PROC MIXED: Same data, same code, different results

[morgodth]

My company does 100% QC on all results programming.  My QC programmer are coming up with different results from a proc mixed (using to get estimates to calcuate a Spearman-Brown Prophecy, but that doesn't matter).  We could not figure out why so eventially I sent him my data set with only the relevent variables and the exact code I used [attached].

1) When I run the code I get a note that "Estimated G matrix is not positive definite." (not good!) and my intercept estimate is zero (0).

2) Whe my colleague runs the EXACT SAME code on the EXACT SAME data, he gets no error and an intercept estimate of 618.93.

I have no idea why!?!?

I sent the code and data to two more colleagues.  To add to frustration, one got my result and the other got my QC person's result.

We all have the same version of SAS (9.4), the same patch, and the same OS (windows).

Why would we be getting different results?

Why would it not be merging on my machine?

Code:

proc mixed data = JMdatii;
class MEMBER_MEMBERID;
model d_stf_score =;
random Int / subject = MEMBER_MEMBERID;
run;

 Attached:

(I tried to attach the data set, but kept geting an error.  If I figure out how to do it then I will in an updated post)

[morgodth]

the data set:

MEMBER_MEMBERID = subject ID

d_stf_score = a score from 0-100 for each subject

[Reeza]

Contact Tech support and include your data.

Without the data no one can really assist. 

I would double check the N of the data to make absolutely sure the same data is going in to each model. I assume each person is starting off from a SAS data set.

[ballardw]

Different hardware? Different math coprocessors can yield different results though usually small. BUT statistics work generates lots of these and accumulated differences may show.

Running on 32bit vs 64bit systems?

[morgodth]

I did contast SAS support, but thought I'd check here to see if anyone else had any ideas or different perspectives until SAS got back to me.  Same N, same data set, and in the output the same 201 patients going into each model.  It's really confusing.

We're all running on 64 bit and with the same HP laptops.

[Reeza]

Can you post the differing outputs?

To attach a dataset you can zip the file and upload it. Without data or results were throwing random guesses out. 

[ShanshanCheng]

Have you solved this problem？I also have the same situation.

[PaigeMiller]

---

@ShanshanCheng wrote:

Have you solved this problem？I also have the same situation.

---

Lots of suggestions for solutions above, please try them.

[ShanshanCheng]

I don't see any solutions above. Are you able to see some things that I can't?

[Reeza]

Couple of things - ensure that categorical variables are character (class variables) and sorted the same on each dataset. 

Out of curiosity, if you run the example the OP posted, what do you get as a result?

[ballardw]

---

@ShanshanCheng wrote:

Have you solved this problem？I also have the same situation.

---

How many of the differences to check have you followed up on? Starting with Operating system versions, hardware differences, co-processors present and version, bitness of the operating system (32 or 64), version of SAS (release and bitness)?

After you verify that all that is identical, then talk to SAS tech support. If there are any differences, that is the likely source.

[ShanshanCheng]

The operating system and SAS version are the same.

if the results are different because of the operating system,such an explanation is difficult to accept.

[ballardw]

---

@ShanshanCheng wrote:

The operating system and SAS version are the same.

if the results are different because of the operating system,such an explanation is difficult to accept.

---

Why is it hard to accept that the operating system has an impact? It controls everything. The instructions on what gets sent to which CPU or the math co-processor bits for example, or the precision of results. Applications call documented parts of the operating systems to do things. If the subsystem changes (bug fix, enhancement, because it's Tuesday) then applications like SAS get to live with the result. IF the systems are well designed, robust and well documented then the OS is unlikely to be a major cause.

I worked in a shop that had for a period of time had to use a specific version of software to prepare files for a printer. Different machines running Windows 95 or Windows 98 (yes it was some time ago) had different results. Out of 16 computers in our work group only ONE could do the file conversion from a Word document and get the expected results. You would think something like a bullet list in document would be pretty stable. Depending on which machine ran the conversion we would get a bullet list with round bullets, the expected result, square bullets, arrows, open triangle or hands pointing to the right. So I am not surprised if major or minor differences in OS can affect output.

I note that there was no response regarding any of the hardware possible impacts.

[morgodth]

I zipped the data (attached)

[morgodth]

output of the two programmers