km) from the farmer’s homestead to the VBAA’s homestead and VBAA
demonstration plot.
\b Demographic controls (2017 ag. year) are: farmer household’s size; indicators for
farmer or spouse related to village chairmen; farmer’s and VBAA’s gender and age;
indicator for farmer household’s and VBAA’s highest education achieved is primary or less.

111

REFERENCES

112

REFERENCES

Ali, D., Bowen, D., & Deininger, K. (2017). Personality Traits, Technology Adoption, and
Technical Efficiency: Evidence from Smallholder Rice Farms in Ghana (Policy Re-
search Working Paper No. 7959). World Bank.

Anderson, J. R., & Feder, G. (2004). Agricultural Extension: Good Intentions and Hard

Realities. The World Bank Research Observer, 19 (1), 41–60.

Bazzani, C., Caputo, V., Nayga, R. M., & Canavari, M. (2017). Revisiting consumers’ valu-
ation for local versus organic food using a non-hypothetical choice experiment: Does
personality matter? Food Quality and Preference, 62, 144–154.

BenYishay, A., & Mobarak, A. M. (2019). Social Learning and Incentives for Experimentation

and Communication. The Review of Economic Studies, 86 (3), 976–1009.

Brandst¨atter, H., Brandst¨atter, V., & Pelka, R. B. (2018). Similarity and Positivity of Per-
sonality Profiles Consistently Predict Relationship Satisfaction in Dyads. Frontiers
in Psychology, 9.

Foster, A. D., & Rosenzweig, M. R. (2010). Microeconomics of Technology Adoption. Annual

Review of Economics, 2 (1), 395–424.

Grebitus, C., Lusk, J. L., & Nayga, R. M. (2013). Explaining differences in real and hypo-
thetical experimental auctions and choice experiments with personality. Journal of
Economic Psychology, 36, 11–26.

Jack, K. (2013). Market innefficiencies and the adoption of agricultural technologies in de-

veloping countries (Working Paper).

Krishnan, P., & Patnam, M. (2014). Neighbors and Extension Agents in Ethiopia: Who
Matters More for Technology Adoption? American Journal of Agricultural Economics,
96 (1), 308–327.

Laajaj, R., Macours, K., Pinzon Hernandez, D. A., Arias, O., Gosling, S. D., Potter, J.,
Rubio-Codina, M., & Vakis, R. (2019). Challenges to capture the big five personality
traits in non-WEIRD populations. Science Advances, 5 (7).

Lachman, M., & Weaver, S. (1997). The Midlife Development Inventory (MIDI) Personality

Scales: Scale Construction and Scoring. Brandeis University. Waltham, MA.

113

Lachman, M., & Weaver, S. (2005). Addendum for MIDI Personality Scales. Brandeis Uni-

versity. Waltham, MA.

Lin, W., Ortega, D. L., Caputo, V., & Lusk, J. L. (2019). Personality traits and consumer ac-
ceptance of controversial food technology: A cross-country investigation of genetically
modified animal products. Food Quality and Preference, 76, 10–19.

Macho-Stadler, I., & P´erez-Castrillo, J. (2001). An Introduction to the Economics of Infor-

mation: Incentives and Contracts (2nd ed.). Oxford University Press.

McKenzie, D. (2012). Beyond baseline and follow-up: The case for more T in experiments.

Journal of Development Economics, 99 (2), 210–221.

Melkani, A., & Mason, N. M. (2018). Tanzania Southern Highlands Farm Input Promotions-
Africa Village-Based Agricultural Advisors Baseline Survey Report. Michigan State
University.

Morgan, S. N. (2018). The Experimental Science of Economic Behavior: Testing Theories of

Participation, Valuation, and Innovation. Michigan State University.

Nakano, Y., Tsusaka, T. W., Aida, T., & Pede, V. O. (2018). Is farmer-to-farmer extension
effective? The impact of training on technology adoption and rice farming productivity
in Tanzania. World Development, 105, 336–351.

Norton, G. W., & Alwang, J. (2020). Changes in Agricultural Extension and Implications
for Farmer Adoption of New Practices. Applied Economic Perspectives and Policy,
aepp.13008.

Ozer, D. J., & Benet-Mart´ınez, V. (2006). Personality and the Prediction of Consequential

Outcomes. Annual Review of Psychology, 57 (1), 401–421.

Papke, L. E., & Wooldridge, J. M. (1996). Econometric methods for fractional response
variables with an application to 401(k) plan participation rates. Journal of Applied
Econometrics, 11, 619–632.

Roberts, B. W. (2009). Back to the future: Personality and Assessment and personality

development. Journal of Research in Personality, 43 (2), 137–145.

Soto, C. J. (2019). How Replicable Are Links Between Personality Traits and Consequential
Life Outcomes? The Life Outcomes of Personality Replication Project. Psychological
Science, 1–17.

114

Van der Linden, D., Dunkel, C. S., Figueredo, A. J., Gurven, M., Von Rueden, C., & Woodley
of Menie, M. A. (2018). How Universal Is the General Factor of Personality? An
Analysis of the Big Five in Forager Farmers of the Bolivian Amazon. Journal of
Cross-Cultural Psychology, 49 (7), 1081–1097.

Weidmann, R., Sch¨onbrodt, F. D., Ledermann, T., & Grob, A. (2017). Concurrent and
longitudinal dyadic polynomial regression analyses of Big Five traits and relationship
satisfaction: Does similarity matter? Journal of Research in Personality, 70, 6–15.

Wilson, K. S., DeRue, D. S., Matta, F. K., Howe, M., & Conlon, D. E. (2016). Personality
Similarity in Negotiations: Testing the Dyadic Effects of Similarity in Interpersonal
Traits and the Use of Emotional Displays on Negotiation Outcomes. Journal of Ap-
plied Psychology, 101 (10), 1405–1421.

Wooldridge, J. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, MA, MIT Press.

115