On Not Interpreting Coefficients

Abstract

A recent recommendation by Holt that coefficients resulting from estimating log-linear and similar models should not be interpreted is argued to be based on lack of clarity about the substantive and theoretical importance of the choice between dummy and effect coding for categorical variables. As a consequence, his recommendation applies only when the main concern is for a balance between parsimony and goodness of fit to the data and is inappropriate when interest lies in correct specification of a model representing the substantive process generating the data.