440 – Missing Data, Imputation, and Calibration
A Predictive Approach to Missing Data from the Exponential Family
Valbona Bejleri
USDA National Agricultural Statistics Service
Darcy Miller
USDA National Agricultural Statistics Service
Kay Turner
USDA National Agricultural Statistics Service
Discrete outcomes are often observed among survey responses, i.e., counts distributed as Poisson, Bernoulli, multinomial, etc. When some responses are missing, the observed data provide a foundation for predicting the unobserved values or estimating some statistic for the full sample. In this paper, data are assumed to be a random sample from a discrete distribution in the exponential family with missing at random (MAR) responses, i.e., the probability of a response missing is unrelated to the value of that response, but could be related to other variables. A distribution-based technique to compute lower and upper bounds for a missing response is developed. The algorithm makes no use of the parameter estimate. Simulations are used to illustrate the technique and to assess its efficiency. Bayesian prediction bounds constructed based on gamma, Jeffreys, and uniform priors are also discussed for comparison purposes. We tested the algorithm using actual data from USDA's National Agricultural Statistics Service's Quarterly Hog Survey.