M. Nthangeni et J. Algina, Type I error rate and power of some alternative methods to the independentsamples t test, EDUC PSYC M, 61(6), 2001, pp. 937-957
Several tests exist for use in treatment-control studies in which a larger
treatment mean may be accompanied by a larger treatment variance. Type I er
ror rates and power were estimated for four of these tests and for the inde
pendent samples t test and the Welch test. Five factors were manipulated: h
eteroscedasticity, total sample size, sample size imbalance, distribution,
and power. The test due to Brownie, Boos, and Hughes-Oliver may not control
the Type I error rate when the data are skewed. Among the other tests, one
due to Conover and Salsburg and one due to Johnson, Verrill, and Moore had
the best power advantage when compared to the independent sample t test an
d the Welch test. In the Johnson et al. test, a parameter that strongly aff
ects power is specified, and a poor choice can reduce power below that for
the Conover and Salsburg test.