Diagnosing Conflict Between Patient Data and Population Priors

Why Compare Patient Data With the Population Model?

An individual concentration may differ from the population-model prediction without indicating a problem. Population models allow patients to vary, and measured concentrations also contain observation error. The diagnostic question is therefore not simply whether an observation differs from a typical prediction. It is whether the observation is surprising relative to the full distribution of observations the model predicts before using that patient’s data.

This comparison is useful because it can flag a mismatch that merits investigation: perhaps the recorded dose or sampling time is wrong, the assay result is unusual, or the individual is not well represented by the model. It does not, by itself, identify the cause. Nor does a surprising concentration prove that the population model is wrong; rare observations occur even when a model is appropriate.

Constructing a Prior-Predictive Residual

Let \(\eta\) denote an individual’s model parameters or random effects, with population prior distribution \(p(\eta)\). At a specified sampling time and under a specified dose history, let the model predict concentration \(C(\eta)\), and let the observed concentration be \(Y\). The observation model describes how \(Y\) varies around \(C(\eta)\). For example, an additive-error model is \(Y=C(\eta)+\epsilon\), where \(\epsilon\) has mean zero and variance \(\sigma^2\), independently of \(\eta\).

The prior-predictive mean and variance follow from the law of total expectation and the law of total variance:

Thus, the predictive variance includes both observation uncertainty and variation in the model-predicted concentration across the population prior. If the model is nonlinear, these quantities can be calculated by integrating over the prior or approximated by drawing parameter values from it and simulating observations. The resulting distribution need not be normal.

When the prior-predictive distribution is adequately approximated by a normal distribution, with mean \(m\) and standard deviation \(s_{\mathrm{pred}}\), a standardized prior-predictive residual is:

This residual reports how far the observed concentration \(y\) lies from the prior-predictive mean, in prior-predictive standard deviation units. It is not automatically a standard normal statistic: that interpretation depends on the predictive distribution being approximately normal and on the model assumptions being adequate. If the distribution is skewed, bounded, or heavy-tailed, a predictive quantile or tail probability can be more informative than a z-score.

For the additive-error model, the variance calculation has a short justification. Conditional on \(\eta\), the expected observation is \(C(\eta)\) and its variance is \(\sigma^2\). Substitution into the total-variance identity gives:

This addition is appropriate when observation error is independent of the individual parameters. Omitting either term can make the predictive distribution too narrow and make ordinary observations appear spuriously conflicting.

Reading Patterns, Not Just Individual Residuals

With multiple observations, calculate each prior-predictive residual using the prediction for its dose history and sampling time. Their signs and sizes can reveal patterns that a single concentration cannot. Several observations consistently above their predictions may suggest a shared mismatch, while one isolated large residual may be more consistent with a local measurement or recording issue. These are clues to investigate, not diagnoses.

Observations from the same patient are often correlated because they share individual parameters, and their errors may also be correlated. Therefore, a collection of marginal z-scores should not automatically be treated as independent evidence. If the joint prior-predictive distribution is available, a vector of residuals \(r=y-m\) can be assessed using its covariance matrix \(S\), for example through the quadratic discrepancy \(r^\mathsf{T}S^{-1}r\). Its reference distribution depends on the assumed joint predictive distribution. Ignoring dependence can exaggerate how much evidence a set of observations provides.

A further distinction matters: prior-predictive residuals assess compatibility before using the patient’s observations to update individual parameters. Residuals calculated after fitting those same observations are conditioned on the data and can be smaller simply because the individual fit has adapted to them. Such post-fit residuals answer a different diagnostic question and should not replace the prior-predictive comparison when asking whether the observations conflict with the population prior.

Posterior Shifts: How the Data Move an Individual Estimate

A concentration that conflicts with the population prior can pull an individual parameter estimate away from its prior center. The size and direction of that movement depend not only on the residual but also on how informative the observation is about the parameter. A low-precision observation can be surprising yet produce a limited update; a precise observation can move the estimate substantially.

For a simple, exactly solvable illustration, let \(\theta\) be a scalar quantity representing the model-predicted concentration for a fixed regimen and sampling time. Assume a normal prior \(\theta\sim N(\mu,\tau^2)\), and an observation model \(Y\mid\theta\sim N(\theta,\sigma^2)\), with \(\theta\) and observation error independent. The posterior mean is:

To see why, the posterior density is proportional to the product of the prior density and the likelihood. Keeping only terms involving \(\theta\), its log density is proportional to:

Collecting the quadratic and linear terms in \(\theta\) gives posterior precision \(1/\tau^2+1/\sigma^2\) and posterior mean \((\mu/\tau^2+y/\sigma^2)/(1/\tau^2+1/\sigma^2)\). Rearranging that mean yields the expression above. The multiplier \(\tau^2/(\tau^2+\sigma^2)\) lies between zero and one: the posterior mean is between the prior mean and the observed value. This result is exact under the stated scalar normal model; it is not a general formula for nonlinear PK models.

Investigating Conflict and Recognizing Failure Modes

A large standardized residual is a reason to check the assumptions and inputs before changing a dose or rejecting the population model. In practice, plausible causes include:

• Incorrect or incomplete dosing history: the recorded amount, administration time, or recent missed dose may not match what occurred.
• Incorrect sample timing: a recorded collection time may be wrong, or the interval from the most recent dose may have been miscalculated.
• Unusual observation error: assay, sample handling, or transcription problems may affect a result.
• Patient-model mismatch: the individual may have characteristics or clinical changes not represented adequately by the model or its population prior.
• Mis-specified variability: the population or observation-error variance may be too small, making the model’s predictive distribution too narrow.

These possibilities have different implications and should not be collapsed into a single conclusion. For instance, a verified outlying assay result suggests a different follow-up from several reproducible concentrations that remain systematically above prediction after the dose history and timing are checked. A concentration should not be discarded merely because it is unusual: doing so can remove genuine information about an individual who differs from the population average.

A normal z-score threshold is only a convenient screen when the predictive distribution is approximately normal. In skewed or heavy-tailed settings, compare the observation with simulated prior-predictive draws or calculate its position in the predictive distribution. For repeated measurements, consider their joint distribution and dependence. If many patients or time points are screened, some apparently extreme values will occur by chance; investigate the pattern and context rather than applying a universal cutoff as an automatic decision rule.

The alternatives serve different purposes. A prior-predictive residual asks whether the observed data were expected before the patient-specific update. A posterior predictive check asks whether a model updated using the data can reproduce relevant features of those data; it may be less sensitive to a prior mismatch because the update has already adapted to the observations. Direct review of records and assays can identify concrete data problems that neither predictive calculation can establish. None of these checks alone proves a causal explanation.

The main limitations are that diagnostics depend on the model and inputs being checked, and that a single residual cannot reveal which parameter or process caused a mismatch. A moderate residual does not guarantee that a model is adequate, while a large residual does not prove that it is defective. Interpret the predictive comparison alongside data verification, the pattern across observations, and clinical context.