Insurance against weather risk: A quasi-experimental evaluation of weather index insurance impacts on productivity and welfare

Hello! I am Yesuf and I would like to communicate to you my work (co-authored with Théophile T. Azomahou) scheduled to be presented at the 2nd International Conference on Evaluating Climate Change and Development. The title of our study is Insurance against weather risk: A quasi-experimental evaluation of weather index insurance impacts on productivity and welfare. 

A bit of background: In light of the substantial rise in weather risk and the evidence that households suffer from the consequences of weather risk, there are now provisions of weather indexed insurance in many developing countries. The hope is that insurance could alter risk-taking behavior of households leading to investment in high risk but remunerative activities. It could also help households cope with weather shocks without depleting their hard accumulated assets. So, we aimed to evaluate the impact of weather index insurance on production investment, productivity and welfare.

Our study approach: We used a unique cross-section of household data from both insurance purchasers and non-purchasers in five villages in northern Ethiopia, where insurance has been sold for the last four years. The insurance is commercially marketed and farm households can purchase the product by paying the premium. The provision of insurance is not randomly assigned, raising the concern of the self-selection problem (that may be due to observable or unobservable characteristics) in evaluating impact. Hence, we have employed a quasi-experimental evaluation design.

Quasi-experimental evaluation design is one of several evaluation techniques often used when randomized evaluation design is absent either due to feasibility or ethical concerns. Specifically, we used econometric matching and instrument variable (IV) techniques.

In matching, we attempted to statistically construct comparison groups of insurance purchasers and non-purchasers that look alike in terms of their observables. By doing so, we hoped to address the problem of self-selection due to observables. We used propensity score matching to construct our comparison groups and estimated the treatment effects. What if the selection problem is due to unobservables? Then our matching estimates may not redress our concern.  This is what led us to apply the IV technique.

In the IV technique, an important requirement is to find an instrument variable that is related to our treatment indicator but unrelated to the outcome variable. When a credible instrument is available, the approach addresses the selection problem due to unobservables.

Finding an instrument variable is often challenging. We looked for variables that can be used as instrument in our survey data and used ?attendance to insurance training sessions? as an instrument. Why? Because it is correlated with the insurance purchase decision but would hardly have an effect on the outcome variables (production investment, productivity and welfare). Have any thoughts about our selection of an instrument variable? Please comment below and let us know!

Our data: We collected a rich set of information covering standard socio-economic data, detailed production data and issues of risk and insurance. We also elicited risk preference data using survey experiments and measured financial literacy of the households through a series of questions. Then, we used the information collected to estimate the impacts using matching and IV techniques

Our findings: We found that those that purchased insurance increased the use of inorganic fertilizer by almost 25 percent. We also observed a significant rise in farm productivity. Our preliminary results suggest the positive benefits of insurance in terms of altering farmers? production behavior and improving productivity. However, we found no evidence on welfare improvements due to insurance. 

Final words on our study: The quasi-experimental design we followed requires less time and resources and is relatively easier to manage. However, we have imposed different assumptions in order to estimate impacts. For instance in the matching, though we have made efforts to sufficiently control for the most important confounding factors, perhaps we may have not balanced the variables (especially those immeasurable) for both groups sufficiently, leading to some bias in our impact estimates.  While taking the limitations into account, the quasi-experimental design has enabled us to get sensible impact estimates.  

Any thoughts or questions on our study? Please let us know in the comments section below