Diagnostics for Misspecified Adaptation Models

Why Adaptation Models Need More Than a Good Overall Fit

An adaptation model can reproduce an average response level while missing the timing of tolerance, compensation, recovery, or rebound. This matters because the same overall error can arise from very different failures: an omitted rebound process may create errors mainly after withdrawal, whereas an incorrect recovery rate may cause errors to accumulate throughout washout. Looking only at an overall fit statistic can conceal both patterns.

Diagnostics compare observations with predictions in ways that retain information about exposure and time. A response-versus-exposure plot asks whether the exposure-response relationship changes systematically across periods or adaptation states. Time-stratified residuals ask whether errors cluster in particular phases. Observed-predicted trajectories show whether the model gets the direction, timing, and duration of changes right. These tools reveal patterns to investigate; none, by itself, proves a biological mechanism.

The diagnostics assume that the observation and prediction refer to the same response definition, time, dose or exposure history, and relevant covariates. Predictions must be calculated using the fitted model and the intended state initialization. If these are mismatched, the resulting residual pattern may reflect a data-processing error rather than a misspecified adaptation structure.

Response Versus Exposure: Look for Time-Dependent Structure

A simple tolerance model can be represented as a drug effect modified by an adaptation state. For illustration, let \(C(t)\) be concentration, \(A(t)\) a nonnegative tolerance state, and \(R(t)\) the response. One possible inhibitory model is

Here \(R_0\) is the untreated response, \(E_{\max}\) is the maximum drug effect in the absence of adaptation, and \(EC_{50}\) is the concentration giving half that effect under this model. At the same concentration, a larger \(A(t)\) attenuates the drug effect, so the predicted response is closer to \(R_0\). The model is only an illustrative structure: actual response direction, link function, and adaptation dynamics must match the application.

A plot of response against concentration can suggest a changing exposure-response relationship. If observations at similar concentrations show systematically different responses during onset, continued dosing, or washout, a model using concentration alone may be inadequate. With a tolerance state, the difference can be expected: \(C(t)\) may be the same while \(A(t)\) differs because the state reflects prior exposure. The resulting loop over time is often called hysteresis.

Therefore, a loop is not automatically evidence of tolerance. A delay between concentration and effect, a changing disease state, measurement timing, or other omitted covariates could also produce time-dependent separation. Marking observations by time or dosing phase helps locate the pattern, but interpretation requires the model assumptions and study design.

Time-Stratified Residuals Locate Model Failure

If the model describes the conditional mean adequately and the residual errors have mean zero, then residuals should not show a systematic mean pattern across time or dosing phase. This gives a basic justification for grouping residuals by meaningful periods. For a stratum \(g\), such as early washout, the average residual is

If \(\bar{e}_g\) is consistently positive, the model tends to underpredict in that stratum; if consistently negative, it tends to overpredict. This is a diagnostic summary, not a hypothesis test by itself. Repeated measurements from the same subject may be correlated, and an average can hide a changing pattern within a stratum. Examine the individual residuals and their timing as well.

Observed-Predicted Trajectories Reveal Timing Errors

Plot observed responses and model predictions against time on the same axes, using the same response scale. A useful trajectory comparison should preserve the dosing and withdrawal times and should make clear whether predictions are population-level or individual-level. A model may match the overall response range but still predict the peak too early, recover too quickly, or miss a post-withdrawal overshoot.

For tolerance-recovery assumptions, inspect the washout trajectory rather than only the response during repeated dosing. If the model assumes exponential recovery, for example, the predicted adaptation state declines smoothly toward its untreated level. Systematic residuals that change sign or magnitude during washout can indicate that a single recovery rate does not represent the observed course. Check the observed and predicted trajectories together: the same residual sequence may be difficult to interpret without the response scale and the timing of the dose changes.

A Practical Diagnostic Sequence

Assumptions, Alternatives, and Failure Modes

These diagnostics are most interpretable when observation timing is adequate, the response is measured consistently, and the model predictions account for relevant dosing and exposure history. A response-versus-exposure plot can be misleading if concentration measurements are sparse or not aligned with response times. Time-stratified residuals can be misleading if strata were selected after seeing the pattern, if one phase has very few observations, or if subject-level correlation is ignored. Trajectories can appear to fit because of a few well-sampled periods while missing a short rebound between observations.

Visual diagnostics complement, rather than replace, quantitative model assessment. A residual summary compresses information and can conceal individual trajectories. A single overall fit measure compresses it further and can reward a model that predicts average levels while missing phase-specific behavior. Formal comparisons of candidate models can help evaluate whether added structure improves fit, but a better fit alone does not establish biological truth. The choice of alternative should also reflect mechanism, parameter identifiability, and whether the design contains observations capable of distinguishing competing explanations.

A common alternative is to fit a more flexible time trend or a separate effect for each phase. Such descriptions may reveal unexplained structure, but they do not necessarily represent a transferable adaptation mechanism. A mechanistic state-variable model can encode prior exposure and recovery, but may be harder to identify when the data do not sample the relevant transitions. The practical goal is not to add complexity whenever residuals are imperfect; it is to determine whether the pattern is reproducible, consequential, and addressed by a defensible model change.

• Mistake: treating any response-exposure loop as proof of tolerance. A loop only shows that response is not described by the plotted concentration alone. Delay, disease progression, measurement timing, or other omitted variables may also matter.
• Mistake: interpreting one large residual as a structural failure. A single residual can arise from ordinary variability. Seek a repeated, phase-related pattern and consider the error scale.
• Mistake: relying only on overall fit. Good average agreement does not rule out systematic underprediction during withdrawal or an incorrect recovery time course.
• Mistake: equating an improved fit with a confirmed mechanism. Improvement supports the candidate as a description of these data; it does not by itself show that the proposed biological process is the cause.