PRICELESS OR PRICEY? “ARBITRARY” CHOICES IN LOG-LINEAR MODELS AND THE “ARBITRARY” COST OF HAVING CHILDREN

Abstract

A popular fix when dealing with zeros in the dependent variable, y, is to add a scalar value, a, within the log transformation, i.e. log (y+a). However, the choice of the scalar value is often seemingly arbitrary. Using data from the Current Population Survey, I step-by-step walk through an empirical investigation of how an additional child in the household affects childcare cost, and I show that the choice of the arbitrary scalar value significantly affects the estimates of a log-linear regression model. For those “special couples” who are mining through data from the Current Population Survey to inform them on life decisions, they can estimate a model to justify any decision by their choice of a. We demonstrate that the best practice may be to forgo the log-linear regression model when dealing with zeros and turn to a Poisson regression.