Reading a Prediction-Error Pattern
Fault isolation is not a search for a one-to-one mapping between a residual pattern and a parameter. Several model failures can produce similar errors. The useful question is narrower: which explanations are consistent with the timing, size, and repetition of the errors, and which additional features of the data can distinguish them? This tutorial applies that reasoning to four common patterns across a PK profile.
Before interpreting a pattern, check dose amount, route, dose and sample times, units, assay records, and the model's input history. Compare observations with predictions made on a consistent basis: population predictions answer a different question from individual predictions that condition on estimated individual effects. A pattern is evidence to investigate, not proof of a particular mechanism. Sparse sampling, shrinkage, a misspecified residual-error model, and influential individuals can all distort the apparent pattern.
The cases below assume that the records have been checked and that observations being compared are sufficiently informative for the stated feature. Even then, a proposed revision is a hypothesis to test against the full data, not a diagnosis made from one plot alone.
Early-Profile Errors: Input Before Disposition
Errors concentrated around the rising phase or peak first raise questions about drug input: absorption rate, lag time, multiple absorption phases, or the dose history. The direction and timing help refine the possibilities. For example, underprediction on the rising limb followed by overprediction near the peak can indicate that the model brings drug into the system too slowly or spreads input over the wrong time window. A shifted peak can instead point toward a timing or lag-time discrepancy.
But early errors are not automatically absorption errors. An incorrect dose time, sample time, or recorded dose can shift the apparent profile. An unsuitable residual model can also make errors look more prominent at particular concentrations. Check those explanations before adding absorption parameters. Compare several individuals and doses: a similar time-locked shape across adequately sampled profiles is more persuasive than one unusual point.
Worked Example: A Peak That Arrives Too Early
Hypothetical data. After an oral dose, a model predicts concentrations of 2.0 mg/L at 1 hour and 5.0 mg/L at 2 hours. The observations are 3.0 mg/L and 4.0 mg/L, respectively. Assume verified sampling times, a verified dose record, and a residual standard deviation of 0.5 mg/L at both times.
At 1 hour, the signed error is \(3.0-2.0=+1.0\) mg/L, giving \(r=1.0/0.5=+2.0\). At 2 hours, it is \(4.0-5.0=-1.0\) mg/L, giving \(r=-1.0/0.5=-2.0\). Thus the model is relatively low early and high later: its predicted rise and peak timing do not match this profile.
A plausible first model check is whether the absorption rate or lag-time specification makes input too delayed. The next check is whether this same directional pattern recurs in other well-sampled profiles. If it does, compare a targeted absorption revision with alternatives such as a second absorption phase, using the observed curve shape and dosing conditions. If it occurs only for one profile, verify the records and consider an individual-level explanation before changing the population absorption model. Two points alone do not establish which absorption mechanism is correct.
Terminal Errors: Separate Slope From Scale
Late errors can arise when the model predicts the terminal decline too quickly or too slowly, but a vertical separation between observed and predicted concentrations is not, by itself, evidence of a wrong terminal slope. With first-order elimination in a one-compartment model after input has ended, concentration has the form \(C(t)=C_0e^{-kt}\). Taking logarithms gives \(\log C(t)=\log C_0-kt\), so the slope of log concentration against time is \(-k\). This interpretation requires the one-compartment, first-order elimination phase to apply over the times being compared.
If the observed decline is consistently shallower than predicted, possible explanations include a missing distribution compartment, absorption-limited terminal behavior after oral dosing, or a change in clearance over time. The route and the rest of the profile matter: absorption-limited terminal behavior is not a viable explanation for an intravenous bolus profile, while evidence of a changing terminal slope after intravenous dosing may motivate evaluation of distribution or time-varying disposition. A scale difference with parallel log-linear declines instead suggests a different kind of discrepancy, such as an incorrect dose or concentration scale.
Worked Example: Late Underprediction After an Intravenous Dose
Hypothetical data. After an intravenous dose, a one-compartment model predicts 4.0 mg/L at 4 hours and 1.0 mg/L at 12 hours. Observations are 4.0 mg/L and 1.4 mg/L. Assume the dose and sample times are verified and the model's predicted log-linear terminal phase is expected to apply at both times.
The predicted terminal rate constant from these points is \(\log(4.0/1.0)/(12-4)=\log(4)/8=0.1733\) per hour. The observed rate estimate is \(\log(4.0/1.4)/8=0.1312\) per hour. The observed concentration at 12 hours is therefore 0.4 mg/L above prediction, and the observed decline is shallower between the two times.
Because administration was intravenous, slow oral absorption cannot explain the tail. A missing distribution phase or a time-dependent disposition process remains plausible, among other possibilities. The two-point calculation does not distinguish these causes, and a single late observation could also reflect measurement error. Inspect denser early-to-late observations and other individuals: a reproducible transition from an early phase to a slower terminal phase supports testing a distribution model; evidence that the decline changes with time or clinical state warrants a different investigation. Do not simply lower clearance to fit one late point without checking whether that revision worsens the rest of the profile.
Dose-Related Errors: Exposure, Route, and Shape
A model may fit one dose range but systematically underpredict exposure at higher doses, or show errors that grow with dose. Saturable elimination is one candidate: when elimination capacity is approached, dose-normalized exposure can rise. For intravenous data, the ratio of dose to AUC is an apparent clearance summary; under linear PK with constant clearance, it should not change systematically with dose, assuming AUC is adequately measured and the same disposition conditions apply.
For oral data, dose-related exposure can also reflect a change in bioavailability. Other alternatives include dose-record errors, concentration-dependent residual error, or differences between dose groups in covariates or clinical state. Compare routes and profile shapes where available. A dose trend in intravenous data is more informative about disposition than a trend in oral data alone, but neither automatically proves saturation. A focused revision should be tested across the dose range and against the time course, not selected from an exposure summary alone.
Worked Example: Higher Dose-Normalized Exposure
Hypothetical data. Two intravenous dose groups have complete AUC measurements: 100 mg gives an AUC of 20 mg·h/L, and 200 mg gives an AUC of 55 mg·h/L. Assume dose records and AUC calculations are reliable and the groups are otherwise comparable for this illustration.
For 100 mg, dose/AUC is \(100/20=5.00\) L/h. For 200 mg, it is \(200/55=3.64\) L/h. The dose doubles, but AUC rises by a factor of \(55/20=2.75\); dose-normalized AUC rises from \(20/100=0.200\) to \(55/200=0.275\) h/L. These summaries are inconsistent with a single constant-clearance linear model if the measurements and comparability assumptions hold.
Because the doses are intravenous, dose-dependent oral bioavailability cannot account for this difference. Saturable elimination is a reasonable candidate to evaluate, but the two summaries alone do not establish it: uncertainty in AUC, different clinical conditions, or data problems could produce a similar contrast. Examine the full concentration-time profiles across dose levels and assess whether a nonlinear disposition model improves the description without creating systematic errors elsewhere.
Within-Subject Errors: Repetition Is Information
When an individual contributes observations on multiple occasions, ask whether errors are scattered or whether a whole occasion shifts in a consistent direction. Repeated positive errors over one occasion and repeated negative errors over another suggest an occasion-level discrepancy more than independent pointwise noise. Possible explanations include a changed dose or administration, an occasion-specific bioavailability or clearance change, or a clinical change not represented in the model. A single occasion is not enough to identify which one.
An interindividual random effect describes persistent differences between individuals; it does not automatically represent changes within the same individual over time. Conversely, occasion-level patterns are not sufficient grounds to add interoccasion variability before checking dose history, time-varying covariates, records, and residual-error behavior. The shape of the pattern matters: a broadly proportional shift with similar timing is different from an occasion-specific change in peak timing or terminal slope.
Worked Example: A Repeated Occasion-Level Shift
Hypothetical data. For one individual after two oral doses on separate occasions, predicted concentrations at 2, 4, and 8 hours are 8.0, 5.0, and 2.0 mg/L on each occasion. Observed concentrations on occasion A are 10.0, 6.2, and 2.5 mg/L; on occasion B they are 6.4, 4.0, and 1.6 mg/L. Assume the same recorded dose and sampling schedule, and initially use the predictions shown for both occasions.
The errors on occasion A are \(+2.0\), \(+1.2\), and \(+0.5\) mg/L. Those on occasion B are \(-1.6\), \(-1.0\), and \(-0.4\) mg/L. Every observation on A is above prediction and every observation on B is below it. The observed-to-predicted ratios are approximately 1.25 on A and 0.80 on B at all three times, so the discrepancy is close to a proportional shift rather than an obvious change in peak timing or decline.
This repetition supports investigating an occasion-level scale factor, such as occasion-specific relative bioavailability, after confirming the administration records and assay data. It does not prove that mechanism: dose delivery or another occasion-related cause could produce a similar pattern. If the discrepancy instead concentrated in the tail, a clearance or time-varying disposition explanation would deserve more attention. A persistent individual effect cannot by itself explain why the same individual shifts in opposite directions between occasions.
Choosing and Checking a Revision
A targeted revision should explain the pattern that prompted it while preserving adequate fit elsewhere. The following comparisons are useful because superficially similar residuals can imply different model changes.
| Pattern to explain | Competing explanations to check | Evidence that helps distinguish them |
|---|---|---|
| Early rise or peak mismatch | Absorption rate or lag time; multiple input phases; incorrect dose or sample timing | Repeated peak timing across profiles; verified records; shape of rise and decline |
| Late underprediction | Missing distribution phase; absorption-limited terminal slope; changing clearance | Route; early-to-late slope behavior; repeatability and clinical time course |
| Exposure error grows with dose | Saturable elimination; dose-dependent bioavailability; dose or group differences | Intravenous versus oral evidence; dose-normalized exposure; profile shapes |
| Errors shift by occasion within one individual | Occasion-level variability; changed administration or covariates; residual-model problem | Same-direction errors across an occasion; dose history; comparison of timing, slope, and scale |
After a proposed change, inspect predictions and residuals by time after dose, dose level, route, individual, and occasion, as appropriate. Check whether the targeted pattern diminishes and whether new patterns appear. Also consider whether the change is supported by the sampling design: parameters that cannot be distinguished with the available data should not be treated as established simply because a more flexible model fits better.
Common Pitfalls and Limits
- Assigning a mechanism from one residual. A single high late concentration does not establish a missing compartment. Look for a recurring, time-structured pattern and assess competing explanations.
- Confusing a scale error with a slope error. A profile uniformly high relative to prediction may differ from one whose log concentration declines too slowly. Compare profile shapes, not just the largest error.
- Calling every dose trend saturation. For oral dosing, changing bioavailability is a competing explanation. Check route and the dose-specific time courses.
- Treating occasion shifts as interindividual variability. Persistent differences between people and changes within the same person are different patterns; establish which is present before revising the random-effects structure.
- Overinterpreting standardized residuals. Standardization depends on the residual model. If that model is wrong, apparent changes in standardized errors may mislead; inspect raw concentration errors as well.
Fault isolation is limited by the information in the design. Sparse early samples may not locate a peak; few late samples may not characterize a terminal slope; dose groups that differ in clinical state may confound dose-related trends. When several explanations remain compatible with the data, report that uncertainty and identify what additional sampling or comparison would be informative rather than making an unsupported mechanistic claim.
Check Your Understanding
For each question, identify the strongest justified conclusion and the next useful check. Solutions follow each question.
- A model predicts 3 mg/L at a verified 1-hour sample and 6 mg/L at 2 hours. The observations are 4 and 5 mg/L. What profile feature is present, and what should be checked before revising absorption?
- After an intravenous dose, observed late concentrations repeatedly decline more slowly than predicted. Is slow oral absorption a viable explanation? Name one other candidate and one useful diagnostic.
- For intravenous doses of 50 and 100 mg, AUCs are 10 and 25 mg·h/L. Calculate dose/AUC at each dose and state what the comparison suggests under the stated assumptions.
- One individual's errors are mostly positive on one occasion and mostly negative on another. Why is adding an interindividual random effect not automatically the right response?
Solutions. (1) The model underpredicts early and overpredicts later, a timing or profile-shape mismatch. Verify dose and sample times and examine repeated profiles before comparing absorption revisions. (2) No: oral absorption cannot explain an intravenous profile. A missing distribution phase or changing disposition is possible; compare the early-to-late slopes across adequately sampled profiles. (3) The ratios are \(50/10=5.00\) L/h and \(100/25=4.00\) L/h. The lower dose/AUC at the higher dose is inconsistent with constant clearance if AUCs are reliable and conditions comparable; it motivates investigation, not a definitive saturation claim. (4) An interindividual effect represents persistent differences among people, whereas the observed direction changes within one person across occasions. Check occasion-specific records and covariates, and consider an occasion-level explanation only if the repeated pattern remains.