The practice of mutual insurance is conditioned by two types of transaction costs: "association" costs in establishing links with insurance partners and "extraction" costs in using these links to implement insurance transfers. Data on insurance-motivated water exchanges among households along two irrigation canals in Pakistan show that households exchange bilaterally with neighbors and family members but the majority exchange with members of tightly knit clusters. We, therefore, develop a model that endogenizes both cluster formation and the quality of insurance in the chosen cluster as a function of the relative importance of association and extraction costs. Full insurance at the community level, the object of most empirical tests of mutual insurance, is seen to be an extreme case. It is consequently not surprising that tests of the hypothesis of full risk pooling at the community level have led to rejection. The Pakistan data support the proposition that the configuration of insurance clusters and the intensity of exchanges within clusters vary with association and extraction costs. These costs are affected by kinship, distance to neighbors, and exposure to risk. Households with larger kinship groups, closer neighbors, and greater risk exposure insure through larger clusters and more intensive exchange.