A1655
Title: Sparse sieve MLE
Authors: Di Liu - StataCorp (United States) [presenting]
Abstract: This paper introduces a Sparse Sieve Maximum Likelihood Estimator (Sparse- SMLE) that utilizes Neyman orthogonal scores for estimating marginal parameters within a copula framework. We employ a Bernstein copula to capture dependence structures between marginals, leveraging a Lasso-type Bernstein sieve approach to es- timate the high-dimensional copula parameters. A key challenge addressed is the reg- ularization bias introduced when applying Lasso-type techniques to achieve sparsity in the copula structure. We derive Neyman orthogonal scores that provide robustness against machine learning bias from the sparse copula estimation, enhancing the consis- tency of the marginal parameter estimates. The estimator is implemented through a concentrated-out likelihood, with Lasso and post-Lasso adjustments to manage dimen- sionality in the nuisance parameters. We propose single-sample and cross-fitted ver- sions of the Neyman orthogonal estimator and assess their finite-sample performance through simulation. Our results demonstrate that the Neyman orthogonal Sparse- SMLE outperforms traditional SMLE under model selection errors and improves esti- mation efficiency over Quasi-MLE.
