Identifiability of Tolerance and Recovery Parameters

Why Identifiability Matters

A tolerance model may fit the observed response closely and still fail to determine which biological parameters produced it. For example, a small response during sustained dosing could reflect strong adaptation, slow recovery, low drug sensitivity, or a changing baseline. These explanations imply different responses to a new dose or a washout period, so distinguishing them matters for interpretation and prediction.

Identifiability concerns what the chosen model and observations can determine. It is not the same as goodness of fit: several parameter combinations can fit equally well. Nor is it the same as precision: collecting more observations of an uninformative pattern may narrow measurement error without separating the parameters.

A Minimal Adaptation Model

Consider a dimensionless adaptation state \(A(t)\), driven by a known, dimensionless exposure input \(u(t)\). A first-order induction-and-recovery model is

Here \(k_{\mathrm{ind}}>0\) and \(k_{\mathrm{rec}}>0\) are rates, both in inverse time. The state \(A(t)\) represents adaptation, not a directly observed response. For illustration, suppose the measured response is

In this observation model, \(B_0\) is baseline response, \(b\) is linear baseline drift per unit time, \(C(t)\) is measured drug concentration, and \(\beta\) is response per concentration in the absence of adaptation. The parameter \(\gamma\) controls how strongly adaptation attenuates the drug response. The residual \(\epsilon(t)\) is measurement error. This simplified model assumes known exposure, a linear concentration-response relation over the observed range before adaptation, a multiplicative sensitivity effect, and a linear baseline drift over the study window. Those assumptions must be credible for the identifiability conclusions to apply.

When \(u(t)=u_0\) is constant on an interval beginning at \(t_0\), solving the state equation gives

This expression separates two features of the trajectory. The asymptotic state under constant input is determined by the ratio \(k_{\mathrm{ind}}u_0/k_{\mathrm{rec}}\); the rate of approach to it is determined by \(k_{\mathrm{rec}}\). Observing only the asymptote therefore discards information about the time scale.

What the Data Can Separate

A useful way to diagnose confounding is to ask which features of the trajectory respond differently to each parameter. The initial rise during exposure, the later plateau, the decay after withdrawal, and the response to a controlled rechallenge do not carry the same information. The following justification makes the role of time-course data explicit.

During a known constant input \(u_0>0\), with \(A(t_0)=0\), the solution is

As \(\Delta\) grows, \(e^{-k_{\mathrm{rec}}\Delta}\) tends to zero, leaving the plateau \(A_\infty=k_{\mathrm{ind}}u_0/k_{\mathrm{rec}}\). Multiplying both rates by the same positive constant leaves this plateau unchanged. Thus plateau-only observations cannot distinguish the two rates. In contrast, the transient term \(e^{-k_{\mathrm{rec}}\Delta}\) changes when \(k_{\mathrm{rec}}\) changes, so observations during the rise can identify the time scale if the transient is adequately sampled and the state scale is known or separately constrained.

After withdrawal, \(u(t)=0\), so for washout time \(\Delta\),

If the adaptation state is measured, two nonzero state values at distinct washout times identify the recovery rate through their ratio. If only drug response is observed, however, the state is latent and the response is also affected by \(\gamma\), concentration, and baseline. A washout with no drug may yield little or no information about tolerance recovery unless an outcome linked to the state is measured during washout or a standardized rechallenge is given.

The observation model adds further confounding. At a single concentration and adaptation state, the drug-related response depends on \(\beta\exp(-\gamma A)\), not on \(\beta\) and \(\gamma\) independently. If adaptation is small, \(\exp(-\gamma A)\) is approximately \(1-\gamma A\). Under sustained exposure, \(A\) approaches \(k_{\mathrm{ind}}u_0/k_{\mathrm{rec}}\), so the attenuation can depend largely on the product \(\gamma k_{\mathrm{ind}}/k_{\mathrm{rec}}\). Independent information about sensitivity, or observations across changing states and concentrations, is needed to separate these contributions.

Baseline Drift and Drug Sensitivity

Baseline drift can mimic a slowly changing drug effect when concentration and time move together. In a locally linear approximation, suppose the observations satisfy \(Y_i=B_0+bt_i+\theta x_i+\epsilon_i\), where \(x_i\) is a known exposure-related predictor and \(\theta\) is its response coefficient. The drift and exposure effect are estimable separately only if the columns for time and exposure are not perfectly collinear (and are sufficiently distinct in noisy data). A fitted line alone cannot reveal which component caused a trend if both predictors have the same pattern.

In the nonlinear model, an unchanging \(C(t)\) during adaptation may also leave \(\beta\) and \(\gamma\) strongly dependent. A single observed response level can be explained by different combinations of baseline, sensitivity, and adaptation. Sampling multiple concentrations at similar adaptation states helps characterize sensitivity; comparing controlled rechallenges at different washout durations helps characterize recovery. Untreated or placebo observations taken over the same calendar time help estimate baseline drift rather than attributing it to drug action.

Designing Informative Exposure and Sampling Patterns

Useful designs create observations in which candidate parameters have different consequences. No single schedule guarantees identifiability, but several principles follow directly from the model:

• Sample the transient, not just the plateau. Observations during induction contain information about the approach rate. Sampling only after steady state mainly informs a rate ratio.
• Include washout and appropriately timed rechallenges. A rechallenge at a fixed concentration after several washout durations can reveal recovery through the changing drug response, provided baseline and exposure are controlled.
• Vary exposure independently of calendar time. Staggered starts, changing input levels, or concurrent controls can reduce alignment between exposure and baseline drift.
• Use more than one concentration when sensitivity is uncertain. Concentration variation can help distinguish a change in drug sensitivity from a time trend, provided the concentration-response form is credible.
• Measure or constrain a hidden component where possible. An independent state marker, sensitivity estimate, or informative control can reduce reliance on the same response curve to estimate every parameter.

Sampling times matter as much as the number of samples. If all induction samples occur long after the approach to plateau, they add little information about the induction time course. If all rechallenges occur after recovery is essentially complete, they cannot distinguish fast from moderately fast recovery. Conversely, extremely early samples may be insensitive to a slow process. Candidate schedules should be assessed against plausible parameter values before data collection.

Diagnostics, Alternatives, and Limitations

Signs of weak identifiability include very large standard errors, strong parameter correlations, estimates that change substantially with initial values, flat likelihood or objective-function profiles, and different parameter sets with nearly identical fitted curves. These are clues about information in the design, not by themselves proof of a particular biological mechanism. Simulating responses under alternative parameter combinations is a direct check: if they are indistinguishable at the planned sampling times, the design cannot reliably separate them.

Several analysis strategies answer complementary questions. Algebraic inspection of the model exposes exact confounding, as in a plateau that depends only on a ratio. Sensitivity or information-matrix methods assess local information around a candidate parameter value. Profile likelihood examines whether a parameter can vary while the fit remains nearly unchanged. Design changes—such as adding a control, transient samples, or rechallenge times—can supply information that a fitting method cannot create.

Fixing a parameter using external evidence can make other parameters estimable, but the resulting estimates are conditional on that fixed value and its relevance to the study. A prior distribution can regularize an estimate, yet a narrow posterior may reflect prior information rather than identification by the observations. Report that distinction. Likewise, a low residual error does not establish that induction, recovery, and sensitivity have each been estimated reliably.

These conclusions are conditional on the stated model. Nonlinear concentration-response behavior, more than one adaptation process, or an inappropriate linear drift assumption can change the apparent parameter trade-offs. The objective here is to diagnose whether the chosen design separates parameters under its assumptions, not to claim that one model structure is uniquely correct.