Help for old friends

This documents describes the changes between mice 2.46.0 and mice 3.0.0. Most code written for versions mice 2.12 - mice 2.46.0 should run unchanged. I have tried to minimize the changes to the function arguments, but it was not possible to remain 100% backward compatible. This document outlines the visible changes, and suggests ways how to adapt old code to mice 3.0.

`mice` function arguments

Argument	2.46.0	3.0.0
method	length `ncol(data)`	length `length(blocks)`
predictorMatrix	`ncol(data)` rows and columns	`length(blocks)` rows, `ncol(data)` columns
blocks	not defined	new argument
visitSequence	`integer`, arbitrary length	`character`, arbitrary length
form	character, length `ncol(data)`	deprecated
formulas	not defined	named list of formulas
blots	not defined	named `list` of `alist`
diagnostics		deprecated
imputationMethod		deprecated, use `method`
defaultImputationMethod		deprecated, use `defaultMethod`
…

No changes were made for the following arguments: data, m, where, post, defaultMethod, maxit, printFlag, seed and data.init.

If blocks is not specified, then each variable is allocated to a separate block. In that case, length(blocks) is identical to ncol(data), and the method in mice 3.0.0 reduces to variable-by-variable imputation, as in mice 2.46.0 and before.

Argument visitSequence may still be specified as integer or numeric, but it will internally be converted into character using the column names in data.

An existing function call to mice using the old form argument may result in an error Argument "formulas" not a list. The advice is to specify the formula as a list, e.g.,

library(mice, warn.conflicts = FALSE)
imp <- mice(nhanes,
  formulas = list(
    hyp ~ bmi,
    chl ~ age + hyp + bmi,
    bmi ~ age + hyp + chl
  ),
  print = FALSE, m = 1, maxit = 1, seed = 1
)
imp$formulas
#> $hyp
#> hyp ~ bmi
#> 
#> $chl
#> chl ~ age + hyp + bmi
#> 
#> $bmi
#> bmi ~ age + hyp + chl

`pool()` function uses `broom`

mice 2.46.0 and before used coef() and vcov() to extract the parameters from the complete-data model. There were two problems with this approach: 1) coef() or vcov() were often not defined for the analysis of interest, 2) their output was not standard across procedures. Older versions of mice therefore needed quite some custom code to extract the parameters and their variance estimates.

mice 3.0.0 uses the broom package for this task. The advantage is that the standardised output. The downside is that (still) many packages offer no defined tidy.xxx() and glance.xxx() functions. In those cases, the user sees the error message Error: No tidy methods for objects of class xxx. See the pool() documentation for what to do in those cases.

`pool()` for mixed models requires `library(broom.mixed)`

mice automatically loads the broom package. Tidiers for mixed models live in the broom.mixed packages and this is NOT automatically loaded. If you want to pool the results of mixed model, issue library(broom.mixed) before calling the pool() function.

Help for old friends

2024-10-16