Identifying the Therapeutic Window From Dose-Response Evidence

From Two Dose Thresholds to a Therapeutic Window

Dose selection often requires two different judgments: how much dose is needed for clinically meaningful benefit, and how much dose can be given before safety becomes unacceptable. Dose-response evidence can inform both judgments, but their thresholds need not be estimated with equal precision or from the same endpoint. The therapeutic window puts them on a common dose scale and makes their separation, or lack of separation, explicit.

A window is not an intrinsic property of a drug independent of context. It depends on the population, regimen, efficacy endpoint, safety endpoint, clinical criteria, and period over which outcomes are assessed. The criteria should be specified before interpreting the fitted dose-response evidence; otherwise, thresholds can be chosen after seeing the results to favor a desired dose.

For example, the efficacy criterion might require at least an 8-percentage-point improvement over placebo on a clinically interpretable response measure. The safety criterion might require that the excess incidence of a specified adverse event not exceed 5 percentage points. These are hypothetical decision criteria, not universal cutoffs. A therapeutic window is meaningful only if both criteria are relevant to the intended treatment context.

The threshold estimates may come from fitted dose-response curves, dose-level estimates, or a prespecified interpolation between doses. They should reflect the evidence for the relevant endpoints, rather than an assumption that efficacy and toxicity share a common curve. A monotone relationship may be a reasonable approximation over a limited dose range, but it must be checked against the data and the scientific setting. Earlier material on diagnosing dose-response model misspecification and evaluating nonmonotonic dose-response is relevant when that assumption is doubtful.

Quantifying Whether the Thresholds Are Separated

The key uncertainty question is not just whether each threshold has a wide interval. It is whether the difference \(W=d_S-d_E\) is positive, near zero, or negative, accounting for the uncertainty and dependence between the two estimates. If the estimated thresholds are approximately jointly normal, the variance of their difference is

This follows from the variance rule for a difference: \(\operatorname{Var}(A-B)=\operatorname{Var}(A)+\operatorname{Var}(B)-2\operatorname{Cov}(A,B)\). The estimated standard error is the square root of this variance. A large-sample 95% confidence interval for \(W\) is \(\hat W \pm 1.96\,SE(\hat W)\), if the joint normal approximation is adequate. A lower limit above zero supports separation under the model and criteria; an interval that includes zero leaves the ordering uncertain. A confidence interval below zero indicates that the estimated safety limit is lower than the efficacy threshold, though uncertainty still matters.

The covariance term matters because the thresholds may be estimated from overlapping participants, shared dose groups, or correlated endpoint models. Ignoring dependence can overstate or understate uncertainty. When threshold estimates are skewed, near a boundary, or derived from complex models, joint bootstrap or simulation draws can be more appropriate than a normal interval. In each draw, calculate the safety threshold minus the efficacy threshold; summarize that distribution directly.

When Threshold Uncertainty Overlaps

When estimated thresholds are close, a point-estimate window can give a false impression of certainty. The direct quantity of interest remains \(W\), with its uncertainty. Separate confidence intervals for \(d_E\) and \(d_S\) are useful descriptions, but their overlap or nonoverlap is not itself a formal test of whether \(W\) is positive. In particular, overlapping marginal intervals do not prove there is no window, and nonoverlapping marginal intervals should not substitute for a joint uncertainty calculation.

A further practical complication is that thresholds are estimated on a continuous dose scale while trials often test only a few dose levels. Interpolation is most defensible between observed doses when the dose-response relationship is adequately described locally. Extrapolating beyond the tested range can yield precise-looking thresholds unsupported by direct evidence. Sparse-dose methods and model averaging can help characterize model uncertainty, but they cannot make unobserved dose regions empirically well established.

Diagnostics, Alternatives, and Limitations

A threshold-based summary is only as credible as the dose-response evidence and decision criteria behind it. Inspect the dose-level estimates and uncertainty, not just the final interval. Warning signs include a threshold beyond the tested dose range, a crossing driven by one imprecise dose group, substantial curvature that conflicts with the interpolation rule, or a safety endpoint with too few events to estimate its dose trend reliably. If an estimated curve crosses a criterion multiple times, a single lower or upper threshold may not describe the acceptable doses; the response pattern needs to be represented without forcing a one-window summary.

The principal alternative summaries answer related but different questions. Reporting \(d_E\) and \(d_S\) separately makes the clinical criteria visible but does not directly show their separation. Comparing marginal confidence intervals is easy to communicate, but does not correctly quantify uncertainty in \(W\). A joint model or paired bootstrap directly estimates the width distribution and preserves dependence, though it relies on adequate model specification and resampling assumptions. Dose-level tables avoid some model dependence but can leave the threshold between tested doses unresolved. These approaches complement rather than replace clinical judgment.

The window also simplifies variation across patients, endpoints, and treatment conditions. A population-level threshold does not imply that all patients benefit or tolerate the same doses. Different safety endpoints may impose different upper limits, and a threshold based on one endpoint does not guarantee that all safety concerns are acceptable. In addition, uncertainty about the chosen dose-response model should be reflected where material; a narrow interval conditional on one model can understate total uncertainty. The introductory material on applying safety constraints to efficacy-based dose selection provides the immediate decision context, while this analysis focuses on how the two threshold estimates and their uncertainty define a window.

A defensible report names the efficacy and safety criteria, gives the threshold estimates and units, explains how they were obtained, and describes uncertainty in their difference. It should also identify whether the thresholds rely on interpolation or extrapolation and state the main assumptions. If uncertainty is substantial, report that directly instead of presenting a narrow dose interval as a firm recommendation.