Article
Farcical Results with Rank Procedures in Unbalanced Designs – Ranks and Pseudo-Ranks
Search Medline for
Authors
Published: | February 26, 2021 |
---|
Outline
Text
Background: If rank methods are used for d>2 samples then farcical results may be obtained in case of unequal sample sizes since the quantities pi = ∫H dFi on which rank procedures are based depend on the relative sample sizes through the definition of the weighted mean distribution function H in the experiment. This undesirable property applies to the well-known Kruskal-Wallis test [1], the trend test by Hettmansperger-Norton [2] as well as for the Akritas-Arnold-Brunner procedures [3] in factorial designs. In two-way layouts interactions may appear or disappear just by changing the ratios of the samples sizes while keeping fixed the underlying distributions. Such designs appear in the so-called sub-group analysis in clinical trials or in clinical epidemiology.
Methods: First some examples are presented and then the theoretical background of the farcical results is investigated by extending the well-known Mann-Whitney effect for two samples [4] to several samples and to factorial designs. It turns out that a nonparametric effect (a quantity not depending on relative samples sizes) could be based on comparisons of the distributions Fi for example, with the unweighted mean distribution G.
Results: The functionals ψi = ∫G dFi describe an effect of the distribution Fi relative to the mean distribution G. They can be estimated by simple plug-in estimators which can easily be computed by rank-like quantities, the so-called pseudo-ranks which have already been discussed in the literature [5]. The asymptotic distribution of these estimators and confidence intervals for ψi (see, e.g., [6] or [7]) are briefly discussed. For equal sample sizes and for two distributions ranks and pseudo-ranks are identical.
Conclusion: It is demonstrated that in case of unequal sample sizes procedures based on the nonparametric effects ψi do not have the farcical properties of rank-based procedures. An example from a register study in multiple sclerosis demonstrates the practical meaning.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
References
- 1.
- Kruskal WH, Wallis WA. The use of ranks in one-criterion variance analysis. Journal of the American Statistical Association. 1952; 47: 583-621.
- 2.
- Hettmansperger TP, Norton RM. Tests for patterned alternatives in k-sample problems. Journal of the American Statistical Association. 1987; 82: 292-299.
- 3.
- Akritas MG, Arnold SF, Brunner E. Nonparametric hypotheses and rank statistics for unbalanced factorial designs. Journal of the American Statistical Association. 1997; 92: 258-265.
- 4.
- Mann HB, Whitney DR. On a test of whether one of two random variables is stochastically larger then the other. Annals of Mathematical Statistics. 1947; 18: 50-60.
- 5.
- Gao X, Alvo M. A Unified Nonparametric Approach for Unbalanced Factorial Designs. Journal of the American Statistical Association. 2005; 100: 926-941.
- 6.
- Thangavelu K, Brunner E. Wilcoxon Mann-Whitney test for stratified samples and Efron's paradox dice. Journal of Statistical Planning and Inference. 2007;137:720-737.
- 7.
- Brunner E, Bathke AC, Konietschke F. Rank- and Pseudo-Rank Procedures for Independent Observations in Factorial Designs. 2019. (Series in Statistics).
- 8.
- Rommer PS, Eichstädt K, Ellenberger D, Flachenecker P, Friede T, Haas J, Kleinschnitz C, Pöhlau D, Rienhoff O, Stahmann A, Zettl UK. Symptomatology and symptomatic treatment in multiple sclerosis: Results from a nationwide MS registry. Multiple Sclerosis Journal. 2018: 1-12. DOI: 10.1177/1352458518799580