Does the testanova command do something equivalent to using MSw
assuming equal variances, or something equivalent to allowing unequal
variances approach after Welch and Satterthwaite?
For example, to test any contrast (psi) where the ANOVA was estimated
by least squares, we can use:
F = (psi)^2 / sum(c^2/n) / MSw
which assumes equal variances.
Omnibust tests in ANOVA may be robust to unequal variances, but
contrasts are not. One way to fix the F above, allowing variance
heterogeneity, is to change MSw and denominator df. I won't begin to
try to type those formulas in here but they are on page 166 in Maxwell
and Delaney 2004 for example, with an SPSS contrast example output on
page 169.
