New Critical Values for the Winsorized t-Test
Michael Lance *
Education Evaluation and Research, Wayne State University, Detroit, MI, 48282, USA.
Piper Farrell-Singleton
Office of Education Improvement and Innovation, Michigan Department of Education, Lansing, MI, 48909, USA.
Shlomo S. Sawilowsky
Education Evaluation and Research, Wayne State University, Detroit, MI, 48282, USA.
*Author to whom correspondence should be addressed.
Abstract
Aims: To determine if (and in which situations) Monte Carlo or asymptotically derived critical values are more robust for the Winsorized t-test.
Study Design: A Monte Carlo simulation via FORTRAN 90 was used to test type I and II error properties across 14 unique distributions for various combinations of sample sizes and effect sizes for alpha = .01 and .05. Both Monte Carlo and asymptotically derived sets of critical values were used. Each combination of parameters was used to run 1 million iterations.
Place and Duration of Study: Windows PC for a duration of 6.5 days (to obtain results generated per each set of iterations).
Methodology: FORTRAN 90 code was used to do the following: For 1 (value) and 10% of n1 + n2, samples were drawn per distribution and Winsorized. Next, t-tests were conducted per the parameters specified above in the study design.
Results: Results generally supported the use of the new table of Monte Carlo derived critical values over the classical asymptotically-derived critical values.
Conclusion: The Monte Carlo-derived Winsorized critical values are generally preferable to asymptotically derived critical values.
Keywords: Outliers, trimming, winsorizing, t-test, critical values, Monte Carlo simulation