Latent Trajectories Versus Observed Measurements

Why the Distinction Matters

A biomarker measured during follow-up is not usually observed without error. A laboratory value can vary because of assay imprecision, short-term biological fluctuation, or specimen handling, even when the subject’s underlying marker level has changed little. If an event hazard is meant to reflect the subject’s evolving marker status, treating each recorded value as that status can make the event association respond to noise rather than to the trajectory of interest.

In the basic shared-random-effects joint model, the longitudinal measurements inform an underlying trajectory, and the event submodel relates risk to that trajectory. The distinction is important: a noisy measurement is evidence about the latent marker level, not necessarily the level itself. This tutorial focuses on that distinction, not on the later likelihood calculations.

A standard simple measurement-error specification is \(\epsilon_{ij}\sim N(0,\sigma_\epsilon^2)\), independent across measurements and independent of the event process conditional on the latent trajectory and shared random effects. The mean-zero assumption says that measurements are centered on the latent value; the variance \(\sigma_\epsilon^2\) quantifies residual measurement variability. These are modeling assumptions, not facts guaranteed by collecting repeated measurements.

The latent trajectory is commonly represented by a longitudinal model with subject-specific random effects, for example a subject-specific intercept and slope. Those random effects induce the trajectory’s variation across subjects and are shared with the event submodel. Under the current-value link, for two subjects with the same baseline hazard and other covariates, the hazard ratio at time \(t\) corresponding to latent marker values differing by \(\Delta m\) is \(\exp(\alpha\Delta m)\). Here \(\alpha\) is the log-hazard association per one-unit increase in the latent marker.

How Measurement Noise Changes the Link

Suppose two subjects have the same latent marker value at a given time but their recorded measurements differ because of error. A model that substitutes the recorded values directly into the hazard link would assign different risks to them. The latent-current-value model does not make that change solely because one assay happened to read high and the other low. Their longitudinal records may still provide different evidence about their latent trajectories, but a single residual deviation is not automatically interpreted as a true marker difference.

There is also a timing problem with directly using observations. Measurements are generally taken at discrete visits, while a hazard is defined over follow-up time. Between visits there may be no current observed value to insert. Carrying the last value forward or interpolating between readings creates an additional rule that can affect the apparent risk relationship. A latent trajectory provides a modeled marker value across time, including between measurement occasions.

A Short Derivation: Averaging a Noisy Plug-In Hazard

The consequences of substituting a noisy value can be made explicit under a Gaussian error model. Consider a hazard multiplier \(\exp(\alpha m_i(t))\) based on the latent value. If instead one uses \(Y_{ij}=m_i(t)+\epsilon_{ij}\), the multiplier becomes \(\exp\{\alpha(m_i(t)+\epsilon_{ij})\}\). Averaging this plug-in multiplier over a new measurement error gives

The second equality uses the moment-generating function of a mean-zero normal random variable with variance \(\sigma_\epsilon^2\). Unless \(\alpha=0\) or \(\sigma_\epsilon^2=0\), the average of the exponentiated noisy value differs from the exponentiated latent value. This is a consequence of the nonlinear exponential link, not evidence that the latent-value model should include this extra factor. It describes what happens when one averages a different, noisy plug-in multiplier over measurement error.

Repeated Measurements and the Latent Trajectory

Repeated readings help distinguish persistent subject-specific levels and trends from occasion-specific deviations. They do not eliminate measurement error, and an individual trajectory cannot generally be known exactly from a finite set of readings. Instead, the longitudinal submodel uses the observed record to estimate or infer the trajectory under its stated assumptions. The event submodel then uses that latent trajectory as the marker history relevant to the current-value association.

Assumptions, Diagnostics, and Alternatives

The latent-versus-observed distinction is useful only when the measurement and event models are credible. The standard formulation relies on assumptions such as the following:

• The longitudinal model represents the systematic marker pattern adequately, including relevant time trends and subject-to-subject variation.
• Conditional on the latent trajectory and modeled random effects, the measurement errors have the specified mean and variability and do not contain unmodeled systematic patterns.
• The current-value association is an appropriate description of how the latent marker relates to the hazard; it is not automatically a causal effect.
• The shared random-effects structure and conditional independence assumptions are adequate for connecting the longitudinal and event processes.

Diagnostics should target the parts of the model that make the latent trajectory meaningful. Examine longitudinal residuals for changing variance, trends over time, or outliers that suggest an inadequate measurement model. Compare observed marker profiles with fitted trajectories, and assess whether the trajectory form is plausible rather than relying on a single average trend. Consider whether assay precision differs across visits or subjects. A pattern of residuals related to time, risk factors, or impending events may indicate that the assumed measurement error or longitudinal structure is inadequate.

Also inspect the visit schedule and its relation to the event process. The simple argument that repeated readings inform a latent history can fail if visits occur more often when a subject is deteriorating and this visit process is not adequately addressed. Sparse measurements can leave trajectories weakly determined; substantial measurement error can make the latent association imprecise. Neither a sophisticated link nor repeated measurements can recover information that the data do not contain without relying on model assumptions.

These approaches are not interchangeable merely because each uses a marker in an event analysis. The latent link targets association with the modeled underlying trajectory. An observed-value approach may instead target association with the value available to a clinician at a visit, which can be scientifically relevant but must be specified and interpreted accordingly. Measurement-error correction is another option when its assumptions and data support it. The key is to state which marker quantity the risk association concerns.

Common Pitfall and Practical Interpretation

A common mistake is to say, “The assay was 9, so the subject’s current marker level was 9; therefore the hazard should use 9.” That reasoning equates a noisy observation with the latent quantity and ignores the measurement model. The correction is to treat the recorded value as an observation centered on the latent trajectory, then use the trajectory in the current-value association if that is the model being fitted.

A second mistake is to assume that replacing observations with a latent trajectory removes all uncertainty. It does not: the trajectory is estimated from finite, noisy data, and its adequacy depends on the model and available measurements. The association describes a conditional statistical relationship under that model. It does not establish that changing the biomarker would cause a change in event risk.