Recent advances in bootstrap test for comparing two or more groups
Alok Kumar Dwivedi
Associate Professor, Division of Biostatistics and Epidemiology Director of Biostatistics and Epidemiology Consulting Lab (BECL) Department of Biomedical Sciences, Paul L. Foster School of Medicine Texas Tech University Health Sciences Center El Paso (TTUHSC EP) Room no-3103A, Medical Science Building 5001 El Paso Drive, El Paso, Texas-79905 Cell: 513-917-4242 ; Office: 915-215-4177 Fax: 915-545-5716
Randomized experimental design is a common design in biomedical
studies which typically requires univariate statistical tests for the analysis. In such studies, investigator generally wants to compare quantitative outcomes according to two or more conditions/groups. However, non-normal data distribution, small sample size, and unequal variances between groups frequently pose analytical challenges in these studies. A number of parametric and nonparametric statistical tests have been recommended under different assumptions and conditions for data analysis. The validity of these statistical tests in terms of statistical power and type I error probability has been questioned especially for non-normal data with small to moderate sample size studies. To overcome this problem, methodologists have started exploring the utility of semi-parametric approaches such bootstrap and permutation tests which combine the benefits of parametric and nonparametric tests. However, these alternative tests also have sample size limitations. Further, there are many versions of bootstrap test which also produced varying statistical powers and type I error probability. We developed pooled bootstrap test, composite bootstrap test, and bootstrap test based on rank t-statistic with counterintuitive resampling method for comparing distributions and means across groups. The performance of these novel bootstrap tests as compared with common parametric and nonparametric statistical tests for comparing means were assessed using extensive simulations under various conditions and using real data examples. The proposed bootstrap tests reasonably controlled type I error probability in most situations. All the proposed bootstrap tests provided greater or equal statistical power for comparing two means as compared with unpaired t-test, Welch t-test, Wilcoxon rank sum test, and permutation t-test in many conditions. Similarly, all the proposed bootstrap tests outperformed over other alternatives for comparing paired means and for comparing more than two means. For exponential family distributions, we recommend to use the pooled bootstrap test while the use of composite bootstrap test or bootstrap based on rank t-statistic should be preferred for extreme variable distributions such as Cauchy or log normal data, or under non- exchangeability conditions. The proposed bootstrap tests should also be used for comparing paired or unpaired groups.