Seeds do not reproduce my results...

eniseg2

Hello dear JASP team!

I've been experimenting with setting seeds for Bayesian analyses. I've set my seed in one data file, and then created another JASP file with the very same data. I re-ran the analyses with the very same seed and got different results - some of my Bayes factors have changed by as much as six! Is this something to worry about?

Kind regards,

Anna

EJ

This is strange -- I'll make some enquiries

E.J.

František

Hi Anna,

thank you very much for contacting us. Could you please describe what specific analysis lead to different results after re-running it with a specified seed?

Cheers,

Frantisek

eniseg2

Hi Frantisek!

These were repeated-measures analyses of covariance. Essentially, I had one JASP file full of analyses and I was struggling to wrap my head around what would be the best analytical approach to choose. Eventually, I settled and wanted to have one neat and tidy file, but keep the untidy file. So I opened the very same data set in JASP in a fresh file, entered the analyses with the same seed and voilà, different results...

Cheers,

Anna

František

Hi Anna,

thanks for the quick response. I tried to reproduce your finding using our test datasets but I did not manage to do so. Could I ask you to save the .JASP file from the analyses as well as a screenshot that marks the results that changed, so I can re-run it on my computer? It would greatly help me in solving this issue.

You can send me the files and screenshot at f.bartos96@gmail.com. I won't share the .JASP file or the screenshot with anyone else.

Cheers,

Frantisek

eniseg2

Hi Frantisek,

I'm sorry, I just saw this. I will email you over the course of the next few days.

Best,

Anna

František

Hi Anna,

no problem, looking forward to your email.

Cheers,

Frantisek

agata_

Hi, I just saw this, I'm playing with the Bayesian regression as well as was wondering whether this has been solved?

Thanks

František

Hi Agata,

Unfortunatelly, I did not receive the email at end and we did not manage to reproduce it internally. If you happen to achive the same problem, let us know please!

Best,

Frantisek

agata_

Hi again, I haven't actually been able to reproduce the error, but I do remember encountering this and I wonder if this can be the question of workflow? Because when I run a regression and then set seed, it doesn't change, so the initial random BF doesn't change. On the other hand, when I set seed before adding predictors, shouldn't it give me a different, non-random result?

František

Hi Agata,

what you describe should not be the case. Especially in Bayesian Linear Regression where the BF are obtained by quadrature and are not dependent on the seed (the seeds matters for the posterior distribution and interval estimates when the MCMC method is selected). I also tried what you described with Bayesian ANOVAs where the BF computation is dependent on seed. Before setting a seed, I get a BF corresponding to a random seed, after changing the seedm the model is automatically recomputed and produces the same BF even if the analysis is refreshed. This is also the case if you set seed before adding predictors into the model - which gives the same result as if I added the predictors before setting the seed. Or did you had a different combination in mind that I did not check?

Cheers,

Frantisek

agata_

I see, I just wanted to make sure, this is good to know. I think that dispels my doubts, thanks!

eniseg2

Hi, I'm really sorry to ring the alarm bells here and then to be offline for so long. After I re-installed JASP, the seed-setting worked just fine, thankfully. I'm not sure what happened but I was transferring the files between Linux at work and my private Mac in my "home office", and maybe something got scrambled there? Now that I've abandoned Linux, I haven't encountered the error again.

František

Hi eniseg2,

thanks for getting back to us and glad to hear that you did not experience a similar issue again.

Regarding the file transfer between MacOS and Linux, I do not know whether that could have caused the issue but I will forward it to our team and we will check it out.

Cheers,

Frantisek

František

So, I got an answer regarding the same seed usage across different OSes.

Setting a seed does not guarantee identical results across different OSes. Even completely deterministic algorithms may return different results across different OSes. There are multiple possible reasons for this, but it is not a concern for e.g. MCMC, as long as the chains converge to the same target distribution (e.g. the quantiles converge to similar values after many, many iterations).

Another thing that differs between OSes is rounding error (i.e., https://en.wikipedia.org/wiki/Round-off_error). This might play more of a role especially in cases where one of the compared models does not describe the data well and its marginal likelihood is hard to estimate. You mentioned that the BF differed by 8, it might sound like a lot, but the relevance is practically irrelevant if the original BF was on the scale of thousands.

Hope this provides additional clarification,

Frantisek

eniseg2

Hi Frantisek,

Thank you so much for getting back to me about this issue. I'll make sure not to convert between Linux and Mac regularly to avoid any issues with replication, and to always export & save an HTML file.

All the best, eniseg