Missing data assumptions in latent class growth models of mental health: a secondary analysis case study of two UK cohorts

Abstract

To mitigate bias caused by selective attrition in cohort studies, researchers have been advised to link cohort data to administrative data sources, as these are not prone to attrition and provide more objectively measured information. However, consent for data linkage is often influenced by mental health status and individuals with mental health issues are frequently underrepresented in administrative data. Alternative approaches should therefore be considered for mental health research. This study examined how different missing data assumptions influence estimates from latent class growth models of mental health trajectories and the implications of those estimates for longitudinal mental health research. Methodological study comparing latent class growth analyses under assumptions of data missing completely at random (complete case analysis), missing at random (standard models and multiple imputation) and missing not at random (delta-adjusted pattern mixture models). Secondary data from young people born in the early 1990s participating in two British cohort studies, the Avon Longitudinal Study of Parents and Children and Longitudinal Study of Young People in England. Latent class growth analysis identified two to three groups of young people who exhibited similar trajectories of mental wellbeing changes from age 14 to 30. Estimated trajectory shapes varied according to the assumed missing data mechanism and frequency of mental health measurements. Findings highlight the importance of sensitivity analyses under alternative missing data assumptions, particularly where missingness may not be at random in longitudinal mental health research.