A colleague, Carl Moons, and I are doing some simulations to study properties of multiple imputation using the MICE S-Plus library and using the S-Plus transcan function. We are are getting some unexpected findings for which we would appreciate some of your insight.
First, our mechanism for setting some of the observations for one of the variables to missing in the simulations is to, with a constant probability, set a value to missing if the values of three other binary variables on that observation are simultaneously zero. The imputation model is an additive model that does include these three other variables. The ultimate Y model is also additive. What do you expect to find if the mechanism for missingness is non-additive but the imputation and outcome models are additive? Second, with the above missingness mechanism we find that multiple imputation using the response variable results in significantly less bias in that model's regression coefficient estimates as expected, but this improvement in bias is more than offset by an increase in variance. Multiple imputations ignoring Y resulted in lower MSE of regression coefficient estimates. Have other people found the same result? Thanks in advance for any thoughts you have. -Frank -- Frank E Harrell Jr Prof. of Biostatistics & Statistics Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences U. Virginia School of Medicine http://hesweb1.med.virginia.edu/biostat
