Diagnostics for Mixed-Effects Survival Models

Why Survival-Model Diagnostics Need Special Care

A mixed-effects survival model can describe event times that vary both with measured covariates and with unobserved subject-level heterogeneity. A diagnostic must therefore make clear which survival distribution it is checking: survival conditional on a subject’s random effect, or survival averaged over the random-effect distribution. Censoring adds another complication. A curve calculated only from observed event times generally does not estimate survival correctly unless censoring is handled under appropriate assumptions.

The checks here focus on calibration: whether the fitted model’s survival probabilities and event-time summaries agree with what is observed. They can reveal systematic overprediction or underprediction, but a close match does not prove that the model is correct. In particular, diagnostics can be weak when samples are small or censoring is extensive.

Observed Versus Predicted Survival

A calibration plot compares estimated observed survival with model-predicted survival over time. For a group of subjects, the observed curve is often estimated by the Kaplan–Meier estimator. Its usual interpretation as the group’s survival curve requires censoring to be independent of event time marginally within that group. If censoring is independent only after conditioning on covariates or random effects, an unadjusted pooled Kaplan–Meier curve need not estimate the group’s marginal survival: mixing over those variables can induce dependence between censoring and event time.

The predicted curve must target the same population as the observed curve. A conditional prediction \(S(t\mid x_i,b_i)\) describes a subject at specified covariates and random effect. A population-averaged prediction integrates over random effects. For a cohort with covariates \(x_1,\ldots,x_n\), one simple standardized model curve is the average of \(S_{\mathrm{pop}}(t\mid x_i)\) over that cohort. This is not necessarily the same as a prediction at the average covariate value.

When censoring depends on measured covariates, possible remedies include comparisons within covariate strata or censoring-adjusted methods such as inverse-probability-of-censoring weighting. Such methods require an appropriate model for censoring and adequate data to estimate it. Random-effect heterogeneity must also be represented in the prediction or adjustment. A conditional independence assumption alone does not justify an ordinary pooled Kaplan–Meier comparison.

Cumulative-Hazard Residuals

A cumulative-hazard residual transforms an event time onto a scale with a simple reference distribution. For a continuous event-time model, define the cumulative hazard \(H(t)=-\log S(t)\). If the correct survival function for a subject’s covariates is known, the transformed event time \(R=H(T)\) follows a unit exponential distribution. In a mixed-effects model, this transformation can use either the conditional survival function at the true random effect or the survival function marginalized over the random-effect distribution. In practice, the latter avoids treating an estimated individual random effect as known.

For a censored subject, the transformed censoring time is observed, but the transformed event time is only known to exceed it. Thus residuals are censored too; they must not be treated as observed events at the censoring values. Under the correctly specified continuous event model, \(P\{H(T)>r\}=e^{-r}\) for \(r\geq 0\), so \(R\sim\mathrm{Exp}(1)\), with survival \(e^{-r}\) and cumulative hazard \(r\).

Plot a nonparametric cumulative-hazard estimate of the transformed residuals against \(r\), with the reference line \(r\). Systematic curvature or persistent departures can indicate a mismatch in the event-time distribution. For example, a curve consistently above the reference suggests more cumulative residual events than expected over that range. Sampling variation, especially in the tail where few subjects remain at risk, can create substantial apparent departures.

An ordinary pooled Kaplan–Meier or cumulative-hazard comparison of residuals still needs marginal independent censoring on the residual scale. Conditional independence given covariates or random effects does not guarantee this after mixing. If censoring independence holds only conditionally, use an appropriate conditional, stratified, or censoring-adjusted analysis and account for the random-effect distribution. Also, residuals formed using empirical Bayes estimates of \(b_i\) are not exactly unit exponential: those estimates are uncertain and often informed by the same event data being checked. Marginal transformations, or simulation-based checks that repeat the fitting and estimation process, help avoid overinterpreting such plug-in residuals.

Event-Time Summaries and Diagnostics

Event-time summaries provide a compact check on clinically interpretable features of the distribution. Common choices include the median survival time and the restricted mean survival time (RMST) up to a prespecified horizon \(\tau\). The RMST is the area under the survival curve from zero to \(\tau\), or equivalently the expected event-free time accumulated up to \(\tau\). Comparing such summaries can show whether a model places the central event time or average event-free time too early or too late.

The Kaplan–Meier median is the earliest time at which the estimated survival is at most 0.5; it is not estimable from a curve that remains above 0.5 throughout observed follow-up. RMST estimated from Kaplan–Meier data relies on censoring assumptions over the interval through \(\tau\). Model-based summaries should be computed for the same target population and horizon as their observed counterparts, and uncertainty should be reported when drawing conclusions. Comparing only observed event times, while ignoring censored subjects, generally biases summaries toward shorter event times.

Practical Workflow, Failure Modes, and Alternatives

A common mistake is to argue that censoring is independent after conditioning on covariates and random effects, then compare one pooled Kaplan–Meier curve with a population-averaged prediction. That conditional statement does not, in general, ensure marginal independence after mixing; the subjects still under observation can have a different covariate or random-effect composition. The correction is to establish marginal independent censoring for the pooled comparison, or to condition, stratify, or adjust for the censoring-related variables while representing heterogeneity appropriately.

Other warning signs include treating censored residuals as events, using conditional predictions as if they described the whole cohort, and interpreting a late-tail discrepancy where very few subjects remain at risk as strong evidence. Alternatives to a simple pooled curve include covariate-stratified comparisons, inverse-probability-of-censoring-weighted estimates, and simulation-based checks. These approaches have their own assumptions: weighting depends on a credible censoring model, and simulations can only reflect the model and estimation process specified. No single diagnostic detects every form of misspecification.