The Gold Standard of Experimental Design | Dose–Response Trials, Model Uncertainty and Target-Dose Selection

1. Define the target-dose question before selecting models

Back to contents

State the population, endpoint, time point, dose range and treatment objective. The target might be the lowest dose achieving a clinically important effect, a dose near the plateau, or the best benefit–risk trade-off. These are different estimands. A statistical signal that some dose works does not by itself select the dose to advance.

2. Separate proof of dose response from estimation

Back to contents

One task tests whether the data show a dose-related signal against a flat response; another estimates the curve and target dose. Multiple-comparison procedures can protect false-positive error for the first task, while modelling pools information for the second. Keeping the stages distinct prevents a flexible curve chosen after seeing the data from creating both the hypothesis and its apparent confirmation.

3. Prespecify a plausible candidate-model set

Back to contents

Candidate curves may include linear, Emax, sigmoid Emax, exponential or logistic shapes, with parameters chosen from prior knowledge and clinically meaningful effect sizes. The set should span genuine uncertainty without becoming an unprincipled catalogue. If every model is nearly identical at the studied doses, model multiplicity adds little; if implausible curves dominate, averaging can dilute interpretation.

4. Choose doses and allocation to reveal shape

Back to contents

Include placebo or an appropriate reference where ethical, cover the range where curvature is expected and avoid clustering all active doses near the presumed maximum. Enough distinct levels are needed to distinguish rising, plateauing and delayed responses. Allocation can be optimised for contrasts or curve precision, but safety, recruitment and exposure constraints remain. A narrow high-dose range cannot identify the minimum effective dose.

5. Use MCP-Mod as a structured workflow, not a black box

Back to contents

Multiple Comparison Procedures and Modelling, or MCP-Mod, first constructs optimal contrasts for the prespecified shapes and tests them with multiplicity control. Supported models are then fitted to estimate the dose–response curve and target. Selection, averaging and uncertainty intervals require declared rules. Software output does not decide clinical relevance or erase uncertainty about untested shapes.

6. Worked example: five doses and two plausible curves

Back to contents

A trial allocates participants to placebo and four active doses. Mean improvement rises quickly at the first two doses and changes little thereafter, but estimates are noisy. Both an Emax and a sigmoid Emax candidate reject a flat response. The Emax fit reaches the prespecified six-point benefit at a lower dose than the sigmoid fit. Instead of advancing the largest observed mean, the team reports target estimates under both models, examines model-averaged uncertainty and checks adverse events. The selected dose is the lowest range supported by efficacy and tolerability, with the disagreement carried forward.

7. Combine efficacy with safety, tolerability and exposure

Back to contents

A response plateau does not imply an optimal dose if adverse events continue to rise. Inspect safety endpoints, discontinuation, pharmacokinetics and relevant subgroups across the same dose range. Avoid selecting efficacy models first and consulting safety only after a favourite dose appears. A target-dose rule should state how clinical benefit, uncertainty and harm are weighed and who owns the final decision.

8. Changed-context transfer check: tutoring intensity

Back to contents

Move from drug dose to weekly hours of a learning intervention. The transferable structure is a multi-level intensity trial with possible diminishing returns and burden. The changed context adds attendance, teacher capacity, contamination and an intensity that participants may not receive as assigned. Models should describe assigned and delivered exposure separately, and the target should balance learning gain against fatigue and feasibility rather than copy pharmaceutical thresholds.

9. Know the limits and common misconceptions

Back to contents

A significant trend does not identify a unique target dose. Model averaging does not protect against every omitted shape, and the best-fitting model can change with one noisy group. Extrapolating beyond the studied range is especially fragile. Small phase-II studies may estimate efficacy better than rare harms. Adaptive changes, missing outcomes and multiplicity beyond the primary endpoint need their own controls.

10. Audit the route from design to dose decision

Back to contents

Publish prior evidence, candidate models and parameters, dose levels, allocation, endpoint, contrasts, multiplicity method, model-selection or averaging rule, target definition, safety rule, missing-data handling, code and diagnostics. Plot observed summaries with every retained curve and uncertainty band. Refit after plausible candidate-set, variance and influential-group changes. The audit should show whether the dose decision survives reasonable model uncertainty.

11. Frequently asked questions

Back to contents

Why not compare every dose with placebo? Pairwise tests use less information about the shared curve and can reward random highs. Does MCP-Mod choose the final dose automatically? No. How many models are needed? Enough to represent scientifically plausible shapes, not every imaginable curve. Must the highest dose be tested? The range must support the target question while remaining safe and ethical. Can safety use another model? Yes, but its assumptions and decision role must be explicit.

12. The gold standard selects a decision, not a decorative curve

Back to contents

A credible dose-finding programme separates signal detection from curve estimation, studies doses that reveal shape and carries several plausible models into the decision. It defines clinical relevance before fitting and integrates safety, tolerability and exposure. The gold standard does not hide disagreement among curves; it uses that disagreement to show what is known, what remains model-dependent and why the selected target is defensible.