PLACE:  319 Snedecor

SPEAKER: Trivellore Raghunathan, Department of Biostatistics and Instititute for Social Research, University of Michigan, Ann Arbor, MI

TITLE: Semiparmetric Approach for Multiple Imputation of Unobserved Values in Longitudinal Studies

ABSTRACT:
Unbalanced data, where not all the individuals are observed at the same time points, is a common feature in many longitudinal studies. There are several possible reasons for observing an unbalanced data. The primary example is due to nonparticipation even though the original design may have called for measuring all the individuals at the same time. Sometimes it may not be feasible to obtain data from all the individuals at the same time due to logistical considerations, and the measurement may be staggered over a period of years. Alternatively, by design the data may not be collected from all the individuals at every wave of data collection. Though the unbalanced data can be handled in some analysis but for several other types of analysis having a complete data with all the individuals measured at comparable time points is critical. A case study that motivated this research involved relating the wealth of the parents during the critical developmental age on their children's development. Similar problems occur when the longitudinal data are used to characterize the exposure information over a life course or during a certain critical period. Multiple imputation approach provides a framework for handling unbalanced data in such instances. This paper discusses a semiparametric approach using spline models, within a Bayesian framework, to create imputations. Multiple imputations are defined as draws from the posterior predictive distribution. Gibbs sampling us used to obtain the draws. The case study described above is analyzed using this approach. A simulation study evaluates the repeated sampling properties of the inferences obtained using this approach.
 

COFFEE: 3:45 p.m., 104 Snedecor Hall