patterns of missing data and SPSS
Brigette
My study was a RCT with an immediate treatment group and a waitlist
control group assessed at 3 time points - pre-treatment, post-
treatment and at a 3 month followup. I have both intermittent
(occassional missing data due to incorrect questionnaire or subscale
completion) and missing data due to dropout. I have greater dropout
from my waitlist control group - a problem that is not so uncommon
from what I understand. I have small n's excluding my missing data -
immediate treatment group n=20 and waitlist control group n=10. I had
3 dropout of my treatment group (2 due to family deaths) and 14
dropout of my waitlist control group.
My problem is really modelling the missingness to determine whether it
is MCAR (probably not), MAR or MNAR. SPSS in missing value analysis
has Little's MCAR test for determining whether data are MCAR and
Separate Variance t-tests for determining whether data are MAR. My
problem is that I am unsure how to enter my variables to determine
these patterns of missingness. My study is one of psychology and we
tend to give questionnaires as measures. I have 6 questionnaires that
were repeated at times 1, 2, and 3. For most of these questionnaires
there are various subscales which purportedly measure slighlty
different things. My question is do I subject each subscale to the
MVA or do I combine all subscales from a particular measure for MVA. I
have tried both ways - getting a pooled Little's MCAR statistic when I
subject all subscales to MVA at once. I also tend to get significant
results when I use all subscales on the Separate Variance t-tests
compared to when I analyse each subscale separately.
I am also unsure whether I should be looking at each group separately
when it comes to analysing missing data patterns?
I also wonder whether my dataset is too small to be using MVA and
producing reliable results?
Unfortunately none of the examples I have read have variables similar
to mine so I am finding this part rather challenging so any help would
be gratefully appreciated.
mcap
You may be unlikely to get answers on the specific MVA analysis
module functions because most of us can't afford the additional cost
and have never used it. Some of these features are part of the base
package in programs like Stata.
That said, there are still plenty of ways to consider missing
values. Tabachnick and Fidell do a nice job in their mutlivariate
book (although I think they use the MVA module also). Have a look if
you get a chance.
You are raising two separate issues here as far as I can see.
Although technically, if someone drops out, their answers could be
considered missing values...they are different. Missing values need
to be handled separately from dropout. You have a very small sample
and a relatively large number of dropouts. There is no way to get
around this with missing data. You need to look at both groups and
decide how the dropouts differ between the groups and how the dropouts
differ with the subjects who stayed in. Your choice of what to look
at and compare depends on what you think is important as the
researcher. There is not much you can do here except explain and
compare. There is no imputation for someone who missed the full
questionnaire at a particular time.
For your user missing values within subscales you have a lot of
options. Look at the patterns in terms of your outcomes and in terms
of your other indep variables. You don't need to do fancy tests
necessarily. As for imputation, there are a variety of strategies -
each with limiations. Whether you look at subscales or the full
scales if you are going to replace depends on what each one is and
whether you think it can be done in theory and your choice of
strategy.
If this is pilot data you may be OK even with everything you don't
have. You will have to explain and justify quite a bit though.
HTH,
Marc