The role of experience and complementarity in technology adoption: Evidence from long-range monsoon forecasts

In 2022 we randomized the provision of forecasts to farmers, and benchmarked the impact of forecasts against that of insurance. Specifically, we worked in 250 villages, assigning 100 villages to receive forecasts, 50 to receive insurance, and 100 to be part of the control group. We selected 5 farmers within each village to be part of our sample. In 2023, we ran the study again. We continued to work in the same 250 villages, but we changed our experimental design. First we assigned half of the control group to receive forecasts for the first time, and we assigned half of the insurance group to receive forecasts an} insurance. The original 100 villages assigned to forecasts continued to receive forecasts for a second year.

2023-05-01

2023-11-30

Demand for forecasts and insurance; farmer beliefs about the monsoon's onset; \emph{ex ante} outcomes; and \emph{ex post} outcomes. Please see PAP doc attached

Demand for forecasts and insurance; farmer beliefs about the monsoon's onset; \emph{ex ante} outcomes; and \emph{ex post} outcomes. Please see PAP doc attached

Please see PAP doc attached

Please see PAP doc attached

In Year 1 we randomized 250 villages (sampling 5 farmers each) in Telangana into one of two treatment arms and a control group. In treatment 1 villages (100), farmers received forecasts, while in treatment 2 villages (50), farmers received insurance as a benchmark. Comparing Treatment 1 to control identifies the impact of forecasts relative to the control group. Comparing Treatment 2 to Treatment 1 identifies the impact of insurance relative to the forecast group, and allows us to benchmark the impact of the forecast relative to another well-known risk-mitigation technology.

We adjust our experimental design in the next agricultural season to understand 1) whether the provision of forecasts has a larger impact when farmers receive them a second time, 2) whether the provision of forecasts affects insurance take-up, and 3) whether forecasts and insurance act as complements or substitutes. First, we provide half of the Year 1 control group with access to forecasts. Second, we provide half of the insurance group from year 1 with access to forecasts. We can make three interesting comparisons with this design. First, we can compare farmers who receive forecasts two years in a row to those who receive it for the first time. This identifies the impact of receiving forecasts two years in a row, and establishes whether farmers who trust the forecast more make larger adjustments to their agricultural practices. Second, we can compare farmers who receive insurance to those who receive forecasts and insurance, to identify how the forecast offer affects farmers’ demand for insurance. Third, we can compare farmers who receive forecasts \emph{and} insurance to those who receive forecasts \emph{or} insurance to identify whether farmers treat the products as complements or substitutes.

In Year 1 we randomized 250 villages (sampling 5 farmers each) in Telangana into one of two treatment arms and a control group. In treatment 1 villages (100), farmers received forecasts, while in treatment 2 villages (50), farmers received insurance as a benchmark. Comparing Treatment 1 to control identifies the impact of forecasts relative to the control group. Comparing Treatment 2 to Treatment 1 identifies the impact of insurance relative to the forecast group, and allows us to benchmark the impact of the forecast relative to another well-known risk-mitigation technology.

We adjust our experimental design in the next agricultural season to understand 1) whether the provision of forecasts has a larger impact when farmers receive them a second time, 2) whether the provision of forecasts affects insurance take-up, and 3) whether forecasts and insurance act as complements or substitutes. First, we provide half of the Year 1 control group with access to forecasts. Second, we provide half of the insurance group from year 1 with access to forecasts. We can make three interesting comparisons with this design. First, we can compare farmers who receive forecasts two years in a row to those who receive it for the first time. This identifies the impact of receiving forecasts two years in a row, and establishes whether farmers who trust the forecast more make larger adjustments to their agricultural practices. Second, we can compare farmers who receive insurance to those who receive forecasts and insurance, to identify how the forecast offer affects farmers’ demand for insurance. Third, we can compare farmers who receive forecasts \emph{and} insurance to those who receive forecasts \emph{or} insurance to identify whether farmers treat the products as complements or substitutes.

Randomization done in office by a computer

Village

Yes

250 villages

1250 (5 households per village)

1250

See Burlig et al. (2022)