Weighted Marker Burden in the Hazard

Why Combine Current Marker Values?

When several biomarkers are measured repeatedly, a survival model can relate the instantaneous event rate to their latent values at the same time. A weighted sum provides a concise way to represent their joint contribution: each marker’s value is multiplied by its own association coefficient, and the resulting contributions are added. The weights determine how a particular marker profile enters the hazard, including whether a higher value raises or lowers it.

The values in this model are the markers’ underlying trajectories, not the noisy measurements recorded at individual visits. This distinction matters: a visit measurement can differ from the latent marker value because of measurement error. The joint model represents the longitudinal and survival processes together, so uncertainty about the latent trajectories is relevant to estimation and interpretation.

Model and Interpretation

Definition: For subject $i$, let $m_{ik}(t)$ be the latent current value of marker $k$ at time $t$, for $k=1,\ldots,K$. A weighted current-marker hazard is $$ \lambda_i(t\mid \mathbf{m}_i(t),\mathbf{w}_i) = \lambda_0(t)\exp\left\{ \boldsymbol{\gamma}^{\mathsf T}\mathbf{w}_i+ \sum_{k=1}^{K}\alpha_k m_{ik}(t) \right\}. $$ Here $\lambda_0(t)$ is the baseline hazard, $\mathbf{w}_i$ is a vector of baseline covariates, $\boldsymbol{\gamma}$ their coefficient vector, and $\alpha_k$ the association coefficient for marker $k$. Define the weighted marker burden $B_i(t)=\sum_{k=1}^{K}\alpha_km_{ik}(t)$. The hazard is conditional on the latent marker values and covariates.

The word burden refers to the sum on the log-hazard scale. It is not necessarily a physical quantity, and it is not a probability. If $\alpha_k>0$, then, holding the other markers and covariates fixed, a one-unit higher value of marker $k$ multiplies the hazard by $\exp(\alpha_k)$. If $\alpha_k<0$, the corresponding multiplier is below one. The units of $\alpha_k$ are the inverse units of marker $k$, so coefficients on differently measured markers should not be compared by size alone.

The sum is linear in the marker values on the log-hazard scale, not on the hazard scale. Consequently, each marker’s contribution is multiplicative in the hazard: the factors $\exp\{\alpha_km_{ik}(t)\}$ multiply to $\exp\{B_i(t)\}$. For a comparison between profiles, the most direct interpretation uses their difference in weighted burden.

Deriving a Profile Hazard Ratio

Consider two marker profiles, $A$ and $B$, at the same time $t$, for subjects with the same baseline covariates. Their baseline hazard and covariate contributions cancel in the ratio. Subtracting their log hazards gives the difference in weighted burdens:

\begin{aligned} \log\left\{\frac{\lambda_A(t)}{\lambda_B(t)}\right\} &= \sum_{k=1}^{K}\alpha_km_{Ak}(t) - \sum_{k=1}^{K}\alpha_km_{Bk}(t)\\ &= \sum_{k=1}^{K}\alpha_k\{m_{Ak}(t)-m_{Bk}(t)\}. \end{aligned}

Exponentiating both sides yields the profile hazard ratio:

\frac{\lambda_A(t)}{\lambda_B(t)} = \exp\left[ \sum_{k=1}^{K}\alpha_k\{m_{Ak}(t)-m_{Bk}(t)\} \right].

This derivation also makes clear what a coefficient comparison means: $\exp(\alpha_k d)$ is the conditional hazard ratio for changing marker $k$ by $d$ units while holding the other marker values and covariates fixed. It is not automatically a realistic comparison if the markers are strongly related and such a change is rarely observed.

Worked Example: Comparing Two Marker Profiles

Suppose a hypothetical fitted model has two markers, with $\alpha_1=0.25$ per unit of marker 1 and $\alpha_2=-0.40$ per unit of marker 2. At a specified time, profile $A$ has latent values $(4,1)$, and profile $B$ has values $(2,3)$. Compare their hazards for otherwise identical subjects.

The weighted burdens are $B_A=0.25(4)-0.40(1)=1.00-0.40=0.60$, and $B_B=0.25(2)-0.40(3)=0.50-1.20=-0.70$. Therefore the log hazard ratio is $0.60-(-0.70)=1.30$, or, equivalently, $0.25(4-2)-0.40(1-3)=0.50+0.80=1.30$.

The profile hazard ratio is $\exp(1.30)=3.669$. Under this model, at that time and conditional on the stated covariates, profile $A$ has about 3.67 times the hazard of profile $B$. This comparison is not a statement that profile $A$ has 3.67 times the event probability.

Worked Example: A Change in One Marker

Suppose, hypothetically, that $\alpha_1=0.30$ per unit of marker 1 and $\alpha_2=-0.20$ per unit of marker 2. Compare two otherwise identical marker profiles that differ only in marker 1, which is 2 units higher in the first profile.

The change in weighted burden is $0.30(2)+(-0.20)(0)=0.60$. Thus the hazard ratio is $\exp(0.60)=1.822$. Conditional on marker 2, the covariates, and the model, the profile with marker 1 higher by 2 units has an estimated hazard about 1.82 times as high at that time.

For a one-unit increase in marker 1, the corresponding multiplier is $\exp(0.30)=1.350$, or a 35.0% higher hazard. The 2-unit change does not produce a 70% increase: the model is multiplicative, so the multiplier is $\exp(0.30)^2=\exp(0.60)$.

How the Weights Shape the Burden

A coefficient’s sign and magnitude determine how a marker value changes the sum. Positive weights add to the log hazard as the marker increases; negative weights subtract from it. But a marker’s contribution must be interpreted relative to a specified reference or profile. The value of $B_i(t)$ alone depends on where zero lies for each marker. The profile hazard ratio, based on a difference in burden, is generally the clearer comparison.

A weighted sum can also conceal very different profiles. For example, with $\alpha_1=0.5$ and $\alpha_2=-0.25$, the profiles $(2,0)$ and $(3,2)$ both have burden 1: $0.5(2)-0.25(0)=1$ and $0.5(3)-0.25(2)=1$. This model assigns them the same marker contribution to the hazard, even though their marker values differ. That equality is a consequence of the model’s form, not proof that the profiles are biologically equivalent.

Worked Example: Equal Burdens From Different Profiles

Consider a hypothetical model with $\alpha_1=0.50$ per unit and $\alpha_2=-0.25$ per unit. At a given time, profile $C$ has values $(2,0)$, while profile $D$ has values $(3,2)$. The subjects have the same covariates.

For profile $C$, $B_C=0.50(2)-0.25(0)=1.00$. For profile $D$, $B_D=0.50(3)-0.25(2)=1.50-0.50=1.00$. Their burden difference is zero, so the model gives a hazard ratio of $\exp(1.00-1.00)=1$ at that time.

The model therefore predicts equal conditional hazards for these profiles, all else equal. It does not imply that the markers are interchangeable, that the profiles have equal future trajectories, or that either marker causes the event rate to change.

Assumptions, Diagnostics, and Alternatives

The interpretation above rests on specific modeling assumptions. The longitudinal submodels must provide meaningful estimates of the latent current marker values; the hazard must be adequately represented by the stated conditional form; and each $\alpha_k$ is constant over time in this specification. In particular, proportional hazards conditional on the current marker values means that the marker effects enter through fixed coefficients rather than time-varying coefficients. The model also assumes the selected markers and covariates adequately represent the conditional association being studied. For likelihood-based inference, censoring must be non-informative conditional on the modeled information, or its mechanism must be modeled appropriately.

Diagnostics should consider both the survival component and the marker panel. Examine whether estimated marker trajectories are supported by observed measurements over the event-time range, whether the fitted model reproduces relevant survival patterns, and whether association estimates are sensitive to influential subjects or modeling choices. Wide uncertainty intervals, large changes in coefficients across plausible specifications, or strong dependence among marker estimates can indicate that the data do not distinguish their separate weights well. A numerically converged fit does not by itself guarantee that the individual weights are precisely identified.

This specification differs from a model that links survival to shared random effects: the present hazard uses the current latent marker values directly, while a shared-random-effects formulation expresses the link through latent subject-specific effects. It also differs from fitting separate marker-specific current-value models, which do not estimate all marker weights jointly in the same hazard. Joint inclusion can adjust each marker’s association for the other included markers, but correlated markers may make those adjusted associations imprecise. Neither approach is universally preferable; the choice should reflect the scientific target and the support of the data.

A weighted current-value association is an association, not a causal effect. Unmeasured confounding, selection, model misspecification, or changes in treatment can affect it. The model also does not, by this equation alone, describe effects of marker rates of change or accumulated exposure. Those require different survival association structures.

Common Pitfall

Wrong reasoning: “The coefficient is 0.30, so the event probability rises by 30% for a one-unit marker increase.”
Correction: The coefficient is on the log-hazard scale. A one-unit increase multiplies the conditional hazard by $\exp(0.30)=1.350$, about a 35.0% increase in hazard. It does not directly specify a percentage-point change in event probability.
Wrong reasoning: “A positive weight proves that increasing the marker causes the event rate to rise.”
Correction: The weight describes a conditional model association. A causal interpretation requires additional design and identification assumptions beyond this hazard specification.

Key takeaway: For two otherwise comparable marker profiles, calculate the difference in their weighted current values, $\sum_k\alpha_k\{m_{Ak}(t)-m_{Bk}(t)\}$, then exponentiate it to obtain the model-implied conditional hazard ratio. Interpret that ratio in context, with attention to marker units, uncertainty, and the assumptions behind the model.

Check Your Understanding

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

With $\alpha_1=0.20$ and $\alpha_2=-0.10$, profile $A$ differs from profile $B$ by 3 units higher on marker 1 and 2 units higher on marker 2. What is the log hazard ratio and hazard ratio?
Why does a coefficient of 0.40 not mean a 40-percentage-point increase in event probability?
What does it mean to interpret $\alpha_1$ while holding the other markers fixed?
Two marker profiles have the same weighted burden. What does the model imply, and what does it not imply?

Solutions. (1) The log hazard ratio is $0.20(3)+(-0.10)(2)=0.40$; the hazard ratio is $\exp(0.40)=1.492$. (2) The coefficient is on the log-hazard scale, so the one-unit hazard multiplier is $\exp(0.40)=1.492$; it is neither an event probability nor a percentage-point change. (3) It is the model-based association for changing marker 1 while fixing the other current marker values and covariates; it need not describe a feasible change if the markers are strongly related. (4) The model assigns the profiles the same conditional marker contribution to the hazard at that time, but does not claim their biology, future trajectories, or causal consequences are equivalent.