Why a Bridge Begins With Definitions
A regional or population bridge asks whether evidence from one setting can inform a decision in another. Before choosing a model or summarizing evidence, the analyst must state exactly which evidence is being carried across, to whom, under what treatment conditions, and for which decision. If those pieces are left implicit, an analysis may be technically polished yet address the wrong question.
For example, evidence about a short-term biomarker response in adults does not automatically answer whether a dose should be recommended for adolescents on a different regimen. The populations, treatment conditions, outcome, or intended decision may differ. Defining the bridge does not establish that extrapolation is valid; it makes the claim testable and reveals what evidence is missing.
“Source” and “target” are roles in a particular question, not permanent labels for countries, trials, or datasets. A trial conducted in one region can be the source for a question about another region, while evidence from that second region could be the source for a different question. The target is defined by the decision to be made, not simply by where future patients live.
Specify the Bridge in Five Parts
A useful specification names five parts. Each should be concrete enough that another analyst could tell whether a proposed dataset or model addresses the same question.
- Source population and setting: identify the people represented by the evidence and the relevant context. Include the population-defining features that matter to the question, such as age range, disease stage, prior treatment, region, care setting, and calendar period. State whether the source is a trial sample, a registry, or another defined evidence base.
- Target population and setting: identify the people for whom the decision is intended, using equally specific descriptors. “Patients in Region B” may be too broad if the decision concerns only treatment-naive adults with a defined disease stage.
- Treatment conditions: specify the intervention and the comparison, if relevant. Describe dose, schedule, route, duration, permitted concomitant care, and any treatment changes that matter to interpreting the evidence. A drug name alone may not define the treatment being compared.
- Outcome and assessment: state what is measured, how it is defined, when it is assessed, and how intercurrent events such as treatment discontinuation are handled in the reported result. A laboratory value, symptom score, response, and survival outcome are not interchangeable labels for “benefit.”
- Decision context: name the action the evidence is meant to inform, who makes it, and the criterion that matters. Examples include selecting a dose, deciding whether a regional authorization is supported, or determining whether further evidence is needed. A decision threshold or required performance should be stated when it is part of the question.
The goal is not to make source and target identical. Bridging exists because some feature differs. The goal is to describe the difference precisely and distinguish it from differences that change the meaning of the evidence. In particular, keep separate: who was studied, what treatment was given, what outcome was measured, and what decision is being considered.
A Short Justification: Why the Outcome Must Match the Decision
Suppose a source result is a response proportion measured at 12 weeks, while the target decision concerns one-year survival. A numerical comparison is not made valid simply by calling both quantities “efficacy.” The quantities have different definitions and time horizons. To use the source result for the target decision, a defensible relationship between them would be needed; otherwise the source does not directly answer the target question.
This can be expressed by denoting the source outcome summary by \(Q_S\) and the target decision-relevant summary by \(Q_T\). A direct comparison requires that they represent the same defined quantity, or that a justified mapping \(g\) connects them:
The equation is a statement of what would be needed, not a claim that such a mapping exists. If \(Q_S\) is a 12-week response proportion and \(Q_T\) is one-year survival, their definitions alone provide no \(g\). Subtracting the two, or treating one as a substitute for the other, has no defined interpretation. This is why outcome definition and assessment time belong in the bridge specification, before modeling begins.
Worked Example: A Regional Efficacy Bridge With an Outcome Mismatch
Hypothetical scenario. A sponsor proposes using a trial in Region A to support use of Drug Z in adults with a particular cancer in Region B. In the source trial, 200 adults received 100 mg once daily with the same background care for 12 weeks. Eighty had a documented tumor response by the trial’s response criteria at week 12. The proposed regional decision, however, concerns whether to recommend the treatment based on one-year overall survival.
Source. Adults with the specified cancer enrolled in the Region A trial; the treatment is 100 mg once daily plus the protocol’s background care. The measured outcome is response at week 12.
Target and decision. Adults with the specified cancer in Region B; the proposed decision is a recommendation informed by one-year overall survival.
What the source number says. The observed source response proportion is
Thus, 40% of the 200 hypothetical trial participants met the stated response definition at week 12. This is a descriptive source result; it is not a one-year survival estimate and does not establish the treatment’s effect in Region B.
Bridge assessment. The treatment description is reasonably explicit, but the decision-relevant outcome does not match the source outcome. No information in the scenario specifies a validated relationship between week-12 response and one-year survival. Therefore, the 40% response proportion cannot by itself answer the stated regional decision. The bridge question must be revised to match an outcome supported by the evidence, or additional evidence connecting the source result to the target survival decision must be identified.
Worked Examples: Finding Other Kinds of Mismatch
Worked Example: A Dose Decision With Different Treatment Conditions
Hypothetical scenario. A source study enrolled 120 adults in Region C and administered Drug Y at 50 mg twice daily for 8 weeks. Its decision-relevant measure was the mean change in a disease score from baseline to week 8. A regional team wants to decide whether to recommend 50 mg once daily for 12 weeks in adults in Region D, where usual concomitant care includes a second medicine not allowed in the source study.
Specify the source. The source is the 120 adults in Region C, treated at 50 mg twice daily for 8 weeks without the second medicine. The outcome is change in disease score at week 8.
Specify the target. The target is adults in Region D receiving 50 mg once daily for 12 weeks with the second medicine permitted. The decision is whether to recommend that regimen.
Locate the mismatches. The dose schedule differs (twice daily versus once daily), the treatment duration differs (8 versus 12 weeks), and concomitant care differs. These are not merely differences in how the question is worded: they define different treatment conditions and could alter the outcome. The source result therefore does not directly describe the target regimen.
Conclusion. The bridge is clearly stated but not directly supported by the source result as described. The analyst should not relabel the source regimen as the target regimen. The decision may require evidence addressing the target conditions or a justified analysis that explicitly connects the differing conditions to the outcome.
Worked Example: A Population Label That Is Too Broad
Hypothetical scenario. A model was developed using 300 adults with moderate disease in Region E. A team proposes to use its predictions to select a dose for “patients in Region F.” The intended decision actually concerns 60- to 75-year-old patients with severe disease who have not previously received the drug.
Clarify the source population. The stated source is 300 adults with moderate disease in Region E. The age range and prior-treatment distribution are not reported in the scenario, so they cannot be assumed.
Clarify the target population. The target is adults aged 60 to 75 with severe disease in Region F who are treatment-naive to the drug. The decision is dose selection for that group.
Identify what is missing. “Patients in Region F” does not specify disease severity, age, or prior treatment; the actual decision does. Also, the scenario gives no source information on age or prior treatment. Thus, the population descriptions cannot yet be compared on those dimensions. The stated source disease severity differs from the target severity, and there is no basis here to say how the model’s predictions apply to severe disease.
Conclusion. Before fitting or applying a model, obtain enough source and target description to determine what populations the evidence represents. The model’s existence does not resolve an undefined or poorly characterized target.
Diagnostics, Alternatives, and Limits
A practical diagnostic is to place the five parts of the source and target side by side, then ask whether each difference changes the interpretation of the evidence or the action being considered. This is a specification check, not a statistical test. It can reveal that the proposed bridge is answerable as stated, answerable only after a defensible link is added, or not answerable with the described evidence.
| Part | Questions to check | Warning sign |
|---|---|---|
| Population and setting | Who is included, where, and under what care conditions? | A broad label hides a target subgroup not represented in the source description. |
| Treatment | Are dose, schedule, duration, comparator, and background care explicit? | The target decision concerns a regimen or co-treatment different from the source. |
| Outcome | Is the same construct measured using the same definition and timing? | A surrogate, symptom, response, and survival result are treated as equivalent without a stated link. |
| Decision | What action is being taken, and what evidence criterion matters? | A descriptive result is presented as if it answers a dose, benefit-risk, or authorization question. |
The main alternatives are not interchangeable shortcuts. A direct bridge uses source evidence for a target question whose key definitions and treatment conditions are sufficiently aligned. A model-based bridge may be considered when differences need to be represented explicitly, but naming a model does not establish that its inputs or assumptions can support the target. Additional target evidence may be needed when a key difference cannot be resolved by the source data or a defensible mapping. In every case, first define the question; only then can the appropriate evidence strategy be judged.
Several failure modes are especially common:
- Using geography as a substitute for a population definition. A region name does not specify age, disease severity, prior treatment, or clinical setting.
- Describing the molecule but not the treatment condition. Dose schedule, duration, comparator, and background care can define materially different interventions.
- Using an outcome label without its measurement details. State the definition, assessment time, and relevant handling of treatment changes or missing observations.
- Starting with the available dataset. The target should be defined by the decision, not narrowed after seeing what the source happens to measure.
- Claiming support because a model produced a prediction. A numerical prediction is useful only if the source, target, treatment, outcome, and decision are adequately connected.
A well-defined bridge still has limits. Specification cannot establish that source evidence is representative of the target, show that a treatment relationship carries across populations, or guarantee that a decision is robust. Those are substantive evidence and assumption questions. This tutorial’s purpose is narrower and prior: make clear what is being transported and what the target decision requires, so that later analyses can address the right problem.
Check Your Understanding
For each question, answer first, then compare with the brief solution below.
- A source trial measures a symptom score at week 4; the target decision concerns hospitalization by month 6. What mismatch must be addressed?
- A report says the source is “Asian patients” and the target is “European patients.” Name two details needed to make those population descriptions more useful.
- A source uses 20 mg once daily for 6 weeks. The proposed target decision concerns 20 mg twice daily for 10 weeks. What part of the bridge differs?
- Why is “the model predicts a 30% response” not, by itself, a complete bridge conclusion?
Solutions. (1) The outcome construct and assessment time differ: a week-4 symptom score is not month-6 hospitalization. A defined, defensible relationship or target-relevant evidence is needed. (2) Examples include age range, disease severity, prior treatment, care setting, and calendar period; the relevant details depend on the decision. (3) The treatment conditions differ in both schedule and duration, so the source regimen is not the target regimen. (4) The statement omits which population and treatment the prediction concerns, how response is defined and timed, and what decision or criterion the 30% is meant to inform.