balance between model interpretability and accuracy

[dailing616]

Thank you for your presentation! Sometimes improving a model using sophisticated methods would inevitably make the model more complex and often less interpretable. I am wondering how do you usually choose the balance between interpretability and accuracy of your models?

[sdurlauf]

I don't have a good way to describing my internal thought processes on modelling. In my mind, models of the type I work on have two somewhat distinct objectives. One is to provide understanding of how certain mechanisms affect some set of individual outcomes. Here there is no claim about empirical accuracy. In fact, one can think about some models as serving the goal of showing what assumptions are needed for certain propositions to be true (for example Pareto Efficiency of equilibrium), so that empirical work can focus on the assumptions. Other models are meant to provide low dimensional approximation of a high dimension reality. Here there is an interplay between the initial model, empirical and modification of the model. You also raise an important issue with respect to interpretability of models. In my own work, this does not really arise since the models are analytically tractable. But can for more complicated ones. I don't have anything useful to say on how to adjudicate the tradeoff!