Shared Versus Cause-Specific Association Constraints

Why Constrain Associations Across Causes?

A longitudinal marker can be associated with more than one type of event, but the association need not be equally strong for every cause. For example, a higher latent marker value might be associated with increased hazards for two event types, while the magnitude of that association differs. A joint model can either impose a common association parameter or estimate a separate parameter for each cause. The choice changes the scientific assumption encoded in the model, not merely its notation.

Let \(T_i\) be subject \(i\)'s event or censoring time, and let \(D_i\in\{1,\ldots,K\}\) denote the event cause when an event occurs. Let \(m_i(t)\) be the subject's latent longitudinal marker value at time \(t\), as defined by the longitudinal submodel. The cause-specific hazard is the instantaneous event rate for cause \(k\) among subjects who have not yet experienced any event. It is conditional on the latent trajectory and any modeled covariates.

Model Forms and What the Constraint Means

For a current-value association, one common specification is

Here \(\lambda_{0k}(t)\) is the baseline hazard for cause \(k\), \(w_i\) is a vector of covariates, \(\gamma_k\) is its cause-specific coefficient vector, and \(b_i\) denotes the latent random effects used in the longitudinal model. The displayed form conditions on the latent effects and trajectory. In a joint likelihood, uncertainty about those effects is accounted for through the longitudinal and event contributions together; the observed, error-prone marker is not simply substituted for \(m_i(t)\).

Under a shared association constraint, \(\alpha_k=\alpha\) for every cause. Under the cause-specific structure, the \(\alpha_k\)'s are unrestricted. For a one-unit increase in the latent marker at a fixed time, the cause-\(k\) hazard ratio is \(\exp(\alpha_k)\), holding the other terms fixed. The shared constraint therefore asserts equality of these log hazard ratios across causes, not equality of the cause-specific hazards themselves.

If the model includes both current value and slope, a corresponding specification is

One may constrain only the current-value parameters to be common, only the slope parameters, or both. A common pair \((\alpha,\xi)\) imposes two equalities across causes; it is not implied by using a shared current-value parameter alone. The units matter: \(\alpha_k\) is per unit of marker, whereas \(\xi_k\) is per unit of marker change per unit of time.

Equality of association parameters does not require equal baseline hazards. For example, \(\lambda_{01}(t)\) can be much larger than \(\lambda_{02}(t)\) even when \(\alpha_1=\alpha_2\). Nor does a common \(\alpha\) make the overall event processes identical: the event causes have distinct cause-specific hazards, and the chance of observing one cause also depends on the competing hazards.

Assumptions Imposed by the Constraint

Both structures rely on the assumptions of the specified joint model. In particular, the longitudinal submodel must represent the latent marker trajectory adequately; the cause-specific hazard form must be suitable for the event process; and censoring must be non-informative conditional on the information included in the model, or be modeled appropriately. The association is a conditional model parameter, not automatically a causal effect.

The shared structure adds a substantive restriction: for the chosen longitudinal feature and link function, a one-unit change has the same log-hazard association for every cause. It does not assert that the causes have equal event rates, equal covariate effects, or identical mechanisms. The separate structure removes that equality restriction, but still assumes that each cause's association follows the chosen functional form—for example, a linear effect of \(m_i(t)\) on the log hazard.

The restriction should be chosen for a reason. A common parameter may be appropriate when a common biological or measurement-scale association is scientifically plausible, or when a more parsimonious model is needed. Separate parameters are appropriate when cause-specific mechanisms plausibly differ and the data can support estimating the additional quantities. Neither structure is guaranteed to be correct merely because it converges.

Testing the Equality Restriction

For \(K\) causes and one association parameter, the shared model is nested within the separate model. The null hypothesis is \(H_0:\alpha_1=\cdots=\alpha_K\); the alternative allows the \(\alpha_k\)'s to differ. If both current-value and slope parameters are constrained, the joint null also includes \(\xi_1=\cdots=\xi_K\). Under standard regularity conditions, a likelihood-ratio test compares the maximized log likelihoods.

When the models use the same data and likelihood, the shared model is correctly nested in the separate model, and the usual large-sample regularity conditions hold, \(G^2\) is approximately chi-square under \(H_0\). For one association parameter across \(K\) causes, the difference in parameter counts is \(K-1\), so \(df=K-1\). If both current-value and slope parameters are constrained across causes, the difference is \(2(K-1)\), provided both sets of parameters are otherwise unrestricted.

Comparing Constraints in Practice

A fully shared structure is parsimonious but can conceal meaningful heterogeneity if the true cause-specific associations differ. A fully separate structure is flexible but estimates more parameters. Intermediate structures are possible: for example, selected causes can share one association while another cause has its own. Such a model represents a specific set of equality restrictions and should be defined explicitly before fitting or testing it.

The likelihood-ratio test assesses the restrictions collectively. It does not identify a reliable cause-specific pattern by itself, and failure to reject equality is not proof that the effects are identical. The uncertainty in each estimate and the scientific size of the differences also matter. A non-significant test can reflect limited information, while a statistically detectable difference may be small on a meaningful hazard-ratio scale.

Diagnostics should include checking that both fits have converged to plausible maxima, that their likelihoods and data are comparable, and that estimates and standard errors are not extreme or unstable. For separate estimates, inspect the signs and magnitudes of \(\hat{\alpha}_k\), their standard errors, and—where available—profile likelihoods. A very large standard error, a flat likelihood profile, or materially different results across plausible starting values warns that the cause-specific structure may be weakly supported by the available data. These checks do not replace the model assumptions or the formal test.

Common Pitfalls and Limits

• Confusing a shared association with a shared hazard. Equal \(\alpha_k\)'s do not make \(\lambda_{ik}(t)\) equal across causes; the baseline hazards and other terms can differ.
• Reading a common coefficient as a common event probability. The coefficient describes a conditional cause-specific hazard association. It does not directly state a cumulative probability of a cause.
• Assuming that failure to reject equality proves equality. The test may have limited power. Report uncertainty and consider the plausible size of cause-specific differences.
• Counting restrictions incorrectly. With \(K\) causes, one association component has \(K-1\) equality restrictions, not \(K\). Two separately constrained components have \(2(K-1)\).
• Treating the more flexible fit as automatically preferable. The separate model cannot have a lower maximized likelihood when truly nested and fitted successfully, but its extra flexibility can yield imprecise estimates and does not guarantee better scientific explanation.

These models are tools for representing and evaluating a particular association structure. They do not settle whether the longitudinal feature was measured without error, whether its functional form is adequate, or whether the association is causal. Those questions require separate scientific and modeling judgments.