Reparameterization of nonlinear statistical models: a case study

The importance of finding appropriate parameterizations for nonlinear statistical models is highlighted. The purpose of this paper is to explore the principles of reparameterization, using an example from real data. It is shown that stable parameterizations allow likelihood-based confidence intervals to be computed. Further, it is noted that the choice of error distribution may seriously affect the estimates and confidence intervals of quantities of interest. The influence of each observation on the estimation of each parameter is displayed for each error model. Multidimensional likelihood contours may be displayed pairwise using profile likelihood computations.