605 – Imputation of Missing Data
Bayesian Multiple Imputation of Count Data with Zero Inflation
Chin-Fang Weng
U.S. Census Bureau
In government survey applications, zero-inflated count data often arise, sometimes with item nonresponse. We consider the problem of imputing missing counts. We assume that the observations/items are missing at random. We also assume the zero-inflated data is a mixture distribution: one component from a distribution degenerate at zero and one from a Poisson distribution. Both components may depend on covariates, which are always observed. We formulate a model for bivariate zero inflated count data and propose a Bayesian imputation scheme for imputing missing items by assigning priors to unknown regression parameters. Using the predictive distribution of missing items given observed items, one can impute each missing item as a random draw from the predictive distribution. We use Markov Chain Monte Carlo to generate imputed values of missing items. Multiple imputations are computed by running the chain long enough to produce multiple realizations of the missing item values. To obtain (nearly) independent draws, the chain must be thinned. We will illustrate its potential with a simulation study and with an analysis of part-time Employment data from the Annual Survey of Public Employment & Payroll (ASPEP).