Network-Based Targeting with Heterogeneous
Agents for Improving Technology Adoption
August 2022
Abstract
How do we use existing social ties to improve the adoption of a new technology? I
explore network-based targeting when the benefits from the technology vary at the
household level, with this heterogeneity in benefits affecting the diffusion of
information. I develop a theoretical framework where initially uninformed agents
engage in DeGroot learning to decide whether or not to get fully informed about a new
technology. Conditional on being fully informed, they then decide
whether or not to adopt the technology. The model predicts the possibility of low
information equilibria where nobody will adopt the new technology even if it is the
efficient choice for some of them, and targeting is needed. My simulations suggest that
the optimal targeting strategy in such a scenario relies on the underlying heterogeneity
in the population. If heterogeneity is low in the benefits from the technology, targeting
based on centrality works well. However, if the population is highly heterogeneous,
centrality-based targeting fails in reaching the population of interest. In such a scenario,
targeting based on the probability of adoption works better if the network is highly
assortative in terms of characteristics determining the heterogeneity. I test these
predictions using data from Malawi and provide evidence supporting my theoretical
model. I argue that population heterogeneity in benefits from a technology matters for
the success or failure of alternative targeting strategies that promote that technology.
JEL Codes: D83, O13, O33, Q16
Keywords: Targeting, Social Network, Technology Adoption, Agriculture
1 Introduction
Technology adoption in agriculture is a driving force of economic growth through its effect on
structural transformation (Bustos et al., 2016). However, the adoption of modern
technologies has been low in developing regions, especially in Sub-Saharan Africa (Bold et al.,
2017). Information constraints are one of the key reasons behind such phenomenon
(Magruder, 2018). How do we use existing social ties to improve the adoption of a new
technology? The literature argues that the answer depends on the underlying diffusion
1
,process. If information diffuses only if a certain threshold of each agent’s connections is
informed, targeting based on existing social ties may be required for widespread adoption. In
such a scenario, the literature recommends targeting agents central to the network (Beaman
et al., 2021a). The recommendation, however, is based on the underlying assumption that
the diffusion only depends on the agents’ positions in the network. What happens if the
agents differ in terms of other characteristics that affect the diffusion process?
This paper investigates network-based targeting strategies for improving technology
adoption. In particular, I focus on the situation where the new technology can be more
beneficial to some agents than others, with this heterogeneity in benefits affecting the
diffusion of information. The benefits can vary across agents due to several possible reasons.
The agents can differ in terms of their education, skills, and ability affecting how much they
can learn about a new technology and use it in practice. They can also vary in terms of other
characteristics, e.g., land quality (for agriculture), size of operation (for both farm and firm
households), access to infrastructure (such as road and irrigation facilities), and access to
other technologies. For my purpose, I consider heterogeneity in benefits to reflect the
existing network structure driven by agent sorting according to their observable and
unobservable characteristics. I explore whether the optimal network-based targeting
strategies vary with the extent of heterogeneity within the network. More specifically, I
concentrate on the relative performance of two targeting strategies: targeting based on
centrality and targeting based on probability of adoption.
I develop a theoretical framework where economic agents participate in a two-stage
decision process. In the first stage, uninformed agents decide whether or not to get fully
informed about a new technology. Since information is costly, the agents engage in DeGroot
learning to make this decision. 1 In the second stage, fully informed agents decide whether or
not to adopt the technology. This framework helps me formalize a scenario where pessimism
regarding the prospect of a new technology will lead to low adoption, even if it is efficient for
many agents to adopt.
Based on my theoretical model, I use simulations to evaluate the relative importance of
different targeting strategies and to generate testable hypotheses. 2 I test these predictions by
combining two different data sources from Malawi. The first one is the replication data
(Beaman et al., 2021b) from a randomized controlled trial (RCT) conducted by Beaman,
1 DeGroot learning refers to a social learning process whereby agents form beliefs/ opinions as a weighted
average of the beliefs/ opinions of people they are linked to (including themselves). Here the weights
correspond to how much the agents are influenced by one another. It is a heuristic, as agents do not account
for the interdependence of beliefs between each of the people they are connected to ( Barnett-Howell and
Mobarak, 2021). Chapter 8.3 of Jackson (2010) contains more information on this type of learning.
2 The use of simulations is not new to the network literature. For example, Bala and Goyal (1998) uses
simulations to generate spatial and temporal patterns of adoption when individuals learn from their neighbors;
Acemoglu et al. (2011) uses simulations to show that innovations might spread further across networks with a
smaller degree of clustering. Similar to Beaman et al. (2021a), I use them to understand the effectiveness of
targeting strategies a few periods down the line.
2
,BenYishay, Magruder, and Mobarak (2021a) (henceforth, BBMM). The second dataset is the
Agricultural Extension Services and Technology Adoption Survey (henceforth, AESTAS) data
(IFPRI, 2021a,b) collected by International Food Policy Research Institute (IFPRI). One of the
reasons existing studies made simplifying assumptions on the structure of heterogeneity in
the population is the difficulty in observing heterogeneity in benefits beforehand. As the
benefits are only realized after adoption, they cannot be factored into the targeting
strategies. I attempt to solve this issue by using AESTAS data to estimate adoption conditional
on observable demographics. This way I can categorize the population in terms of their
propensity to adopt a new technology. I calculate households’ probability of adoption in the
BBMM data using estimates from the AESTAS sample. BBMM data is used as their
experiment relies on exploiting the centrality of seeds to improve the adoption of a
technology suitable for my analysis, thus including all other information that I need. I exploit
both the village-level and experimental variations in the BBMM data to test my hypotheses.
My simulations indicate that the relative performance of different targeting strategies
depends on the degree of heterogeneity in a network. Centrality-based targeting strategies
should be less effective in settings where the agents vary significantly in terms of their true
benefits from adopting a technology. In such settings, targeting based on the likelihood of
adoption should perform better if the network is highly assortative in terms of characteristics
determining the benefits. The intuition behind such a result lies in the characteristics of the
central seeds in a network.3 Central seeds are, by definition, the most well-connected people
in a network. Thus, selecting them would maximize the diffusion if diffusion depends only on
the agents’ positions in the network. If agents vary in terms of other characteristics that
affect diffusion, we need to consider this heterogeneity for effective diffusion. Centrality-
based targeting fails to consider this heterogeneity. In an assortative network, central seeds
also represent the average network characteristics. In a setting where a new technology
applies to only a certain sub-section of the population, targeting based on centrality becomes
more likely to fail in reaching the population of interest. Targeting the population of interest
works better in such a scenario.
Reduced form results show evidence in favor of my hypothesis. Exploring villagelevel
variations in the BBMM data, I show that the positive effect of seeds’ centrality on the
adoption of pit planting decrease with an increase in village-level heterogeneity in terms of
probability of adoption. Simultaneously, the negative effect of seeds’ probability of adoption
decreases with increased village-level heterogeneity. Weaker, but similar results are found
when I shift my focus to exploring experimental variations.
My study makes three contributions to the existing literature. First, I provide evidence
(both theoretical and empirical) that the success of network-based targeting strategies
depends on population-level heterogeneity. Diffusion of information via networks is the key
3 In network literature, information entry points are termed seeds.
3
, to increasing technology adoption (Besley and Case, 1993; Foster and Rosenzweig, 1995;
Conley and Udry, 2010; Krishnan and Patnam, 2013). In recent years, there have been several
studies focusing on the role of networks in the diffusion of technologies. 4A growing
proportion of these studies explore the most effective way to use social networks to improve
technology adoption (e.g., Banerjee et al., 2013; BenYishay and Mobarak, 2018). A few of
these studies explore the role of the underlying diffusion process in designing the most
effective targeting policies (e.g., Beaman et al., 2021a; Akbarpour et al., 2021). However,
these studies assume existing network ties to be the only factor characterizing diffusion.
Thus, for diffusion, households are assumed to be homogeneous in terms of other
characteristics. In the current study, I consider the population to be heterogeneous in terms
of the benefits they get from the new technology, with this heterogeneity having a direct
effect on the effectiveness of targeting strategies. In such a scenario, I show evidence that
optimal targeting strategies may differ from the ones prescribed in the existing literature. In
particular, the effectiveness of a targeting policy will vary depending on population-level
heterogeneity in terms of the benefits of the new technology. Considering population-level
heterogeneity in social learning itself is not new (e.g., Munshi, 2004; Bandiera and Rasul,
2006; Conley and Udry, 2010).5However, to the best of my knowledge, the current study is
the first to consider the consequences of population-level heterogeneity on network-based
targeting strategies.
Second, my theoretical framework helps formalize the scenario where agents learn from
their network about a technology that is more beneficial to some of them than others.
Existing studies consider technologies to be equally beneficial to everyone. The adoption may
still differ due to heterogeneity in costs. But these heterogeneous costs are assumed to be
known by the agents and thus do not require learning. 6 Thus, simplifying assumptions are
made such that the learning involves the characteristics that are similar for all the agents and
not the characteristics that differentiate them. This assumption helps us to focus on a
problem where the agents are collectively trying to uncover some hidden characteristics of
interest (e.g., in the theoretical models of Acemoglu et al., 2008 and Golub and Jackson,
2010). More importantly, a consequence of this assumption is that the diffusion of
knowledge regarding the technology depends only on the agent-level heterogeneity in
network ties. In many scenarios, however, agents face heterogeneous benefits in adopting a
new technology (Suri, 2011). For example, the performance of some agricultural practices
may depend on the land quality.7 Thus, the benefits of some technologies may vary
4 See Cheng (2021) for a review of the existing literature.
5 Using the data from Indian Green Revolution, Munshi (2004) finds that information flows are weaker for
rice growers than wheat growers as rice-growing regions are more heterogeneous. Bandiera and Rasul (2006)
observe network effects on technology adoption to vary based on the number of adopters in the network for
sunflower production in Mozambique. Conley and Udry (2010) finds that only novice farmers learn from their
veteran neighbors about the use of fertilizers for pineapple production in Ghana. 6Even if the heterogeneous
costs are unknown to the agents, there is no possibility of learning from
4
Agents for Improving Technology Adoption
August 2022
Abstract
How do we use existing social ties to improve the adoption of a new technology? I
explore network-based targeting when the benefits from the technology vary at the
household level, with this heterogeneity in benefits affecting the diffusion of
information. I develop a theoretical framework where initially uninformed agents
engage in DeGroot learning to decide whether or not to get fully informed about a new
technology. Conditional on being fully informed, they then decide
whether or not to adopt the technology. The model predicts the possibility of low
information equilibria where nobody will adopt the new technology even if it is the
efficient choice for some of them, and targeting is needed. My simulations suggest that
the optimal targeting strategy in such a scenario relies on the underlying heterogeneity
in the population. If heterogeneity is low in the benefits from the technology, targeting
based on centrality works well. However, if the population is highly heterogeneous,
centrality-based targeting fails in reaching the population of interest. In such a scenario,
targeting based on the probability of adoption works better if the network is highly
assortative in terms of characteristics determining the heterogeneity. I test these
predictions using data from Malawi and provide evidence supporting my theoretical
model. I argue that population heterogeneity in benefits from a technology matters for
the success or failure of alternative targeting strategies that promote that technology.
JEL Codes: D83, O13, O33, Q16
Keywords: Targeting, Social Network, Technology Adoption, Agriculture
1 Introduction
Technology adoption in agriculture is a driving force of economic growth through its effect on
structural transformation (Bustos et al., 2016). However, the adoption of modern
technologies has been low in developing regions, especially in Sub-Saharan Africa (Bold et al.,
2017). Information constraints are one of the key reasons behind such phenomenon
(Magruder, 2018). How do we use existing social ties to improve the adoption of a new
technology? The literature argues that the answer depends on the underlying diffusion
1
,process. If information diffuses only if a certain threshold of each agent’s connections is
informed, targeting based on existing social ties may be required for widespread adoption. In
such a scenario, the literature recommends targeting agents central to the network (Beaman
et al., 2021a). The recommendation, however, is based on the underlying assumption that
the diffusion only depends on the agents’ positions in the network. What happens if the
agents differ in terms of other characteristics that affect the diffusion process?
This paper investigates network-based targeting strategies for improving technology
adoption. In particular, I focus on the situation where the new technology can be more
beneficial to some agents than others, with this heterogeneity in benefits affecting the
diffusion of information. The benefits can vary across agents due to several possible reasons.
The agents can differ in terms of their education, skills, and ability affecting how much they
can learn about a new technology and use it in practice. They can also vary in terms of other
characteristics, e.g., land quality (for agriculture), size of operation (for both farm and firm
households), access to infrastructure (such as road and irrigation facilities), and access to
other technologies. For my purpose, I consider heterogeneity in benefits to reflect the
existing network structure driven by agent sorting according to their observable and
unobservable characteristics. I explore whether the optimal network-based targeting
strategies vary with the extent of heterogeneity within the network. More specifically, I
concentrate on the relative performance of two targeting strategies: targeting based on
centrality and targeting based on probability of adoption.
I develop a theoretical framework where economic agents participate in a two-stage
decision process. In the first stage, uninformed agents decide whether or not to get fully
informed about a new technology. Since information is costly, the agents engage in DeGroot
learning to make this decision. 1 In the second stage, fully informed agents decide whether or
not to adopt the technology. This framework helps me formalize a scenario where pessimism
regarding the prospect of a new technology will lead to low adoption, even if it is efficient for
many agents to adopt.
Based on my theoretical model, I use simulations to evaluate the relative importance of
different targeting strategies and to generate testable hypotheses. 2 I test these predictions by
combining two different data sources from Malawi. The first one is the replication data
(Beaman et al., 2021b) from a randomized controlled trial (RCT) conducted by Beaman,
1 DeGroot learning refers to a social learning process whereby agents form beliefs/ opinions as a weighted
average of the beliefs/ opinions of people they are linked to (including themselves). Here the weights
correspond to how much the agents are influenced by one another. It is a heuristic, as agents do not account
for the interdependence of beliefs between each of the people they are connected to ( Barnett-Howell and
Mobarak, 2021). Chapter 8.3 of Jackson (2010) contains more information on this type of learning.
2 The use of simulations is not new to the network literature. For example, Bala and Goyal (1998) uses
simulations to generate spatial and temporal patterns of adoption when individuals learn from their neighbors;
Acemoglu et al. (2011) uses simulations to show that innovations might spread further across networks with a
smaller degree of clustering. Similar to Beaman et al. (2021a), I use them to understand the effectiveness of
targeting strategies a few periods down the line.
2
,BenYishay, Magruder, and Mobarak (2021a) (henceforth, BBMM). The second dataset is the
Agricultural Extension Services and Technology Adoption Survey (henceforth, AESTAS) data
(IFPRI, 2021a,b) collected by International Food Policy Research Institute (IFPRI). One of the
reasons existing studies made simplifying assumptions on the structure of heterogeneity in
the population is the difficulty in observing heterogeneity in benefits beforehand. As the
benefits are only realized after adoption, they cannot be factored into the targeting
strategies. I attempt to solve this issue by using AESTAS data to estimate adoption conditional
on observable demographics. This way I can categorize the population in terms of their
propensity to adopt a new technology. I calculate households’ probability of adoption in the
BBMM data using estimates from the AESTAS sample. BBMM data is used as their
experiment relies on exploiting the centrality of seeds to improve the adoption of a
technology suitable for my analysis, thus including all other information that I need. I exploit
both the village-level and experimental variations in the BBMM data to test my hypotheses.
My simulations indicate that the relative performance of different targeting strategies
depends on the degree of heterogeneity in a network. Centrality-based targeting strategies
should be less effective in settings where the agents vary significantly in terms of their true
benefits from adopting a technology. In such settings, targeting based on the likelihood of
adoption should perform better if the network is highly assortative in terms of characteristics
determining the benefits. The intuition behind such a result lies in the characteristics of the
central seeds in a network.3 Central seeds are, by definition, the most well-connected people
in a network. Thus, selecting them would maximize the diffusion if diffusion depends only on
the agents’ positions in the network. If agents vary in terms of other characteristics that
affect diffusion, we need to consider this heterogeneity for effective diffusion. Centrality-
based targeting fails to consider this heterogeneity. In an assortative network, central seeds
also represent the average network characteristics. In a setting where a new technology
applies to only a certain sub-section of the population, targeting based on centrality becomes
more likely to fail in reaching the population of interest. Targeting the population of interest
works better in such a scenario.
Reduced form results show evidence in favor of my hypothesis. Exploring villagelevel
variations in the BBMM data, I show that the positive effect of seeds’ centrality on the
adoption of pit planting decrease with an increase in village-level heterogeneity in terms of
probability of adoption. Simultaneously, the negative effect of seeds’ probability of adoption
decreases with increased village-level heterogeneity. Weaker, but similar results are found
when I shift my focus to exploring experimental variations.
My study makes three contributions to the existing literature. First, I provide evidence
(both theoretical and empirical) that the success of network-based targeting strategies
depends on population-level heterogeneity. Diffusion of information via networks is the key
3 In network literature, information entry points are termed seeds.
3
, to increasing technology adoption (Besley and Case, 1993; Foster and Rosenzweig, 1995;
Conley and Udry, 2010; Krishnan and Patnam, 2013). In recent years, there have been several
studies focusing on the role of networks in the diffusion of technologies. 4A growing
proportion of these studies explore the most effective way to use social networks to improve
technology adoption (e.g., Banerjee et al., 2013; BenYishay and Mobarak, 2018). A few of
these studies explore the role of the underlying diffusion process in designing the most
effective targeting policies (e.g., Beaman et al., 2021a; Akbarpour et al., 2021). However,
these studies assume existing network ties to be the only factor characterizing diffusion.
Thus, for diffusion, households are assumed to be homogeneous in terms of other
characteristics. In the current study, I consider the population to be heterogeneous in terms
of the benefits they get from the new technology, with this heterogeneity having a direct
effect on the effectiveness of targeting strategies. In such a scenario, I show evidence that
optimal targeting strategies may differ from the ones prescribed in the existing literature. In
particular, the effectiveness of a targeting policy will vary depending on population-level
heterogeneity in terms of the benefits of the new technology. Considering population-level
heterogeneity in social learning itself is not new (e.g., Munshi, 2004; Bandiera and Rasul,
2006; Conley and Udry, 2010).5However, to the best of my knowledge, the current study is
the first to consider the consequences of population-level heterogeneity on network-based
targeting strategies.
Second, my theoretical framework helps formalize the scenario where agents learn from
their network about a technology that is more beneficial to some of them than others.
Existing studies consider technologies to be equally beneficial to everyone. The adoption may
still differ due to heterogeneity in costs. But these heterogeneous costs are assumed to be
known by the agents and thus do not require learning. 6 Thus, simplifying assumptions are
made such that the learning involves the characteristics that are similar for all the agents and
not the characteristics that differentiate them. This assumption helps us to focus on a
problem where the agents are collectively trying to uncover some hidden characteristics of
interest (e.g., in the theoretical models of Acemoglu et al., 2008 and Golub and Jackson,
2010). More importantly, a consequence of this assumption is that the diffusion of
knowledge regarding the technology depends only on the agent-level heterogeneity in
network ties. In many scenarios, however, agents face heterogeneous benefits in adopting a
new technology (Suri, 2011). For example, the performance of some agricultural practices
may depend on the land quality.7 Thus, the benefits of some technologies may vary
4 See Cheng (2021) for a review of the existing literature.
5 Using the data from Indian Green Revolution, Munshi (2004) finds that information flows are weaker for
rice growers than wheat growers as rice-growing regions are more heterogeneous. Bandiera and Rasul (2006)
observe network effects on technology adoption to vary based on the number of adopters in the network for
sunflower production in Mozambique. Conley and Udry (2010) finds that only novice farmers learn from their
veteran neighbors about the use of fertilizers for pineapple production in Ghana. 6Even if the heterogeneous
costs are unknown to the agents, there is no possibility of learning from
4
