Cumulative Effects for a Constant Marker Trajectory

Why a Constant Marker Still Has a Cumulative Effect

A marker can remain unchanged over time and still make a cumulative association with the event hazard change over time. The reason is that cumulative exposure adds the marker’s level over the entire elapsed interval: a constant positive level contributes a little more exposure with every additional unit of time. In a hazard model that uses this accumulated exposure, the marker’s hazard contribution therefore depends on both its level and how long it has persisted.

This tutorial considers the specific case in which an individual’s latent marker trajectory is constant over the interval of interest. The constant-trajectory assumption makes the integral straightforward, but its interpretation matters: the cumulative term is not simply the marker’s current value, and its hazard contribution is not a statement that the marker itself has increased.

Model and Constant-Trajectory Formula

Definition: Let $m_i(s)$ be individual $i$’s latent marker value at time $s$, and let $t=0$ be the chosen origin for accumulation. The cumulative marker exposure at time $t$ is $$ A_i(t)=\int_0^t m_i(s)\,ds. $$ If the latent marker is constant over this interval, so that $m_i(s)=c_i$ for every $s$ from $0$ to $t$, then $$ A_i(t)=c_i t. $$

Here, $c_i$ is the individual’s constant latent marker level, not necessarily a single value shared by all individuals. The exposure has units of marker-units multiplied by time-units. For example, a marker measured in milligrams per liter and time measured in years gives cumulative exposure in milligram-years per liter.

For specificity, consider a proportional-hazards model conditional on the latent marker trajectory and covariates $w_i$:

h_i(t\mid c_i,w_i) = h_0(t)\exp\left\{\gamma^\top w_i+\alpha A_i(t)\right\} = h_0(t)\exp\left\{\gamma^\top w_i+\alpha c_i t\right\}.

In this expression, $h_0(t)$ is the baseline hazard, $\gamma$ is the coefficient vector for the covariates, and $\alpha$ is the cumulative-association parameter. The model specifies a hazard conditional on the latent level and covariates; it does not substitute a noisy marker measurement for $c_i$. The proportional-hazards form also assumes that, conditional on these quantities, the cumulative-marker term enters the log hazard additively as written.

The term $\alpha A_i(t)=\alpha c_i t$ is the cumulative marker’s contribution to the log hazard. Thus the hazard multiplier attributable to this term is $\exp\{\alpha c_i t\}$. If $\alpha>0$ and $c_i>0$, that multiplier grows with elapsed time. If $\alpha<0$ and $c_i>0$, it declines with elapsed time. These statements concern the model’s marker contribution; the full hazard may also change through $h_0(t)$ and other covariates.

Derivation and Hazard-Ratio Interpretation

The constant-trajectory result follows directly from integration. Since $c_i$ does not depend on $s$, it can be taken outside the integral:

A_i(t)=\int_0^t c_i\,ds =c_i\int_0^t 1\,ds =c_i(t-0) =c_i t.

Equivalently, dividing the interval into $K$ equal subintervals of width $\Delta s=t/K$, each contributes approximately $c_i\Delta s$. Summing gives $Kc_i(t/K)=c_i t$; because the marker is constant, this sum equals the integral for any such partition.

At the same time $t$, compare two individuals with constant latent levels $c_1$ and $c_0$, holding their covariates equal. Their hazard ratio is

\frac{h_1(t)}{h_0^{*}(t)} = \exp\left\{\alpha t(c_1-c_0)\right\}.

In this ratio, $h_0^{*}(t)$ denotes the hazard for the comparison individual, not the baseline hazard function $h_0(t)$. The ratio follows by cancelling the common baseline hazard and covariate term. It compares the model-implied hazards at a specified time; it does not compare event probabilities, and it does not imply that changing a marker causes a change in hazard.

The elapsed time is essential. If $c_1-c_0$ and $\alpha$ are fixed, the log hazard ratio is proportional to $t$, while the hazard ratio itself is exponential in $t$. At $t=0$, the accumulated exposure is zero for everyone, so the cumulative term contributes no between-person contrast at that instant.

Worked Example: Calculating Exposure and Its Hazard Multiplier

Suppose a hypothetical individual has a constant latent marker level $c=0.8$ marker-units over three years. Let the cumulative-association parameter be $\alpha=0.25$ per marker-year. Calculate cumulative exposure and the hazard multiplier from the cumulative term at year 3.

The constant-trajectory formula gives

A(3)=c(3)=(0.8\ \text{marker-units})(3\ \text{years}) =2.4\ \text{marker-years}.

The cumulative term in the log hazard is

\alpha A(3) = (0.25\ \text{per marker-year})(2.4\ \text{marker-years}) =0.60.

Therefore, the multiplier contributed by this term is

\exp(0.60)\approx 1.822.

Conditional on the model and other terms, this individual’s hazard at year 3 is multiplied by about 1.82 relative to the same hazard with the cumulative term set to zero. This is a model-component comparison, not a claim that the individual’s overall hazard is 1.82 times a hazard at year 0.

Worked Example: Comparing Two Constant Marker Levels

In a hypothetical model, two individuals have the same covariates and are evaluated at year 4. Their constant latent marker levels are $c_1=1.1$ and $c_0=0.7$ marker-units, and $\alpha=0.20$ per marker-year. Calculate and interpret their hazard ratio at year 4.

The difference in cumulative exposures is

A_1(4)-A_0(4) = (1.1)(4)-(0.7)(4) = 4.4-2.8 = 1.6\ \text{marker-years}.

Using the same-time hazard-ratio formula,

\frac{h_1(4)}{h_0^{*}(4)} = \exp\{\alpha(4)(c_1-c_0)\} = \exp\{(0.20)(4)(1.1-0.7)\} = \exp(0.32) \approx 1.377.

At year 4, the model-implied hazard for the individual with the higher constant marker level is about 1.38 times the hazard for the individual with the lower level, conditional on equal covariates and the stated model. The contrast depends on elapsed time: at year 2 the corresponding ratio would be $\exp\{(0.20)(2)(0.4)\}=\exp(0.16)\approx1.174$.

Worked Example: How the Contribution Accumulates Over Time

Consider a hypothetical constant latent marker level of $c=1.2$ marker-units and $\alpha=0.15$ per marker-year. Calculate the cumulative exposure and cumulative-term multiplier at years 2 and 5. Explain what increases and what remains constant.

At year 2,

A(2)=(1.2)(2)=2.4\ \text{marker-years}, \qquad \exp\{\alpha A(2)\} = \exp\{(0.15)(2.4)\} = \exp(0.36) \approx 1.433.

At year 5,

A(5)=(1.2)(5)=6.0\ \text{marker-years}, \qquad \exp\{\alpha A(5)\} = \exp\{(0.15)(6.0)\} = \exp(0.90) \approx 2.460.

The marker level remains $1.2$ at both times, but cumulative exposure rises from $2.4$ to $6.0$ marker-years, and the cumulative-term multiplier rises accordingly. These multipliers describe the cumulative association component at each time. They are not, by themselves, a hazard ratio comparing year 5 with year 2, since the baseline hazard and other time-dependent features may differ between those times.

Assumptions, Diagnostics, and Alternatives

The calculation is simple, but using it in an analysis requires more than arithmetic. The constant-trajectory specification assumes that the latent marker level does not change over the accumulation interval. It also requires a clearly defined time origin and a consistent time scale: the duration $t$ in $c_i t$ must be measured in the units used to define $\alpha$. The hazard model assumes that the cumulative exposure enters through the specified log-linear term and that the comparison is made conditional on the modeled covariates and latent quantity.

In a joint model, repeated marker observations are typically noisy measurements of an underlying trajectory. A flat sequence of observed values does not establish that the latent trajectory is exactly constant, just as small observed fluctuations do not necessarily establish a changing latent trajectory. Diagnostics should therefore examine whether the fitted longitudinal component supports a stable level over the relevant interval. Useful checks include plotting observed measurements and fitted latent trajectories against time, assessing residual patterns for systematic upward or downward trends, and checking whether the constant specification leaves time-related structure unexplained.

If the trajectory is not plausibly constant, using $c_i t$ can misstate accumulated exposure. In particular, systematic changes in the latent marker can make the true integral differ from the rectangle formed by its initial or representative level times elapsed time. The linear-trajectory case is a distinct calculation and is not covered here.

A current-value association uses the marker level at time $t$, whereas the cumulative association uses the integral of marker values over the interval. For a constant trajectory, the current value stays at $c_i$, but cumulative exposure grows as $c_i t$. An average exposure over the interval, $A_i(t)/t$ for $t>0$, stays at $c_i$ under this assumption; it therefore does not represent the same time-growing quantity as the cumulative total. These alternatives answer different modeling questions rather than being interchangeable versions of one formula.

Common pitfall: It is incorrect to say that a constant marker has no cumulative effect because its level does not change. The marker’s value stays fixed, but its integral grows with the duration of exposure: $A_i(t)=c_i t$. The correction is to distinguish the marker trajectory from its accumulated area over time.

A second pitfall is to interpret $\exp\{\alpha c_i t\}$ as the whole hazard or as a comparison across different times. It is only the multiplier for the cumulative term. To form a hazard ratio between individuals, compare them at the same time and account for the other model terms; to compare hazards across times also requires the baseline hazard and any other time-varying components.

The constant-marker formula is useful as a transparent benchmark: it shows exactly why duration matters and provides a check on units and implementation. Its limitation is equally clear. It cannot represent changing latent marker levels, and it does not remove uncertainty in estimating those levels or in estimating $\alpha$. In applications, conclusions should be based on the fitted joint model and its uncertainty, not on a single plug-in calculation treated as known truth.

Check Your Understanding

Answer each question, then compare with the solutions below.

A constant latent marker level is $0.6$ marker-units for 5 years. What is cumulative exposure?
With $\alpha=0.10$ per marker-year, what is the cumulative-term multiplier for $c=0.6$ at year 5?
Two individuals have constant levels $0.9$ and $0.5$, equal covariates, and are compared at year 3. If $\alpha=0.25$ per marker-year, calculate the hazard ratio for the higher-level individual.
Why can the cumulative-term multiplier at year 5 not generally be interpreted as the hazard ratio comparing year 5 with year 2?
Name one diagnostic sign that would make the constant latent trajectory assumption questionable.

Solutions. (1) $A(5)=ct=(0.6)(5)=3.0$ marker-years. (2) The log-hazard contribution is $(0.10)(3.0)=0.30$, so the multiplier is $\exp(0.30)\approx1.350$. (3) The same-time ratio is $\exp\{(0.25)(3)(0.9-0.5)\}=\exp(0.30)\approx1.350$; conditional on equal covariates and the model, the higher-level individual’s hazard is about 1.35 times the other’s at year 3. (4) The multiplier applies only to the cumulative term. Comparing hazards at different times also involves the baseline hazard and potentially other time-dependent components. (5) For example, a fitted latent trajectory with a clear sustained upward or downward trend, or residuals with systematic time patterns, would call the constant-trajectory assumption into question.