PROC LIFEREG Viya vs Enterprise Guide

[sasuser12345671]

Our company recently converted from Enterprise Guide to Viya. We use PROC LIFEREG for our model fitting processes and noticed that we get different output in Viya than we did in Enterprise Guide with the same data and parameter specifications. Were there any changes to PROC LIFEREG in Viya that would cause these differences?

Below is an example of the code we're running:

proc lifereg data=paid_trended outest=outdata;

    model total_paid*censor(0)= /d=gamma;

    where total_paid > 1600;

    output out=a p=median std=s;

run;

[Rick_SAS]

>  we get different output in Viya than we did in Enterprise Guide 

1. It sounds like you moved from SAS 9.4 to SAS Viya, is that correct?
2. Can you describe the differences (or post images of them)? Are the differences in tables? Graphs? The OUTEST= data set? If the difference appears in tables, use ODS TRACE ON to discover the names of the tables.
3. Do you see the differences for a large variety of data sets or only for one data set?

[sasuser12345671]

1. We moved from SAS 8.3 to SAS Viya

2. The model parameters that come out of the fitting process are different (intercept, scale, shape), even if we use the same dataset. The OUTEST= dataset is where we've noticed the differences.

3. This is the only dataset that we run this process on, so we're unable to test across different datasets. 

[Rick_SAS]

OK. Can you post the before/after estimates? I'd like to know if they are close (for example, differ in the 4th decimal place) or are vastly different.

I assume the server that SAS runs on has also changed? For example, you used to run on a Windows Workstation but now you are running on a Linux server on AWS in the cloud?

[sasuser12345671]

Yes, the SAS server also changed. 

Below are the parameters we got using EG:

And below are the parameters we got using Viya:

[Tom]

---

@sasuser12345671 wrote:

Yes, the SAS server also changed. 

Below are the parameters we got using EG:

And below are the parameters we got using Viya:

---

That does not look different to me.

PS  I doubt you were previous using 8.3 of SAS.  I don't think there was such a version and if so it was replaced by 9.0 over 22 years ago.  Perhaps you meant to say you were using an 8.3 release of the tool (Enterprise Guide) that you used to submit your code to SAS.

[Rick_SAS]

This is all conjecture, but I think you'll find that the predictions from both models are essentially equivalent. It is also likely the differences between the predicted values from these models are smaller than the errors in the data.

Recall that the parameter estimates are the solution of a maximum likelihood optimization. When we optimize a function, we iterate until some criteria are met. If you add the ITPRINT option to the MODEL statement, you can compare the optimization path used by the old and new version of SAS. My guess is that you will see a very similar path, with minor differences explained by the newer chips on the new server.