Events

Dr. Kostas Florios
Postdoctoral Fellow at Athens University of Economics and Business
Consultant Engineer at n & k Technology

In this talk, we focus on two leading examples of semiparametric estimators, namely the maximum score estimator (MS) and the censored least absolute deviations (CLAD) estimator, handling the complicated computational issues that arise during their estimation, such as multiple local optima. There is much value in computing these and other similar estimators exactly, in the meaning of having an optimality guarantee for the obtained point estimates. If not, the statistical properties of the estimators are questionable, since they depend on having obtained the global solution to the optimization problem at hand. A suitable exact method to solve optimization problems that correspond to the estimation procedure of similar estimators is the Mixed Integer Programming approach (MIP). MIP was first introduced in econometric estimation in 2008 by our team for the case of the exact computation of the maximum score estimator. A second application of MIP to econometric estimation was published more recently for the case of the CLAD estimator in 2019 again by our team. Since 2008, other researchers have proposed similar MIP models for related estimators, such as the Sum of Absolute Deviations estimator or the maximum rank correlation estimator. Also, variable selection has been addressed by other researchers in the context of the maximum score estimator. In this talk we will review the MIP approach as applied to the MS and CLAD estimator computation problems. Then, we will briefly discuss metaheuristics (e.g., tailored variants of Simulated Annealing and/or Tabu Search) that do not provide a theoretical guarantee of optimality for the obtained estimates but can handle large sample size and/or number of estimated parameters, beyond the current practical capability of MIP. Finally, we will discuss open issues on which our team works on, such as the development of a MIP model for the computation of the CLAD estimator for panel data with fixed effects or the integration of the LASSO operator within the CLAD estimator computation. For most of these estimators, open-source code and executables are provided on the website and GitHub of our team.