Multiple imputations by chained equations (MICE) procedure changes non-imputed categorical variables in dataset

[Francis Vergunst]

I’m running a multiple imputations by chained equations (MICE) procedure. I have a sample of 1374 subjects with 5 variables that require imputation, with the maximum number of missing values for a given variable around 250. I’m pooling across 50 datasets.

For some reason the N(%) of my categorical variables changes after the imputation. This occurs whether I include the variable in the MI procedure or not. For example, if I run a crosstab on sex before the MI procedure, I get 48% of the sample as male, while after the procedure it is 52%. The original dataset (_mi_m==0) doesn't change; it's when I check the N(%) across all datasets or run my analysis that the variables have changed. I thought the MI procedure keeps unimputed variables the same across all datasets? What am I missing?
I’ve copied a syntax example below.

Many thanks


*(1)
mi set mlong

*(2)
mi misstable summarize advmerej advperej adpmt01 adpjt01 djbes01
mi misstable patterns advmerej advperej adpmt01 adpjt01 djbes01

*(3)
cor advmerej advperej adpmt01 adpjt01 djbes01

*(4)
mi register imputed advmerej advperej adpmt01 adpjt01 djbes01

*(5)
mi impute chained (regress)advmerej (regress)advperej (regress)adpmt01 (regress)adpjt01 (regress)djbes01 = sex, add(50) rseed (53421) force

mi estimate: logistic sex rmint02b advmerej advperej adpmt01 adpjt01 djbes01

[FernandoRios]

Hi Francis
The MICE procedure its not changing the "non-imputed" data. What is happening is that when you use
mi set mlong
new observations are created, but only for cases with missing values. This is done, i believe, to save space.
I personally like the "mi set flong" This replicates entire datasets, not only the observations with missing data. Thus, if you do your simple cross tab, you will find nothing has changed.
HTH
Fernando

[Francis Vergunst]

Thanks Fernando - that's really helpful, it works perfectly now. Francis