I'm trying to check that the proportional hazards assumption is satisfied with all my variables in my Cox model. I used 2 methods to do this, but they give different results. First method: Add time-dependent variable to the original model (in this case, a product of a variable of interest and logarithm of time variable was added for each covariate). prog phreg data=data covsandwich;
class drinker meds;
model time*event(0)=drinker meds comorbs age avg_glucose avg_creatinine t_drinker t_meds t_comorbs t_age t_avg_glucose t_avg_creatinine;
t_drinker=drinker*log(time);
t_meds=meds*log(time);
t_comorbs=comorbs*log(time);
t_age=age*log(time);
t_avg_glucose=avg_glucose*log(time);
t_avg_creatinine=avg_creatinine*log(time);
id pracid;
run; This method showed that proportional hazards assumption was satisfied for all variables because none of these time-dependent variables were significant. Second method: for continuous variables - plots of Shoenfeld residuals against the time variable for categorical variables - plots of log-minus-log of the survival function against the time variable, stratified by the category levels. e.g. proc lifetest data=data plots=(s, lls);
time time*event(0);
strata drinker;
run; This method showed that all the continous variables satisfied the proportional hazards assumption (graghs showed straight line), whilst it did not for one of the categorical variable (drinker) as the lines of the survival function crossed. The first method and second method showed slightly different results (all covariates satisfy the PHA in the first, whilst it doesn't for the second). Could this be because the second analysis is unadjusted? I wonder whether it's fair to conclude that the PH assumption is satisfied as indicated by the first method, or whether it isn't because of the second method. Many thanks for your help in advance.
... View more