Tyler Smith a écrit :
> Hi,
>
> I'm trying to make sense of the options for multiple comparisons
> options in R. I've found the following options:
[ Snip ... ]
> As I understand it, there is no universal consensus as to which test
> is best.
There is no such thing. Each of the procedures is aimed at some subset
of the possible contrasts you may want to test. From my limited
knowledge, with some(I hope not too gross) simplifications :
- Tukey HSD will enable you to test the p(p-1)/2 pair differences one
can create with p groups ;
- Dunnett's procedure is made to compare (p-1) "treatments" to a common
control ;
- Scheffé's procedure is applicable to *any* ("reasonable") set of
contrasts you can form ;
- Newman-Keuls : aims to create separate subset of groups (but has
serious conceptual and technical flaws ! Don't do that nunless you know
what you're doing...).
- etc ...
Those procedures have the same hypothesis as the base ANOVA :
homoscedasticity, normality of residuals, etc ... Their robustness to a
small departure friom these conditions vary. As a last resort, a set of
non-parametric comparisons with the (overly conservative) Bonferroni
adjustment may help (but less power !).
You'll have to refer to the subject matter to make a choice.
> TukeyHSD appears to be appropriate for
> balanced designs. I
> have an unbalanced design to analyze.
Therefore, no aov(..) (hence no TukeyHSD(aov(...))) for you...
> I can use glht, but can someone
> tell me what each option is actually calculating? A reference to a
> paper that describes the procedures would be great, but I'm afraid I
> the one offered in ?glht[1] is beyond me.
Google ("multiple comparisons") will offer you some dubious and quite a
few good references...
HTH
Emmanuel Charpentier
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
