Data analytics
Prof. Dr. habil. Rouven E. Haschka
Nonparametric Bayesian Inference for Copula-Based Endogeneity Corrections
DFG project (project number: 568088747)
Causal inference is a central topic in empirical research, especially when endogenous regressors are present. Classical approaches to correct for endogeneity are based on instrumental variables (IV), but require strong and valid instruments - which are often hard to find in practice.
This project develops an IV-free approach to correct for endogenous bias by modeling the joint distribution of the structural error and the endogenous regressors non-parametrically Bayesian using a copula function. While parametric methods such as the frequentist copula approach by Park and Gupta (2012) have mainly been used to date, the proposed approach deliberately dispenses with such restrictions.
The distribution of the structural error is modeled flexibly via Dirichlet process-mixture priors, while the copula function is captured by an infinite mixture. This eliminates the need for restrictive distributional assumptions, and estimation uncertainty can be fully accounted for. Advances in Markov chain Monte Carlo simulation also enable efficient practical implementation without the need for asymptotic approximations.
The project thus offers a general, robust and flexible methodology for IV-free endogeneity correction and contributes to the further development of modern causal inference methods.