Slope Association: Definition and Interpretation

Why Associate the Hazard With a Slope?

A marker’s level describes where a subject’s latent trajectory is at a particular time. Its slope describes how quickly that trajectory is changing at that time. These answer different questions: two subjects can have the same marker level but different rates of change, or have different levels while changing at the same rate. A slope association allows the event hazard to depend on that rate of change.

The key quantity is not the derivative of a noisy observed measurement series. It is the instantaneous derivative of the subject-specific latent marker trajectory. The model links that derivative to the hazard, conditional on the model’s other terms. The coefficient therefore describes a hazard effect per unit difference in rate of change, not per unit difference in marker level.

Definition and Model

A slope-only association term can be written as

Here \(h_i(t\mid b_i)\) is subject \(i\)’s hazard at time \(t\), conditional on the subject-specific latent effects \(b_i\); \(h_0(t)\) is the baseline hazard; \(\boldsymbol{w}_i\) contains other covariates with coefficient vector \(\boldsymbol{\gamma}\); and \(\alpha_s\) is the slope-association coefficient. The latent trajectory, and hence its derivative, may depend on \(b_i\). The displayed model isolates the slope term; other association terms could be present in a different model.

The hazard is conditional on the modeled covariates and latent effects. The model assumes that the log hazard is linear in \(m_i'(t)\), with a constant coefficient \(\alpha_s\), and that the hazard ratio for a fixed slope difference does not otherwise vary with time. It also requires that \(m_i(t)\) be differentiable over the times at which the slope is used, and that the slope be defined using the time and marker scales chosen for the analysis.

Interpreting the Coefficient

Compare two subjects at the same time, with the same values of the other hazard covariates, and suppose their latent slopes differ by \(\Delta\). The ratio of their hazards is

Thus, \(\exp(\alpha_s\Delta)\) is the hazard ratio for a \(\Delta\)-unit difference in latent slope, under the model. A positive \(\alpha_s\) means that a higher slope is associated with a higher hazard, all else equal; a negative \(\alpha_s\) means that a higher slope is associated with a lower hazard. The direction refers to the signed rate of change: a higher slope can mean a faster increase, or a less rapid decrease.

Units are essential. If the marker is measured in units \(U\) and time in years, \(m_i'(t)\) has units \(U/\text{year}\). Since the exponent must be dimensionless, \(\alpha_s\) has reciprocal units, \(\text{year}/U\). The coefficient itself is not a hazard ratio. For a one-unit-per-year slope difference, the hazard ratio is \(\exp(\alpha_s)\); for any other difference, use \(\exp(\alpha_s\Delta)\).

The comparison is associational, not automatically causal. It describes a conditional hazard contrast implied by the fitted joint model. It does not establish that changing a subject’s marker slope would change that subject’s event hazard.

Worked Interpretations

Justification of the Hazard-Ratio Formula

The hazard-ratio interpretation follows directly from the model. At a fixed time, take the hazards for two subjects with the same values of \(\boldsymbol{w}\). Their common baseline hazard and covariate contribution cancel in the ratio:

Writing the slope difference as \(\Delta=m_i'(t)-m_j'(t)\) gives \(\exp(\alpha_s\Delta)\). This is a ratio of instantaneous hazards, not a ratio of event probabilities over an interval. The interpretation also depends on the stated comparison: if other covariates differ, their contributions must be included rather than canceled.

Assumptions, Diagnostics, and Alternatives

The interpretation is meaningful only when the latent slope is well-defined on the scale used in the model and the hazard specification is suitable. In particular, the slope link assumes a linear contribution of \(m_i'(t)\) to the log hazard and a time-constant association coefficient. The proportional-hazards formulation further means that, for a fixed slope contrast, the modeled hazard ratio is constant over time. These are modeling assumptions, not consequences of using a derivative.

Assessment should focus on whether the model is representing the data adequately, rather than treating a numerical slope estimate as self-validating. Relevant checks include whether the marker trajectory fits the repeated measurements, whether the assumed trajectory is smooth enough for a derivative to be meaningful, and whether the event-model residual diagnostics indicate lack of fit. Investigators should also consider whether the slope effect may vary with time or be nonlinear; a single constant \(\alpha_s\) does not capture every possible relationship.

A common mistake is to calculate a difference between two noisy observed measurements and call it the subject’s slope. Such a difference can be strongly affected by measurement error, the spacing of observations, and short-term fluctuation. In the joint model, the association term is defined using the derivative of the latent trajectory, not an unadjusted observed difference. That latent slope is estimated through the longitudinal model; it is not simply read off from one pair of recorded values.

Slope association also answers a different scientific question from the main alternatives. A current-value association relates the hazard to the latent marker level at time \(t\); a cumulative association relates it to an accumulated feature of the marker history. A slope association instead links the hazard to the marker’s instantaneous rate of change. These structures are not interchangeable: the best choice depends on the intended scientific meaning and the adequacy of the resulting model. A slope-only link can omit relevant information about marker level, while adding multiple association terms requires care because their effects may be difficult to distinguish in the data.

A further limitation is that the estimated slope can be sensitive to how the longitudinal trajectory is modeled, especially where measurements are sparse or the trajectory is poorly supported. If time or marker units change, the numerical coefficient changes too; its hazard-ratio interpretation must use the corresponding slope contrast in those units. Slope association describes a conditional model relationship, not an intervention effect.