In explaining wage or income by personal attributes (e.g. educational attainment, age, and ethnicity) in a regression model, many researchers choose to use the log of wage or income as the dependent variable and then to estimate the unknown coefficients by some version of the least-squares method. We call this approach the conventional approach. Using the micro data of the 2005-2007 American Community Survey and Taiwan's 2001-2010 Manpower Utilization Survey, we show that the conventional approach has the serious shortcoming of under-predicting the observed wage structure in the space spanned by the values of the explanatory variables. In addition to revealing the reason for the under-prediction problem and linking the severity of this problem to wage variability, we present a nonlinear approach that does not have this shortcoming. We also offer a SAS module for carrying out the estimation task in the nonlinear approach.
uneven distortions log-transformation of dependent variable regression model wage structure income nonlinear estimation