Random Slopes for Individual Change

Why Allow Subjects to Change at Different Rates?

A random intercept allows subjects to start at different biomarker levels while sharing the same time trend. That can be too restrictive when some subjects’ biomarker levels rise or fall faster than others’. A random slope for time represents this additional source of heterogeneity: subjects can differ in their rates of change, not only in their baselines.

This distinction matters when interpreting longitudinal data. A population-average decline does not imply that every subject declines at exactly that rate. A random-slope model describes the spread of subject-specific rates around the average, while retaining a fixed effect for the average trend.

Model and Assumptions

Let \(Y_{ij}\) be the biomarker measurement for subject \(i\) at visit \(j\), and let \(t_{ij}\) be time since a chosen reference point, measured in a specified unit such as months. A linear model with a random intercept and random time slope is

Here, \(\beta_0+\beta_1t\) is the population-average mean trajectory under the model. The subject-specific intercept is \(\beta_0+b_{0i}\), and the subject-specific slope is \(\beta_1+b_{1i}\). Thus \(b_{0i}\) is subject \(i\)’s deviation from the population intercept, while \(b_{1i}\) is that subject’s deviation from the population rate of change.

For the standard Gaussian random-effects model, assume the subject effects have mean zero and a common covariance matrix, and that the within-subject errors have mean zero and variance \(\sigma^2\):

The random-effects vectors are assumed independent across subjects. Conditional on the random effects and the visit times, residual errors are assumed independent across observations and independent of the random effects in this basic specification. These assumptions make the model a useful starting point; a different residual structure may be needed when measurements have additional serial dependence or changing residual variance. The random-slope interpretation also assumes that a linear time trend is a reasonable summary over the follow-up period being modeled.

The random effects have mean zero, so the average subject-specific intercept and slope equal \(\beta_0\) and \(\beta_1\), respectively, under this model. The key slope-variation parameter is \(\tau_{11}\). Its square root, \(\sqrt{\tau_{11}}\), is the standard deviation of subject-specific slopes. If the biomarker is measured in mg/L and time in months, slopes are in mg/L per month and \(\tau_{11}\) is in \((\text{mg/L per month})^2\).

What the Random Slope Says About Change

For a given subject, the model’s latent mean at time \(t\), excluding residual error, is \(\beta_0+b_{0i}+(\beta_1+b_{1i})t\). Between two times \(t_a\) and \(t_b\), the subject’s latent change is therefore

The intercept cancels: differences in latent change over this interval come from slope differences. Since \(b_{1i}\) has mean zero and variance \(\tau_{11}\), the population-average latent change is \(\beta_1(t_b-t_a)\), and the between-subject variance of latent change is \((t_b-t_a)^2\tau_{11}\). This result concerns variation in the underlying subject trajectories, not measurement error around those trajectories.

Time Coding and the Meaning of the Intercept

The time origin determines what the intercept represents. If \(t=0\) is baseline, \(\beta_0+b_{0i}\) is subject \(i\)’s latent baseline level. If time is recentered so that zero corresponds to month 3, the intercept instead describes the latent level at month 3. The subject-specific slope remains the rate of change per month: shifting the time origin changes the intercept’s interpretation, not the rate.

Choosing a clinically meaningful time origin often makes model coefficients easier to explain. A time unit must also be stated: a slope per month has a different numerical value from the same rate expressed per year. The random-slope variance changes by the square of the corresponding unit conversion, while its substantive interpretation remains between-subject variation in rate of change.

Justification: Why Slope Variance Measures Change Heterogeneity

The interval-change expression shows precisely what the random slope contributes. Subtract the latent level at \(t_a\) from the latent level at \(t_b\); the intercept terms cancel, leaving \((\beta_1+b_{1i})\Delta t\), where \(\Delta t=t_b-t_a\). The fixed component is common to the population, and the subject-varying component is \(b_{1i}\Delta t\). Therefore,

This derivation explains why \(\tau_{11}\) is variation in rate, whereas \((\Delta t)^2\tau_{11}\) is variation in latent change over a particular interval. The longer the interval, the greater the model-implied spread in accumulated change, in proportion to its length squared on the variance scale.

Diagnostics, Alternatives, and Limitations

A random slope adds flexibility, but it is not automatically warranted. Examine whether the assumed linear trend is plausible over the observed time range, whether there are enough subjects and repeated measurements to distinguish slope variation from residual noise, and whether model fitting is stable. Useful warning signs include nonconvergence, implausible estimates, a random-slope variance estimated near zero, or a singular fit. Sparse visits or very similar visit times within subjects can make individual slope variation difficult to estimate. Residual plots and comparisons of observed trajectories with fitted trajectories can also reveal systematic curvature or changing residual spread that a linear random-slope model does not address.

A random-intercept-only model is simpler and assumes subjects share one time slope; it may be reasonable when slope heterogeneity is negligible or unsupported by the data. A model with separate fixed slopes for every subject can describe the observed sample, but does not use a shared distribution to represent subject-to-subject variation and can require many parameters. The random-slope model is a compromise: it estimates a population trend and a distribution of subject-level deviations. Model comparison should account for the inferential method and the specific question; a more complicated model is not better merely because it has more parameters.

The basic specification has limitations. A single linear random slope may not capture nonlinear trajectories, abrupt changes, or groups with qualitatively different time patterns. The normal random-effects assumption may be a poor description of the population, especially with few subjects or influential trajectories. Random-slope estimates are also tied to the time scale and follow-up range used in the model. In longitudinal and time-to-event analyses, this model describes the biomarker trajectory component; by itself it does not specify how that trajectory is associated with an event time.

Common Pitfall

A second interpretation error is to call \(\sqrt{\tau_{11}}\) a standard error for the population slope. It is instead the model-based standard deviation of subject-specific slope deviations. Uncertainty in estimating \(\beta_1\) is a separate quantity.