We have a cluster randomised trial with a small number of clusters, The primary endpoint is measured at follow-up and we have missing data. We proposed to conduct a linear mixed model including fitting 'treatment' ,a random intercept for cluster and adjusting for various baseline measures and covariates. We are also using Kenward-Roger correction for degrees of freedom (due to small sample ). After complete-case analysis, we noticed our intraclass correlation coefficient (ICC) is close to 0.
To deal with missing data, we ran 40 imputations to get 40 compelete datasets accounting for clustering. We ran linear mixed models for each imputation then combined the results ( treatment paremeter estimate, differences, lsmeans, etc) using Rubin's rules through the SAS procedure 'proc mianalyze'.
However, one of the options for 'proc mianalyze' is to specify complete data degress of freedom (EDF=). And we noticed that degrees of freedom for the treatment effect changes for each imputation due to the use of ddfm=KR in our mixed model. Especially we have cases where ICC equals 0 for some imputations and ICC not equal to 0 for other imputations. For those who had ICC equal to 0, degrees of freedom is about 200 but those who had ICC not equal 0, degrees of freedom is about 6. Therefore, we are not sure what EDF to put for 'proc mianalyze' statement. It would be great if anyone could give us some suggestions. Many thanks!