What an Exposure-Response Result Can Establish
An exposure-response analysis relates a measure of drug exposure to an outcome, such as efficacy or a safety event. Exposure might be represented by area under the concentration-time curve, peak concentration, trough concentration, or another defined summary over a specified time window. The central review question is not only whether the fitted relationship is statistically or clinically notable. It is what comparison the data and analysis can support.
Exposure is often not assigned directly. Patients assigned the same dose can have different exposures because of clearance, body size, adherence, interacting medicines, or other factors. Those factors may also be related to the outcome. Further, doses may be selected or adjusted in response to a patient's prognosis or evolving clinical condition. Consequently, an exposure-response relationship may be useful for prediction without showing that an intervention that changes exposure would change the outcome.
A causal contrast can be written as the difference between the mean outcomes that would occur under two specified exposure interventions, \(e_1\) and \(e_0\):
Here, \(Y(e)\) denotes the outcome that would occur under intervention \(e\). This notation requires the exposure intervention to be meaningful and sufficiently specified: “set exposure to \(e\)” must correspond to a feasible intervention, with a defined timing and exposure measure. A model predicting outcomes at two exposure values is not by itself evidence for this contrast.
Assumptions Behind a Causal Interpretation
Causal interpretation requires more than a well-fitting exposure-response curve. In a simplified setting with a baseline exposure and outcome, the usual requirements include consistency (the observed outcome corresponds to the intervention actually received), exchangeability (no unaccounted common causes of exposure and outcome after conditioning on measured covariates), positivity (each exposure being compared is possible across the relevant covariate profiles), and adequate measurement and model specification. These assumptions are substantive and cannot be proved merely by fitting the model.
In pharmacometrics reports, a particularly important source of non-exchangeability is dose selection. Investigators or clinicians may assign higher doses to patients expected to benefit more, or escalate doses for patients whose response is inadequate. In either case, patient characteristics or clinical history can affect both exposure and outcome. In addition, within a dose group, clearance or adherence may relate exposure to prognosis. Randomization of assigned dose supports a causal comparison of assigned dose under the trial's conditions; it does not automatically randomize achieved exposure.
A simple derivation shows how an omitted common cause can distort an exposure slope. Suppose the outcome follows \(Y=\beta E+\gamma U+\varepsilon\), where \(E\) is exposure, \(U\) is an unmeasured prognostic factor, and \(\varepsilon\) has zero covariance with \(E\). If a linear regression omits \(U\), its population slope is:
Unless \(\gamma=0\) or \(E\) and \(U\) are uncorrelated, the fitted slope need not equal the exposure effect \(\beta\). Adjusting for measured baseline covariates can help when those covariates are appropriate common causes, but it does not remove bias from unmeasured causes, poor measurement, or a misspecified relationship.
Worked Example: A Pooled Association Reverses the Within-Group Pattern
Consider a hypothetical observational report comparing response rates in patients with high versus low exposure. Baseline disease severity may affect both treatment decisions and response. The counts are:
| Baseline severity | High exposure | Low exposure |
|---|---|---|
| Mild | 45 of 50 responded (90%) | 9 of 10 responded (90%) |
| Severe | 2 of 10 responded (20%) | 18 of 30 responded (60%) |
The pooled response rate at high exposure is \(47/60=0.783\), or 78.3%. At low exposure it is \(27/40=0.675\), or 67.5%. The pooled difference is \(0.783-0.675=0.108\), an apparent 10.8 percentage-point advantage for high exposure.
That pooled comparison hides the severity imbalance: 50 of 60 high-exposure patients (83.3%) had mild disease, compared with 10 of 40 low-exposure patients (25.0%). Within the mild group, response rates are equal; within the severe group, the high-exposure rate is lower by 40 percentage points. Thus, these data do not support the simple pooled claim that high exposure improves response. Severity and the way exposure was selected need scrutiny. Even the stratified comparisons are observational and do not alone establish causal effects.
Overlap, Timing, and Model-Based Extrapolation
A fitted curve can produce a prediction at any exposure value allowed by its mathematical form. That does not mean the data support every prediction equally. Exposure overlap asks whether patients with relevant covariate profiles are observed across the exposure range being compared. If low exposures occur only in one dose group and high exposures only in another, it may be difficult to separate exposure from dose group, patient characteristics, or other differences. Predictions in sparsely represented regions, gaps between observed ranges, or beyond the observed range rely increasingly on model assumptions.
The reviewer should inspect the exposure distributions overall and by important baseline factors, dose group, and relevant clinical history. A pooled range can look broad while specific subgroups have little or no overlap. Lack of overlap does not make a model-based prediction impossible, but it weakens empirical support for comparisons across the gap. The report should identify where the data are informative and where the curve is being driven chiefly by its assumed functional form.
Worked Example: A Prediction Across a Gap in Observed Exposure
Suppose a hypothetical analysis summarizes two dose groups using their mean exposure and mean response. The lower-dose group has mean exposure 1.5 exposure units and mean response 45 outcome units; the higher-dose group has mean exposure 4.5 exposure units and mean response 69 outcome units. A linear model fitted to these two summaries has slope:
The fitted line has intercept \(45-(8\times1.5)=33\), so its prediction at exposure 3 is \(33+(8\times3)=57\) outcome units. This is a model-based prediction. If observed exposures were concentrated from 1 to 2 in the lower-dose group and from 4 to 5 in the higher-dose group, exposure 3 lies in an unobserved gap. The model supplies the prediction by assuming a linear relationship across that gap; there are no observations there to check that assumption directly.
The calculation neither proves a causal effect of increasing exposure nor shows that the line is appropriate. A reviewer should ask whether individual-level data support the shape, whether groups differ in prognostic factors, and whether uncertainty around the prediction is reported. The slope's units and the prediction's meaning should also match the outcome and exposure definitions in the report.
Timing is equally important. Exposure used to explain an outcome should be defined over a window that precedes that outcome, or the analysis should otherwise account for their temporal relationship. If a patient's worsening condition leads to a dose increase, then later high exposure can be associated with poor outcome because deterioration preceded the increase. A contemporaneous summary may obscure this sequence or invite reverse-causal interpretation.
Repeated measures create an additional complication: a patient's earlier exposure may affect clinical status, while that status may influence later dose and exposure. Simply adjusting for the latest status can be inadequate. Depending on the question and data, longitudinal causal methods may be needed, with explicit assumptions about measured time-varying causes and treatment history. A standard exposure-response model should not be presented as having resolved this problem unless its specification and interpretation justify that conclusion.
Worked Example: Escalation After Early Deterioration
In a hypothetical study, 20 patients have their status assessed at week 2. Ten have deteriorated by then and receive dose escalation, resulting in higher exposure later in treatment. At the end of treatment, 8 of those 10 patients have a poor outcome. The other 10 patients did not deteriorate by week 2 and are not escalated; 2 of them have a poor outcome.
A crude comparison reports poor outcomes in 80% of the later high-exposure group, \(8/10=0.80\), and 20% of the lower-exposure group, \(2/10=0.20\). The crude difference is \(0.80-0.20=0.60\), or 60 percentage points higher in the high-exposure group.
But the status that prompted escalation was measured before the later exposure. Early deterioration predicts the final outcome and determines who receives higher exposure. The crude difference therefore cannot be interpreted as evidence that high exposure caused poorer outcomes. Nor does this example establish that escalation helped or had no effect: it lacks the counterfactual comparison needed to determine what would have happened to the deteriorating patients without escalation. A sound review asks whether exposure and outcome windows are ordered appropriately and whether the analysis handles the clinical history that drives dose changes.
Reviewing the Claim and Its Alternatives
The appropriate interpretation depends on the analysis and the decision question. A randomized assigned-dose comparison is often the clearer causal comparison for the effect of assignment to those doses, though it may not answer the effect of achieved exposure. An exposure-response model can characterize a relationship over the observed data and support prediction in populations and settings sufficiently similar to those represented. A causal analysis of achieved exposure requires additional design or modeling assumptions. These approaches answer related but distinct questions.
| Approach | What it can address | Key limitation for causal interpretation |
|---|---|---|
| Randomized assigned-dose comparison | Outcome contrast by randomized dose assignment | Achieved exposure varies within dose groups and is not itself randomized |
| Exposure-response model | Modeled outcome pattern across observed exposure | Confounding, dose selection, limited overlap, or model form can drive the pattern |
| Longitudinal causal analysis | A specified intervention contrast with evolving exposure and clinical history | Requires suitable data, measured confounders, positivity, and correct assumptions |
In a report, the conclusion should be no broader than the evidence. State the exposure metric and time window, the analyzed population, the outcome, the range with data support, and whether the result is an association, prediction, or causal estimate. Where the intended decision concerns changing a dose, explain how the exposure analysis relates to randomized dose evidence and why its assumptions are credible. A precise limitation is more useful than a generic statement that “confounding is possible.”
Diagnostics and Common Pitfalls
Warning signs include a steep association driven by a few patients at exposure extremes; different exposure distributions across dose groups or clinical subgroups; a curve that changes materially when its functional form changes; an exposure window that overlaps the outcome window; and dose escalation triggered by prior lack of response. These findings do not automatically invalidate the analysis. They indicate that the reviewer should narrow the claim, examine support in the data, or request a clearer justification of the model and interpretation.
- Wrong reasoning: “The exposure-response curve is smooth and fits the observed outcomes, so increasing exposure will improve response.” Correction: Fit and predictive agreement do not establish exchangeability, adequate overlap, correct timing, or a causal effect.
- Wrong reasoning: “Dose was randomized, so the relationship between achieved exposure and outcome is randomized.” Correction: Randomization applies to assigned dose; achieved exposure can still depend on patient characteristics, adherence, and clinical history.
- Wrong reasoning: “The model predicts an outcome at this exposure, so that exposure is supported by observed patients.” Correction: A prediction may be model-based extrapolation or interpolation across a data gap. Check the observed exposure distribution and state the limitation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A randomized trial assigns two doses, but individual exposure varies substantially within each dose group. What does randomization directly support, and what does it not automatically establish about exposure?
- A report predicts response at an exposure value not represented in one important subgroup. What should a reviewer check before treating that prediction as well supported?
- A clinician increases dose after early worsening, and later high exposure is associated with poor outcome. Why is the association not enough to conclude that high exposure caused harm?
- What is the difference between a predictive exposure-response association and a causal exposure effect?
Solutions. (1) Randomization directly supports a comparison by assigned dose under the trial conditions; it does not randomize achieved exposure within dose groups. (2) Check subgroup-specific exposure support, overlap, the model's assumed shape, uncertainty, and whether the prediction depends on extrapolation. (3) Early worsening precedes and influences escalation and also predicts later outcome, so the crude association is confounded by clinical history; it cannot establish the effect of high exposure. (4) A predictive association describes outcomes expected under the fitted model and relevant data distribution; a causal effect compares outcomes under specified exposure interventions and requires assumptions that identify that intervention contrast.