Why Terminal Patterns Deserve Separate Attention
A concentration-time profile can fit well over most of its range and still systematically miss the late observations. That matters because late concentrations help characterize the terminal decline, and a persistent prediction bias there may be consistent with a mismatch in the elimination rate. But a late residual pattern is a diagnostic clue, not a direct measurement of the cause: distribution behavior, absorption, residual-error structure, sampling, and the individual fit itself can all affect what the plot shows.
The first useful distinction is between a pattern in the mean of residuals and a change in their spread. A residual trend that grows in one direction across terminal time may indicate systematic prediction bias. Residuals that become more variable at higher concentrations, without a consistent signed trend, may instead reflect an inadequately represented concentration-dependent error variance.
An individual overlay places a subject's observed concentrations alongside that subject's predicted concentration-time curve. Reading the overlay together with residuals is important: the overlay shows the size and timing of the discrepancy in concentration units, while a residual plot makes its direction and persistence easier to compare across times or subjects. Here, “terminal” means the late portion of the profile in which the decline is approximately log-linear, after earlier absorption or distribution behavior no longer dominates the observed slope.
How a Terminal Residual Trend Relates to Elimination
A simple one-compartment model with first-order elimination has a log-linear post-input concentration profile. If \(A\) is the concentration scale at the start of the terminal segment and \(k_e\) is the elimination rate constant, then \(C(t)=A\exp(-k_e t)\). This relation is a useful local diagnostic approximation; it is not a claim that every late profile in a multicompartment model directly reveals the parameter called \(k_e\).
Suppose the observed terminal concentrations have effective scale \(A_{\mathrm{obs}}\) and decline rate \(k_{\mathrm{obs}}\), while the individual predictions have scale \(A_{\mathrm{pred}}\) and decline rate \(k_{\mathrm{pred}}\). Subtracting their log profiles gives
The intercept can reflect a difference in scale, but the slope of this log-residual trend is the difference between the predicted and observed terminal decline rates. If residuals become increasingly negative with time, the predicted decline is slower than the observed decline under this approximation. If they become increasingly positive, the predicted decline is faster. A roughly constant offset, by contrast, is not by itself evidence of a rate mismatch.
Worked Examples: Reading the Pattern in Context
Worked Example: A Terminal Decline That Is Faster Than Predicted
Consider a hypothetical subject whose observed concentration is \(10.0\ \mathrm{mg/L}\) at 4 hours. At 8 hours it is \(10\exp(-0.8383)\ \mathrm{mg/L}\), approximately \(4.324\ \mathrm{mg/L}\). Suppose the individual prediction is anchored at \(10.0\ \mathrm{mg/L}\) at 4 hours and uses a terminal rate of \(k_{\mathrm{pred}}=0.160\ \mathrm{h}^{-1}\). Assume for this calculation that both times are in a log-linear terminal segment, concentrations are positive, and there are no intervening doses.
The observed apparent decline rate between 4 and 8 hours is
The predicted concentration at 8 hours is \(10\exp[-0.160(8-4)]=10\exp(-0.64)\approx5.273\ \mathrm{mg/L}\). Therefore the log residual at 8 hours is
The residual is zero at the 4-hour anchor and negative at 8 hours: the model predicts too much concentration late in this interval. The observed apparent decline rate, \(0.2096\ \mathrm{h}^{-1}\), exceeds the predicted rate, \(0.160\ \mathrm{h}^{-1}\). That is consistent with a model terminal decline that is too slow. With only two points this is a descriptive comparison, not convincing evidence of a persistent residual trend; several informative terminal observations and inspection of other subjects would be needed.
Worked Example: Early Curvature Does Not Set the Terminal Rate
Suppose a hypothetical profile follows \(C(t)=100\exp(-1.0t)+20\exp(-0.20t)\ \mathrm{mg/L}\), with time in hours. The two declining components make the early profile curved on a log scale. Assume the late observations are sufficiently precise and the times are correctly recorded. At 4 hours, the concentration is approximately \(10.818\ \mathrm{mg/L}\); at 8 hours it is approximately \(4.071\ \mathrm{mg/L}\); and at 12 hours it is approximately \(1.815\ \mathrm{mg/L}\).
The apparent rate from 4 to 8 hours is
whereas from 8 to 12 hours it is
The later interval is much closer to the slow component's \(0.20\ \mathrm{h}^{-1}\) rate. Extrapolating that late component back to 4 hours gives \(20\exp(-0.20\times4)\approx8.987\ \mathrm{mg/L}\), below the observed \(10.818\ \mathrm{mg/L}\), because the faster component still contributes at 4 hours. Thus a discrepancy that is largest earlier and diminishes in the late segment is not the same pattern as a persistent terminal rate mismatch. The overlay and residuals should be read by phase; an early deviation should not be “fixed” by changing the terminal rate without evidence from the late observations.
Worked Example: Concentration-Dependent Error Without a Terminal Mean Trend
Suppose hypothetical individual predictions are \(10.0,\ 5.0,\ 2.0\ \mathrm{mg/L}\), and observations are \(11.0,\ 5.5,\ 2.2\ \mathrm{mg/L}\), at successively later times. Assume the values are positive and the specified error model is intended to allow variability proportional to concentration. The raw residuals are \(1.0,\ 0.5,\ 0.2\ \mathrm{mg/L}\), which shrink as concentrations shrink. However, the relative residual is \(0.10\) at each time, and each log residual is
There is no progressive signed trend in these log residuals: all three observations are about 10% above their predictions. The smaller raw residuals at lower concentrations do not show that the elimination rate is changing. In a proportional-error setting, assess residuals on a scale that reflects the error model; a pattern in residual spread can otherwise be mistaken for time-dependent structural bias. Conversely, a consistent increase or decrease in signed, appropriately scaled residuals through the terminal period warrants investigation of the predicted profile, not just its variance.
A Practical Diagnostic Sequence
A disciplined reading keeps the time pattern, concentration scale, and individual profile connected. Use individual predictions to ask whether the mismatch is specific to a few subjects or repeated across them; use residual plots to assess whether the discrepancy is signed and progressive. A population-level trend can be obscured by heterogeneous subject fits, while one unusual individual should not be generalized to the full population without checking influence and data quality.
Check dose times, sampling times, and the profile's shape. Do not treat an interval dominated by absorption or visible curvature as a reliable terminal slope.
Determine whether observations repeatedly fall above or below the individual curves, and whether the signed discrepancy grows through late time rather than appearing as a constant offset.
Inspect residuals on a scale appropriate to the residual-error model and compare subjects with adequate late sampling. Look for concentration-related spread that could arise from the variance model.
Review distribution curvature, absorption timing, assay or study structure, dose history, and whether a small number of influential subjects dominate the apparent trend.
This interpretation complements earlier material on residual plots by time, individual overlays, and proportional or combined error. Compared with relying only on a numerical goodness-of-fit summary, a time-ordered overlay can expose where the model misses and whether that miss is terminal. Compared with a residual plot alone, the overlay preserves the concentration scale and makes it easier to notice phase-specific curvature. Neither display replaces checking data and model assumptions, and neither by itself selects a new structural model.
Assumptions, Failure Modes, and Limits
The simple rate interpretation requires a meaningful terminal segment: the observed decline should be approximately log-linear over the selected interval, and the compared predictions should refer to the same subject, dose history, and time scale. For a one-compartment model after input has ended, the terminal slope corresponds to the elimination rate constant. In multicompartment models, however, the terminal observed slope may be a hybrid disposition rate rather than the structural elimination parameter itself. A terminal residual trend can therefore identify prediction bias without uniquely identifying which model parameter should change.
Oral profiles require particular caution. If absorption is slow relative to elimination, the late observed decline can be governed by absorption rather than by the elimination process; a terminal slope alone cannot separate those mechanisms. Likewise, an apparent late trend may result from incorrect dose or sample times, an unmodeled additional dose, assay effects, or values near a quantification limit. Sparse late sampling makes the estimated slope unstable, and shrinkage or a small number of observations can make individual predictions and residuals less informative.
| Observed display pattern | Interpretation to investigate | Important caution |
|---|---|---|
| Signed log residuals progressively negative late | Predicted decline may be slower than observed | Confirm a genuine terminal segment and adequate late data |
| Signed log residuals progressively positive late | Predicted decline may be faster than observed | Check input timing and competing explanations before changing elimination |
| Largest discrepancy early, fading later | Earlier distribution or input behavior may dominate | Do not use the early slope as the terminal rate |
| Raw residual spread changes with concentration but relative or scaled residuals do not show a signed trend | Error variance may depend on concentration | Variance pattern alone does not establish structural mean bias |
A common mistake is to see several negative late residuals and conclude, “elimination is too slow, so increase the elimination parameter.” The sign alone does not justify that conclusion: a constant negative offset may reflect scale, while a late pattern may be driven by distribution, absorption, data timing, or a misspecified error model. The correction is to establish the direction of the trend across a credible terminal interval, compare observed and predicted decline rates, check other subjects and residual scales, and describe the result as evidence consistent with a mismatch rather than proof of its cause.
Check Your Understanding
Answer each question, then compare with the solutions.
- If observed concentrations decline faster than individual predictions in a credible log-linear terminal segment, what direction of trend is expected in log residuals defined as log observed minus log predicted?
- A profile is strongly curved from 1 to 4 hours but nearly log-linear from 8 to 16 hours. Which interval is more appropriate for a terminal-rate comparison, and why?
- Raw residuals get smaller as concentration decreases, but relative residuals are roughly constant and show no signed time trend. Does this alone support an elimination-rate mismatch?
- Why can a late observed slope in an oral profile fail to identify the elimination rate?
- Two subjects show a terminal trend, while most adequately sampled subjects do not. What should be checked before making a population-wide conclusion?
Solutions. (1) The log residuals should trend downward, becoming more negative, because \(k_{\mathrm{pred}}-k_{\mathrm{obs}}<0\) when the observed decline is faster. (2) The 8-to-16-hour segment is the better candidate because its near log-linearity is more consistent with terminal decline; the earlier curved interval may reflect other profile phases. (3) No. Constant relative residuals with no signed time trend can be compatible with proportional error; the change in raw residual size alone does not establish a terminal mean bias. (4) Slow absorption can govern the observed late decline, so the terminal slope may reflect absorption rather than elimination. (5) Check data timing and dose history, influential-subject effects, late sampling adequacy, individual fit uncertainty, and whether the pattern is specific to a covariate, study, or occasion before generalizing.