Using imputation models already fitted to impute missing values in new individuals?

[eusebe]

Hi,

I was wondering if it is possible (and if yes, how) to implement one of the methods described in https://onlinelibrary.wiley.com/doi/10.1002/sim.8682 (open access) for the imputation of missing predictor values in new individuals with missing data. More precisely, I would like to implement the method 6, which use the vector of parameter estimates for each of the fully conditional models as derived in a development dataset (where missing values were imputed by MICE) to impute missing data in new individuals. The method is described in the last paragraph of section 2.3.
I guess this is not straightforward, but is it possible using the mice package?

Best,
David

[stefvanbuuren]

There was an old request for this feature at #32. See the discussion - and our solution using the ignore argument - to split the data into a train set (on which we fit the imputation model) and a test set (on which we apply the imputation model). Would that be useful for your case?

[eusebe]

I'm going to dive in, thanks!

[stefvanbuuren]

Here is a small reprex demonstrating two ways in which we may fit the imputation model in the train data, and apply it to the test data.

library("mice", warn.conflicts = FALSE)

set.seed(123)
ignore <- sample(c(TRUE, FALSE), size = 25, replace = TRUE, prob = c(0.3, 0.7))

# scenario 1: train and test in the same dataset
imp <- mice(nhanes2, m = 2, ignore = ignore, print = FALSE)
imp.test1 <- filter(imp, ignore)
imp.test1$data
#>      age  bmi  hyp chl
#> 2  40-59 22.7   no 187
#> 4  60-99   NA <NA>  NA
#> 5  20-39 20.4   no 113
#> 8  20-39 30.1   no 187
#> 11 20-39   NA <NA>  NA
#> 16 20-39   NA <NA>  NA
#> 20 60-99 25.5  yes  NA
#> 21 20-39   NA <NA>  NA
#> 24 60-99 24.9   no  NA
complete(imp.test1, 1)
#>      age  bmi hyp chl
#> 2  40-59 22.7  no 187
#> 4  60-99 21.7 yes 184
#> 5  20-39 20.4  no 113
#> 8  20-39 30.1  no 187
#> 11 20-39 28.7  no 229
#> 16 20-39 26.3  no 184
#> 20 60-99 25.5 yes 186
#> 21 20-39 28.7  no 199
#> 24 60-99 24.9  no 218
complete(imp.test1, 2)
#>      age  bmi hyp chl
#> 2  40-59 22.7  no 187
#> 4  60-99 27.4 yes 218
#> 5  20-39 20.4  no 113
#> 8  20-39 30.1  no 187
#> 11 20-39 28.7  no 204
#> 16 20-39 29.6  no 206
#> 20 60-99 25.5 yes 284
#> 21 20-39 22.0 yes 131
#> 24 60-99 24.9  no 229

# scenario 2: train and test in separate datasets
traindata <- nhanes2[!ignore, ]
testdata <- nhanes2[ignore, ]
imp.train <- mice(traindata, m = 2, print = FALSE)
imp.test2 <- mice.mids(imp.train, newdata = testdata)
#> 
#>  iter imp variable
#>   6   1  bmi  hyp  chl
#>   6   2  bmi  hyp  chl
complete(imp.test2, 1)
#>      age  bmi hyp chl
#> 2  40-59 22.7  no 187
#> 4  60-99 27.4  no 206
#> 5  20-39 20.4  no 113
#> 8  20-39 30.1  no 187
#> 11 20-39 26.3  no 118
#> 16 20-39 29.6  no 238
#> 20 60-99 25.5 yes 199
#> 21 20-39 27.4  no 187
#> 24 60-99 24.9  no 206
complete(imp.test2, 2)
#>      age  bmi hyp chl
#> 2  40-59 22.7  no 187
#> 4  60-99 27.5 yes 184
#> 5  20-39 20.4  no 113
#> 8  20-39 30.1  no 187
#> 11 20-39 29.6  no 187
#> 16 20-39 33.2  no 218
#> 20 60-99 25.5 yes 184
#> 21 20-39 28.7  no 238
#> 24 60-99 24.9  no 184

^{Created on 2021-12-06 by the reprex package (v2.0.1)}