Dear Allstat,
Thank you to all who replied in response to my query. The general
conclusions are summarized below
(1) Multiple Imputation + GEE
First, carry out multiple imputation of the missing values, say 3 or 4
imputations, yielding "complete" datasets.
Second, analyze the "completed" datasets using GEE
There is a chapter in the book Missing Data in Clinical Studies by
<http://eu.wiley.com/WileyCDA/Section/id-302479.html?query=Geert+Molenberghs
> Geert Molenberghs,
<http://eu.wiley.com/WileyCDA/Section/id-302479.html?query=Michael+Kenward>
Michael Kenward which talks about combining multiple imputation and GEE
(2) Weighted GEE
This falls under the so-called Selection Models.
Firstly, model the missingness (dropout) process using multiple logistic
regression under the MAR assumption, i.e. using baseline covariates,
baseline and previous measurements of your response.
Secondly, calculate weights for the dropout process proportional to the
inverse probability of dropout from the logistic model for dropout.
Lastly, fit a GEE model weighted by the weights obtained from the logistic
model for dropout.
This method is valid under MAR assumption even if the correlation model is
misspecified, provided the model for estimating the probability for missing
response is correctly specified.
(Robins, Rotnitzky and Zhao 1995)
Best Wishes
Clare
From: Clare Rutterford [mailto:[log in to unmask]]
Sent: 03 November 2009 12:29
To: [log in to unmask]
Subject: Question :GEE with missing data
Dear Allstat,
I have a longitudinal dataset from a 2 treatment group clinical trial with
baseline and 4 follow up timepoints. I may have missing baseline covariates
and follow up data.
As I am interested in the average effect of the treatment groups I wish to
analyse using a GEE model. As the missing data is likely to be Missing At
Random (MAR) I have read that GEE models are likely to provide biased
estimates.
My thoughts were to multiply impute the missing data and then carry out a
GEE model.
I would be interested to hear your views on whether this is the best
approach (if so how best to impute the missing data?) , or are there
alternative approaches.
Best Wishes
Clare Rutterford
|