Random Slopes and Slope Association Parameters

Two Parameters, Two Different Questions

A longitudinal trajectory can differ across subjects in how quickly the marker changes. A time-to-event model can also relate those differences in rate of change to event risk. These are connected ideas, but they are not the same parameter: random-slope variation describes how much slopes vary, whereas a slope association coefficient describes how the hazard changes with slope.

The distinction matters for interpretation and estimation. A population may have substantial variation in marker slopes without any modeled relationship between slope and event risk. Conversely, a slope association can be scientifically meaningful only when the data provide information about subject-specific slopes. The joint model represents these roles separately.

Model and Assumptions

For subject \(i\), let \(Y_{ij}\) be the measured marker at time \(t_{ij}\). A linear mixed trajectory is

Here \(m_i(t)\) is the subject-specific mean marker trajectory, \(\beta_0\) and \(\beta_1\) are population fixed effects, and \(b_{0i}\) and \(b_{1i}\) are the subject's random intercept and random slope. The residual \(\epsilon_{ij}\) represents measurement-level variation around that trajectory. Under this linear model, the subject-specific slope is the derivative

A common specification assumes \((b_{0i},b_{1i})^\mathsf{T}\sim N_2(0,D)\), where

Thus \(\tau_{11}=\operatorname{Var}(b_{1i})\) is the between-subject variance in deviations from the population slope, and \(\sqrt{\tau_{11}}\) is the standard deviation of those deviations. The covariance \(\tau_{01}\) describes how intercept and slope deviations vary together. Neither quantity is the slope association coefficient.

For illustration, suppose the event hazard conditional on the random effects and covariates is

Here \(h_0(t)\) is the baseline hazard, \(w_i\) is a vector of observed covariates, and \(\gamma\) is its coefficient vector. The coefficient \(\alpha_v\) links the current marker value to the hazard; \(\alpha_s\) links the subject-specific slope to the hazard. A model may omit one or both association terms, but this tutorial focuses on distinguishing \(\alpha_s\) from \(\tau_{11}\).

This specification relies on assumptions that should be made explicit. The trajectory is linear over the time range being modeled; the random effects follow the stated distribution; and, commonly, the measurement errors are mean-zero with a specified variance and independent across measurements conditional on the random effects. The hazard is proportional in the stated covariates and latent trajectory features. The joint model also assumes that, conditional on the shared random effects and observed covariates, the longitudinal measurement process and event process have no remaining dependence relevant to the model. Inference additionally relies on an appropriate censoring and observation-process assumption. A poor fit to any of these components can distort the estimated association.

Why the Parameters Must Not Be Confused

The random slope \(b_{1i}\) is a latent subject-level quantity; \(\tau_{11}\) describes its population distribution. By contrast, \(\alpha_s\) is a hazard-model coefficient. It does not say how variable slopes are, but how hazard changes across slope values under the conditional model.

This difference is clear in the slope contrast. Compare two otherwise identical subjects at the same time, with the same current marker value and covariates, but slopes \(s_1\) and \(s_0\). Their conditional hazard ratio is

The common terms, including the baseline hazard and current-value contribution, cancel. This calculation shows why \(\alpha_s\), rather than \(\tau_{11}\), determines the hazard ratio for a specified slope contrast. If current marker values differ too, the log hazard contrast also includes \(\alpha_v\) times their difference.

The units help keep the interpretations straight. If the marker is measured in units \(U\) and time in units \(T\), then a slope is measured in \(U/T\). The random-slope variance has units \((U/T)^2\); \(\alpha_s\) has units \(T/U\), so that \(\alpha_s m_i'(t)\) is dimensionless. A hazard ratio is reported for a stated slope contrast and time unit.

Calculating and Interpreting a Slope-Linked Hazard Ratio

A slope association is most directly interpreted through a contrast. The coefficient's sign alone is not enough: the direction must be connected to how the slope is defined, the slope contrast, and the conditional nature of the comparison. In particular, a positive \(\alpha_s\) means that a higher slope is associated with a higher conditional hazard, holding the other terms fixed. It does not establish a causal effect.

When both current value and slope are included, the contributions must be kept separate. For two subjects with covariates held equal, their conditional log hazard ratio is

This expression makes the conditioning explicit. If the current values differ, the resulting contrast is not a slope-only comparison. A slope interpretation that ignores a difference in current values may attribute too much of the modeled hazard contrast to slope.

How the Joint Model Uses the Two Components

The longitudinal measurements inform the distribution of subject-specific trajectories, including \(\tau_{11}\). The event times and censoring information contribute information about how those latent trajectories relate to event risk, including \(\alpha_s\). Because individual slopes are not observed directly, the model estimates them through the longitudinal data while fitting the event process jointly. A measured change between two noisy observations is not automatically the subject's slope.

A brief algebraic check shows the different roles of the two parameters. Since \(m_i'(t)=\beta_1+b_{1i}\), the slope term in the log hazard can be written as

The first term is common to subjects at a given model parameter value; the second varies with the subject's random slope. The variance of that random deviation is \(\tau_{11}\), while its multiplier in the log hazard is \(\alpha_s\). A larger \(\tau_{11}\) means more spread in slopes; a larger positive \(\alpha_s\) means a stronger upward hazard gradient across those slopes. Neither parameter can be substituted for the other.

There is also an important data-information qualification. If \(\tau_{11}=0\), all subjects have the same slope under this model, so there is no between-subject slope contrast from which to learn a slope association. In that boundary case, \(\alpha_s\) is generally not separately identified through between-subject slope variation. Even when \(\tau_{11}>0\), sparse or noisy longitudinal measurements can make individual slopes uncertain and weaken estimation of \(\alpha_s\).

Model Comparisons, Diagnostics, and Limitations

Several model choices answer different questions:

In practice, assess whether the longitudinal trajectory captures the observed patterns, whether residual behavior is compatible with the measurement model, and whether the random-effects distribution is plausible. For the event component, examine whether the specified proportional-hazards structure and association form are adequate. Check convergence, uncertainty in variance estimates, and sensitivity to plausible model specifications. Very few measurements per subject, measurements concentrated in a narrow time range, or little variation in estimated slopes can make \(\alpha_s\) difficult to estimate. A strong relationship between current value and slope can also make their separate hazard contributions imprecise.

Do not diagnose a slope link by simply plotting raw pairwise differences against event outcomes: noisy differences can be unstable and can misrepresent the latent trajectory. Likewise, a nonzero estimated \(\alpha_s\) does not mean that changing a subject's slope would change that subject's event risk. It is a conditional association under the joint model and its assumptions. The linear slope is also a simplifying summary; if the trajectory is not approximately linear over the relevant range, the slope term may be misleading. More complex trajectory specifications require their own careful interpretation.

Common Mistake