Evaluating Dose-Ranging Designs With Emax Models

Why dose-arm design matters

A dose-ranging trial must do more than detect a difference from control. It should provide information about how response changes across doses and support a useful choice among the doses that could be taken forward. The locations of the dose arms determine which parts of that relationship are observed; the allocation determines how precisely each part is measured. A design that estimates the average response well at one dose may still be poor for describing the curve or selecting a dose that meets a clinical target.

Simulation makes these trade-offs explicit. Candidate designs can be evaluated under the same plausible data-generating scenarios, analysis method, total sample size, and selection rule. The conclusions are conditional on those assumptions, so the assumed dose-response shape and range of possible parameter values need to be examined rather than treated as known facts.

An Emax data-generating model

For a basic Emax model, let \(d\) denote dose, with placebo represented by \(d=0\). Orient the response so that a larger value means greater benefit. The expected response at dose \(d\) is

Here \(E_0\) is the mean response on placebo, \(E_{\max}\) is the maximum increase above placebo as dose becomes large, and \(ED_{50}\) is the dose that produces half of that maximum increase. With \(E_{\max}>0\) and \(ED_{50}>0\), the curve rises with dose and approaches \(E_0+E_{\max}\), but does not reach that asymptote at any finite dose.

A simple individual-level data-generating model is \(Y_{ij}=\mu(d_j)+\epsilon_{ij}\), where \(Y_{ij}\) is the response for participant \(i\) in arm \(j\), and \(\epsilon_{ij}\) is a residual error with mean zero and variance \(\sigma^2\). For the worked calculations below, assume independent normally distributed residuals with common variance, a fixed response assessment time, no missing outcomes, and a correctly specified Emax mean model. These are simplifying assumptions, not requirements for all dose-ranging simulations. If the intended trial includes between-participant variability, dropout, or other features, those should be represented consistently across candidate designs when comparing them.

The design is described by doses \(d_1,\ldots,d_K\) and allocations \(n_1,\ldots,n_K\), where \(\sum_{k=1}^{K} n_k=N\). Placebo should be included when the trial needs to estimate the placebo response or quantify treatment effects relative to placebo. A simulation comparison should state whether the same total \(N\) is held fixed and whether changes in dose placement also change allocation.

Why dose placement affects estimation

The Emax curve is nonlinear: low doses may lie on its rising portion, while high doses may lie near its plateau. A trial restricted to low doses may see little curvature and have difficulty distinguishing a large \(E_{\max}\) with a large \(ED_{50}\) from a smaller \(E_{\max}\) with a smaller \(ED_{50}\). A trial containing only high doses may show that response is near a plateau but provide little evidence about where the rise occurred. Both patterns can make parameter estimates unstable even when individual arm means are measured precisely.

A short sensitivity calculation explains why doses around \(ED_{50}\) can help estimate the location of the curve’s rise. Holding \(E_0\) and \(E_{\max}\) fixed, the magnitude of the derivative of the mean with respect to \(ED_{50}\) is

For positive \(d\), differentiate the dose-dependent term \(g(d)=d/(ED_{50}+d)^2\):

Thus \(g(d)\) increases up to \(d=ED_{50}\) and decreases above it. This does not prove that a design with a dose exactly at the true \(ED_{50}\) is always best: \(ED_{50}\) is not known when designing the trial, and \(E_0\), \(E_{\max}\), residual variation, and parameter correlations also affect precision. It does show why a useful range commonly includes doses near the plausible location of the curve’s rise, as well as doses low enough to anchor placebo and high enough to reveal possible saturation.

Allocation and simulation-based comparison

With independent residuals of common variance \(\sigma^2\), the sample mean in arm \(k\) has standard error \(\sigma/\sqrt{n_k}\). For two arms \(a\) and \(b\), the estimated difference in arm means has standard error

This expression illustrates an allocation trade-off. More participants in an arm reduce its contribution to uncertainty, but taking them from another arm increases uncertainty there. Equal allocation is a useful reference design, not a universal optimum. If estimation of placebo-relative effects is central, a larger placebo group may improve their precision; if the goal is to distinguish the curve’s shape, very small allocations to informative dose arms may be a weakness.

For a fair simulation comparison, hold constant the assumed trial features that are not under evaluation, such as total sample size, outcome model, analysis procedure, and replicate count. Change the dose placements, allocations, or both according to the comparison question. Evaluate more than one plausible parameter scenario: a design can look strong when the curve rises within its dose range and weak when \(ED_{50}\) lies above it or the response plateaus early.

For estimation, relevant summaries include bias in fitted \(E_{\max}\) and \(ED_{50}\), interval coverage, interval width, and the precision of predicted mean responses at doses of interest. Parameter estimation and prediction are not interchangeable: a model can estimate the mean response adequately over the observed range while estimating the asymptotic \(E_{\max}\) poorly. For selection, define success in relation to the trial’s actual objective, such as selecting a dose that reaches a prespecified response target. A selection summary should make clear what counts as success and what happens when no tested dose meets the rule.

Evaluating a dose-selection rule

A dose-selection rule must be specified before interpreting simulation results. One transparent example is a target-based rule: choose the lowest tested dose whose fitted mean response is at least a prespecified target \(\Delta^*\) above placebo. If no tested dose qualifies, select no dose. The restriction to tested doses avoids treating a model-based extrapolation as if it were direct trial evidence.

In a simulation replicate, the rule is applied to fitted responses. Its performance is then judged against the data-generating truth. For example, a selected dose may be classified as adequate if its true mean response is at least \(\Delta^*\). Other useful summaries include the proportion of replicates with no selection, the proportion selecting an adequate dose, and how often the selected dose is higher than the lowest dose that truly meets the target. These summaries answer different questions: a high adequate-selection rate does not by itself show that the rule avoids unnecessarily high doses.

Diagnostics, alternatives, and limitations

Inspect whether the simulated and fitted curves behave as intended across the full range, not only whether a model-fitting routine reports success. Warning signs include frequent nonconvergence, estimates of \(ED_{50}\) or \(E_{\max}\) at imposed bounds, very wide intervals, and strong changes in parameter estimates when a dose arm is removed. Check predicted responses at tested doses against the observed arm patterns in simulated replicates. If estimates are unstable or dose-response patterns are systematically missed, the design comparison may be revealing insufficient dose coverage or a mismatch between the generating and fitted models.

A common mistake is to choose the design with the most precise estimate of \(E_{\max}\) in one assumed scenario and call it the best dose-ranging design. That reasoning ignores whether the assumed plateau is observed, whether other plausible curves produce different results, and whether the design can select an adequate dose. The correction is to prespecify the goals and compare estimation and selection measures across scenarios. The introductory material on estimating operating characteristics from replicates and quantifying Monte Carlo error provides the basis for interpreting those simulated summaries.

An Emax model is not the only way to describe dose response. A categorical-dose analysis estimates separate arm means without imposing a smooth curve; this can be a useful comparator when model shape is uncertain, but it does not directly estimate Emax parameters and may use the dose information less efficiently. A simpler linear dose trend may be easier to estimate over a limited range, but can misrepresent curvature or a plateau. The design evaluation should reflect the planned primary analysis and can include plausible alternative models to assess sensitivity, without assuming that any one curve is guaranteed to be correct.

The basic calculations here assume a fixed assessment time, a single response outcome, independent residuals, and a common residual variance. Realistic simulations may need to represent heterogeneity, correlated pharmacokinetic and pharmacodynamic variability, adherence, dropout, or misspecification when these features are relevant to the question. Dose selection based on a response target also does not, by itself, incorporate safety, exposure, feasibility, or later clinical evidence. Those considerations can change which dose is suitable for development.