different results by sas releases

[geneshackman]

I want to ask, are there any listings or discussions of getting different results from different releases of SAS 9.4? We seem to be getting slightly different results from proc surveyfreq, depending on whether the person uses The TS1M5, X64_10PRO and the TS1M6, X64_10PRO  versus the TS1M3, X64_8PRO.

For the first set, the computer is creating rows of observation for certain indicators that shouldn’t have those rows. In the second set, when two people ran the same program, 0.84% (276 out of 33008) of our observations have a confidence limit(s) that differs by a decimal place. 

All of us are running exactly the same program. 

I don't really want to share the program, it's too complicated, the data set isn't really sharable. I just want to ask if there have been reports of people getting slightly different results from specific procs, when they use different releases of SAS.

Thanks

[Rick_SAS]

Since you don't want to share the code or data on the forum, you might want to contact SAS Technical Support. They can tell you whether there were known changes that would explain what you are seeing.

In general, the answer to your question is yes. There can be differences between releases. Sometimes bugs are discovered and fixed, both in SAS and in computational libraries that SAS uses. Technical Support can provide more information about your specific case.

[ballardw · Accepted solution]

Did you run the program on the exact same computer? Hardware can have an effect based on specific math co-processors and such.

Plus, SAS is always tweaking code to improve things though typically I wouldn't expect drastic differences. And how much is an actual "decimal point change". Please provide a few specific values.

The display formats may well be showing the difference between lower order decimals than you see. Example:

data example;
  x = 14.0051;
  y = 14.0049;
  put x= 6.2 y=6.2;
run;

The actual values between the x and y are .0002. But the format chosen here displays them as:

x=14.01 y=14.00

which you might think is a 0.01 difference. So without data and the methodology of exactly how you compared results it is difficult to see if you actually have a "serious" difference in values.

The only thing I see that makes me wonder what you may be doing is the "For the first set, the computer is creating rows of observation for certain indicators that shouldn’t have those rows." Which unfortunately does bring up data content as well as actual code to address.

If you are getting output that you shouldn't that would be an item for Tech Support to address. They will want code and data to diagnose things.