Tuesday, December 1st, 4:10pm, 1043 Gladys Valley Hall
Speaker: Christiana Drake (Dept. Statistics, UC Davis)
Title: Non-ignorable non-response, random subsets of non-respondents and combined odds ratios
Abstract: Longitudinal studies often experience loss to follow-up. The validity of inference depends on the missingness mechanism. If the missingness mechanism depends on unobserved data, unbiased estimation requires further information. The information from random sub-samples of dropouts whose responses are obtained, can be used to model the data using selection or mixture models, both of which are identifiable in this case. Mixture models result in regression parameters that are conditional on the response status. When the outcome variable is binary and logistic regression is used to model the outcome conditional on response status, the question arises, how to combine the logodds ratios from responders and non-responders. We discuss the conditions required for the conditional logodds ratio to be the same for responders and non-responders. We also propose a weighted average of the conditional logodds ratios when its value differs for responders and non-responders and study its properties.
Christiana Drake, Department of Statistics, UC Davis
Travis Loux, College of Public Health and Social Justice, St. Louis University