A Chain Logistic Model for Analysing Categorical Survey Data Involving Both Unit and Item Non-responses
by Guoqi Qian
Abstract: In this talk I will present a chain logistic regression model for analysing
categorical survey data involving both unit non-responses and non-ignorable
item non-responses. This model results from a statistical consulting problem
brought by Rebecca Bentley of The Australian Research Centre in Sex, Health and Society. The model follows the maximum likelihood approach of Ibrahim, Lipsitz and Chen (1999, Journal of the Royal Statistical Society, Series B, vol. 61, pp 173-190). However it deals with more realistic and complicated data which are often seen in multi-stage sampling surveys. A comprehensive procedure for estimation, testing and computing of this model has been developed. The computing of the model is based on the EM or MCEM algorithm. Employment of the maximum likelihood approach ensures that the standard MLE-based inference methods can be applied to this model and the results can be easily explained. The model has been applied to analyse a real survey data and has revealed some new results that can not be obtained by using a model ignoring non-responses. Problems of variable selection and model selection arisen from this model will also be discussed in the talk.
For More Information: Department of Mathematics and Statistics tel. 03 8344 5550