Shrinkage for Categorical Regressors

TL;DR

This paper proposes a new regularization method for categorical regressors that improves estimation accuracy by shrinking group means towards informative estimates, with theoretical analysis and practical demonstrations showing its advantages.

Contribution

It introduces a flexible, closed-form shrinkage estimator for categorical regressors, deriving optimal penalties and demonstrating asymptotic and finite-sample improvements over traditional methods.

Findings

Plug-in estimator outperforms OLS in large samples with more than three groups.

Monte Carlo simulations show robust finite-sample improvements.

Real data applications illustrate practical benefits in panel and difference-in-differences analyses.

Abstract

This paper introduces a flexible regularization approach that reduces point estimation risk of group means stemming from e.g. categorical regressors, (quasi-)experimental data or panel data models. The loss function is penalized by adding weighted squared l2-norm differences between group location parameters and informative first-stage estimates. Under quadratic loss, the penalized estimation problem has a simple interpretable closed-form solution that nests methods established in the literature on ridge regression, discretized support smoothing kernels and model averaging methods. We derive risk-optimal penalty parameters and propose a plug-in approach for estimation. The large sample properties are analyzed in an asymptotic local to zero framework by introducing a class of sequences for close and distant systems of locations that is sufficient for describing a large range of data generating processes. We provide the asymptotic distributions of the shrinkage estimators under different penalization schemes. The proposed plug-in estimator uniformly dominates the ordinary least squares in terms of asymptotic risk if the number of groups is larger than three. Monte Carlo simulations reveal robust improvements over standard methods in finite samples. Real data examples of estimating time trends in a panel and a difference-in-differences study illustrate potential applications.