Why the Dose Grid and Allocation Matter
Dose-selection evidence depends not only on the observed responses, but also on which doses were tested and how many participants received each one. A study with only placebo and one active dose may estimate a treatment contrast, but it cannot show how response changes across several active doses. A study with many dose levels but very few participants in some arms may offer broad coverage while estimating particular contrasts imprecisely. These are different limitations, and neither is repaired simply by fitting a more elaborate curve.
The practical question is not just whether a dose-response model can be fitted. It is whether the available data support the claim being made: for example, that a tested dose improves the endpoint versus placebo, that one tested dose is preferable to another, or that an untested dose is the best choice. The further a claim reaches beyond observed comparisons, the more it relies on assumptions that the study design may not be able to check.
What Unequal Allocation Changes
Suppose two randomized, independent arms have sample sizes \(n_1\) and \(n_0\), sample means \(\bar{x}_1\) and \(\bar{x}_0\), and sample standard deviations \(s_1\) and \(s_0\). The estimated difference is \(\bar{x}_1-\bar{x}_0\). Its estimated standard error is
This expression assumes independent randomized groups and uses the observed within-arm variability to estimate uncertainty. Randomization supports interpreting the contrast as a treatment effect for the trial population, subject to the study’s conduct and estimand. Unequal allocation does not, by itself, bias the difference in sample means. It does change its precision: with similar variability, fewer participants in an arm generally mean a larger contribution to the standard error.
For a rough 95% confidence interval with sufficiently large samples, one can use the normal approximation \((\bar{x}_1-\bar{x}_0)\pm1.96SE\). This approximation is not automatically reliable with very small arms or strongly non-normal outcomes; an analysis should use an appropriate small-sample method when needed. A confidence interval describes uncertainty about the contrast, not the probability that the fixed true contrast lies inside the computed interval.
Worked Example: Comparing Two Active Doses With Sparse Arms
Hypothetical data. A randomized trial measures improvement in a clinical score, in points, with larger values indicating more improvement. The placebo arm has \(n=40\), mean improvement 3.0 points, and standard deviation 8.0 points. The low- and high-dose arms each have \(n=10\), standard deviation 8.0 points, and mean improvements of 8.0 and 9.5 points, respectively. Assume independent arms and use the large-sample normal approximation for illustration.
The low-dose estimate versus placebo is \(8.0-3.0=5.0\) points. Its standard error is
The approximate 95% confidence interval is
For high dose versus placebo, the estimated difference is \(9.5-3.0=6.5\) points. The arm sizes and standard deviations are the same as for the low-dose contrast, so its standard error is also 2.828 points. Its interval is \(6.5\pm5.543=(0.957,\ 12.043)\) points.
The direct high- versus low-dose estimate is \(9.5-8.0=1.5\) points, with
The approximate 95% interval is \(1.5\pm1.96(3.578)=(-5.513,\ 8.513)\) points. The point estimates increase across the three arms, but the direct comparison is imprecise. These data do not establish that high dose is better than low dose. Nor do three dose levels, only two of them active, give a detailed account of the response shape between or beyond the tested doses.
Why a Few Dose Levels Cannot Reveal Every Curve
A dose-response curve is a model for how the expected outcome changes with dose. A small number of tested doses can estimate responses at those doses, but many different curves may pass near the same observed arm means. With only placebo and one active dose, a dose trend cannot be separated from a single placebo contrast without imposing a model. With a few active levels, the data may still provide little information about curvature, a plateau, or behavior in an untested interval.
An estimate for an untested dose is therefore not simply another observed arm mean. It is a prediction based on assumptions about the curve. Interpolation refers to prediction between tested doses; extrapolation refers to prediction outside the tested range. Interpolation can be more defensible than extrapolation when the doses bracket the target, but it still depends on the assumed response shape and on adequate data near the target. The introductory material on choosing among tested doses and interpolating between them provides the broader decision context; here the key issue is how sparse coverage restricts the strength of a claim.
Worked Example: A Missing Dose Region Makes Selection Model-Dependent
Hypothetical data. Suppose estimated mean improvement is 0 points at placebo, 5 points at dose 10 mg, and 6 points at dose 40 mg. There are no tested doses between 10 and 40 mg. For this illustration, treat these estimates as exact so that the effect of curve assumptions is visible; actual sampling uncertainty would add further uncertainty.
A straight line between the two active-dose estimates gives a prediction at 20 mg of
A saturating \(E_{\max}\) curve, \(E(d)=E_{\max}d/(ED_{50}+d)\), can also pass through the active-dose estimates. Requiring \(E(10)=5\) and \(E(40)=6\) gives \(ED_{50}=20/7=2.857\) mg and \(E_{\max}=45/7=6.429\) points. It predicts
A quadratic \(E(d)=ad+bd^2\), constrained to pass through \((10,5)\) and \((40,6)\) as well as \((0,0)\), has \(a=37/60\) and \(b=-7/600\). It predicts \(E(20)=(37/60)(20)-(7/600)(20^2)=7.667\) points. That quadratic is not necessarily a scientifically suitable model, particularly outside the observed range; its role here is to show that the three observed values alone do not determine the response at 20 mg.
If a decision threshold were 6 points of improvement, the linear and \(E_{\max}\) predictions at 20 mg fall below it, while the quadratic prediction exceeds it. The threshold conclusion is therefore not supported by the observed arm means alone. It depends on which curve is assumed, and in real data it would also need to account for uncertainty in the estimates and model parameters. Testing a dose within the gap would provide direct information that curve fitting cannot manufacture.
Allocation, Precision, and Design Trade-Offs
For a single comparison with equal outcome variability, balanced allocation minimizes the variance of the difference when the total sample size is fixed. To see why, let the common variance be \(\sigma^2\) and let \(n_1+n_0=N\). Then
For fixed \(N\), the product \(n_1n_0\) is largest when the allocation is balanced, so the variance is smallest then. This result concerns one comparison under equal variances; it does not imply that every multi-arm dose-ranging trial should allocate equally. A design may deliberately allocate more participants to placebo, a promising dose, or a dose requiring greater precision. The choice should follow the comparisons and decisions the study needs to support.
Worked Example: Quantifying the Precision Cost of Imbalance
Hypothetical data. Compare one active dose with placebo in a trial with 100 participants total. Suppose both arms have standard deviation 10 endpoint units, and the estimated active-minus-placebo difference is 6 units. Compare an allocation of 20 active and 80 placebo participants with an allocation of 50 in each arm. Use the normal approximation to illustrate precision.
With 20 active and 80 placebo participants,
The approximate 95% interval is \(6\pm1.96(2.500)=(1.100,\ 10.900)\) units. With 50 participants per arm,
and the interval is \(6\pm1.96(2.000)=(2.080,\ 9.920)\) units. The estimated contrast is unchanged; the balanced allocation gives a narrower interval under the stated assumptions. Unequal allocation has not made the estimate biased, but it has reduced precision for this comparison. In a multi-dose trial, how much this matters depends on which contrasts are central and whether a control arm is shared across comparisons.
What Claims Remain Supportable?
The strength and scope of a dose-selection conclusion should match the design’s information. A tested-dose contrast against placebo can be supported even when the dose grid is sparse, provided the comparison is adequately estimated and the analysis is appropriate. A comparison between two tested doses can also be supported, but its uncertainty should be reported directly rather than inferred from separate placebo comparisons. By contrast, sparse levels or a missing region may not support a precise optimum, a detailed curve shape, or a confident claim about an untested dose.
Unequal allocation calls for attention to precision, not an automatic rejection of the results. Sparse coverage calls for attention to what the observed doses identify, not a reflexive conclusion that no selection is possible. A selection can remain defensible among tested doses when the relevant evidence and decision criteria distinguish them sufficiently. If the evidence does not distinguish them, the appropriate conclusion is uncertainty between those options—not a claim that the doses are equivalent.
The main alternatives address different needs. A design with more dose levels can improve coverage of the dose range, though adding levels without adequate participants per arm can leave contrasts weakly estimated. More balanced allocation can improve a particular two-arm comparison under similar variances, but may reduce the sample size available for other comparisons. A model-based analysis can summarize a plausible curve and estimate responses between tested levels, but it introduces assumptions that need to be checked and cannot replace missing observations. A comparison limited to tested doses makes fewer modeling demands, but cannot answer questions about untested doses directly.
| Design feature | What it mainly limits | Claim requiring caution |
|---|---|---|
| Few tested dose levels | Information about response shape | Curvature, plateau, or precise optimum |
| Unequal arm sizes | Precision of contrasts involving smaller arms | Ranking doses when intervals are wide |
| Untested interval within the dose range | Direct evidence in that region | Threshold claims or selection of a dose in the gap |
| No doses beyond the tested range | Evidence for behavior outside observed doses | Extrapolated response or optimum beyond the range |
Diagnostics and Common Pitfalls
Before making a selection claim, inspect the dose-specific sample sizes and outcome variability, the observed arm estimates and their uncertainty, and the location of the tested levels relative to the proposed dose. Check whether the key decision depends on a small arm or a large gap. For model-based predictions, compare predictions under scientifically plausible alternatives and examine whether the decision changes. Strong disagreement between plausible models is evidence that the data do not pin down the conclusion, not a reason to report only the preferred model.
A frequent mistake is to say, “The high-dose mean is largest, so high dose is best.” This reasoning treats a noisy point estimate as if it were a precise ranking and ignores the direct high-versus-low uncertainty. The correction is to quantify the relevant contrast, account for its uncertainty, and state whether the result supports superiority, leaves the ranking unresolved, or supports only a broader selection decision. Another mistake is to describe an untested-dose prediction as an observed treatment effect. Label it as a model-based prediction and state which assumptions generate it.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Two independent arms have equal standard deviations. One arm has 25 participants and the other has 100. Does the unequal allocation by itself bias the difference in sample means? What does it affect?
- A trial tests placebo and one active dose. Can those arms establish whether the response plateaus at a higher dose? Explain.
- A proposed dose lies in a large untested gap, and two plausible dose-response models predict different outcomes there. What is the most defensible interpretation?
- A high-dose arm has a larger estimated mean response than a low-dose arm, but the confidence interval for the direct high-minus-low contrast includes zero and meaningful benefit in either direction. What can be concluded about their ranking?
Solutions. (1) No. Under independent randomized allocation and appropriate estimation, unequal arm sizes do not by themselves bias the difference in means. They affect its precision; with equal variability, the smaller arm contributes more to the standard error through its \(s^2/n\) term. (2) No. Placebo and one active dose can estimate a contrast at that dose, but they provide no observed response at a higher active dose with which to assess a plateau. Any such claim would require additional data or strong model assumptions. (3) The outcome at the proposed dose is model-dependent and is not directly observed. Report the range of plausible predictions and the assumptions behind them; if the selection changes across plausible models, the available design does not support a robust choice there. (4) The point estimate favors high dose, but the direct comparison is too imprecise to establish that high dose is better or that the doses are equivalent. State that their ranking remains unresolved by this comparison.