December 1, 1989
It is common in empirical studies using nonlinear models to estimate the mean response by evaluating the nonlinear response function at the mean value of its argument(s). However, this procedure conceptually is flawed if the response function has significant curvature in the neighborhood of the mean. Ideally, one should evaluate the estimated response function for each of the estimated responses. In general, there will be some nonzero approximation error if one instead simply evaluates the response function at the mean of the independent variable(s). Furthermore, if the variability is significant in the independent variable(s) of interest, the approximation error of using the 'evaluate at the mean' procedure increases. This paper examines the magnitude of the approximation error, and attempts to identify situations in which somewhat more computationally intensive procedures should be used.
Read More
|
Download Report