Hello,
I've performed multiple imputation and fit my statistical model to each imputed set. Unfortunately, my statistical model is complicated and fails to converge some of the time. As a result, when I apply PROC MIANALYZE, I get an error message that parameter estimates are missing for some imputed sets.
Does anyone know whether there's an option to pool the estimates that DID converge? Is my best bet to pool the estimates by hand?
Thanks in advance for the input!
Madeline
You would have to manually drop the estimates from the models that didn't converge. Once you do that you should be able to use Proc MIANALYZE to combine the estimates from the rest of the models that did. The only caveat is that MIANALYZE requires that there be consecutive numbering of the values in the _IMPUTATION_ variable so you will also need to re-number them once you omit the models that did not converge.
Hello,
It's a mystery to me why you posed this question in the 'Graphics' board under 'Programming'.
The 'Statistical Procedures' board under 'Analytics' is the appropriate place. Can somebody with super-powers move this thread to the 'Statistical Procedures' board under 'Analytics'? Thanks.
To answer your question:
Is your statistical model a mixed model (fit with PROC MIXED or PROC GLIMMIX or PROC NLMIXED)? I would work on the convergence issue first. By using a PARMS statement for example. Or by using an NLOPTIONS statement.
Don't try to do the PROC MIANALYZE by hand (that is with regular programming, not on paper obviously) for the parameter estimates you do have. It's a nice exercise to find out how exactly PROC MIANALYZE works behind the scenes but it takes a LOT OF TIME.
Cheers,
Koen
So sorry about that. I forgot to change the settings from the last question I posed.
Thank you for your response. I've already worked on the convergence problem, and it's probably as good as it's going to get. I'm using PROC NLMIXED, and the model has three random effects, so I'm not surprised there are convergence issues.
I suppose this leaves me with no choice but to obtain the pooled estimates "by hand".
You would have to manually drop the estimates from the models that didn't converge. Once you do that you should be able to use Proc MIANALYZE to combine the estimates from the rest of the models that did. The only caveat is that MIANALYZE requires that there be consecutive numbering of the values in the _IMPUTATION_ variable so you will also need to re-number them once you omit the models that did not converge.
Thank you! I considered re-numbering and moved to obtaining the pooled estimates "by hand" when I couldn't figure out how to re-number by replication (this is a simulation with 500 replications each containing 20 imputed sets).
Build your skills. Make connections. Enjoy creative freedom. Maybe change the world. Registration is now open through August 30th. Visit the SAS Hackathon homepage.
Register today!ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.
Find more tutorials on the SAS Users YouTube channel.