Information Diffusion in Agricultural Social Networks under Externalities: A Pest in Honduras

An individual/personal training session with selected farmers in communities affected y the yellow aphid in Honduras. These trainings are aimed to transmit three key messages or lessons: i) how to identify and monitor the yellow aphid in sorghum and maize, ii) how to control the yellow aphid once it has affected the crop, and iii) how do pest spatial externalities affect the community, not just the individual, and how can they be prevented through cooperation.
Solutions will be shared with the participants through multimedia and interactive learning activities that are adapted to the local context. We will also provide each participant with the tools to implement these solutions: a calendar that contains the key messages from the training and serves as a tool to keep their scouting records, and yellow traps. In addition, each training participant receives a set of calendars and yellow traps for identified members in their community, to which we ask them to deliver them to and communicate the lessons learned during the training.

2018-08-01

2019-10-31

Identification and Knowledge: Correctly identifies yellow aphid from different images, Names unique characteristics of yellow aphid

Use of Solutions: Monitoring, alertness and diligence: Reports performing regular scouting for yellow aphid, Keeps a yellow aphid scouting, record on their calendar, Estimates number of yellow aphid from images, Correctly identifies number of areas and places were to perform yellow aphid scouting

Use of Resources: Yellow traps and home-made insecticide solution, Identifies yellow traps and detergent solution as effective measures to control yellow aphids, Correctly knows how install yellow traps (spacing and placement), Correctly identifies ingredients and quantities to make sustainable detergent solution to use as insecticide, Installed yellow traps in the previous agricultural season, Applied detergent solution as insecticide if exposed to yellow aphid, Knowledge of management of infected plants and weeds to minimize further contagion

Collective action and externalities: Received information or heard of yellow aphid infestation in the community from neighbors, Provided information of yellow aphid infestation in the community to their neighbors, Received training and calendar tool package from an injection point trained in the experiment, Received training, calendar tool package, or information on yellow aphid from a non-injection point (other trained by an injection point of the experiment), Used calendar tool to record neighbors s/he informed and/or trained in the previous agriculture season

Effectiveness of training in final outcomes: Probability of having had a yellow aphid problem in the past agricultural season, Probability of planting and not harvesting due to yellow aphid, Quantity of sorghum lost on field and quantity harvested, Sorghum productivity or yields, Food security: Probability of being hungry and availability of grains in the household

To compare the effectiveness of the different methods on reaching most farmers and achieve coordination, we will test three different injection points based on defined characteristics and positionality in the social networks. These three injection points will be compared with a group of communities where random farmers initially receive the information. In addition, these four treatment conditions will be compared to a control group, where no farmer will receive the training.
The network characteristics that define the injection points are:
a. Local Leaders: social capital and institutions
b. Diffusion central: network theory/statistics
c. Community nominated: the people’s voice
d. Random farmers: natural diffusion

Local Leaders: social capital and institutions arm
For this treatment arm we will rely on existing institutions on these communities, which are a natural benchmark as they are the typical entry points in traditional extension programs. For such local institutions we will consider existing local organizations, such as political committees or water joints. The injection points will be the presidents of such organizations. These leaders are easier and less costly to approach, given that the institutional structure is already in place. Local leaders are appointed by political reasons. They may or may not have a central positionality in the local network. Thus, the scope of their message diffusion could be outperformed by more central individuals.
Diffusion central: network theory/statistics arm
For this treatment arm we calculate a centrality measure for all network members and identify the members that appear to be the most central. Different centrality measures have been studied and tested in the literature and for this study we use the centrality measure proposed by Banerjee et al (2013). Banerjee et. al (2013) develop a model to measure the effectiveness of every node as an injection point to increase diffusion of a newly available microfinance program. The authors find that the centrality measure they call “communication centrality” is the one that most strongly predicts adoption. This communication centrality is costly to estimate, so the authors develop an easily computable proxy that they call “diffusion centrality”. Banerjee et al (2013) show that diffusion centrality can be calculated by: DC(g;q,T)=[∑_(t=1)^T▒(qg)^t ].1 , where T are the iterations of information passing, q the probability of information passing, and g the adjacency matrix. If T=1, diffusion centrality is proportional to degree centrality. As T→∞, it becomes proportional to Katz-Bonacich centrality or eigen-vector centrality, depending on whether q is smaller or not than the inverse of the first eigen-value of the adjacency matrix. In the intermediate region of T, the indicator differs from existing measures. Any method that does not rely on the estimation of their model requires an appropriate choice of q. The authors suggest an intermediate value of q, given by the inverse of the first eigenvalue of the adjacency matrix, λ_1 (g). This is, the critical value of q for which the entries of (qg)^Ttend to 0 as T grows if q < 1/λ_1 and some entries diverge if q>1/λ_1.
We calculate this centrality measure for all network members and sort them by this measure. We have information on a sample of the network from 20 randomly selected farmers in each community (where each community has between 20 and 140 farmers). Given that we do not have the full network information, the measure of centrality might be biased. However, the rank order of the farmers in our observed network is preserved and, for identifying the injection points, this is a sufficient statistic.
Community nominated: the people’s voice arm
For this treatment arm we will check how does asking people to tell us who will be the best person to disseminate information compares to the other treatments. To obtain such information, we will ask members of the community who they think is the most suited farmer to spread out information about a new technological innovation. Banerjee et. al (2016) find that asking a random sample of individuals to nominate best suited people to spread information highly correlates with diffusion centrality; and that these nominees are more central than traditional leaders and geographical central individuals.
Given that we deal with small communities, individuals may even know about other important characteristics of their fellow farmers. For example, they might know of farmers who are more prone to innovate, have more experience, know better about agriculture, are more charismatic; all important characteristics that might influence diffusion and adoption. The idea is that this treatment arm can incorporates both centrality and such other (unobservable to the researcher) characteristics.
Random farmers: natural diffusion arm
To have a benchmark for this comparison, we will use random farmers as injection points. Given that we have network information for each community, in a group of them we will randomly choose a farmer to deliver the information to and compare its message reach with the other groups. How effective this group performs is a central part of this study.

Randomization done in office computer on baseline data.

Communities (caseríos)

Yes

221 communities

3,025

a. Local Leaders: social capital and institutions: 28 communities
b. Diffusion central: network theory/statistics: 28 communities
c. Community nominated: the people’s voice: 28 communities
d. Random farmers: natural diffusion: 27 communities
e. Pure control: 110 communities

We calculated the minimum detectable effect on a set of outcome variables: whether the farmer has faced infestation of any of the two pests, if farmer is able to identify them, if correctly identifies the pest (from pictures presented), if mentions some measure to control such pest, total crop production and total crop area. We varied the number of clusters per treatment arm and found that using 25 clusters (or communities) per treatment arm will allow us to detect a change of between 4 to 11 percent points in the mentioned binary outcome variables.