Tumor Volume and Progression in Metastatic Melanoma

Why Link Tumor Volume to Progression or Death?

In metastatic melanoma, tumor volume may be assessed repeatedly during treatment, while the clinically important event is often disease progression or death. These data are connected: a patient's underlying tumor burden may be associated with their event risk, and worsening disease can also affect whether later volume measurements are available. An analysis that treats each measurement as independent, or uses only the last recorded volume, can miss this connection.

A joint model represents the repeated measurements and the event time together. Its central feature here is a link between the patient's latent tumor-volume trajectory and the instantaneous hazard of progression or death. The trajectory is latent because individual measurements vary around it through measurement error.

Definition: Let \(T_i\) be patient \(i\)'s time from treatment start to progression or death, whichever occurs first. Let \(V_i(t)\) be the patient's underlying tumor volume at month \(t\). We model repeated measurements of a transformed volume and link its latent mean trajectory \(m_i(t)\) to the hazard of \(T_i\). The association parameter describes how the hazard changes with the latent tumor-volume level, conditional on the model's other terms.

A Joint Model for Tumor Volume and Event Time

For positive tumor volumes measured in cubic millimeters, let \(V_{ij}^{\mathrm{obs}}\) denote the observed volume at time \(t_{ij}\), and define the dimensionless transformed measurement \(Y_{ij}=\ln\{V_{ij}^{\mathrm{obs}}/(100\ \mathrm{mm}^3)\}\), where \(t_{ij}\) is the month of the \(j\)th measurement for patient \(i\). The reference volume makes the logarithm dimensionless. Under the measurement model, \(E(Y_{ij}\mid b_i)=m_i(t_{ij})\). A useful linear mixed model is

Y_{ij}=m_i(t_{ij})+\epsilon_{ij},\qquad m_i(t)=\beta_0+\beta_1t+\beta_2A_i+\beta_3A_it+b_{0i}+b_{1i}t.

Here \(A_i=1\) denotes the treatment group of interest and \(A_i=0\) the reference group; \(b_{0i}\) and \(b_{1i}\) allow patients to have different starting levels and slopes. A standard assumption is that \((b_{0i},b_{1i})^\mathsf{T}\sim N_2(0,D)\), and that measurement errors \(\epsilon_{ij}\sim N(0,\sigma^2)\) are independent of one another and of the random effects. The model can be extended if the trajectory is not adequately described by a straight line.

Let \(U_i=\min(T_i,C_i)\) be the observed time, where \(C_i\) is the censoring time, and let \(\Delta_i=1\) if progression or death was observed and \(0\) if the observation was censored. One current-value association model is

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

The baseline hazard \(h_0(t)\) describes the hazard over time for a patient at the reference covariate values and with \(m_i(t)=0\). \(X_i\) denotes additional baseline covariates, \(\gamma\) is the treatment coefficient conditional on the other model terms, and \(\alpha\) is the tumor-volume association parameter. The hazard has units of events per month if time is measured in months. Because \(m_i(t)\) is a log volume relative to the reference volume, \(\alpha\) is the change in log hazard per one-unit increase in that log-relative volume.

Model assumptions: The longitudinal mean and error distributions are adequately specified; the random-effects distribution captures relevant between-patient variation; and the event hazard is correctly linked to the latent current value under the chosen baseline-hazard model. Censoring must be independent of the event process conditional on the included information. The measurement schedule should be noninformative after conditioning on the modeled history and covariates, or its informativeness must be addressed. These assumptions are substantive, not guaranteed by fitting the model.

The likelihood makes the joint structure explicit. We assume that, conditional on the random effects and covariates, the longitudinal measurements and event process are independent. Given the random effects, the longitudinal measurements contribute their normal densities and the event-time record contributes a hazard-survival term. These contributions are multiplied and then averaged over the random-effects distribution:

L_i=\int \left\{\prod_j f_Y\bigl(Y_{ij}\mid b_i\bigr)\right\} h_i(U_i\mid b_i)^{\Delta_i} \exp\left\{-\int_0^{U_i}h_i(s\mid b_i)\,ds\right\} \phi(b_i;D)\,db_i.

This integration matters: the observed event time informs the plausible latent trajectory, and the repeated volume measurements inform the trajectory used to describe the event hazard. Simply inserting one noisy observed volume into a survival model is not the same procedure.

Interpreting the Fitted Association

At two times for the same treatment and baseline covariates, a difference \(d\) in latent log-relative tumor volume corresponds to a hazard ratio \(\exp(\alpha d)\), provided the baseline hazard and other terms are equal. In particular, a one-unit increase in \(m_i(t)\) corresponds to multiplying the volume by \(e\), about 2.718, relative to the reference volume. This is not a one-cubic-millimeter increase.

The association is conditional on the model and the latent trajectory. It does not by itself establish that changing tumor volume would cause a change in event risk. Also, \(\gamma\) compares treatment groups at the same modeled tumor-volume level; it is not automatically the overall treatment effect, because treatment may also be associated with different volume trajectories.

Worked Example: Hazard Ratio for a Volume Difference

Suppose a fitted model gives \(\hat{\alpha}=0.70\). Compare two otherwise similar patients at the same treatment status and time, where one patient's latent tumor volume is 20% higher than the other's. Find the model-implied hazard ratio.

A 20% increase in volume corresponds to a difference in the log-relative volume of \(\ln(1.20)=0.1823\). Therefore,

\widehat{\mathrm{HR}} =\exp\{\hat{\alpha}\ln(1.20)\} =\exp(0.70\times0.1823) =\exp(0.1276) \approx1.136.

Under this fitted model, the patient with 20% higher latent tumor volume has an estimated instantaneous progression-or-death hazard about 1.136 times as high, or 13.6% higher, at that time, conditional on the stated comparison. This is a hazard comparison, not a comparison of the probability of an event by a specified month.

Following the Association Across Treatment Follow-Up

A current-value association gives a time-specific hazard contribution: as \(m_i(t)\) changes, so does the factor \(\exp\{\alpha m_i(t)\}\). To calculate a survival probability, however, one must accumulate the hazard over the whole interval. For a patient with specified covariates and trajectory, conditional survival through month \(t\) is

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

The expression is conditional on the patient's random effects. A population-level survival prediction generally averages conditional survival over the random-effects distribution; it is not obtained simply by substituting an average trajectory into the conditional formula.

Worked Example: Conditional Survival Through Month 6

For illustration, suppose a hypothetical fitted model has a constant baseline hazard \(h_0(t)=0.04\) per month, \(\gamma=-0.30\), and \(\alpha=0.70\). Consider a patient with \(A_i=1\), no other covariate contribution, and latent trajectory \(m_i(t)=0.20+0.08t\). Calculate the model-implied conditional survival through month 6.

The hazard at month 6 is

h_i(6)=0.04\exp\{-0.30+0.70(0.20+0.08\times6)\} =0.04\exp(0.176) \approx0.04770\text{ per month}.

For survival through the interval, integrate rather than multiplying the month-6 hazard by six. Since the exponent in the hazard is \(-0.16+0.056t\),

\begin{aligned} H_i(6) &=\int_0^6 0.04\exp(-0.16+0.056s)\,ds\\ &=0.04\exp(-0.16)\frac{\exp(0.056\times6)-1}{0.056}\\ &\approx0.243067. \end{aligned}

Thus \(S_i(6)=\exp(-0.243067)\approx0.7842\). Under these hypothetical parameters and this specified trajectory, the model gives about 0.784 conditional probability of remaining free of progression or death through month 6. This is a conditional model-based prediction, not an empirical estimate from a real melanoma study.

Worked Example: Comparing Trajectories at Month 8

Suppose two hypothetical patients have the same treatment status and baseline covariates. Their fitted latent trajectories are \(m_1(t)=0.10+0.02t\) and \(m_2(t)=0.10+0.07t\), with \(\hat{\alpha}=0.70\). Compare their hazards at month 8.

At month 8, \(m_1(8)=0.26\) and \(m_2(8)=0.66\), so \(m_1(8)-m_2(8)=-0.40\). Their latent volume ratio is \(\exp(-0.40)\approx0.6703\). The hazard ratio for patient 1 relative to patient 2 is

\frac{h_1(8)}{h_2(8)} =\exp\{0.70(0.26-0.66)\} =\exp(-0.28) \approx0.7558.

At month 8, the model therefore assigns patient 1 about 0.756 times the hazard of patient 2, conditional on the shared treatment and covariates. The comparison is time-specific: at month 0 the two trajectories coincide, so this association term gives a hazard ratio of 1 at that time.

Estimation, Diagnostics, and Alternatives

In practice, the joint likelihood is usually maximized numerically, with integration over the random effects handled by an appropriate numerical method. The association estimate and its uncertainty should be interpreted only after checking whether the fitted model is plausible for these data. Useful checks include:

Plot observed log-relative volumes and fitted trajectories by patient and treatment group. Systematic curvature or treatment-specific patterns left in residuals suggest that a linear time trend may be inadequate.
Inspect residual spread over follow-up and check whether unusually large errors or outlying measurements dominate the fit. A normal, constant-variance error assumption may fail.
Check convergence warnings, sensitivity to starting values, and whether the estimated random-effects covariance matrix is plausible. Near-singular covariance estimates can make patient-specific trajectories and the association unstable.
Assess whether the chosen baseline hazard and proportional-effects form describe the event-time pattern. A time-varying association or a different trajectory shape may be needed if the fitted model systematically misses the observed event pattern.
Compare the measurement schedule and reasons for missed assessments across patients. If deteriorating patients are measured more often, or measurements stop for reasons related to unobserved health status, the assumed visit or censoring process may be inadequate.

The current-value link is one reasonable scientific choice, not a universal law. It assumes the instantaneous hazard depends on the latent volume at the current time. If the scientific question instead concerns rate of change or accumulated burden, a slope or cumulative-history association may be more appropriate, but those are different models with different interpretations.

Approach	What it uses	Main caution
Joint model	Repeated measurements and event time, linked through a latent trajectory	Requires assumptions about trajectory, random effects, event hazard, and observation process
Time-dependent survival model using observed volume	Recorded volume values as time-varying covariates	Measurement error and irregular measurement times can distort the association; the recorded value is not the latent trajectory
Landmark analysis	Volume measured at a chosen time among patients still at risk then	Answers a conditional prediction question from that landmark and does not use the full trajectory in the same way
Separate longitudinal and event analyses	Each outcome modeled on its own	Does not directly represent the association between the latent marker course and event risk

A joint model is particularly useful when repeated measurements are noisy and related to event risk, but it does not remove bias automatically. It can be sensitive to misspecification, and a fitted association should not be presented as a causal effect or as proof that tumor volume is a surrogate endpoint. A surrogate claim requires additional evidence and a different scientific argument.

Common Pitfall

Common mistake: “The fitted hazard ratio is 2.01 per one-unit increase in the observed tumor volume.” This is wrong when \(\alpha\) multiplies \(m_i(t)\), because \(m_i(t)\) is the latent log-relative volume, not volume in cubic millimeters. With \(\hat{\alpha}=0.70\), the hazard ratio for a one-unit increase in \(m_i(t)\) is \(\exp(0.70)\approx2.014\). A one-unit increase in \(m_i(t)\) means the underlying volume is multiplied by \(e\), not increased by one unit of physical volume. For a 20% increase, use \(\exp\{0.70\ln(1.20)\}\approx1.136\).

Why the Link Matters

In this setting, a joint model can use the repeated tumor-volume record to estimate a smoother latent trajectory and relate that trajectory to progression or death over follow-up. Reporting the transformation, time units, event definition, association scale, and conditioning terms makes the interpretation reproducible. Just as importantly, examining assumptions and alternative explanations prevents a statistical association from being mistaken for a causal treatment effect.

Check Your Understanding

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

If \(\hat{\alpha}=0.50\), what is the model-implied hazard ratio for a 10% increase in latent tumor volume, holding other terms fixed?
Why does a hazard at month 6 alone not determine survival through month 6?
At the same time and covariate values, patient A has a latent log-relative volume 0.30 lower than patient B. If \(\hat{\alpha}=0.70\), calculate the hazard ratio for A relative to B.
Name one diagnostic warning that could indicate the longitudinal model is inadequate, and explain why it matters for the association estimate.

Solutions. (1) The log-volume difference is \(\ln(1.10)=0.09531\), so the hazard ratio is \(\exp\{0.50(0.09531)\}=\exp(0.04766)\approx1.049\). (2) Survival depends on the cumulative hazard, \(\int_0^6h_i(s)\,ds\), so the hazard over the entire interval must be integrated. (3) The ratio is \(\exp\{0.70(-0.30)\}=\exp(-0.21)\approx0.8106\); the model-implied hazard for A is about 0.811 times B's. (4) For example, systematic curvature in residual plots suggests a linear trajectory may not capture the tumor-volume course. Since that fitted trajectory enters the event hazard, its misspecification can distort the estimated association.