mi estimate and xtmixed - model does not converge

[fran]

After running a multiple imputation using mi ice, I used the xtmixed command as follows:

xi: mi estimate: xtmixed fQPR_tot bQPR_tot i.Intervention i.wave || team: , mle variance

however, the model crushes and I receive the following error message:

model did not converge on
m=37

I run it again with set trace on but the error message is not helpful (r498) in letting me know what is actually wrong with m = 37.

In a previous thread, Stas Kolenikov suggested that one could "Re-impute that particular replicate (I heard some people call it implicate)."

I'd be grateful on any suggestion on how to achieve this.

Cheers,
Fran

[Timothy Mak]

Which version of Stata have you got? If you can run mi estimate, your version should be at least 11. This means that xi should be unnecessary.

I don't think Stata has a function for replacing a particular imputed dataset with another one. However, if you use mi impute, you can use the add(1) option to add another imputed dataset. You can then make it into wide format and replace the problematic imputation with the new. (Then delete the new dataset.)

But I don't see why you should re-impute a particular dataset if it is problematic. Why not just tell people that one of the datasets was problematic, and so we just use the remaining 49 datasets for inference (assuming you have 50 to start with)? To do this, you only have to specify the errorok option when running mi estimate. But perhaps Stas can advise us on the benefits of "implication".

Best,
Tim

[fran]

Thanks Tim - this is really helpful!