mice does not support post-processing for blocks of variables

[Garyf20]

The processing of "post" is missed for multivariate imputation.

[gerkovink]

Cannot reproduce on mice v3.13. Could you present a reprex?

set.seed(123)
library(mice)
library(magrittr)
version()
#> [1] "mice 3.13.0 2021-01-26 /Library/Frameworks/R.framework/Versions/4.0/Resources/library"

ini <- mice(boys, maxit = 0)
meth <- ini$meth
meth["tv"] <- "norm"
post <- ini$post
post["tv"] <- "imp[[j]][, i] <- squeeze(imp[[j]][, i], c(3, 7))"
post["bmi"] <- "imp[[j]][, i] <- squeeze(imp[[j]][, i], c(28, Inf))"
post["hgt"] <- "imp[[j]][, i] <- squeeze(imp[[j]][, i], c(50, 55))"
imp <- mice(boys, meth=meth, post=post, print=FALSE)

imp$imp$tv %>% range
#> [1] 3 7
imp$imp$bmi %>% range
#> [1] 28 28
imp$imp$hgt %>% range
#> [1] 50 55

^{Created on 2021-03-25 by the reprex package (v1.0.0)}

[gerkovink]

Please note that post-processing only works for imputed values; the observations are not processed.

[stefvanbuuren]

This is not a bug. mice does not support post-processing for blocks of variables. The length of the post vector equals the number of variables', and post works only in combination with a series of univariate models.

I found it hard to imagine a generic way to implement post-processing under block-wise imputation models. I'll be interested to hear of any suggestions, preferably including practical examples that illustrate the need for such block-wise post-processing.

[stefvanbuuren]

Commit 8f78a27 adds the explanation to the documentation.

[Garyf20]

Thank you for your response. I also have the following questions. As pointed out in the block notes, variables within a block are imputed by a multivariate imputation method. The problem is that the "make.method" function picks up a univariate imputation method. So my questions are as follows. 1) In that default method case, is there any difference between blocks with more than 1 variable setting and 1 variable in 1 block setting? 2) If the results are similar. Using blocks plus default methods will have the drawback that only 1 default method can be applied to a block. Also, it is confusing that the default methods don't contain a multivariate one. And the univariate methods can be applied to the blocks with more than 1 variable. Looking forward to your response. Thank you. Best, Gary

…

On Wed, Mar 31, 2021 at 10:54 AM Stef van Buuren ***@***.***> wrote: Commit 8f78a27 <8f78a27> adds the explanation to the documentation. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#326 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARN45ODSOCNPRZGAVLIUSEDTGMZRRANCNFSM4Z2A7NYQ> .

[stefvanbuuren]

The mice philosophy is to solve the multivariate imputation problem by repeatedly applying univariate variable-by-variable imputation. So mice solves the multivariate imputation problem, and nothing special is needed to do that.

The blocks facility is more of an experimental feature meant for cases where you wish model two or more variables jointly, so you would impute two or more values per row by one method.

Hope this helps.

[stefvanbuuren]

Question: In that default method case, is there any difference between blocks with more than 1 variable setting and 1 variable in 1 block setting?

No difference. mice simply repeats any given univariate method within the block. Be aware that it is possible to change the sequence in which variables are imputed. See example below.

suppressPackageStartupMessages(library(mice))

imp1 <- mice(nhanes, seed = 1, m = 1, maxit = 2, print = FALSE)
imp2 <- mice(nhanes, blocks = list(c("bmi", "hyp"), "chl"), m = 1, maxit = 2, seed = 1, print = FALSE)
identical(complete(imp1, 1), complete(imp2, 1))
#> [1] TRUE

imp3 <- mice(nhanes, blocks = list(c("hyp", "bmi"), "chl"), m = 1, maxit = 2, seed = 1, print = FALSE)
identical(complete(imp1, 1), complete(imp3, 1))
#> [1] FALSE

imp4 <- mice(nhanes, visitSequence = c("hyp", "bmi", "chl"), m = 1, maxit = 2, seed = 1, print = FALSE)
identical(complete(imp3, 1), complete(imp4, 1))
#> [1] TRUE

^{Created on 2021-04-01 by the reprex package (v1.0.0)}

[stefvanbuuren]

Comment: If the results are similar. Using blocks plus default methods will have the drawback that only 1 default method can be applied to a block.

That's correct, but you should think of it as a feature rather than as a drawback. The whole idea of block imputation is that you can use one method (be it univariate or multivariate) to impute multiple variables at the same time.

[Garyf20]

Yes. For the users who want to use blocks, is it possible to assign a default multivariate imputation method (jomoImpute) to the blocks with more than 1 variable? Best, Gary

…

On Thu, Apr 1, 2021 at 9:00 AM Stef van Buuren ***@***.***> wrote: Comment: *If the results are similar. Using blocks plus default methods will have the drawback that only 1 default method can be applied to a block.* That's correct, but you should think of it as a feature rather than as a drawback. The whole idea of block imputation is that you can use one method (be it univariate or multivariate) to impute multiple variables at the same time. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#326 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARN45OAYN2TXVHRTT5FBXMLTGRU5FANCNFSM4Z2A7NYQ> .

[stefvanbuuren]

I would do that by

mice(nhanes, blocks = list(c("bmi", "hyp"), "chl"), method = c("jomoImpute", "pmm"))

I also tried to set defaultMethod but that errors on chl, probably because it is univariate. But the above just works fine.