Residual Signatures of Tolerance and Rebound

Why repeated dosing and washout matter

A model can describe the first dose adequately yet systematically miss later responses. If the drug effect becomes smaller during repeated exposure, a model with an unchanged exposure-response relationship may predict too much effect on later cycles. After dosing stops, the observed response may also move beyond baseline in the direction opposite to the drug effect. A model that only allows the drug effect to fade back to baseline may then miss a rebound pattern.

Residuals help identify these mismatches, but a pattern is a diagnostic clue, not proof of a biological mechanism. The same residual sequence can arise from incorrect exposure predictions, an inadequately represented baseline or time trend, delayed response, or unmodeled dependence. The key is to compare observations at relevant times and exposures, and to check competing explanations before attributing a pattern to tolerance or counter-regulation.

For an endpoint that increases under treatment, \(q\) can be the response minus baseline. For an endpoint that decreases under treatment, \(q\) can be baseline minus response. Apply the same direction convention to observed and predicted values. This avoids a sign error: a negative residual in the direction-coded response means the predicted drug effect is larger than the observed drug effect, regardless of whether the raw endpoint rises or falls with treatment.

Residual signatures across dosing cycles

In a model without adaptation, the predicted effect at a given exposure is expected to be similar across dosing cycles, after accounting for the model's specified time course, baseline, and other covariates. A possible tolerance signature is therefore a shift toward negative residuals on later cycles at comparable exposure: the model predicts an effect that is too large relative to what is observed. The comparison should be made at comparable concentrations or exposure histories, not merely at the same clock time after each dose.

A useful conceptual decomposition makes the sign clear. Suppose that on the direction-coded scale the fixed model predicts \(q_{\text{fixed}}\), while an unmodeled attenuation \(A\) reduces the actual drug effect. If \(A\) is positive, then the observed response can be written as \(q_{\text{obs}}=q_{\text{fixed}}-A+\epsilon\), where \(\epsilon\) is observation error. The residual is therefore \(e=q_{\text{obs}}-q_{\text{fixed}}=-A+\epsilon\). Across observations, attenuation can produce a negative residual trend if it is not overwhelmed by noise or other model errors. This algebra explains the expected direction; it does not establish that tolerance caused the trend.

The relevant comparison is often within an individual across cycles. A population average can conceal opposite individual patterns, while an individual plot can make a cycle-related shift visible. However, repeated measurements from one person are not independent replicates. A run of negative residuals should not be treated as a collection of independent confirmations; account for residual scale and serial dependence when judging how persuasive the pattern is.

Washout residuals and rebound

Washout provides a second informative phase. Under a model in which the drug effect simply disappears as exposure declines, the direction-coded response is predicted to return toward its baseline value, usually zero after baseline adjustment. A rebound pattern occurs when the response moves beyond baseline in the direction opposite to the drug effect. Relative to a model predicting return to baseline, that opposite-direction response gives a negative residual on the direction-coded scale.

A possible mechanism is counter-regulation: a process opposing the drug effect develops during treatment and remains temporarily active as drug exposure falls. A simple representation is \(q_{\text{obs}}=q_{\text{drug}}-R+\epsilon\), where \(R\) is an opposing response expressed in direction-coded units. If the model predicts only \(q_{\text{drug}}\), then \(e=-R+\epsilon\). During washout, the drug contribution may become small while \(R\) persists, making the observed response negative relative to baseline and the residual negative relative to the model prediction. This is a mechanistic interpretation to investigate, not something that residual sign alone identifies.

Timing matters. A negative residual shortly after the last dose may instead reflect delayed drug effect, an incorrect effect-compartment or turnover time course, or an incorrect baseline trajectory. Earlier material on delayed-effect misspecification and turnover and disease-progression misspecification addresses these broad alternatives. For this diagnostic, inspect whether the residual pattern is specifically linked to repeated exposure or washout after the model's stated time course and baseline have been checked.

Check exposure and scale before assigning a PD explanation

An apparent later-cycle PD mismatch can be induced by an exposure error. If the model uses a concentration prediction that is too high, it may overpredict effect even when the response at the actual concentration follows the assumed PD relationship. The following hypothetical calculation separates these two residual comparisons.

Residual scale also matters. With heteroscedastic error, equal-sized raw residuals can have different significance depending on the expected variability at that response or exposure. When a model specifies a conditional standard deviation, a standardized residual can be expressed as \(r_i=e_i/SE_i\), where \(SE_i\) is the model's standard deviation for observation \(i\). Use the error structure appropriate to the response scale; residuals for transformed or bounded responses require the corresponding diagnostic scale. Earlier material on heteroscedastic PD error and residuals for transformed and bounded responses covers these considerations.

A practical diagnostic sequence

Comparison, failure modes, and limitations

A cycle-stratified residual plot is especially useful for seeing systematic shifts across repeated dosing. A residual-versus-time plot shows when the shift emerges, while a residual-versus-concentration plot asks whether it remains at comparable exposure. Observed-predicted plots can show calibration problems but may obscure a cycle-specific pattern if all phases are pooled. Predictive checks can compare the observed sequence with replicated sequences from the fitted model, but they do not by themselves identify tolerance or rebound. These approaches complement one another: no single plot establishes a biological cause.

Several failure modes deserve particular attention:

• Calling any negative residual tolerance. A negative residual only says the observation is below prediction. Compare exposure, baseline, response timing, and error scale before proposing adaptation.
• Comparing cycles at the same clock time only. Concentrations may differ across cycles because of accumulation or changing exposure. A valid comparison needs comparable exposure or an explicit model that accounts for it.
• Ignoring the response direction. For a drug that lowers a raw endpoint, a negative raw-scale residual does not automatically mean the drug effect was overpredicted. Define the direction-coded scale first.
• Calling return toward baseline rebound. A response that merely approaches baseline is not an opposite-direction overshoot. Rebound requires a response beyond baseline in the direction opposite to the drug effect, interpreted against the model's prediction.
• Treating residual runs as independent evidence. Serially correlated observations can produce runs of the same sign. Consider the dependence structure and inspect the individual and population patterns together.

Residual signatures have important limits. Their visibility depends on sampling during the relevant cycles and washout, reliable exposure information, adequate baseline characterization, and enough repeated observations to distinguish a pattern from noise. Sparse data can miss transient attenuation or rebound. Between-individual variation can also make a population-level average misleading. Finally, several biological processes can produce similar residuals; a proposed adaptation model should be evaluated against the full time course and scientifically plausible alternatives, rather than selected from residual sign alone.