multiple imputation data analysis conundrum

[Zdaniuk, Bozena-3]

Hello everyone,
I want to run a post - treatment analysis of my longitudinal data. We have already done the analyses (using MIXED) of baseline (T1), post-treatment (T2) and six-month followup (T3). Now, we are interested in T2 (post-treatment) versus T4 (12-month followup). I want to use multiple imputation for T2 missing data (subject dropout) in order to maintain ITT approach. I want to use 2x2 mixed (1 within-, 1 between-subject factor with 2 levels each) ANOVA to compare two treatments on how they change between T2 and T4.
I know I can accomplish it by first doing Multiple Imputation in SPSS and then running MIXED analysis equivalent to such an ANOVA (I think SPSS does not support GLM ANOVAS for multiply imputed data sets). But here is my conundrum:
My assumption #1 is that I should include baseline variables in the multiple imputation procedure in order to maximize the quality of imputed T2 data.

My assumption #2 is that I should first create the long format file, then do Multiple Imputation, and then run the MIXED syntax on the imputed files.

If my assumptions 1 and 2 are correct, how do I include baseline in the long file for the imputation but then ignore it for the MIXED procedure run on the imputed files?  I will need a time index and I will have three rows of data for each subject instead of  two...
Are my assumptions correct? If yes, what can I do?
thanks so much in advance!
bozena


=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Andy W]

I'm not on expert on MI for longitudinal data, but I think most folks do that
in wide format (and then if necessary put constraints on the predicted
values, or take advantage of the monotonic missing nature of the data in the
imputation process). I imagine you could do the imputation in wide format,
then use VARSTOCASES to turn into long format and estimate your models.
(Worst case you have to combine whatever results yourself instead of SPSS
spitting whatever pooled estimate for me.) If you post your model other
folks on the list can give better advice whether it can be implemented in
wide format.

If you insist on the long data format, you may just use LAG to create a new
variable (so you only have two rows). You may just have for both T2 and T4 a
variable that equals baseline, but in long format your imputed value for T2
won't have anything to do with the imputed value for T4. (Unfortunately SPSS
does not allow passive transformations in the imputation procedure.)





-----
Andy W
[hidden email]
http://andrewpwheeler.wordpress.com/
--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD