One common problem in longitudinal studies is the presence of missing data. Intensive longitudinal data involving repeated assessments of constructs such as emotions are particularly prone to nonignorable missingness, namely, missingness where the missing data mechanism depends on unobserved information. For example, in measuring day-to-day changes in negative emotions, it is possible that the participants may opt not to report their feelings on the days with heightened negative emotions, thereby leading to nonignorable missingness when modeling emotion processes. Even though myriad approaches for handling missing data have been proposed and tested in the literature, few studies have investigated the tenability and utility of these approaches when used with intensive longitudinal data.
In this project, we consider and illustrate two different approaches for coping with missingness in fitting multivariate time series models, including multiple imputation (MI) approaches and a Bayesian approach. For the MI approaches, we compared a full MI approach, in which all missing variables – including dependent/endogenous variables as well as covariates – are imputed simultaneously, and a partial MI approach, in which missing covariates are imputed with multiple imputation, while missingness in the dependent variables is handled via full information maximum likelihood estimation. The Bayesian approach simultaneously models missingness in the dependent variables and covariates in the context of the VAR model. The approaches are examined in Monte Carlo simulation studies under the assumptions of missing completely at random (MCAR), missing at random (MAR), and not missing at random (NMAR). The methods are also demonstrated with longitudinal dyadic data from a previously published study on couples’ self-reported conflict resolution (Schermerhorn, Chow & Cummings, 2010).