Establishing a Minimum Effective Dose

What Does “Minimum Effective Dose” Mean?

A dose is not “effective” in isolation: the conclusion depends on the endpoint, the target population, the assessment time, and the amount of benefit judged worthwhile. A statistical test of whether an effect differs from zero does not by itself specify a clinically relevant threshold. Here, the efficacy criterion is a prespecified benefit of at least \(\delta\) units on an endpoint scale oriented so that larger values represent greater benefit.

Let \(M(d)\) denote the expected treatment benefit at dose \(d\), relative to the chosen reference, for the estimand and assessment time of interest. For example, \(M(d)\) might be the placebo-adjusted mean reduction in symptom severity, expressed in scale points so that a larger reduction is more beneficial. The target is the lowest dose meeting \(M(d)\geq\delta\). This is an efficacy target; it does not by itself establish that the dose is acceptably safe or suitable for use.

A common sufficient condition for the minimum to exist is that \(M\) is continuous and \(A\) is nonempty. In that case \(A\) is a closed subset of \([0,\infty)\), and every nonempty closed subset of this domain has a minimum. If continuity is not justified, or the model is defined only over a restricted dose domain, attainment must be checked rather than presumed. A monotone increasing dose-response function is convenient because it gives a single crossing, but monotonicity is not part of the definition. With a nonmonotonic response, the lowest qualifying dose must be found across the full modeled domain.

Estimating a Threshold Dose From a Dose-Response Model

The criterion \(\delta\) should be specified before interpreting a fitted curve. It is defined in the units of the benefit measure, not in units of dose. The model, in turn, must describe the same estimand and time point. For an increasing Emax model with positive parameters,

where \(E_{\max}\) is the limiting benefit and \(ED_{50}\) is the dose producing half of that limiting benefit. The model is continuous and increasing for \(E_{\max}>0\) and \(ED_{50}>0\). If \(0<\delta<E_{\max}\), solving \(M(d)=\delta\) gives a finite, attained threshold:

The condition \(\delta<E_{\max}\) matters. If \(\delta=E_{\max}\), the curve approaches the criterion only as dose tends to infinity and no finite dose reaches it. If \(\delta>E_{\max}\), no dose meets the criterion under this model. A computed dose is therefore meaningful only after checking that the target is attainable and that the fitted model’s assumptions are credible.

Why Uncertainty in the Dose Estimate Matters

The threshold estimate is a function of estimated model parameters. Uncertainty in those parameters therefore induces uncertainty in \(d^*\), and uncertainty can be especially large when the target is close to the model’s limiting benefit. A confidence interval for \(d^*\) describes uncertainty in the threshold under the model and inferential procedure; it is not a probability statement about a fixed parameter being inside the realized interval.

For a smooth function of approximately normally distributed parameter estimates, a first-order delta-method approximation propagates their covariance. Write \(g(E_{\max},ED_{50})=ED_{50}\delta/(E_{\max}-\delta)\). Its partial derivatives are

If \(\widehat{\mathrm{Cov}}\) is the estimated covariance matrix of \((\widehat E_{\max},\widehat{ED}_{50})\), the approximate variance is the gradient of \(g\), evaluated at the estimates, multiplied by that covariance matrix and the transposed gradient. A symmetric interval based on this approximation can be inadequate if the dose distribution is strongly skewed, approaches a parameter boundary, or includes cases in which the criterion is unattainable. Profile likelihood or a suitable parametric bootstrap can better represent such cases; those methods still depend on a credible model and appropriate inferential conditions.

What a Dose Grid Can—and Cannot—Locate

Dose-ranging studies observe responses at selected doses, not at every possible dose. The lowest tested dose with an estimated benefit at or above \(\delta\) is a grid-based summary. It is generally not the same quantity as the minimum effective dose on a continuous dose scale. If a monotone, continuous true response is below the criterion at one dose and above it at a higher dose, the crossing lies between those doses. But estimated group means are noisy, and the grid alone does not establish either the true response at its dose levels or the curve between them.

A dose grid is poorly informative about the target if no tested doses bracket it, if there are few doses around the crossing, or if uncertainty in the fitted curve is wide relative to the spacing of the doses. A target close to \(E_{\max}\) is particularly difficult: the inverse formula has a denominator \(E_{\max}-\delta\), which becomes small as \(\delta\) approaches \(E_{\max}\). The resulting threshold may be highly sensitive to modest changes in the estimated plateau. A profile likelihood that extends to very high doses, strong dependence on model choice, or materially different estimates under plausible curves are warnings that the data have not localized the target reliably.

Assumptions, Diagnostics, and Common Mistakes

A minimum effective dose is conditional on the definition of benefit and on the model used to estimate it. Before interpreting a result, check that the effect direction and units are explicit; the criterion is clinically justified and fixed independently of the fitted crossing; and the estimated dose is within a domain where the model and dose administration are meaningful. The response model should also be checked against the observed dose-response evidence. Residual patterns, systematic underprediction or overprediction at particular doses, implausible parameter estimates, and strong changes under reasonable alternative models can signal misspecification. Model-averaging and misspecification diagnostics are relevant when plausible models disagree, but do not replace a clear efficacy criterion.

Monotonicity deserves special attention. The Emax inversion above presumes a monotone increasing curve. If the response is nonmonotonic, a single algebraic crossing may not identify the lowest qualifying dose: the threshold set may have multiple separated portions. The definition still applies, but the fitted response must be searched over the entire dose domain and attainment checked. Likewise, if the maximum modeled benefit does not reach \(\delta\), reporting a finite dose by extrapolating the Emax inverse is not justified.

• Mistake: treating the infimum as automatically qualifying. If the set of qualifying doses is \((1,\infty)\), its infimum is 1, but 1 is not in the set. Correction: call \(d^*\) a minimum only when the threshold set has a minimum; otherwise report the infimum as a boundary and state that it does not meet the criterion.
• Mistake: equating statistical significance with the efficacy threshold. Evidence that an effect differs from zero does not show that it reaches \(\delta\). Correction: estimate the dose at which the defined benefit criterion is reached, with uncertainty.
• Mistake: calling the lowest successful tested dose the true threshold. This ignores untested doses and sampling uncertainty. Correction: distinguish the lowest tested dose whose estimate meets the criterion from the model-based threshold on a continuous scale.
• Mistake: trusting a precise-looking inverse near the plateau. When \(\delta\) is close to \(E_{\max}\), small changes in \(E_{\max}\) can cause large changes in \(d^*\). Correction: inspect uncertainty and model sensitivity, and report when the data cannot support a reliable finite threshold.

The practical alternatives answer different questions. A tested-dose rule is transparent and avoids interpolation, but its answer depends on the dose grid and may be conservative or coarse. Model-based inversion uses information across doses and estimates a continuous threshold, but relies on the chosen curve and its assumptions. Profile likelihood or bootstrap methods can better reflect asymmetric uncertainty than a local normal approximation, but cannot create information absent from the dose-ranging data. The appropriate report states both the model-based estimate and its uncertainty, and says plainly when the available evidence does not locate the threshold reliably.