Risk or Loss Aversion? Evidence from Personnel Records

I use data on individual pay and output from Safelite Corp. to estimate a principal-agent model with moral hazard in which the agent has preferences that approximate various specifications of reference dependence, diminishing sensitivity, and loss aversion, as well as the standard case of CARA utility. Estimating the preference parameters via GMM on the incentive and participation constraints shows that the specification with both loss aversion and diminishing sensitivity fits the data much better that any of the other nested specifications. In addition, goodness-of-fit simulations suggest that the model with loss aversion and diminishing sensitivity has root mean squared error that is about 18 percent smaller than the model where preferences approximate the CARA type.