|
Abstract:
|
This presentation focuses on computationally feasible ideas on developing imputation models. Our main motivation is to employ these ideas in data structures with unique complexities such as large number of incompletely-observed categorical variables, multilevel data and/or clustered data with multiple membership. Our computational routines modify computationally advantageous Gaussian-based methods that use Markov Chain Monte Carlo techniques to sample from posterior predictive distribution of missing data. Specifically, we propose rounding rules to be used with these existing imputation methods, allowing practitioners to obtain usable imputation with negligible biases. These rules are calibrated in the sense that values re-imputed for observed data have distributions similar to those of the observed data. The methodology is demonstrated using a sample data from the NewYork Cancer Registry database.
|
