Translating Exposure-Response Evidence Into Targets

From an Exposure–Response Curve to a Clinical Target

An exposure–response model describes how an outcome changes across exposure levels. A dose-selection analysis needs a further step: it must state what outcome would count as adequate efficacy, and by when. Only then can the model be used to identify the exposure level associated with that outcome. The resulting target is not itself a dose, and a modeled exposure–response association is not automatically proof that changing exposure will cause the predicted benefit.

The earlier material on Defining the Dose-Selection Estimand establishes why the target must match the intended population and clinical question. Here, the emphasis is on making the efficacy criterion explicit: which endpoint, which response threshold, and which assessment time. These choices determine the meaning of any exposure target.

A response criterion might be a mean improvement of at least 5 points on a validated scale by week 12, or a probability of clinical response of at least 70% by week 12. These are different targets. The first concerns a mean; the second concerns the chance of a defined individual-level response. A target should also identify the population and endpoint definition to which it applies.

The exposure measure must be defined consistently with the evidence supporting the model. It could be a summary of concentration over a period or another exposure quantity, but the target is only interpretable if its units, calculation, and relevant time period are clear. The response threshold should come from clinical knowledge—such as a validated meaningful-change criterion or a clinically justified response definition—not merely from a convenient feature of the fitted curve.

Assumptions Before Translating a Target

A model-based translation is conditional on both statistical and clinical assumptions. Before using an exposure–response curve to calculate a target, check that:

• The endpoint and threshold are clinically justified. The measurement, direction of benefit, and threshold correspond to a meaningful patient or clinical outcome.
• The time horizon is specified. A response criterion at week 4 is not interchangeable with the same criterion at week 12 unless the model and clinical question justify that equivalence.
• The model represents the relevant population. Predictions should address the intended population, including any needed covariate standardization, rather than silently describing a different group.
• The exposure–response relationship is sufficiently supported. The data should cover exposures relevant to the proposed target, and model fit and uncertainty should be examined. A target far beyond observed exposure is an extrapolation.
• The interpretation of exposure is defensible. Exposure may be associated with prognosis, adherence, disease severity, or other factors. The model and supporting evidence must address plausible alternative explanations before the relationship is interpreted as the effect of changing exposure.
• The response model is appropriate for the endpoint. A model for mean change does not directly answer a question about the probability of individual response, and vice versa.

These conditions do not guarantee that a target is clinically useful, but they make its meaning and evidential limits assessable. In particular, a mathematically solvable equation cannot supply a clinical rationale for the threshold used in that equation.

Translating a Mean-Response Threshold

Suppose \(B(E)\) is the model-predicted mean benefit at a specified time, expressed so that larger values indicate more benefit, and \(E\) is a defined exposure measure. A common saturating relationship is the Emax model:

Here, \(B_{\max}\) is the maximum modeled benefit, in endpoint units, and \(EC_{50}\) is the exposure associated with half of that maximum, in exposure units. If the required mean benefit is \(B^*\), the target solves \(B(E)\geq B^*\). For a threshold below the model’s maximum, the boundary exposure is obtained by rearranging the equation.

This solution exists at a finite, nonnegative exposure only when \(0\leq B^*<B_{\max}\). If \(B^*=B_{\max}\), the model approaches the threshold only as exposure increases without bound; if \(B^*>B_{\max}\), the model cannot reach the threshold at any exposure. This is a useful check before reporting a numerical target.

Probability-Based Response Targets

When the clinical criterion defines whether each person is a responder, the target may instead be expressed as a probability. Let \(R(\tau)=1\) indicate that a person meets the prespecified response definition by time \(\tau\). A model can then predict \(P(R(\tau)=1\mid E=e)\), or a version standardized to a specified population. The target exposure is the exposure at which this predicted probability reaches a stated criterion.

This probability is not the same as a mean change. It depends on the response definition and, where the model includes them, the distribution of individual differences and covariates. State whether a prediction is conditional on particular covariates or averaged over a specified population. Otherwise, the target can appear to describe all patients when it actually describes a selected profile.

Why the Time Horizon Changes the Target

An exposure–response target is incomplete without its assessment time when response develops over time. An exposure that produces a particular eventual response may not produce it early enough to meet a short-term clinical objective. A time-dependent model makes this dependence explicit. For example, consider a hypothetical mean-benefit model:

Benefit is in points, exposure \(E\) is in mg·h/L, and time \(t\) is in weeks. The first factor is the exposure-dependent limiting benefit; the second describes the fraction of that limiting benefit reached by time \(t\). This form assumes the same time-course parameter across exposure levels and a response that approaches its limit monotonically. Those assumptions would need support before applying the model clinically.

Evidence, Diagnostics, and Alternatives

A proposed target is more defensible when several kinds of evidence align: the threshold has a clinical rationale; the endpoint is measured reliably; the assessment time reflects the clinical decision; and the exposure–response data support the model in the region where the target lies. Supporting evidence may include clinical outcome studies, validated response definitions, relevant prior clinical knowledge, and checks that the modeled relationship is not mainly a consequence of confounding or selective exposure patterns.

Diagnostics should focus on the translation as well as the fitted model. Check whether the target lies within observed exposure support, whether the fitted curve captures the data near the threshold, and whether the prediction changes substantially under plausible alternative model specifications. Examine uncertainty in model parameters and in the translated target. A narrow-looking point estimate does not resolve uncertainty about the clinical threshold or the model’s applicability.

These approaches answer related but nonidentical questions. A mean-response target is appropriate when the clinical decision is framed around average benefit; a probability-based target is appropriate when the decision requires a specified fraction of people to meet a response definition. Neither should be substituted for the other without a clinical and statistical justification. Direct clinical evidence can anchor the meaning of a threshold, while an exposure–response model can describe how predicted outcomes vary with exposure; each has limits when used alone.

Common warning signs include a target above the model’s attainable maximum, a target far beyond observed exposure, a threshold chosen after viewing the curve without clinical justification, and a time horizon omitted from the result. Another pitfall is to treat an exposure association as causal: patients with higher exposure may differ in adherence, disease severity, or other predictors of outcome. The correction is to describe what the fitted model predicts, assess plausible sources of bias, and claim a causal treatment effect only when the study design and analysis support that interpretation.