Get in the Zone: The Risk-Adjusted Welfare Effects of Machine Learning to Define Index Insurance Zones

Insurance programs serve to protect people against downside risks. Within the subset of agricultural insurance that use agronomic conditions to trigger payouts (i.e., index insurance), many programs are designed to protect against risks experienced across a given space, i.e., an index insurance zone. Often, administrative boundaries are used to delineate these zones. However, such boundaries may not necessarily reflect natural breaks in agronomic conditions or account for variation across space. As a result, some farmers may lose out because their yields do not correlate with other farmers in their admin area. Redrawing zone boundaries to group more homogeneous farmers together thus offers the possibility of improving the economic value of index insurance.

To examine how admin boundaries compare to alternatives that use observable agronomic conditions, this work evaluates how zone boundaries influence the expected value of a stylized index insurance program to a given set of smallholder farmers. Drawing on yield data from >5,000 Kenyan maize producers combined with satellite-based estimates of agronomic conditions such as rooting depth and biomass across the growing season from 2016-2019, we use economic approaches to value negative yield deviations -- when people suffer the most -- and estimate changes in value associated with data-driven (e.g., satellite data and cluster based) zones compared with ones based on admin boundaries.

We have found that zones generated with remotely-sensed data offer a similar or slightly lower value proposition to risk averse farmers compared to that of admin zones, assuming a fixed number of zones and no additional cost of adding new zones. However, using data-driven zones opens the door to evaluating zones larger than administrative boundaries as well as considering the welfare trade-offs that these boundaries imply. Changing the number of zones can be economically meaningful, as each new zone often imposes high costs for local yield data collection. These results may influence the viability of privately offered programs or the sampling protocols of publicly subsidized programs to account for variability within administrative regions. We conclude with additional suggestions to improve the value proposition of index insurance to the insured.