Design From the Prediction That Needs to Be Checked
A bridging study is useful when its design makes a meaningful comparison possible. Recruiting more participants from a new region may make a local treatment estimate more precise, but it does not necessarily test whether a model-based prediction transports to that region. A targeted design starts by stating the prediction at issue and then selects participants, doses, measurements, and covariates that allow the prediction to be checked.
The relevant gap might concern a different distribution of a known effect modifier, a possible regional difference in exposure, or uncertainty about whether a source-population treatment effect applies under local trial conduct. Earlier material on exposure matching, transported exposure-response relationships, endpoint harmonization, and trial conduct differences provides tools for describing these concerns. The design task here is to turn a specified concern into an informative study, without assuming that collecting local data alone resolves it.
Before choosing the design, specify the prediction in operational terms: the target population, treatment contrast or dose regimen, endpoint, analysis time, and prediction from the source model. Also state what size of disagreement would matter clinically or for the development decision. If the target estimand is a treatment contrast, it can be written as the mean outcome under treatment minus the mean outcome under control in the target population. The direction of the contrast must be stated consistently; for example, a negative difference may indicate fewer events under treatment.
Assumptions to Make Explicit
Design recommendations depend on assumptions. State them before using the study to assess transport:
- A defined target: The target population and clinical use are specified, and the study sample is relevant to that target or can be related to it using recorded characteristics.
- Comparable quantities: The endpoint, follow-up time, treatment contrast, and analysis scale are sufficiently aligned between the source prediction and the bridging data. Otherwise, a difference may reflect incompatible definitions rather than transport failure.
- Relevant covariates are measured: Covariates considered important to the stated gap are collected with compatible definitions and adequate coverage. Measuring a covariate does not guarantee that every modifier is known.
- Relevant treatment support: The study includes doses or exposures that bear on the proposed target regimen. Extrapolating far beyond observed or supported exposure ranges cannot be repaired merely by increasing the sample at unrelated doses.
- Interpretable comparison: Randomization, when feasible, supports comparison of treatment groups within the bridging study. If differences in conduct or care are part of the concern, the design must measure or control those features well enough to interpret the comparison.
These assumptions do not prove that a model transports. They clarify what the study can test and what would remain uncertain. For example, a study that enrolls only participants with a covariate value rarely found in the target population may estimate an effect in that subgroup well, yet provide a weak estimate of the target-population effect unless the subgroup’s role in the target is accounted for.
Choose the Population to Match the Gap
Use a target-representative sample when the primary question is whether the overall predicted effect applies to the target population under its expected covariate distribution. Consider deliberate enrichment when the concern is concentrated in a subgroup, such as patients with high clearance or a particular baseline-risk profile. Enrichment can improve the opportunity to assess that subgroup, but the study should retain enough information about other target-relevant groups to interpret an overall target-population prediction.
When an enriched sample differs from the target distribution, report stratum-specific results and define how they relate to the target. A weighted target estimate may be appropriate if the target stratum proportions are known and the needed stratum-specific effects are estimated. If a subgroup is absent, or has too little representation for a usable estimate, no reweighting can create evidence about its treatment effect.
Worked Example: Enriching for a Suspected Effect Modifier
A source model predicts a treatment effect, measured as the treatment-minus-control difference in event risk, of −8 percentage points in patients with typical clearance and −14 percentage points in patients with high clearance. In the target population, 70% have typical clearance and 30% have high clearance. The stated gap is whether the high-clearance prediction transports.
A study sampling only the target proportions would include relatively fewer high-clearance patients. To examine the subgroup directly, investigators instead enroll equal numbers in the two clearance strata and randomize treatment versus control within each stratum. Suppose the resulting local estimates are −7 percentage points for typical clearance and −15 percentage points for high clearance. Equal enrollment makes the unweighted average \((−7−15)/2=−11\) percentage points, but that average does not represent the target population.
Standardizing the observed stratum estimates to the target proportions gives \(0.70(−7)+0.30(−15)=−4.9−4.5=−9.4\) percentage points. The similarly standardized prediction is \(0.70(−8)+0.30(−14)=−5.6−4.2=−9.8\) percentage points. The target-standardized difference is \(−9.4−(−9.8)=0.4\) percentage points. This point estimate is close to the prediction, while the separate high-clearance estimate directly addresses the stated subgroup concern. Precision and uncertainty still matter: the point estimates alone do not establish that the prediction is sufficiently accurate.
This design relies on correctly measured clearance strata, adequate representation and randomized comparisons within each stratum, and applicable target proportions. If the stratum estimates are noisy, or clearance is only a proxy for the true modifier, the study may not settle the transport question.
Select a Dose Range That Covers the Target Use
Dose selection should be tied to the target regimen and the exposure range expected in the target population. A study at a dose that is locally familiar but produces exposures well below or above those relevant to intended use may fail to assess the proposed dose prediction. Where exposure is the basis for bridging, use the available pharmacokinetic and exposure-response model to identify a range that includes the relevant target exposures and, when ethically and clinically appropriate, enough spread to assess whether the predicted response pattern holds across that range.
A narrow dose range can be reasonable if the question is limited to one regimen and exposure variability is adequately represented. A broader range may be needed if the prediction concerns dose-response behavior or if covariates are expected to change exposure. It should not extend into doses that are unsafe or clinically irrelevant merely to create a wider range. The design must also avoid treating an exposure prediction as a substitute for observing the clinical endpoint that supports the intended claim.
Worked Example: Choosing Dose Levels From Target Exposure
Suppose a source model predicts that a dose of 100 mg produces a median exposure of 20 exposure units in the source population. For the target population, the available model predicts 25% higher clearance at the same dose. Assume, for this planning calculation, that exposure is inversely proportional to clearance and proportional to dose over the range considered. The predicted target exposure at 100 mg is therefore \(20/1.25=16\) units.
Under the same proportionality assumption, the target exposure at dose \(D\) is \(16(D/100)\) units. A clinically relevant target exposure interval of 18 to 24 units corresponds to doses from \(100(18/16)=112.5\) mg to \(100(24/16)=150\) mg. Candidate randomized dose levels of 125 mg and 150 mg predict target exposures of \(16(1.25)=20\) and \(16(1.50)=24\) units, respectively. Including a 100 mg arm would add a lower exposure of 16 units, if that dose is appropriate and safe.
The design should not assume that these arithmetic predictions are exact. The dose choice depends on the stated linear exposure assumptions, the reliability of the clearance prediction, and the safety and clinical relevance of the candidate doses. If exposure is nonlinear, the calculation must use the supported nonlinear model instead.
Harmonize the Endpoint and Measure the Covariates
The endpoint should match the prediction’s clinical meaning, measurement scale, and time point. If the source estimate concerns change from baseline at week 12, a different endpoint definition or follow-up time may not provide a direct check. Use common operational definitions, assessment schedules, and adjudication procedures where possible. If differences are unavoidable, specify how they will be handled before data are examined; otherwise, observed disagreement can be difficult to distinguish from endpoint or conduct differences.
Collect covariates because they answer a design question, not simply because they are available. Prioritize variables that define the target population, are implicated in the transportability gap, or are needed to interpret treatment and exposure differences. Define them consistently across populations, including units, timing, and any categories used in analysis. Pre-specifying their role helps prevent a post hoc search through many variables from being mistaken for a planned assessment of the stated gap.
Worked Example: Aligning Endpoint and Covariate Collection
A source model predicts a treatment-minus-control difference of −12 percentage points in the proportion of patients meeting a response definition at week 12. The target trial could instead use a week-8 assessment, but the response can change between weeks 8 and 12. A week-8 result would therefore not directly test the week-12 prediction. The design should retain the week-12 endpoint or specify a justified model-based link between the two time points before interpreting the comparison.
The stated concern is higher baseline severity in the target population. Collect baseline severity using the same scale and timing as in the source data, and ensure the enrollment plan includes the severity range found in the target. If treatment is randomized, assess the treatment contrast within relevant severity strata or use a pre-specified analysis that relates the estimate to the target severity distribution. Recording severity without adequate coverage or without a defined role in the analysis does not address the concern.
The primary endpoint remains the week-12 response proportion, so the study evaluates the stated prediction rather than silently substituting a different outcome. The severity measurement supports interpretation of whether the target population is represented; it does not alone demonstrate that severity is the only relevant effect modifier.
Precision Is Not the Same as Validation
A useful design distinguishes two aims. A precision-oriented study primarily reduces uncertainty in a local estimate. A prediction-checking study is arranged so the local evidence can be compared with a pre-specified prediction on the same population, treatment contrast, endpoint, and scale. A study may do both, but a larger sample cannot compensate for a mismatched endpoint, absent subgroup, or irrelevant dose range.
For a quantitative illustration, let \(\widehat{\Delta}_L\) be the estimated local treatment contrast and \(\widehat{\Delta}_P\) the source-model prediction for the target population. Define the discrepancy \(D=\widehat{\Delta}_L-\widehat{\Delta}_P\). If the two estimates are independent, with standard errors \(SE_L\) and \(SE_P\), then \(SE_D=\sqrt{SE_L^2+SE_P^2}\). If they are not independent, their covariance must also be accounted for. A pre-specified clinically acceptable discrepancy margin is needed to interpret how close is close enough; merely failing to find a statistically significant difference does not demonstrate adequate agreement.
The following derivation shows why a narrow interval is only a precision statement. For a balanced two-arm comparison with \(n\) participants per arm, equal outcome standard deviation \(\sigma\), and independent observations, the variance of the difference in sample means is \(\sigma^2/n+\sigma^2/n\). Therefore its standard error is \(\sigma\sqrt{2/n}\). A conventional approximate 95% confidence interval has half-width \(1.96\sigma\sqrt{2/n}\). Reducing this half-width improves precision, but says nothing by itself about whether the estimated local contrast agrees with a transport prediction.
Worked Example: Checking a Prediction Against a Local Estimate
Suppose the source model predicts a week-12 treatment-minus-control response difference of −12 percentage points, with standard error 1.0 percentage point. A harmonized randomized bridging study estimates −10.5 percentage points, with standard error 2.0 percentage points. Assume the prediction and local estimate are independent for this illustration.
The discrepancy is \(D=−10.5−(−12)=1.5\) percentage points. Its standard error is \(\sqrt{2.0^2+1.0^2}=\sqrt{5}=2.236\) percentage points. An approximate 95% interval for the discrepancy is \(1.5\pm1.96(2.236)=1.5\pm4.383\), or approximately −2.88 to 5.88 percentage points.
If a margin of ±6 percentage points was specified in advance as the largest acceptable discrepancy, this interval lies within the margin. Under the assumptions used to construct the interval, the result supports agreement within that specified margin. It does not prove the model is correct for every outcome or subgroup. If the interval instead extended beyond a margin boundary, the study would not establish agreement within the margin, even if a conventional test found no statistically significant difference from zero.
Compare Design Options and Diagnose Weaknesses
No single design is best for every gap. The appropriate choice depends on whether the decision concerns an average target effect, a particular subgroup, a dose prediction, or a measurement or conduct difference.
| Design choice | Most useful when | Main limitation |
|---|---|---|
| Target-representative enrollment | The primary aim is an overall target-population estimate. | A rare but important subgroup may be too small for a useful assessment. |
| Enrichment of a specified subgroup | The gap concerns a group expected to differ, such as high-clearance patients. | The raw overall estimate may not represent the target distribution; subgroup estimates may remain imprecise. |
| Several relevant dose levels | The prediction concerns exposure matching or response over a dose range. | Unnecessary or unsupported doses add complexity and may introduce safety concerns. |
| Focused local replication at one regimen | The decision concerns one clinically intended regimen. | It may not assess transport across other doses or exposures. |
Warning signs include poor representation of the target covariate range, exposure levels that do not reach the intended use, different endpoint definitions or assessment times, and treatment groups that cannot be compared fairly. A further warning sign is a design whose stated success criterion is only a small standard error or a non-significant difference from the model prediction. Before data collection, check that the intended estimate is identifiable from the planned arms and measurements, and that the result could change or support the relevant development decision.
A targeted bridging study still has limitations. It cannot rule out unmeasured effect modifiers, guarantee transport to populations outside its observed range, or resolve model-form uncertainty solely by adding participants. Its conclusions should be restricted to the target, treatment conditions, and endpoints actually supported by the design. If the study is enriched, report how the enrichment affects the overall estimate; if the dose range is narrow, do not generalize the result across unsupported doses.
Name the target population, treatment contrast or regimen, endpoint, and prediction that is uncertain.
Represent the target or enrich the implicated subgroup, and include clinically relevant treatment or exposure conditions.
Harmonize the endpoint and collect the covariates needed to interpret the specified transport concern.
Define the local-to-prediction comparison and a clinically meaningful discrepancy criterion; do not substitute lack of statistical significance for evidence of agreement.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study enrolls only typical-clearance patients to check a prediction specifically about high-clearance patients. What is the central design problem?
- A source prediction concerns a week-12 endpoint, but a local study measures only week 8. Why might a precise local estimate fail to validate the prediction?
- A balanced two-arm study plans 25 participants per arm, and the outcome standard deviation is 10 units. Calculate the approximate standard error and 95% confidence-interval half-width for the difference in means.
- A local estimate does not differ significantly from the model prediction, but its confidence interval includes discrepancies larger than the pre-specified acceptable margin. What conclusion is warranted?
Solutions. (1) The study lacks direct evidence in the subgroup named by the gap. A typical-clearance estimate cannot establish whether the high-clearance prediction applies; the design should include enough high-clearance participants to assess it. (2) The local result is for a different time point and may not estimate the same quantity. Without a pre-specified and justified link between weeks 8 and 12, precision at week 8 does not establish agreement with the week-12 prediction. (3) \(SE=10\sqrt{2/25}=10\sqrt{0.08}=2.828\) units. The approximate 95% half-width is \(1.96(2.828)=5.543\) units. This describes precision under the assumptions of independent observations and equal standard deviations; it does not by itself validate a prediction. (4) Agreement within the margin has not been established. A non-significant difference is not evidence that all clinically unacceptable discrepancies have been excluded.