elementary method functions for multiple imputation with mice

[avehtari]

Summary:

This is an idea for a "new developer friendly" project.
mice (Multivariate Imputation by Chained Equations) package
https://cran.r-https://cran.r-project.org/package=mice
allows user defined elementary imputation methods. It would be quite easy to make functions which would allow use of rstanarm, too.

Description:

Implement Stan versions of elementary imputation methods

• mice.impute.stan.norm
• mice.impute.stan.logreg
• mice.impute.stan.polr
• mice.impute.stan.2l.norm
• mice.impute.stan.2lonly.norm
  based on the corresponding existing functions (without .stan in the name). The existing norm function uses a fixed conjugate prior and analytic posterior, logreg uses glm.fit, polr uses a function from MASS package, and 2l functions use Gibbs sampling. rstanarm would provide more flexibility on priors and better inference.

[bgoodri]

Yeah, there was a postdoc at Columbia who was doing these things for the mice and mi packages. One difficulty is asking the users for priors when each variable with missingness has a turn at being the "outcome". Another is the MCMC is a bit slow for this, so they were doing optimization, which is not quite as reliable.

…

On Fri, Feb 10, 2017 at 6:07 AM, Aki Vehtari ***@***.***> wrote: Summary: This is an idea for a "new developer friendly" project. mice (Multivariate Imputation by Chained Equations) package https://cran.r-https://cran.r-project.org/package=mice allows user defined elementary imputation methods. It would be quite easy to make functions which would allow use of rstanarm, too. Description: Implement Stan versions of elementary imputation methods - mice.impute.stan.norm - mice.impute.stan.logreg - mice.impute.stan.polr - mice.impute.stan.2l.norm - mice.impute.stan.2lonly.norm based on the corresponding existing functions (without .stan in the name). The existing norm function uses a fixed conjugate prior and analytic posterior, logreg uses glm.fit, polr uses a function from MASS package, and 2l functions use Gibbs sampling. rstanarm would provide more flexibility on priors and better inference. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#164>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADOrqtpkZCJvAP1XFU6C8NL-l_bg8WXAks5rbEThgaJpZM4L9QuT> .

[avehtari]

I know people who used Stan for mice style imputation using their own code for chaining. They didn't care it was slow, they wanted the best possible inference. I know there can be also problems of checking all diagnostics for each model fit etc.