A1029
Title: Weak instrumental variables due to nonlinearities in panel data: A super learner control function estimator
Authors: Monika Avila Marquez - University of Geneva (Switzerland) [presenting]
Abstract: A triangular structural panel data model is proposed with additive separable individual-specific effects to estimate the causal effect of a covariate on an outcome in the presence of unobservable confounders, some of which are time-invariant. In this context, a linear reduced-form equation may be problematic when the conditional mean of the endogenous covariate and instrumental variables is nonlinear, as ignoring such nonlinearity can lead to weak instruments. To address this, a triangular simultaneous equation model with a linear structural equation and a nonlinear reduced-form equation is introduced. The parameter of interest is the structural coefficient on the endogenous variable. Identification is achieved under standard exclusion restrictions using a control function approach. An estimator called the super learner control function estimator (SLCFE) is developed. The method involves two main steps and sample splitting across the individual dimension. First, the control function is estimated using a super learner; then, this estimate is used to correct for endogeneity in the structural equation. Consistency of the estimator is established, and its performance is evaluated through Monte Carlo simulations. Results show that SLCFE significantly outperforms conventional Within 2SLS estimators, especially when the reduced-form relationship is nonlinear.
