[R] Trouble Computing Type III SS in a Cox Regression

[Terry Therneau]

You've missed the point of my earlier post, which is that "type III" is not an answerable 
question.

    1. There are lots of ways to compare Cox models, LRT is normally considered the most 
reliable by serious authors.  There is usually not much difference between score, Wald, 
and LRT tests though, and the other two are more convenient in many situations.

    2. "Type III" is a question that can't be addressed. SAS prints something out with 
that label, but since they don't document what it is, and people with in-depth knowlegde 
of Cox models (like me) cannot figure out what a sensible definition could actually be, 
there is nowhere to go.  "How to do this in R" can't be answered.  (It has nothing to do 
with interactions.)

   3. If you have customers who think that the earth is flat, global warming is a 
conspiracy, or that type III has special meaning this is a re-education issue, and I can't 
much help with that.

Terry T.

On 04/25/2013 07:59 AM, Paul Miller wrote
> Hi Dr. Therneau,
>
> Thanks for your reply to my question. I'm aware that many on the list do not like type III SS. I'm not particularly attached to the idea of using them but often produce output for others who see value in type III SS.
>
> You mention the problems with type III SS when testing interactions. I don't think we'll be doing that here though. So my type III SS could just as easily be called type II SS I think. If the SS I'm calculating are essentially type II SS, is that still problematic for a Cox model?
>
> People using type III SS generally want a measure of whether or not a variable is contributing something to their model or if it could just as easily be discarded. Is there a better way of addressing this question than by using type III (or perhaps type II) SS?
>
> A series of model comparisons using a LRT might be the answer. If it is, is there an efficient way of implementing this approach when there are many predictors? Another approach might be to run models through step or stepAIC in order to determine which predictors are useful and to discard the rest. Is that likely to be any good?
>
> Thanks,
>
> Paul