(Reply copied to edstat).
Let me recapitulate your problem, and then ask for clarification.
You have a 2x2 design, both factors within subjects, and an unspecified
number N of subjects. For reasons beyond your control, some 10% of the
cells (this must mean 10% of the 4N observations, since you could hardly
have 10% of the four 2x2 cells). You seek a means of supplying what I
shall call fictitious information for those missing observations.
Question 1. How many missing cells are there?
Question 2. Are more than one cell missing for each subject?
If the answer to Question 2 is "No", I may have a method of addressing
your problem, although I'd be surprised if it met all your criteria
("good/statistically sound/widely used"). It certainly would NOT help
estimate the interaction effect (unless I can think of a suitable
modification), but would at least permit you to use more of your data.
(Just how bad is the situation if you should simply discard, at least
for a preliminary analysis, all the Ss for whom you have missing cells?)
If more than one observation is missing for a subject, it might be
necessary to discard that subject. In which case, how many Ss are you
left with? And depending on the pattern of observations missing, it
might be possible to include such a subject in a one-way sub-analysis
dealing with only one of your two factors, which might be reasonable
should you find no interaction between the two factors.
On Thu, 21 Aug 2003, Elvis Castello wrote (edited):
> does anybody know a good/statistically sound/widely used method to
> estimate missing values in a repeated measures design? References?
> Software to apply the method?
>
> I tested subjects on reading times of words belonging to 4
> conditions, resulting from 2 within-subject variables with two
> levels:
[I infer the following arrangement -- DFB]
Low frequency words High frequency words
participant Long words Short words Long words Short words
> 1 < mean RTs of 10 words in that condition >
> 2
> 3
>
> However, due to a prior phase (not important) some of my cells are
> empty (10%), approximately MAR.
Sorry, I don't recognize the acronym. "MAR"?
> I don't want to use multiple imputation,
Not clear that one could, actually. Do you have information that
would permit it?
> so I want to impute one value for each cell, and then use standard
> ANOVA statistics (I am aware that this results in an underestimation
> of within-condition variance) to test the effect of word length
> (short vs. long), word frequency and their interaction. I read about
> the EM algorithm (Little & Rubin, 1987) in SPSS which is good for
> imputation when variables are independently, but what in such
> repeated measures designs? Can (or should) I use the xtra info on
> the interdependence of the variables in the imputation procedure? If
> so, which procedure does this?
-----------------------------------------------------------------------
Donald F. Burrill [EMAIL PROTECTED]
56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
