Joint Model for Longitudinal and Time-to-Event Data
Also known as: joint model, shared random effects model, longitudinal-survival joint model, Joint Model (Boylamsal + Sağkalım Birleşik Model)
The joint model for longitudinal and time-to-event data, formalised by Tsiatis and Davidian in 2004 and extended comprehensively by Rizopoulos in 2012, simultaneously estimates a mixed-effects model for repeatedly measured biomarkers and a survival model for the time to an event, linking the two processes through shared random effects. It resolves two major problems that simpler approaches cannot handle: informative dropout from longitudinal studies and the endogeneity of time-varying biomarkers used as covariates in a Cox model.
Key highlights
- Jointly handles informative dropout and endogenous time-varying covariates, which are the two main sources of bias in separate analyses.
- Uses the full observed biomarker trajectory to predict the event hazard at each moment, giving more efficient estimates than last-observation-carried-forward or simple time-dependent Cox approaches.
- Produces dynamic survival predictions: given a subject's biomarker history up to time s, the model yields personalised conditional survival probabilities for future times, which is clinically actionable.
- Propagates uncertainty from the longitudinal model into the survival model rather than treating estimated values as fixed.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use the joint model when you have longitudinal data with repeated biomarker measurements and a time-to-event outcome and at least one of the following problems is present: (1) dropout from the longitudinal component is informative — patients leave the study because their condition is worsening; (2) you want to use a time-varying biomarker as a covariate in a survival model but the standard time-dependent Cox approach is biased because the biomarker is measured with error or is endogenous. The method requires a longitudinal submodel that fits reasonably well (typically linear mixed effects), a survival submodel (Cox or parametric), a minimum of approximately 100 subjects with at least three repeated measurements per subject, and the assumption that the shared random-effects structure correctly captures the dependence between the two processes.
Strengths & limitations
- Jointly handles informative dropout and endogenous time-varying covariates, which are the two main sources of bias in separate analyses.
- Uses the full observed biomarker trajectory to predict the event hazard at each moment, giving more efficient estimates than last-observation-carried-forward or simple time-dependent Cox approaches.
- Produces dynamic survival predictions: given a subject's biomarker history up to time s, the model yields personalised conditional survival probabilities for future times, which is clinically actionable.
- Propagates uncertainty from the longitudinal model into the survival model rather than treating estimated values as fixed.
- Computationally intensive; EM or MCMC estimation can take minutes to hours for moderate datasets and may fail to converge with small samples or complex random-effects structures.
- Requires at least approximately 100 subjects and at least three longitudinal measurements per subject; it is not suitable for short follow-up or sparse measurement schedules.
- The association structure (current value, current slope, cumulative effect) must be chosen carefully — misspecification can bias both submodels.
- Interpretation of the alpha parameter requires care: it links the true latent trajectory to the hazard, not the observed measurement.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does the joint model differ from a time-dependent Cox model?
In a time-dependent Cox model you plug the last observed biomarker value directly into the hazard function as if it were the true value at that moment. This approach is biased when the biomarker is measured with error (attenuation bias) or when it is endogenous — that is, when the same process that drives the biomarker also drives the event. The joint model avoids both problems by treating the true biomarker trajectory as a latent process inferred from all longitudinal measurements, and by linking that latent process to the hazard through shared random effects estimated simultaneously with both submodels.
What is informative dropout and why does it matter?
Informative dropout occurs when the reason a subject leaves the longitudinal study is related to the outcome being measured. For example, a patient with worsening kidney function may miss clinic visits or die before the next scheduled measurement. If you analyse only the observed measurements, the late-stage trajectories are missing precisely for the sickest patients, so the average trajectory looks artificially better than the truth. The joint model explicitly accounts for this by linking the dropout process (the survival submodel) to the true biomarker trajectory through shared random effects.
What association structures can be used to link the two submodels?
The most common association is the current true value of the biomarker at time t, which enters the hazard directly. Alternatives include the current slope of the trajectory (the rate of change at time t), the area under the trajectory up to time t (cumulative exposure), or combinations of these. The choice should be guided by subject-matter reasoning and compared using information criteria such as DIC or WAIC in the Bayesian setting.
How many repeated measurements do I need per subject?
A generally accepted rule of thumb is at least three longitudinal measurements per subject, with at least 100 subjects in total. Fewer measurements leave the subject-specific random effects poorly identified, which destabilises both the longitudinal submodel and the association parameter. Studies with only one or two measurements per subject should use a simpler approach such as a time-dependent Cox model, accepting its limitations.
Sources
- 1.Rizopoulos, D. (2012). Joint Models for Longitudinal and Time-to-Event Data. CRC Press.
- 2.Tsiatis, A.A. & Davidian, M. (2004). Joint Modeling of Longitudinal and Time-to-Event Data: An Overview. Statistica Sinica, 14(3), 809–834.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Joint Model for Longitudinal and Survival Data. ScholarGate. https://scholargate.app/survival/joint-model-survival