In Simple Words
Think of trying to test how a new vaccine affects an entire town's virus spread by giving it to random families. If you just picked bigger towns to give more vaccines to, you couldn't tell if the results came from the vaccine or town size. Instead, by randomly selecting specific social clubs (castes) of varying sizes within each town and offering vaccines only to members of those chosen clubs, you create random variations in total town coverage without biasing the test by town size.
Observed insurance-offer coverage
- Cultivator households
- Village range: 0%–53%
- Landless worker households
- Village range: 0%–100%
Observed design coverage across villages, not insurance purchase or outcome effects. Eligibility was limited by caste size.
See source →Research question
Estimating general equilibrium market price spillovers requires random variation in product market density across communities without introducing endogenous correlations with village population size or social structure.
Participants
118 unique castes (jatis) with >50 households across 42 randomized treatment villages out of 63 REDS survey villages in India.
Study Design
Two-stage randomized design stratifying insurance offers by sub-caste population thresholds (>50 members), assigning 93 treatment castes and 25 control castes to create exogenous variation in village treatment density (0% to 53% for cultivators; 0% to 100% for laborers).
Results
Conditioning on eligibility population shares eliminates endogeneity bias, reducing correlations between insurance marketing proportions and village characteristics (such as caste concentration and total households) to statistically insignificant levels (p-values > 0.33).
Takeaway
Design general equilibrium field experiments by randomizing treatment across social network sub-units of varying sizes to generate exogenous community-level coverage variation.