LMM - Problem with missing data

#1
Hi!

I've been working on thses stats for a long moment and I still can't figure out what is the best way to deal with this problem.

Participants in a clinical study have written in a diary emotions and other variables trough a therapy for OCD. I am using LMM in SPSS to analyze a set of emotions (anxiety, sadness, joy) in relation with a treatment variable (outcome) accross therapy sessions (20 weeks) and I am using a very small N (8 persons).

Now, I know I can perform LMM even if there are few missing data, like someone skipped a day here and there, but the problem is that some people haven't rated some emotions. For example, 8 out of 8 rated anxiety and sadness but only 5 rate joy.

If I perform an analysis with all emotions as covariates, SPSS only calculates for the 5 persons who rated all emotions. But I'm wondering if I can analyze all emotions in different models. Is it possible? Or is there a way I can perform an anlysis with everything in one model and not losing participants?

Examples of syntax:

**For everything in one model**
MIXED
IP BY Sujet WITH Semaines Anxiete Tristesse Joie Colère
/PRINT = SOLUTION TESTCOV G
/SAVE = FIXPRED SEFIXP DFFIXP PRED SEPRED DFPRED RESID
/CRITERIA = CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
/METHOD = ML
/FIXED = Anxiete Tristesse Joie Colère Semaines Semaines*Anxiete Semaines*Tristesse Semaines*Joie
Semaines*Colère | SSTYPE(3) NOINT
/REPEATED=Semaines | SUBJECT(Sujet) COVTYPE(AR1)
/RANDOM = intercept Semaines | COVTYPE(UN) SUBJECT(Sujet).

**One emotion per model**
MIXED
IP BY Sujet WITH Anxiete Semaines
/PRINT = SOLUTION TESTCOV G
/SAVE = FIXPRED SEFIXP DFFIXP PRED SEPRED DFPRED RESID
/CRITERIA = CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
/METHOD = ML
/REPEATED=Semaines | SUBJECT(Sujet) COVTYPE(AR1)
/FIXED = Semaines Anxiete Semaines*Anxiete | SSTYPE(3) NOINT
/RANDOM = intercept Semaines | COVTYPE(UN) SUBJECT(Sujet).

And if I do them seperatly, how can I interpret the effect of time, since I get different results with each model?

Help MUCH appreciated :D
 

noetsi

Fortran must die
#2
The default for most software is to eliminate any case on which there is any missing data. You can chose to change the default, but you need to ask yourself the following question. Is the lack of response truly an oversight (that is random) or are people deliberately not answering some questions for reasons that are of interest to your research. If the later is true and you ignore that data is missing (that is you use the responses they did put down and ignore that they did not answer other questions), I would think it would bias your results.

Tabachnick and Fidel ("using multivariate statistics" 5th ed do a good job of covering missing data including software issues.
 
#3
THank you for your answer!

Well, they chose to enumerate some emotions, so technically they didn't 'choose' to not rate those, as they weren't mandatory (I didn't do the design!!), so it's more random as in 'they didn't think of rating joy or disgust', rather than not wanting to.

If I ever want to change the default, I would I do that? I have read about it, but never really found out how to do it.

THanks again!
 
#4
Basically, what I want to know is: is there anyway the LMM analysis in SPSS can analyze a model with some variables complete and some variables having missing data (ex. 2 out of 8 persons haven't rated joy).

In other words, can LMM analyze a model and take in account data from everyone for some variables and data of only some participants for other variavles (those who have complete data for those variables).