Justifying a Dose When Efficacy and Safety Favor Different Doses

When the Efficacy-Favoring Dose Is Not the Safety-Favoring Dose

A higher dose may improve an efficacy endpoint more than a lower dose while also producing more adverse events. The efficacy ranking alone does not establish which dose should be used. Nor does observing more adverse events at the higher dose, by itself, show that its benefit is not worth the risk. A defensible dose recommendation makes the clinical tradeoff explicit: what efficacy is sufficient, how much additional efficacy would justify a higher dose, and what level of safety risk is acceptable?

The key is to distinguish three claims. A dose can meet a minimum efficacy requirement; its safety risk can be below a defined limit; and the extra efficacy of a higher dose can be too small to justify its extra risk. These claims require different comparisons. They should be tied to the relevant clinical outcomes, treatment duration, patient population, and decision context rather than inferred from which estimate is numerically largest.

State the Decision Criteria Before Comparing Doses

Let \(E(d)\) denote the efficacy effect at dose \(d\), with larger values representing better efficacy. Let \(R(d)\) denote the risk of a specified safety event over a defined period, with larger values representing greater risk. Contrasts must use consistent definitions and reference groups. For instance, the excess event risk at dose \(d\) relative to placebo is \(R(d)-R(0)\), while the efficacy gain at a higher dose \(d_H\) relative to a lower dose \(d_L\) is \(E(d_H)-E(d_L)\).

A transparent decision can specify three criteria. First, require the selected dose to meet a minimum efficacy threshold relative to placebo or another relevant comparator. Second, require its safety risk to remain below a maximum acceptable limit. Third, if choosing the lower dose, define how much efficacy it may give up relative to the higher dose. That last criterion prevents “the doses are not significantly different” from being mistaken for evidence that their efficacy is sufficiently similar.

Here \(\operatorname{LCB}\) and \(\operatorname{UCB}\) are lower and upper confidence bounds from an analysis appropriate to the endpoint and design. These inequalities describe one conservative decision rule, not a universal regulatory formula. The thresholds and confidence level must be justified for the clinical context. If safety is instead framed as an absolute risk limit, compare an upper bound for \(R(d)\) with that limit. The relevant formulation should be chosen in advance and kept consistent.

A common convention is to use one-sided 95% bounds. Under an approximately normal sampling distribution with a valid standard error, an estimate \(\hat\theta\) has lower and upper bounds \(\hat\theta-1.645SE\) and \(\hat\theta+1.645SE\), respectively. This calculation relies on the analysis providing an appropriate estimate and standard error. Model-based contrasts should use their model-based standard errors, including covariance between dose estimates; one should not assume that separate estimates are independent unless that assumption is justified.

Why the Confidence-Bound Rule Matches the Decision

Suppose the clinical question is whether a lower dose meets a minimum efficacy threshold \(M_E\). The relevant contrast is \(\theta=E(d_L)-E(0)\). If a one-sided lower confidence bound for \(\theta\) exceeds \(M_E\), the data support an effect greater than the minimum under the stated confidence procedure. Likewise, if the question is whether the higher dose adds more than a tolerable efficacy loss \(\delta_E\), define \(\Delta=E(d_H)-E(d_L)\). A one-sided upper bound below \(\delta_E\) supports the conclusion that the added efficacy is smaller than the prespecified margin.

The same logic applies to safety. For excess risk \(\rho=R(d)-R(0)\), an upper bound below \(M_S\) supports the conclusion that excess risk is below the chosen limit. A point estimate below the limit is not enough when its uncertainty extends above it. Conversely, an upper bound above the limit does not prove that the risk exceeds the limit; it means the data do not establish that it is below the limit using this rule.

These are confidence-procedure statements, not probabilities that a fixed parameter lies inside a particular interval. Their interpretation depends on a suitable estimand, a sound statistical model or estimation method, and adequate handling of missing data, multiplicity, and trial conduct. In a randomized comparison, the dose groups should be compared using the prespecified population and follow-up period. If estimates are adjusted or model-based, their interpretation should remain aligned with the clinical contrast the threshold was designed to assess.

When the Higher Dose Can Still Be Defensible

A safety-favoring lower dose does not automatically win. If the higher dose produces an incremental efficacy gain that is clinically important and its risk remains below a justified safety limit, selecting the higher dose can be defensible. This argument should show both that the higher dose clears the efficacy requirement and that the safety cost is acceptable. A larger efficacy estimate without a clinical threshold does not establish that the extra benefit is worth the extra risk.

When the Evidence Does Not Resolve the Tradeoff

Sometimes the lower dose clearly meets its efficacy requirement, while the higher dose has a promising efficacy estimate but uncertain safety. The correct conclusion may be that the available evidence supports the lower dose, while leaving open whether the higher dose would be acceptable with more precise evidence. Failure to establish safety below a limit is not proof that the limit is exceeded. Equally, a reassuring point estimate is not enough to establish that it is below the limit.

Diagnostics, Alternatives, and Common Pitfalls

Before using a dose-justification rule, check that the efficacy and safety contrasts correspond to the same relevant population and treatment period. Confirm the endpoint direction, units, reference group, and event definition. Inspect confidence intervals as well as point estimates, and verify that the standard errors correspond to the actual dose contrasts. A threshold based on a short exposure period cannot automatically be applied to a materially longer treatment duration. If the analysis changes substantially under reasonable endpoint definitions or missing-data assumptions, the recommendation should disclose that sensitivity rather than present a single contrast as definitive.

There are several alternatives to a threshold-based argument, but each answers a different question. An efficacy-only choice selects the dose with the largest efficacy estimate and may ignore an important safety cost. A safety-first rule can exclude a dose that exceeds a hard clinical constraint, but it does not by itself identify the best dose among those that satisfy the constraint. A single composite score can combine benefits and harms, but its result depends on how outcomes are scaled and weighted. A threshold-based argument instead makes the boundaries visible and explains which dose meets them. It does not replace clinical judgment; it makes the judgment inspectable.

• Pitfall: “The higher dose was not significantly better, so the lower dose is equally effective.” A nonsignificant difference does not show that the efficacy loss is smaller than a clinical margin. Specify \(\delta_E\) and assess the appropriate confidence bound for the dose contrast.
• Pitfall: “The safety estimate is below the limit, so the dose is safe.” The estimate may be imprecise. Compare an appropriate upper bound with the limit, and describe a failed criterion as uncertainty when the evidence does not establish that risk exceeds the limit.
• Pitfall: “The higher dose has the best efficacy estimate, so it is the best dose.” A numerical ranking does not quantify whether the incremental benefit is clinically worthwhile or whether the safety cost is acceptable.
• Pitfall: choosing thresholds after seeing which dose passes. This weakens the interpretability of the decision. Thresholds should be justified in clinical terms and specified before using the trial results to select among doses.

Threshold-based dose justification has limits. A threshold may not capture all patient preferences, the evidence may be too imprecise to satisfy either criterion, and safety risks can differ in severity even when their frequencies are similar. A recommendation should therefore state the population and conditions to which it applies, identify unresolved uncertainties, and avoid implying that passing a specified criterion proves the absence of other risks.