Current-Value Association with Noisy Measurements

Why the Current Value Is Not the Latest Observation

A longitudinal marker is usually measured intermittently, and each recorded value contains both information about the subject’s underlying trajectory and measurement noise. A current-value association is meant to connect the event hazard to the subject’s underlying marker level at time \(t\), not to the accidental upward or downward error in one measurement. The distinction matters especially when measurements are sparse, imprecise, or taken at times that do not coincide with the event-risk time of interest.

The introductory material on current-value links defines the association through the marker value at the time the hazard is evaluated. Here the central issue is what that value represents. In a joint model, the hazard uses the latent trajectory specified by the longitudinal model. The observed marker informs estimation of that trajectory, but it is not itself the hazard covariate.

Measurement and Hazard Models

A standard linear mixed longitudinal submodel can be written as

Here \(\mathbf{x}_i(t)\) and \(\mathbf{z}_i(t)\) are specified covariate vectors, \(\boldsymbol{\beta}\) is a vector of fixed effects, and \(\mathbf{b}_i\) is a vector of subject-specific random effects. A common assumption is that \(\epsilon_{ij}\) has mean zero and variance \(\sigma_\epsilon^2\), is independent across measurements conditional on the random effects, and is independent of the random effects. The random effects describe between-subject trajectory variation; the errors describe measurement-level variation around each trajectory.

For a current-value link, a typical conditional hazard model is

The hazard is conditional on the subject-specific trajectory, represented through \(\mathbf{b}_i\), and on the included covariates \(\mathbf{w}_i\). The baseline hazard is \(h_0(t)\), and \(\alpha\) is the current-value association parameter. If the marker is measured in units \(U\), then \(\alpha\) has units \(U^{-1}\), so that \(\alpha m_i(t)\) is dimensionless. A hazard ratio for two otherwise comparable subjects whose latent marker values differ by \(\Delta\) at time \(t\) is \(\exp(\alpha\Delta)\).

These models require more than a convenient formula. The longitudinal mean and random-effects structure must adequately describe the marker trajectories; the measurement-error distribution and variance must be reasonable; and the hazard model must correctly specify how the latent trajectory and other covariates relate to event risk. The usual shared-random-effects formulation also assumes that, conditional on the modeled random effects and covariates, the longitudinal measurement errors do not carry additional information about the event process. If measurement timing or measurement error itself is related to event risk in ways not represented in the model, the stated conditional interpretation may fail.

Why Replacing the Latent Value Changes the Model

Suppose for a moment that a measurement is taken exactly at time \(t\), with \(Y_i(t)=m_i(t)+\epsilon_i(t)\). Replacing \(m_i(t)\) by \(Y_i(t)\) changes the hazard multiplier from \(\exp\{\alpha m_i(t)\}\) to \(\exp\{\alpha[m_i(t)+\epsilon_i(t)]\}\). The ratio of the substituted multiplier to the intended multiplier is therefore

Unless \(\alpha=0\) or the measurement error is exactly zero, this ratio varies with the error. Thus a raw measurement injects a random, subject- and occasion-specific perturbation directly into the hazard. If errors are normal with mean zero and variance \(\sigma_\epsilon^2\), the expected substituted multiplier conditional on \(m_i(t)\) is

This follows from the moment-generating function of a normal random variable: for \(\epsilon\sim N(0,\sigma_\epsilon^2)\), \(E[\exp(\alpha\epsilon)]=\exp(\alpha^2\sigma_\epsilon^2/2)\). The extra factor shows that even mean-zero error need not average out after exponentiation. This calculation is a justification for avoiding the substitution; it is not a general formula for the bias in an estimated association parameter. In a fitted survival model, risk sets, event times, covariates, and latent-trajectory uncertainty all affect estimation, so the direction and amount of coefficient bias are not determined by this factor alone.

Model-Estimated Does Not Mean Known Exactly

The latent value \(m_i(t)\) is not directly observed. It is defined by the longitudinal model and estimated using the available repeated measurements, covariates, and—in a joint analysis—the event information through the shared model. The joint likelihood accounts for the fact that a subject's random effects are unknown rather than treating a noisy measurement as the true trajectory. In the hazard submodel, \(m_i(t)\) is the model-defined value; it is not replaced by a preliminary point prediction and then treated as error-free.

A subject-specific fitted trajectory can be useful for displaying or summarizing results, but a point prediction does not remove uncertainty about the trajectory. For example, a subject with few noisy measurements may have a predicted value pulled toward the population trajectory, with substantial remaining uncertainty. Joint estimation propagates latent-effect uncertainty through the model; a two-stage procedure that first predicts a marker value and then fits a survival model as if that prediction were known can understate uncertainty and can produce different estimates.

Alternatives, Diagnostics, and Limits

The main alternatives differ in what they treat as the hazard covariate. Their consequences are not interchangeable.

Several checks help assess whether the latent-value link is credible. Examine whether the longitudinal model describes the observed trajectories, including systematic residual patterns, changes in variability over time, and unusually influential observations. Compare the estimated measurement-error variation with the variation in fitted trajectories: when error is large relative to between-subject trajectory differences, individual latent values will be weakly determined. Check whether the assumed error distribution and independence structure are plausible. Finally, assess whether the specified time pattern and hazard association are adequate; a good measurement model cannot compensate for a misspecified hazard model.

The approach has important limitations. The latent marker is only as credible as the model used to define it. A misspecified mean trajectory, random-effects distribution, measurement-error structure, or hazard link can yield misleading estimates. If the timing or frequency of measurements depends on unmodeled health changes or event risk, treating observation times as ignorable may also be inappropriate. Sparse measurements can make the trajectory between visits—and especially near or after the last visit—highly uncertain. The latent-value formulation avoids equating a noisy observation with the marker, but it does not make the unobserved marker directly known.

A Common Mistake

The practical distinction is not between using data and ignoring data. The observed measurements are essential: they inform the trajectory parameters and subject-specific effects. The distinction is between using measurements as evidence about the latent marker and declaring one noisy measurement to be the marker itself.