logged events suggest only one value of a variable may have linear dependence, how to solve this?

[ABQTrap]

Dear all, I tried to impute missing values by MICE, there was warning message:
Number of logged events: 25
impute_training$loggedEvents
it im dep meth out
1 1 1 IGF1 pmm subject_ID59
2 1 2 IGF1 pmm subject_ID59
3 1 3 IGF1 pmm subject_ID59
4 1 4 IGF1 pmm subject_ID59
5 1 5 IGF1 pmm subject_ID59
6 2 1 IGF1 pmm subject_ID59
7 2 2 IGF1 pmm subject_ID59
8 2 3 IGF1 pmm subject_ID59
9 2 4 IGF1 pmm subject_ID59
10 2 5 IGF1 pmm subject_ID59
11 3 1 IGF1 pmm subject_ID59
12 3 2 IGF1 pmm subject_ID59
13 3 3 IGF1 pmm subject_ID59
14 3 4 IGF1 pmm subject_ID59
15 3 5 IGF1 pmm subject_ID59
16 4 1 IGF1 pmm subject_ID59
17 4 2 IGF1 pmm subject_ID59
18 4 3 IGF1 pmm subject_ID59
19 4 4 IGF1 pmm subject_ID59
20 4 5 IGF1 pmm subject_ID59
21 5 1 IGF1 pmm subject_ID59
22 5 2 IGF1 pmm subject_ID59
23 5 3 IGF1 pmm subject_ID59
24 5 4 IGF1 pmm subject_ID59
25 5 5 IGF1 pmm subject_ID59

the logged events suggested that the value of subject_ID59 of the variable subject_ID may have linear dependence problem, so did it mean that only this peticular value have collinear problem? If I don't want to exclude the whole variable subject_ID using predictor matrix, what should I do? Many thanks.

[stefvanbuuren]

You didn't say how you run your model, but did I suspect your variable subject_ID has class character or factor. In that case mice expands it into dummy variables. My advice is to keep subject_ID out of the imputation model. See below for a reprex and solution.

library(mice, warn.conflicts = FALSE)

# show problem
df <- mice::nhanes
df$subjid <- rownames(df)
imp1 <- mice(df, print = FALSE, seed = 1)
#> Warning: Number of logged events: 99
head(imp1$loggedEvents)
#>   it im dep     meth
#> 1  0  0     constant
#> 2  1  1 bmi      pmm
#> 3  1  1 bmi      pmm
#> 4  1  1 hyp      pmm
#> 5  1  1 hyp      pmm
#> 6  1  1 chl      pmm
#>                                                                                                        out
#> 1                                                                                                   subjid
#> 2                                                      df set to 1. # observed cases: 16  # predictors: 28
#> 3 age, hyp, subjid10, subjid11, subjid12, subjid13, subjid16, subjid21, subjid3, subjid4, subjid5, subjid6
#> 4                                                      df set to 1. # observed cases: 17  # predictors: 28
#> 5               age, bmi, chl, subjid10, subjid11, subjid12, subjid16, subjid2, subjid21, subjid4, subjid6
#> 6                                                      df set to 1. # observed cases: 15  # predictors: 28

# repair problem
pred <- make.predictorMatrix(df)
pred[, "subjid"] <- 0
imp2 <- mice(df, pred = pred, print = FALSE, seed = 1)
head(imp2$loggedEvents)
#> NULL

^{Created on 2023-12-27 with reprex v2.0.2}

[ABQTrap]

Dear stefvanbuuren, thank you for your reply. You are absolutely right, subject_ID has class of factor. I followed your advice, and logged events became NULL. I highly apprecitate your help.

[ABQTrap]

Dear stefvanbuuren, you mentioned that mice expands the variables with class of factor or character into dummy variables, does this mean that I should keep out other variables with class of factor out during imputing data, although logged events showed no warning message? I'm new to MICE, thanks a lot.

[stefvanbuuren]

No, keep any factors that are informative in the imputation model. But ID-variable is administrative, not related to the data, and has many different categories, so you better remove that one from the predictors (you still keep it in the data).