Hi all,
I'm examining data from a quite complex designed experiment, where 4% of
the response observations are missing. I perfectly know that the missing
mechanism depends on the values of covariates (all available) and
results in very large values of the response being missing. The full
records are therefore biased with respect to the original design whose
effects are not comparable anymore. For this reason I'm thinking of
re-constructing the missing response data, though standard multiple
imputations (as those available in SAS proc MI) do not apply in this
case, due to the informativeness of the missing mechanism. Unfortunatly
my bibliography about missing data in experiments stops at chapter 2 of
Little and Rubin's 1987 (!) book. Does anybody have any idea?
--
Alessio Pollice
Dipartimento di Scienze Statistiche
Università delgi Studi di Bari
Via C. Rosalba 53
70124 Bari, ITALY
|