Stefan Andréas Sperlich 
Université de Genève

Causal inference with Varying Coefficient Models

Abstract:
The flexibility of semiparametric varying coefficient models is exploited to model heterogeneity that cannot be captured by additive isotonic random deviations from the mean. This is of particular interest for causal analysis when heterogeneity in returns is important. Although they are more restrictive than fully nonparametric methods, varying coefficient models are a powerful compromise between the still popular and typically used linear methods and nonparametrics. Where unobserved heterogeneity is a major concern, they allow to substantially improve in both the interpretation of their estimates and the credibility of the so‐called instrumental methods where applied which presently are predominant in empirical economic studies. To this aim, all considerations and developments should be embedded in the modeling of structural equations for which varying coefficients can play a key role. While spline methods are very popular for computational reasons, kernel based smooth backfitting is the maybe most developed method regarding asymptotic theory.