Non-integers derived through multiple imputation: consequences for GEE analysis (PROC GENMOD)?

[rebeccachau]

I have a cross-sectional GEE model clustered by facility (nursing home). The dv is depression score for each participant, and there are a series of predictors (activities of daily living, mini mental state, physical pain, sleep).

I have completed multivariate normal multiple imputation, and this has generated non-integer responses. Consequently, when I run the GEE analysis in SAS I get the following error message:

"WARNING: Some observations with invalid response values have been deleted. The response was less than or equal to zero for the

Gamma or Inverse Gaussian distributions or less than zero for the Negative Binomial or Poisson distributions."

This warning appears 4 times, for 4 of the 20 imputations. Because the imputation process estimates so much additional data that is then pooled, I am inclined to think that this warning can be ignored. Any advice?

Below is my code:

PROC GENMOD DATA=WORK.IMPORT;
CLASS FAC;
MODEL GDS=ADL MMSE NHP PEFIT SD MSPSP / DIST=POISSON LINK=LOG COVB;
BY _Imputation_;
REPEATED SUBJECT=FAC;
ODS OUTPUT
ParameterEstimates=a_mvn;
SCALE=1;
RUN;

PROC MIANALYZE parms=a_mvn;
modeleffects intercept ADL MMSE NHP PEFIT SD MSPSP;
RUN;

[SAS_Rob]

I assume you are using Proc MI to do the imputation.  If so, you might consider using the MIN= and ROUND= option to impute only integers.  While MI does not specifically impute Poisson variables, in most cases it would be robust enough to handle using the MIN= and ROUND= options to make it Poisson.