Interpreting Joint-Model Associations Without Causal Claims

Why Interpretation Requires Care

A joint model can connect a repeatedly measured biomarker to the cause-specific hazard of an event. That connection can improve prognosis: among people with the same modeled covariates and latent health characteristics, a higher biomarker value may signal a higher instantaneous rate of a particular event. But a predictive association is not automatically the effect of changing the biomarker. The biomarker may be a consequence or indicator of underlying health, and that underlying health may also affect the event process.

This distinction matters especially in competing-risks settings. A cause-specific hazard describes the instantaneous event rate for one cause among people who have not yet experienced any competing event. It is conditional on remaining event-free up to that time. A coefficient in a joint model therefore has a specific conditional meaning; it is not, by itself, a comparison of event probabilities under interventions.

What a Cause-Specific Association Coefficient Says

For illustration, let \(m_i(t)\) denote the modeled underlying longitudinal marker value for individual \(i\), and let \(b_i\) denote that individual’s shared random effects. For event cause \(k\), consider the current-value association model

Here \(\lambda_{0k}(t)\) is the cause-specific baseline hazard, \(w_i\) is a vector of event-process covariates, and \(\alpha_k\) is the cause-specific current-value association parameter. The model states that, at a given time, the cause-\(k\) hazard is compared among individuals with the same modeled covariates and random effects. It also presumes the specified hazard form is adequate, the relevant event causes and times are correctly recorded, and the longitudinal and event processes are represented by the chosen joint model.

For a difference \(\Delta\) in the modeled marker value, the conditional cause-specific hazard ratio is \(\exp(\alpha_k\Delta)\). The unit of \(\alpha_k\) is the reciprocal of the marker’s unit; the hazard ratio is dimensionless. This is an association on the hazard scale, conditional on the model. It does not say that intervening to raise the marker by \(\Delta\) would multiply a person’s hazard by that ratio.

How Shared Latent Health Can Create Association

A shared random effect can represent persistent individual differences that influence both the longitudinal trajectory and event risk. For instance, an unobserved health state may lower a biomarker and raise the hazard. In that case, the marker and event process are statistically linked because they share a cause. Even if the marker itself has no effect on the event, the observed marker can still be prognostic.

A joint model represents dependence through its specified shared effects and association structure. That is valuable for describing the data and accounting for dependence between repeated measurements and event times. But including a random effect does not guarantee that every common cause has been captured, that its distribution is correct, or that the marker is unconfounded. Moreover, conditioning on a latent random effect in a model is not equivalent to a randomized intervention on the marker.

Why Marginal and Conditional Associations Can Differ

The model’s conditional hazard ratio need not equal a population-average contrast. To see the issue, suppose individuals have different latent hazards. As time passes, those at higher risk tend to experience the event earlier. The people who remain event-free in a group therefore become a selected subset. A marginal cause-specific hazard averages over the latent characteristics of those still at risk, and that composition can change over time.

For a latent state \(U\), the marginal cause-specific hazard among those still event-free can be written as an average of individual hazards weighted by the distribution of \(U\) among survivors:

The weights \(P(U=u\mid M,T\ge t)\) need not equal their baseline values \(P(U=u\mid M)\). Thus, even when a conditional model has a stable coefficient, the corresponding marginal association may vary with time. In competing risks, experiencing another cause also removes a person from the risk set for cause \(k\), so the cause-specific hazard is not itself the cumulative probability of cause \(k\).

This distinction is one reason to state exactly what a coefficient estimates. A cause-specific joint-model coefficient is a conditional parameter in the specified model. A cumulative incidence is a population probability that depends on the cause-specific hazards together. Neither should be casually substituted for the other.

A Short Derivation and Its Interpretation

The hazard ratio formula follows directly by comparing two conditional hazards at the same time, for individuals with the same \(w_i\) and \(b_i\), but marker values differing by \(\Delta\). Denote those marker values by \(m\) and \(m+\Delta\). Then

The baseline hazard and covariate term cancel, yielding a conditional association on the hazard scale. This algebra establishes the model-based comparison; it does not establish that \(m\) can be changed independently of the factors that determine it. A causal interpretation would require a well-defined intervention and additional identification assumptions, such as consistency, exchangeability given measured pre-intervention covariates, and positivity for the intervention levels being compared. It would also require a method that targets the causal estimand of interest.

For a biomarker that is a manifestation of underlying disease, “set the biomarker to a higher value” may not even define a single intervention: different treatments could change the marker through different pathways and have different effects on the event. In that setting, a biomarker coefficient can still have prognostic value while lacking a straightforward intervention interpretation.

Diagnostics, Alternatives, and Limitations

Interpretation is only as credible as the model and data that support it. Check whether the longitudinal trajectory and residual variation describe the repeated measurements adequately, whether the chosen random-effects distribution is plausible, and whether the event-time specification fits each cause. Also examine whether estimates are stable under reasonable alternative association structures and baseline-hazard choices. Sparse events for a cause can make its association estimate unstable; a narrow-looking numerical output should not be taken as proof that the latent structure is well identified.

A useful diagnostic question is whether conclusions change materially when plausible unmeasured-health dependence or random-effects assumptions are varied. If they do, report that sensitivity rather than presenting one fitted coefficient as definitive. A joint model cannot diagnose every unmeasured common cause from the observed data alone. Nor does good predictive performance validate a causal interpretation.

The main alternative depends on the question. For prediction, a validated joint model may be appropriate, with performance assessed in the intended population. For an intervention question, the analyst must define the intervention and target outcome first, then use a causal design and assumptions suited to that question. Simply changing the label on \(\alpha_k\), or adding more covariates without justification, does not change the estimand.

Common Pitfall

A related mistake is to describe a cause-specific hazard ratio as a ratio of cumulative event probabilities. A cause-specific hazard concerns an instantaneous rate among people still event-free; cumulative incidence is a probability over time and reflects competing events as well. State the scale and risk set whenever reporting an association.