Testing Regional Consistency of Dose Response

What Regional Consistency Means

When dose-response evidence is considered across regions, the question is not simply whether outcomes differ between regions. Regions may differ in baseline risk, covariate distribution, dose assignment, or drug exposure, even when the treatment response follows the same pattern. The relevant question is whether the dose-response relationship differs across regions on a defined outcome scale, after accounting for the model and covariates used in the comparison.

A regional interaction represents a difference in the dose-response relationship. Its meaning depends on the model: an interaction with dose in a linear model compares slopes; interactions with dose indicators compare dose-specific effects; and interactions with nonlinear dose-response terms compare curve parameters or predictions. A nonsignificant interaction test does not establish that the curves are sufficiently similar for a bridging decision. That requires interpreting uncertainty against a prespecified, clinically meaningful difference.

Modeling Regional Differences

For a simple continuous outcome, let \(Y\) be an outcome for an individual receiving dose \(d\), and let \(R=1\) indicate one region while \(R=0\) indicates a reference region. A linear dose-response model with a regional intercept and slope difference is

Here, \(\beta\) is the dose slope in the reference region, \(\gamma\) is the regional difference in expected outcome at dose zero, and \(\delta\) is the difference between regional dose slopes. The slope in the second region is \(\beta+\delta\). If the scientific question is whether the slopes differ, the null hypothesis is \(H_0:\delta=0\). If the question is whether the entire curves are identical, both the intercept and slope differences matter; in that case, the joint null is \(H_0:\gamma=0,\ \delta=0\).

For nonlinear curves, dose can be represented by a set of basis functions, such as dose categories, a polynomial, or a prespecified pharmacological function. A regional interaction is then a region-specific coefficient for one or more of those functions. The test of equal curves is a joint test that all relevant interaction coefficients are zero. A comparison should state the outcome scale and the dose range: two curves can be close at studied doses but diverge beyond them.

Interaction Estimates and Uncertainty

For a single interaction coefficient \(\hat{\delta}\) with standard error \(SE(\hat{\delta})\), a two-sided Wald test uses \(z=\hat{\delta}/SE(\hat{\delta})\). A large-sample 95% confidence interval is \(\hat{\delta}\pm1.96SE(\hat{\delta})\). The interval describes values compatible with the data and model at that confidence level; it is not a probability statement about the fixed, unknown coefficient. Its units are the units of outcome per unit of dose.

For several interaction coefficients, collect their estimates in a vector \(\hat{\boldsymbol{\delta}}\), and let \(V\) be their estimated covariance matrix. If there are \(k\) coefficients, the Wald statistic for the joint null is

This joint test assesses whether the interaction terms, taken together, differ from zero. It does not quantify whether any difference is clinically important. For that, examine confidence intervals for interpretable contrasts, such as the regional difference in predicted response at a selected dose. If acceptable differences are defined by a margin, evidence of equivalence requires the relevant interval to lie wholly within that margin; a conventional test that fails to reject equality is not an equivalence test.

Joint Comparisons for Curves

A single slope interaction is not enough when the dose-response curve is nonlinear. One direct approach is to estimate regional treatment effects at a set of doses and compare them jointly. Let \(\delta_j\) be the difference between regions in treatment effect at dose \(j\). The equality-of-curves null is \(H_0:\delta_1=\cdots=\delta_k=0\). Because estimates at different doses may be correlated, the joint test must use their covariance matrix rather than treating each contrast as independent unless independence is justified.

Checking Apparent Inconsistency

An interaction is evidence about the modeled outcome relationship, not a diagnosis of why regions differ. Before interpreting it, examine whether the comparison is supported by the data. Plot observed outcomes and fitted curves by region, with uncertainty bands, and inspect residual patterns against dose and fitted values. Look for curvature left unexplained by a linear model, influential observations, sparse dose ranges, and regions with little or no overlap at relevant doses. If one region contributes few participants, its curve may be imprecise; instability can produce a wide interval or a large estimated interaction.

Dose is not always the most informative scale. If the same dose produces different exposures across regions, a dose-scale interaction can arise even when the exposure-response relationship is shared. Exposure matching and transport of exposure-response relationships are useful earlier considerations when deciding which scale is relevant. Exposure is often measured after treatment begins, so an exposure-adjusted analysis also needs a clear estimand and appropriate assumptions; it should not automatically replace the randomized dose comparison.

Baseline risk also affects how differences appear. For a binary endpoint, a shared relative effect can correspond to different absolute risk differences in regions with different control risks. Therefore, state whether the comparison is on an absolute-risk, relative-risk, odds, or another scale. A difference on one scale does not necessarily imply a difference on another, and changes in baseline risk alone should not be described as a changed dose-response effect without specifying the model scale.

Alternatives, Limitations, and Common Pitfalls

A regression with interaction terms is a direct way to estimate and test regional differences while adjusting for specified covariates. A model with separate curves by region is more flexible, but may be imprecise when regional data are sparse. Comparing standardized predictions can answer a population-level question by predicting each region’s outcomes under a common covariate distribution; the result depends on the selected target population and adequate covariate overlap. Likelihood-ratio tests can compare nested models, while Wald tests use estimates and their covariance matrix. These approaches can differ in finite samples, particularly when estimates are imprecise or the model is complex.

No test compensates for inadequate dose overlap, an incorrect curve form, incompatible endpoint definitions, or unmeasured differences that matter to the comparison. A p-value also does not measure the probability that curves are equal or the size of a clinically acceptable difference. Interpret estimates, uncertainty, data support, outcome scale, and the intended decision together.

• Mistake: “The interaction p-value is above 0.05, so the regions have the same dose response.” Correction: Failure to reject is not proof of equality. Inspect the interval and ask whether it excludes differences that would matter.
• Mistake: “The dose interaction proves the drug works differently by region.” Correction: It estimates a difference on the model’s chosen scale. Exposure, baseline risk, model form, and sampling uncertainty may explain or qualify the pattern.
• Mistake: “The regional curves are consistent because their point estimates are close.” Correction: Point estimates without uncertainty can conceal substantial imprecision, especially in a small regional sample.