[R] t.test() with missing values

[Peter Dalgaard]

S Ellison wrote:
>   
>>>> Peter Dalgaard <P.Dalgaard at biostat.ku.dk> 14/09/2007 09:26:16 >>>
>>>>         
>>> So what can I do now to solve my problem?
>>>
>>> Do you think I should not use paired=TRUE?
>>>       
>> You *can* only use it when you have pairs, and you must do it then, to
>> correct for intra-pair correlation. The drawback is that it looks only
>> at complete pairs, throwing away all the singlets. It is possible to
>> recover the information from the singlets , basically by combining a
>> paired test for the pairs and an unpaired one for the singlets. (Someone
>> must have written this down, but I'm afraid I don't have a nice reference).
>>     
>
> Question: Could you achieve this kind of outcome with lme? stack the two groups, mark the observations y by subject (ie the pair ID) and group (treatment, presumably), and do something like
>
> anova(lme(y~group, data=d, random=~1|subj, na.action=na.omit))
>
> Or is that just disguising one of those nasty unbalanced 2-way anova problems?
>   
Yes, but....

I don't think lme() will do better than what you can do by hand: Get two
independent estimates of mu1-mu2 (one estimate from the pairs and one
from the singlets), compute a weighted average using the s.e.'s and test
that against zero (possibly after testing them for equality for good
measure). This is easy if you use a plug-in approach: first assume that
the s.e. are known, then plug in their empirical value. The tricky bit
is to calculate the DF in the style of Welch's test.

-- 
   O__  ---- Peter Dalgaard             Øster Farimagsgade 5, Entr.B
  c/ /'_ --- Dept. of Biostatistics     PO Box 2099, 1014 Cph. K
 (*) \(*) -- University of Copenhagen   Denmark          Ph:  (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)                  FAX: (+45) 35327907