ABSTRACT
We consider a control group versus treatment group experimental design and assert that powering the design for a potential treatment effect that is represented by a pure shift of the control group distribution is usually unrealistic. Instead, we propose the use of a mixture model as the design alternative in anticipation that there might be a sub-population in the treated population whose responses come from the same control group distribution. When the responses in the treatment group follow a mixture model, the sample size found by the traditional pure shift alternative based method is demonstrated to be under-powered. We develop a new sample size formula for the Wilcoxon test statistic and propose a more general definition of the treatment effect. Method of moment estimators of the treatment effect are proposed and their bias and mean squared error properties are evaluated.
Acknowledgements
The authors appreciate helpful comments from two referees and an associate editor.
Disclosure statement
No potential conflict of interest was reported by the authors.
ORCID
Daniel R. Jeske http://orcid.org/0000-0002-0214-7992