Hi Jeremy and Brian,
Thanks very much for your detailed and really helpful answers.
To clarify, mine is a pre-post study, no control group, only two data
points in time.
The Editor's comment was phrased as:
"The data analysed represent only a proportion of those who signed up
for the intervention and they are clearly different from those that
did not complete, an intention to treat analysis would have taken
little time and added better evidence of the true effectiveness of
the intervention".
Below are two more answers responses that I have received.
1. I do think though that intent-to-treat analysis is very
conservative and probably loses you lots of power - it assumes that
all the data that you had at baseline is the same at follow-up, which
might be wildly unrealistic (e.g. if there is a main effect of time).
Generally, I think the analysis should try to include all cases, no
matter how much missing data, and use imputation or full information
maximum likelihood to deal with the missing data.
This article is useful as a general one on the principles behind ITTA:
<http://us.mc1135.mail.yahoo.com/mc/http://www.bmj.com/cgi/content/full/319/7211/670>http://www.bmj.com/cgi/content/full/319/7211/670.
This is kind of sophisticated, probably unecessarily so, but possibly
worth a look:
<http://us.mc1135.mail.yahoo.com/mc/http://www.nesug.org/Proceedings/nesug00/st/st9008.pdf>http://www.nesug.org/Proceedings/nesug00/st/st9008.pdf
This seems a pretty pursuasive argument for a full information
maximum likelihood-type analysis (what they refer to as mixed model):
<http://us.mc1135.mail.yahoo.com/mc/http://www.rti.org/pubs/mr-0009-0904-chakraborty.pdf>http://www.rti.org/pubs/mr-0009-0904-chakraborty.pdf
AND
2. Analytic approaches (last-observation-carried-forward, multiple imputation)
can be used to reduce, but not remove the effect of withdrawal but should be
done. I would use multiple imputation of post treatment outcomes on the 25%
of missing cases as well as a last value carried forward analysis. If you
had used the SDQ you might use the Value Added formula that estimates using
a simple regression mode what the values are likely to be without treatment
- and I would have used that but only if the time two measures were at 6
months. You would have already done an analysis that looks at the
differences between the dropouts and the people who stay in the treatment
and enter those predictors as covariates. Again, not a perfect solution but
better than no
I understand from the above suggestions that I mainly have 2 options:
1. To try, as Brian suggested, using the baseline data as follow-up
data for those who did not provide any data at the end of treatment
(conservative, but necessary).
2. The other approach may be to impute the follow-up data at end of
treatment, using linear regression. As predictors I could use
baseline data at time 1,
and any other variables where dropouts and completers differ. For
example, I could impute CBCL internalising problems at End of
treatment based on CBCL internalising problems scores at baseline,
severity of psychosocial stressors, overall level of functioning
(GAF), where there were significant differences between the analysed
group and the group we didn't have data for.
Is this reasonable?
I understand other options would be to simulate the data using more
complicated approaches (at least as they appear to me!), but could
one do such a thing with SPSS (the classic question...)
Many thanks again
Ioanna
|