Joint Models vs. Cox: Which Works Best in Real‑World Studies?

In studies that track health markers over time and link them to when people experience an event, researchers often use “joint models. ” These methods combine two types of data: repeated measurements and the time until an event happens. They are seen as a fair way to see how a marker, like kidney function, affects the chance of death or illness. But do the software tools we use actually give reliable results? To answer this, a team ran computer experiments that mimicked an older‑adult kidney study. They tested several popular R packages, most of them left on their default settings, and looked at how the number of events (e. g. , deaths) and the amount of repeated data influenced accuracy.The results varied a lot. One package, called JM, often gave biased estimates and sometimes failed to finish the calculations. Another, joineRML, performed better: it produced estimates close to the true values and ran successfully in most situations. However, both of these frequentist packages tended to shrink the impact of baseline characteristics (like age or gender) in their survival models. A Bayesian package, JMbayes2, usually did well. Still, when the data had fewer than about 70 events or when the ratio of observations to events dropped below two, it struggled to converge and its estimates could be off. Standard methods—time‑varying Cox regression and a two‑stage approach that separates the longitudinal part from the event part—showed more bias than JMbayes2 in some cases. Yet they were more stable overall, converging in most scenarios. In short, if you want unbiased estimates and your data are rich enough, joineRML or JMbayes2 can be good choices. If you have limited events or fewer repeated measures, simpler methods might be safer.