Extrapolation and Boundary Risks in Dose Selection

Why a Dose Estimate Can Outrun the Evidence

A fitted dose-response model can return a dose even when that dose was never studied. It can also return a parameter estimate exactly at a constraint imposed by the analyst or software. Both results may be numerically well-defined, but neither automatically means the corresponding dose is supported by the data. The central task here is to separate what the model calculates from what the observed dose range and model assumptions justify.

This distinction matters in dose selection because dose-response curves can be shallow, nonlinear, or poorly characterized near their upper plateau. Earlier material on weak identification near dose-response plateaus considers weak information in that setting; here the emphasis is on whether the proposed dose is inside the range studied and on what to conclude when estimation reaches a boundary.

A recommendation may also be described as “at a boundary” for more than one reason. A selected dose can equal the largest dose studied, or the maximum dose permitted in an optimization. Separately, a fitted model parameter can equal a lower or upper constraint used in estimation. These are different situations: a dose boundary concerns the decision range, whereas a parameter boundary concerns the fitted model. In either case, the boundary deserves scrutiny rather than automatic acceptance.

Checking an Inverse Dose Estimate Against the Studied Range

For a concrete calculation, consider the Emax mean-response model

where \(E(D)\) is the expected response at dose \(D\), \(E_0\) is the expected response at zero dose, \(E_{\max}>0\) is the maximum increase above \(E_0\), and \(ED_{50}>0\) is the dose producing half of that increase. This model assumes that the mean response follows the specified curve over the dose range being interpreted. It does not establish that the curve remains appropriate outside the studied range. The response measurements must also be on a scale for which the model’s mean and asymptote are meaningful.

To find the dose associated with a target mean response \(T\), set \(E(D)=T\) and solve:

For a finite positive solution under this model, the target must satisfy \(E_0<T<E_0+E_{\max}\). A target at or above the asymptote cannot be reached at a finite dose, and a target at or below \(E_0\) does not define a positive target-attaining dose in this calculation. Even when these conditions hold, the inverse estimate must still be compared with the observed dose support.

This example illustrates why the arithmetic and the evidence check are separate steps. A curve can produce fitted values at each dose, but noticeable residual patterns can call its adequacy into question. A target dose calculated from that curve inherits the model’s limitations; it is not made reliable merely by being reported to several decimal places.

Extrapolation and the Sensitivity of Target Doses

When an inverse estimate lies beyond \(D_{\max}\), the model is being asked to predict beyond the doses that informed its fit. The prediction may be plausible, but the data do not directly show the response at that proposed dose. The further the estimate lies beyond the observed range, the more its interpretation depends on the assumed curve shape and the stability of the fitted parameters.

The target-dose calculation also shows why estimates near the asymptote can be especially sensitive. Differentiating the inverse formula with respect to the target gives

To see this, write the target-dependent factor as \((T-E_0)/(E_0+E_{\max}-T)\). Applying the quotient rule gives a numerator of \((E_0+E_{\max}-T)+(T-E_0)=E_{\max}\); multiplying by \(ED_{50}\) yields the displayed derivative. As \(T\) approaches \(E_0+E_{\max}\) from below, the denominator approaches zero and the derivative grows without bound. Small changes in the target can therefore produce large changes in the calculated dose near the asymptote.

For the parameters above, at \(T=55\) units the derivative is \(50(60)/(70-55)^2=3000/225=13.33\) mg per response unit. At \(T=65\) units it is \(3000/(70-65)^2=120.0\) mg per response unit. The latter calculation indicates much greater local sensitivity: a small change in the target, or in the fitted curve that determines the target’s position relative to the asymptote, can shift the inverse dose substantially. This is a diagnostic of fragility, not a substitute for examining the uncertainty analysis.

What a Parameter Boundary Does—and Does Not—Tell You

Model fitting often uses constraints to keep parameters in meaningful ranges or to help numerical optimization. For example, an Emax model commonly requires \(E_{\max}>0\) and \(ED_{50}>0\). An analyst may also impose finite computational bounds, such as \(1\leq ED_{50}\leq 200\) mg. A fitted value at one of these imposed endpoints is a warning that the unconstrained optimum may lie outside the search region, or that the data provide little information to locate the parameter precisely. The endpoint itself is not evidence that the biological parameter truly equals that value.

A dose recommendation at a dose limit needs a related but distinct interpretation. If an optimization restricted to 0–100 mg selects 100 mg, that says only that 100 mg is the best point within the specified decision range under the chosen objective. It does not establish that the response would continue to improve above 100 mg or that 100 mg is a global optimum over a wider range. Likewise, a recommendation equal to \(D_{\max}\) is at the edge of observed support even if no fitted parameter is at a boundary.

Diagnostics, Alternatives, and Limitations

Several checks help determine whether a dose result is being driven by extrapolation or a boundary rather than by stable evidence:

• Mark the support explicitly. Report the lowest and highest studied doses alongside the estimate, and state how far outside the range it lies.
• Inspect fitted values and residuals by dose. Large discrepancies or a systematic pattern may indicate that the assumed curve is not describing the observed data adequately.
• Check optimization status and constraints. Identify parameters at bounds, inspect convergence messages, and determine whether changing the search bounds changes the estimate.
• Examine dose sensitivity. Small changes in the target or fitted parameters that cause large changes in the inverse dose indicate a fragile recommendation.
• Compare with the studied design. A model fitted at widely spaced or sparse dose levels may provide little direct support for the shape between, or beyond, those levels.

The main alternative to extrapolating a fitted curve is to restrict the recommendation to a dose range supported by the study or to obtain data at doses that clarify the uncertain region. A bounded optimization can be useful when the decision must be restricted to a defined range, but the bound should be reported and the result interpreted as conditional on it. Profile likelihood, bootstrap distributions, or posterior distributions—methods covered in earlier material—can describe uncertainty in a target dose, but none can turn unsupported extrapolation into direct evidence. Their conclusions still depend on the model and data.

There are important limitations to these checks. A dose inside the observed range can still be weakly supported, and an estimate outside the range is not automatically wrong. The concern is that its justification relies increasingly on assumptions not tested by observations near the proposed dose. Conversely, a parameter at a bound is a diagnostic signal, not proof of model failure: the bound might be appropriate and the data might genuinely favor a value near it. The appropriate conclusion depends on the profile, fit, parameterization, and scientific context.

A common mistake is to treat every optimizer output as a recommendation: “The model returned 150 mg, so 150 mg is the supported dose,” or “The fitted parameter is 200 mg, so its value is known.” The correction is to ask separately whether the target is mathematically attainable, whether the dose lies within observed support, whether the fit is adequate, and whether the estimate is constrained by a boundary. A numerical answer is a calculation; evidence support is an interpretation that requires these checks.