Why a PD Delay Can Leave a Residual Signature
A PD model may predict response as an immediate function of the current concentration. If the response changes more slowly than concentration, that representation can be wrong even when the concentration profile is well described. One clue is a systematic change in the sign of the response errors over the dosing trajectory: predictions may be too high while concentration is rising and too low after concentration has begun to fall. The pattern is informative because it links errors to the timing of exposure, not merely to their overall size.
A residual pattern is evidence of a mismatch, not a direct measurement of a biological delay. Baseline errors, an incorrect concentration-effect shape, time trends, or other model deficiencies can also produce structured errors. The diagnostic question here is narrower: do response errors depend on the position in the concentration-time trajectory in a way consistent with an inadequately represented delay?
Definitions and Conditions for Interpreting the Pattern
The most useful comparison is often between observations at similar concentrations on the rising and falling portions of a concentration-time curve. In a direct-effect model, the same current concentration generally implies the same mean effect, given the model's other predictors. If the actual response lags exposure, the response at a given concentration can differ by trajectory phase. For a response that increases with exposure, it may still be relatively low while concentration is rising, then remain relatively high while concentration is falling.
Interpret this comparison only after checking that the predictions and residuals are appropriate for the data structure. The response scale, variance model, baseline terms, subject-level predictions, and sampling schedule all matter. A signed residual on the original scale may be difficult to compare across observations if variability changes with the predicted response; a variance-standardized residual can help in that situation. Residuals from individual conditional predictions and from population-level predictions answer different questions, so a pattern should be checked at a level that matches the diagnostic aim.
Why Phase-Specific Mean Residuals Can Indicate Misspecification
Write the true conditional mean as \(m_{\text{true},ij}\), and express the observation as \(Y_{ij}=m_{\text{true},ij}+\epsilon_{ij}\), where \(\epsilon_{ij}\) is the remaining observation error. The fitted residual is then the difference between the true mean and the fitted mean, plus observation error. Taking a conditional expectation by phase gives the key justification:
If \(E(\epsilon_{ij}\mid\text{phase})=0\), the phase-specific mean residual estimates the phase-specific average discrepancy between the true and fitted means. It does not generally equal one unaveraged difference \(m_{\text{true}}-m_{\text{fit}}\): those means can vary across concentration, subject, covariates, and time even within the same phase. A practical estimate is the average of the observed residuals in that phase, interpreted with those differences in composition in mind.
For a direct-effect model that predicts the same response at a matched concentration on both the rising and falling limbs, a delayed response can create a phase contrast: residuals tend to be negative on the rising limb and positive on the falling limb when larger exposure increases response. A repeatable contrast after accounting for concentration and other predictors supports investigating the delay representation. It does not by itself identify the biological process responsible.
Worked Example: Matched Concentrations on Rising and Falling Limbs
Suppose a hypothetical drug has a PD response that increases with exposure. A fitted immediate concentration-effect model predicts responses of 2, 4, and 6 response units at concentrations of 1, 2, and 3 mg/L, respectively. The following hypothetical observations were collected at those same concentrations on both limbs of the concentration-time curve.
| Concentration (mg/L) | Predicted response | Observed rising limb | Rising residual | Observed falling limb | Falling residual |
|---|---|---|---|---|---|
| 1 | 2.0 | 1.4 | -0.6 | 2.6 | 0.6 |
| 2 | 4.0 | 3.3 | -0.7 | 5.0 | 1.0 |
| 3 | 6.0 | 5.2 | -0.8 | 7.1 | 1.1 |
At each concentration, calculate the signed residual as observed minus predicted. For example, at 2 mg/L on the rising limb, \(e=3.3-4.0=-0.7\) response units; on the falling limb, \(e=5.0-4.0=1.0\) response units. The average rising-limb residual is \((-0.6-0.7-0.8)/3=-0.700\) response units. The average falling-limb residual is \((0.6+1.0+1.1)/3=0.900\) response units, giving a phase contrast of \(0.900-(-0.700)=1.60\) response units.
The sign change and the positive falling-minus-rising contrast are consistent with delayed response for an effect that increases with exposure: the immediate model overpredicts on the way up and underpredicts on the way down. Because this small hypothetical table has no uncertainty calculation and may not represent other subjects or intervals, it is a diagnostic illustration, not a formal test or proof of a delay.
Compare Across Dosing Intervals, Not Just Within One Curve
A single interval can be misleading. Uneven sampling, an unusual concentration profile, or one subject's influential observations may create an apparent phase contrast. If repeated intervals are available, examine whether a comparable residual pattern recurs. Also check whether the contrast remains when comparing observations with similar concentrations, dose history, and relevant subject characteristics. A late observation after repeated dosing is not necessarily comparable with a late observation after the first dose if accumulation or the preceding exposure history differs.
Worked Example: Checking Repeatability Across Two Intervals
In a hypothetical study, an immediate-effect model is used for a response that increases with exposure. For two dosing intervals, the fitted model's residuals at comparable early post-dose times are \(-0.8\) and \(-1.2\) response units. At comparable late, falling-concentration times, the residuals are \(1.4\) and \(0.6\) response units. Assume these are signed residuals computed as observed minus predicted.
The early mean is \((-0.8-1.2)/2=-1.0\) response unit, and the late mean is \((1.4+0.6)/2=1.0\) response unit. The late-minus-early contrast is \(1.0-(-1.0)=2.0\) response units. Both intervals show negative early residuals and positive late residuals, so the direction of the pattern repeats.
This repetition makes a phase-related mismatch more plausible than a pattern driven by one interval alone. It still does not establish that a delay is the only cause. The intervals must be comparable in the relevant predictors, and residual dependence across observations must be considered when assessing the uncertainty of a summary. Repeated points from one subject do not provide the same information as independent subjects.
Phase Averages Are Averages Over Observations
Worked Example: Avoiding an Incorrect Conditional-Mean Shortcut
Consider two hypothetical rising-phase observations. For them, the true-minus-fitted mean discrepancies are 1 and 3 response units, and the observation errors are \(0.2\) and \(-0.2\) response units. The residuals are therefore \(1+0.2=1.2\) and \(3-0.2=2.8\) response units. Their mean residual is \((1.2+2.8)/2=2.0\) response units.
With errors averaging zero in this phase, the conditional mean residual corresponds to the average true-minus-fitted discrepancy, \((1+3)/2=2.0\) response units. It is not equal to either observation's discrepancy individually. More generally, the valid relationship is \(E(e\mid\text{phase})=E(m_{\text{true}}-m_{\text{fit}}\mid\text{phase})\) when the phase-conditional mean error is zero. Replacing the right-hand side with an unaveraged difference is unjustified unless the means are fixed within the conditioning set.
Diagnostics, Alternatives, and Limitations
Use residuals against time and concentration alongside phase comparisons. A useful display can distinguish rising from falling observations, mark dose times, and show the concentration trajectory. If the model represents delay adequately, a systematic early-negative/late-positive pattern should be reduced, not merely shifted to a different time region. Consider whether the pattern occurs across subjects or is concentrated in a few; whether it persists at similar concentrations; and whether it appears in both individual and population-oriented checks. Residual autocorrelation can make a smooth run of signs look more persuasive than its effective information warrants.
Other explanations should be considered before changing the delay structure. An incorrect concentration-effect curve can leave residuals that vary with concentration. A misspecified baseline or placebo component can create time-related errors; a changing response unrelated to concentration can resemble a phase pattern. Incorrect exposure predictions can also displace the apparent timing of the effect. These alternatives matter because the residual pattern identifies where predictions fail, not which parameter or mechanism must be changed.
The main alternatives to a direct concentration-effect representation include models that allow an indirect time course, such as an effect compartment or a response-turnover structure. These structures encode different assumptions about how response follows exposure. This diagnostic does not select among them: a phase contrast can motivate checking candidate representations, while the scientific mechanism, data density, parameter identifiability, and predictive performance determine which are defensible. Very sparse sampling, little overlap in concentration between rising and falling phases, or substantial residual noise can make the phase comparison inconclusive.
A common mistake is to say, “The mean residual is positive late, so the drug must have a delayed effect.” The correction is that positive late residuals show average underprediction in the sampled late-phase observations; delay is one possible explanation. Check the corresponding early-phase errors, concentration matching, interval repeatability, baseline and time structure, and uncertainty before interpreting the pattern. A lack of a visible contrast is not conclusive evidence of no delay when the sampling design has little power to reveal it.
Mark rising and falling concentration observations within dose intervals, and state whether increasing exposure is expected to increase or decrease the response.
Use observed minus predicted response and choose a residual scale that accounts for the observation-error model.
Inspect matched or similar concentrations and relevant predictors; summarize phase averages without treating them as a single fixed mean difference.
Look for repetition across intervals and subjects, and assess time, baseline, concentration-shape, exposure, and serial-dependence explanations before concluding that the delay representation is inadequate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- An increasing response has an average early residual of \(-0.5\) response units and an average late residual of \(0.7\) response units. What is the late-minus-early contrast, and what pattern does it suggest?
- Why is \(E(e\mid\text{phase})\) not generally equal to one unaveraged value of \(m_{\text{true}}-m_{\text{fit}}\)?
- A positive late residual appears in one interval, but early residuals are near zero and the pattern does not recur. What is a cautious interpretation?
- Why can matching observations on concentration help diagnose a delayed response?
Solutions. 1. The contrast is \(0.7-(-0.5)=1.2\) response units. It is a negative-early, positive-late pattern consistent with under-represented delay for an increasing response, but it does not prove that explanation. 2. The conditional expectation averages over observations in the phase, and the true and fitted means can vary across concentration, subjects, and covariates. Under zero conditional-mean observation error, the result is \(E(m_{\text{true}}-m_{\text{fit}}\mid\text{phase})\). 3. The model underpredicts on average for that sampled late phase, but the evidence for a systematic phase signature is weak; examine sampling, predictors, and alternative time-related explanations. 4. A direct-effect model may assign the same prediction to the same concentration on both limbs, whereas a delayed response can differ by trajectory phase. Matching concentration helps separate this phase contrast from an error caused only by different concentration values.