Detailansicht
Testing nonparametric functionals in factorial designs | Prof. Markus Pauly (Universität Ulm)
Most existing tests for nonparametric factorial designs are based on
ranks and hypotheses are formulated in terms of distribution functions.
However, especially in heterogeneous settings, null hypotheses
formulated in terms of parameters or effect measures and corresponding
confidence intervals would be of more interest. In this talk, we explain
that the effect measures, underlying existing rank-based procedures, may
either lead to possibly paradox results or/and depend on sample sizes
(except in the case of completely balanced designs). Moreover, we point
out that this undesirable property may particularly cause problems in
interpretation of effects but also for inference. Thus, we propagate to
work with unweighted nonparametric effect measures that can be motivated
from so-called pseudo-ranks. We then introduce novel test procedures
that are suitable for testing hypotheses formulated in these effects in
general uni- and multivariate factorial designs and analyze their large
and small sample properties theoretically and in simulations. We note
that the R-package rankFD performing the computations in general
univariate factorial designs can be downloaded from CRAN.
References.
Brunner, E, Konietschke, F, Bathke, A and Pauly, M (2018). Ranks and
Pseudo-Ranks - Paradoxical Results of Rank-Tests. Submitted Preprint.
Brunner, Edgar, Konietschke, Frank, Pauly, Markus and Puri, Madan L.
(2017). Rank-Based Procedures in Factorial Designs: Hypotheses for
Nonparametric Treatment Effects. Journal of the Royal Statistical
Society - Series B 79, 1463–1485.
Dobler, D, Friedrich, S and Pauly, M (2017). Nonparametric MANOVA in
Mann-Whitney effects. Submitted Preprint.
Umlauft, M, Placzek, M, Konietschke, F, and Pauly, M (2017). Wild
Bootstrapping Rank-Based Procedures: Multiple Testing in Nonparametric
Split-Plot Designs. Submitted Preprint.
Keywords: Factorial designs, Kruskal–Wallis test, Rank-based Methods,
Multivariate Ordinal Data, Wild Bootstrap.
Einladung von Prof. Werner Brannath