[R] Post-hoc test on ANCOVA

[David Winsemius]

On Feb 1, 2012, at 4:34 PM, Mark Na wrote:

> Dear R-helpers,
>
> I have an ANCOVA with a significant effect of the factor, which has
> three levels. I wish to determine which of the levels are different
> from each other but, because my model was fitted with lm(), I cannot
> use TukeyHSD.
>
> For some reason, I get different results (no significant effect of the
> factor) when I fit the model using aov()

... which is just a wrapper to "lm" ...'

> so, for the moment, I am
> using lm().

Doesn't that seem to violate a common sense rule? You should not pick  
a test (or in this test one particular formulation to the same  
function) just because it gives you the answer you want, but rather  
because it is appropriate for your data situation.

>
> Could anyone point me to a test and associated R function that will
> work on a fitted lm() or glm()?

Given what little you have said about your design and data analysis  
program, and what sort of heuristics you seem to use for choosing your  
tests, it seems that it might be unwise to give suggestions.

-- 
David Winsemius, MD
West Hartford, CT