DMS Statistics and Data Science Seminar

Speaker: Yuming Zhang, University of Geneva

Title: A Simulation-based Estimation Framework for Generalized Partially Linear Models

Abstract: Significant increase of data size and rapid growth of model complexity have led to unreliable performance of many existing statistical methods due to their computational and numerical limitations. For example, in settings where the number of parameters diverges with

The sample size, consistent estimators can be difficult to obtain due to the analytical and/or computational challenges typically entailed. For semiparametric models, the unknown smooth functions that need to be treated nonparametrically further increase the estimation complexity as kernel or spline approximations are often necessary. We propose a simulation-based estimation framework for generalized partially linear models. This proposed framework provides a simple strategy to construct estimators for parameters of increasing dimensions with desirable statistical properties, including consistency, asymptotic normality and finite sample bias correction. Moreover, the proposed framework allows to account for various specificities commonly encountered, such as truncation, misclassification or outliers. The computational efficiency can also be significantly improved compared to classical methods. These theoretical results are exemplified with simulation studies on a partially linear logistic regression model.