05-14-2016 08:35 AM
I have 6000 patients with 10 treatment choice. I would like to see which factors are accociated with the doctor's decision about the treatment. There are 25 factors total.
I am wondering how this should be modeld , what analysis result should be used.
Thank you very much,
05-14-2016 09:19 AM
It sounds like the response variable is a categorical variable with 10 levels. You can analyze data like that by using a generalized logistic model. In SAS you can use PROC LOGISTIC for the analysis. There is an example of a generlized logit model in the documentation for PROC LOGISTIC, along with an explanation of the output, so copy that example. A main-effects model will look something like
MODEL treatment = x1 x2 x3 ... x25 / link=glogit;
where x1-x25 are the 25 factors.
05-14-2016 05:12 PM
You could also consider a discriminant analysis, if you consider each treatment option a group.
You need to factor in the doctor though, there tends to be a bias by provider, not sure how that would be handled. Maybe just as a categorical variable? I would suggest anonymizing the docs names from yourself and until you finalize your results to prevent personal bias.
05-14-2016 10:09 PM - edited 05-14-2016 10:11 PM
You could investigate decision tree procedures. The easiest is the partition platform in JMP. In SAS/STAT, you should try proc HPSPLIT. The skeleton code would look like
proc hpsplit data=myData; target treatment / level=nom; input fact1 fact2 fact3 / level=int; /* List of continuous factors */ input fact4 - fact25 / level=nom; /* List of categorical factors */ output nodestats=myNodes importance=myVars; run;s
05-16-2016 10:45 AM
You might also have to consider the patient's health insurance as some insurance sources may not pay, or only partially pay, for certain procedures and that can influence a provider's choice.