Missing Values with Count Data

#1
So, I have a question that I cannot figure out. I've looked in the literature and can't really find a solution. Maybe I'm looking in the wrong place, but could use some direction.

I have an outcome variable that is a count (i.e. # of deaths). There is missing data that I would like to address. However, the solutions I typically use (maximum likelihood or multiple imputation) produce values that are not an integer. A negative binomial regression requires an integer.

Is there a technique that anyone is aware of that I can use to replace missing values that will result in an integer? As a long shot, would it be acceptable to round the non-integers to integers?

Thanks!

Dave
 

Dason

Ambassador to the humans
#2
If you're getting non-integer values when doing maximum likelihood imputation my guess is that you're assuming normally distributed data. Change the likelihood to the proper distribution and ML should give you reasonable values.
 
#3
Thanks! Unfortunately I am not familiar with missing value analysis for non-normal data. I use SPSS V. 22 (the only one my organization buys!). The options for EM are "normal", "mixed normal", and "student's t". Mixed normal has a field for mixture proportion and standard deviation ratio. Stduents t has a field for degrees of freedom.

I appreciate you taking the time to respond and don't expect you to walk me through how to make the proper adjustment. However, are you able to point me to a source that can assist me with making these adjustments? Thanks!
 
#5
Round the value to the closest integer!

If you have lots of values and few missing, then that does not matter. If you have few values and many missing then you have probably other strange selection difficulties.

But I would not be surprised if the EM algorithm calculated an expected value as a non-integer for a discrete distribution.