04-15-2015 02:09 PM
I am working on models to predict hospital readmission. The data were arranged by admissions (claim ID) with a flag to indicate whether there was readmission followed. A member (person) could have multiple admissions, which meant one member ID could correspond to multiple claim IDs. How do I take this repeated measure into account in SAS EM? I did not find any place I could define repeated variable in regression, decision tree or neural network. I suspect probably variable role in SAS EM could clarify this? But how? When I raised this question in one SAS user group meeting, people from bank system told me that my scenario is equivalent to theirs in which they have person level data as well as account level data (one person could have multiple accounts). But they could not answer my question.
Does anyone have any experience? Thanks.
04-15-2015 05:09 PM
I'm not an expert on recurring events, but I'd approach the problem like this:
Suppose each individual patient record were coded 1,2,...n based on the 'episode'. Examine whether the factors affecting probability of readmission are the same across episodes.
You can investigate this by building stratified models of readmission for each episode (you might need to combine the highest numbers of episodes since you're likely to see relatively few patients with really large numbers of episodes). Some basic visual exploration may help as well.
If there isn't a systematic relationship between episode and the factors affecting readmission, then you can simply ignore the episodes, and build a single model treating all patient records equally. If your ability to predict readmission does depend on the episode, then you could include it as an model input or go with stratified models. (I'd try try both and see which gives the best predictions in validation.)
Hope this helps,
04-15-2015 08:36 PM
Outside of E-miner, PHREG would handle this (proc logistic has STRATA too, depending on your version). I have not done this inside E-miner, though.
Perhaps you can think about your problem like a Cox regression and model time-until-readmission. E-miner does this very well using the Survival node (under the Applications toolbar) which I believe uses PHREG behind the scenes.
04-16-2015 04:03 PM
Repeated measures analysis is not available in EM. I would look into PROC MIXED and PROC GLIMMIX which would enable you to specify specific correlation structures for the observations that come from the same person.
Note that if you have a very large data set with large number of predictors, fitting mixed models can be computationally very exhaustive. If that is the case, try HPMIXED which might provide batter performance.