Biomarker Slope and Recurrent-Event Intensity

Why the Biomarker Slope Can Matter

A biomarker value describes a person’s current measured state after accounting for measurement error. Its slope describes how that latent state is changing. Two people with the same current marker level may have different recurrence rates if one is improving and the other is deteriorating. A recurrent-event model can represent this distinction by including both the latent marker value and its instantaneous slope in the event intensity.

The slope is not a difference between two noisy observed measurements. It is the derivative of the underlying longitudinal trajectory. This distinction is important: noisy measurements can make raw differences unstable, while a fitted trajectory provides a model-based estimate of the underlying rate of change. As in the earlier material on current biomarker value and recurrent-event intensity, the event process is linked to the latent marker process rather than to each observed marker value as if measured without error.

Model and Conditional Interpretation

Definition: Let $N_i(t)$ count recurrent events for subject $i$ by time $t$, and let $Y_i(t)$ indicate that the subject is under observation and at risk immediately before $t$. Let $m_i(t)$ be the subject’s latent biomarker trajectory, and let $m_i'(t)$ be its instantaneous derivative with respect to time. A slope-augmented recurrent-event intensity is $$ \lambda_i(t\mid b_i)=Y_i(t)\lambda_0(t)\exp\left\{\gamma^\top w_i+\alpha_m m_i(t)+\alpha_s m_i'(t)+\kappa^\top b_i\right\}. $$ Here $\lambda_0(t)$ is a baseline intensity, $w_i$ is a vector of event-model covariates, and $b_i$ is a vector of shared random effects. The parameters $\alpha_m$ and $\alpha_s$ describe the conditional associations with current latent marker value and latent slope, respectively. The term $\kappa^\top b_i$ allows the shared random effects to contribute directly to recurrence intensity.

This is a conditional intensity: given the latent random effects and the observed history immediately before time $t$, it describes the instantaneous expected event count per unit time. The factor $Y_i(t)$ sets the intensity to zero when a subject is not at risk. The model is often specified as a multiplicative intensity model for a counting process; it does not require recurrent events to be independent after ignoring the subject’s history and latent effects.

The latent trajectory may be represented by a mixed-effects model such as

m_i(t)=x_i(t)^\top\beta+z_i(t)^\top b_i, \qquad m_i'(t)=\frac{d x_i(t)^\top\beta}{dt}+\frac{d z_i(t)^\top b_i}{dt}.

Thus the slope in the event model is obtained by differentiating the fitted latent mean trajectory, not by substituting the most recent observed marker. The derivative must exist at the time of interest. For a piecewise trajectory with a change point, the relevant derivative may differ on either side of that point.

When the other terms are held fixed, increasing the slope by $\Delta s$ changes the conditional intensity by the factor $\exp(\alpha_s\Delta s)$. A positive slope means the marker is increasing, and a negative slope means it is decreasing; whether that represents improvement or deterioration depends on the marker’s clinical meaning. The coefficient’s units are time per marker unit, so that $\alpha_s m_i'(t)$ is dimensionless. For example, if the marker is measured in mg/L and time in years, $\alpha_s$ is measured in years per mg/L.

“Conditional” matters. The interpretation compares intensities at the same time with current marker value, modeled covariates, and shared random effects held fixed. If a model’s random effects determine the slope completely, then a contrast that changes slope while keeping those same random effects fixed may not be an empirically available comparison. In that setting, the coefficient still belongs to the specified conditional model, but its interpretation relies on the model’s supported trajectory contrasts and it may be difficult to estimate separately from the random-effects contribution.

From Trajectory to Intensity

Worked Example: Calculating the Latent Slope and Intensity

Suppose a hypothetical biomarker is measured in mg/L and time is measured in years. For one subject, the fitted latent trajectory is $m_i(t)=12+0.6t-0.05t^2+b_{0i}+b_{1i}t$, with $b_{0i}=-1$ mg/L and $b_{1i}=0.1$ mg/L per year. At $t=4$ years, suppose $Y_i(t)=1$, $\lambda_0(t)=0.08$ events per year, $\gamma^\top w_i=0.20$, $\alpha_m=0.10$ per mg/L, $\alpha_s=-0.70$ years per mg/L, and $\kappa^\top b_i=-0.12$. Calculate the current latent marker, slope, and conditional intensity.

First, substitute $t=4$ into the trajectory:

m_i(4)=12+0.6(4)-0.05(4^2)-1+0.1(4) =12+2.4-0.8-1+0.4=13.0\ \text{mg/L}.

Differentiate the trajectory with respect to time:

m_i'(t)=0.6-0.10t+b_{1i}, \qquad m_i'(4)=0.6-0.10(4)+0.1=0.30\ \text{mg/L per year}.

The marker is increasing at 0.30 mg/L per year at this time. Substitute the values into the intensity model:

\begin{aligned} \lambda_i(4\mid b_i) &=0.08\exp\{0.20+0.10(13.0)-0.70(0.30)-0.12\}\\ &=0.08\exp(1.17) =0.08(3.222) \approx 0.2578\ \text{events per year}. \end{aligned}

This is a model-based conditional intensity at year 4, not a prediction that the subject will experience exactly 0.2578 events during that year. The intensity describes the instantaneous event rate at that time, conditional on the stated covariates and random effects.

Rate-Ratio Contrasts and Estimation

The conditional rate-ratio interpretation follows by dividing two intensities at the same time. If the two modeled states differ in marker value by $\Delta m$, slope by $\Delta s$, covariates by $\Delta w$, and random effects by $\Delta b$, while both are at risk, their intensity ratio is

\frac{\lambda_i^{(2)}(t\mid b_i^{(2)})} {\lambda_i^{(1)}(t\mid b_i^{(1)})} = \exp\left\{ \gamma^\top\Delta w+ \alpha_m\Delta m+ \alpha_s\Delta s+ \kappa^\top\Delta b \right\}.

The baseline intensity and at-risk indicator cancel when both states have the same $Y_i(t)=1$ and share the same baseline at time $t$. For a slope-only contrast, set the other differences to zero. The resulting ratio is $\exp(\alpha_s\Delta s)$. If $\alpha_s<0$, a more positive slope is associated with a lower conditional intensity, all else equal. If a negative slope represents deterioration, that same negative coefficient implies a higher intensity for more rapidly declining marker trajectories.

A common inferential target is $H_0:\alpha_s=0$. For an estimated coefficient $\hat{\alpha}_s$ with standard error $SE(\hat{\alpha}_s)$, a large-sample Wald statistic is $z=\hat{\alpha}_s/SE(\hat{\alpha}_s)$. This uses an approximately normal sampling distribution for the estimator under the null. In joint models, the standard error must reflect estimation of the longitudinal and event submodels and their dependence; treating the fitted latent slope as known generally understates uncertainty.

Worked Example: Interpreting a Slope Contrast

In a hypothetical joint model, the biomarker is coded so that a declining value represents deterioration. Suppose the fitted slope coefficient is $\hat{\alpha}_s=-0.50$ years per marker unit. Compare two at-risk states with the same current marker value, event covariates, and shared random effects, but with slopes of $-0.4$ and $+0.6$ marker units per year.

The slope difference from the first state to the second is $+0.6-(-0.4)=1.0$ marker unit per year. The conditional intensity ratio is

\frac{\lambda(\text{slope }+0.6)}{\lambda(\text{slope }-0.4)} =\exp\{-0.50(1.0)\} =\exp(-0.50) \approx 0.6065.

The model therefore associates the more positive slope with an intensity about 0.607 times as large, or about 39.3% lower, than the less positive slope, conditional on the stated quantities. In the reverse direction, the rate ratio is $\exp(0.50)\approx1.649$: the declining state has an intensity about 1.65 times as large in this contrast. The direction is consistent with a negative coefficient when decline indicates deterioration. This is an association specified by the model, not evidence that changing the marker slope would itself cause a change in recurrence.

The numerical contrast illustrates the coefficient’s interpretation; it does not guarantee that both slope values occur with the same random effects in a particular fitted trajectory model. That feasibility depends on the longitudinal design and trajectory specification.

Worked Example: Wald Test and Confidence Interval

Suppose a hypothetical fit gives $\hat{\alpha}_s=-0.55$ years per marker unit with $SE(\hat{\alpha}_s)=0.20$ years per marker unit. Test $H_0:\alpha_s=0$ against a two-sided alternative at $\alpha=0.05$, and interpret the estimated slope rate ratio for a 1 marker-unit-per-year increase.

The Wald statistic is

z=\frac{-0.55}{0.20}=-2.75.

Under the null and the large-sample normal approximation, the two-sided p-value is $2P(Z\leq-2.75)\approx0.0060$, where $Z\sim N(0,1)$. Since $0.0060<0.05$, the data provide evidence against a zero slope association under this model and approximation.

A 95% Wald confidence interval for $\alpha_s$ is

-0.55\pm1.96(0.20) =-0.55\pm0.392 =(-0.942,-0.158)\ \text{years per marker unit}.

For a slope increase of 1 marker unit per year, exponentiating the coefficient and interval gives an estimated conditional rate ratio of $\exp(-0.55)\approx0.577$, with 95% interval $(\exp(-0.942),\exp(-0.158))\approx(0.390,0.854)$. Under the model, that slope increase is associated with an estimated 42.3% lower intensity, with the interval corresponding to reductions of about 14.6% to 61.0%. The interval is a procedure that, under its assumptions, would cover the parameter in 95% of repeated samples; it is not a 95% probability statement about the fixed parameter.

Justification, Diagnostics, and Alternatives

The slope rate-ratio result follows directly from the log-linear intensity specification. Consider two conditional states with identical time, at-risk status, covariates, current marker value, and random effects, differing only in slope by $\Delta s$. Their log intensities differ by

\log\lambda^{(2)}(t)-\log\lambda^{(1)}(t) =\alpha_s\{m^{(2)\prime}(t)-m^{(1)\prime}(t)\} =\alpha_s\Delta s.

Exponentiating both sides gives $\lambda^{(2)}(t)/\lambda^{(1)}(t)=\exp(\alpha_s\Delta s)$. The interpretation therefore depends on the log-linear intensity form and on holding the other modeled terms fixed; it is not a general property of all recurrent-event models.

The main assumptions include a correctly specified latent trajectory and its derivative, a suitable baseline intensity and event-process structure, and an appropriate distributional model for shared random effects. The observation and risk processes must also be represented correctly. If a terminal event removes a subject from the recurrence risk set, that cessation must be handled in the at-risk process and joint model as appropriate; the slope term does not by itself resolve informative termination.

Diagnostics should examine whether the longitudinal trajectory fits the observed measurements over time, including whether its shape is plausible between measurement occasions. Check that the assumed trajectory is differentiable where the slope is used and that its derivative has a meaningful scale. For the event component, examine fit over time and by event history, and assess whether the slope association is sensitive to plausible choices of longitudinal trajectory and baseline intensity. Large standard errors, strong dependence among $\hat{\alpha}_m$, $\hat{\alpha}_s$, and random-effect association estimates, or estimates that change markedly under small model changes can signal weak separation of these effects.

A frequent failure is to calculate slopes from adjacent raw measurements and insert them into the intensity as though they were the latent slopes. Measurement error can create extreme differences, and unequal observation intervals make such differences incomparable unless divided by elapsed time. More fundamentally, plugging in a noisy estimated slope as if known ignores its uncertainty and can distort both the association estimate and its standard error. The joint model instead links the event intensity to the latent derivative while accounting for uncertainty in the longitudinal process.

The principal alternative within this family is a value-only association model, which sets $\alpha_s=0$. It asks whether current marker level is associated with recurrence intensity, but does not represent additional association with the rate of marker change. A slope-only specification omits current value and may be suitable only if that is scientifically and statistically defensible. Including both permits distinct conditional associations but can make them difficult to separate when marker value and slope are strongly related in the available data. A slope term also differs from cumulative biomarker exposure: it concerns instantaneous change, not accumulated past marker burden.

Common pitfall: A negative slope coefficient does not automatically mean that “deterioration lowers recurrence.” First determine which marker direction represents deterioration. If decline is deterioration, then a negative slope multiplied by a negative $\alpha_s$ contributes positively to the log intensity. Also verify whether the interpretation is conditional on current marker value and shared random effects.

Check Your Understanding

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

A marker is measured in mmol/L and time in months. What are the units of its derivative and of $\alpha_s$ in a log-linear intensity?
For $\alpha_s=0.3$ months per marker unit, calculate the conditional rate ratio for a slope increase of 2 marker units per month, with all other terms fixed.
A biomarker decreases as a disease worsens, and the estimated slope coefficient is negative. What association does the model imply for a more rapidly declining marker?
Why is a difference between two observed marker readings generally not interchangeable with the latent slope in the event model?

Solutions. (1) The slope has units mmol/L per month. Since $\alpha_s m_i'(t)$ must be dimensionless, $\alpha_s$ has units month per (mmol/L), equivalently months·L/mmol. (2) The ratio is $\exp(0.3\times2)=\exp(0.6)\approx1.822$. At the same current value and other modeled terms, the specified slope increase is associated with an intensity about 1.82 times as large. (3) A more rapidly declining marker has a more negative slope. Multiplying that change by a negative coefficient increases the slope contribution to the log intensity, so the model associates more rapid decline with higher conditional intensity, all else equal. (4) The observed difference includes measurement error, depends on the time between readings, and does not directly estimate the instantaneous derivative. Treating it as known also ignores uncertainty in the latent trajectory and its slope.