scienceneutral
Joint Models vs. Cox: Which Works Best in Real‑World Studies?
Berlin, GermanyFriday, May 15, 2026
Researchers often rely on joint models to link repeated health measurements with time‑to‑event outcomes, such as kidney function and mortality. These models combine two data streams—longitudinal markers and event times—to estimate how a marker influences the risk of death or illness.
Study Design
A team simulated an older‑adult kidney study and benchmarked several popular R packages, most run with default settings. They varied:
- Number of events (e.g., deaths)
- Amount of repeated data
to assess how these factors influence estimation accuracy.
Key Findings
| Package | Performance Highlights |
|---|---|
| JM | Frequently produced biased estimates and sometimes failed to converge. |
| joineRML | Delivered accurate estimates, converging successfully in most scenarios; however, it tends to shrink the influence of baseline covariates (age, gender). |
| JMbayes2 | Generally robust; struggles with convergence when events < 70 or observation‑to‑event ratio < 2, potentially biasing results. |
| Standard Methods (time‑varying Cox & two‑stage) | Showed higher bias than JMbayes2 in some cases but were more stable, converging across most settings. |
Practical Takeaways
- Rich Data: If you have ample events and repeated measures, joineRML or JMbayes2 can provide unbiased estimates.
- Sparse Data: With limited events or fewer observations, simpler methods (time‑varying Cox or two‑stage) may be safer due to their stability.
Actions
flag content
