Marker-Specific Current-Value Associations

Why Give Each Marker Its Own Association?

When several biomarkers are measured repeatedly, each may carry information about a subject’s event risk. A joint model can connect the event hazard to the markers’ underlying, time-specific values rather than to noisy measurements taken at particular visits. Giving each marker its own coefficient asks a focused question: what is the association of one marker’s current value with the hazard, conditional on the other markers’ current values?

That conditional qualification matters. A marker-specific coefficient is not generally the same as the coefficient from a model using that marker alone. If markers move together, the model estimates their associations while accounting for their overlap. The result is a conditional association on the specified marker scales, not by itself a causal effect.

Model Definition and Interpretation

Let \(i=1,\ldots,n\) index subjects and \(k=1,\ldots,K\) index markers. Write \(Y_{ikj}\) for the observed value of marker \(k\) for subject \(i\) at visit time \(t_{ikj}\). Let \(m_{ik}(t)\) denote the latent, or underlying, value of marker \(k\) for subject \(i\) at time \(t\), as specified by the longitudinal model. Let \(T_i\) be the event time, \(C_i\) the censoring time, \(U_i=\min(T_i,C_i)\) the observed follow-up time, and \(\delta_i=I(T_i\le C_i)\) the event indicator. Let \(\mathbf w_i\) be a vector of baseline covariates with coefficient vector \(\boldsymbol\gamma\).

For two markers, the linear predictor contains \(\alpha_1m_{i1}(t)+\alpha_2m_{i2}(t)\). Holding the other marker values and covariates fixed, increasing marker \(k\) by one unit multiplies the hazard by \(\exp(\alpha_k)\). Thus, \(\alpha_k\) is a log-hazard change per one-unit increase in marker \(k\), conditional on the other current marker values. Its interpretation depends on the marker’s measurement scale.

The “current value” is the model-estimated latent value at time \(t\), not necessarily the most recent observed measurement. This distinction allows the event model to use the longitudinal process while accounting for measurement noise and irregular visit times through the longitudinal model.

Assumptions Behind the Interpretation

The coefficient interpretation follows from the model only when its components are appropriately specified. The hazard form assumes proportionality conditional on the marker values: the log hazard is linear in those values, and each \(\alpha_k\) is constant over time unless the model explicitly includes time-varying effects. The longitudinal models must represent the marker trajectories adequately over the relevant follow-up period. In a multivariate linear mixed model, for example, each marker has a specified mean trajectory and random effects, and the random effects can have a joint covariance structure.

The joint model also needs a defensible dependence structure. Conditional on the subject-specific latent effects, the longitudinal observations and event process are modeled using their specified distributions. If censoring is treated as noninformative, the analysis assumes that censoring does not carry additional information about the event process after conditioning on the modeled information. These assumptions do not follow merely from having repeated measurements; they must be considered in the study context.

The markers must be distinguishable enough in the observed data to estimate their separate contributions. If two latent marker trajectories are nearly redundant, the model may have difficulty determining which marker accounts for the conditional association. A fitted model can still return estimates, but the estimates may be imprecise or sensitive to modeling choices.

Why the Coefficient Is a Conditional Association

The interpretation can be justified directly by taking a ratio of hazards. Consider two marker-value vectors for otherwise identical covariate values at the same time, denoted by \(\mathbf m\) and \(\mathbf m'\). The shared baseline hazard and the covariate term cancel:

If only marker \(r\) changes by \(d\) units, then \(m_r-m'_r=d\) and all other differences are zero. The ratio reduces to \(\exp(\alpha_r d)\). This establishes both the one-unit hazard ratio \(\exp(\alpha_r)\) and the condition that other marker values are held fixed.

This calculation is a comparison of hazards at a time, not a comparison of cumulative event probabilities. A hazard describes the instantaneous event rate among subjects still at risk. The hazard ratio does not, on its own, give a risk difference or an event probability.

How the Latent Values Enter the Joint Model

Let \(\mathbf b_i\) collect the subject-specific random effects across markers, with density \(p(\mathbf b_i)\). The longitudinal model gives a conditional density \(p(\mathbf Y_i\mid\mathbf b_i)\), where \(\mathbf Y_i\) denotes the observations for subject \(i\). Given the random effects, the marker trajectories determine the hazard. With \(S_i(t\mid\mathbf b_i)=\exp\{-\int_0^t h_i(u\mid\mathbf b_i)\,du\}\), the event-time contribution for an observed time \(U_i\) and event indicator \(\delta_i\) is \(h_i(U_i\mid\mathbf b_i)^{\delta_i}S_i(U_i\mid\mathbf b_i)\). Integrating over the random effects gives the subject’s joint contribution:

This expression shows why the current-value association uses latent trajectories: the same random effects help describe repeated marker observations and determine the event hazard over follow-up. The likelihood integrates over uncertainty in those subject-specific effects rather than treating an estimated trajectory as known without error. The particular longitudinal distributions, random-effects covariance, and baseline-hazard specification must still be chosen and checked as part of fitting the joint model.

Diagnostics, Alternatives, and Limitations

Marker-specific estimates require particular care when markers are strongly correlated. A useful diagnostic is to inspect coefficient uncertainty, correlations among estimated association parameters, and sensitivity of estimates to reasonable changes in the longitudinal or random-effects specification. Large standard errors, unstable signs across fits, or broad profile likelihoods can indicate that the data do not separate the markers’ conditional contributions well. These symptoms are evidence of weak information, not proof that either marker has no association.

Assess whether the assumed hazard relationship is plausible. For example, a constant log-hazard change per marker unit may be inadequate if the association changes over time or is nonlinear in marker value. Examine model fit and residual-based diagnostics appropriate to the fitted survival model, along with whether the longitudinal model reproduces observed marker patterns. Also check that event information is adequate relative to the number of association parameters and that numerical integration or other fitting procedures have converged reliably.

The main alternative considered here is a shared-random-effects link, in which the hazard is connected to subject-specific random effects rather than including a separate coefficient for each current marker value. That link asks how latent subject-level trajectory features relate to the hazard. The current-value formulation instead asks for conditional associations with the markers’ underlying values at time \(t\). These are different scientific specifications; neither is automatically preferable. Fitting each marker in a separate survival model is also not equivalent to estimating marker-specific coefficients jointly, because separate models do not condition on the other marker.

The current-value model has limitations. Its coefficients are conditional on the included markers, covariates, trajectory models, and hazard form. They may be difficult to interpret when markers have very different scales or are highly correlated. A current-value association also does not represent every possible way a marker history could relate to event risk. Finally, estimated associations should not be described causally without a separate causal argument and assumptions that justify such an interpretation.

Common Mistake

• Wrong reasoning: “The coefficient for marker 1 is positive, so increasing marker 1 causes the event hazard to rise, regardless of marker 2.”
• Correction: The coefficient describes a conditional association per unit of marker 1, holding the other modeled marker values and covariates fixed. It does not establish causation, and the estimate may be uncertain or difficult to separate from marker 2 when the trajectories are strongly related.
• Wrong reasoning: “A hazard ratio of 1.8 means the event probability is 80% higher.”
• Correction: It means the instantaneous hazard is 1.8 times as large for the stated conditional comparison. A probability or cumulative risk comparison requires additional information, including the baseline hazard and follow-up time.