R help - Feb 2007 - Coxph and ordered factors

Dear useRs,

I am trying to fit a Cox PH model on survival data from a lung cancer 
dataset. I would like to include the patient staging (I-IV) as a 
covariate. For this I use the following function:
coxph(Surv(time,status) ~ stage)
The staging information is a categorical variable, and it is important 
to take the ordering into account (I<II<III<IV). Does the coxph
function
handle correctly ordered factors, or is it preferable to recode the 
covariate into numerical values (1-4)?

Thank you very much for your help,

Florent

-- 
--------------------------------------------------
		Dr Florent BATY
Pulmonary Gene Research, Universit?tsspital Basel
   Petersgraben 4, CH-4031 Basel, Switzerland
 tel: +41 61 265 57 27 - fax: +41 61 265 45 87

No, coxph does not have any special code to deal with ordered factors.
 You can do it "by hand" by using the pool-adjacent-violators
algorithm
    1. Fit the data with stage as a categorical
    2. Verify that the coefficients are ordered
       if so, you are done
       if not, assume that they were
       	     0 = stage 1  (implicit coef)
       	     .1= stage 2
       	     .5 = stage 3
       	     .4 = stage 4
       	  Create a new class variable that combines stages 3 and 4, and
       	  return to step 1.
       	  
       	  
       	  Terry Therneau

Dear Terry,

Thanks for your answer. I realised that depending on whether or not I am 
using an ordered factor I have some substantial changes in the output of 
function coxph:

Here is the output I obtain when using variable 'stage2' as an ordered 
factor:

Call:
coxph(formula = Surv(survTime, status) ~ stage2, data = survData)


            coef exp(coef) se(coef)       z       p
stage2.L  2.3430    10.413    0.709  3.3033 0.00096
stage2.Q -0.0498     0.951    0.629 -0.0792 0.94000
stage2.C -0.0468     0.954    0.541 -0.0866 0.93000

Likelihood ratio test=30.2  on 3 df, p=1.25e-06  n= 61

Here is the output when using 'stage2' without ordering:

Call:
coxph(formula = Surv(survTime, status) ~ stage2, data = survData)


        coef exp(coef) se(coef)     z      p
stage22 1.06      2.87     1.16 0.914 0.3600
stage23 2.17      8.73     1.10 1.975 0.0480
stage24 3.12     22.70     1.03 3.037 0.0024

Likelihood ratio test=30.2  on 3 df, p=1.25e-06  n= 61

Finally here is the output when coding 'stage2' as a numerical variable:

Call:
coxph(formula = Surv(survTime, status) ~ stage2, data = survData)


       coef exp(coef) se(coef)    z       p
stage2 1.03      2.79    0.226 4.53 5.8e-06

Likelihood ratio test=30.2  on 1 df, p=3.94e-08  n= 61


It seems that when using ordered factors a model with linear, quadratic 
and cubic terms is fitted. Is that something expected? Can't we 
specifically focus on a simpler model including only linear terms? I am 
not sure to have a full understanding of what is actually done when 
factors are ordered.

Best wishes,

Florent

-- 
--------------------------------------------------
		Dr Florent BATY
Pulmonary Gene Research, Universit?tsspital Basel
   Petersgraben 4, CH-4031 Basel, Switzerland
 tel: +41 61 265 57 27 - fax: +41 61 265 45 87