Translating Clinical Objectives Into Dose-Selection Criteria

From Clinical Intent to a Dose Rule

A dose is not justified simply because its efficacy estimate is larger than that of a lower dose, or because a safety result is not statistically significant. A defensible choice starts with a clinical objective: what improvement would matter to patients in this treatment setting, and what level of harm would be unacceptable? Those statements must be translated into criteria that can be compared with the evidence for each candidate dose.

This translation is a decision task, not a search for a universal “best” threshold. The acceptable balance can differ between a short course for a severe condition and long-term treatment of a mildly symptomatic condition. It also depends on the available alternatives, the intended patient population, the treatment duration, and how outcomes are measured. The earlier material on exposure-response therapeutic windows helps characterize where efficacy and safety may overlap; here the focus is on expressing the clinical objective as a usable dose-selection rule.

For example, “the treatment should work well and be safe” is not operational. A more useful objective might require at least a 4-point improvement on a specified symptom scale relative to control, while limiting the excess risk of a specified adverse event to no more than 1.5 percentage points over a stated treatment period. The numbers must be justified clinically, not chosen because they make an observed dose pass.

Define What Counts as Benefit and Harm

First state the contrast and its direction. Suppose a positive efficacy quantity \(\Delta_E(d)\) represents benefit at dose \(d\) relative to a comparator. A minimum worthwhile benefit is denoted \(\delta_{\min}\); the basic efficacy requirement is then \(\Delta_E(d)\geq\delta_{\min}\). If lower outcome values indicate improvement, the contrast must be defined with the sign reversed or described in a way that still makes positive values favorable. This prevents a threshold from being applied in the wrong direction.

For safety, let \(\Delta_S(d)\) be the increase in risk of a specified adverse outcome at dose \(d\) relative to the comparator. A maximum acceptable excess risk \(r_{\max}\) gives the requirement \(\Delta_S(d)\leq r_{\max}\). This is not interchangeable with a limit on the absolute event risk. If a clinical objective concerns absolute risk, write an absolute-risk ceiling instead. Relative risks, risk differences, and event rates answer different questions, so the criterion and the evidence must use the same measure.

The clinical context should be concrete enough to interpret both quantities. State the target population, dose and regimen, comparator, treatment duration, and outcome measurement. A 2-percentage-point excess risk may be acceptable for a brief treatment of a life-threatening illness but unacceptable for years of treatment of a mild condition. Likewise, a change on a symptom scale is interpretable only when the scale, direction, and assessment period are known.

Make Uncertainty Part of the Rule

A point estimate alone can conceal substantial uncertainty. One practical approach is to require a lower confidence bound for benefit to exceed the minimum worthwhile benefit, and an upper confidence bound for harm to remain below its maximum acceptable level. In a simple normal-approximation setting, these one-sided bounds are:

Here, \(\widehat{\Delta}_E\) and \(\widehat{\Delta}_S\) are estimated effects, \(SE_E\) and \(SE_S\) are their standard errors, and \(z_{1-\alpha}\) is the standard normal quantile for the chosen one-sided confidence level. Under the stated approximation, the efficacy gate is \(L_E\geq\delta_{\min}\), and the safety gate is \(U_S\leq r_{\max}\). This rule requires evidence not only that the estimated effect is favorable, but that a lower bound for benefit is large enough and an upper bound for harm is small enough.

The bound formulas follow directly from the assumed sampling distribution. If an estimate is approximately normal with standard error \(SE\), then \((\widehat{\Delta}-\Delta)/SE\) is approximately standard normal. Subtracting \(z_{1-\alpha}SE\) gives a lower one-sided confidence bound for a benefit contrast; adding it gives an upper bound for a harm contrast. The confidence level describes the long-run coverage of the procedure under its assumptions. It does not mean there is a \(1-\alpha\) probability that the fixed true effect lies in the calculated interval.

These formulas are not appropriate by default for every endpoint. Their use depends on a suitable sampling approximation and a valid standard error. For rare events, sparse data, or complex model-based estimates, a normal approximation may be poor; use an appropriate interval method and check its assumptions. The clinical logic of the gate remains, but the calculation of its uncertainty must fit the data and analysis.

Worked Examples

Phase 2 and Phase 3 Use

In Phase 2, dose selection often aims to identify one or more reasonable candidates for later study. Criteria should still have clinical meaning, but uncertainty may be substantial because exposure, response, and safety data are limited. A Phase 2 rule can distinguish a minimum signal worth pursuing from a dose that is already sufficiently supported for confirmation. It should also identify evidence that would stop advancement, such as a clearly unacceptable risk. A dose that does not pass a stringent confidence-bound gate may remain uncertain rather than proven ineffective; the decision depends on the stated purpose of the screen and the totality of evidence.

Phase 3 dose selection has higher consequences because a dose is intended for larger, confirmatory studies and potentially clinical use. The criteria and evidence standard should be consistent with the intended claim and treatment context. This commonly calls for stronger support that worthwhile benefit is achieved without exceeding acceptable safety limits. Reusing a permissive screening rule without adjustment can overstate what Phase 2 evidence establishes. Conversely, a rigid threshold should not be treated as a substitute for clinical judgment when evidence is limited or outcomes differ in severity.

A useful decision table makes the rule transparent. It can show, for each candidate dose, the efficacy estimate and bound, the relevant safety estimate and bound, whether each gate passes, and the resulting action. The action should be tied to the development decision: for example, “advance for further evaluation,” “do not advance because the safety limit is exceeded,” or “insufficient evidence to classify.” Keeping “fails a gate” separate from “insufficient evidence” avoids interpreting uncertainty as proof of no benefit or proof of harm.

Diagnostics, Alternatives, and Limitations

Before applying a criterion, check that the evidence corresponds to its intended context. A mismatch in population, comparator, follow-up, dose regimen, or outcome definition can make a numerically precise result clinically irrelevant. Check the direction and units of every contrast, and confirm whether the threshold applies to an absolute outcome, a difference, or a ratio. For model-based results, assess whether the model adequately describes the observed efficacy and safety patterns; poor fit or unstable estimates weaken the decision even if a calculated bound crosses its threshold.

A common mistake is to say, “The benefit was statistically significant, so the dose is clinically worthwhile.” Statistical significance tests whether data are inconsistent with a specified null value under a model; it does not establish that an effect exceeds a clinically meaningful minimum. The correction is to state the clinical threshold explicitly and compare the relevant estimate, with its uncertainty, to that threshold. The reverse mistake also occurs: a non-significant safety result is treated as proof that risk is acceptable. A wide upper confidence bound may still include clinically unacceptable harm.

A second mistake is to combine benefit and harm using an unexplained single score. A weighted benefit-risk score can be useful when weights reflect explicit clinical preferences, but its result may hide which safety limit failed or which benefit was judged important. A set of separate, prespecified gates is often easier to communicate. Neither approach is universally best: a constrained rule is clear when some harms must not exceed defined limits, whereas a utility-based approach can represent trade-offs when preferences are defensibly quantified. In either case, thresholds and weights should be justified in advance and checked for sensitivity to reasonable alternatives.

Other limitations matter. Multiple endpoints and repeated dose comparisons can make apparent threshold crossings sensitive to selective reporting. Small Phase 2 samples may produce unstable safety bounds, particularly for rare events. A threshold that is appropriate for one population or duration may not transfer to another. Finally, a binary pass/fail label can obscure how close the result is to a boundary. Report the estimates, uncertainty, units, and decision rationale, not just the label.