Rating and yield predicting procedures for the Group Risk Federal Crop Insurance Program: A nonparametric approach

In 1994, the Group Risk Federal Crop Insurance Program, termed the Group Risk Plan (GRP), was introduced by the Federal Crop Insurance Corporation. Under this program, losses are measured on the basis of an areas' mean yield. The main objective of this dissertation is to improve the accuracy of GRP premium rates, thereby reducing program inefficiency and inequity. In the past, premium rates have been constructed from double exponential smoothing models and the Gaussian distribution. This dissertation employs both nonparametric kernel density estimators and mixture distributions in an attempt to remove any program inefficiencies or inequities induced by distributional assumptions. Seminonparametric maximum likelihood techniques and Stein-rule estimators are employed to exploit population characteristics among county yield distributions.;National Agricultural Statistical Service county mean yield data for barley, corn, cotton, sorghum, soybeans, and, wheat are used in the analysis. Prediction bias was found to be significant with the current double exponential smoothing model. This was not surprising since a high percentage of counties were optimally modelled by a specification nested within an ARIMA (0,1,2) specification. Distributional bias was found as existing rates were statistically different from "pooled" nonparametric rates, Stein nonparametric rates, seminonparametric rates, and empirical rates. For the most part, all nonparametric rates are significantly higher than existing rates. This is to be expected if the underlying distribution is a mixture of two Gaussians where the second distribution lives on the lower tail of the first distribution. Gaussianity test results are also consistent with the conjecture that conditional yield distributions follow a mixture of two Gaussians.;A simulation analysis was undertaken to consider possible ramifications to program efficiency and equity from intertemporal adverse selection responses. The upper bound on long-run loss ratios ranged from 1.25 to 1.61 when the conditional yield distribution is properly defined. However, when the additional error resulting from a two step ahead prediction is not accounted for (current methods), the upper bound on long-run loss ratios ranged from 1.51 to 1.77.