A clinical researcher uses each patient's most recently observed biomarker value (carried forward) as a time-varying covariate in a standard Cox model. Which bias does a joint model specifically correct?
AThe Cox model cannot handle time-varying covariates and requires only baseline measurements
BPatients who deteriorate fastest have fewer later measurements, so their true current biomarker status is systematically underrepresented, biasing the association between biomarker and event
CThe Cox model underestimates sample size by excluding patients who died before any measurements were taken
DCompeting events cannot be incorporated into a Cox model with time-varying covariates
This is informative dropout: the reason measurements stop is correlated with the outcome. Patients who are deteriorating most rapidly are more likely to miss visits and die, meaning their last observed value reflects an earlier, less severe state than their true current status. This biases the biomarker-event association downward. Joint models resolve this by modeling the true underlying trajectory as a latent process, borrowing information across all observed time points simultaneously rather than relying on the last observed value.
Question 2 Multiple Choice
In a joint model for competing events (e.g., cancer-specific death vs. cardiovascular death), how can the same longitudinal biomarker relate to the two competing outcomes?
AIt must have the same association with both outcomes — otherwise the model is misspecified
BIt enters the model only for the primary outcome; competing events are treated as independent censoring
CThe longitudinal trajectory can have different associations with each competing event, estimated simultaneously while respecting the competing risks structure
DThe biomarker trajectory predicts whichever event occurs first, with no outcome-specific distinction
In the competing events extension, the survival submodel becomes a competing risks model with cause-specific (or subdistribution) hazards for each event type. The longitudinal trajectory can have a strong association with one cause (e.g., rising PSA predicts cancer-specific death) but a weak or null association with another (e.g., cardiovascular death). These associations are estimated separately within the joint model, which is precisely why cause-specific joint models provide more nuanced clinical information than simple joint models.
Question 3 True / False
In a joint model, the longitudinal and survival submodels are estimated separately and their results are combined in a second post-hoc stage.
TTrue
FFalse
Answer: False
This describes a two-stage approach, not a joint model. In a true joint model, the two submodels are linked through shared random effects and estimated simultaneously. This joint estimation is the whole point: the same individual-level latent parameters that describe the biomarker trajectory enter the hazard model directly, so measurement error in the biomarker is properly accounted for and the two submodels borrow strength from each other. Two-stage approaches that estimate the longitudinal submodel first and plug in its predictions produce biased results.
Question 4 True / False
Dynamic event prediction in a joint model improves as longitudinal measurements accumulate because the estimated true biomarker trajectory becomes more precise over time.
TTrue
FFalse
Answer: True
Early in follow-up, with only a few observations, the estimated trajectory (intercept and slope of the individual random effect) is uncertain, leading to wide prediction intervals. As more measurements are observed, the individual trajectory shape becomes apparent, shrinking uncertainty in the survival submodel's hazard estimates. This is the clinically important feature of joint models: unlike Cox models where baseline covariates are fixed, joint models allow risk predictions to be dynamically updated at each patient visit as new biomarker data arrive.
Question 5 Short Answer
Why is informative dropout a fundamental problem for standard survival models with longitudinal biomarkers, and what feature of the joint model addresses it?
Think about your answer, then reveal below.
Model answer: In informative dropout, measurements stop not randomly but because the patient is deteriorating — the reason for missingness is correlated with the outcome. A standard model using observed biomarker values treats missingness as random, systematically underestimating the biomarker-event association for patients with the worst trajectories. Joint models address this by specifying a longitudinal submodel for the true underlying biomarker trajectory as a continuous latent process, linked to the survival submodel through shared random effects. The hazard model uses the estimated true trajectory, not the noisy observed values, so it properly accounts for the fact that patients with the steepest trajectories are the most likely to die earliest.
The key move is replacing the observed (noisy, incomplete) biomarker measurements with the estimated latent trajectory. This simultaneously handles measurement error (the observed values are noisy proxies for the true biological process) and informative dropout (the trajectory is estimated even when later observations are missing because the patient died or deteriorated). The shared random effects structure is the mechanism: the same individual parameters that predict the biomarker trajectory also directly enter the hazard function.