Liver Stiffness and Cirrhosis Decompensation

Why the Event Definition Matters

Liver stiffness measurements can be collected repeatedly, while cirrhosis decompensation is a clinical event that occurs at a particular time. A joint model connects the underlying stiffness trajectory to the time until an event, accounting for the fact that measurements from one patient are related and that follow-up may end when the event occurs. But the model’s association is only as clinically meaningful as its event definition. “Decompensation” must be specified in terms of observable, adjudicated events before interpreting a hazard ratio.

For the hypothetical analysis below, define the endpoint as time from study entry to the first documented occurrence of ascites, variceal bleeding, or hepatic encephalopathy, according to prespecified clinical criteria. These three components make a composite endpoint; the model’s association concerns the first occurrence of any one of them, not a generic biological state called “worsening cirrhosis.” The components may differ in severity, ascertainment, and relationship to stiffness. A single composite association does not show that the association is the same for each component.

Model for Log Liver Stiffness and First Decompensation

Let \(Y_{ij}\) be the measured liver stiffness, in kPa, for patient \(i\) at time \(t_{ij}\), measured in years since study entry. Because stiffness is positive and often analyzed on a logarithmic scale, set \(y_{ij}=\log(Y_{ij})\). A linear mixed model for the repeated log measurements is

Here \(m_i(t)\) is the patient-specific latent mean log stiffness, \(\beta_0\) and \(\beta_1\) describe the population trajectory, and \(b_{0i}\) and \(b_{1i}\) are patient-specific random intercept and slope. A common specification assumes the random-effects vector \((b_{0i},b_{1i})\) follows a mean-zero bivariate normal distribution with an estimated covariance matrix. Conditional on the random effects, measurement errors \(\epsilon_{ij}\) have mean zero, constant variance, and are independent across measurement occasions; they are also independent of the random effects.

Let \(T_i\) be time to the first qualifying decompensation event and let \(\delta_i=1\) if that event is observed, with \(\delta_i=0\) if observation is right-censored before it. A current-value hazard model is

where \(h_0(t)\) is an unspecified baseline hazard, \(w_i\) is a vector of observed covariates, and \(\gamma\) is its coefficient vector. The parameter \(\alpha\) links latent log stiffness to the instantaneous hazard of the defined first event. The hazard is conditional on the patient still being event-free at \(t\), the modeled covariates, and the shared random effects. This is an association model, not evidence that changing stiffness would cause the event.

Assumptions and Interpretation of the Association

Interpretation relies on the model being a reasonable description of the data. The longitudinal mean trajectory must adequately represent changes in log stiffness over the observed time range; the assumed distribution and covariance structure for random effects and the measurement-error model must be adequate. The event model assumes proportional hazards conditional on the modeled quantities, and the effect of the latent value is linear on the log-hazard scale. The current-value formulation assumes that the latent stiffness at time \(t\), rather than only a past value or rate of change, is the quantity linked to the hazard.

The shared-effects structure must also be suitable: after conditioning on modeled covariates and the shared latent trajectory, the remaining dependence between repeated measurements and event times should be sufficiently represented. For ordinary right-censoring, censoring must be conditionally noninformative for the event-time likelihood given the modeled information. If death before a qualifying decompensation is possible, it is a competing event, not a decompensation-free observation that can automatically be treated as ordinary independent censoring. A cause-specific hazard analysis may censor at death under its assumptions, but its hazard interpretation is not the same as a cumulative probability of decompensation in the presence of death.

The scale of the association is especially important. Since \(m_i(t)\) is on the natural-log kPa scale, a one-unit increase in \(m_i(t)\) corresponds to multiplying latent stiffness by \(e\), not increasing stiffness by one kPa. If two otherwise comparable patients differ in latent log stiffness by \(d\), the model-implied hazard ratio is \(e^{\alpha d}\). For a doubling of latent stiffness, \(d=\log 2\), so the hazard ratio is \(e^{\alpha\log 2}\). This conversion gives the estimate a more clinically legible scale.

Uncertainty and the Meaning of a Composite Event

A confidence interval for \(\alpha\) is on the log-hazard scale. Transforming its endpoints gives an interval for a hazard ratio, but the transformation must use the contrast of interest. For a doubling, multiply the endpoints by \(\log 2\) before exponentiating. The resulting interval quantifies uncertainty in the model-based relative hazard for that contrast; it is not an interval for an individual’s event probability.

Estimation, Diagnostics, and Alternatives

Joint estimation uses the repeated stiffness values and event or censoring information together. In outline, the likelihood integrates over the unobserved patient-specific random effects: for each patient it combines the longitudinal measurement contribution with the event-time contribution conditional on those effects, then averages over their distribution. The estimated association is therefore not obtained by simply regressing observed event times on the last recorded stiffness value.

A practical analysis should inspect observed trajectories and fitted trajectories on the log scale, residual patterns, and the estimated random-effects distribution. Check whether a linear time trend is plausible across the time range: systematic curvature or distinct phases can indicate that the longitudinal mean is inadequate. Examine whether residual variability changes with time or fitted value, and whether influential patients drive the estimated association. For the survival component, assess whether a single proportional-hazards association is plausible, including whether the stiffness association appears to change over follow-up. Also verify endpoint coding, event dates, and reasons for loss to follow-up. Sparse measurements near event times can make the current latent value weakly identified even if the model converges.

The shared-parameter joint model is useful when repeated measurements are noisy and the outcome process can relate to the underlying trajectory; it can account for that dependence under its assumptions. A Cox model using a single baseline stiffness measurement is simpler, but it does not represent subsequent trajectory information. A time-updated Cox analysis using the most recent observed measurement is another alternative, but it can be vulnerable to measurement error and to informative observation or dropout unless those processes are handled appropriately. Two-stage approaches that first estimate trajectories and then use estimated values in an event model may not fully propagate uncertainty from the first stage. These approaches can answer useful questions, but they are not interchangeable with joint modeling.

The joint model also has limitations. A fitted \(\alpha\) does not prove that stiffness causes decompensation; disease severity, treatment, and other unmeasured factors may affect both. A composite endpoint can obscure differences among ascites, bleeding, and encephalopathy, and its association may depend on which component occurs most often. Liver stiffness itself can be influenced by factors other than fibrosis, such as acute inflammation or congestion; the model does not automatically distinguish these mechanisms. Results also depend on how the measurement scale, time course, event definition, and competing events are specified.

A Short Derivation of the Hazard-Ratio Contrast

For two otherwise comparable patients at time \(t\), let their latent log stiffness values be \(m_A(t)\) and \(m_B(t)\). The hazard model gives

The baseline hazard and shared covariate contribution cancel because they are the same for both patients. If patient A has twice patient B’s latent stiffness, then \(m_A(t)-m_B(t)=\log(2)\), since \(m\) is the log of stiffness. Substitution yields \(\exp\{\alpha\log(2)\}\). This derivation also shows why the hazard ratio describes a relative instantaneous rate among patients still event-free, not the ratio of their cumulative risks over an interval.

Common Pitfalls

• Treating a composite as one uniform clinical event. The wrong reasoning is that any positive association applies equally to ascites, bleeding, and encephalopathy. The correction is to report the modeled endpoint as first occurrence of the specified composite and, when clinically important, examine components with an appropriate separate analysis.
• Calling the association causal. A shared-parameter model addresses dependence induced by the modeled latent trajectory; it does not remove all confounding or prove that stiffness changes cause decompensation.
• Reading \(\alpha\) as a per-kPa effect. With log stiffness in the model, \(\alpha\) is per one-unit increase in log stiffness. For a doubling, use \(\exp\{\alpha\log(2)\}\).
• Ignoring death before decompensation. Death may prevent a later decompensation from being observed. Treating it as ordinary censoring without considering the target estimand and assumptions can misrepresent the clinical question.