Policy Research Working Paper                  9532




                  The Demand for Advice
         Theory and Empirical Evidence from Farmers
                    in Sub-Saharan Africa

                           Dominik Naeher
                          Matthias Schündeln




Development Economics
Development Policy Team
February 2021
Policy Research Working Paper 9532


  Abstract
 Low levels of investment into modern technologies, and                             available free of charge, but positive effects on production
 limited use of measures that have low monetary cost but                            are only realized if farmers devote attention to listening to
 the potential for high yields, are often regarded as obstacles                     and implementing the provided advice. Modeling farmers
 to further agricultural development. This paper investigates                       as rational decision makers facing scarce attention, the study
 farmers’ demand for one such measure, namely agricul-                              identifies the circumstances under which farmers may opti-
 tural advisory services. These have modest (most frequently                        mally abstain from demanding advisory services. The model
 zero) monetary user cost but, according to some recent                             complements the insights of other theories commonly used
 research, have the potential to result in large increases of                       to explain suboptimal farm decisions and outcomes, and
 yields. Yet, demand for these extension services is often low.                     generates testable predictions, which are consistent with
 This study proposes that costly attention may be part of the                       empirical evidence based on a large farm-level panel dataset
 explanation for this. In the model, advisory services are                          from Sub-Saharan Africa.




 This paper is a product of the Development Policy Team, Development Economics. It is part of a larger effort by the
 World Bank to provide open access to its research and make a contribution to development policy discussions around the
 world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may
 be contacted at dominik.naeher@ucd.ie.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
          The demand for advice: Theory and empirical
          evidence from farmers in Sub-Saharan Africa*
                    Dominik Naeher„                             undeln…
                                                    Matthias Sch¨




Keywords: economic development, advice, agricultural extension, rational inattention,
Sub-Saharan Africa


JEL Codes: D91, O13, Q16



  * The  authors thank Eric Edmonds, Mirko Wiederholt, and anonymous referees for extensive comments
and suggestions, and Sandro Ambuehl, Luc Christiaensen, Rema Hanna, and Sara Savastano for helpful
discussions.
   „ Corresponding author. University College Dublin, Ireland (e-mail: dominik.naeher@ucd.ie).
   … Goethe University Frankfurt, Germany (e-mail: schuendeln@wiwi.uni-frankfurt.de).
1       Introduction
Increases in productivity are key to growth and poverty reduction, yet productivity in
many domains is lacking far behind technological possibilities. This is particularly true
for agricultural productivity in many parts of the developing world, especially in Sub-
Saharan Africa (Jama and Pizarro, 2008; World Bank, 2008). One puzzling observation is
that farmers do not make use of measures that have relatively low costs but are believed
to offer high returns. In particular, modern farm inputs, such as fertilizer and improved
seeds, have been demonstrated in a variety of contexts to have significant expected returns,
yet are frequently not adopted and, conditional on adoption, inputs are used in suboptimal
amounts. To help farmers overcome the underlying constraints, a major policy instrument
used is the provision of agricultural extension and advisory services.1
      A growing body of empirical evidence suggests that these services can have signifi-
cant positive effects, including on farmers’ management practices and adoption of better
technologies (Cole and Fernando, 2013; Buehren et al., 2019), crop yields (Casaburi et al.,
2014), profits (Bandiera et al., 2018), and poverty reduction (Dercon et al., 2009). Yet,
demand for extension services is often low. In our dataset, which consists of nationally
representative samples of farmers in three African countries, only 19% of farmers actively
solicit agricultural advice (including from other farmers). The fairly low interest in agri-
cultural advice despite the apparent potential for high returns is particularly puzzling in
light of the low (most frequently zero) monetary user cost associated with receiving such
advice.2
      In this paper, we use a rational inattention model to argue that costly attention may
be part of the explanation for this. Our model centers around the idea that farmers (like
anybody else) face a limited capacity to attend to information, which causes attention to
be a scarce resource.3 Given that participation in extension programs is only worthwhile
if farmers devote sufficient attention to listening to and implementing the provided ad-
vice, the decision to request extension services will depend on the amount of attention
that farmers are willing to devote to their agricultural production process. Based on
    1
      According to different sources, there are between half a million and one million agricultural extension
workers worldwide, of which 90% are located in developing countries (Feder, 2005; Anderson and Feder,
2007).
    2
      In our sample, 92% of respondents report having paid nothing for the advice that they received.
Also note that the share of districts in which at least one farmer reports having received agricultural
advice is 95% (84% of district-years), which suggests that extension services are widely available in the
countries we study. This view is also in line with the findings of other studies. For example, surveying
recent studies on extension services in Malawi, Ragasa (2018) report that three-quarters of Malawian
farm households received agricultural advice in the last two years, and half of households received advice
in the last 12 months.
    3
      This feature is in line with a large body of psychological and experimental evidence on the limits of
human cognition (DellaVigna, 2009; Caplin and Dean, 2013; World Bank, 2015), as well as a growing
literature on the link between poverty and lack of mental resources (Banerjee and Mullainathan, 2008;
Mani et al., 2013; Haushofer and Fehr, 2014).



                                                     2
this feature, we model household decision making in two domains, agriculture and non-
agriculture, and identify the circumstances under which households will optimally abstain
from demanding agricultural advice. In line with our theoretical results, we argue that
considering costly attention as a binding constraint for farmers can contribute to a better
understanding of the determinants of farmers’ demand for advice.
     One particular theoretical result, which is helpful in distinguishing the proposed
channel from other theories commonly used in the literature to explain suboptimal farm
decisions and outcomes, is that shocks in the non-agricultural domain may increase de-
mand for agricultural advice. This prediction runs counter to much of the recent literature
on psychology and poverty (e.g., Banerjee and Mullainathan, 2008; Haushofer and Fehr,
2014), according to which negative shocks in the non-agricultural domain would tend to
absorb cognitive bandwidth, leading to less attention paid to agricultural production and
thus less use of agricultural advice. The intuition is as follows. In the model, agricultural
advisory services are available to farmers free of charge and farmers expect a positive
effect on their production when participating in such programs. Absorbing advice and
implementing the activities suggested by extension workers, however, requires attention,
which is a costly resource as it can also be used to attend to decisions in other areas of
the farmer’s life (leading to expected utility gains in these domains). Shocks in the non-
agricultural domain may raise the marginal utility of income, and consequently increase
farmers’ willingness to devote costly attention to the agricultural domain with the goal of
generating additional income, thus increasing the probability that extension services are
demanded.4
     As an example, consider the following illustration. The household head is busy
spending time with and thinking about a lot of things that he cannot avoid, but he would
typically spend some of his time and attention on attending village meetings and discussing
issues of village-level development. This is at the costs of being able to spend time on
meeting the extension agent and spending some of his scarce attention on following up
on what the extension agent tells him. Now some money is stolen or his house burns
down. Now he is poorer and every dollar counts more. He hopes that spending time
with the extension agent eventually provides additional dollars through better agricultural
practices. Thus, instead of going to the village meeting and spending attentional resources
on village-development issues, the household head now spends time and attention on
meeting with the extension agent and following the received advice.
   4
     Our model is based on rational inattention (Sims, 2003), which uses entropy to model costly attention.
However, the prediction that non-agricultural shocks can lead to increased demand for agricultural advice
arises independently of the specific structure of attention assumed in the rational inattention literature,
and could therefore also be generated by a simpler model that is not based on entropy. As we discuss
below, there are several advantages of using the rational inattention approach to model farmers’ demand
for advice.




                                                    3
     Relying on detailed farm-level data on shocks (both related and not directly related
to agriculture) and the use of extension services from Sub-Saharan Africa, we then demon-
strate that empirical observations are consistent with the predictions of our model.5 In
particular, we show that there is statistically strong and robust evidence that, in line
with the model, non-agricultural shocks are indeed associated with increased demand for
extension. For this purpose, we split shocks into two categories, namely shocks that are
directly related to agriculture (e.g., pest outbreaks and irregular rainfalls) and shocks
that are not directly related to agriculture (e.g., shocks to farmers’ wealth or health).
In addition, our data allow us to separately identify extension services that are actively
requested by farmers from unsolicited extension services of which the household is merely
a passive recipient.
     Given that the agricultural and non-agricultural domains are likely not separable
in the context we study (see Benjamin, 1992; LaFave and Thomas, 2016), one should
be worried that there may be other channels than the one implied by our model that
account for the observed link between non-agricultural shocks and demand for extension.
Our data allow us to assess the empirical plausibility of several other potential channels in
this context, but we do not find strong evidence in favor of those alternative explanations.
In particular, there is evidence that suggests that the effect of non-agricultural shocks is
not due to an increased need for credit (e.g., driven by shocks to non-farm income and
remittances) or because non-agricultural shocks (some of which may affect household
labor supply) work through a demand for advice on labor inputs. Overall, our modeling
insights are more broadly applicable and the empirical support for the model provided in
this paper suggests that scarce attention may be part of explanations for phenomena in
economic development, including beyond the specific application studied here.
     The remainder of the paper is organized as follows. Section 2 discusses related lit-
erature and provides further background on agricultural extension services. Section 3
presents the model and derives predictions about farmers’ demand for agricultural ad-
vice. Section 4 provides empirical evidence on the predictions of the model and addresses
alternative explanations. Section 5 concludes.


2       Related literature and background on extension
        services
Our paper contributes to several different literatures. First, we add to the recent literature
in development economics that tries to explain low levels of technology adoption and low
input use in agriculture, in particular in Sub-Saharan Africa. Various channels have
    5
     The empirical work is based on survey data from the Living Standard Measurement Study - Inte-
grated Surveys on Agriculture (LSMS-ISA) database (World Bank, 2018).



                                                4
been considered in this literature, including market imperfections related to finance and
insurance (Moser and Barrett, 2006; Dercon and Christiaensen, 2011; Karlan et al., 2014),
uncertainty about quality and returns (Suri, 2011; Bold et al., 2017), individual and
social learning (Foster and Rosenzweig, 1995; Bandiera and Rasul, 2006; Conley and
Udry, 2010; Hanna et al., 2014), high cost of providing information if schooling levels
are low (Schultz, 1964), limited attention and other informational frictions (Cole and
Fernando, 2013; Casaburi et al., 2014; Van Campenhout et al., 2017; Naeher, 2020), as
well as behavioral biases (Duflo et al., 2011). Indeed, some of these channels have been
shown to be important across a variety of contexts. While some of the related papers
have highlighted the role of scarce mental resources in determining farm outcomes, this
paper suggests a new specific channel based on (imperfect) demand for advice through
which limited attention affects agricultural input choices and productivity, and provides
empirical support for this channel.
     Second, we contribute to the literature on the determinants and constraints of the use
of agricultural extension services and their returns. The theoretical power to transform
agriculture in developing countries has prominently been described by Schultz (1964). An
important question in this literature is why extension services are not demanded more
widely by farmers despite a growing body of evidence that extension services can have
positive effects (see the studies cited below). One possible explanation is that the findings
in these studies that show fairly large returns to extension cannot be generalized beyond
the specific settings they study. Further, benefits may be heterogeneous, and extension
may benefit only a small share of farmers. For example, the content (see Duflo et al.,
2008) or the timing (Cole and Fernando, 2013) of advice provided may not be adequate
for all households, especially in contexts where there is large information loss along the
advice chain so that extension field staff do not fully understand themselves the advice
they are to provide (Niu and Ragasa, 2018). We do not rule out the possibility that
low (expected) returns to extension are part of the explanation of low demand. Instead,
we introduce a novel mechanism to this discussion which we argue may be part of the
explanation for the low demand for extension.
     Third, we add to the body of literature that studies the role of attentional constraints
for economic decision making. Such constraints can take different forms. For example,
attention has been modeled as a stimulus-driven allocation process, emphasizing the im-
portance of salient aspects of different pieces of information over the true informational
value they carry (Bordalo et al., 2013; K˝oszegi and Szeidl, 2013).6 In this paper, we follow
   6
    In these models, attention allocation is exogenous to the decision maker and determined by external
signals rather than by optimizing behavior. However, such models appear to be unsuited for explaining
the positive correlation between farm-unrelated shocks and demand for agricultural advice that we find in
the data. It seems reasonable to assume that non-agricultural shocks increase the salience of the affected
(non-agricultural) domain, which in these models would reduce rather than increase farmers’ attention
to agriculture, including obtaining agricultural advice.



                                                   5
the literature on rational inattention started by Sims (2003), which assumes that people
allocate their attention optimally across different pieces of information, incorporating the
costs associated with acquiring and processing information (for recent surveys of the ra-
tional inattention literature, see Handel and Schwartzstein, 2018; Mackowiak et al., 2018).
It should be noted, however, that the central predictions of our model (i.e., those that we
test empirically) arise independently of the assumed structure of information processing
and could therefore also be generated by a simpler model with costly mental effort or time
needed to obtain and process information, not based on entropy reduction.7 Nevertheless,
using a rational inattention approach to model farmers’ demand for advice offers several
advantages. First, in the rational inattention literature agents are allowed to choose not
only the precision of the signals (i.e., the allocated amount of attention), but also the
distribution from which signals are drawn. The structure of information is therefore en-
dogenously determined in these models rather than exogenously imposed, which reduces
the number of required assumptions. Second, the rational inattention literature makes
use of concepts from the field of information theory (most importantly entropy) which are
well-understood and for which a wide range of mathematical tools are readily available.
Finally, there is a growing body of empirical evidence providing support for the rational
inattention approach (Gabaix et al., 2006; Caplin and Dean, 2013; Goecke et al., 2013;
Bartoˇ s et al., 2016; Ambuehl et al., 2018).8
     Until now, applications of rational inattention have focused primarily on rich coun-
tries. By using the same approach to model farmers’ demand for agricultural advice and
extension services, this paper extends the analysis based on rational inattention to the
development context.9 The literature provides several reasons why constraints on infor-
mation processing may be particularly relevant in the context of less-developed economies.
First, people in the developing world (and poor farmers in particular) tend to have less
access to information in pre-processed forms than people in richer countries, e.g., because
of limited access to media and tools such as online search engines. Second, the poor are
often unable to benefit from distraction-saving goods and services, such as stable electric-
   7
     Note that what is termed “attention” in the literature (and in our model) usually comprises both
cognitive effort and time use more generally, since both are needed for processing information. As pointed
out by Banerjee and Mullainathan (2008), the relevant aspect of time use in this context should be thought
of as the “quality of time (i.e., mental focus)” rather than the “quantity of time” (p. 493).
   8
     For a more detailed overview of different approaches to attention used in economics, including a
comprehensive discussion of the assumptions underlying the rational inattention approach, see Handel
and Schwartzstein (2018).
   9
     Applications of rational inattention initially focused on macroeconomic contexts such as monetary
transmission (Mankiw and Reis, 2002; Ma´     ckowiak and Wiederholt, 2009), consumption dynamics (Luo,
2008; Tutino, 2013), and business cycles (Ma´     ckowiak and Wiederholt, 2015). Further studies apply
rational inattention in the areas of finance (Van Nieuwerburgh and Veldkamp, 2010; Kacperczyk et al.,
2016), industrial organization (Sallee, 2014; Martin, 2017), and labor (Bartoˇ  s et al., 2016; Acharya and
Wee, 2020). In the development context, some existing studies investigate the role of other (i.e., not
entropy-based) forms of attention (Banerjee and Mullainathan, 2008; Beaman et al., 2014; Hanna et al.,
2014).



                                                    6
ity and water supply, and thus have to dedicate more time and mental effort to everyday
life problems (Banerjee and Mullainathan, 2008; Bick et al., 2018). Finally, an increasing
body of evidence suggests that there exists a direct adverse effect of poverty on cognitive
functioning, because poverty-related concerns can induce stress and thereby deteriorate
available mental resources (Mani et al., 2013; Haushofer and Fehr, 2014). All of these
factors tend to increase the mental costs for the poor to make well-informed choices and
thus exacerbate the severity of limitations in cognitive capacity which all humans face.
      Finally, the paper relates to the literature that studies ex-post strategies to deal with
shocks, e.g., through increasing labor supply (e.g., Kochar, 1995; Jayachandran, 2006),
taking children out of school (Jensen, 2000), or relying on informal insurance networks
(Udry, 1995; Dercon, 2002). We show that seeking advice or participating in available
training opportunities is one possible way to (at least partially) compensate for incurred
welfare losses due to adverse shocks.

Background on extension services. Agricultural extension is commonly seen as a
key component in increasing productivity and triggering sustainable economic growth in
developing regions around the world. Many policy-related studies particularly emphasize
the role that advisory services can play in reaching marginalized farmers, reducing food
insecurity, and breaking patterns of persistent rural poverty (Chipeta, 2006; World Bank,
2008). In addition, agricultural extension is often perceived as an important instrument to
address new challenges related to environmental degradation and climate change (IFAD,
2013; FAO, 2014).
      The economic literature finds mixed evidence of returns of extension services.10 It
is, in general, not clear whether this is due to variation in the studied programs or to
methodological challenges associated with evaluating programs in the absence of exoge-
nous variation (Birkhaeuser et al., 1991; Anderson and Feder, 2007). Traditionally, most
often advice provided to farmers takes the form of field visits by extension staff or other
local delivery agents to farm households. Various studies find that these forms of advice
can indeed have significant effects. For example, Dercon et al. (2009) find that receiving
at least one extension visit reduces headcount poverty by 9.8 percentage points and in-
creases consumption growth by 7.1 percentage points in rural Ethiopia. Bandiera et al.
(2018) study a program in Uganda that involves local farmers as delivery agents. Farmers
in treatment villages grow 17% more marketable crops, have 40% higher profits, and per
capita consumption expenditure is 22% higher. Other modern approaches to provide ad-
vice employ information and communication technology (ICT), such as agricultural apps,
consulting hotlines, and SMS-based reminders (Aker, 2011; Cole and Fernando, 2013;
  10
    For example, in the survey by Birkhaeuser et al. (1991) 36 of 48 reviewed studies find positive and
significant impacts of extension services.




                                                  7
Casaburi et al., 2014).11 More efforts to increase farmers’ access to agricultural advice are
currently under way, e.g., the ‘Netflix for Agriculture’ initiative by Fabregas et al. (2017).
     According to Anderson and Feder (2007), 80% of the world’s extension services are
publicly funded and delivered by civil servants. While in the past such programs have
predominantly been characterized by top-down and supply-driven approaches, the focus
has shifted in recent years toward making extension more demand driven (Anderson,
2007; Davis, 2008). Several factors have contributed to this development. First, the
collapse of the Training and Visit (T&V) system12 in the late 1990s has led to the rise of
a more pluralistic model of providing and financing extension services, involving stronger
decentralization, privatization, and involvement of NGOs and farmer-based organizations.
This process has shifted the focus to the demand side of extension services by emphasizing
the importance of increasing farmers’ voice and participation as compared to traditional
top-down approaches (Rivera and Alex, 2005; Birner and Anderson, 2007).13 Second, the
transformation in the agricultural extension sector towards demand-driven approaches has
been linked to a more general paradigm shift in public sector reform toward responsive
governance, which advocates accountability and empowerment to increase the effectiveness
of public service provision (including in many other sectors such as health and education;
see United Nations, 2005; Birner and Anderson, 2007). Finally, the rapid diffusion of
information and communication technologies has contributed to the promotion of demand-
driven approaches to extension, by reducing the cost and providing new possibilities for
farmers to access agricultural advice based on their own individual-specific needs.


3      Theory: Demand for agricultural advice under costly
       attention
This section presents a stylized model that is able to explain (i) why farmers may rationally
decide not to participate in extension services, and (ii) why non-agricultural shocks can
increase farmers’ demand for agricultural advice and participation in extension services.
The model builds upon the following three main assumptions. First, farmers have access
to sources of agricultural advice (e.g., extension services), and believe that engaging with
these sources has a positive effect on production. Second, requesting and benefiting from
  11
     Examples of such technologies include mobile phone apps and hotlines that offer farmers the possi-
bility to acquire information about farm practices, weather, and relevant prices, as well as online systems
where farmers can send photos or show crops affected by diseases to a web camera in order to receive
advice on treatment. A survey of ICT-based agricultural extension programs is provided in Aker (2011).
  12
     The T&V system was a public extension model promoted by the World Bank from 1975 until 1995 to
increase the adoption of “Green Revolution” technologies (mainly high-yielding seed varieties, fertilizer,
and other agrochemicals) in more than 70 countries (Anderson, 2007).
  13
     In part, this may also have been driven by the need to achieve greater scope for cost recovery in
order to facilitate privatization and contracting of extension services, which are core elements of the new
pluralistic model (Anderson, 2007).


                                                    8
advice requires farmers to be attentive. Third, attention is a scarce resource and farmers
have to allocate their attention between different areas of their life, including those not
related to farming activities. Based on these features, we demonstrate how not only
agricultural shocks (such as pests and irregular rainfalls), but also shocks which are not
directly related to agriculture may be linked to farmers’ demand for extension.14


3.1     Model
We study the decision problem of a farmer who faces choice problems in two distinct
domains. The first domain captures decisions in the farmer’s production process, e.g.,
related to crop types, usage of inputs, and other agricultural tasks. We refer to this as
the agricultural domain (or simply A). The second domain captures all other decisions
which require the farmer’s attention, e.g., how to deal with the sickness or death of a
household member. This is referred to as the non-agricultural (N) domain. For each
domain, the farmer’s choices are summarized in a single action, denoted by ai ∈ R, with
i ∈ {A, N }. Let the optimal action in each domain be given by

                                              a∗
                                               i = φzi ,                                           (1)

                                           2
where the random variable zi ∼ N (0, σz     i
                                              ) describes the fundamental state of the economy
which is initially unobserved by the farmer.15 For example, one may think of zA as cap-
turing all relevant factors that determine the returns to different agricultural production
choices, including current prices, demand for certain crops, and soil conditions. Let u∗      i
                                                                                    ∗
denote the maximal payoff for domain i associated with the optimal action ai . Choosing
                                                                        ∗
a suboptimal action, ai = a∗ i , leads to a payoff that is smaller than ui . For example, in the
agricultural domain, this may be the case if the farmer uses the wrong type of pesticide
or a suboptimal quantity of fertilizer given the prevailing conditions (as captured by the
realization of zA ).
     In addition, we allow for exogenous events, denoted by ϑi , which directly reduce the
payoff in each domain. For example, ϑA may capture harvest loss due to bad weather, and
ϑN may capture theft of money. In contrast to zi , these shocks cannot be compensated
for by the farmer through selecting appropriate actions. Instead, any realization of ϑi
unequal to zero induces a direct shift in the farmer’s level of utility. The highest possible
payoff in each domain is thus given by

                                           u∗    ∗
                                            i = ui + ϑi ,                                          (2)
   14
      The model heavily draws on existing work in the rational inattention literature, particularly the
framework presented by Ma´   ckowiak and Wiederholt (2009) and Wiederholt (2010). Where possible, we
follow the notation used by these authors.
   15
      To keep the analysis tractable, we focus on the case where fundamentals are independent across
domains, i.e., nothing can be learned about a fundamental by paying attention to the other fundamental.


                                                  9
which we write as u∗  i (ϑ) to simplify notation. The farmer seeks to select the optimal
action in both domains, maximizing a concave utility function over the sum of payoffs
received from aA and aN .
     To make better decisions, the farmer can devote attention to available sources of infor-
mation and thereby reduce uncertainty about the realized states zi and associated optimal
actions. For example, in the agricultural domain, this may correspond to the farmer con-
tacting an extension officer, participating in a field school, or asking a neighboring farmer
for advice. In modeling attention, we follow the literature on rational inattention started
by Sims (2003) and quantify attention as reduction in Shannon entropy, a measure of the
unpredictability of a random variable’s realization.16 Paying attention can be modeled as
receiving a noisy signal
                                         si = zi + i                                      (3)

about the realization of zi , where i ∼ N (0, σ 2i ).17 In this case, the prior entropy of the
random variable zi and the posterior entropy that results from paying attention to the
fundamental are defined as

                             1           2                               1           2
                 H (zi ) =     log2 (2πeσz   )   and     H (zi |si ) =     log2 (2πeσz i | si
                                                                                              ).        (4)
                             2             i
                                                                         2

Based on these two equations, the amount of attention, κi , devoted to the choice problem
associated with each domain can be quantified as

                                        κi = H (zi ) − H (zi |si ).                                     (5)

This equation states that the more attention is devoted to domain i, the better the
quality of the acquired signal, and thus the larger the reduction in uncertainty (measured
in entropy) about the realized state zi .18
     Capturing the idea that attention is a scarce resource, the farmer faces in each domain
the cost µi κi , where µi > 0 denotes the unit cost of attention (e.g., opportunity cost of
mental energy). Note that µi may differ across domains, because paying attention to one
decision may be more demanding than paying attention to another decision, e.g., due to
experience or education.
  16
     It has been pointed out in the literature that the concept of rational inattention relies on relatively
strong assumptions about people’s ability to focus their attention on those pieces of information that
are most worth attending to (Handel and Schwartzstein, 2018; Kremer et al., 2019). On the other hand,
a growing body of empirical evidence supports the rational inattention approach (Gabaix et al., 2006;
Caplin and Dean, 2013; Goecke et al., 2013; Bartoˇ  s et al., 2016; Ambuehl et al., 2018).
  17
     As shown by Sims (2003), this represents the signal structure that an agent who can freely set the
distribution of the signal would choose in the case of a quadratic objective function and normal priors.
Normality of the signals is therefore an outcome of the model itself rather than an additional assumption.
  18
     Notice that this differs from standard models in economics based on rational expectations, which
assume that agents are able to be perfectly attentive to all available information (i.e., agents can process
information instantaneously and without any additional cost).



                                                    10
     Overall, the decision problem of the farmer consists of choosing how much attention
to devote to each of the two domains, and which actions to perform conditional on the
received signals. The timing of the model is such that the farmer first observes the
realizations of the shocks ϑi , then chooses the allocation of attention, receives the signals,
and finally selects the actions.
     Formally, this can be described as a decision problem with two stages. In the second
stage, the farmer chooses the actions with the highest expected payoff given the received
signals:
                           max       E [u (aA , aN , zA , zN , ϑ) |sA , sN ] ,             (6)
                          aA ,aN ∈R

                          ˆi denote the optimal action for domain i ∈ {A, N } chosen by
where ϑ = (ϑA + ϑN ). Let a
the farmer, i.e.
                            ˆi = arg max E [u (ai , zi , ϑ) |si ] .
                            a                                                       (7)
                                        ai ∈R

In the first stage, the farmer observes the realization of the shocks ϑi and chooses the
allocation of attention. In doing so, the farmer maximizes the expected utility resulting
from the action chosen in the second stage less the cost of attention. Formally, in the first
stage, the farmer solves
                             max E [u (ˆ  ai , zi , ϑ)] − µi κi ,                        (8)
                               κi ≥0

subject to the constraint on information processing (5) and equation (7). The farmer
anticipates the actions chosen in the second stage and the associated expected payoffs
when deciding how much attention to devote to each domain.


3.2    Solution and predictions
In deriving the farmer’s optimal behavior, we focus on the case of a strictly concave utility
function under quadratic approximation. This is a common approach in the literature on
                              ckowiak and Wiederholt, 2009; Gabaix, 2014). The solution
rational inattention (e.g., Ma´
to the farmer’s optimization problem is formally derived in Appendix A. Based on the
resulting optimal allocation of attention, the model gives rise to a number of predictions
which are summarized in the following two propositions.

Proposition 1. A rationally inattentive farmer will pay more attention to the agricultural
domain and thus be more likely to request advice, (i) the more costly are mistakes in aA ,
(ii) the larger the prior uncertainty about the optimal action a∗
                                                                A , and (iii) the smaller the
cost of paying attention to the fundamental zA .

Proof. See Appendix A.

    According to the results in Proposition 1, factors which make it easier for farmers
to pay attention to agricultural advice, or which cause suboptimal cultivation practices


                                                11
to be relatively more costly, will make it more likely that farmers request advice. This
suggests that we should expect household characteristics which affect the costliness of
paying attention to advice to be correlated with farmers’ demand for extension services.
For example, characteristics such as high age and low education are often argued to be
positively related to individuals’ attentional costs (Greenwood and Parasuraman, 1991;
Ambuehl et al., 2018). In addition, one might expect demand for agricultural advice to
be positively correlated with household size, given that households with fewer members
feature less resources (such as time and cognitive capacity) and are thus more constrained
in their capacity to participate in extension programs (other factors equal).
     Notice that the results in Proposition 1 are exclusively based on parameters associ-
ated with the agricultural domain itself. In addition, the overarching utility function in
the farmer’s objective, which connects both domains through the channel of imperfect at-
tention, gives rise to cross-domain effects, i.e., effects on farmers’ demand for agricultural
advice arising from parameters of the non-agricultural domain. This feature of the model
is summarized in Proposition 2.
Proposition 2. If attention is costly (i.e., µi > 0), negative shocks in any of the two
domains will cause the farmer to increase the amount of attention allocated to aA , and
thus raise the probability that agricultural advice is requested.

Proof. See Appendix A.
     The mechanism behind the results in Proposition 2 is as follows. In equilibrium, the
farmer allocates the amount of attention to each domain for which the marginal cost of
attention (captured by the parameter µi ) equals the marginal return (i.e., the increase
in expected utility that results from a better signal due to more attention). A negative
shock to the farmer’s income or wealth lowers the absolute level of utility, which, under a
concave utility function, leads to a larger marginal utility of income from farm activity.
Therefore, the farmer is more willing to devote costly attention to advice that can help to
raise productivity. Hence, the model predicts that farmers who are negatively affected by
shocks will be more willing to devote attention to sources of agricultural advice, and thus
be more likely to actively demand extension services. Importantly, this holds for both
agricultural shocks and shocks that are not directly related to farming.19


3.3     Discussion
In the case of agricultural shocks, the predicted effect in Proposition 2 seems relatively
unsurprising, and could in principle also be generated by other existing models. This
  19
    Notice that, because the model is symmetric in the two domains, a negative shock to income in any
domain will also increase the amount of attention paid to aN . Since we are mainly interested in studying
the determinants of farmers’ demand for agricultural extension, the empirical part below focuses on
the cross-domain effect described in Proposition 2. Note also, that the model makes no assumption on
whether the shocks ϑi are correlated across the two domains.


                                                   12
applies less to shocks which are not directly related to agriculture. In the large body
of literature that uses learning models to explain agricultural production decisions (e.g.,
Foster and Rosenzweig, 1995; Munshi, 2004; Conley and Udry, 2010), farmers are typically
perceived as facing an initial lack of knowledge about optimal input targets, which can be
progressively overcome through learning (based either on own experimentation or learning
from others). In such an environment, agricultural shocks that affect optimal input targets
(e.g., pest outbreak or change in the prices of agricultural inputs) will create a renewed
need for farmers to learn, and thus tend to increase the demand for advice. However,
these models would predict no such effect for shocks that are unrelated to farming, i.e.,
shocks that do not affect optimal cultivation practices.
      Other studies in this context stress the importance of constraints resulting from
prevailing market imperfections in developing countries, but perceive farmers as rational
decision makers under full information (Schultz, 1964; Bardhan and Udry, 1999). An
important insight of these studies is that farmers’ optimal production choices may not
be perfectly separable from consumption choices (as would be the case in the absence
of market imperfections). This non-separability implies that optimal decisions in the
agricultural domain may also be affected by shocks originating in the non-agricultural
domain. For example, this would be the case if shocks to households’ off-farm income
also lead to changes in the optimal level of modern agricultural input use (which would
not happen if farmers had access to perfect credit markets), or if optimal farm decisions
are sensitive to shocks that affect the availability of family labor (which would not be the
case if farmers could hire from perfect labor markets).
      However, it is important to note that, due to their strong assumptions on rational
decision making, these models face difficulties in explaining demand for advisory services
in general. In particular, it remains unclear from these models why, in the absence of
any information frictions, there would be a need for agricultural advice in the first place.
Furthermore, assuming that there was a need, these models are unsuited to explain why
farmers in Sub-Saharan Africa are not more widely making use of available extension
services, especially in situations where those services are available free of charge.
      Finally, it should be noted that the prediction in Proposition 2 runs counter to much
of the recent literature on psychology and poverty (e.g., Banerjee and Mullainathan,
2008; Mani et al., 2013; Haushofer and Fehr, 2014), according to which negative shocks
in the non-agricultural domain would tend to absorb cognitive bandwidth, leading to less
attention paid to agricultural production and thus less use of agricultural advice. When
testing the predictions of the model empirically, these effects clearly work against us.
To the extent that we still find evidence in line with Proposition 2, the fact that our




                                            13
model based on rational inattention does not capture these cross-domain effects in terms
of available cognitive bandwidth is therefore less of a concern.20
     In sum, standard modeling approaches, e.g. based on learning or credit constraints,
have difficulties explaining the combination of low levels of demand for extension and
a (positive) reaction of demand for extension to non-agricultural shocks. On the other
hand, our model with costly attention would be consistent with non-agricultural shocks
explaining changes in demand for extension and we verify this prediction of the model
below. Because it is conceivable that non-agricultural shocks are linked to demand for
extension through other channels (e.g., due to non-separability), the empirical part below
also discusses the evidence in light of some alternative theories.


4      Empirical evidence
In this section, we first show that there is statistically strong and robust evidence that, in
line with the model, non-agricultural shocks are indeed associated with increased demand
for extension. In addition, we assess the plausibility of some alternative explanations
which might also account for the observed link. Our model is based on the assumption
that agricultural advice has positive effects on farmers’ production. To complement the
existing evidence, cited in the introduction, we therefore also investigate the relationship
between agricultural advice, farm decisions, and outcomes in our data (see Appendix D).
While our data allow us to improve upon prior literature that also uses observational data
to study the effects of extension services, we stress that the observational nature of the
data limits our ability to make causal claims.


4.1     Data and descriptive statistics
The empirical work is based on survey data from three countries in Sub-Saharan Africa
(Malawi, Nigeria, and Uganda), which were collected as part of the World Bank’s Living
Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA).21 In each
country, households are selected in a nationally representative way and tracked over time,
such that we can use a panel dataset consisting of three rounds of surveys in Malawi
(2010-11, 2013-14, 2016-17) and Nigeria (2010-11, 2012-13, 2015-16), and four rounds in
Uganda (2009-10, 2010-11, 2011-12, 2013-14).22 We focus on households that engage in
  20
     The same argument applies to the concern that farmers’ demand for advice could also be explained
based on a similar model using time rather than attention as the limiting factor, where for some types
of non-agricultural shocks (e.g., illness of a family member) the opportunity cost of time would increase
and thus demand for agricultural advice would tend to decline.
  21
     The same data are used by Deininger et al. (2017), Gollin and Udry (2017), and Sheahan and Barrett
(2017).
  22
     Detailed information about the design and implementation of the surveys in each country can be
found at World Bank (2018). While the LSMS-ISA database covers more countries, only the data from
the three countries we use here allow for separately identifying extension contacts that were actively


                                                   14
agricultural activities and report having cultivated at least one plot in a given season (for
countries with two cropping seasons, i.e., Malawi and Uganda, we focus our analysis on the
main season). Starting with 36,014 household-wave observations in the original dataset,
dropping 9,359 observations that report not having engaged in agricultural activities, and
then dropping 1,381 observations without cultivated plots, leaves us with a maximum of
25,274 household-wave observations, comprising 11,154 households and 50,478 plot-wave
observations. In all of the main regression analyses we work with household fixed effects,
which requires two rounds of data. After dropping 748 observations with missing data
on the use of agricultural advice, 162 observations with missing data on shocks, and 755
observations with missing data on one of the key household level control variables, the
subsample of households that appear in the data for at least two rounds consists of 20,233
household-wave observations.
     The rest of this section describes the construction of the key variables used to test
the model’s predictions. Additional information on the construction of variables and dif-
ferences in survey designs across countries can be found in Appendix B, together with a
complete list of all variables and basic summary statistics (see Table A1).

Extension services. In all three countries, the LSMS-ISA surveys collect detailed in-
formation on whether households received advice related to farming activities, as well as
the number of extension contacts and the sources of advice (e.g., governmental extension
service, NGOs, or other farmers). The dummy variable Advice indicates whether a house-
hold has received advice related to farming at least once over the past 12 months. The
dummy variable Extension considers only advice that was obtained through sources other
than media (i.e., sources such as TV, radio, or flyers). The variable Contacts contains
the number of extension contacts, which includes visits to the household as well as visits
by the household to the source.23
     In addition, the data allow us to separately identify extension contacts that were
actively solicited by farmers from extension services for which the household was merely
a passive recipient. The former is captured by the variable Solicited Contacts, which is
the sum of visits to the household solicited by the farmer, and visits by the household to
the source. The dummy variable Solicited Extension equals one if the household had at
least one such actively demanded contact in the past 12 months.
     Table 1 provides a first impression of farmers’ exposure to extension services in the
studied countries. In this table we focus on the variable Extension, i.e., on dimensions
solicited by farmers from extension services for which households were merely passive recipients. As we
are interested in studying farmers’ own demand for extension, we focus on these three countries.
  23
     It should be noted that while, in all three countries, the survey questions underlying our extension
variables refer exclusively to received advice and information, we cannot exclude the possibility that some
farmers maintain contacts with extension services also in order to receive other forms of support (e.g.,
financial or in-kind support). Note also, that extension services will likely be heterogeneous even within
countries.


                                                    15
                 Table 1: Exposure to agricultural advice in last 12 months

                                                                      Full
Description                                                                Malawi Nigeria Uganda
                                                                   Sample
Households who received agricultural advice (% all HH-year obs.)     28.7     46.5     16.7      26.0
   of those:
   Average number of contacts                                         4.4      4.1      5.4       4.1
   More than 2 contacts (%)                                          50.6     52.8     60.9      41.1
   More than 10 contacts (%)                                          8.5      5.6     12.5       9.1
   Received advice rated as useful or very usefula (%)               84.6     80.0     91.2      87.0
   Received advice rated as useless or bada (%)                       8.0      8.8      3.3       9.9
   Paid in order to receive advice (%)                                7.6      0.9     14.2      13.5
Households who solicited advice (% all HH-year obs.)                 19.5     35.0     12.4      15.3
Households who solicited advice (% HH-year obs. with advice)         76.0     89.3     76.5      60.1
Share of solicited contacts (% all contacts)                         73.6     80.8     64.2      75.8
Observations                                                       24,526    7,222    8,844    8,460

Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: Agricultural advice measures in-person advice and excludes advice received through media (TV,
radio, flyers, etc.). a Exact wording differs between countries (see Appendix B).


of advice that exclude advice obtained through media. On average, 29% of farmers
report having received such advice. In Malawi (which generates 32% of its GDP in the
agricultural sector and is known for its relatively strong policy focus on agriculture, see
OECD/FAO, 2016), the number is 46%. Conditional on receiving advice, farmers have an
average of 4 to 5 contacts a year. While 51% of farmers receiving advice report more than
two contacts a year, only a few farmers seem to be involved in regular extension programs
(such as monthly village or field school meetings). The majority of farmers (85%) that
receive advice rate the received advice as useful, and less than 10% of farmers report
having received advice that was useless. On average, less than 8% of households that
received advice paid anything for it.24 In addition, Table 1 indicates that in all countries
the majority of extension contacts are solicited by farmers. The share of solicited contacts
ranges between 64% in Nigeria to over 80% in Malawi (possibly reflecting the fact that
Malawi features a relatively decentralized and demand-driven extension landscape; see
Davis, 2008). More details are provided in Table A2 in the appendix.
     Table 2 provides additional insights into the topics on which farmers receive advice.
Almost all farmers (91%) who have been in contact with extension services received ad-
vice on agricultural production and processing. To a large extent, this involves advice
about modern agricultural inputs such as new seed varieties, fertilizer, and pesticides. In
addition, a third of farmers report having received advice on livestock production and
animal care. With respect to different sources of agricultural advice, Table A3 in the
appendix shows that government extension services are responsible for 79 to 94% of ex-
tension provided in Uganda and between 32 and 49% in Malawi. In Nigeria, where public
  24
    Conditional on paying for advice, the average amounts paid are USD 6 in Uganda, USD 9 in Nigeria,
and USD 18 in Malawi.


                                                 16
                                         Table 2: Topics of advice

Descriptiona                                     Full Sample       Malawi         Nigeria         Uganda
Agricultural production and processing                    91.3        92.6            80.1           96.3
    New seed varieties                                    50.7        51.3            49.2
    Fertilizer                                            52.6        50.0            58.9
    Pest control                                          29.0        28.3            30.8
    Composting/manure                                     41.8        50.3            21.0
    Irrigation                                            26.5        34.5             6.9
Marketing and crop sales                                  25.9        17.2            24.6           39.9
Livestock production                                      32.8        20.2            26.7           55.8
Animal diseases and vaccination                           30.5        18.8            28.1           50.0
Fishery                                                    8.1         7.9             1.9           12.4
Forestry                                                  15.7        21.3             2.3
Access to credit                                          12.8        15.8             5.5
Other                                                      2.1         2.8             0.5
Observations (HHs who received advice)                   6,877       3,336           1,369          2,172

Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: Numbers are percentages of households who received advice on indicated topic, conditional on
having received in-person advice (i.e., excluding advice received through media). a Exact wording differs
between countries.


extension programs seem to be less prevalent, farmers report receiving advice relatively
more often from private extension services and from other farmers.

Shocks. Data on shocks are available at the household level and include detailed infor-
mation on whether households have been negatively affected by any shocks over the past
year and the type of each shock.25 The latter is collected based on a list of around 20
different events, which are very similar across countries. To test the predictions of the
model, we classify shocks as either directly related to agriculture (AgShock ) or not directly
related to agriculture (NonAgShock ). Table 3 provides a complete list of the shock items
that appear in the data and the assigned categories for our baseline regressions.26 As can
be seen in the table, agricultural shocks include events such as “irregular rains”, “un-
usual high level of crop pests”, and “livestock disease”. Non-agricultural shocks include
events such as “illness of a household member”, “theft of money”, and “end of regular
assistance/remittances”.27 Of course, any classification of the shock items listed in Table
3 into two distinct groups related to agriculture and not directly related to agriculture
  25
     As described above, there are several years between panel rounds, and these are inconsistently spaced
over time and countries. Therefore, we do not work with lagged shocks, as these are (i) several years,
and therefore, in our view, too far in the past, and (ii) irregularly spaced.
  26
     The shock items “other” and “increase in the prices for food” are not classified in our baseline
regressions since we found it difficult to attribute them to either of the two shock variables. As shown
below, our results are robust to including “increase in the prices for food” in AgShock or in NonAgShock.
  27
     As can be seen in Table 3, the extent to which households report experiencing negative shocks
differs considerably between countries. For example, in Nigeria only 12% of farmers report having been
negatively affected by agricultural shocks (on average during one year), while in Malawi and Uganda the
corresponding percentages are 82% and 41%, respectively.


                                                    17
                                                              Table 3: Shocks to households by country

                                                                     Percentage of household-wave obs.                         Percentage of households with
     Descriptiona
                                                                  with shock (on average during one year)                transition in shock (across survey rounds)
                                                      Malawi            Nigeria         Uganda          Full Sampleb    Malawi Nigeria Uganda Full Sampleb

     Agricultural shock (AgShock )                      82.1             12.0            40.6               43.3          20.5     24.3     43.2         28.9
     Drought/irregular rains                            60.6              2.8            35.9               31.9          34.4      6.2     42.3         28.3
     Floods/landslides                                  11.4              5.5             4.8                7.0          11.3     11.5     8.5          10.5
     Crop pests or disease                              14.4              0.8             3.2                5.8          15.9      2.0      6.4          8.7
     Livestock disease                                  14.8              1.8            1.7                5.7           16.3     4.5      3.7           8.7
     Increase in price of agricultural inputs           60.3              2.2             1.7               19.6          36.2      5.4     3.3          16.4
     Fall in price of agricultural output               26.8              0.7             1.6                8.9          27.1      1.7     3.4          11.8
     Non-agricultural shock (N onAgShock )              40.1             14.1            15.7               22.5          32.5     27.3     26.0         28.9
     Health
     Illness or accident of household member            16.1             2.6              8.1                8.6          17.0     6.1      15.8         13.3
     Death or disability of household member            9.4              3.7              3.1                5.2          9.9      8.5      5.8          8.2




18
     Income
     Loss of employment                                  4.5             0.2              0.1                1.5          3.1      0.4       0.2          1.4
     Reduction in non-farm income                       10.0             2.2              0.3                3.9          9.0      4.8       0.8          5.1
     End of assistance/remittances                       8.9             1.9                                 5.2          9.1      4.6                    7.0
     Crime
     Theft of money/valuables                           10.5             2.2              4.1                5.4          10.2     5.1       8.0          8.0
     Conflict/violence                                   7.6              0.5              1.1                2.9          7.1      1.1       2.6          3.8
     Other
     Fire/earthquake/other damage                       6.6              2.4              0.8                3.1          6.6      5.8       1.6          4.8
     Not classified
     Increase in prices for food                        59.7             8.2                                32.2          39.3     18.0                  29.8
     Other (not specified)                               11.4             1.2              2.5                4.7          11.6      2.8      5.2          6.9
     Observations                                      7,517            8,578           8,686              24,781        4,151    3,340     3,487       10,978

     Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: a Exact wording differs between countries. b The full sample is obtained by pooling data across countries and waves.
will be imperfect. We therefore explore several alternative classifications as part of the
robustness tests.

Household characteristics. We control for unobserved factors that are fixed at the
household level via the inclusion of household fixed effects. In addition, we control for
time-varying characteristics, including household size and composition as well as educa-
tional level, age, and gender of the household head.


4.2    Testing the model’s predictions
The regression model we estimate is given by

      Extensionit = α0 + α1 AgShock it + α2 N onAgShock it + α3 Xit + µi + ωt +    it ,   (9)

where the dependent variable is a measure of farmers’ exposure to extension services
(in some specifications, we restrict the dependent variable to services that were actively
requested by farmers). AgShock it is a dummy variable that captures shocks which affect
optimal cultivation practices and are therefore directly related to farmers’ need for advice
(e.g., to inquire about optimal pest control after a pest outbreak). N onAgShock it is
a dummy that corresponds to shocks which have a negative effect on farmers’ income
or wealth, but do not constitute an immediate reason to request advice about farming
activities (e.g., theft of money or non-farm business failure). The terms µi and ωt denote
household and wave fixed effects. Finally, Xit is a vector of time-varying household-level
variables.
     The hypothesis we test is that both coefficients α1 and α2 in equation (9) are positive
(see Proposition 2). In addition, the model predicts that characteristics that make it
easier for farmers to pay attention to agricultural advice, or factors that cause suboptimal
cultivation practices to be relatively more costly, will make it more likely that farmers
request advice (Proposition 1). Therefore, we expect household characteristics such as
education and household size to be correlated with farmers’ demand for extension (with
the directions discussed in Section 3).
     Our empirical strategy relies on the comparability of households that receive non-
agricultural shocks and those that do not receive shocks. Table A4 in the appendix shows
mean characteristics of households that report different types of shocks and characteristics
of households without shocks. The comparison suggests reasonably comparable groups.
In addition, our regressions capture fixed differences between households through the
included household fixed effects.
     Table 4 reports estimates of the regression model specified in equation (9) for various
measures of farmers’ exposure to agricultural advice. In the first three columns, the
dependent variable is an indicator that equals one if in a given season the household

                                            19
received any advice related to farming (i.e., from any of the sources listed in Table A3).
The dependent variable in columns (4) to (6) excludes advice that is obtained through
media (TV, radio, flyers, etc.). In columns (7) to (9), the indicator for extension is further
restricted to solicited advice, i.e., the dummy is only equal to one if the household actively
requested advice.
     The results in Table 4 show that, irrespective of whether a set of household controls
is included or not, both agricultural shocks and non-agricultural shocks are positively
correlated with receiving agricultural advice (full sets of dummies are included as controls
in columns 3, 6, and 9). In all specifications, the estimated coefficients of the two shock
indicators are statistically highly significant. As discussed in more detail below, the results
are robust to different ways of clustering standard errors and to alternative definitions
of the considered shock and extension variables. We thus interpret the findings in Table
4 as being in line with the predictions of the model (as described in Proposition 2) and
providing empirical support for the suggested limited attention channel.
     Although the results in Table 4 are in line with the predictions of our model, we also
note that they imply that the quantitative magnitude of the link between non-agricultural
shocks and demand for advice is modest. More specifically, the estimated coefficients of
NonAgShock in columns (7) to (9) suggest that one out of 40 (or 2.5%) of farmers actively
seeks agricultural advice in response to a non-agricultural shock. While these estimates
are consistent with attentional constraints being part of the low demand for extension,
they also highlight that costly attention is only one of several pieces of the puzzle to
explain the low demand for extension.
     Furthermore, the results in columns (2), (5), and (8) of Table 4 show that farmers’
decisions to request advice and participate in extension programs are positively corre-
lated with household size. Of course, we cannot interpret this finding as causal, but the
positive correlation is consistent with the model. From the perspective of the model,
this finding could be attributed to the fact that additional household members represent
additional resources (such as time and cognitive capacity), which are useful in obtaining
and processing information. Thus, when households that are similar along other charac-
teristics are compared, those with more members will find it relatively easier to request
and implement agricultural advice. In contrast, households with fewer members may be
more constrained in their capacity to participate in extension programs, which will tend
to reduce the number of contacts and lead to a lower probability of requesting advice.28
  28
    A similar line of reasoning is used by Gido et al. (2015) to rationalize the finding of a negative
link between off-farm income and extension contacts in Kenya. Regarding the positive coefficient of the
number of children, notice that the mechanism would only apply to children old enough to participate
in working on the farm, thus representing additional resources rather than a liability (which, among the
rural households in our sample, might start at a fairly young age). However, for very young children
there might be another explanation in line with our model. A newborn baby might have a similar effect
as a negative shock to household income, i.e., marginal utility of additional income increases as part of
the income is diverted to the baby.


                                                   20
                                                                Table 4: Test of model predictions

                                                     Advice                                     Extension                                Solicited Extension
                                         (1)          (2)             (3)            (4)           (5)            (6)            (7)              (8)            (9)
                                           ∗∗∗            ∗∗∗            ∗∗∗           ∗∗∗            ∗∗∗            ∗∗∗           ∗∗∗              ∗∗∗
     AgShock                          0.094          0.089          0.087         0.079          0.074         0.072          0.065            0.062           0.061∗∗∗
                                       (0.005)        (0.005)        (0.005)       (0.002)        (0.002)       (0.002)        (0.001)          (0.001)         (0.001)
     NonAgShock                       0.033∗∗         0.033∗∗       0.033∗∗       0.031∗∗        0.032∗∗       0.030∗∗        0.025∗∗∗         0.026∗∗∗        0.025∗∗∗
                                       (0.030)        (0.024)        (0.022)       (0.030)        (0.025)       (0.029)        (0.008)          (0.007)         (0.007)
     Adults                                          0.020∗∗∗                                    0.022∗∗∗                                      0.020∗∗∗
                                                      (0.002)                                     (0.001)                                       (0.000)
     Children                                        0.013∗∗∗                                    0.010∗∗∗                                      0.010∗∗∗
                                                      (0.000)                                     (0.003)                                       (0.001)
     Primary educ. (head)                               0.012                                      0.015                                         0.003
                                                      (0.576)                                     (0.374)                                       (0.895)
     Secondary educ. (head)                            -0.008                                      -0.003                                        0.001
                                                      (0.767)                                     (0.897)                                       (0.980)
     Age head                                         0.012∗∗                                       0.009                                       0.008∗




21
                                                      (0.015)                                     (0.106)                                       (0.078)
     Age head sq.                                    -0.000∗∗                                      -0.000                                      -0.000∗
                                                      (0.015)                                     (0.130)                                       (0.070)
     Male head                                         0.027                                        0.012                                        0.013
                                                      (0.363)                                     (0.666)                                       (0.492)
     HH FE                               yes           yes            yes            yes           yes            yes            yes             yes             yes
     Wave FE                             yes           yes            yes            yes           yes            yes            yes             yes             yes
     Full set of control dummies                                      yes                                         yes                                            yes
     Observations                      20,392         20,392        20,392         20,392         20,392        20,392         19,653           19,653         19,653
     R-squared (within)                 0.024         0.030          0.039          0.020         0.024          0.033          0.011           0.017           0.026
     R-squared (total)                 0.074          0.031         0.031          0.049          0.025         0.026          0.037            0.024          0.027
     Mean dep. var.                     0.335          0.335         0.335          0.285          0.285         0.285          0.197            0.197          0.197

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined as
     region×rural). The omitted category for education is “no schooling”. Full sets of control dummies are included for the following variables: adults, children,
     education of HH head, age of head, male head.
While there seems to be a positive relationship between age of the household head and
access to extension services, we do not find significant results for gender and education
of the household head, nor for the highest educational level of any household member.29
The latter may not be surprising, given that we control for household fixed effects, which
implies that there is little variation remaining in education variables.
     Our model does not consider heterogeneity in the agricultural shocks. The model
assumes that advice is valuable with any agricultural shock. Yet, empirically, the timing
and the type of shocks likely matters. Shocks early in the season may be more amenable
to action and consequently the value of advice may be higher than in the case of shocks
later in the season. Similarly, some shocks may be such that even with advice there is
not much the farmer can do (e.g. if the field is flooded for a large part of the season).
Unfortunately, we do not observe the precise timing of shocks and seeking advice. We
also cannot cleanly identify shocks for which the value of advice would be high. However,
both of these issues will likely work against us, i.e., will make it less likely that we find
a correlation. To the extent that we still find a correlation between shocks and seeking
advice, the inability to distinguish the timing of shocks better is therefore less of a concern.
     One may be worried about the role of individual components of the aggregate non-
agricultural shock measure. Therefore, in Table 5 the dummy variable NonAgShock is split
into individual types of shocks, focusing separately on shocks related to health, income,
and crime. The first column in Table 5 repeats the estimation results from the main
specification (column 9 in Table 4). In column (2), only shocks related to health are used
to construct the non-agricultural shock dummy variable. Columns (3) and (4) repeat the
same approach for shocks related to income and crime. In column (5), all three types
of non-agricultural shocks are included simultaneously. Overall, the results in Table 5
indicate that the link between non-agricultural shocks and farmers’ demand for extension
is statistically robust, even for separate categories of non-agricultural shocks.
     We also perform placebo regressions. Specifically, we repeat the central regressions
in Table 4 with the dependent variable Unsolicited Extension, which equals one for house-
holds that received only advice which they did not actively seek themselves. The idea
is that, according to our theoretical model, unsolicited advice should not be affected by
non-agricultural shocks. The results of these placebo regressions are shown in Table A9 in
Appendix C. According to these results, there is no statistically significant link between
non-agricultural shocks and unsolicited extension services. These findings suggest that it
is not an unobserved variable that drives the correlation between non-agricultural shocks
and extension services generally. Instead, this supports the hypothesis that there is a
causal link from experiencing a non-agricultural shock to actively looking for advice.
  29
     Note that age of the household head does not simply pick up time trends when including household
and wave fixed effects, but also varies over time when the household head changes (e.g., the head dies
and the spouse or one of the children becomes the new head).



                                                 22
                  Table 5: Test of model predictions: Disaggregated shocks

                                    (1)             (2)            (3)             (4)            (5)
                                 Solicited       Solicited      Solicited       Solicited      Solicited
                                 Extension       Extension      Extension       Extension      Extension
AgShock                           0.061∗∗∗        0.063∗∗∗       0.061∗∗∗        0.062∗∗∗       0.060∗∗∗
                                   (0.001)         (0.001)        (0.001)         (0.001)        (0.001)
NonAgShock                        0.025∗∗∗
                                   (0.007)
NonAgShock (Health only)                          0.022∗∗                                        0.017∗
                                                  (0.030)                                        (0.066)
NonAgShock (Income only)                                         0.055∗∗∗                        0.049∗∗
                                                                  (0.008)                        (0.013)
NonAgShock (Crime only)                                                          0.046∗∗         0.041∗∗
                                                                                 (0.020)         (0.041)
HH FE                                yes            yes             yes            yes             yes
Wave FE                              yes            yes             yes            yes             yes
Controls                             yes            yes             yes            yes             yes
Observations                       19,653          19,653         19,653          19,653         19,653
R-squared (within)                  0.026           0.025          0.026           0.026         0.027

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at
the stratum level (consisting of 24 clusters defined as region×rural). Included shock items are as follows:
Health (“Illness or accident of HH member”, “Death or disability of HH member”), Income (“Loss of
employment”, “Reduction in non-farm income”, “End of assistance/remittances”), Crime (“Theft of
money/valuables”, “Conflict/violence”). As controls, full sets of dummies are included for the following
variables: adults, children, education of HH head, age of head, male head.


     Additional robustness checks with respect to clustering of standard errors, alternative
variable definitions, and exclusion of household fixed effects are reported in Appendix C.
As shown in Table A5, the model predictions hold when standard errors are clustered
along both stratum and wave dimensions (“2-way”), when clustered at the district level
(there are 158 districts), as well as for the cluster wild bootstrap procedure described in
Cameron et al. (2008) to take into account the small number of clusters when clustering at
the “stratum” level. Table A6 tests the robustness of the model predictions to alternative
variable definitions. In columns (1) and (4), the extension indicators exclude advice which
was provided by other farmers, because this constitutes a somewhat different approach
to extension. In columns (2) and (5), the non-agricultural shock variable also contains
the shock item “increase in prices for food”, which was previously not classified. In
columns (3) and (6), the food price shock item is instead included in the agricultural shock
variable. For all specifications, the shock variables remain significant. Table A7 shows
that the model predictions also hold when no household fixed effects and no household-
level controls are included (see columns 1-4; columns differ in their way of dealing with
clustering), and when only the household-level controls are included but not the household
fixed effects (columns 5-8). In particular, comparing the results in Table A7 with those
in Table 4 shows that adding household fixed effects changes the point estimates of the

                                                   23
two shock variables but does not change the significance levels much (note that while
Table A7 reports results only for the dependent variable Advice, the findings are similar
when we repeat the same procedure for Extension and Solicited Extension as dependent
variables).


4.3    Discussion of alternative interpretations
One concern in assessing the predictions of the model is that non-agricultural shocks
may also affect farmers’ demand for extension by creating a need for advice on how to
best operate under reduced availability of resources for purchasing farm inputs. This
concern is particularly relevant given that the agricultural and non-agricultural domains
are likely not separable in the context we study (Benjamin, 1992; LaFave and Thomas,
2016). For example, it might be the case that adverse shocks to households’ wealth or
non-farm income lead to changes in the optimal level of modern agricultural input use
(if farmers have imperfect access to credit), or that optimal farm decisions are sensitive
to shocks that affect the availability of family labor (see Section 3). In addition, farmers
might contact extension workers after a non-agricultural shock simply because they hope
to receive monetary or in-kind support rather than agricultural advice.
      While our data do not allow us to completely rule out the possibility that some
farmers solicit extension services for other reasons than requesting advice (see also foot-
note 23), we are able to obtain some useful insights by studying the link between shocks
and farmers’ demand for extension using disaggregated advice topics. In Table A8 in
Appendix C, the considered dependent variables capture specific topics for which farm-
ers demand advice (i.e., the dependent variables are only equal to one if the household
actively requested advice of a particular category). While agricultural shocks show signif-
icant coefficients for all topics except “fishery”, the non-agricultural shock variable is only
significant for ‘agricultural production and processing’ and ‘animal diseases and vaccina-
tion’. In particular, this suggests that the link between shocks and farmers’ demand for
extension is not due to an increased need for advice related to credit (see the insignificant
results in column 7).
      One might also wonder whether the link between incurred shocks and demand for
extension may be due to reverse causality, i.e., an effect of received advice on the occur-
rence (or at least the perception) of shocks. While in the case of agricultural shocks we
believe this might be a reasonable concern, it seems unlikely that this would also apply to
non-agricultural shocks (which constitute the main prediction of the model that we test).
      Another concern might be that farmers do not request extension services if these
are believed to have small (or maybe negative) effects on profits. In addition to the
evidence cited in the introduction, we therefore also perform an econometric analysis of the
association between extension services and farm-level outcomes based on our data. The


                                            24
panel structure of the LSMS-ISA data allows us to make some progress in addressing the
endogeneity problems facing existing related studies that are based on observational data
(see the survey papers provided by Birkhaeuser et al., 1991; Evenson, 2001), in particular
through our ability to include farm fixed effects and proxies for agricultural shocks. Yet,
obvious endogeneity concerns remain, due to the possibility that time-varying shocks –
other than those that we can control for – affect both demand for extension and outcomes,
and the results should be interpreted as suggestive correlations in the data. Results are
reported in Appendix D. We find that access to extension services is positively associated
with a number of relevant farm outcomes. This particularly holds for agricultural advisory
services that are provided at farmers’ own request. The quantitative magnitudes of our
estimates imply that receiving agricultural advice is associated with a 7% larger value of
harvest and 8% higher profits (defined as value of harvest net of costs for seeds, fertilizer,
and agrochemicals).30 Thus, as long as farmers’ beliefs are consistent with the outcomes
observed in the data, there is no reason to assume that farmers abstain from requesting
extension services because they believe that the provided advice is not useful (also recall
that more than 80% of farmers that received advice rate it as useful or very useful in our
sample; see Table 1).
     In addition to the ability to distinguish between solicited and unsolicited advice, the
data also allow an analysis of potential channels through which extension might affect
farm outcomes. We find a significant link between extension services and use of modern
farm inputs (mainly fertilizer and improved seeds), which are commonly argued to be
important determinants of agricultural productivity (Evenson and Gollin, 2003; Morris
et al., 2007; Duflo et al., 2011).
     Another concern might be that the observed correlation between shocks and received
extension services may be driven by supply-side effects rather than the proposed mech-
anism through increases in farmers’ demand for advice. For example, this might be the
case if access to extension services was affected by limited supply (e.g., because these
services are often understaffed), and the supply of extension is partly in response to ad-
verse shocks. While to some extent this could be captured through fixed effects in the
analysis, our data do not allow us to fully rule out the possibility that (part of) the ob-
served correlation between shocks and received extension services is due to such supply
side effects.
     Finally, non-agricultural shocks may affect farmers’ demand for extension by creating
a need for advice on optimal farm practices with respect to labor inputs. Yet, in our
analysis of the relation between receiving advice and farmers’ production choices, we do
  30
    We stress again that due to the endogeneity concerns stated above, we cannot interpret the estimates
as causal. Note further that irrespective of these quantitative estimates, the mere observation of a
positive association between shocks and solicited advice might be interpreted as an indication of farmers’
perception that returns to participating in extension programs are indeed positive (otherwise they would
not ask for extension in light of negative shocks).


                                                   25
not find evidence of a significant link between the amount of labor farmers allocate to
their plots (including both family and hired labor) and participation in extension services
(see Table A12 in Appendix D). Thus, our data seem to provide no empirical support for
this concern.


5    Conclusion
In this paper, we study demand for advice in an agricultural setting. We first lay out
a novel channel to explain farmers’ demand for agricultural extension services. Second,
we provide empirical evidence that is consistent with implications of the model. The
empirical work makes use of a large recently collected panel dataset of farmers from three
countries in Sub-Saharan Africa.
     The theoretical insights we provide on the determinants of farmers’ demand for agri-
cultural advice and extension services center around the idea that participation in ex-
tension programs is only worthwhile if farmers devote sufficient attention to demanding,
listening to, and implementing received advice. Based on this feature, we model farmers
as rational decision makers facing a limited capacity to process information, causing at-
tention to be a scarce resource. Our model gives rise to several interesting predictions.
First, farmers’ decision to request extension services will depend on the amount of at-
tention they are willing to devote to their agricultural production process, as opposed to
other areas of life. Second, negative shocks to household income or wealth can result in
larger demand for agricultural advice. Importantly, this holds even for shocks which do
not directly affect optimal farming practices and thus do not constitute an immediate
reason for farmers to request agricultural advice.
     Overall, the model suggests a specific channel through which limited attention affects
agricultural input choices and productivity, based on (imperfect) demand for advice. The
derived insights complement the existing literature in this context and can contribute to
explanations for low demand for extension service. Specifically, the implications of our
model complement the insights of Banerjee and Mullainathan (2008), who argue that
“people may not be able to fully attend to their jobs if they are also worrying about
problems at home, and being distracted in this way reduces productivity” (p. 489).
The model we propose explains how scarce attention translates into low productivity
even if outside help is available at low monetary cost. On the other hand, the model
implies a link between shocks and increased attention to agricultural production: Non-
agricultural shocks can raise farmers’ marginal utility of additional income, consequently
increasing their willingness to devote costly attention to agricultural production, leading
to a higher probability that extension services are demanded. Given that participation in
such programs translates into better production outcomes and eventually higher incomes,



                                            26
this mechanism may also be seen as adding to the spectrum of ways by which the poor
respond to adverse shocks and ex post cope with risk.
     When we assess the predictions of the model empirically, we find statistically strong
evidence for the predicted link between non-agricultural shocks and farmers’ demand
for extension. The results are robust to various alternative specifications and also hold
when individual types of non-agricultural shocks, such as health, income, and crime-
related shocks, are separately considered. Yet, the size of the estimates suggests that the
link between non-agricultural shocks and demand for advice is of a modest quantitative
magnitude. Thus, while the empirical results are consistent with attentional constraints
being part of the low demand for agricultural advice, costly attention is apparently only
one piece of the puzzle to explain the low demand for advice. The empirical work also
provides some evidence that suggests that the effect of non-agricultural shocks is not due
to an increased need for credit or because non-agricultural shocks, some of which may
affect household labor supply, work through a demand for advice on labor inputs. Yet,
our analysis is limited by the available data and we clearly cannot rule out all plausible
alternative explanations. More empirical research to investigate scarce attention as a
possible limiting factor for further development in agriculture is warranted.
     Overall, the findings in this paper highlight the possibility that farmers’ optimal
demand for advice and the decision to participate in extension programs are affected by
constraints to information processing due to costly attention. Just as findings by Drexler
et al. (2014) in the context of financial literacy suggest that simpler rules may sometimes
be more helpful, our results point to the importance of designing advisory services in ways
that minimize the cognitive burden associated with requesting and absorbing advice.
Our results can thus be seen as providing support for recent initiatives which aim at
making agricultural information more easily accessible to farmers in developing countries,
including by increasing the availability of mobile phone-based services (Cole and Fernando,
2013; Casaburi et al., 2014) and developing new tools for farmers to obtain personalized,
real-time advice from an interactive online database (Fabregas et al., 2017). In addition,
our findings suggest that the timing of offering extension services matters, as farmers may
be more willing to devote attention to listening to and implementing advice when they
are facing a more pressing need to increase agricultural production – even if this need was
caused by events unrelated to farming.


References
Acharya, S., and S. L. Wee. 2020. “Rational inattention in hiring decisions.” American
  Economic Journal: Macroeconomics 12 (1): 1–40.
Aker, J. C. 2011. “Dial “A” for agriculture: A review of information and communi-
  cation technologies for agricultural extension in developing countries.” Agricultural
  Economics 42 (6): 631–647.


                                            27
Ambuehl, S., A. Ockenfels, and C. Stewart. 2018. “Attention and selection effects.”.
 Working paper.
Anderson, J. R. 2007. “Agricultural advisory services.” Background paper for the World
 Development Report 2008 .
Anderson, J. R., and G. Feder. 2007. “Agricultural extension.” Handbook of Agricultural
 Economics 3: 2343–2378.
Bandiera, O., R. Burgess, E. Deserrano, R. Morel, I. Rasul, and M. Sulaiman. 2018.
  “Social ties and the delivery of development programs.”. Working paper.
Bandiera, O., and I. Rasul. 2006. “Social networks and technology adoption in Northern
  Mozambique.” Economic Journal 116 (514): 869–902.
Banerjee, A., and S. Mullainathan. 2008. “Limited attention and income distribution.”
  American Economic Review 98 (2): 489–493.
Bardhan, P., and C. Udry. 1999. Development Microeconomics. Oxford University Press.
     s, V., M. Bauer, J. Chytilov´
Bartoˇ                                        ejka. 2016. “Attention discrimination:
                                 a, and F. Matˇ
  Theory and field experiments with monitoring information acquisition.” American
  Economic Review 106 (6): 1437–1475.
Beaman, L., J. Magruder, and J. Robinson. 2014. “Minding small change among small
  firms in Kenya.” Journal of Development Economics 108 (2014): 69–86.
Benin, S., E. Nkonya, G. Okecho, J. Randriamamonjy, E. Kato, G. Lubade, and M. Ky-
  otalimye. 2011. “Returns to spending on agricultural extension: The case of the Na-
  tional Agricultural Advisory Services (NAADS) program of Uganda.” Agricultural Eco-
  nomics 42 (2): 249–267.
Benjamin, D. 1992. “Household composition, labor markets, and labor demand: Testing
  for separation in agricultural household models.” Econometrica 60 (2): 287–322.
                      undeln, and D. Lagakos. 2018. “How do hours worked vary with in-
Bick, A., N. Fuchs-Sch¨
  come? Cross-country evidence and implications.” American Economic Review 108 (1):
  170–99.
Birkhaeuser, D., R. Evenson, and G. Feder. 1991. “The economic impact of agricultural
  extension: A review.” Economic Development and Cultural Change 39 (3): 607–650.
Birner, R., and J. R. Anderson. 2007. “How to make agricultural extension demand-
  driven? The case of India’s agricultural extension policy.” IFPRI Discussion Paper
  00729 .
Bold, T., K. Kaizzi, J. Svensson, and D. Yanagizawa-Drott. 2017. “Lemon technolo-
  gies and adoption: Measurement, theory and evidence from agricultural markets in
  Uganda.” Quarterly Journal of Economics 132 (3): 1055–1100.
Bordalo, P., N. Gennaioli, and A. Shleifer. 2013. “Salience and consumer choice.” Journal
  of Political Economy 121 (5): 803–843.
Buehren, N., M. Goldstein, E. Molina, and J. Vaillant. 2019. “The impact of strengthening
  agricultural extension services on women farmers: Evidence from Ethiopia.” Agricul-
  tural Economics 50 (4): 407–419.



                                           28
Cameron, A. C., J. B. Gelbach, and D. L. Miller. 2008. “Bootstrap-based improvements
  for inference with clustered errors.” Review of Economics and Statistics 90 (3): 414–
  427.
Caplin, A., and M. Dean. 2013. “Rational inattention and state dependent stochastic
  choice.”. Unpublished working paper.
Casaburi, L., M. Kremer, S. Mullainathan, and R. Ramrattan. 2014. “Harnessing ICT to
  increase agricultural production: Evidence from Kenya.”. Unpublished working paper.
                                                                     atel Group,
Chipeta, S. 2006. Demand Driven Agricultural Advisory Services. Neuchˆ
 Lindau.
Cole, S., and A. N. Fernando. 2013. “The value of advice: Evidence from mobile phone-
  based agricultural extension.”. Unpublished working paper.
Conley, T. G., and C. R. Udry. 2010. “Learning about a new technology: Pineapple in
  Ghana.” American Economic Review 100 (1): 35–69.
Davis, K. E. 2008. “Extension in Sub-Saharan Africa: Overview and assessment of past
  and current models, and future prospects.” Journal of International Agricultural and
  Extension Education 15 (3): 15–28.
Deininger, K., S. Savastano, and F. Xia. 2017. “Smallholders’ land access in Sub-Saharan
  Africa: A new landscape?.” Food Policy 67: 78–92.
DellaVigna, S. 2009. “Psychology and economics: Evidence from the field.” Journal of
  Economic Literature 47 (2): 315–372.
Dercon, S. 2002. “Income risk, coping strategies, and safety nets.” World Bank Research
  Observer 17 (2): 141–166.
Dercon, S., and L. Christiaensen. 2011. “Consumption risk, technology adoption and
  poverty traps: Evidence from Ethiopia.” Journal of Development Economics 96 (2):
  159–173.
Dercon, S., D. O. Gilligan, J. Hoddinott, and T. Woldehanna. 2009. “The impact of agri-
  cultural extension and roads on poverty and consumption growth in fifteen Ethiopian
  villages.” American Journal of Agricultural Economics 91 (4): 1007–1021.
Drexler, A., G. Fischer, and A. Schoar. 2014. “Keeping it simple: Financial literacy and
  rules of thumb.” American Economic Journal: Applied Economics 6 (2): 1–31.
Duflo, E., M. Kremer, and J. Robinson. 2008. “How high are rates of return to fertilizer?
 Evidence from field experiments in Kenya.” American Economic Review 98 (2): 482–
 488.
Duflo, E., M. Kremer, and J. Robinson. 2011. “Nudging farmers to use fertilizer: Theory
 and experimental evidence from Kenya.” American Economic Review 101 (6): 2350–
 2390.
Evenson, R., and D. Gollin. 2003. “Assessing the impact of the Green Revolution, 1960
  to 2000.” Science 300 (5620): 758–762.
Evenson, R. E. 2001. “Economic impacts of agricultural research and extension.” Hand-
  book of Agricultural Economics 1: 573–628.



                                          29
Fabregas, R., M. Kremer, J. Robinson, and F. Schilbach. 2017. “Netflix for agri-
  culture.”. Work in progress, URL: https://www.atai-research.org/wp-content/
  uploads/2017/06/Netflix-for-Agriculture-Prof.-Michael-Kremer.pdf (last ac-
  cessed on October 18, 2019).
FAO. 2014. The state of food and agriculture: Innovation in family farming. Food and
  Agriculture Organization of the United Nations, Rome.
Feder, G. 2005. “The challenges facing agricultural extension - and a new opportunity.”
  New Agriculturalist . Available at: http://new-ag.info/en/view/point.php?a=1253
  (last accessed on October 18, 2019).
Foster, A. D., and M. R. Rosenzweig. 1995. “Learning by doing and learning from oth-
  ers: Human capital and technical change in agriculture.” Journal of Political Econ-
  omy 103 (6): 1176–1209.
Gabaix, X. 2014. “A sparsity-based model of bounded rationality.” Quarterly Journal of
 Economics 129 (4): 1661–1710.
Gabaix, X., D. Laibson, G. Moloche, and S. Weinberg. 2006. “Costly information ac-
 quisition: Experimental analysis of a boundedly rational model.” American Economic
 Review 96 (4): 1043–1068.
Gido, E. O., K. W. Sibiko, O. I. Ayuya, and J. K. Mwangi. 2015. “Demand for agricultural
  extension services among small-scale maize farmers: Micro-level evidence from Kenya.”
  Journal of Agricultural Education and Extension 21 (2): 177–192.
Goecke, H., W. J. Luhan, and M. W. Roos. 2013. “Rational inattentiveness in a forecasting
 experiment.” Journal of Economic Behavior and Organization 94 (2013): 80–89.
Gollin, D., and C. Udry. 2017. “Heterogeneity, measurement error, and misallocation:
 Evidence from African agriculture.”. Unpublished working paper.
Greenwood, P., and R. Parasuraman. 1991. “Effects of aging on the speed and attentional
  cost of cognitive operations.” Developmental Neuropsychology 7 (4): 421–434.
Handel, B., and J. Schwartzstein. 2018. “Frictions or mental gaps: What’s behind the
  information we (don’t) use and when do we care?.” Journal of Economic Perspec-
  tives 32 (1): 155–78.
Hanna, R., S. Mullainathan, and J. Schwartzstein. 2014. “Learning through noticing:
  Theory and evidence from a field experiment.” Quarterly Journal of Economics 129 (3):
  1311–1353.
Haushofer, J., and E. Fehr. 2014. “On the psychology of poverty.” Science 344 (6186):
  862–867.
IFAD. 2013. Smallholders, food security, and the environment. International Fund for
  Agricultural Development, Rome.
Jama, B., and G. Pizarro. 2008. “Agriculture in Africa: Strategies to improve and
  sustain smallholder production systems.” Annals of the New York Academy of Sci-
  ences 1136 (1): 218–232.
Jayachandran, S. 2006. “Selling labor low: Wage responses to productivity shocks in
  developing countries.” Journal of Political Economy 114 (3): 538–575.



                                           30
Jensen, R. 2000. “Agricultural volatility and investments in children.” American Eco-
  nomic Review 90 (2): 399–404.
Kacperczyk, M., S. Van Nieuwerburgh, and L. Veldkamp. 2016. “A rational theory of
 mutual funds’ attention allocation.” Econometrica 84 (2): 571–626.
Karlan, D., R. Osei, I. Osei-Akoto, and C. Udry. 2014. “Agricultural decisions after
 relaxing credit and risk constraints.” Quarterly Journal of Economics 129 (2): 597–
 652.
Kochar, A. 1995. “Explaining household vulnerability to idiosyncratic income shocks.”
 American Economic Review 85 (2): 159–164.
 oszegi, B., and A. Szeidl. 2013. “A model of focusing in economic choice.” Quarterly
K˝
 Journal of Economics 128 (1): 53–104.
Kremer, M., G. Rao, and F. Schilbach. 2019. “Behavioral development economics.” In
  Handbook of Behavioral Economics, edited by D. Bernheim, S. DellaVigna, and D. Laib-
  son. Elsevier.
LaFave, D., and D. Thomas. 2016. “Farms, families, and markets: New evidence on
  completeness of markets in agricultural settings.” Econometrica 84 (5): 1917–1960.
Luo, Y. 2008. “Consumption dynamics under information processing constraints.” Review
  of Economic Dynamics 11 (2): 366–385.
Mackowiak, B., F. Matejka, and M. Wiederholt. 2018. “Survey: Rational inattention, a
 disciplined behavioral model.”. Working paper.
  ckowiak, B., and M. Wiederholt. 2009. “Optimal sticky prices under rational inatten-
Ma´
 tion.” American Economic Review 99 (3): 769–803.
  ckowiak, B., and M. Wiederholt. 2015. “Business cycle dynamics under rational inat-
Ma´
 tention.” Review of Economic Studies 82 (4): 1502–1532.
Mani, A., S. Mullainathan, E. Shafir, and J. Zhao. 2013. “Poverty impedes cognitive
 function.” Science 341 (6149): 976–980.
Mankiw, N. G., and R. Reis. 2002. “Sticky information versus sticky prices: A proposal to
 replace the New Keynesian Phillips curve.” Quarterly Journal of Economics 117 (4):
 1295–1328.
Martin, D. 2017. “Strategic pricing with rational inattention to quality.” Games and
 Economic Behavior 104 (2017): 131–145.
Morris, M., V. Kelly, R. Kopicki, and D. Byerlee. 2007. Fertilizer use in African agricul-
 ture: Lessons learned and good practice guidelines. World Bank, Washington, DC.
Moser, C. M., and C. B. Barrett. 2006. “The complex dynamics of smallholder technology
 adoption: The case of SRI in Madagascar.” Agricultural Economics 35 (3): 373–388.
Munshi, K. 2004. “Social learning in a heterogeneous population: Technology diffusion in
 the Indian Green Revolution.” Journal of Development Economics 73 (1): 185–213.
Naeher, D. 2020. “Technology adoption under costly information processing.”. Unpub-
  lished working paper.



                                           31
Niu, C., and C. Ragasa. 2018. “Selective attention and information loss in the lab-to-farm
  knowledge chain: The case of Malawian agricultural extension programs.” Agricultural
  Systems 165: 147–163.
OECD/FAO. 2016. Agricultural Outlook 2016-2025. OECD Publishing, Paris.
Ragasa, C. 2018. Supply of and demand for agricultural extension services in Malawi: A
  synthesis. International Food Policy Research Institute, Washington, DC.
Rivera, W., and G. Alex. 2005. Extension reform for rural development: Case studies of
  international initiatives. World Bank, Washington, DC.
Sallee, J. M. 2014. “Rational inattention and energy efficiency.” Journal of Law and
  Economics 57 (3): 781–820.
Schultz, T. W. 1964. Transforming Traditional Agriculture. Yale University Press.
Sheahan, M., and C. B. Barrett. 2017. “Ten striking facts about agricultural input use in
  Sub-Saharan Africa.” Food Policy 67 (2017): 12–25.
Sims, C. A. 2003. “Implications of rational inattention.” Journal of Monetary Eco-
  nomics 50 (3): 665–690.
Suri, T. 2011. “Selection and comparative advantage in technology adoption.” Econo-
  metrica 79 (1): 159–209.
Tutino, A. 2013. “Rationally inattentive consumption choices.” Review of Economic
  Dynamics 16 (3): 421–439.
Udry, C. 1995. “Risk and saving in Northern Nigeria.” American Economic Review 85 (5):
 1287–1300.
United Nations. 2005. World Public Sector Report 2005: Unlocking the Human Potential
 for Public Sector Performance. United Nations, New York.
Van Campenhout, B., W. Walukano, F. Nattembo, L. Nazziwa-Nviiri, and J. Blom. 2017.
  “The role of information in agricultural technology adoption: Experimental evidence
  from rice farmers in Uganda.” IFPRI Discussion Paper 01684 .
Van Nieuwerburgh, S., and L. Veldkamp. 2010. “Information acquisition and under-
  diversification.” Review of Economic Studies 77 (2): 779–805.
Wiederholt, M. 2010. “Rational inattention.” Prepared for The New Palgrave Dictionary
 of Economics .
World Bank. 2008. World Development Report 2008: Agriculture for Development. World
 Bank, Washington, DC.
World Bank. 2015. World Development Report 2015: Mind, society, and behavior. World
 Bank, Washington, DC.
World Bank. 2018. Living Standard Measurement Study - Integrated Surveys on
 Agriculture (LSMS-ISA). URL: http://surveys.worldbank.org/lsms/programs/
 integrated-surveys-agriculture-ISA (last accessed on October 18, 2019).




                                           32
APPENDIX

A      Proof of propositions
In deriving the farmer’s optimal behavior, we make the following simplifying restrictions
and assumptions. First, we focus on the case where the utility function u(.) is strictly con-
cave, such that negative shocks ϑi that decrease the overall level of utility lead to a larger
marginal utility of additional payoff. Second, we assume that u(.) is twice continuously
differentiable and the second derivative with respect to the chosen action takes a larger ab-
solute value at lower levels of payoff. Third, we derive the optimal behavior for a quadratic
approximation of the utility function around the solution of the model when zi = 0 for
                                                    ˆϑ
i ∈ {A, N }, i.e., around the point (ai , zi , ϑ) = a               ˆϑ
                                                      i , 0, ϑ with a i = arg maxai ∈R u ai , 0, ϑ .
Let this point be denoted as π , and let u ˜(.) denote the second-order Taylor approximation
of the function u(.) around π . The approximation is given by

                                       ∂u                    ∂ 2u       (ai − aˆϑ
                                                                                i)
                                                                                   2
            u
            ˜ (ai , zi , ϑ) = u(π ) +      (π )(ai − aˆϑ
                                                       i ) +       (π )                                   (A.1)
                                      ∂ai                    ∂a2 i           2
                                 2                             2
                                ∂ u                          ∂ u
                             +         (π )(ai − aˆϑ
                                                   i )zi +           (π )(ai − aˆϑ
                                                                                 i )(ϑ − ϑ) + r,
                               ∂ai ∂zi                      ∂ai ∂ϑ

where r contains terms that are independent of ai . Using this approximation, the utility
loss associated with choosing a suboptimal action ai = a∗
                                                        i is given by


                                               ∂u
         ˜ (a∗
         u   i , zi , ϑ) − u
                           ˜ (ai , zi , ϑ) =       (π )(a∗
                                                         i − ai )                                         (A.2)
                                               ∂ai
                                                  1 ∂ 2u            2
                                               + · 2 (π ) a∗              2
                                                                   i − ai + 2ˆ aϑ         ∗
                                                                                 i (ai − ai )
                                                  2 ∂ai
                                                     ∂ 2u              ∂ 2u
                                               +            (π )zi +         (π )(ϑ − ϑ) (a∗  i − ai ).
                                                    ∂ai ∂zi           ∂ai ∂ϑ

As we are approximating around a local maximum, it holds that ∂u/∂ai = 0 and ∂ 2 u/∂a2i <
                                         ˜(.) in equation (A.1) with respect to ai shows
0. Furthermore, taking the derivative of u
that the optimal action is determined by

                ∂u
                 ˜   ∂ 2u                ∂ 2u             ∂ 2u              !
                                  ˆϑ
                    = 2 (π )(ai − a i)+         (π )zi +        (π )(ϑ − ϑ) = 0.                          (A.3)
                ∂ai  ∂ai                ∂ai ∂zi          ∂ai ∂ϑ

         ˆ11 = ∂ 2 u/∂a2
Denoting u                                                   ˆ13 = ∂ 2 u/∂ai ∂ϑ(π ), this implies
                               ˆ12 = ∂ 2 u/∂ai ∂zi (π ), and u
                       i (π ), u
that
                                           ˆ12
                                           u         ˆ13
                                                     u
                               a∗    ˆϑ
                                 i = a i −     zi −      (ϑ − ϑ).                           (A.4)
                                           ˆ11
                                           u         ˆ11
                                                     u




                                                         33
Using these results, equation (A.2) can be simplified to

                                      ˆ11 ∗ 2
                                      u
     ˜ (a∗
     u   i , zi , ϑ) − u
                       ˜ (ai , zi , ϑ) =   ai − a2      aϑ
                                                   i + 2ˆ
                                                                   ∗
                                                                         ˆ11 (a∗
                                                          i (ai − ai ) − u        ˆϑ
                                                                               i −a
                                                                                         ∗
                                                                                    i )(ai − ai )
                                       2
                                        ˆ11
                                        u
                                     =−     (ai − a∗ 2
                                                   i) .                                         (A.5)
                                         2

Denoting the utility u˜ (a∗                                                 ∗
                           i , zi , ϑ) obtained from the optimal action as ui (ϑ), it thus follows
that
                                                         ωi (ϑ)
                              ˜ (ai , zi , ϑ) = u∗
                              u                  i (ϑ) −        (ai − a∗ 2
                                                                       i) ,                  (A.6)
                                                           2
where ωi (ϑ) = |∂ 2 u/∂a2i (π )| depends negatively on ϑ due to the made assumptions about
the properties of u. Working with the second-order approximation of the utility function
in equation (A.6) greatly simplifies the further derivation, as it allows us to use the result
from Ma´ ckowiak and Wiederholt (2009) that the optimal action taken by a rationally
inattentive agent who faces a quadratic objective function is simply given by the expected
optimal action conditional on the received signal (i.e., there are no higher order effects
involved, such as those related to risk aversion, when choosing actions). The decision
problem of the farmer thus simplifies to

                                                           ωi (ϑ)
                    max                    Es u∗
                                               i (ϑ) −            ( ai − a∗
                                                                          i)
                                                                            2
                                                                              −              µi κi    (A.7)
                  κA ,κN ≥0                                  2
                               i∈{A,N }                                           i∈{A,N }




                                           s.t. a∗
                                                 i = φzi                                              (A.8)
                                               ai = E [ a∗
                                                         i |si ]                                      (A.9)
                                               si = zi +     i                                       (A.10)
                                               κi = H (zi ) − H (zi |si )                            (A.11)

where condition (A.9) corresponds to the choice in the second stage of the original decision
problem. The optimal allocation of attention can be derived as follows. First, combining
equations (A.8) - (A.10) yields

            ωi (ϑ)               ωi (ϑ) 2                             ωi (ϑ) 2 2
       Es          (ai − a∗
                          i)
                             2
                               =       φ E (E [zi |si ] − zi )2 |si =       φ σzi |si .              (A.12)
              2                    2                                    2

Using the expressions for the prior and posterior entropy provided in (4) to transform
equation (A.11) leads to
                                                 2
                                      1        σz   i
                                 κi = log2     2
                                                        ,
                                      2       σz i |s i


which means that
                                                2           2 −2κi
                                               σz i |s i
                                                         = σz i
                                                                2  .                                 (A.13)


                                                        34
Using these two results, the optimization problem given by (A.7)-(A.11) can be written
as
                                          ωi (ϑ) 2 2 −2κi
                  max            u∗
                                  i (ϑ) −       φ σzi 2   −       µi κi         (A.14)
                 κA ,κN ≥0                  2
                             i∈{A,N }                                          i∈{A,N }

Taking the derivative with respect to κi yields the first-order condition

                                               2 −2κi
                             {κi } : ωi (ϑ)φ2 σz i
                                                   2  ln(2) = µi .                           (A.15)

Solving for κi leads to the result
                           
                            1 log       ωi (ϑ)φ2  2
                                                i σzi ln(2)        ωi (ϑ)φ2  2
                                                                          i σzi ln(2)
                             2       2          µi
                                                              if          µi
                                                                                        ≥1
                    κ∗
                     i =                                                                     (A.16)
                           0                                 else

The numerator of the term inside the logarithm is the marginal increase in utility as-
sociated with paying attention. The denominator, µi , is the marginal cost of attention.
The farmer will only devote a positive amount of attention to domain i if the marginal
benefit at κ = 0 exceeds the marginal cost. For a concave utility function with the as-
sumed properties, it holds that the marginal utility of additional payoff is higher at lower
levels of utility. Therefore, a negative shock to payoff, either in the agricultural or the
non-agricultural domain, will shift the farmer to a part of the utility function where ωi (ϑ)
takes a larger value, causing κ∗ A to increase (since the model is symmetric in the two
domains, the same effects also applies to κ∗ N ). In the same way, it follows from equation
               ∗
(A.16) that κA will be larger with more costly mistakes (i.e., ωA ↑ for reasons other than
shocks ϑ), larger prior uncertainty about the optimal action (φ2    2
                                                                 A σzA ↑), and a smaller cost
of being attentive to the agricultural domain (µA ↓).


B     Data and variables
The data in our source, the World Bank’s LSMS-ISA database, are collected in collab-
oration with the national statistics offices of the participating countries in Sub-Saharan
Africa, often building upon existing surveys on agriculture in these countries. While the
LSMS-ISA project seeks to provide information in a comparable manner across countries,
the survey designs sometimes involve considerable differences in the type of collected in-
formation and the way questions are asked. As far as possible, we restrict our analysis
to variables for which the available information is comparable across countries. The fol-
lowing provides additional information on the construction of variables and handling of
the unique survey design in each country. For more details on the data collection process
and utilized tools, we refer to the original survey documents and enumerator manuals




                                                      35
available at World Bank (2018).

Agricultural seasons. For countries with two cropping seasons within a year (Malawi
and Uganda), we focus our analysis only on the main season. In Nigeria, the survey in
each season is split into two visits (post-planting and post-harvest), and some of the in-
formation is collected separately for the time “since the new year” (first visit) and “since
the last interview” (second visit). In computing variables over one agricultural season, we
combine the information from these two visits (e.g., the number of extension contacts is
calculated as the sum of reported contacts across both visits).

Local currencies and units of measurement. For all countries, local currencies are
transformed into US dollars using yearly average exchange rates from the World Develop-
ment Indicator database. Information on quantities and area that was collected based on
local units of measurement is transformed into standard units by using (as far as avail-
able) the conversion factors provided in the LSMS-ISA database.

Plots and farm size. All surveys collect data on plot areas based both on farmer-
reported values and GPS measurement. In Malawi and Nigeria, the data are collected
at the plot level. In Uganda, the corresponding unit of observation is called parcel (we
follow the approach of other authors and aggregate information across plots belonging to
the same parcel; see Sheahan and Barrett, 2017). For our analysis, we use GPS measures
as the main basis. Since GPS estimates are not available for all plots (e.g., due to flood-
ing, security concerns, or because plots are located too far away from the household),
we complement the data with farmer-reported plot areas where necessary. The variable
Cultivated Area (which we use both in the construction of other variables and as a depen-
dent variable itself), includes only plots which are cultivated by the household in a given
season (i.e., excluding fallow, pasture, and forest land, as well as plots that are given or
rented out by the household).

Perceived quality of received advice. The information in Table 1 on whether received
advice was rated as useful by farmers was obtained as follows. In all surveys, farmers are
asked to rate the quality of the advice received. In Malawi and Nigeria, the question
is stated as “how useful was the advice/information received?”. In Uganda, the corre-
sponding question is “How would you rate the advice received?”. The underlying scales
differ between countries. To make farmers’ responses comparable across countries, we
classified responses into two categories: “useful or very useful” and “useless or bad”. The
former category comprises the responses “useful” and “very useful” in Malawi, “somewhat
useful” and “very useful” in Nigeria, and “average” and “good” in Uganda. Responses
grouped together as “useless or bad” are “not very useful” and “useless” in Malawi, “not


                                            36
very useful”, “not useful”, and “harmful” in Nigeria, and “bad” in Uganda.

Use of modern farm inputs. All surveys collect data on used seeds, fertilizer, agro-
chemicals (mainly pesticides and herbicides), and irrigation. However, the exact informa-
tion differs between countries and, in some instances, between different survey rounds of
the same country. This applies particularly to the data on improved seed varieties, where
most of the earlier survey rounds do not allow us to distinguish between different types of
seeds. Therefore, our indicator variable for improved seeds is only available for a subset
of survey rounds. For fertilizer, agrochemicals, manure, and irrigation we observe binary
input use decisions at the plot level, i.e., whether farmers used the respective input (in
any quantity) on individual plots or parcels. Based on this plot-level information, the
household indicators which we use in the analysis are set equal to one if farmers report
having used the respective input on at least one plot during a given season.
     For seeds, fertilizer and agrochemicals we also observe used quantities. The corre-
sponding household-level variables are obtained by summing up reported quantities across
individual plots belonging to a household. In addition, we also observe the costs farmers
incurred for these three inputs, including costs for purchasing, transportation to the farm,
and obtaining input coupons (in Malawi). In most surveys, data on incurred costs are
only available for inputs which have been purchased in the same season. In these cases,
we use the available data to calculate country- and wave-specific prices for each input,
which we then use to assign values to those inputs which are reported to be left over from
the previous season.

Farm output and productivity. The variable Harvest is based on farmer-reported
estimates of the value of harvested crops for individual plots. In Nigeria, farmers are
asked directly what would have been the total value of their harvested crops, if everything
had been sold at current market prices. In Malawi and Uganda, the total value of harvest
is based on separate data about harvested quantities and prices of sold crops. The latter
is used to calculate country- and wave-specific median market prices for each individual
crop, which are then used to assign monetary values to farmers’ harvests. This approach
is used because many farmers sell only a small fraction of their harvested quantities of
a given crop, and it is not clear whether they could have sold the remaining (unsold)
produce for the same price.
     Agricultural productivity (Harvest per ha ) is computed by dividing, for each plot,
the value of harvest by cultivated area, and then calculating the weighted average at the
household level using plot areas as weights. In addition, we also observe quantities and
values of a number of farm inputs. We use these data to construct a proxy for farmers’
profits, namely Net Harvest, which captures farmers’ total value of harvest net of the



                                            37
estimated value of used seeds, fertilizer, and agrochemicals.

Farm labor input. We observe the amount of labor input provided by family members
as well as the incurred costs for hired labor. The latter is based on household-specific
information about average daily wages paid, number of hired persons, and number of
days for which laborers were hired. To capture the cost for both types of labor in a single
variable (Cost Labor ), we use the information on costs for hired labor to calculate country-
wave median wages for two categories (harvest and non-harvest activities), which we then
use to assign a monetary value to provided family labor (i.e., assuming the same wage for
family labor as for hired labor for each category). We do not account for differences in
the composition of hired or family labor with respect to gender and age (i.e., adults or
children).




                                             38
                                   Table A1: List of variables

                        Full
Variable                        Malawi    Nigeria    Uganda       Descriptionb
                      samplea

HH characteristics:
Adults                  2.89      2.61     3.25          2.79    Number of adults in HH
Children                2.90      2.51     2.98          3.16    Number of children in HH
Education (head)                                                 Educational level of HH head
 No schooling           0.26      0.17     0.40          0.20    D: Never attended school
 Primary educ.          0.51      0.60     0.35          0.59    D: Primary or Quranic school
 Secondary educ.        0.23      0.23     0.24          0.21    D: Secondary education or more
Age head                47.8      43.6     51.9          47.2    Age of HH head (years)
Male head               0.78      0.76     0.87          0.70    D: HH head is male
Farm outcomes:
Harvest                502.0     289.6     882.1         308.7   Value of harvest (USD)
Net Harvest            433.6     228.7     750.4         295.8   Value of harvest net of input costs (USD)
Production choices:
Cultivated Area         1.35      1.12     1.18           1.70   Total cultivated plot area (hectare)
Fertilizer             0.40      0.79      0.42          0.04    D: HH uses inorganic fertilizer
Agrochemicals           0.19     0.05       0.40          0.11   D: HH uses agrochemicals
Improved Seeds          0.33      0.56      0.17          0.18   D: HH uses improved seeds
Manure                  0.15      0.22      0.11          0.12   D: HH uses organic fertilizer
Irrigation             0.02      0.01      0.03          0.02    D: HH uses irrigation
Qty. Fertilizer         67.5      96.5     108.8           1.9   Quantity of used fertilizer (kg)
Qty. Agrochem.           2.7      1.8       5.5           0.7    Quantity of used agrochemicals (kg)
Cost Fertilizer         30.1      41.4     48.9           1.9    Value of used fertilizer (USD)
Cost Agrochem.          8.8       2.9      20.7           1.8    Value of used agrochemicals (USD)
Cost Seeds              28.8      15.8      59.5           9.1   Value of used seeds (USD)
Cost Labor             281.8     107.6     552.6         170.2   Value of family and hired labor (USD)
Extension:
Advice                  0.34      0.63     0.19          0.26    D: Received advice in past year
Extension               0.29      0.47     0.17          0.26    D: Received advice (excl. media)
Solicited Extension     0.20      0.35     0.12          0.15    D: Actively requested advice
Contacts                1.12      1.58     0.88          1.03    Number of extension contacts
Solicited Contacts      0.81      1.25     0.53          0.78    Number of actively requested contacts
Shocks:
AgShock                 0.43      0.82     0.12          0.41    D: Agricultural shock in past year
NonAgShock              0.22      0.40     0.14          0.16    D: Non-agricultural shock in past year
Observations (max.)    25,274    7,519     8,874         8,881

Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: Numbers are mean values. Included survey rounds are: Malawi (2010-11, 2013-14, 2016-17),
Nigeria (2010-11, 2012-13, 2015-16), and Uganda (2009-10, 2010-11, 2011-12, 2013-14). Only households
with at least one cultivated plot are included. a The full sample is obtained by pooling data across
countries and waves. b “D:” indicates dummy variables. Local currencies are transformed into US dollars
using yearly average exchange rates from the World Development Indicator database.




                                                    39
                                          Table A2: Exposure to agricultural advice by country and survey round

                                                                       Malawi                       Nigeria                               Uganda
     Description
                                                              wave 1   wave 2   wave 3    wave 1    wave 2     wave 3    wave 1        wave 2  wave 3     wave 4

     HHs who received agricultural advice (%)                   32.1     47.0      69.0      16.9      13.4      19.7      34.9           21.9     28.1      21.0
        of those:
        avg. number of contacts                                  3.7      4.4       3.9       5.3       6.2       4.9       5.0            4.4      3.1       3.8
        more than 2 contacts (%)                                50.1     56.0      50.0      56.5      63.2      63.2      46.8           47.2     36.2      34.9
        more than 10 contacts (%)                                4.6      6.5       5.3      12.0      17.6       9.4      11.8           13.2      6.4       5.8
        received advice rated as useful or very usefula (%)     44.6     91.6      91.9      86.9      87.5      97.9      88.9           78.0     87.0      92.8




40
        received advice rated as useless or bada (%)             7.0      9.7       8.9       5.8       2.3       1.8       7.5           10.6     13.4       8.4
        paid in order to receive advice (%)                      0.9      1.1       0.7      10.6      22.2      12.1      19.0           18.0     10.7       6.6
     HHs who actively solicited advice (%)                      26.1     36.9      49.5      11.5      10.8      14.9      22.4           11.9     15.5      12.4
     Share of solicited contacts (%)                            80.6     82.5      78.0      60.6      70.3      62.9      75.5           74.0     78.4      75.1
     Observations                                              2,592    3,029     1,601     3,072     2,925     2,847     1,833          2,080    2,158     2,389

     Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
                                                                                                                                 a
     Notes: Agricultural advice measures in-person advice and excludes advice received through media (TV, radio, flyers, etc.).       Exact wording differs between
     countries.
                                                      Table A3: Sources of advice by country and survey round

                                                                   Malawi                            Nigeria                                Uganda
     Descriptiona
                                                         wave 1    wave 2      wave 3     wave 1     wave 2      wave 3     wave 1      wave 2   wave 3        wave 4

     Government extension service                          49.3       43.9       32.3        27.5       16.2        18.9       78.6        86.4       93.7       86.7
     Private extension service, input supplier              1.9        2.7        2.6         9.8       11.6        13.4        7.5         2.0        1.8        4.4
     NGO, lead farmer and farmer field day/school            9.1       12.9        8.7         5.1       11.0        13.2       17.5        12.5        7.6       12.7
     Farmer cooperative/association                         1.8        1.7        1.0         5.3        1.6         4.0        7.2         5.5        1.0        4.6
     Other farmer (neighbor, relative, large-scale)        11.6       19.1       34.2        46.9       51.4        55.1        4.7         1.1        0.3        1.2




41
     Village agricultural extension meeting                 6.9        8.2        4.3         5.9        5.9        12.7
     Media (TV, radio, flyers)                              40.7       53.8       35.8        23.5       25.8        15.6
     Other                                                  1.9        2.8        1.4         4.3        3.9         3.0        7.5         5.7        1.3         2.6
     Observations (HHs who received advice)               1,119      1,966      1,429        510         438        597         640        455         606        502

     Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: Numbers are percentages of households who received advice from indicated source, conditional on having received any advice. a Exact wording differs between
     countries.
Table A4: Mean values of household characteristics across households with different (tran-
sitory) shocks

                   (1)      (2)          (3)             (4)      (5)       (6)          (7)         (8)
                                               No       With         With transition in:
                   No    With      With    transition transition                          Both
                                                                 AgShock NonAgShock
                  shock AgShock NonAgShock in shocks in shocks                           shocks
Adults          3.02        2.79         2.86           2.86     2.92       2.91        2.91        2.89
Children        2.76       2.96          2.88           2.68     3.05      3.17         2.95        3.10
Primary educ.   0.41       0.57          0.54           0.48     0.53      0.53         0.54        0.55
Secondary educ. 0.26       0.21         0.22            0.24     0.21      0.22         0.20        0.20
Age head        50.03      46.34        47.34           47.41    48.08     47.87        48.45       48.39
Male head       0.82       0.75          0.76           0.79     0.76      0.77         0.75        0.76
HH-Wave Obs.      6,545   15,737       10,804           10,647   14,134    9,486        9,042       4,394
Households        3,152   6,805         4,504           6,085    4,893     3,168        3,171       1,446
P-value (H0 : Demeaned    values are jointly equal to            0.093     0.127        0.502       0.838
    demeaned values in    column (4)a )
P-value (H0 : Demeaned    values are jointly equal to            0.560     0.393        0.689       0.894
    demeaned values in    column (1)a )

Source: Authors’ calculation based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: Columns show mean values for subsets of households as follows. Column (1): households
without any reported shock across all survey rounds, (2): households with at least one agricultural
shock across survey rounds, (3): households with at least one nonagricultural shock across survey
rounds, (4): households without transition in any shock, (5): households with at least one transition
in AgShock or NonAgShock , (6): households with transition in AgShock , (7): households with transi-
tion in N onAgshock , (8): households with transitions in both AgShock and N onAgshock . a The last two
rows report p-values for joint F -tests on the equality of demeaned values across the subset of households
included in the corresponding column and those included in column (4) and column (1), respectively,
using wild cluster bootstrap standard errors (10,000 repetitions) at the stratum level (consisting of 24
clusters defined as region×rural). Demeaned values are demeaned taking the mean of the respective
variable at the household level over up to four waves. For details on the construction of AgShock and
N onAgshock , see Table 3.




                                                    42
C      Robustness of tests of model predictions

     Table A5: Test of model predictions: Robustness to different ways of clustering

                          Advice                            Extension                  Solicited Extension
                 (1)       (2)        (3)        (4)          (5)         (6)        (7)        (8)      (9)
AgShock        0.087∗    0.087∗∗∗   0.087∗∗∗   0.072∗       0.072∗∗∗    0.072∗∗∗   0.061∗∗   0.061∗∗∗   0.061∗∗∗
               (0.098)    (0.000)    (0.002)   (0.084)       (0.000)     (0.004)   (0.039)    (0.000)    (0.000)
NonAgShock     0.033∗    0.033∗∗∗   0.033∗∗∗   0.030∗       0.030∗∗     0.030∗∗∗   0.025∗    0.025∗∗    0.025∗∗∗
               (0.096)    (0.007)    (0.006)   (0.078)      (0.020)      (0.008)   (0.058)   (0.025)     (0.006)
HH FE            yes       yes        yes       yes           yes         yes        yes       yes        yes
Wave FE          yes       yes        yes       yes           yes         yes        yes       yes        yes
Controls         yes       yes        yes       yes           yes         yes        yes       yes        yes
Clustera        2-way    district     wild     2-way        district     wild      2-way     district    wild
Observations    20,392   20,392      20,392    20,392       20,392      20,392     19,653    19,653     19,653

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. a Clusters and associated
numbers of clusters are as follows: “district” (158); “2-way”: stratum (24) × wave (up to 4); “wild”:
wild bootstrap (1,000 repetitions) for stratum (24). As controls, full sets of dummies are included for the
following variables: adults, children, education of HH head, age of head, male head.




                                                       43
   Table A6: Test of model predictions: Robustness to alternative variable definitions

                               (1)           (2)              (3)          (4)          (5)         (6)
                            Extension     Extension        Extension   Sol. Exten.   Solicited   Solicited
                            (no peers)                                 (no peers)    Extension   Extension
AgShock                      0.053∗∗∗      0.067∗∗∗                     0.055∗∗∗     0.056∗∗∗
                              (0.004)       (0.002)                      (0.001)      (0.001)

NonAgShock                    0.030∗∗                       0.031∗∗     0.020∗∗                  0.025∗∗∗
                              (0.016)                       (0.025)     (0.022)                   (0.007)

NonAgShock (incl. food)                    0.041∗∗∗                                  0.041∗∗∗
                                            (0.007)                                   (0.000)

AgShock (incl. food)                                       0.064∗∗∗                              0.059∗∗∗
                                                            (0.003)                               (0.000)
HH FE                           yes          yes              yes          yes          yes         yes
Wave FE                         yes          yes              yes          yes          yes         yes
Controls                        yes          yes              yes          yes          yes         yes
Observations                  20,391        20,392          20,392       20,260       19,653      19,653
R-squared (within)             0.020         0.034           0.033        0.018        0.027       0.026

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at
the stratum level (consisting of 24 clusters defined as region×rural). As controls, full sets of dummies are
included for the following variables: adults, children, education of HH head, age of head, male head.




                                                      44
    Table A7: Test of model predictions: Robustness to exclusion of HH fixed effects

                         (1)        (2)        (3)          (4)        (5)        (6)        (7)       (8)
                        Advice     Advice     Advice       Advice     Advice     Advice     Advice    Advice
AgShock                 0.252∗∗∗   0.252∗∗    0.252∗∗      0.252∗∗∗   0.250∗∗∗   0.250∗∗    0.250∗∗   0.250∗∗∗
                         (0.001)   (0.016)    (0.035)       (0.000)    (0.000)   (0.014)    (0.031)    (0.000)
NonAgShock              0.095∗∗∗   0.095∗∗∗   0.095∗∗      0.095∗∗∗   0.095∗∗∗   0.095∗∗∗   0.095∗∗   0.095∗∗∗
                         (0.000)    (0.000)   (0.042)       (0.000)    (0.000)    (0.000)   (0.046)    (0.000)
Adults                                                                0.013∗∗    0.013∗      0.013    0.013∗∗∗
                                                                      (0.032)    (0.054)    (0.102)    (0.000)
Children                                                               0.002      0.002      0.002     0.002
                                                                      (0.606)    (0.654)    (0.602)   (0.515)
Primary educ. (head)                                                   0.031      0.031      0.031     0.031
                                                                      (0.497)    (0.638)    (0.519)   (0.312)
Second. educ. (head)                                                   0.051      0.051      0.051     0.051
                                                                      (0.277)    (0.448)    (0.327)   (0.128)
Age head                                                               -0.001     -0.001     -0.001    -0.001
                                                                      (0.665)    (0.632)    (0.713)   (0.657)
Age head sq.                                                           -0.000     -0.000     -0.000    -0.000
                                                                      (0.570)    (0.526)    (0.611)   (0.573)
Male head                                                              0.031      0.031      0.031     0.031
                                                                      (0.192)    (0.262)    (0.274)   (0.120)
HH FE                     no         no         no           no         no         no         no        no
Wave FE                   yes        yes        yes          yes        yes        yes        yes       yes
Clustera                stratum     wild      2-way        district   stratum     wild      2-way     district
Observations             20,392    20,392     20,392       20,392      20,392    20,392     20,392    20,392
R-squared                 0.090     0.090      0.090        0.090      0.101      0.101      0.101     0.101

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. a Clusters and associated numbers
of clusters are as follows: “stratum” (24); “wild”: wild bootstrap (1,000 repetitions) for stratum (24);
“2-way”: stratum (24) × wave (up to 4); “district” (158). The omitted category for education is “no
schooling”.




                                                      45
            Table A8: Test of model predictions: Disaggregated topics of advice

                         (1)          (2)         (3)          (4)        (5)     (6)       (7)
                     Agric. prod. Marketing Livestock Animal diseases                     Access
                                                                        Fishery Forestry
                     & processing & crop sales production & vaccination                  to credit
AgShock                0.069∗∗∗       0.038∗∗∗     0.034∗∗        0.031∗∗∗       -0.004 0.023∗∗     0.009∗∗
                        (0.002)        (0.001)     (0.022)         (0.006)      (0.671) (0.035)     (0.038)
NonAgShock              0.027∗∗        0.010         0.022         0.017∗        -0.003  0.014       0.001
                        (0.022)       (0.446)       (0.123)        (0.096)      (0.611) (0.372)     (0.961)
HH FE                     yes           yes           yes            yes          yes       yes       yes
Wave FE                   yes           yes           yes            yes          yes       yes       yes
Controls                  yes           yes           yes            yes          yes       yes       yes
Observations            20,351        20,350        20,350         20,350        20,350   13,697a   13,697a
R-squared (within)       0.033        0.016          0.029          0.020         0.015    0.033     0.028
Mean dep. var.          0.254         0.072          0.095          0.087         0.023    0.044     0.036

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the
stratum level (consisting of 24 clusters defined as region×rural). The dependent variables are only equal
to one if the household actively requested advice. a Numbers of observations in columns (6) and (7) are
smaller, because the categories “Forestry” and “Access to credit” are not available in Uganda (see list of
topics of advice in Table 2). As controls, full sets of dummies are included for the following variables:
adults, children, education of HH head, age of head, male head.




                                                    46
                                       Table A9: Placebo test

                                             (1)                        (2)                       (3)
                                          Unsolicited                Unsolicited               Unsolicited
                                          Extension                  Extension                 Extension
AgShock                                       0.004                      0.004                    0.003
                                             (0.579)                   (0.612)                   (0.715)
NonAgShock                                    0.007                     0.007                     0.007
                                             (0.344)                   (0.336)                   (0.360)
Adults                                                                  -0.000
                                                                       (0.828)
Children                                                                 0.001
                                                                       (0.708)
Primary educ. (head)                                                     0.008
                                                                       (0.307)
Second. educ. (head)                                                    -0.010
                                                                       (0.534)
Age head                                                                -0.001
                                                                       (0.717)
Age head sq.                                                             0.000
                                                                       (0.597)
Male head                                                                0.006
                                                                       (0.671)
HH FE                                          yes                        yes                      yes
Wave FE                                        yes                        yes                      yes
Full set of control dummies                                                                        yes
Observations                                 20,392                    20,392                     20,392
R-squared (within)                            0.003                     0.004                      0.012

Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at
the stratum level (consisting of 24 clusters defined as region×rural). The omitted category for education
is “no schooling”. Full sets of control dummies are included for the following variables: adults, children,
education of HH head, age of head, male head.




                                                      47
D      The relation between agricultural advice, input
       choices, and farm outcomes
This appendix investigates the association between demand-driven advisory services and
farm decisions and outcomes. For this purpose, we estimate a number of regressions of
the form

              yit = β0 + β1 Extensionit + β2 AgShock it + β3 Xit + µi + ωt +           it ,      (D.1)

where the dependent variables we consider belong to one of the two following groups. In
the first group are measures of farm performance, such as total value of harvest, value of
harvest per hectare, and value of harvest net of input costs. Second, we are interested in
specific channels through which provided agricultural advice may affect these outcomes.
In investigating potential channels, we focus on the link between extension services and
farmers’ decisions related to the amount of cultivated land, allocation of labor, and use
of modern farm inputs.
     For this analysis we use our observational data, and unobserved factors will likely
be correlated with both exposure to extension services and outcomes.31 Relative to most
existing work based on observational data (see Birkhaeuser et al., 1991; Evenson, 2001;
Benin et al., 2011), the panel structure of our dataset allows us to make some progress
towards dealing with endogeneity, by making it possible to control for unobserved time-
invariant household characteristics via the included household fixed effects. Further, we
can also control to some extent for time varying unobservables by including indicators
of agricultural shocks incurred by households. Finally, we also include a number of time
varying household characteristics which are likely to be important (e.g., household size
and education). Nevertheless, we stress that our ability to make causal claims is limited.
     To measure farm performance, the dataset contains information about the total value
of harvested crops (estimated by farmers) and about the incurred costs for inputs, includ-
ing seeds, fertilizer, and other agrochemicals (see Appendix B for details on how these
costs are calculated). This information is used to construct a measure of agricultural
productivity (value of harvest per hectare of cultivated land) and a proxy for farmers’
profits (defined as the value of harvest net of the estimated value of used seeds, fertilizer,
and agrochemicals). As potential channels, we investigate the effect of extension services
on the size of cultivated land, allocation of labor, and farmers’ decision to adopt modern
agricultural input technologies. Regarding the latter, we focus on use of improved seed
  31
    The dominant view in the literature seems to be that simple OLS estimates of the effects of extension
on farm outcomes tend to be biased upward, as more productive farmers also have more exposure to
extension services (Birkhaeuser et al., 1991; Evenson, 2001).




                                                  48
varieties, inorganic fertilizer, and other agrochemicals (pesticides, herbicides, etc.).32 In
addition to binary indicators for use of individual modern inputs in a given season, we
also observe the applied quantities of fertilizer and agrochemicals. In order to limit the
influence of outliers, most of our regressions involving farm outputs or input costs use
logarithmic transformations of the original values as dependent variables. In addition, in
some specifications (indicated in the notes to tables) we exclude outliers that we identify
as those observations for which either value of harvest or net value of harvest per hectare
is in the top or bottom 1% of the distribution.
     Table A10 reports estimates for the regression model specified in equation (D.1),
both with and without including non-agricultural shocks. For both specifications, results
are obtained for two different outcome variables. In the first five columns, the dependent
variable is the logarithm of farmer-reported total value of harvest. In the second half of
the table, the dependent variable is the logarithm of value of harvest net of input costs
(including costs for seeds, fertilizer, and agrochemicals).33
     The estimates in Table A10 show a consistently positive correlation between the two
outcome variables and indicators of received extension services. In all specifications, the
coefficients of the extension variables are statistically significant. The quantitative role
of extension services suggested by these estimates seems reasonable: Receiving (solicited)
agricultural advice is associated with a 7% larger value of harvest and 8% higher value of
harvest net of input costs.
     We note that the coefficient for NonAgShock does not change much once the extension
variable is included. This suggests that the direct effect of the non-agricultural shocks
on seeking advice is, although statistically significant, small in absolute magnitude. In
addition, the fact that there are significant effects of non-agricultural shocks on input
choices (see Tables A12 and A13), even after controlling for extension seeking behavior,
highlights a previously not mentioned implication of the theory. Theory also predicts
that other actions that bring high returns but need attention receive more attention in
response to a non-agricultural shock. This may include using modern inputs (without
asking for advice). Therefore, the fact that non-agricultural shocks predict input use is
also consistent with the theory.
     Furthermore, we find that the sizes of the coefficients for the extension variables
are considerably larger when we do not control for unobserved household characteristics
using fixed effects (not shown here). This suggests an upward bias in simple OLS estimates
caused by positive selection of more productive farmers into extension programs, which
  32
     Information on improved seeds is only available for a subset of survey rounds, since most of the earlier
LSMS-ISA surveys do not distinguish between different types of seeds.
  33
     Note that after dropping outliers as described in the text, of the logarithmic dependent variables
used in Tables A10 and A12, only “cost of inputs” and “cost of labor” have remaining zeros. To be able
to include observations with zero costs, we take the logarithms after adding 0.01 to the original variables.
This transformation affects 2,472 observations in Table A12, columns (1) to (5), and 78 observations in
columns (6) to (10), respectively.


                                                     49
                                                              Table A10: Extension and farm outcomes

                                                          log(Harvest)                                                            log(Net Harvest)
                                 (1)            (2)           (3)             (4)           (5)            (6)           (7)             (8)           (9)          (10)
     AgShock                   -0.078∗∗      -0.082∗∗       -0.083∗∗∗      -0.081∗∗       -0.082∗∗      -0.081∗∗       -0.086∗∗       -0.086∗∗       -0.085∗∗     -0.086∗∗
                                (0.017)       (0.013)        (0.010)        (0.015)        (0.011)       (0.029)        (0.022)        (0.019)        (0.023)      (0.019)

     NonAgShock                 0.019                         0.016                        0.017          0.010                         0.008                       0.008
                               (0.526)                       (0.578)                      (0.555)        (0.735)                       (0.801)                     (0.779)

     Extension                               0.079∗∗∗       0.078∗∗∗                                                   0.079∗∗∗       0.078∗∗∗
                                              (0.001)        (0.001)                                                    (0.007)        (0.007)




50
     Solicited Extension                                                   0.068∗∗∗       0.067∗∗∗                                                   0.083∗∗       0.083∗∗
                                                                            (0.006)        (0.006)                                                   (0.020)       (0.019)

     HH FE                       yes            yes            yes            yes           yes            yes           yes            yes            yes           yes
     Wave FE                     yes            yes            yes            yes           yes            yes           yes            yes            yes           yes
     Controls                    yes            yes            yes            yes           yes            yes           yes            yes            yes           yes
     Observations              18,112         18,112         18,112         18,112        18,112         17,483         17,483         17,483        17,483        17,483
     R-squared (within)        0.028          0.029          0.029           0.029        0.029          0.025          0.026           0.026         0.026         0.026

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined as
     region×rural). Outliers are excluded as explained in the text. As controls, full sets of dummies are included for the following variables: adults, children, education
     of HH head, age of head, male head.
coincides with the perception in the literature described above (see footnote 31). As an
aside, note that the agricultural shock variable is indeed significantly negatively correlated
with harvest value in Table A10, which confirms the usefulness of this (self-reported) shock
variable for the above analysis.
     There are several possible channels through which extension services may affect farm
outcomes. First, it may be the case that access to advice simply motivates farmers to
cultivate more land. In this case, total value of harvest would tend to increase, while
productivity (i.e., harvest per unit of land) might remain unaffected. Second, extension
services may indeed help farmers to achieve higher productivity. An important channel
through which this may be accomplished is by facilitating the adoption and use of modern
farm inputs, such as improved seed varieties, fertilizer, and pesticides, which are known
to be an important factor in determining agricultural productivity (Evenson and Gollin,
2003; Morris et al., 2007; Duflo et al., 2011).
     Tables A11 and A12 present estimation results which suggest that particularly the
latter channel through use of modern inputs may be at work. According to the results in
columns (2) to (5) of Table A11, farmers who receive extension do not tend to cultivate
more (or less) land.34 At the same time, these farmers are able to generate 53 to 58 dollars
in additional harvest value per hectare of land, an increase of 6 to 7% relative to the mean
(columns 7 to 10).
     As can be seen in columns (2) to (5) of Table A12, farmers who receive advice
spend more money on modern farm inputs (the included costs are for seeds, fertilizer,
and agrochemicals). This may be seen as in line with the figures reported in Table 2,
which show that a large share of the provided advice from extension services is related to
agricultural input use. Finally, columns (7) to (10) of Table A12 suggest that extension
services are not significantly associated with the amount of labor farmers allocate to their
plots (including both family and hired labor).
     Tables A13 and A14 provide further insights into the link between extension and mod-
ern input use. As shown in the first four columns of Table A13, receiving advice through
extension is associated with a slight increase in the number of agricultural inputs used
(considered inputs are improved seeds, irrigation, manure, fertilizer, and agrochemicals).
When looking at binary adoption decisions for individual inputs, improved seeds, fertilizer,
and agrochemicals show significant coefficients. For fertilizer and agrochemicals, we also
observe applied quantities (Table A14). While the coefficients are not always statistically
significant, they are generally positive and of plausible magnitude. Overall, the results
in Tables A11, A12, A13, and A14 are in line with the view that extension services can
have positive effects on farmers’ productivity by increasing the use of modern farm inputs.

  34
    In the data, cultivated area changes frequently over time, which may be explained by the large share
of plots which farmers report to rent or use for free.



                                                  51
                                                     Table A11: Extension and production choices (Part 1)

                                              Cultivated Area (in ha)                                                   Harvest per ha (US$)
                               (1)          (2)         (3)         (4)          (5)           (6)             (7)               (8)             (9)         (10)
     AgShock                  -0.136       -0.105      -0.139       -0.100      -0.134      -127.360∗      -135.291∗∗        -131.667∗∗    -135.104∗∗     -131.476∗∗
                             (0.585)      (0.670)     (0.579)      (0.683)     (0.589)       (0.051)         (0.043)           (0.046)       (0.043)        (0.046)

     NonAgShock               0.562                    0.561                    0.563        -58.779∗                         -61.453∗                     -60.792∗
                             (0.131)                  (0.136)                  (0.133)        (0.066)                          (0.056)                      (0.060)

     Extension                             0.058       0.040                                                53.129∗∗          56.036∗∗
                                          (0.747)     (0.831)                                                (0.046)           (0.037)

     Solicited Extension                                            -0.019      -0.037                                                         55.188∗∗    57.717∗∗




52
                                                                   (0.918)     (0.850)                                                          (0.049)     (0.044)

     HH FE                     yes          yes         yes          yes         yes           yes            yes               yes              yes          yes
     Wave FE                   yes          yes         yes          yes         yes           yes            yes               yes              yes          yes
     Controls                  yes          yes         yes          yes         yes           yes            yes               yes              yes          yes
     Observations             18,652      18,652       18,652      18,652       18,652       14,893          14,893           14,893           14,893       14,893
     R-squared (within)        0.023       0.023       0.023        0.023       0.023         0.023           0.023            0.024            0.023        0.024
     Mean dep. var.            1.40        1.40         1.40        1.40         1.40        871.13          871.13           871.13           871.13       871.13

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined as
     region×rural). Outliers are excluded as explained in the text. Local currencies are transformed into US dollars using yearly average exchange rates from the
     World Development Indicator database. As controls, full sets of dummies are included for the following variables: adults, children, education of HH head, age of
     head, male head.
                                                        Table A12: Extension and production choices (Part 2)

                                                          log(Cost Inputs)                                                   log(Cost Labor)
                                (1)           (2)               (3)            (4)         (5)         (6)           (7)           (8)           (9)        (10)
                                      ∗             ∗                                ∗           ∗
     AgShock                  0.255         0.242              0.235         0.246       0.238        -0.063        -0.058        -0.060        -0.062      -0.064
                              (0.079)       (0.091)           (0.102)        (0.088)     (0.099)     (0.511)       (0.536)       (0.520)       (0.510)     (0.495)

     NonAgShock               0.127∗∗                         0.117∗∗                    0.120∗∗      0.030                       0.031                     0.030
                              (0.020)                         (0.025)                    (0.023)     (0.431)                     (0.391)                   (0.421)

     Extension                             0.300∗∗∗          0.296∗∗∗                                               -0.047        -0.048
                                            (0.001)           (0.001)                                              (0.489)       (0.473)

     Solicited Extension                                                     0.270∗∗∗    0.266∗∗∗                                               0.011       0.011




53
                                                                              (0.001)     (0.001)                                              (0.868)     (0.878)

     HH FE                      yes           yes               yes            yes         yes         yes           yes           yes          yes          yes
     Wave FE                    yes           yes               yes            yes         yes         yes           yes           yes          yes          yes
     Controls                   yes           yes               yes            yes         yes         yes           yes           yes          yes          yes
     Observations             18,680        18,680            18,680         18,680      18,680       18,385       18,385        18,385        18,385      18,385
     R-squared (within)        0.033         0.035             0.035         0.034        0.034        0.132        0.132        0.132          0.131       0.132

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined
     as region×rural). Logarithms are obtained after adding 0.01 to the original variables. Outliers are excluded as explained in the text. Local currencies are
     transformed into US dollars using yearly average exchange rates from the World Development Indicator database. As controls, full sets of dummies are included
     for the following variables: adults, children, education of HH head, age of head, male head.
                                                        Table A13: Extension and modern input use (Part 1)

                                  Number of Inputs                       Improved Seeds                           Irrigation                                Manure
                            (1)     (2)     (3)         (4)       (5)     (6)     (7)         (8)       (9)     (10)      (11)    (12)      (13)         (14)    (15)           (16)
                                                                                                                                                   ∗           ∗          ∗
     AgShock               0.019   0.016   0.019   0.016 -0.017 -0.018 -0.017 -0.018       0.001 0.001 0.001 0.001 0.030          0.028   0.029   0.027∗
                          (0.529) (0.599) (0.514) (0.584) (0.120) (0.112) (0.124) (0.114) (0.643) (0.668) (0.629) (0.656) (0.070) (0.090) (0.076) (0.096)

     NonAgShock                    0.051∗∗            0.052∗∗            0.014∗∗            0.015∗∗∗            0.001             0.001                0.033∗∗∗               0.033∗∗∗
                                   (0.020)            (0.019)            (0.010)             (0.005)           (0.820)           (0.816)                (0.009)                (0.009)

     Extension            0.113∗∗ 0.111∗∗                       0.040∗∗ 0.039∗∗                         -0.000 -0.000                       0.019   0.018
                          (0.014) (0.015)                       (0.038) (0.040)                        (0.968) (0.962)                     (0.284) (0.307)

     Solicited Exten.                        0.112∗∗ 0.110∗∗                       0.042∗∗ 0.041∗∗                        -0.001 -0.001                             0.028   0.027




54
                                             (0.017) (0.017)                       (0.038) (0.040)                       (0.869) (0.863)                           (0.118) (0.129)

     HH FE                  yes      yes      yes       yes      yes       yes      yes       yes       yes      yes      yes      yes      yes          yes        yes         yes
     Wave FE                yes      yes      yes       yes      yes       yes      yes       yes       yes      yes      yes      yes      yes          yes        yes         yes
     Controls               yes      yes      yes       yes      yes       yes      yes       yes       yes      yes      yes      yes      yes          yes        yes         yes
     Observations       18,693     18,693    18,693   18,693    13,469   13,469    13,469   13,469     18,545 18,545 18,545 18,545 18,609              18,609      18,609     18,609
     R-squared (within) 0.070       0.070     0.069    0.070     0.023    0.023     0.023    0.023     0.016 0.016 0.016 0.016 0.078                    0.079       0.078      0.079
     Mean dep. var.      1.015      1.015     1.015    1.015     0.318    0.318     0.318    0.318      0.022 0.022 0.022 0.022 0.148                   0.148       0.148      0.148

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined as
     region×rural). Outliers are excluded as explained in the text. As controls, full sets of dummies are included for the following variables: adults, children, education
     of HH head, age of head, male head.
                                                       Table A14: Extension and modern input use (Part 2)

                                    Fertilizer: Any                  Fertilizer: log(Quantity)                 Agrochemicals: Any             Agrochemicals: log(Quantity)
                            (1)       (2)       (3)      (4)       (5)      (6)       (7)      (8)       (9)      (10)      (11)    (12)      (13)    (14)      (15)    (16)

     AgShock               -0.002    -0.003 -0.002 -0.002 -0.058 -0.059 -0.058 -0.059 -0.001 -0.001 0.000 -0.000 0.001 -0.004 0.005 0.000
                          (0.854)   (0.848) (0.872) (0.865) (0.646) (0.640) (0.644) (0.638) (0.962) (0.926) (0.984) (0.978) (0.990) (0.944) (0.930) (0.998)

     NonAgShock                      0.001              0.001              0.008              0.009               0.008             0.009            0.079∗            0.082∗
                                    (0.892)            (0.870)            (0.901)            (0.891)             (0.259)           (0.230)           (0.063)           (0.053)

     Extension            0.023∗∗∗ 0.023∗∗∗                      0.179∗∗∗ 0.179∗∗                      0.039∗∗ 0.039∗∗                       0.249∗∗ 0.247∗∗
                           (0.002) (0.003)                        (0.009) (0.010)                      (0.032) (0.033)                       (0.039) (0.041)

     Solicited Exten.                         0.020∗∗ 0.020∗∗                       0.192∗∗ 0.192∗                          0.031 0.030                         0.203 0.200




55
                                              (0.020) (0.022)                       (0.049) (0.052)                        (0.152) (0.153)                     (0.137) (0.140)

     HH FE                  yes       yes       yes      yes       yes      yes      yes      yes       yes       yes       yes      yes      yes      yes      yes      yes
     Wave FE                yes       yes       yes      yes       yes      yes      yes      yes       yes       yes       yes      yes      yes      yes      yes      yes
     Controls               yes       yes       yes      yes       yes      yes      yes      yes       yes       yes       yes      yes      yes      yes      yes      yes
     Observations          18,620   18,620    18,620   18,620    18,589   18,589    18,589   18,589    18,665    18,665    18,665 18,665 18,631      18,631 18,631 18,631
     R-squared (within)     0.015    0.015     0.014    0.014     0.014   0.014      0.014   0.014     0.023     0.023      0.022 0.022   0.025       0.025  0.024 0.024
     Mean dep. var.         0.399    0.399     0.399    0.399                                          0.219     0.219     0.219 0.219

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors are clustered at the stratum level (consisting of 24 clusters defined as
     region×rural). Logarithms are obtained after adding 0.01 to the original variables. Outliers are excluded as explained in the text. As controls, full sets of dummies
     are included for the following variables: adults, children, education of HH head, age of head, male head.
Robustness. Table A15 tests the robustness of the link between extension and farm
outcomes (as reported in Table A10) with respect to clustering and handling of missing
observations. In particular the data on harvest values contains many observations where
information is missing for some individual plots or crops within households. In the analysis
so far, these cases have been treated as zero when summing up over all crops and plots
of a household. The results in Table A15 show that the link between extension and farm
outcomes remains significant when the analysis is restricted to those households for which
information on harvest is complete (i.e., available for all crops and all plots).
     Table A16 repeats a similar exercise for the link between extension and farmers’ pro-
duction choices (as reported in Tables A11 and A12). The dependent variables Cultivated
Area and Harvest per ha are based on the amount of cultivated area, which (similar to
the data on harvest) involves missing values for individual plots within households. The
results in columns (1) to (8) in Table A16 show that excluding households for which this
is the case leads to similar coefficients as before, though significance levels in columns (5)
to (8) drop. As a result, the coefficients of the extension variables in these columns be-
come slightly insignificant (with p-values around 0.13). For all other estimates which have
previously been significant, the results remain significant at least at the 10% significance
level.
     Table A17 tests the robustness of the link between extension and use of modern
farm inputs (as reported in Tables A13 and A14) with respect to two-way clustering of
standard errors. As significance levels tend to become smaller, only some of the coefficients
for extension remain statistically significant. In particular, the results in Table A17 show
that the coefficients for extension and solicited extension remain statistically significant at
the 10% level for use of improved seeds and fertilizer, whereas the results for agrochemicals
become insignificant.




                                             56
                                                                  Table A15: Extension and farm outcomes: Robustness

                                                                 log(Harvest)                                                                  log(Net Harvest)
                          (1)           (2)          (3)          (4)          (5)           (6)          (7)          (8)          (9)          (10)        (11)        (12)        (13)         (14)       (15)    (16)
                                ∗∗∗           ∗∗∗          ∗∗           ∗∗           ∗∗∗           ∗∗∗          ∗∗           ∗∗           ∗∗            ∗∗          ∗           ∗           ∗∗          ∗∗
     AgShock            -0.114   -0.115  -0.114 -0.115 -0.113    -0.114  -0.113 -0.114 -0.102 -0.101 -0.102 -0.101 -0.102 -0.101          -0.102 -0.101∗
                         (0.002) (0.001) (0.036) (0.028) (0.002) (0.001) (0.039) (0.030) (0.024) (0.021) (0.096) (0.085) (0.024) (0.020) (0.101) (0.090)

     NonAgShock                        0.008                     0.008                      0.009                     0.009                      -0.007                  -0.007                   -0.007             -0.007
                                      (0.802)                   (0.848)                    (0.802)                   (0.849)                    (0.862)                 (0.892)                  (0.855)            (0.888)

     Extension          0.105∗∗∗ 0.104∗∗∗ 0.105∗∗ 0.104∗∗                                                                         0.128∗∗∗ 0.128∗∗∗ 0.128∗∗ 0.128∗∗
                         (0.000) (0.000) (0.013) (0.014)                                                                           (0.000) (0.000) (0.027) (0.028)

     Solicited Exten.                                                        0.098∗∗∗ 0.098∗∗∗ 0.098∗∗ 0.098∗∗                                                                      0.138∗∗∗ 0.138∗∗∗ 0.138∗∗ 0.138∗∗




57
                                                                              (0.000) (0.000) (0.019) (0.020)                                                                        (0.000) (0.000) (0.011) (0.011)

     HH FE                yes          yes           yes         yes           yes          yes           yes         yes           yes           yes        yes         yes          yes         yes        yes     yes
     Wave FE              yes          yes           yes         yes           yes          yes           yes         yes           yes           yes        yes         yes          yes         yes        yes     yes
     Controls             yes          yes           yes         yes           yes          yes           yes         yes           yes           yes        yes         yes          yes         yes        yes     yes
     Sample restr.a        yes     yes               yes         yes            yes     yes               yes         yes     yes     yes   yes   yes     yes     yes   yes   yes
     Clusters           stratum stratum             2-way       2-way        stratum stratum             2-way       2-way stratum stratum 2-way 2-way stratum stratum 2-way 2-way
     Observations        13,648  13,648             13,648      13,648        13,648  13,648             13,648      13,648 13,199 13,199 13,199 13,199 13,199 13,199 13,199 13,199

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Standard errors clustered at the stratum level consist of 24 clusters defined as region×rural.
     Two-way (“2-way”) clusters are defined as stratum (24) × wave (up to 4). Outliers are excluded as explained in the text. a Observations are dropped if information
     on harvest involves missings. As controls, full sets of dummies are included for the following variables: adults, children, education of HH head, age of head, male
     head.
                                                    Table A16: Extension and production choices: Robustness

                            Cultivated Area (in ha)                   Harvest per ha (US$)                      log(Cost Inputs)                     log(Cost Labor)
                         (1)      (2)     (3)      (4)        (5)        (6)       (7)         (8)       (9)      (10)    (11)      (12)      (13)     (14)    (15)     (16)
     AgShock             -0.144 -0.178 -0.139 -0.174 -132.819 -128.563 -132.578 -128.311 0.242 0.235     0.246   0.238 -0.058 -0.060 -0.062 -0.064
                        (0.577) (0.496) (0.587) (0.504) (0.128) (0.134) (0.130) (0.136) (0.169) (0.178) (0.169) (0.177) (0.569) (0.555) (0.543) (0.530)

     NonAgShock                  0.581             0.583              -74.128∗∗             -73.436∗∗            0.117              0.120             0.031             0.030
                                (0.165)           (0.164)              (0.041)               (0.038)            (0.153)            (0.148)           (0.500)           (0.523)

     Extension           0.121 0.103                        52.594     56.132                           0.300∗ 0.296∗                         -0.047 -0.048
                        (0.412) (0.504)                     (0.126)    (0.125)                          (0.051) (0.051)                      (0.510) (0.500)

     Solicited Exten.                      0.060 0.042                            53.639     56.754                       0.270∗∗ 0.266∗∗                       0.011 0.011
                                          (0.677) (0.787)                         (0.134)    (0.134)                      (0.045) (0.046)                      (0.880) (0.890)




58
     HH FE               yes      yes      yes      yes      yes         yes       yes         yes       yes      yes      yes      yes       yes      yes      yes      yes
     Wave FE             yes      yes      yes      yes      yes         yes       yes         yes       yes      yes      yes      yes       yes      yes      yes      yes
     Controls            yes      yes      yes      yes      yes         yes       yes         yes       yes      yes      yes      yes       yes      yes      yes      yes
     Sample restr.a      yes    yes    yes    yes            yes         yes       yes        yes
     Cluster            2-way 2-way 2-way 2-way             2-way       2-way     2-way      2-way      2-way 2-way       2-way    2-way     2-way 2-way 2-way 2-way
     Observations       18,415 18,415 18,415 18,415         14,722      14,722    14,722     14,722     18,680 18,680     18,680   18,680    18,385 18,385 18,385 18,385

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Two-way (“2-way”) clusters are defined as stratum (24 entities based on region×rural)
     × wave (up to 4). Logarithms are obtained after adding 0.01 to the original variables. Outliers are excluded as explained in the text. a Observations are
     dropped if information on cultivated area involves missings. Local currencies are transformed into US dollars using yearly average exchange rates from the World
     Development Indicator database. As controls, full sets of dummies are included for the following variables: adults, children, education of HH head, age of head,
     male head.
                                                      Table A17: Extension and modern input use: Robustness

                                    Improved                       Fertilizer:              Fertilizer:               Agrochemicals:              Agrochemicals:
                                      Seeds                          Any                  log(Quantity)                    Any                    log(Quantity)
                              (1)              (2)           (3)                 (4)     (5)           (6)           (7)           (8)           (9)           (10)
     AgShock                 -0.018         -0.018          -0.003           -0.002     -0.059        -0.059        -0.001        -0.000        -0.004         0.000
                            (0.474)        (0.478)         (0.848)          (0.866)    (0.616)       (0.616)       (0.921)       (0.976)       (0.945)        (0.998)

     NonAgShock            0.014∗∗∗        0.015∗∗∗         0.001            0.001      0.008         0.009         0.008         0.009         0.079          0.082
                            (0.010)         (0.006)        (0.848)          (0.822)    (0.855)       (0.850)       (0.284)       (0.266)       (0.133)        (0.126)

     Extension              0.039∗                         0.023∗∗                     0.179∗                       0.039                       0.247
                            (0.092)                        (0.044)                     (0.060)                     (0.122)                     (0.131)




59
     Solicited Exten.                      0.041∗                           0.020∗                   0.192∗                       0.030                        0.200
                                           (0.077)                          (0.067)                  (0.094)                     (0.204)                      (0.196)

     HHFE                     yes              yes           yes                 yes     yes           yes           yes           yes           yes           yes
     Wave FE                  yes              yes           yes                 yes     yes           yes           yes           yes           yes           yes
     Controls                 yes              yes           yes                 yes     yes           yes           yes           yes           yes           yes
     Cluster                2-way           2-way           2-way            2-way     2-way         2-way         2-way          2-way         2-way         2-way
     Observations           13,469          13,469          18,620           18,620    18,589        18,589        18,665         18,665        18,631        18,631

     Source: Authors’ analysis based on survey data from the LSMS-ISA database (World Bank, 2018).
     Notes: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. p -values in parentheses. Two-way (“2-way”) clusters are defined as stratum (24 entities based on region×rural)
     × wave (up to 4). Logarithms are obtained after adding 0.01 to the original variables. Outliers are excluded as explained in the text. As controls, full sets of
     dummies are included for the following variables: adults, children, education of HH head, age of head, male head.