Latent Marker Values After the Last Measurement

Why a Latent Marker May Be Needed After the Last Measurement

In longitudinal studies, marker measurements are often intermittent, while an event time can occur between scheduled visits or after the final recorded marker value. A joint model can still define the subject’s underlying marker level at that time: the fitted mixed model supplies a trajectory, and the trajectory can be evaluated at an unobserved time. This is an extrapolation, not a recovered measurement. Its credibility depends on whether the trajectory model and its assumptions remain reasonable beyond the subject’s observed measurement times.

Let \(Y_{ij}\) be the marker measurement for subject \(i\) at time \(t_{ij}\), and let \(m_i(t)\) denote that subject’s latent marker trajectory. A common mixed-model specification is

Here \(x_i(t)\) and \(z_i(t)\) are known covariate vectors, \(\beta\) is a vector of fixed effects, and \(b_i\) is the subject-specific random-effects vector. The residual \(\epsilon_{ij}\) represents measurement-level variation around the latent trajectory. For example, a random-intercept-and-slope model has \(z_i(t)=(1,t)^\top\), so a subject’s latent marker can vary in both level and rate of change.

The latent trajectory is a model-defined quantity. It does not claim that the marker was measured at \(t^*\), nor that a biological process must continue according to the fitted curve. If an event occurs at \(t^*\), the model can assign a latent marker value at that event time, provided the trajectory is defined there. When the event terminates follow-up or makes later measurement impossible, the value remains unobserved.

How Observed Measurements Inform a Later Value

A prediction beyond the last visit should reflect what the subject’s measurements say about their random effects, while also allowing for uncertainty about those effects. To show this calculation explicitly, suppose the model parameters are treated as known and

with \(b_i\) independent of \(\epsilon_i\). Here \(Y_i\) collects the subject’s observed marker values, and \(X_i\) and \(Z_i\) contain the corresponding fixed- and random-effects covariates. Under these Gaussian assumptions, the conditional distribution of \(b_i\) given the observed markers is normal. Its mean and covariance are

At a target time \(t^*\), write \(x_i^*=x_i(t^*)\) and \(z_i^*=z_i(t^*)\). Then the conditional mean and variance of the latent marker are

These expressions make an important distinction: the conditional mean is a point prediction, while the conditional variance describes uncertainty in the latent trajectory under the specified model. A prediction interval for a future observed marker is wider because it includes residual variation as well.

Why Extrapolation Uncertainty Can Increase

The uncertainty calculation follows directly from the linear form of the trajectory. If \(b_i\mid Y_i\) has covariance \(V_i\), then the random part of \(m_i(t^*)\) is \(z_i^{*\top}b_i\). Therefore,

For a random intercept and slope, \(z_i^*=(1,t^*)^\top\). If \(V_i\) has entries \(V_{00},V_{01},V_{11}\), the variance is \(V_{00}+2t^*V_{01}+(t^*)^2V_{11}\). Thus uncertainty about the subject’s slope can contribute a term that grows quadratically with time. Extrapolating farther from observed visits can be much less certain even when the fitted mean curve looks smooth.

This calculation treats \(\beta\), \(D\), and \(R\) as known. In practice they are estimated, so a complete uncertainty assessment should also account for parameter-estimation uncertainty. Conditional intervals that hold fitted parameters fixed can be narrower than intervals that incorporate this additional source of uncertainty.

Assumptions, Diagnostics, and Alternatives

An extrapolated value is only as credible as the assumptions supporting the trajectory and the information used to estimate it. In particular, the calculation assumes that the chosen time scale and functional form adequately represent the marker trajectory beyond the observed window. A linear random-slope model assumes a straight-line subject-specific trend; it cannot capture unmodeled curvature merely because the target time is important. The Gaussian random-effects and residual assumptions in the derivation above are needed for those exact conditional normal formulas. Other mixed models may use different distributions, but still require a specified, defensible trajectory and uncertainty calculation.

A second assumption concerns why the marker stopped being measured. If measurement times or dropout depend on the marker after conditioning on modeled information, the observed measurements may not adequately represent the latent trajectories of subjects who are no longer measured. In a joint model, event information can also contribute information about random effects through the survival component. A prediction based only on the longitudinal measurements conditions on \(Y_i\); a joint-model prediction that uses event history conditions on both the marker data and the event history under the specified joint model. These are different conditional predictions. Neither creates an observed marker value where none was collected.

Useful checks include plotting observed marker values and fitted trajectories over time, checking whether residual patterns suggest curvature or changing variability, and examining whether the model fits subjects whose measurements cover different portions of follow-up. For extrapolation specifically, compare predictions under plausible alternative trajectory forms, such as a linear trend and a suitably specified nonlinear trend, and report how predictions and intervals change. If the target time lies far beyond the observed time range, label the prediction as highly model-dependent rather than presenting it as if it were interpolated between visits.

A common alternative is to carry the last observed marker forward. This avoids extending a fitted slope, but assumes the marker remains constant after the last visit and does not provide a principled account of trajectory uncertainty. Interpolation between observations uses measurements on both sides of a target time; it is not available when the target is later than the final observation. A mixed-model extrapolation uses the fitted population trajectory and subject-specific information, but its validity depends on the trajectory continuing to be appropriate. The alternatives answer different questions and rely on different assumptions.

The central limitation is that observations do not directly verify a trajectory after the final measurement. Model checks can reveal a poor fit within the observed range, but cannot prove that a chosen trajectory remains valid beyond it. Extrapolated values are therefore useful model-based quantities in the joint analysis, not substitutes for post-event measurements or guarantees about an individual’s unobserved biology.