Why Triangulate Regional Evidence?
A regional extrapolation argument is rarely supported by one result alone. Pharmacokinetic (PK) evidence can indicate whether a dose produces comparable exposure; exposure-response evidence can connect exposure to benefit and harm; biomarker evidence can provide a related biological signal; clinical outcomes address patient benefit directly; and clinical-practice evidence can reveal differences in care or endpoint measurement. These streams answer different questions. Their agreement can strengthen an argument, but only when the evidence is relevant to the same decision and its dependencies are understood.
Triangulation is not a rule that several favorable results automatically add up to proof. Nor is it necessarily a calculation that averages all available estimates. It is a structured assessment of whether multiple lines of evidence support a coherent interpretation, whether they rely on genuinely different information, and whether any disagreement has a plausible explanation. The earlier material on transporting PK models, exposure-response relationships, biomarkers, and clinical outcomes supplies the component analyses; the task here is to assess them together.
A useful organizing question is: What would have to be true for this evidence stream to support the target-region conclusion? For example, similar exposure supports a dose bridge only if exposure is relevant to response and the exposure-response relationship can be transported. Similar biomarker effects help only if the biomarker is informative about the clinical effect in this setting. Direct clinical outcomes remain important, but their interpretation may depend on trial conduct, endpoint definitions, precision, and baseline risk.
Evidence Streams and Their Dependencies
Before judging convergence, state the target claim precisely. It might be that the same dose gives an acceptable exposure in the target region, that the treatment’s relative effect is sufficiently similar, or that the expected benefit-risk profile is acceptable under local practice. These are related but not interchangeable claims. A PK comparison alone does not establish clinical efficacy, and a similar biomarker response does not by itself establish a similar clinical outcome.
| Evidence stream | Question it can inform | Important limitation |
|---|---|---|
| PK | Does the dose produce comparable exposure, accounting for relevant covariates? | Comparable exposure is not enough if exposure-response differs or the exposure range is inadequate. |
| Exposure-response | Does the modeled relationship predict a similar response at target-region exposure? | Prediction depends on model form, measured covariates, and transportability of the relationship. |
| Biomarker | Is a biological response consistent with the proposed mechanism and clinical effect? | A biomarker may not capture the full clinical effect or may have different clinical meaning across regions. |
| Clinical outcome | Are patient outcomes consistent with the proposed effect? | Precision, endpoint comparability, baseline risk, and trial conduct affect interpretation. |
| Clinical practice | Could local care, adherence, rescue treatment, or measurement alter observed outcomes or implementation? | Practice differences may affect both the true benefit-risk context and the observed endpoint. |
Evidence streams are often dependent. A PK estimate and an exposure-response prediction may use the same participants and concentration measurements. An exposure-response model and a biomarker analysis may share subjects, covariates, or a common mechanistic assumption. A clinical outcome and biomarker result may come from the same trial, so random variation in treatment groups can affect both. A practice difference may influence adherence, exposure, biomarker response, and the clinical endpoint. Thus, five reported results need not represent five independent confirmations.
A dependence map helps make this concrete. Draw each evidence stream as a node, then connect two nodes when they share participants, samples, endpoints, model assumptions, or causal pathways. Connections do not make evidence useless; they change how much independent confirmation it provides. If a numerical combination is justified, covariance must be considered. If there is no defensible common estimand or covariance model, keep the evidence streams distinct and synthesize them qualitatively rather than inventing a precision-enhancing calculation.
Worked Evidence Assessments
Worked Example: Interpreting a Regional Exposure Comparison
Suppose a hypothetical study compares dose-normalized AUC, a measure of drug exposure, between a source region and a target region. The estimated target-to-source geometric mean ratio is 1.08, and the estimated standard error of its logarithm is 0.0790. For illustration, assume the protocol specified an exposure comparison range of 0.80 to 1.25. The logarithmic estimate is \(\log(1.08)=0.0770\). A 90% interval on the log scale is the estimate plus or minus \(1.645\) standard errors:
Exponentiating the endpoints gives an interval for the ratio of approximately \((0.948,\ 1.230)\). This interval lies within the stated comparison range. Under the study’s sampling and modeling assumptions, the exposure comparison is therefore consistent with that operational criterion. It does not prove that exposure is identical, establish clinical equivalence, or show that response at a given exposure is transportable. Those conclusions require additional evidence and assumptions.
Worked Example: Accounting for Dependence Between Two Estimates
Imagine two hypothetical analyses estimate the same regional contrast on a log relative-risk scale. One exposure-response analysis gives \(0.06\), and one clinical-outcome analysis gives \(0.10\). Each has standard error \(0.10\). They use overlapping trial participants, and their estimated error correlation is \(\rho=0.50\). If, for this illustration, the two estimates are averaged with equal weights, the average is \(0.08\). The variance of an average of two estimates is:
Substituting the values gives:
A normal-approximation 95% interval for the average is \(0.08 \pm 1.96(0.08660)\), or approximately \((-0.0897,\ 0.2497)\). Exponentiating gives a relative-risk ratio of approximately \(1.083\), with interval \((0.914,\ 1.284)\). The estimate is compatible with a range of regional contrasts; it does not establish equality. If the estimates were incorrectly treated as independent, the standard error would be \(\sqrt{(0.01+0.01)/4}=0.07071\), producing an artificially narrower interval. This calculation is illustrative, not a recommendation to average evidence streams: a common estimand, defensible weights, and a credible covariance estimate are required.
Worked Example: Reconciling Several Evidence Streams
Consider a hypothetical program evaluating a treatment in a target region. The PK comparison estimates a target-to-source exposure ratio of 1.04, with a 90% interval from 0.95 to 1.14. A transported exposure-response model predicts a target-region relative risk of 0.78, with a 95% prediction interval from 0.65 to 0.94. The estimated regional difference in a prespecified biomarker effect is \(-0.02\) standard deviations, with a 95% interval from \(-0.12\) to \(0.08\). A local clinical-outcome analysis estimates a relative risk of 0.91, with a 95% interval from 0.72 to 1.15. Local rescue treatment is also more frequent: 28% of participants in the target-region trial versus 9% in the source-region trial.
The PK result supports similar exposure in the studied settings, subject to the adequacy of the covariate model and exposure measurements. The exposure-response prediction and biomarker result point in a direction consistent with benefit, but both depend on assumptions about how those relationships carry over. The local outcome estimate is less favorable as a point estimate, yet its interval includes 0.78; it is imprecise rather than clear evidence of a different effect. Higher rescue-treatment use could alter the endpoint, dilute an apparent treatment contrast, or reflect differences in patient management. It is a reason to investigate endpoint handling and practice, not a numerical correction by itself.
Baseline risk also affects absolute benefit without necessarily changing relative treatment effect. If the target-region control risk is 0.12 and the predicted relative risk is 0.78, the predicted treated risk is \(0.12(0.78)=0.0936\), giving an absolute risk reduction of \(0.12-0.0936=0.0264\), or 2.64 percentage points. At a control risk of 0.20, the same relative risk predicts \(0.20(0.78)=0.156\), an absolute reduction of \(0.044\), or 4.40 percentage points. A regional difference in absolute benefit should therefore not automatically be interpreted as a difference in relative treatment effect.
The defensible conclusion is conditional: the streams are broadly compatible with benefit, while local outcome precision and the rescue-treatment difference require attention. The evidence does not independently confirm the prediction five times, because PK, exposure-response, biomarker, and clinical results may share participants, model assumptions, or pathways. A stronger conclusion would require checking those dependencies and whether the local endpoint remains interpretable under different rescue practices.
Justifying a Dependence Adjustment
The covariance term in the second example follows directly from the variance rule for a sum. For two estimates with variances \(SE_1^2\) and \(SE_2^2\), and covariance \(\operatorname{Cov}(\hat{\theta}_1,\hat{\theta}_2)\), expanding the squared deviation of their sum gives the two variances plus twice the covariance. Dividing by four gives the variance of their average. If their correlation is \(\rho\), then covariance equals \(\rho SE_1SE_2\), yielding the displayed formula. Positive correlation increases uncertainty relative to the independent case; negative correlation decreases it. Treating correlated evidence as independent typically overstates the amount of information.
This derivation applies only when the two quantities are estimates of a common parameter on a compatible scale and the covariance is meaningful. Evidence streams that address different estimands—such as exposure similarity and clinical benefit—cannot be combined with this formula just because both are numerical. Their relationship is usually assessed through the scientific structure linking them, not by averaging their numbers.
Diagnostics, Alternatives, and Limitations
A practical triangulation assessment should check for failures that can make apparent convergence misleading or apparent conflict uninformative. Useful diagnostics include:
- Estimand alignment: confirm that regions, populations, doses, outcomes, follow-up periods, and treatment contrasts answer the intended target question.
- Support and measurement: check whether the target region has relevant covariate and exposure ranges, and whether assays, endpoints, and data collection are comparable.
- Model dependence: identify shared participants, samples, covariates, fitted relationships, and mechanistic assumptions; describe how each affects the evidence.
- Precision and discordance: distinguish a noisy estimate from a precise contradiction. Compare effect estimates and intervals, not just whether a result crosses a significance threshold.
- Practice and implementation: assess whether adherence, rescue therapy, concomitant treatment, or endpoint assessment differs enough to change exposure or observed outcomes.
If exposure evidence agrees but clinical outcomes do not, examine response-model transportability, outcome definitions, practice, and the precision of the outcome estimate. If biomarkers agree while clinical outcomes diverge, ask whether the biomarker is validated for the clinical question and whether clinical pathways not represented by the biomarker differ. If all streams agree but depend on the same trial and assumptions, report that convergence as internally coherent but not as independent replication.
The main alternative is to base the conclusion primarily on one evidence stream, often direct clinical outcomes. That can be appropriate when the outcome evidence is sufficiently informative and relevant, but it may leave mechanisms of regional difference unexplained. A mechanistic chain from PK through exposure-response and biomarkers can explain why extrapolation seems plausible, but cannot replace clinical evidence when those intermediate measures do not fully determine patient benefit. Formal pooling or borrowing is another possible approach when there is a justified common estimand and a specified statistical model; it is not a substitute for mapping dependencies and is not warranted merely because several estimates are available.
Triangulation has important limits. It cannot eliminate unmeasured effect modification, repair incompatible endpoints, or establish that a biomarker is a surrogate. A qualitative evidence map may be more honest than a single summary number when the evidence streams address different questions. Conversely, a conclusion that all evidence is “consistent” can conceal weak precision or common bias. State what supports the conclusion, what assumptions it depends on, and what observation would change the interpretation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why does a similar regional exposure estimate not, by itself, establish similar clinical efficacy?
- Two estimates each have standard error 0.10 and correlation 0.50. Is the standard error of their equal-weight average greater or less than it would be under independence? Give the value.
- A target region has a higher control-group event risk than the source region, while the relative risk is the same. Must the treatment effect differ by region? Explain.
- PK, biomarker, and clinical outcome analyses all use participants from one trial. What does their agreement establish, and what does it not establish?
- A local clinical estimate is less favorable than a transported prediction, but its interval is wide and includes that prediction. What is an appropriate interpretation?
Solutions. (1) Similar exposure supports only the PK part of the argument; efficacy also depends on transportability of the exposure-response relationship, outcome relevance, and other regional factors. (2) It is greater: the variance is \([0.01+0.01+2(0.50)(0.10)(0.10)]/4=0.0075\), so \(SE=\sqrt{0.0075}=0.08660\), compared with 0.07071 under independence. (3) No. With the same relative risk, a higher baseline risk gives a larger absolute risk reduction, but that alone does not imply a different relative treatment effect. (4) Agreement supports a coherent interpretation across related measurements, but shared participants and assumptions mean it is not three independent replications. (5) The result is imprecise and compatible with the prediction; it is not clear evidence either that the prediction is correct or that the regional effect differs. Investigate precision, endpoint comparability, and practice differences.