During the validation of one of our studies work, the study team noticed a discrepancy in some of the output of the PHREG procedure. In particular, when using PHREG to analyze time to multiple events, we noticed some differences in statistics between SAS 9.1.3 and SAS 9.2.
Sample code:
proc phreg data=phreg.testdata covs(aggregate);
model (startdt, enddt)* censor(0) = trt / ties=efron r1;
strata strata1 strata2 strata3;
id usubjid;
ods output globaltests=g1 parameterestimates=p1 fitstatistics=f1;
run;
|
|
Std Error of Treatment
|
Score (Sandwich)
P-value
|
Wald (Sandwich)
P-value
|
|
SAS 9.13
|
0.19461
|
0.5768
|
0.5841
|
|
SAS 9.2
|
0.19544
|
0.5737
|
0.5857
|
By working with SAS technical support, we discovered the reason for the difference:
- Prior to SAS9.2 (TS1M0[1st version of 9.2]), the score residuals are computed by the Breslow approximation regardless of the TIES= option.
- Starting in SAS9.2 (TS2M0), the score residuals are computed by the Efron approximation when TIES=EFRON is specified.
- The robust variance estimates are based on the score residuals and therefore explain the differences between 9.1.3 and 9.2 for TIES=EFRON. If you use the default TIES=BRESLOW, you won't see any differences.
I wanted to share this finding with other statistical programmers and data analysts who are doing this type of survival analysis. Hopefully it will save you some time.