Assessing for PH - should I test one variable at at time? I know i'll need to use parametric

[Denise Vella]

Hello there, I am using COX regression - Outcome: Readmissiontohospital. I've got covariates : age, gender, experience, institution, socialequality.

I'm fairly sure this won't show PH hazards. Would you recommend doing a PH hazard test for each covariate separately as seen in the covariate1 code? Then if it satisfied PH - then include in coxregression in the adjust. If it doesn't either use interaction or stop wasting time (as I know it wont satisfy COX regression and use parametric)


You may be wondering, why I'm bothering to test each covariate to see if it assessed PH when i'M FAILY sure a parametric test would be better, however, ultimately I would like to compare the results from COX to a parametric test and show parametric test is a better way to test- hence this hassle.

Now with regards to assessing PH for each covariate, would you recommend performing the goodness of fit, graphical interpretation for EACH covariate then adjusting as seen in (phase 2)

stcox readmission covariate1
estate phtest //schofields 
stphplot,by(readmission) adjust(covariate1) //adjusting if satisfy PH (phase 2)

[Tiago Pereira]

Hi, Denise.

I recommend this book (or a more recent version), because it addresses exactly what you need:

Kleinbaum, D. G., & Klein, M. (1996). Survival analysis a self-learning text. Springer.

[Denise Vella]

Hi Tiago Pereira thanks for your answer - this is the book I'm reading. However, it's not clear
In his mathematical section on p 171 - he takes different variables separately into his models and assesses PH (log WBC assessed) followed by Treatment then followed by gender (which does not satisfy PH)
Each one is considered separately and independently (hence my idea)

However in the computer appendix - for the graphical assessment
again he only takes clinic as the covariate and assesses it independently (p 535)

stphplot,by (clinic) //here we look to see if clinic satisfies PH

Despite explaining what we should be looking out for, he does not explain or go into if we should do the above for other covariates eg prison
Later on he says to use stcoxkm (to look at observed vs predicted) but only does this for the clinic variable. Ths is another test done to assess for PH

Later on he says to use this:

stphplot, strata(clinic) adjust(prison dose)  //Does this mean the variables after adjust are those than have already been proven to be ok for PH?

Also the statistical GOF shown on p 546

stcox prison dose clinic, schoenfeld(schoen*) scaledsch(scaled*)  //where here all the covariates are taken together, is this taken into consideration that all variables have been tested graphically and thus taken all together

So ultimately my question is:
1. Is it silly to test for each covariate with stphplot , or is this wasting my time ?
2. Ultimately I will apply the GOF with all variables together


Finally I know it does not satisfy the cox regression and will need to use parametric but would like to show differences in the output.

The book was written by two separate authors, with Daivd having written the mathematical sections nd computer appendix by the second author hence why they may not marry up

[Denise Vella]

Actually the answer to the above should be :

1. Fit model first (stcox) - ALL covariates
2. Calculate aic
3. Check PH - GOF test using estat (all variables). If one variable violates it can be double check subjectively with KM observed vs predicted using stphplot separately with each variable on its own

Depending on how many variables violate it decide to go for Parametric test




That book is excellent but it can sometimes be too complex and one can get bogged down with text and details




After reading the book here’s a site i found that brings together everything in one page




https://onezero.blog/survival-analys...and-aft-model/