A New Method of Resampling Resting Nonparametric Hypotheses: Balanced Randomization Tests

• Activity Number: 262
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Oumar Thiero* and Sudesh Srivastav and John J. Lefante and Frances Mather
• Companies: Tulane University and Tulane University and Tulane University and Tulane University
• Keywords: Resampling tests ; Bootstrap ; Monte Carlo tests ; permutation tests ; Balanced Randomization ; Balanced Sampling design

Abstract

Resampling methods(RM):Monte Carlo(MC), Bootstrap Approach(BA) are used as a non-parametric method for comparing and/or estimating population parameters(PP). However, relatively large random samples needs to perform MC and BA methods, in order to achieve exact distributions for PP. We developed an efficient(EF) RM comparing two PP, using a balanced and controlled sampling design with small number of resamples: Balanced Randomization(BR) tests. Multiple datasets varying under probability distributions and sample sizes were simulated from known data to compare the accuracy/consistency(AC) and EF of RM. The AC comparison were performed by the correlation(R) "exact p-value(PV)vs. RM PV" and 95[99]% of PV containing the true PV. Screen tests parameters(ST) were used to compare EF. The R are higher for BR and MC, (increased with an increased sample size), much less for BA, and most pronounced for skewed distributions. Furthermore, the relative proportion of 95[99]% AC for BR/MC=3[1.3]% and BR/BA=20[15]% (p< .001). The ST of the BR method were shown to have a slight advantage in most situations. Therefore, BR is more EF and AC for comparing two groups being less computationally intensive.