Non-linear regression with discrete explanatory variables, with an application to the earnings function (Q913417)

The paper presents an estimation technique applicable to nonlinear regression with discrete explanatory variables: a problem very often faced in social sciences. It has been shown that a nonlinear regression involving discrete variables can be expressed as a finite polynomial of a linear function of the regressions employing linear separators whose existence is always ensured. However, a proper choice of linear separators is essential. A linear separator is considered to be the best, if it generates or yields the polynomial whose order is the least one. A two-stage estimation procedure has been developed and various model specification tests are also proposed for testing degree of the polynomial and existence of linear separators. Considering an earning function, the procedure is to start with an a priori degree of the polynomial and to apply the Wald-test to reduce the degree. This procedure is sequential in nature until the degree of the polynomial is accepted. In this manner, the degree of the polynomial is determined by the data. The cross-validation approach might be a useful alternative though its consistency has not been tested.