How to use correlation threshold argument to avoid covariate dropping

[thomasklausch2]

When the correlation between target variable and a covariate is too high, the covariate is dropped and the event is logged, e.g.

set.seed(2021)
n = 1e3
x = rnorm(n)
y = x + rnorm(n,0,.1)
cor(y,x)
w = rbinom(n,1, 1/(1+exp(-2-x)) )
mean(w)
y[w==0] = NA
cor(y,x,use = 'complete.obs')
library(mice)
m = mice(data.frame(x,y), m = 5, maxit = 10)
head(m$loggedEvents)

In section 9.1.5 of your book (https://stefvanbuuren.name/fimd/sec-toomany.html) it is specified that this happens if the correlation is larger than .999. However in my example the correlation is just below this threshold, still the variable x gets dropped. Furthermore it is stated in the same section that the threshold argument can be used to increase this threshold. However, this arguments seems not to be working in the way I expect, as it does 'nothing'.

m = mice(data.frame(x,y), m = 5, maxit = 10, threshold = .9999)

Could you clarify how this setting is handled?

[stefvanbuuren]

The threshold parameter feeds into find.collinear() defined in https://github.com/amices/mice/blob/master/R/internal.R. This function is run once at initialisation.

The related remove.lindep() function is run before each imputation in the sampler, and uses the maxcor parameter, which is 0.99 by default.

Hope this help.

[stefvanbuuren]

Related: #225, #234, #278, #150

[thomasklausch2]

Thank you. Would it be possible that you show for my example how to change the threshold? I have now tried threshold=1, max.cor = 1 and remove.collinear = F - none of these work in the example.

[stefvanbuuren]

Did you try maxcor?

[thomasklausch2]

OK solved. Thanks!