Slopes and Recent-Trend Predictors at the Landmark

Why Use a Slope at a Landmark?

A biomarker value at a landmark summarizes a patient's current measured level, but it does not describe the direction or pace of change leading up to that time. Two patients can have the same measured value at the landmark even though one has been stable and the other has been declining rapidly. A slope predictor represents that recent rate of change and can be used, alongside other information available at the landmark, to predict subsequent event risk.

The timing restriction is essential: the slope must be estimated from measurements available on or before the landmark. The outcome is then defined among patients eligible at that landmark, such as those still event-free. A slope built using measurements taken after the landmark would use future information and would not be a valid predictor for this prediction task.

The slope has units of biomarker units per unit of time. For example, a slope of \(-0.4\) mg/L per month means that, over the fitted period, the patient's biomarker decreased at an estimated average rate of 0.4 mg/L per month. It is not the biomarker level itself, and it is not automatically a causal measure of disease progression.

Calculating a Patient-Specific Slope

Let \(Y_{ij}\) be patient \(i\)'s biomarker measurement at time \(t_{ij}\), for \(j=1,\ldots,m_i\), where all included times satisfy \(t_{ij}\leq L\). With a linear trend fitted to these measurements, the estimated slope is

where \(\bar{t}_i\) and \(\bar{Y}_i\) are the mean measurement time and mean biomarker value for patient \(i\)'s included observations. The slope can be estimated only if there are at least two observations at distinct times. Using actual times matters: a difference of one week between visits is not equivalent to a difference of one month.

It is often convenient to center time at the landmark, writing \(x_{ij}=t_{ij}-L\). Then the landmark is time zero and all included \(x_{ij}\leq 0\). Centering changes the intercept's interpretation but does not change the estimated slope: subtracting the same constant \(L\) from every time leaves all deviations from the mean unchanged.

A slope may use all available pre-landmark measurements or only a prespecified recent window, such as the last three months before \(L\). The first summarizes the average trend over a longer history; the second targets a more recent trend but uses fewer observations. The window should be selected based on the prediction question, not chosen after inspecting which choice gives the strongest association with the outcome.

Choosing the Time Window and Interpreting the Outcome Model

A slope is defined relative to the observations and time window used to estimate it. If the scientific question concerns a recent change, a recent-window slope may be more relevant than a slope averaged over a long history. If the trend is approximately linear throughout the recorded period, the longer window can provide a more stable summary. These are different predictors, not interchangeable calculations of one uniquely defined quantity.

For a time-to-event outcome, let \(T_i\) be the event time and consider patients who remain event-free at \(L\). Their residual event time is \(T_i-L\). A landmark Cox model can include the estimated slope as a predictor:

Here \(u\) is time since the landmark, \(h_{0,L}(u)\) is the baseline hazard for the landmark risk set, and \(Z_i(L)\) denotes other covariates known at \(L\). The model assumes proportional hazards conditional on these predictors: a patient's relative hazard, compared with another patient's, does not vary with \(u\) when their predictors are fixed. A one-unit increase in slope is associated with a hazard ratio of \(\exp(\gamma)\), holding the other modeled predictors fixed. The association is conditional and predictive; it does not by itself establish that changing the biomarker slope would change event risk.

The outcome-model assumptions are separate from the calculation of the slope. For the Cox model's usual interpretation and inference, the proportional-hazards form and predictor specification should be adequate, and censoring should be independent of the event process conditional on modeled information. The landmark risk set must also be defined correctly: patients who experienced the event before \(L\) are not at risk for a subsequent event after \(L\).

Why the Slope Can Add Information—and Where It Can Fail

Last-value and change-from-baseline predictors, covered in the earlier material on landmark predictors, answer different questions from a slope. The last value describes the measured level at the landmark, while change from baseline describes the difference between two selected measurements. A slope uses the pattern across a set of measurements and expresses change per unit time. It can distinguish patients with the same last value but different preceding trajectories, though it may also be redundant with level or change summaries.

The linear slope is most straightforward when the biomarker follows an approximately linear trend over the chosen window and observations are available at distinct times. A sharp change, a curved trajectory, or an isolated extreme value can make one constant slope a poor summary. A slope based on only two observations is simply their difference divided by elapsed time; it is calculable, but it cannot reveal whether the biomarker changed smoothly between those visits.

For a standard linear regression interpretation of the within-patient trend, one may assume that the expected biomarker value changes linearly with time over the selected window and that deviations from that mean trend have mean zero. Usual regression-based uncertainty calculations additionally rely on assumptions about the residual distribution and variance. Even when the goal is prediction rather than inference about a patient's true biological rate, sparse measurements can make the estimated predictor unstable. The slope is an estimate, not a directly observed patient characteristic.

A useful workflow is to plot each patient's pre-landmark measurements against actual time, check whether a straight-line summary is reasonable, and inspect whether a single observation dominates the fitted trend. At the outcome-model stage, assess whether the slope's relationship to risk is plausibly represented by the specified term and evaluate model fit and prediction performance using appropriate validation. Apparent improvement in the same data used to fit the model is not, by itself, evidence of reliable predictive gain.

Common Mistakes and Limitations

• Using future information. Including a measurement after \(L\) leaks information unavailable at prediction time. Restrict slope estimation to measurements at or before the landmark.
• Regressing on visit number instead of time. If measurement intervals differ, a per-visit slope does not have a consistent rate interpretation. Use actual elapsed times and report the resulting time units.
• Changing the window after seeing the outcome results. This can make the reported association overly optimistic. Prespecify the window or compare alternatives with suitable validation.
• Reading a slope as a guaranteed future trend. It summarizes the fitted pre-landmark period; it does not assert that the biomarker will continue changing at the same rate after \(L\).
• Interpreting a hazard ratio as a risk ratio. A Cox-model hazard ratio compares instantaneous event rates among those still event-free at a given time. It is not generally the ratio of cumulative event probabilities.

A slope alone also does not solve every prediction problem. It may add little when the last value and recent change already summarize the relevant information, and its usefulness depends on adequate pre-landmark measurements. Prediction based on a slope should be judged by whether it improves performance for the intended landmark population and prediction horizon, not merely by whether its fitted coefficient differs from zero.