When PH model assumption is violated I have some questions about cox PH  model. We have a dataset contains follow-up time, status (death or not)/ censing indicator , and some explanatory variables (all categorical variables). We want to fit the Cox PH model to predict the risk. 
We checked the PH assumption  with R ‘cox.zph’ function, from the result, all individual tests have large p-values (all greater than 0.1, with one 0.042 and one 0.059) however the global test has p-value=0.02. Is that means our data is  the PH assumption badly, and we cannot use the traditional PH model?
But we still fitted the Cox PH model using selected covariates anyway, then I checked the deviance, result is attached, the deviance plot has two distinct clusters,  the one above are those individual that dead(not censored), the one below are those individuals not dead (censored). (the censoring rate in this data is about 90%......) . What this plot indicates? Can we use some method to address the problem and still use PH model? Or we shouldn’t use PH model at all?
Thanks for your help!!!
Best wishes,
 A: Your deviance residuals demonstrate that as the linear predictor increases, the risk of an event decreases. The separation of the censored observations from the events is not a problem necessarily. 
If you want to examine the proportional hazards assumption graphically, consider plots of the Schoenfeld residuals against time. Schoenfeld residuals have an intuitive interpretation of the difference in model estimated (covariate-wise) probability between an individual having an event and all persons remaining in the risk set at a given time. If a non-random pattern is evident upon plotting these residuals against time, then you may have violated the PH assumption. 
If the PH assumption is violated, you may be able to either stratify the model by the problem variable or include an interaction between time and the problem variable to adjust for its time-dependent effect. Note, the stratification solution is better for confounders than main effects, as stratification prevents interpretation of the stratifying variable.  
