Behind the curve and beyond

Parameter coefficients from non-linear models are inherently difficult to interpret, and scholars frequently opt for computing and comparing predicted probabilities for variables of interest. In an influential article, Hanmer and Ozan Kalkan (2013) discuss the two most common approaches, the average case respectively observed values approach, and make a strong case for the latter. In this paper, I propose a further refinement of the observed values approach for the purpose of computing predicted probability changes. This refinement concerns the use of counterfactual values for the independent variable of interest. I demonstrate that accounting for non-linearities with regards to the variable of interest is important to avoid estimation biases. I also discuss the implications of this insight for estimating average treatment effects from observational data.