Page 1 of 1

Anomaly when imputing a Normal response

Posted: Tue Mar 04, 2014 9:17 pm
by MannyGomez
I'm faced with an unusual situation. I have used Realcom to impute one (hierarchical) Normally distributed variable (mean 200, 95%CI: 25 to 650, and no negative values). After imputation, I checked the imputed values and found that a substantial number of them were negative and very far from zero, such as -3000 or -2000 (!). Since the variable was approx Normal and sample size>5000, I found this very strange. I computed multilevel MICE in R (using the same imputation model), and the imputed values were well within the range of the observed values.

Potential candidates that may have caused the problem in Realcom:

1- I have only 5 clusters (level-2 identifier). Although this did not cause any problem for multilevel MICE.
2- The variables included in the model do not explain much of the missingness. In other words, is Realcom more sensitive to misspecification than multivel MICE?
3- Any computational issue? I have considered burnin=1000, no. of imputations=50, MCMC iterations=10,000.

Which of these may have caused the problem? Any other suspicion?

Many thanks, Manny

Re: Anomaly when imputing a Normal response

Posted: Wed Mar 05, 2014 11:22 am
by ChrisCharlton
That does look odd. Would it be possible to send either myself or Harvey Goldstein (email address here: an example dataset where this occurs so that we can investigate further?

Re: Anomaly when imputing a Normal response

Posted: Wed Mar 05, 2014 5:55 pm
by MannyGomez
Okay, will do.
When possible, could I have your views on the other issue I raised in the MLwiN user forum on the mixture distribution?