Data contains NAs after imputation, no evidence for constant values or collinearity

[Michael-Covell]

I have NAs after I run mice() when I "skip" imputation of some variables using this line of code...
meth[c('revenue','units','category','menu','menu_agg')] <- ""

There is no warning message about constant values and the correlation matrix (among numeric variables) looks like this...

dfr <- df[,c('broilers_boosted','revenue','units')]
cor(dfr, use = "pair")

| broilers_boosted | revenue | units
1.0000000 | 0.5961082 | 0.6432247
0.5961082 | 1.0000000 | 0.6547681
0.6432247 | 0.6547681 | 1.0000000

I am interested only in imputing values for broilers_boosted. Here is the full code I am running:
Example.txt

df <- read.csv("df.txt)"
ini <- mice(df, maxit = 0)
pred = ini$predictorMatrix
pred[c("revenue",'units', 'category', 'menu', 'menu_agg'),] <- 0
pred[,c('category', 'menu', 'menu_agg')] <- 0
meth <- ini$meth
meth["broilers_boosted"] <- "norm"
meth[c('revenue','units','category','menu','menu_agg')] <- "" #only part of the data is imputed when I include this line
xyz <- mice(df, predictorMatrix=pred, method=meth)
colSums(is.na(complete(xyz)))

Again, I am interested only in imputing values for broilers_boosted, but if (and only if) I skip imputation of the other variables, mice() only partially imputes data for broilers_boosted such that there are still NAs after imputation. Is this a bug?

[stefvanbuuren]

This is normal behaviour. The line

meth[c('revenue','units','category','menu','menu_agg')] <- ""

instruct mice to skip imputation for some variables, but these later are needed to impute broilers_boosted. Hence if they contain missing values in some rows, then mice cannot calculate imputations for these rows, and NAs will result.

You would need to remove the predictors from the imputation model for broilers_boosted if you do not want this from happening, e.g. by setting

pred["broilers_boosted", c("revenue",'units', 'category', 'menu', 'menu_agg')] <- 0

[thomvolker]

If any of the predictors have missing values, it is not possible to create an imputation, because the fitted values cannot be calculated if any of the predictors have missing values. Using the predictorMatrix argument, you could exclude the predictors with missings, if that is something what you want. Alternatively, you can impute all variables.

The second question is hard to answer, because I don't know what your intended analysis is. Most analyses won't work if you have missing values in the predictors, and I don't see why you would want that to do something like that, but in principle you could do that, yes.

[maria-jal]

Thanks for your reply!

In the main analyses, I would like to test for measurement invariance for a construct A across 3 groups. In the next step, I will test the association of this construct A with some criteria B and C. These criteria could be used in the imputation process as predictors but they have some missings. It feels like cheating to use imputations of these criteria in the main analyses. I am not sure which strategy would be adequate:

1. Don't use the criteria B and C as predictors and do not impute them in the imputation process
2. Use them as predictors, impute them in the imputation process and use all imputations in the main analyses
3. Like 2 but use the original variables B and C instead of the imputations in the main analyses

[gerkovink]

It feels like cheating to use imputations of these criteria in the main analyses.

Can you elaborate on why you feel this is the case? I would argue that not modeling the uncertainty associated with the missingness is cheating. After all, you're more than happy to model the analyses, but you would be hesitant to use the same or similar modeling techniques to infer about the missingness problem? There is a nice discussion of this persistent thought in Aboyami, Gelman and Levy (2011)

So, I would do option 2

[stefvanbuuren]

Pros en cons of options:

1. Not recommended because omission of B and C introduces a bias towards zero in the association analysis.
2. Unbiases w.r.t. association analysis. Uses all cases. Take care to allow for differences between the A group during imputation, e.g. by including an additive or interaction term, or by imputing in three groups separately.
3. As 2, but with reduced power.

[maria-jal]

Thank you all for your replies!