Small-Sample Methods for Heteroscedasticity-Robust Hypothesis Tests in Linear Regression: A Review and Evaluation

• In Event: 52.034 - Single-Case Designs and Small Sample Issues

Abstract

Linear regressions are one of the most ubiquitous statistical tools in educational research and across the social sciences. Classical procedures for inference in linear regressions become biased in the presence of heteroskedastic errors. A well-known solution for handling heteroskedasticity is to use heteroskedasticity-consistent covariance matrix estimators (HCCMEs) for constructing hypothesis tests or CIs. While such tests are guaranteed to have correct rejection rates asymptotically, their performance can degrade (becoming overly liberal) in small samples. A variety of small-sample corrections have been proposed to improve the performance of tests based on HCCMEs, including Satterthwaite, Edgeworth, and saddlepoint approximations. This paper provides a comprehensive review and simulation study evaluating the proposed small-sample methods, with the aim of providing recommendations for practice.

Authors

• James Eric Pustejovsky, The University of Texas - Austin

• Gleb Furman, The University of Texas - Austin