Serum Albumin and Decompensation in Cirrhosis

Why Link Albumin to an Event-Time Outcome?

Serum albumin is measured repeatedly in people with cirrhosis, while their clinical course is followed until decompensation, death, or the end of observation. Albumin can reflect underlying disease severity, but it can also change with treatment, fluid status, nutrition, and acute illness. The statistical question is therefore not simply whether a low observed value accompanies a bad outcome. It is whether a patient’s underlying albumin trajectory is associated with their instantaneous risk of the event, after accounting for measurement error, treatment, and other modeled covariates.

In this case study, the endpoint is the first occurrence of cirrhosis decompensation or death. The composite must be defined before analysis: a person is counted as having the event at the first qualifying occurrence, and a person without either event is censored at their last event-free follow-up. The joint model below is a way to estimate an association between the longitudinal marker and this event-time outcome. It does not, by itself, establish that changing albumin would change the risk of decompensation or death.

A Current-Value Joint Model for Albumin

Definition: Let $Y_{ij}$ be the observed serum albumin for patient $i$ at time $t_{ij}$, measured in g/L. A longitudinal submodel represents it as $Y_{ij}=m_i(t_{ij})+\epsilon_{ij}$, where $m_i(t)$ is the patient’s underlying albumin trajectory and $\epsilon_{ij}$ is measurement error. A current-value survival submodel relates the event hazard to that trajectory: $$ h_i(t\mid b_i)=h_0(t)\exp\{\gamma_A A_i+\gamma^\mathsf{T}X_i+\alpha m_i(t)\}. $$ Here $A_i$ indicates treatment group, $X_i$ denotes other covariates, $b_i$ denotes patient-specific random effects in the longitudinal model, and $h_0(t)$ is the baseline hazard. The parameter $\alpha$ describes the association between the underlying albumin value at time $t$ and the instantaneous event hazard, conditional on the other modeled terms.

One possible longitudinal specification is

m_i(t)=\beta_0+\beta_1t+\beta_A A_i+\beta_{At}A_it+b_{0i}+b_{1i}t, \qquad \epsilon_{ij}\sim N(0,\sigma^2).

The random intercept $b_{0i}$ and random slope $b_{1i}$ allow patients to differ in their underlying albumin levels and rates of change. The fixed treatment terms allow average trajectories to differ by treatment group. Time is measured from a defined origin, such as enrollment, in years; the albumin outcome is in g/L. A different time scale or a nonlinear trajectory may be more appropriate in another application.

The model requires more than a choice of equations. The trajectory form should represent the systematic time pattern adequately, the random-effects distribution and residual-error structure should be reasonable, and the survival model should describe the event hazard adequately. Conditional on the random effects and modeled covariates, the longitudinal errors and event process are assumed to be independent in the way specified by the joint model. Censoring is assumed non-informative conditional on the modeled information. If treatment changes over time, its timing and relation to prior albumin and disease status must be handled appropriately; a baseline treatment indicator alone does not represent a changing treatment history.

How the Joint Likelihood Connects the Two Outcomes

A patient’s repeated albumin measurements and event time are not treated as two unrelated data sources. Conditional on that patient’s random effects, the model evaluates both the longitudinal data and the event-time contribution. For an event or censoring time $T_i$, let $\delta_i=1$ indicate that the event was observed and $\delta_i=0$ indicate censoring. The event-time contribution conditional on $b_i$ is $h_i(T_i\mid b_i)^{\delta_i}S_i(T_i\mid b_i)$, where $S_i$ is the corresponding conditional survival function. The patient’s likelihood contribution integrates over the random effects:

L_i= \int \left\{\prod_j f_Y(Y_{ij}\mid b_i)\right\} h_i(T_i\mid b_i)^{\delta_i}S_i(T_i\mid b_i) f_b(b_i)\,db_i .

This integration is central. Patients with persistently lower underlying albumin can have different event hazards, and the shared random effects connect the marker record to the survival contribution. A separate analysis that treats every observed albumin value as error-free may not represent that connection.

The interpretation of $\alpha$ follows by comparing two underlying albumin values at the same time while holding treatment and other covariates fixed. If one value is $d$ g/L higher than another, the ratio of their modeled instantaneous hazards is

\frac{h_i(t\mid m_i(t)+d)}{h_i(t\mid m_i(t))} =\exp(\alpha d).

Thus, if $\alpha$ is negative, higher underlying albumin is associated with a lower instantaneous hazard, conditional on the model. This is an association on the hazard scale—not a statement about the probability of the event by a particular time.

Worked Example: Separating the Underlying Trajectory From Measurement Error

Suppose a hypothetical patient in the control group has observed albumin values of 34, 32, and 31 g/L at years 0, 1, and 2. A fitted longitudinal model for the cohort gives, for this patient, $\beta_0=34$, $\beta_1=-1.1$ g/L per year, $b_{0i}=-1.0$ g/L, and $b_{1i}=-0.3$ g/L per year. These are illustrative fitted values, not measurements from a real study.

At year 2, the model’s underlying trajectory is $$ m_i(2)=34+(-1.1)(2)+(-1.0)+(-0.3)(2)=30.2\ \text{g/L}. $$ The observed value is 31 g/L, so the residual at that visit is $31-30.2=0.8$ g/L. The model does not equate the single observed value with the underlying value; it separates the estimated trajectory from visit-level deviation. This distinction matters when the survival model uses $m_i(t)$, rather than a noisy $Y_{ij}$, as its marker association.

Interpreting the Estimated Association

Albumin is measured in g/L, so $\alpha$ is on a per-g/L scale. A more clinically readable contrast can be reported for a 5 g/L difference. If the model estimates $\hat{\alpha}=-0.0750$ per g/L, the estimated hazard ratio for a 5 g/L higher underlying albumin value is

\exp\{5(-0.0750)\}=\exp(-0.375)=0.6873.

In this hypothetical model, at the same time and conditional on treatment and the other included covariates, a 5 g/L higher underlying albumin value is associated with an estimated instantaneous hazard that is 0.6873 times as large, or about 31.3% lower. This describes a conditional association, not a 31.3% reduction in event probability or proof that raising albumin would reduce risk.

Worked Example: A Confidence Interval for the 5 g/L Contrast

Suppose the hypothetical estimate is $\hat{\alpha}=-0.0750$ per g/L with $SE=0.0180$ per g/L. A Wald 95% confidence interval for $\alpha$ is the estimate plus or minus 1.96 standard errors: $$ -0.0750\pm1.96(0.0180) =-0.0750\pm0.03528 \approx[-0.1103,-0.0397]. $$ The corresponding interval for the hazard ratio per 5 g/L higher albumin is obtained by multiplying both endpoints by 5 and exponentiating: $$ [\exp\{5(-0.1103)\},\exp\{5(-0.0397)\}] =[\exp(-0.5515),\exp(-0.1985)] \approx[0.5761,0.8200]. $$

Because the interval is below 1, these hypothetical results provide evidence of a negative conditional association at the 5% level under the Wald approximation. The interval is compatible with hazard ratios from about 0.5761 to 0.8200 per 5 g/L higher underlying albumin. It is a confidence interval for a model-based association parameter, not a 95% probability statement about the fixed parameter and not an interval for an individual patient’s event probability.

Distinguishing Disease Severity From Treatment-Related Change

A lower albumin trajectory may partly reflect more severe disease, while treatment may also alter albumin over time. The treatment terms in the longitudinal model describe modeled differences in trajectories; the treatment term $\gamma_A$ in the hazard model describes a treatment association conditional on the modeled albumin trajectory and covariates. These are distinct quantities. In particular, $\gamma_A$ is not automatically the total treatment effect, because it does not by itself summarize all pathways through which treatment could relate to the outcome.

Worked Example: Interpreting a Treatment-Related Trajectory Difference

Suppose the fitted longitudinal model gives $\hat{\beta}_A=1.2$ g/L and $\hat{\beta}_{At}=0.4$ g/L per year. In this model, the treatment-group difference in mean underlying albumin at time $t$ is $\beta_A+\beta_{At}t$. At year 2, the estimated difference is $$ 1.2+(0.4)(2)=2.0\ \text{g/L}. $$ Thus, the treatment group’s modeled mean underlying albumin is 2.0 g/L higher at year 2 than the control group’s, conditional on the model.

This difference is not, on its own, evidence that treatment improved clinical outcomes by raising albumin. If the estimated marker association were $\hat{\alpha}=-0.0750$ per g/L, a 2.0 g/L difference corresponds to a marker-associated hazard ratio of $\exp\{(-0.0750)(2.0)\}=\exp(-0.15)=0.8607$, holding other modeled terms fixed. That is the model’s association for the marker contrast. A causal decomposition of treatment effects requires additional assumptions about treatment assignment, confounding, and the mediator-outcome relation; a joint model alone does not supply those assumptions.

This distinction is especially important when treatment is given in response to worsening disease. If prior albumin or clinical status influences subsequent treatment, treatment may be confounded by indication, and prior marker values can affect later treatment decisions. Simply adding a treatment indicator to the model does not necessarily resolve that feedback. A randomized baseline treatment assignment makes some treatment comparisons more defensible, but the interpretation of albumin as a mediator still requires care.

Diagnostics, Alternatives, and Limitations

Model checking should address both submodels and their connection. Examine observed albumin trajectories alongside fitted trajectories, residual patterns over time, and whether variability changes with fitted albumin or time. Check whether the assumed time trend misses curvature or abrupt changes; in cirrhosis, acute episodes or treatment changes may make a straight-line trajectory inadequate. For the survival part, assess whether the chosen baseline hazard and proportional-hazards structure are plausible and compare predicted with observed event patterns. Check convergence, parameter uncertainty, and sensitivity of $\hat{\alpha}$ to reasonable alternatives for the trajectory or association structure.

The model can still fail if important changes in disease status or treatment are omitted, if censoring depends on unmodeled prognosis, or if the assumed trajectory is too rigid. Sparse albumin measurement near an event can make the estimated current value rely heavily on the trajectory assumptions. The association parameter may also be imprecise when few events occur or when albumin trajectories vary little. A joint model accounts for the specified shared structure; it does not automatically correct every source of bias, prove causation, or make the composite endpoint clinically homogeneous.

Approach	Strength	Important limitation
Joint longitudinal-survival model	Links a latent albumin trajectory and event time through shared model components.	Results depend on the trajectory, hazard, and dependence assumptions.
Time-updated Cox model using observed albumin	Uses measured values as time-varying covariates and can be straightforward to fit.	Observed values contain measurement error and may not represent the underlying trajectory.
Separate mixed model and survival model	Allows familiar analyses of each outcome separately.	A plug-in trajectory may fail to propagate uncertainty or account for the shared longitudinal-survival structure.

The appropriate alternative depends on the scientific question and data structure. A time-updated Cox model may be reasonable when measurements are sufficiently accurate and its assumptions are suitable. A joint model is useful when the underlying trajectory is of interest and repeated measurements and event times are plausibly linked, but it should be presented as a model-based analysis with sensitivity checks.

A Common Interpretation Pitfall

Wrong reasoning: “The estimated hazard ratio is 0.687 per 5 g/L, so increasing a patient’s albumin by 5 g/L will reduce that patient’s probability of decompensation or death by 31.3%.”
Correction: The model estimates a conditional association between underlying albumin and the instantaneous hazard. The contrast is not a causal treatment effect and is not a relative change in event probability. Treatment effects and causal mediation require additional design and assumptions.

Key takeaway: In a current-value joint model, the albumin coefficient links the estimated underlying albumin level to the event hazard. Interpret it with its units, conditioning variables, and model assumptions, and distinguish this association from treatment-related changes and causal effects.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If $\hat{\alpha}=-0.060$ per g/L, what is the estimated hazard ratio for a 5 g/L higher underlying albumin?
If a 95% confidence interval for $\alpha$ is $[-0.040,0.020]$ per g/L, what is the corresponding interval for the hazard ratio per 5 g/L higher albumin, and does it include 1?
If $\hat{\beta}_A=0.8$ g/L and $\hat{\beta}_{At}=-0.3$ g/L per year, what is the modeled treatment-group difference in underlying albumin at year 2?
Why is the hazard-model coefficient for treatment not automatically the total causal effect of treatment?

Solutions. 1. $\exp\{5(-0.060)\}=\exp(-0.30)=0.7408$, an estimated conditional hazard ratio of 0.7408 for the 5 g/L contrast. 2. Exponentiate 5 times each endpoint: $[\exp(-0.20),\exp(0.10)]=[0.8187,1.105]$. It includes 1, so this Wald interval does not exclude no association on the hazard-ratio scale. 3. $0.8+(-0.3)(2)=0.2$ g/L. 4. The coefficient is conditional on the modeled albumin trajectory and covariates; it does not by itself include all treatment pathways or establish that treatment assignment is unconfounded. Additional causal assumptions and an appropriate design are needed for a total causal interpretation.