This is the eighth in this year’s series of posts by PhD students on the job market.
Researchers and development practitioners often believe small, rural villages throughout the developing world are tightly-knit communities where everybody knows everything about their neighbors. Notably, we often suppose that people in these villages know when their neighbors fall on hard times (like a poor harvest or illness of a household earner), and then may even use this information to assist others within the community.
Policymakers who are designing social protection programs may find it useful to know who is neediest right now in order to better target benefits. Given that household income is notoriously hard to measure directly, policymakers normally use easily observable proxies (such as asset ownership and household demographics) to predict which households are currently most in need. But these proxies tend to be fairly static -- after all, whether or not a family owns a house doesn’t change often -- and therefore may not provide information on transitory changes in income. In this case, asking the community what they know about who is currently neediest could help policymakers target benefits to households facing transitory shocks, if community members can indeed provide more “up-to-date” information on their neighbors’ welfare.
It is intuitive to believe that people can easily observe their neighbors’ current welfare, yet we have little direct evidence this is true. While some researchers have found a positive correlation between the households that a community identifies as “poorest” and survey-based measures of consumption, which may capture transitory income changes, (e.g., Alderman 2002, Alatas et al. 2012), it is hard to tease out whether community members are simply relying on information about their neighbors’ observable assets, which is also correlated with survey-measured consumption.
In my job market paper, my coauthors and I directly elicit the welfare information community members have about their neighbors in Central Java, using a survey and experimental tasks. Crucially, instead of inferring what information the community has by simply comparing the list of beneficiaries selected by the community to those selected by various survey-measured benchmarks, we ask community members to directly rank their neighbors based on specific benchmarks. We ask participants to provide information about two types of benchmarks: 1) those that likely capture the transitory welfare shocks less observable to centralized policymakers (per capita expenditures, neediness, and recent shocks faced), and 2) assets, which capture longer-term welfare. We then compare this directly elicited information to community reports of which households should receive benefits, to understand which types of welfare attributes (static or transitory) community members use to suggest program beneficiaries to policymakers.
Setting
The participants are 300 randomly-chosen families in 10 randomly-chosen communities in the Purworejo Regency of Central Java, Indonesia. Communities have 50 to 70 households, are socially homogeneous (Muslim, ethnically Javanese), and farm rice as the main source of income. Community members are socially engaged, with 97% reporting participation in at least one community organization. Importantly, community members also know each other well; when participants were asked how well they know 20 randomly chosen families in their community, on average they identified 13.1 as families they know very well, and have at least some familiarity with almost all families.
Exercise
We conducted a household visit to each family, during which a participant completed a survey and a set of experimental tasks. The survey asked participants to report on their own consumption, asset ownership, and recent negative shocks. Directly afterward, we also asked participants to report on whether three other randomly chosen families in their community had recently faced a negative shock. If the participant indicated that another family had indeed experienced a shock, we asked them to identify the shock type (e.g., poor harvest, household member’s death).
After completing the survey, participants completed 4 experimental tasks in which they ranked 10 other families in their community in response to various prompts and incentive schemes. We chose the set of 10 families randomly, but selected more families the participant reported knowing well. We kept the same set of 10 families in each task. The 4 tasks are as follows:
● Targeting Task:
o Prompt: “Rank families from ‘poorest’ to ‘richest.’” (We do not suggest any specific welfare benchmark/criteria for participants to use.)
o Incentive: A random number of households ranked as “poorest” will receive a $1 cash transfer (a program benefit).
● Information Elicitation Tasks:
o Prompts: Rank families based on various welfare benchmarks:
▪ Consumption: Rank from “Spends least per person on food per week” to “Spends most per person on food per week”
▪ Neediness: Rank from “Most needy of additional money” to “Least needy of additional money”
▪ Assets: Rank from “Least asset wealth” to “Most Asset Wealth”
o Incentives: The participant herself receives a payoff that is increasing in the accuracy of the information provided, specifically, in the level of concordance between her reported ranking based on a benchmark and the ranking we calculate using the self-reported survey data for the same set of families under the same benchmark. Specifically, we use the survey data to calculate:
▪ Consumption: Total per capita expenditures on food
▪ Neediness: An estimated index of marginal utility of expenditures, constructed using data on consumption expenditures and household demographic characteristics (Ligon 2020).
▪ Assets: A principal component-based asset index (and total value of land owned for robustness).
We calculate concordance between a participant’s ranking and the survey-based ranking of the same families using the Spearman rank correlation coefficient, which quantifies the similarity between the two ordinal rankings. A Spearman coefficient of 1 would mean that the participant and survey rank the families in the same order, and a Spearman coefficient of -1 would mean that the participant and survey rank the families in the exact opposite order.
Result 1: Individuals have little dynamic information about their neighbors
Overall, the evidence suggests that community members have little “dynamic” welfare information about their neighbors. Figure 1 presents the average Spearman correlation between participant-assessed and survey-assessed rankings. If we take the survey as the “truth,” these low correlations suggest that, on average, participants have very little welfare information of any kind about others. Notably, the average correlations for dynamic benchmarks, like consumption and neediness, are significantly lower than those for the static asset benchmark, regardless of whether we compare participant asset rankings to the survey-collected asset index or total land value.
Additionally, participants did not know whether the three randomly chosen families they were asked about during the survey had faced a recent shock. In 43% of cases, participants refused to even offer a report because they “did not know.” Of those that did report, participants were no better at identifying which families faced a shock than random guessing.
Result 2: Individuals use information about each other’s assets to target
Next, we compare participants’ responses in the Targeting and Information Elicitation tasks to understand which types of welfare information participants actually consider when identifying poor households to receive benefits. We find that participants mostly use information about assets when making targeting decisions. Figure 2 shows that the Spearman correlation between participant-assessed assets and targeting rankings is 0.89, significantly higher than for the consumption and neediness benchmarks (though participants’ consumption and neediness rankings are generally correlated with their asset rankings). In qualitative justifications of their targeting task decisions, participants mostly reference differences in land ownership, a vital productive asset in this context.
Policy Implications
Our results suggest that in some local communities, community members may not actually have much welfare information about others. This is especially true for the more dynamic, shock-sensitive benchmarks for which we might think the community would have a comparative advantage over a remote policymaker. This result has potentially important implications. For example, Albania and Côte d’Ivoire used community-based targeting to help target cash transfers to beneficiaries in the wake of the COVID-19 pandemic (Gentilini et al. 2021). If community information is not as useful for targeting as we previously thought, we may want to consider other methods when targeting similar programs that help individuals smooth transitory shocks. At the same time, asking community members may still be an inexpensive way to capture information on longer-term poverty in a way that allows for community agency in policy decisions. Additionally, our results may depend on the setting, pointing to the need for further work on the information contained in community-based targeting in different contexts.
Carly Trachtman is a 6th year PhD candidate at the University of California, Berkeley. They will be on the job market in the 2021-2022 year.
Join the Conversation