Common Misreadings Of Joint-Model Residuals

Why the Sampling Structure Matters

A residual is a discrepancy between observed data and a model-based expectation. It is useful only when interpreted against the process that produced it. In a joint model, observations from one person share latent random effects, event times may be censored, and fitted parameters are estimated from the same subjects being checked. Consequently, a residual is not automatically an independent draw from a standard reference distribution. A large value can signal model misfit, but it can also reflect conditioning, shrinkage, censoring, or the chance of seeing an extreme among many checks.

For a simple shared-parameter joint model, write the longitudinal outcome for person \(i\) at visit \(j\) as \(Y_{ij}\), with conditional mean \(m_i(t_{ij};b_i)\) and residual error \(\epsilon_{ij}\). Here \(b_i\) is a person-specific random effect. A possible event hazard conditional on the same random effect is \(h_i(t\mid b_i)\). The precise model may be more elaborate, but these components suffice to clarify common misreadings. The earlier material on residual diagnostics and posterior predictive checks establishes tools for examining model fit; the focus here is how not to overinterpret individual residual quantities.

These definitions already signal three distinct sources of caution. A conditional residual depends on what is held fixed; an empirical Bayes estimate is a shrunken estimate rather than an observed latent quantity; and a censored event contributes information about survival up to censoring, not a known event time. Isolated extremes add a fourth issue: many residuals are examined, often with dependence among them.

Conditional Residuals Are Not Marginal Prediction Errors

Suppose the longitudinal model is \(Y_{ij}=m_i(t_{ij};b_i)+\epsilon_{ij}\), with \(\epsilon_{ij}\) having mean zero and variance \(\sigma^2\) conditional on \(b_i\). A raw conditional residual is \(e_{ij}=Y_{ij}-m_i(t_{ij};\widehat b_i)\), where \(\widehat b_i\) is an estimated random effect. It asks whether a measurement differs from the fitted trajectory for that particular person. It does not ask whether the measurement differs from the population-average trajectory.

If the conditional error is normal with known parameters and the true \(b_i\) is fixed, then \([Y_{ij}-m_i(t_{ij};b_i)]/\sigma=\epsilon_{ij}/\sigma\) has a standard normal distribution under the model. In practice, however, \(b_i\) and model parameters are estimated, often using the same measurements whose residuals are being inspected. A residual based on \(\widehat b_i\) therefore does not generally have that simple reference distribution. The fitted subject trajectory can absorb some of the person's unusual pattern, making its conditional residuals look less extreme than marginal prediction errors would.

The distinction is especially important when a diagnostic compares conditional residuals with marginal predictions. Marginal predictions average over the random-effects distribution; conditional predictions use a subject-specific effect. They answer different questions and have different predictive variability. A residual that is ordinary conditionally may be unusual relative to the population distribution, or vice versa. State which prediction is being assessed before interpreting the residual.

Empirical Bayes Random Effects Are Shrunken

Empirical Bayes estimates are often plotted or inspected as if they were observed values of \(b_i\). They are not. The data for a person may be sparse or noisy, so the estimate is pulled toward the population mean (usually zero for a centered random effect). The amount of shrinkage varies with the information available for that person and the estimated random-effects distribution.

A simple normal calculation shows the mechanism. Suppose \(b\sim N(0,1)\), and a subject-specific summary \(Z\) satisfies \(Z\mid b\sim N(b,0.25)\), independently of other subjects conditional on \(b\). Completing the square in the normal posterior density gives a posterior variance \(1/(1+4)=0.2\) and posterior mean \(0.8Z\). Thus the estimate is only 80% of the observed summary in this example.

A narrow spread of empirical Bayes effects can therefore be expected even when the random-effects distribution is correctly specified. Conversely, a few large estimates do not, by themselves, establish a heavy-tailed population distribution: their magnitude must be considered alongside the amount of subject-specific information and uncertainty. For distributional diagnostics, use procedures that account for estimation and shrinkage, such as the simulation-based checks developed for random-effects distributions, rather than treating estimates as raw independent draws.

Censoring Changes What an Event Residual Says

Let \(T_i\) be the observed follow-up time, the minimum of the event time and censoring time, and let \(\delta_i=1\) when the event is observed and \(0\) when follow-up ends by censoring. Given a subject's random effect and covariates, define the fitted cumulative hazard through time \(t\) as \(H_i(t)=\int_0^t h_i(u\mid b_i)\,du\). The basic martingale residual is \(M_i=\delta_i-H_i(T_i)\), with fitted quantities substituted in practice.

For an observed event, the event count is one, so the residual is \(1-H_i(T_i)\). For a censored person, the observed event count is zero, so it is \(-H_i(T_i)\). A negative residual for a censored person does not mean that a future event was observed earlier than it should have been. It reflects survival without an event through the observed follow-up, compared with the model's cumulative hazard over that period. Censoring time is not the unobserved event time.

The martingale-residual construction is motivated by the counting-process difference between observed events and their fitted cumulative intensity while at risk. Under a correctly specified intensity and appropriate independent censoring, that process has mean-zero increments in the relevant filtration. This property does not make a single fitted residual normally distributed: event counts are discrete, follow-up varies, and parameters and random effects are estimated. Censoring assumptions matter as well. If censoring is informative given the variables and random effects included in the model, the event-process diagnostic may be misleading because the observed risk process no longer reflects the assumed observation mechanism.

An Isolated Extreme Is Not a Verdict

Even if individual residuals had a standard normal reference distribution, examining many residuals makes at least one apparently large value unsurprising. Under the idealized assumption of \(n\) independent standard normal residuals, the probability that at least one has absolute value above 2 is \(1-(1-0.0455)^n\), since 0.0455 is approximately the two-sided standard normal tail probability beyond 2. With \(n=10\), this is about 37.2%. This calculation is an illustration, not a universal joint-model calibration: repeated measurements within a person are dependent, and fitted residuals are not generally independent standard normals.

Practical Reading and Failure Modes

A diagnostic is most informative when its target and sampling unit are explicit. Ask whether the check concerns within-person measurement errors, subject-level heterogeneity, the event process, or joint dependence. Also ask whether it is conditional on estimated random effects, and whether the reference distribution accounts for fitting and censoring. The earlier material on simulation envelopes and bootstrap uncertainty provides tools for calibration; here the central point is that calibration must correspond to the quantity being plotted.

Common failure modes include standardizing a conditional residual by \(\widehat\sigma\) and assuming that this alone gives a standard normal score; comparing the spread of empirical Bayes effects directly with the latent random-effects spread; assigning a censored person an event residual as if the event had occurred at censoring; and declaring a model invalid because one of many inspected values crosses a familiar cutoff. The corrections are to identify the conditioning, recognize shrinkage, preserve the distinction between event and censoring, and evaluate patterns using an appropriate reference distribution.

These cautions do not make residuals unhelpful. A cluster of large residuals at late visits can suggest time-dependent longitudinal misfit; systematic event residual patterns can point toward hazard misfit; and a subject with an extreme biomarker pattern and an unusual event history may motivate a joint check. Such signals are stronger when they recur in a meaningful structure or persist under a calibrated comparison than when they consist of one uncontextualized value.