When Efficacy Is Still Rising
A dose-response curve that continues upward across the tested doses can make a Phase 3 recommendation feel unfinished: if efficacy is higher at the highest dose, why not simply select that dose? The answer depends on what the evidence says about the size and uncertainty of the added benefit, whether the dose meets safety constraints, and what decision rule is appropriate for development. A rising trend alone does not establish that the highest tested dose is best, nor does it show that efficacy would keep increasing beyond the studied range.
The aim is not to force a plateau-based conclusion when the data do not support one. It is to state a defensible dose rationale: identify a dose with meaningful evidence of benefit, show how relevant higher-dose increments compare with clinical criteria, apply the stated safety limits, and describe the remaining uncertainty. This is a decision under uncertainty, not proof that the selected dose is a mathematical or biological optimum.
A useful starting point is to define the clinical benefit scale and its direction. In the examples below, benefit is a positive, placebo-adjusted reduction in a symptom score, measured in points; larger values are better. Let \(\hat B(d)\) be the estimated benefit at dose \(d\), and let \(SE\{\hat B(d)\}\) denote its standard error. For two doses \(d_H>d_L\), the estimated incremental benefit is \(\hat\Delta=\hat B(d_H)-\hat B(d_L)\). Its uncertainty must be calculated from the variance of the contrast, accounting for the covariance between dose estimates—not by treating estimates that share a placebo group as necessarily independent.
Quantifying the Increment and Its Uncertainty
For an approximately normally distributed model-based estimate, a two-sided 95% confidence interval for the incremental benefit is \(\hat\Delta \pm 1.96SE(\hat\Delta)\). This approximation is appropriate when the fitted model and standard error are adequate and the sampling distribution of the contrast is approximately normal. The interval describes values reasonably compatible with the data and model; it is not a 95% probability statement about the fixed, true increment.
The key comparison is often between the interval and a pre-specified threshold for clinical relevance, denoted here by \(M\), in the same units as benefit. For example, if \(M=1.0\) symptom-score point, an interval entirely above 1.0 supports an increment exceeding that threshold under the model. An interval that includes both zero and values above 1.0 leaves open whether there is any increment and whether it is clinically important. A failure to rule out zero is not an equivalence finding.
The variance in this expression includes the covariance between the two estimates. If both dose estimates use the same placebo group, their errors can be correlated. The standard error reported for a fitted contrast, or calculated from the fitted covariance matrix, is therefore preferable to subtracting estimates and combining their separate standard errors as if they were independent.
Worked Example: A Higher Dose Adds Uncertain Benefit
In a hypothetical trial, estimated placebo-adjusted symptom-score reductions are 4.0 points at 10 mg and 5.4 points at 20 mg. A fitted model estimates the 20 mg versus 10 mg increment as 1.4 points, with \(SE(\hat\Delta)=0.70\) points. The pre-specified threshold for a clinically meaningful increment is \(M=1.5\) points.
The contrast is \(5.4-4.0=1.4\) points. Its approximate 95% confidence interval is \(1.4\pm1.96(0.70)=1.4\pm1.372\), or \([0.028,2.772]\) points.
The point estimate is positive, and the interval is just above zero at its lower end, but the interval includes values below and above 1.5 points. The data therefore do not establish that the higher dose adds a clinically meaningful 1.5-point benefit. They also do not exclude an increment larger than 1.5 points. This is an uncertain comparison, not proof of no added benefit and not a basis for calling either dose optimal.
Work From Benefit to a Constrained Dose Choice
A transparent rationale separates what is estimated from what is decided. First identify doses with evidence of benefit relative to placebo and compare that benefit with the clinical threshold relevant to the indication. Then examine the increments between doses that could affect the decision. Finally, apply the safety constraints that were defined for the decision. A threshold should be chosen for clinical or development reasons before interpreting the result; setting it after seeing which dose wins can make the rationale appear more certain than the evidence warrants.
Specify what a positive benefit means, the units, and the clinically relevant minimum for benefit or for an incremental dose increase.
Report dose-specific estimates and uncertainty, then calculate the relevant between-dose contrasts with their model-based standard errors.
Use the pre-specified safety criterion, including the uncertainty measure it requires; do not equate an observed rate with a known risk.
Explain why the chosen dose is supported for Phase 3, what higher doses suggest, and why the evidence does not establish an optimum or a plateau.
Worked Example: Selecting a Dose Under a Safety Constraint
Consider a hypothetical study with 200 participants per group. Placebo-adjusted symptom-score reductions are 4.0 points at 10 mg and 6.2 points at 20 mg; the minimum clinically relevant benefit versus placebo is 3.0 points. A dose is considered acceptable only if the upper limit of an approximate 95% confidence interval for its excess risk of a specified adverse event, relative to placebo, is no more than 10 percentage points. The observed event rates are 10% on placebo, 12% at 10 mg and 15% at 20 mg.
Both doses exceed the minimum benefit on the point estimates: 4.0 points at 10 mg and 6.2 points at 20 mg, each greater than 3.0 points. The efficacy increment for 20 mg versus 10 mg is \(6.2-4.0=2.2\) points. Suppose the fitted model gives \(SE(\hat\Delta)=0.80\) points. The 95% interval is \(2.2\pm1.96(0.80)=2.2\pm1.568\), or \([0.632,3.768]\) points. It supports a positive estimated increment, but its lower limit is below 1.5 points, so the data do not establish that the increment exceeds a 1.5-point clinical threshold.
At 10 mg, the observed excess event risk is \(0.12-0.10=0.02\). Using the independent-binomial standard error for these two groups, \(SE=\sqrt{0.12(0.88)/200+0.10(0.90)/200}=0.0313\). The approximate upper 95% limit is \(0.02+1.96(0.0313)=0.0813\), or 8.13 percentage points, which meets the 10-point rule.
At 20 mg, the observed excess is \(0.15-0.10=0.05\), and \(SE=\sqrt{0.15(0.85)/200+0.10(0.90)/200}=0.0330\). Its upper limit is \(0.05+1.96(0.0330)=0.1147\), or 11.47 percentage points, which exceeds the rule. Under this stated criterion, 10 mg is the supported Phase 3 choice: it exceeds the minimum benefit and satisfies the safety constraint. This does not show 10 mg is an efficacy optimum; efficacy estimates are higher at 20 mg, but the safety criterion is not met there.
What the Dose Rationale Can—and Cannot—Claim
A higher tested dose may be a defensible choice when its estimated benefit is meaningfully greater than that of a lower dose, the uncertainty is compatible with that claim, and the selected dose satisfies the applicable safety constraints. Conversely, a lower dose may be justified when it already provides relevant benefit and a higher dose fails a pre-specified constraint. In either case, the argument should make explicit which evidence is decisive rather than relying on the fact that one estimate is numerically largest.
A plateau-based argument asks whether additional increases in dose produce clinically negligible increments. When the evidence does not establish that, describe the response as still increasing over the tested range or as having no clear plateau. Do not turn “no plateau demonstrated” into “the maximum tested dose is optimal.” The maximum tested dose is simply the upper dose studied; the data do not determine what would happen at untested doses.
Worked Example: Choosing a Supported Dose Without Claiming an Optimum
In a hypothetical trial, estimated benefits are 3.8, 5.1 and 6.7 symptom-score points at 10, 20 and 40 mg. All three doses satisfy a pre-specified safety rule. The clinically relevant benefit versus placebo is 3.0 points, and a 1.0-point increment between active doses is considered important. The fitted 40 mg versus 10 mg contrast is 2.9 points, with \(SE=0.75\) points.
The 40 mg versus 10 mg interval is \(2.9\pm1.96(0.75)=2.9\pm1.470\), or \([1.430,4.370]\) points. Its lower limit exceeds the 1.0-point threshold. Thus, in this example, the estimated total benefit at 40 mg is meaningfully greater than at 10 mg under the stated criterion. The 40 mg estimate also exceeds the 3.0-point minimum benefit, and safety is acceptable under the stated rule.
The estimates rise at every tested dose, so these results do not demonstrate a plateau. They can support selecting 40 mg for Phase 3 under the stated benefit and safety criteria, but they do not establish that 40 mg is the optimum among untested doses or that further increases would add benefit. The rationale is based on evidence within the studied range, not an extrapolated maximum.
Diagnostics, Alternatives, and Common Pitfalls
Before relying on a dose comparison, check that the endpoint direction and units are consistent across doses, the fitted model represents the observed dose-response pattern adequately, and the contrast standard error reflects the design and shared controls. Review estimates and intervals rather than only a trend test or p-value. Sparse data at an upper dose, influential observations, poor model fit, or substantial missing data can make the apparent rise fragile. If the estimated pattern is nonmonotonic, do not summarize it as a simple continued increase without explaining the discrepancy.
The main alternatives are not interchangeable. Selecting the maximum tested dose is simple but ignores whether its extra benefit is clinically relevant or whether it meets safety constraints. Selecting the lowest dose that reaches a benefit threshold can be transparent, but it requires a meaningful threshold and does not prove that higher doses add nothing. A plateau-based selection is appropriate only when evidence supports clinically negligible increments; a non-significant contrast alone does not establish that condition. A model-based dose choice can use the full response pattern, but its conclusions depend on model fit and should remain within the dose range and assumptions supported by the data.
The introductory material on dose-response trends in confirmatory trials provides the context for interpreting a fitted trend. For this decision, a trend is not itself a dose recommendation: the clinically relevant contrasts, uncertainty and safety constraints determine what can be justified. A rationale should also disclose the limitations of the evidence, including imprecision, the number and spacing of tested doses, the duration of observation, and whether the endpoint captures the benefit relevant to Phase 3.
Check Your Understanding
For each item, give the conclusion supported by the stated evidence. Solutions follow each question.
- A high-versus-low dose increment is estimated as 0.8 points with \(SE=0.60\) points. The clinically meaningful increment is 1.0 point. Calculate an approximate 95% interval and interpret it.
Solution: \(0.8\pm1.96(0.60)=0.8\pm1.176\), giving \([-0.376,1.976]\) points. The interval includes zero and values above 1.0, so neither an increment nor a clinically meaningful increment is established; a meaningful benefit remains possible. - Why is a non-significant high-versus-low dose contrast not enough to conclude that the dose-response curve has plateaued?
Solution: A non-significant result means the data do not rule out a zero contrast at the chosen significance level. It does not show that the entire plausible range of increments is clinically negligible. A plateau claim requires evidence sufficiently precise to support negligible increments against a defined clinical threshold. - A chosen dose has an estimated placebo-adjusted benefit of 4.2 points, with a 95% interval of \([3.1,5.3]\) points. The pre-specified minimum benefit is 3.0 points. What does this establish, and what does it not establish about higher doses?
Solution: Under the model, the interval lies above the minimum, supporting benefit greater than 3.0 points at this dose. It says nothing by itself about whether a higher dose adds meaningful benefit, meets safety constraints, or is optimal. - Observed adverse-event excess at a dose is 6 percentage points, but the upper confidence limit is 12 points. The pre-specified maximum permitted upper limit is 10 points. Does the dose satisfy the constraint?
Solution: No. The criterion applies to the upper confidence limit, and 12 percentage points exceeds the 10-point maximum. The observed excess of 6 points does not override the uncertainty-based rule.