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

Abstract

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

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

1. Introduction

Increases in productivity are key to growth and poverty reduction, yet productivity in many domains is lagging 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 significant 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 percent of farmers actively solicit agricultural advice (including from other farmers). The fairly low interest in agricultural 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 advice, 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 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 demand 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 cost 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.

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 demonstrate 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 rainfall) 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 literature and provides further background on agricultural extension services. Section 3 presents the model and derives predictions about farmers’ demand for agricultural advice. 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 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, Mullainathan, and Schwartzstein 2014), the high cost of providing information if schooling levels are low (Schultz 1964), and 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, Kremer, and Robinson 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, Kremer, and Robinson 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 for 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 importance of salient aspects of different pieces of information over the true informational value they carry (Bordalo, Gennaioli, and Shleifer 2013; Koszegi and Szeidl 2013).⁶ In this paper we follow 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 rational inattention literature, see Handel and Schwartzstein 2018; Mackowiak, Matejka, and Wiederholt 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 endogenously 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) that 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, Luhan, and Roos 2013; Bartos et al. 2016; Ambuehl, Ockenfels, and Stewart 2018).⁸

Until now, applications of rational inattention have focused primarily on rich countries. 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 information 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 electricity and water supply, and thus have to dedicate more time and mental effort to everyday life problems (Banerjee and Mullainathan 2008; Bick, Fuchs-Schündeln, and Lagakos 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 they exacerbate the severity of limitations in cognitive capacity that 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 exogenous variation (Birkhaeuser, Evenson, and Feder 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 increases 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 percent more marketable crops, have 40 percent higher profits, and per capita consumption expenditure is 22 percent higher. Other modern approaches to provide advice employ information and communication technology (ICT), such as agricultural apps, consulting hotlines, and SMS-based reminders (Aker 2011; Cole and Fernando 2013; 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 percent 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 2008; 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 the voice and participation of farmers as compared to traditional top-down approaches (Rivera and Alex 2005; Birner and Anderson 2007).¹³ Second, the transformation in the agricultural extension sector toward 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 (a) why farmers may rationally decide not to participate in extension services and (b) 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 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 rainfall) but also shocks that 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 that 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 |$a_{i} \in \mathbb {R}$|⁠, with i ∈ {A, N}. Let the optimal action in each domain be given by

$$\begin{eqnarray*} a_{i}^{*} = \phi z_{i}, \end{eqnarray*}$$

where the random variable |$z_{i} \sim \mathcal {N}(0, \sigma _{z_{i}}^{2})$| describes the fundamental state of the economy, which is initially unobserved by the farmer.¹⁵ For example, one may think of z_A as capturing all relevant factors that determine the returns to different agricultural production choices, including current prices, demand for certain crops, and soil conditions. Let |$\overline{u}_{i}^{*}$| denote the maximal payoff for domain i associated with the optimal action |$a_{i}^{*}$|⁠. Choosing a suboptimal action (⁠|$a_{i} \ne a_{i}^{*}$|⁠) leads to a payoff that is smaller than |$\overline{u}_{i}^{*}$|⁠. 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 z_A).

In addition, we allow for exogenous events, denoted by ϑ_i, which directly reduce the payoff in each domain. For example, ϑ_A may capture harvest loss due to bad weather and ϑ_N may capture theft of money. In contrast to z_i, these shocks cannot be compensated for by the farmer through selecting appropriate actions. Instead, any realization of ϑ_i 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

$$\begin{eqnarray*} {u}_{i}^{*}=\overline{u}_{i}^{*}+\vartheta _{i}, \end{eqnarray*}$$

which we write as |${u}_{i}^{*}(\vartheta )$| 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 a_A and a_N.

To make better decisions, the farmer can devote attention to available sources of information and thereby reduce uncertainty about the realized states z_i and associated optimal actions. For example, in the agricultural domain this may correspond to the farmer contacting 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

$$\begin{eqnarray*} s_{i} = z_{i} + \epsilon _{i} \end{eqnarray*}$$

about the realization of z_i, where |$\epsilon _{i} \sim \mathcal {N}(0, \sigma _{\epsilon _{i}}^{2})$|⁠.¹⁷ In this case, the prior entropy of the random variable z_i and the posterior entropy that results from paying attention to the fundamental are defined as

$$\begin{equation} H(z_{i}) = \frac{1}{2} \log _{2} \left(2 \pi e \sigma ^{2}_{z_{i}}\right) \quad {and} \quad H(z_{i}|s_{i}) = \frac{1}{2} \log _{2} \left(2 \pi e \sigma ^{2}_{z_{i}|s_{i}}\right). \end{equation}$$

(1)

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

$$\begin{equation} \kappa _{i} = H(z_{i}) - H(z_{i} | s_{i}). \end{equation}$$

(2)

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 z_i.¹⁸

Capturing the idea that attention is a scarce resource, the farmer faces in each domain the cost μ_iκ_i, where μ_i > 0 denotes the unit cost of attention (e.g., opportunity cost of mental energy). Note that μ_i 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.

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 ϑ_i, 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:

$$\begin{eqnarray*} \max _{a_{A}, a_{N} \in \mathbb {R}} E \left[ u \left( a_{A}, a_{N} , z_{A}, z_{N} , \vartheta \right) \vert s_{A}, s_{N}\right], \end{eqnarray*}$$

where ϑ = (ϑ_A + ϑ_N). Let |$\hat{a}_{i}$| denote the optimal action for domain i ∈ {A, N} chosen by the farmer, i.e.,

$$\begin{equation} \hat{a}_{i} = \arg \max _{a_{i} \in \mathbb {R}} E \left[ u \left( a_{i} , z_{i} , \vartheta \right) \vert s_{i}\right] . \end{equation}$$

(3)

In the first stage the farmer observes the realization of the shocks ϑ_i 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

$$\begin{eqnarray*} \max _{\kappa _{i} \ge 0} E \left[ u \left( \hat{a}_{i} , z_{i} , \vartheta \right) \right] - \sum \mu _{i} \kappa _{i} , \end{eqnarray*}$$

subject to the constraint on information processing (2) and equation (3). 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 rational inattention (e.g., Maćkowiak and Wiederholt 2009; Gabaix 2014). The solution to the farmer’s optimization problem is formally derived in the supplementary online appendix S1. 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 (a) the more costly are mistakes in a_A, (b) the larger the prior uncertainty about the optimal action |$a^{*}_{A}$|⁠, and (c) the smaller the cost of paying attention to the fundamental z_A.

 

Proof.

See the supplementary online appendix S1.

According to the results in Proposition 1, factors that make it easier for farmers to pay attention to agricultural advice, or that cause suboptimal cultivation practices to be relatively more costly, will make it more likely that farmers request advice. This suggests that we should expect household characteristics that 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, Ockenfels, and Stewart 2018). In addition, one might expect demand for agricultural advice to be positively correlated with household size, given that households with fewer members feature fewer 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 associated 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 attention, 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., μ_i > 0), negative shocks in any of the two domains will cause the farmer to increase the amount of attention allocated to a_A, and thus raise the probability that agricultural advice is requested.

 

Proof.

See the supplementary online appendix S1.

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 μ_i) 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 applies less to shocks that 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 on 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 led 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 were 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 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 summary, 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 that 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 the supplementary online appendix S4). 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 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 differences in survey designs across countries can be found in the supplementary online appendix S2, together with a complete list of all variables and basic summary statistics (see table S2.1).

Extension services. In all three countries, the LSMS-ISA surveys collect detailed information 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 household 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 1 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 of advice that exclude advice obtained through media. On average, 29 percent of farmers report having received such advice. In Malawi (which generates 32 percent 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 percent. Conditional on receiving advice, farmers have an average of 4 to 5 contacts a year. While 51 percent of farmers receiving advice report more than 2 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 percent) that receive advice rate the received advice as useful, and less than 10 percent of farmers report having received advice that was useless. On average, less than 8 percent 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 percent in Nigeria to over 80 percent 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 S2.2 in the supplementary online appendix.

Table 1.

Exposure to Agricultural Advice in Last 12 Months

Description	Full sample	Malawi	Nigeria	Uganda
Households who received agricultural advice (% all household-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 household-year obs.)	19.5	35.0	12.4	15.3
Households who solicited advice (% household-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 Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

Note: Agricultural advice measures in-person advice and excludes advice received through media (TV, radio, flyers, etc.). ^aExact wording differs between countries (see the supplementary online appendix Data and Variables).

Table 2 provides additional insights into the topics on which farmers receive advice. Almost all farmers (91 percent) who have been in contact with extension services received advice 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 S2.3 in the supplementary online appendix shows that government extension services are responsible for 79 to 94 percent of extension provided in Uganda and between 32 and 49 percent in Malawi. In Nigeria, where public extension programs seem to be less prevalent, farmers report receiving advice relatively more often from private extension services and from other farmers.

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 (households who received advice)	6,877	3,336	1,369	2,172

Source: Authors’ calculation based on survey data from the Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

Note: 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). ^aExact wording differs between countries.

Shocks. Data on shocks are available at the household level and include detailed information 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,” “unusual 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 will be imperfect. We therefore explore several alternative classifications as part of the robustness tests.

Table 3.

Shocks to Households by Country

	Percentage of household-wave obs.				Percentage of households with
	with shock (on average during one year)				transition in shock (across survey rounds)
Descriptiona	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 (NonAgShock)	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
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 Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

Note: ^aExact wording differs between countries. ^bThe full sample is obtained by pooling data across countries and waves.

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 educational level, age, and gender of the household head.

4.2. Testing the Model’s Predictions

The regression model we estimate is given by

$$\begin{equation} \textrm{Extension}_{it} = \alpha _{0} + \alpha _{1} \textrm{AgShock}_{it} + \alpha _{2} \textrm{NonAgShock}_{it} + \alpha _{3} X_{it} + \mu _{i} + \omega _{t} + \epsilon _{it}, \end{equation}$$

(4)

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). The term AgShock_it is a dummy variable that captures shocks that 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). The term NonAgShock_it is a dummy variable that corresponds to shocks that 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 μ_i and ω_t denote household and wave fixed effects. Finally, X_it is a vector of time-varying household-level variables.

The hypothesis we test is that both coefficients α₁ and α₂ in equation (4) 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 S2.4 in the supplementary online 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 (4) for various measures of farmers’ exposure to agricultural advice. In the first three columns, the dependent variable is an indicator that equals 1 if in a given season the household received any advice related to farming (i.e., from any of the sources listed in table S2.3). 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 1 if the household actively requested advice.

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*	—
		(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)
Household 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 Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

Note: 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 household head, age of head, male head. *p < 0.10, **p < 0.05, ***p < 0.01. p-values in parentheses.

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 1 out of 40 (or 2.5 percent) 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 correlated 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 characteristics 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 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 types of shocks likely matter. 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.

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)
Household 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 Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

Note: 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 household member,” “death or disability of household 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 household head, age of head, male head. *p < 0.10, **p < 0.05, ***p < 0.01. p-values in parentheses.

We also perform placebo regressions. Specifically, we repeat the central regressions in table 4 with the dependent variable Unsolicited Extension, which equals 1 for households that received only advice that 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 S3.5 in the supplementary online appendix. 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.

Additional robustness checks with respect to clustering of standard errors, alternative variable definitions, and exclusion of household fixed effects are reported in the supplementary online appendix S3. As shown in table S3.1, the model predictions hold when standard errors are clustered along both stratum and wave dimensions (“2-way”) and when clustered at the district level (there are 158 districts), as well as for the cluster wild bootstrap procedure described in Cameron, Gelbach, and Miller (2008) to take into account the small number of clusters when clustering at the “stratum” level. Table S3.2 tests the robustness of the model predictions to alternative variable definitions. In columns (1) and (4), the extension indicators exclude advice that 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 S3.3 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 S3.3 with those in table 4 shows that adding household fixed effects changes the point estimates of the two shock variables but does not change the significance levels much (note that while table S3.3 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 footnote 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 S3.4 in the supplementary online appendix, the considered dependent variables capture specific topics for which farmers demand advice (i.e., the dependent variables are only equal to 1 if the household actively requested advice of a particular category). While agricultural shocks show significant coefficients for all topics except “fishery,” the non-agricultural shock variable is only significant for “agricultural production and processing” and “animal diseases and vaccination.” 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 occurrence (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 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, Evenson, and Feder 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 the supplementary online appendix S4. 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 percent larger value of harvest and 8 percent 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 percent 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, Kremer, and Robinson 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 mechanism 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 adverse 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 observed 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 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 S4.3 in the supplementary online appendix). 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 agricultural advice and extension services center around the idea that participation in extension 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 attention to be a scarce resource. Our model gives rise to several interesting predictions. First, farmers’ decisions to request extension services will depend on the amount of attention 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 that 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, p. 489), 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.” 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, 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, Fischer, and Schoar (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 that 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.

Footnotes

1

According to different sources, there are between half a million and one million agricultural extension workers worldwide, of which 90 percent are located in developing countries (Feder 2005; Anderson and Feder 2007).

2

In our sample, 92 percent 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 percent (84 percent 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) reports 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).

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.

5

The empirical work is based on survey data from the Living Standard Measurement Study - Integrated Surveys on Agriculture (LSMS-ISA) database (World Bank 2018).

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 to 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.

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, p. 493), 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”.

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; Maćkowiak and Wiederholt 2009), consumption dynamics (Luo 2008; Tutino 2013), and business cycles (Maćkowiak and Wiederholt 2015). Further studies apply rational inattention in the areas of finance (Van Nieuwerburgh and Veldkamp 2010; Kacperczyk, Van Nieuwerburgh, and Veldkamp 2016), industrial organization (Sallee 2014; Martin 2017), and labor (Bartos 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, Magruder, and Robinson 2014; Hanna, Mullainathan, and Schwartzstein 2014).

10

For example, in the survey by Birkhaeuser, Evenson, and Feder (1991), 36 of 48 reviewed studies find positive and significant impacts of extension services.

11

Examples of such technologies include mobile phone apps and hotlines that offer farmers the possibility 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 2008).

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 2008).

14

The model draws heavily on existing work in the rational inattention literature, particularly the framework presented by Maćkowiak 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.

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, Rao, and Schilbach 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, Luhan, and Roos 2013; Bartos et al. 2016; Ambuehl, Ockenfels, and Stewart 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).

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 a_N. 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 ϑ_i are correlated across the two domains.

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, Savastano, and Xia (2017), Gollin and Udry (2021), 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 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.

24

Conditional on paying for advice, the average amounts paid are USD 6 in Uganda, USD 9 in Nigeria, and USD 18 in Malawi.

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 (a) several years, and therefore, in our view, too far in the past, and (b) 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 percent 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 percent and 41 percent, respectively.

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 to a negative shock to household income, i.e., marginal utility of additional income increases as part of the income is diverted to the baby.

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).

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).

Notes

Dominik Naeher (corresponding author) is an assistant professor at University College Dublin, Dublin, Ireland; his email address is [email protected]. Matthias Schündeln is a professor at Goethe University Frankfurt, Frankfurt, Germany; his email address is [email protected]. The authors thank the editor Eric Edmonds, anonymous referees, and Mirko Wiederholt for extensive comments and suggestions, and Sandro Ambuehl, Luc Christiaensen, Rema Hanna, and Sara Savastano for helpful discussions. A supplementary online appendix is available with this article at the World Bank Economic Review website.