(82) A practice-oriented guide to statistical inference in linear modeling for non-normal or heteroskedastic error distributions
Date:
Contributors: Weber, H., Huber, S. E., & Arendasy, M.
Venue: 53rd DGPs Congress/15th ÖGP Conference, Vienna, Austria, September 16-19, 2024
Abstract: Researchers are frequently encouraged to use well known parametric methods for data analysis, even though their actual data often fails to satisfy the necessary assumptions. Even though a variety of alternative methods to ordinary least squares regression (OLS) exist, they are not available in all software packages or simply lack popularity and exposure. Using OLS with corrected standard errors (SE) or bootstrap methods might be a more viable solution in practice since they are better known, readily available and do not force researchers to familiarize themselves with a different estimator. The planned simulation study will compare how type-I error, power, SE bias and confidence interval coverage differ between the classic OLS SE, type 3 and type 4 of the heteroskedasticity-consistent (HC) SEs, and two bootstrap methods, case resampling and wild bootstrap. For the latter two types of confidence intervals (percentile and bias-corrected and accelerated) will be considered. The simulated data will vary in sample size, degree of non-normality, and degree of heteroskedasticity. HC methods, and similarly the wild bootstrap, are expected to outperform OLS when the errors are normally distributed but heteroskedastic. In contrast, some previous studies have suggested OLS to work well even for larger degrees of non-normality. The goal of this study is to advise researchers under which typically encountered circumstances standard OLS regression is still appropriate and when post-hoc corrections or bootstrap methods should be preferred.