Skip to main content

# Table 2 Overview of the 4 modern longitudinal analytic methods

Model Questions Advantages Disadvantages
GEE What is the averaged outcome trajectory for the population? (Trajectory of averages) Parameter estimates robust to misspecification of the covariance structure.
Both time-invariant and time-varying predictors can be studied.
No individual level inference
Assumes missing data to be missing completely at random (MCAR), which may not be true for many longitudinal studies.
MRM What is the outcome trajectory of the individual?
What is the average outcome trajectory for the population? (Average of trajectories)
Individual level inference possible with the incorporation of random effects.
Both time-invariant and time-varying predictors can be studied.
Assumes missing data to be missing at random (MAR), which is more likely in longitudinal studies.
Misspecification of covariance structure may bias parameter estimates45
LCTAa Are there distinct subgroups within the study population?
What are the trajectories of the identifiable subgroups within the population?
Objectively identifies latent distinct subgroups within a heterogenous population.
Able to use time-invariant factors to predict group membership. Able to study effects of time-varying covariates in different ways (depending on question and underlying theoretical framework)
Assumes data to be missing at random (MAR).
Complex and time-consuming computing procedures. Interpretation of time-varying covariates can be challenging depending on the formulation.
Joint Modelb What are the trajectories of (multiple) outcomes of interest?
What is the correlation between the outcome trajectories of interest (i.e., are the trajectories concordant or discordant)?
Multiple outcome trajectories of disparate nature (e.g., continuous with binary, binary-poisson, continuous-survival) can be studied simultaneously.
Objective determination of the longitudinal correlation of the trajectories.
Joint model with time-to-dropout may be used as a means to adjust for data missing not at random (MNAR).
Modeling procedures can be complex with increasing number and kinds of outcomes modeled jointly.
1. aUsually modeled with MRM as the base model
2. bMRM may be used as the base model for continuous, binary and count data. Proportional hazard is used for time-to-event outcomes