Group-based trajectory modeling under non-random attrition: A sensitivity analysis and application to frailty trajectories

Group-based trajectory modeling (GBTM) is used to identify trajectories but suffers from attrition bias when dropout is nonrandom. We evaluate an extended GBTM that models attrition as a latent-class process. Monte Carlo simulations compare the extended model with conventional GBTM across trajectory separation and missing-data mechanisms, and test a parsimonious version with a constant dropout rate within classes. conventional GBTM is adequate for separated trajectories but biased when they overlap and attrition is nonrandom. The extended GBTM remains unbiased, as shown in Manitoba frailty data. Implications: modeling attrition improves robustness; the parsimonious extension remains reliable under complex dropout. • Compared extended GBTM with conventional GBTM under simulated non-random attrition. • Conventional GBTM bias emerged when trajectory classes overlapped and attrition was non-random; robust only under well-separated trajectories. • Parsimonious extended GBTM (constant dropout rate within classes) consistently produced unbiased estimates of class proportions and trajectory shapes across all scenarios. • Empirical application to frailty data reclassified 11% of the sample into a high-risk worsening group, illustrating the practical impact of accounting for attrition.