Weak Identification Near Dose-Response Plateaus

Why a Plateau Can Leave the Target Dose Uncertain

In dose optimization, a fitted dose-response curve may appear to approach a clear maximum while leaving the dose needed to attain a high target poorly determined. The reason is geometric: close to the plateau, a large change in dose can produce only a small change in predicted response. Data can therefore support a broad range of doses as plausible solutions to the same target, even when the curve's maximum response seems reasonably well estimated.

The issue is weak identification: the observed data do not sharply distinguish among values of the target-attaining dose. It is not the same as saying that no estimate can be calculated. A point estimate may exist, but small changes in the target or fitted curve can move it substantially, and its uncertainty interval may be wide. This tutorial uses the Emax model and focuses on diagnosing that behavior, rather than choosing a dose under decision constraints.

The Emax Dose and Its Sensitivity Near the Plateau

Consider the increasing Emax model

where \(d\) is dose, \(E_0\) is the mean response at zero dose, \(E_{\max}>0\) is the additional response at the plateau, and \(ED_{50}>0\) is the dose producing half of that additional response. The asymptotic mean response is \(E_0+E_{\max}\). For this model, assume the mean structure is appropriate over the studied doses and the response observations are modeled with a likelihood whose parameters can be estimated from the data. The equation concerns the mean response; it does not assert that every individual response equals that mean.

Let \(T\) be an absolute target response. Solving \(\mu(d_T)=T\) gives

A finite positive solution requires \(E_0<T<E_0+E_{\max}\). As \(T\) approaches the plateau from below, the denominator approaches zero, so the target dose can grow rapidly. Differentiating with respect to the target gives

Thus, for fixed model parameters, target sensitivity increases sharply as the target approaches the asymptote. In real estimation, the parameters are uncertain too; the formula is a useful diagnostic of the geometry, not a substitute for propagating that uncertainty.

Using a Profile to Diagnose Identification

A profile likelihood examines how well different candidate target doses fit the data while allowing the other model parameters to adjust. For each candidate \(d_T\), impose the constraint \(\mu(d_T)=T\), maximize the likelihood over the remaining free parameters, and compare that constrained maximum with the unconstrained maximum. A profile that rises slowly as the candidate dose moves away from its optimum indicates that the data do not strongly distinguish those doses.

For a one-dimensional target dose, a commonly used approximate 95% likelihood-ratio confidence set contains values satisfying a profile deviance no greater than 3.84. This cutoff uses the approximate chi-square reference with one degree of freedom and relies on regularity conditions, including adequate identifiability and an interior parameter value. Near a plateau those conditions may be questionable, so the cutoff is a diagnostic convention, not a guarantee of exact 95% coverage. The profile should be inspected, and its results interpreted alongside the design and model fit.

Why the Profile Can Be Flat

Near a plateau, dose-response predictions change slowly with dose. Consequently, shifting a candidate target dose can often be offset by changes in curve parameters, such as \(E_{\max}\) or \(ED_{50}\), with little loss of likelihood. This produces a shallow profile or a ridge of parameter combinations that fit nearly as well. The profile is useful precisely because it allows nuisance parameters to readjust instead of pretending they are known.

A short justification follows from the target-dose equation. Write \(A=E_0+E_{\max}\). Then \(d_T=ED_{50}(T-E_0)/(A-T)\). If \(A-T\) is small, a small change in \(A\) has a large effect on the denominator. In fact,

The magnitude also grows as \(A-T\) shrinks. Thus uncertainty in the plateau estimate can create substantial uncertainty in the target dose, even if other fitted quantities appear stable. This is one reason to inspect the dose profile rather than infer precision from the point estimate alone.

Diagnostics, Alternatives, and Limitations

Assess the profile over a sufficiently wide dose grid and inspect its shape, not only its minimum. Warning signs include a broad region under the likelihood-ratio cutoff, a profile that remains nearly flat at the edge of the scanned range, marked asymmetry, or an interval endpoint that changes substantially when the grid is refined. A profile that does not close on one side does not provide a finite endpoint from that scan; it is not evidence that the endpoint equals the last dose evaluated. Check that the model fits the observed dose-response pattern and that the target is below plausible plateau values. A poor structural fit can make a profile misleading as a measure of uncertainty within the chosen model.

The delta method, bootstrap distributions, and posterior distributions are complementary approaches covered in the earlier material on target-dose uncertainty. The delta method summarizes local uncertainty and can be unreliable when the target-dose transformation is strongly nonlinear. A bootstrap can show skewness or long tails, but its results depend on the resampling design and may be unstable when target estimates are frequently extreme or undefined. A posterior distribution incorporates prior information and can be informative when likelihood information is weak, but its concentration then depends in part on the prior. A profile makes the likelihood support across candidate doses visible; it does not include prior information and does not, by itself, diagnose model misspecification.

No single diagnostic repairs weak identification. More observations near the plateau may help estimate the asymptote, while observations across informative parts of the curve may help separate its parameters. However, adding doses or selecting a study design is a separate design question. This tutorial's central limitation is that its diagnostics are conditional on the chosen Emax structure and likelihood. A narrow profile cannot rule out a wrong model, and a wide profile should not be converted into a falsely precise point recommendation.

Common Mistake

Also avoid interpreting a dose beyond the largest studied dose as though the data directly observed the response there. Whether and how to address such extrapolation is a distinct issue; a profile diagnoses likelihood support under the fitted model, not the clinical credibility of every dose in its range.