Handling Missing Data in Intensive Longitudinal Data with Mixed Missing Mechanisms

Intensive longitudinal studies (ILS) are particularly prone to high levels of missing data compared to cross-sectional or traditional panel designs. Missingness may arise concurrently from the data collection process (MCAR, MAR, or MNAR) and from the handling of unequal measurement intervals (often treated as MAR), resulting in mixed missing mechanisms. Despite their prevalence in applied research, little is known about how missing-data handling methods perform in ILS under such conditions. This study evaluates the performance of the Kalman filter and Bayesian selection models across varying levels of overall missingness and mixtures of ignorable and non-ignorable mechanisms using two simulation studies within a DSEM framework. Results indicate that the Kalman filter performs well under ignorable mixed missingness (MCAR/MAR), tolerating up to 70% missingness under the lower sample-size condition examined (100 individuals and 50 measurement occasions). When non-ignorable missingness is present, estimation performance deteriorates as the proportion of MNAR increases, and correctly specified Bayesian selection models consistently outperform alternative approaches. However, even under optimal specification, the tolerable proportion of MNAR missingness remains low, averaging approximately 3%. These findings provide simulation-based guidance for selecting missing-data handling methods in ILS and highlight the critical role of missingness mechanisms in shaping estimation performance.