A Shared Random-effects Transition Model for Longitudinal Binary Data with Informative Missingness

• Activity Number: 390
• Type: Topic Contributed
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Jinhui Li*+ and Yingnian Wu and Xiaowei Yang
• Companies: University of California, Los Angeles and University of California, Los Angeles and BayesSoft, Inc.
• Keywords: repeated count measures ; shared random-effects ; missingness mechanism ; Markov transition models ; log-linear model ; multinomial-logit regression

Abstract

In substance abuse research, repeated measures counting the numbers of use of certain substances are commonly seen in studies adopting longitudinal design. Missing values here are usually nonignorable in practice because nonresponses or dropouts are related to the underlying count processes. Here, we propose a Markov transition model with shared random-effects, which links the propensity of transition between counts and the probability of being either intermittent missing or dropout. The count data are viewed as samples from a Markov chain with transition probabilities: P(i,j) = prob(Y(t)=j|Y(t-1)=i,r), where i and j equal to integers indicating the recent count of some event, Y(t) and Y(t-1) represent two measures collected on time t and t-1 (t=1, ., T), and r indicates the shared random-effects. Assuming a Poisson distribution for the count repeated measures, the transition probabilities are estimated using a loglinear model. The missingness mechanism is modeled with a multinomial-logit regression with three states (observed, intermittent missing, or dropout).