A new specification of health-insurance-plan choice is developed that uses a nonparametric functional form for the loss function used to evaluate insurance premiums and uncertain out-of-pocket expenditure. The approach is implemented on data resulting from one firm's implementation of a new health plan with three distinct health insurance options. Health expenditure differences greater than 500 percent are observed, indicating extremely strong biased selection. Plan choices are consistent with a convex loss function for small losses, which suggests that consumers underweight high-loss/low-probability outcomes relative to low-loss/high-probability ones. Copyright 1989 by MIT Press.