Abstract:
|
When conducting research synthesis, the studies to be combined often do not measure the same set of variables, which create missing data. When the studies are longitudinal, missing data can occur on either the observation-level (time-varying) or the subject-level (non-time-varying). The focus of missing data methods is usually on missing observation-level variables. We focus on missing subject-level variables. We develop a new joint model for multiple imputation of missing subject-level variables that models subject- and observation-level variables with distributions in the exponential family. Our model is built within the generalized linear models framework and uses normally distributed latent variables to account for dependence on both the subject- and observation-levels. When compared via simulation, the performance of our model is similar to or better than existing approaches for imputing missing subject-level variables with normal, Bernoulli, Poisson, and multinomial distributions. We illustrate our method by applying it to combine two longitudinal studies on the psychological and social effects of pediatric traumatic brain injury.
|