Stata set version not always working (can't rollback some changes from updates)

[Simon Turner]

G'day Statalisters,

Short version: when I set version to an older version, it does not always give me the same results as it does from a previous update.

Medium version: I have code that runs a mixed regression with reml and in my do-file I am interested in the autocorrelation estimate (lag 1) so until recently (sometime in the last few months) you had to take the tanh(<r_atr1_cons estimate in output table>)
Sometime in the last few months an update (not even a version change, just an update) has changed the output table so that the rho estimate is directly in the output table.
This is an excellent change BUT my old code now outputs the wrong values (because now it is running tanh of the actual value).
I have tried setting my version back to a previous version, but with no luck.

Is there a way to rollback an update??? I had hoped that rolling back to a previous version entirely would work, but not in this instance.

I checked by loading my old Stata 16 and Stata 17 on my old laptop (without doing any updates!), ran the code, then updated both.
The Stata 16 (with updates) remained the same.
The Stata 17 (with updates) had changed, but setting version back did not help.

I'm assuming that therefore there was an update specific to the Stata 17 mixed command that may have come out at the same time as the Stata 18 release?

Really frustrating as code that I've included with published papers won't give the same results now even with version control.

Thanks for any advice or comments on this!

Cheers,
Simon.

[Nick Cox]

In essence you can't rollback updates.

Also, the version command is not a time machine. It doesn't rollback e.g. bug fixes or what is considered an improvement otherwise. What it does is documented positively at help version and in intent it is to let you do whatever you might have a really good reason for preferring.

Otherwise put, there is no sense in which version # gets you back to all the behaviour of that version, as if all the code for each previous version was contained within, or alongside, the present.

So if the version command doesn't do what you want, that it is almost always deliberate. In this particular case, I am not in any sense expert on or even familiar with that kind of modelling, but it sounds as if you have a good case to being able to retrieve previous behaviour, so I would get in touch with StataCorp technical support if there isn't an answer from a StataCorp person here in the next day or so.

[Simon Turner]

Thanks Nick, appreciate your comments.

I totally get what you're saying about what version can and can't do, I think in my (maybe strange?) situation the output of the command has effectively changed though the underlying method has not, so the code still "works" and is not "broken" but since the use of the code in further work relies on consistent outputs (which have changed) the end result is that my outputs are now "incorrect" due to an update.
I've sent off a message to technical support as suggested (with some more details) so we'll see what the preferred advice is =)
For now I'll just run my code on my old laptop with Stata 16 until I start an entirely new project where I can update the code and move to a later version!

Again thanks for the feedback,
Cheers,
Simon.

In essence you can't rollback updates.

Also, the version command is not a time machine. It doesn't rollback e.g. bug fixes or what is considered an improvement otherwise. What it does is documented positively at help version and in intent it is to let you do whatever you might have a really good reason for preferring.

Otherwise put, there is no sense in which version # gets you back to all the behaviour of that version, as if all the code for each previous version was contained within, or alongside, the present.

So if the version command doesn't do what you want, that it is almost always deliberate. In this particular case, I am not in any sense expert on or even familiar with that kind of modelling, but it sounds as if you have a good case to being able to retrieve previous behaviour, so I would get in touch with StataCorp technical support if there isn't an answer from a StataCorp person here in the next day or so.