Joint Models for Longitudinal and Time-to-Event Data: Linking Repeated Biomarkers to Clinical Outcomes

Clinical biomarkers are often observed repeatedly before an event occurs. A kidney function marker may decline before hospitalization, a tumor marker may change during treatment, or a symptom score may track disease progression until death or relapse. An analysis that models the biomarker trajectory and the event time separately may ignore the fact that the two processes are connected. It may also mishandle informative dropout: patients who experience the event or become too ill to return for measurement are not missing at random from the longitudinal record.

A joint model addresses this structure by specifying two linked statistical processes. The longitudinal submodel describes how the biomarker evolves over time, while the survival submodel describes the hazard of the clinical event. A shared random effect, the current latent biomarker value, the slope, or another association structure connects the components. The result is not simply a larger regression model; it is a model for two related observation processes with distinct likelihoods, timing, and interpretation.

Joint models are useful for dynamic prognosis, disease monitoring, treatment response, and informative dropout. They require careful time alignment, clinically defensible functional forms, and validation of both the longitudinal trajectory and event-risk predictions. They should not be used merely because a dataset contains repeated measurements. The model must answer a clearly stated question about how longitudinal information relates to future outcomes.

The two components of a joint model

The longitudinal component models repeated biomarker measurements for subject i at measurement time t. A common formulation uses a mixed-effects model with fixed effects for the population-average trajectory and random effects for subject-specific deviations. The fixed-effects part may include time, baseline covariates, treatment, time-by-treatment interactions, and nonlinear time terms. Random intercepts and slopes allow subjects to begin at different levels and change at different rates.

The measurement model should reflect the scale and distribution of the biomarker. A continuous approximately Gaussian marker may use a linear mixed model, while a binary, ordinal, or count marker may require a generalized mixed model. Transformation, heteroscedasticity, serial correlation, and measurement error can materially change the meaning of the latent trajectory. A model that treats a skewed biomarker as normally distributed without assessment may produce poorly calibrated predictions even if the survival component appears plausible.

The survival component models time to the event, often with a relative-risk structure. The baseline hazard may be left unspecified through a Cox-type formulation or represented with flexible splines or piecewise functions. Baseline covariates and treatment can enter the hazard directly. The longitudinal process then enters through an association term, which must be defined rather than assumed.

Association structures: what does the biomarker contribute?

The simplest association structure links the hazard to the current underlying biomarker level. It asks whether a higher latent marker value at time t is associated with a different instantaneous event hazard at t. This can be clinically intuitive when the current disease burden is the main prognostic signal.

A slope association links the hazard to the current rate of biomarker change. It asks whether rapid worsening or improvement predicts the event, even after accounting for the current level. This may be important when trajectory acceleration contains information beyond the latest measurement. A level-and-slope structure includes both terms, but it requires enough repeated data to estimate them reliably.

Other structures can link cumulative exposure, a weighted history, or a latent class to the event process. The choice should be driven by the scientific question and measurement schedule. A highly flexible association may fit the sample but be difficult to validate or explain. Researchers should report the association parameter in terms of the latent biomarker scale and state whether it represents a level, slope, cumulative history, or another feature.

Why separate analyses can be misleading

Fitting a mixed-effects model for the biomarker and a separate Cox model for survival can be adequate for some descriptive questions, but it does not automatically account for the dependence between the processes. If event occurrence causes measurement to stop, the observed longitudinal data are selectively observed. Treating dropout as unrelated to the event can distort the estimated trajectory, especially near the event.

A time-dependent Cox model that uses observed biomarker values can also create problems. Measurements may be recorded irregularly, may contain measurement error, and may only be available for patients who remain under observation. Using an observed value as if it were the true current state can attenuate or distort associations. A joint model uses a latent trajectory to separate the underlying signal from measurement noise, although this protection depends on model specification.

Separate modeling is not always wrong. If the biomarker is measured according to a fixed schedule independent of outcome risk, missingness is plausibly ignorable under the chosen model, and the analysis is explicitly descriptive, separate models may be reasonable. The issue is not whether joint modeling is universally superior; it is whether the data-generating and observation processes support the chosen analysis.

Time alignment and event-history discipline

The biomarker time scale must be aligned with the event time scale. Measurements should be assigned their actual observation times, and the survival model should use only information available before the prediction or risk assessment. A value recorded after a hospitalization cannot be used to claim that the marker predicted the hospitalization unless the research question concerns a later event.

Delayed entry, left truncation, and landmark eligibility require explicit handling. In a transplant cohort, for example, patients may enter the risk set only after surviving to transplantation. In a monitoring study, patients may have different numbers of measurements before the event. The time origin, entry rule, measurement window, and censoring time should appear in the protocol and analysis code.

Informative observation processes may require additional modeling. A patient who visits frequently may have more opportunities for biomarker measurement and may differ systematically from a patient who is too ill or too well to attend. The joint model can link the observed trajectory to the event process, but it does not automatically solve informative visit scheduling. If visit intensity is itself informative, a shared-parameter, joint longitudinal-visit-event model or sensitivity analysis may be needed.

Dynamic prediction from joint models

One important use of joint models is dynamic prediction. At a prediction time s, the model uses the patient’s biomarker history and survival status up to s to estimate the probability of an event by a future horizon s plus h. As new measurements arrive, the prediction can be updated. This differs from a baseline model that freezes all predictors at time zero.

Dynamic predictions should state the target population, prediction time, horizon, event definition, and information set. If death is a competing event for hospitalization, the predicted quantity should distinguish cause-specific survival from cumulative incidence. If a patient has no recent measurement, the model should specify whether prediction uses the latent trajectory, an imputed value, a measurement window, or a reduced information set.

Prediction performance must be evaluated dynamically. Discrimination can be assessed at clinically relevant horizons, but calibration is central: predicted probabilities should be compared with observed risks among patients eligible at the prediction time. Brier scores, prediction error curves, calibration-in-the-large, calibration slope, and flexible calibration plots can complement time-dependent AUC. Performance should be reported across meaningful time points and subgroups rather than only at the time point with the strongest result.

Model checking and sensitivity analysis

Model checking should cover both submodels and their connection. For the longitudinal component, inspect residuals, subject-specific trajectories, time functional form, serial correlation, and distributional assumptions. For the survival component, assess proportionality where relevant, baseline hazard flexibility, influential observations, and event-count support. For the association structure, compare clinically plausible alternatives rather than selecting solely by a single information criterion.

Useful sensitivity analyses may vary the time functional form, random-effects structure, association parameter, measurement-error assumptions, censoring treatment, and missing-data strategy. A slope term that is unstable when measurement times are sparse should not be presented as a robust clinical mechanism. Similarly, a joint model with a highly flexible latent trajectory may overfit if the number of events and repeated observations is limited.

External validation is especially important for dynamic prediction. The longitudinal measurement process, follow-up intensity, event incidence, and treatment pathways may differ between datasets. A model can retain discrimination while losing calibration if the biomarker distribution or baseline hazard shifts. Validation should reproduce the intended measurement timing and prediction horizon rather than using only baseline variables in a different model.

Evidence summary table

Actionable Steps: Build and evaluate a joint model

Common failure modes

A frequent error is treating the association parameter as a treatment effect. A strong relationship between a rising biomarker and event risk may be prognostic, but it does not show that intervening on the biomarker would reduce events. Causal interpretation requires additional assumptions and a design that supports the causal question.

Another error is using a two-stage workflow without accounting for uncertainty from the first stage. A mixed-effects trajectory is estimated first and then inserted into a survival model as if it were observed without error. This can understate uncertainty and distort standard errors. Two-stage methods can be useful with appropriate corrections or bootstrap procedures, but the limitation must be reported.

A third error is forcing an event-time association onto an irregular observation process. If measurement visits are triggered by symptoms, the observed marker history may reflect both disease progression and clinical attention. The visit process should be examined, and a sensitivity analysis may be more honest than a single overconfident shared-parameter model.

A fourth error is reporting only the longitudinal trajectory or only the survival hazard. A joint model must be interpretable as a linked system. Report the marker effects, event effects, association structure, uncertainty, and the dynamic prediction output if prediction is the intended use.

A fifth error is presenting a dynamic prediction as a universal score. Predictions are conditional on the risk set, data availability, model assumptions, and horizon. A model that is useful after a recent biomarker measurement may not apply to a patient with a different measurement schedule or a different treatment pathway.

Reporting and workflow considerations

A transparent manuscript should include a diagram or clear description of the longitudinal and event processes, the time origin, measurement schedule, censoring, event definition, and association structure. Report the number of subjects, measurements, events, measurement-related dropouts, and patients contributing to each prediction horizon. Explain how missing measurements and irregular visits were handled.

For dynamic prediction, report the prediction time and horizon rather than only a generic “future risk.” Include calibration and discrimination at clinically relevant times. If a competing event is present, state whether the output is a cause-specific quantity, cumulative incidence, or a multi-state probability. If the study is observational, distinguish prognostic association from causal effect and describe confounding control separately.

Conclusion

Joint models provide a principled framework for studying repeated biomarkers and time-to-event outcomes as connected processes. Their main contribution is not simply combining two datasets. It is representing how a latent longitudinal trajectory, an event process, measurement timing, and informative dropout may interact. With explicit association structures, disciplined time alignment, appropriate sensitivity analyses, and dynamic validation, joint models can improve prognostic understanding and patient-specific risk prediction. Their results remain conditional on the data-generating process and should not be presented as causal evidence without a design that supports causal interpretation.