When using PROC PHREG, you can output Schoenfeld residuals to assess the proportional hazards (PH) assumption in Cox regression models with the output statement. This behavior is supported for standard Cox models, but it is not supported when you use the RANDOM statement (frailty models).
This behavior applies to SAS/STAT® PROC PHREG across supported versions.
Common symptoms of this issue include the following:
- an OUTPUT statement creates an empty data set
- the SAS log displays NOTE: The OUTPUT statement is ignored for a frailty model analysis
When you specify the RANDOM statement, PROC PHREG fits a shared frailty model using a penalized partial likelihood approach. In this context, residual‑based diagnostics are not supported.
When you use RANDOM, the following statements are ignored:
Workaround
Method 1: Use Covariate × Time Interactions
When frailty or clustering is important, explicitly model non‑proportionality using time‑dependent effects. A significant interaction indicates violation of the PH assumption.
Method 2: Use a Marginal Cox Model (Without RANDOM)
To generate Schoenfeld residuals, fit the model without the frailty term, which allows assessment of PH assumptions marginally (without clustering).
Additional Information
Schoenfeld Residuals in PROC PHREG
See the Residuals documentation for residuals descriptions (including Schoenfeld residuals).
Frailty Model Limitations in PROC PHREG
See the Frailty Model documentation, which states that the OUTPUT, ASSESS, AND BASELINE statements are ignored when you specify RANDOM.