IRT 2PL not replicable across Stata 16 MP on Linux vs Stata 16.1 IC on a PC

[Andy dB]

Hi everyone,

Also posted here. The subject says it all: Results from Stata's internal -irt 2PL- command don't replicate, depending on whether I run the same with Stata 16 MP on Linux vs with Stata 16.1 IC on a PC. Note how the data, version, commands, and seed are exactly the same.

With Stata 16 MP, on Linux:



with Stata 16.1 IC on a PC:



The same issue remains after also setting the sort seed, and sorting on a unique ID variable (with the stable option).

I'd very much appreciate any help,

Andy

[Joseph Coveney]

I don't know whether this is the answer, but there have been some bug fixes to -irt 2pl- since the initial release (16.0), and I think that just specifying -version 16.0: irt 2pl . . .- won't undo the fixes that have occurred in your Version 16.1 Stata and so the output, if it's affected by any of them, wouldn't necessarily match what you get with the older Version 16.0 installation despite the prefix.

Just what are the differences? Have you saved the two resulting datasets and compared the predictions?

[Andy dB]

Thanks a lot! This post explains it: Stata's ML procedure doesn't produce replicable results across platforms. We can close this thread.