Step-down trend tests for identifying the minimum effective dose

Many authors, most recently Tamhane, Hochberg, and Dunnett (18), have studied the problem of determining the minimum effective dose in dose-response studies. Based on past research and on findings from their own extensive simulation study, which covered a wide range of balanced normal homoscedastic situations, Tamhane et al. recommended a procedure they called SD2L, since it exhibited good performance in almost all the situations they studied. This method is a step-down procedure with a simple linear contrast-based trend test at each step. In this paper, we demonstrate that replacing the linear contrast trend test by Bartholomew's test leads to a procedure, SD2B, that consistently outperforms SD2L. In addition to the balanced normal homoscedastic framework, the finite sample performance of these procedures is also explored under unbalanced and/or heteroscedastic conditions. A third procedure, SD2W, which replaces the linear contrast test by Welch's test, offers some improvement over SD2B in a few heteroscedastic situations. In many cases, the increase in efficiency of SD2B and SD2W over SD2L exceeds 10%.