Marker-Specific Slope Associations

Why Model Marker Slopes?

A marker’s current value describes its level at a particular time. Its instantaneous slope describes how quickly that level is changing at that time. These can convey different information about a person’s clinical course: two people may have the same marker value now, while one is improving and the other is deteriorating. A joint model can represent the association between survival and each marker’s rate of change, rather than relying only on the marker levels.

With multiple markers, the slope of each latent trajectory can enter the hazard through its own coefficient. This allows the model to estimate, for example, whether a faster rise in one marker is associated with a higher hazard while a faster rise in another is associated with a lower hazard. The associations are conditional on the other modeled marker slopes and covariates; they are not, by themselves, causal effects.

What the Slope Coefficients Mean

The hazard is conditional on the subject’s modeled latent trajectories and covariates. The longitudinal observations are not substituted directly as if they were error-free slopes; the slope term is computed from the fitted latent trajectory. This matters because a derivative of noisy measurements taken at irregular times can be unstable, whereas the joint model represents the trajectory and its uncertainty through the longitudinal model.

The units of \(\dot m_{ik}(t)\) are marker units per unit time. Therefore, \(\alpha_k\) has reciprocal units: time per marker unit. For example, if a marker is measured in milligrams per litre and time in years, its slope is in milligrams per litre per year, and \(\alpha_k\) is measured in years per milligram per litre. The product \(\alpha_k\dot m_{ik}(t)\) is dimensionless, as required in an exponential hazard model.

For two otherwise identical subjects at the same time, with the same covariates and the same slopes for all markers except marker \(k\), the conditional hazard ratio for a slope difference \(\Delta\) is \[ \frac{h_i(t\mid \dot m_{ik}(t)=s+\Delta)} {h_i(t\mid \dot m_{ik}(t)=s)} =\exp(\alpha_k\Delta). \] Thus, \(\exp(\alpha_k)\) compares hazards for a one-unit difference in slope, where “one unit” means one marker unit per time unit. It does not compare a one-unit difference in marker level. A positive \(\alpha_k\) means a more positive slope is associated with a higher conditional hazard, holding the other modeled quantities fixed. A negative \(\alpha_k\) means a more positive slope is associated with a lower conditional hazard. The meaning of “more positive” applies to both rising and falling profiles: a change from \(-0.5\) to \(-0.2\) is a more positive slope, even though both trajectories are falling.

Justifying the Hazard Ratio

The comparison follows directly from the specified hazard model. Let two subjects have the same covariates and the same marker slopes except that their slope for marker \(k\) differs by \(\Delta\). In the ratio of their hazards, the common baseline hazard and common covariate contribution cancel. The other marker-slope terms also cancel. The remaining log-hazard difference is \(\alpha_k\Delta\), so exponentiating gives the hazard ratio.

This derivation also clarifies the comparison’s conditions: it is conditional on equal values of the other hazard predictors and on the specified latent slope values. If several slopes differ, each contributes to the log-hazard difference.

Rising, Falling, and Flat Profiles

A slope coefficient is not a statement that “increases are harmful” or “decreases are protective” without specifying the coefficient’s sign and the comparison. With a positive coefficient, a positive slope contributes positively to the log hazard relative to a flat slope; a negative slope contributes negatively. With a negative coefficient, those directions reverse. The coefficient describes how the log hazard changes as the slope becomes more positive, not the effect of moving from any particular marker value to another.

Obtaining Slopes from the Longitudinal Model

The derivative in the hazard model comes from the chosen longitudinal trajectory specification. For a linear trajectory, the instantaneous slope is constant. For a nonlinear trajectory, it can vary with time, so the same coefficient may contribute differently to the hazard at different times for the same subject. In either case, the derivative must be taken with respect to the time variable used in the model.

Assumptions, Diagnostics, and Alternatives

Interpretation depends on assumptions from both parts of the joint model. The longitudinal submodels must provide an adequate representation of each latent trajectory over the times at which the hazard is evaluated, and the trajectory must be differentiable wherever its slope is used. The survival component assumes the stated hazard form: the baseline hazard and covariates are modeled appropriately, and each slope enters linearly on the log-hazard scale with a constant association coefficient unless the model explicitly specifies otherwise. The joint model must also represent dependence between marker histories and event times adequately; censoring should be non-informative conditional on the modeled information and assumptions.

Diagnostics should examine more than whether an optimizer returned estimates. Check convergence and identifiability evidence, including whether slope-coefficient estimates are stable and their uncertainty is plausible. Compare fitted longitudinal trajectories with observed marker histories, especially near event times and in regions with sparse observations, where derivatives can be poorly determined. Examine whether estimated slopes or hazard contributions are dominated by a small number of subjects. A coefficient with a very large standard error, unstable estimates across reasonable fits, or extreme sensitivity to a few subjects can indicate weak information about the slope association. A good-looking longitudinal fit alone does not establish that the survival association is well identified.

A slope association is one possible link between longitudinal and survival processes. A current-value association instead relates the hazard to latent marker levels; a shared-random-effects specification links the hazard to longitudinal random effects. These alternatives answer different modeling questions and need not produce the same association estimate. Choosing among them requires a scientific rationale and an adequate fit, not simply selecting whichever gives the strongest association. A model with slope terms alone also does not adjust for current marker levels unless those levels are included separately.

A common mistake is to interpret a positive slope coefficient as meaning that every rising trajectory has a higher hazard than every falling trajectory. That conclusion ignores the fitted coefficient’s sign, the sizes of the slopes, and the other terms in the hazard. The correct comparison is explicit: for a slope difference \(\Delta\), calculate \(\exp(\alpha_k\Delta)\) while holding the other predictors fixed. A further limitation is that these are conditional model-based associations, not causal effects. Unmeasured factors, misspecified trajectories, or inadequate handling of informative observation and censoring can undermine a causal interpretation.