PROC PHREG hazard ratio is different when the same code is executed in Windows and in Linux

[mbertol4]

Hi, we’re wondering about the difference in results between Windows and Linux using PROC PHREG to get the hazard ratio. The 5th decimal place is different. The input datasets are identical in both cases. Do you know about any documentation available that can explain this behavior?

[SteveDenham · Accepted solution]

You will probably want to look at the documentation for floating point representation for the two operating systems.  If those differ (and I assume they will), then very small differences in the binary representation of the data may accumulate during iterative procedures to the point that the estimates at convergence of location and variance differ a bit, and thus you get differences in the hazard ratio and any associated interval estimates or p values.

SteveDenham