ALT Trajectories and Hepatitis-Related Progression

Why ALT and Cirrhosis Are Considered Together

In chronic viral hepatitis, alanine aminotransferase (ALT) is measured repeatedly, often before and during antiviral treatment. Its values may fluctuate: treatment can reduce viral activity and ALT, while biological variation and measurement error can produce short-term rises and falls. Cirrhosis, in contrast, is a longer-term clinical outcome that may occur after years of follow-up. A single ALT measurement can miss the pattern relevant to that outcome, and an analysis of recorded ALT values alone does not directly describe time to cirrhosis.

A joint model links a subject’s underlying, or latent, ALT trajectory to the hazard of cirrhosis. It separates the estimated trajectory from noisy measurements and can account for the fact that marker observations stop when the event occurs. The link describes an association under model assumptions; it does not by itself show that lowering ALT would prevent cirrhosis.

Modeling ALT on a Log Scale

Let \(T_i\) be the time from study entry to cirrhosis for subject \(i\), and let \(C_i\) be the censoring time. The observed follow-up time is \(U_i=\min(T_i,C_i)\), and \(\delta_i=1\) if cirrhosis is observed and \(0\) if follow-up is censored. Let \(Y_{ij}\) be the log of the ALT-to-upper-limit-of-normal ratio at visit \(j\), at time \(t_{ij}\). Thus, \(Y_{ij}=\log(\mathrm{ALT}_{ij}/\mathrm{ULN})\), a dimensionless quantity. A value of zero corresponds to ALT equal to the upper limit of normal.

One useful illustrative longitudinal model includes a subject-specific intercept and slope and a treatment indicator \(D_i(t)\), which equals one while the subject is on the treatment of interest:

Y_{ij}=m_i(t_{ij})+\varepsilon_{ij}, \qquad m_i(t)=\beta_0+\beta_1t+\beta_2D_i(t)+b_{0i}+b_{1i}t.

Here \(m_i(t)\) is the latent mean log ALT ratio at time \(t\); \(\beta_0,\beta_1,\beta_2\) are fixed effects; \(b_{0i}\) and \(b_{1i}\) are subject-specific random effects; and \(\varepsilon_{ij}\) is measurement-level residual variation. For this model, assume \((b_{0i},b_{1i})\) has a mean-zero bivariate normal distribution with covariance matrix \(D\), and that the residuals are independent \(N(0,\sigma^2)\), independent of the random effects. These are modeling assumptions, not facts guaranteed by the data. More complex models may be needed if residuals are correlated, treatment effects change over time, or trajectories are nonlinear.

The treatment term represents a modeled difference in log ALT while treatment is recorded, conditional on the other terms. It does not automatically estimate a causal treatment effect. Treatment initiation may depend on prior ALT, disease severity, or other prognostic factors. A causal interpretation requires an appropriate treatment-assignment strategy and assumptions beyond this trajectory model.

Definition: The exponentiated latent trajectory, \(\exp\{m_i(t)\}\), is the geometric-mean ALT-to-ULN ratio implied by the mean log trajectory. It is not generally the arithmetic expected ALT-to-ULN ratio. Under the stated Gaussian residual model, conditional on the random effects and covariates, \(E\{\mathrm{ALT}/\mathrm{ULN}\mid b_i,D_i(t)\}=\exp\{m_i(t)+\sigma^2/2\}\).

The distinction matters because exponentiation is nonlinear. In general, the exponential of an expected log value is not the expected value on the original scale. If the scientific target is an arithmetic mean ALT ratio, the residual distribution and its variance must be incorporated rather than simply exponentiating \(m_i(t)\).

Worked Example: A Treatment-Related ALT Change

Suppose a hypothetical fitted model has \(\beta_0=0.40\), \(\beta_1=0.12\) per year, and \(\beta_2=-0.50\). For a subject with \(b_{0i}=b_{1i}=0\), consider the trajectory before treatment at \(t=0\), and after treatment begins at \(t=2\) years.

Before treatment, \(D_i(0)=0\), so \(m_i(0)=0.40+0.12(0)-0.50(0)=0.40\). The exponentiated latent trajectory is \(\exp(0.40)=1.492\), or about 1.49 times the ULN on the geometric-mean ratio scale.

At two years on treatment, \(D_i(2)=1\), so \(m_i(2)=0.40+0.12(2)-0.50(1)=0.14\). Thus \(\exp(0.14)=1.150\), about 1.15 times the ULN on the exponentiated latent-trajectory scale. This comparison does not say that the arithmetic expected ALT ratio is 1.15. For example, if the fitted residual variance were \(\sigma^2=0.09\), the conditional arithmetic expectation at two years would instead be \(\exp(0.14+0.09/2)=\exp(0.185)=1.203\). Neither calculation establishes a causal treatment effect.

Linking the Trajectory to Cirrhosis

A shared-trajectory model relates the latent ALT path to the instantaneous hazard of cirrhosis. One specification is

h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\mathsf{T}Z_i(t)+\alpha m_i(t)\}.

Here \(h_0(t)\) is the baseline hazard, \(Z_i(t)\) contains measured covariates (which may include treatment or other time-varying covariates), \(\gamma\) is their coefficient vector, and \(\alpha\) is the association parameter linking the current latent log ALT ratio to the hazard. The hazard is conditional on the random effects and covariates. The proportional-hazards form assumes that the ALT contribution to the log hazard is \(\alpha m_i(t)\); it does not allow that contribution to vary arbitrarily with time unless the model is extended.

For two otherwise comparable subjects at the same time, with latent log trajectories differing by \(\Delta m\), the conditional hazard ratio is \(\exp(\alpha\Delta m)\). In particular, a one-unit increase in \(m_i(t)\)—a multiplicative \(e\)-fold difference in the underlying ALT-to-ULN ratio—corresponds to a conditional hazard ratio of \(\exp(\alpha)\). This is an association conditional on the model covariates and random effects, not a comparison of randomly assigned ALT levels.

Worked Example: Interpreting the Shared-Trajectory Association

Suppose, hypothetically, that a fitted model estimates \(\alpha=1.30\). At a given follow-up time, compare two otherwise equal subjects whose latent log ALT ratios are \(0.60\) and \(0.10\). Their difference is \(\Delta m=0.60-0.10=0.50\).

The conditional hazard ratio is \(\exp(1.30\times0.50)=\exp(0.65)=1.916\), or about 1.92. Under the model, the subject with the higher latent log ALT ratio has about 1.92 times the instantaneous cirrhosis hazard at that time, conditional on the specified covariates and random effects. This is not a statement that the probability of cirrhosis is 1.92 times as large, and it does not establish that ALT itself causes progression.

Likelihood and Conditional Survival

Let \(S_i(t\mid b_i)\) denote the conditional probability of remaining free of cirrhosis through time \(t\). Since the hazard is the instantaneous event rate among subjects still event-free, conditional survival follows by integrating that hazard:

S_i(t\mid b_i)= \exp\left\{-\int_0^t h_i(u\mid b_i)\,du\right\}.

For an observed event, the conditional event-time contribution is \(h_i(U_i\mid b_i)S_i(U_i\mid b_i)\); for censoring, it is \(S_i(U_i\mid b_i)\). Both cases can be written as \(h_i(U_i\mid b_i)^{\delta_i}S_i(U_i\mid b_i)\). The joint likelihood also includes the longitudinal measurement densities. Because the random effects are unobserved, the subject-level contribution integrates over their distribution:

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

This expression shows why a joint model is not equivalent to plugging a noisy observed ALT value into a survival model. The event and measurement data jointly inform the latent trajectory and its association with the event process. Inference typically requires numerical integration or other computational methods for the random effects.

Worked Example: Conditional Event Probability With a Constant Trajectory

For transparent arithmetic, suppose a hypothetical subject has a constant latent trajectory \(m_i(t)=0.20\) over the next two years, no additional covariate contribution, and an exponential baseline hazard \(h_0(t)=0.04\) per year. Let \(\alpha=0.80\). Under these assumptions the conditional hazard is constant:

h_i(t\mid b_i)=0.04\exp\{0.80(0.20)\} =0.04\exp(0.16)=0.0469404348\ \text{per year}.

The conditional probability of cirrhosis within two years is \(1-\exp\{-2(0.0469404348)\}=1-\exp(-0.0938808696)\approx0.08961\), or about 8.96%. This is a model-based probability conditional on the stated trajectory and assumptions. If the trajectory varied over time, the hazard would also vary and the integral of the time-varying hazard, not a single hazard multiplied by two years, would be required.

Assumptions, Diagnostics, and Alternatives

Interpretation depends on how the data were collected and how the model was specified. In particular, the analysis should address the following conditions.

Longitudinal model adequacy. The mean trajectory, treatment timing, random-effects distribution, and residual structure should be plausible. Plot observed log ALT values and fitted subject trajectories; inspect residual patterns against time and fitted values, and check for curvature or residual autocorrelation.
Event model adequacy. The baseline hazard and proportional-hazards link must reasonably describe the event data. Examine whether the ALT association appears to vary with follow-up time and whether important prognostic covariates are omitted.
Censoring and observation process. Censoring should be independent of the unobserved event time conditional on modeled information. The visit process should be noninformative given the model, or explicitly modeled if visit frequency depends on evolving health or ALT. More frequent testing among worsening patients can otherwise distort the apparent trajectory.
Correct event and measurement definitions. The time origin, cirrhosis diagnosis rule, units, and ULN should be defined consistently. Treatment changes and ALT measurements must be correctly ordered in time. An ALT fluctuation may reflect inflammation or measurement variation and need not represent a contemporaneous change in fibrosis.

A common alternative is a time-dependent survival model using the most recent observed ALT. It is simpler, but a noisy last measurement may not represent the underlying trajectory, and visit timing can affect which value is used. A two-stage analysis first estimates individual trajectories and then uses the estimates in a survival model; it can be easier to implement, but treating estimated trajectories as known generally fails to carry their uncertainty into the event analysis. Joint modeling estimates the longitudinal and event components together, at the cost of stronger assumptions and more demanding computation.

The shared-current-value model is not the only possible joint model. The event process could be associated with a slope, cumulative exposure, or another feature of the latent trajectory. The choice should reflect the scientific question and be prespecified where possible; choosing a link after inspecting many options risks overstating evidence. If death can occur before cirrhosis and prevents its observation, treating that death as ordinary independent censoring may be inappropriate; the event structure must match the question and data.

Common Pitfalls and Interpretation

A frequent mistake is to exponentiate a fitted mean log ALT and label the result the arithmetic expected ALT ratio. That is not generally correct: with Gaussian log-scale residual variance \(\sigma^2\), the arithmetic expectation includes the factor \(\exp(\sigma^2/2)\). Another mistake is to call a negative treatment coefficient proof that treatment prevents cirrhosis. In the longitudinal model it describes a conditional difference in the log ALT trajectory, while treatment assignment may be related to disease severity. Likewise, \(\alpha\) describes the modeled association between latent ALT and hazard; it does not establish a biological mechanism or causal effect.

Key takeaway: The joint model distinguishes noisy serial ALT observations from a latent trajectory and links that trajectory to cirrhosis hazard. Interpret exponentiated log-scale quantities on the correct scale, condition conclusions on the model assumptions, and avoid causal claims from association alone.

Check Your Understanding

Try each question, then check the concise solutions below.

If \(m_i(t)=0.30\), what is the exponentiated latent ALT ratio, and what does it represent?
If \(\alpha=0.70\) and two subjects differ by \(\Delta m=0.40\), calculate their conditional hazard ratio.
Under Gaussian log-scale residuals with \(\sigma^2=0.16\), what is the arithmetic expected ALT ratio when \(m_i(t)=0.20\)?
Why can using only the most recent observed ALT produce a different analysis from a shared-trajectory joint model?

Solutions. (1) \(\exp(0.30)=1.350\), the exponentiated latent trajectory, or geometric-mean ALT-to-ULN ratio; it is not automatically the arithmetic expected ratio. (2) \(\exp(0.70\times0.40)=\exp(0.28)=1.323\), a conditional hazard ratio of about 1.32. (3) \(\exp(0.20+0.16/2)=\exp(0.28)=1.323\). (4) The last observation is noisy and depends on visit timing; the joint model estimates a latent trajectory, links it to the event process, and accounts for uncertainty in that trajectory under its assumptions.