Anonymous posted on Wednesday, October 20, 2004 - 8:47 am
I have run a basic mixture model of the distribution of a single variable following the example of Everitt and Hand (1981).
I then used this model to stratify a two-part model of insurance payements.
to identify the mixture classes I used four quarterly measures of a health severity score. I wanted to use quarterly measures instead of an annual measure to distinguish between those with acute and chronic illness.
In some instances, a patient was missing data from a particular quarter. I initially tried to estimate the model with these values missing at random, but the MPLUS was not able to find a solution.
Faced with that, I went to a second best last value carried forward approach LVCF, and estimated the equation accoridingly.
I submitted the manuscript and recieved the following reviewer question:
"If missing values were replaced using the last value carried forward (LVCF) method, then this is problematic because LVCF inflates the nomimal sample size and spuriously increases estimate precision. See the Rubin and Little (2002) article for more information. A more appropriate method would be to include a covariate for number of months enrolled or weighting the data appropriately."
I'm not sure if the reviewer is implying that I should have months as well as health severity scores in the mixture model or what he means?
Its a very big sample, and the means of the clusters are very different. Further, I get essentially the same clusters if I run the data on a single annual measure in which there would be no missing data elements. Long story short, I don't think this extra precision is leading me to select erroneous classes.
However, I need to intelligently respond to the reviewer's comments. Can someone educate me on potential reduction in SE's of my mixture means that result from LVCF and advise on an intelligent approach to respond to this question?
I would like to see the Mplus run that you say did not find a solution. If you send the input/output and data to email@example.com, I would be happy to take a look at it.
Anonymous posted on Wednesday, October 20, 2004 - 2:04 pm
Its really not an issue of the Mplus output, it would be a very large task to go back and redo the analysis using a different imputation method. What I am really looking for is a good discussion of the bias related to LVCF to be ableto appropriately judge if the issue is substantial enough to warrent going back and redoing what I've done.
My feeling is that its not, but it would be great to know for sure.