Why Endpoint Differences Need an Explicit Bridge
A Phase 2 dose may have been selected using a continuous symptom score, while Phase 3 evaluates a responder endpoint, a clinician rating, or the same measure at a later time. The dose rationale cannot simply carry over the Phase 2 effect as if these outcomes were identical. It needs to explain how evidence on the earlier endpoint supports a prediction about the later one, and where that prediction depends on assumptions rather than direct observation.
The distinction matters because an endpoint is not just a label. Its scale, direction, assessment time, threshold, analysis population, and handling of events such as discontinuation can all change the quantity being estimated. A mean difference on a continuous scale does not, by itself, determine the proportion meeting a responder threshold. Nor does an effect at one visit establish the effect at another.
Begin by writing down the Phase 2 and Phase 3 estimands in context: the treatment and comparator, population, endpoint definition, assessment time, and summary measure. Check whether the difference is only in measurement scale, or also in threshold, timing, population, missing-data handling, or the meaning of a response. A bridge across one difference does not automatically address the others.
From a Continuous Endpoint to a Responder Endpoint
A common bridge starts with a continuous measure and predicts the probability that a patient crosses a clinically defined threshold. Suppose \(X\) is improvement from baseline, with larger values representing greater improvement, and a responder is defined by \(X\geq c\). If the improvement distribution in treatment arm \(a\) is normal with mean \(\mu_a\) and standard deviation \(\sigma_a\), then:
where \(\Phi\) is the standard normal cumulative distribution function. The assumptions are consequential: the modeled distribution should be a reasonable description of individual outcomes, the threshold must have the same clinical meaning in both settings, and the population and assessment time must align. If the treatment changes variability or produces a skewed or multimodal distribution, a normal model with a common standard deviation can misstate the responder rate.
Worked Example: Translating a Mean Difference Into a Responder Difference
In a hypothetical Phase 2 study, improvement at the Phase 3 assessment time is modeled as normally distributed in both arms. Placebo has mean improvement 18 points and the selected dose has mean improvement 24 points; both arms have standard deviation 12 points. Phase 3 defines response as improvement of at least 22 points. Calculate the model-implied response rates and risk difference.
Check the bridge assumptions. For this calculation, assume the score has the same definition and direction in both phases, the analysis populations and assessment times align, the within-arm distributions are adequately approximated by normal distributions, and the standard deviation is 12 points in both arms. These are hypothetical modeling assumptions, not established facts about a real study.
For placebo, the standardized threshold is \((22-18)/12=0.3333\). Thus:
For the selected dose, the standardized threshold is \((22-24)/12=-0.1667\). Therefore:
Using the unrounded probabilities, the model-implied risk difference is approximately \(0.5661838-0.3694413=0.1967425\), or 19.67 percentage points. This is a prediction under the stated distributional assumptions, not an observed Phase 3 treatment effect. The six-point difference in mean improvement does not itself equal a 19.67-point responder difference; that number also depends on the threshold and the outcome distributions.
A useful justification follows directly from the probability calculation. Standardizing \(X\) within arm \(a\) gives \(Z=(X-\mu_a)/\sigma_a\), so the event \(X\geq c\) is the event \(Z\geq(c-\mu_a)/\sigma_a\). Under the normality assumption, \(Z\) follows a standard normal distribution, which yields the formula above. This also shows why a mean shift alone is insufficient: the threshold’s location relative to the mean and the standard deviation both affect the tail probability.
Bridging Different Measurement Scales
When Phase 2 and Phase 3 use distinct continuous measures, paired measurements in the same participants can support a crosswalk. A simple linear model is \(Y=a+bX+\varepsilon\), where \(X\) is the Phase 2 measure and \(Y\) the Phase 3 measure. If the relationship is common across treatment groups and \(E(\varepsilon\mid X)=0\), then the model predicts a difference in mean \(Y\) of \(b\Delta_X\) when the mean difference in \(X\) is \(\Delta_X\). This is a statement about expected values, not proof that an individual-level change in one measure causes a change in the other.
To use this relationship for a responder endpoint on \(Y\), more is needed than the predicted mean. One needs an adequate model for the distribution of \(Y\) in each treatment arm. For example, if \(X\) is normally distributed within each arm and the residual \(\varepsilon\) is independent, normal, and has the same variance in both arms, then the marginal variance of \(Y\) is \(b^2\sigma_X^2+\sigma_\varepsilon^2\). The responder probability can then be calculated from the resulting arm-specific distribution. If these assumptions do not hold, a distributional model suited to the data may be needed instead.
Worked Example: Crosswalking Two Continuous Scales
In a hypothetical calibration sample of 120 patients, both a symptom-improvement scale \(X\) and a functioning scale \(Y\) are measured at the same visit. Higher values on both scales mean better status. The sample summaries are \(\bar{x}=10\) points, \(s_X=8\) points, \(\bar{y}=50\) points, \(s_Y=12\) points, and correlation \(r=0.75\). A Phase 2 comparison estimates that the selected dose improves \(X\) by 4 points more than placebo. Use a linear crosswalk to predict the mean difference on \(Y\), then illustrate what extra assumptions would be needed to predict response at \(Y\geq55\).
Check the bridge assumptions. Assume the correlation and scale summaries describe the within-treatment relationship, not merely a pooled association induced by group differences; the relationship is approximately linear; and its slope applies in the Phase 2 and Phase 3 populations. Also assume, for the responder illustration only, normal distributions and a common \(X\) standard deviation of 8 points in each arm.
The regression slope is \(b=r(s_Y/s_X)\), so:
The intercept is \(a=\bar{y}-b\bar{x}=50-(1.125)(10)=38.75\). A four-point treatment difference on \(X\) therefore predicts a mean difference on \(Y\) of:
That predicted mean difference alone does not give the proportion with \(Y\geq55\). For an illustrative distributional bridge, the residual standard deviation is \(s_Y\sqrt{1-r^2}=12\sqrt{1-0.75^2}=7.937\) points. Under the additional assumptions above, the marginal standard deviation of \(Y\) in each arm is \(\sqrt{(1.125\times8)^2+7.937^2}=\sqrt{144}=12\) points. If the placebo mean on \(Y\) is 50, the predicted dose mean is 54.5. The implied responder rates are:
The resulting model-implied difference is about \(0.4834-0.3385=0.1449\), or 14.49 percentage points, calculated from the unrounded probabilities. This illustrative prediction relies on substantially more than the linear mean crosswalk: it requires the specified residual and marginal distributions, common variability, and transportability of the relationship. A calibration sample should also be checked for residual patterns and predictive performance before such a prediction is used to support a dose rationale.
Keep Endpoint Bridging Separate From Time Bridging
An endpoint can differ in both definition and assessment time. These are separate links in the argument. A scale crosswalk addresses how measurements relate at a given time; it does not show that a Phase 2 effect persists to a later Phase 3 visit. Conversely, evidence that an effect persists on one measure does not by itself map that measure to a different endpoint.
Worked Example: Accounting for a Later Assessment Time
In a hypothetical development program, Phase 2 estimates a six-point dose-placebo difference in continuous improvement at week 8. Phase 3 uses a responder threshold at week 12. A relevant longitudinal model, supported by randomized observations through week 12 in a comparable population, estimates that the week-12 treatment difference is 75% of the week-8 difference. The same model predicts a week-12 placebo mean improvement of 18 points and a standard deviation of 12 points in both arms. Phase 3 response is improvement of at least 22 points. Calculate the implied week-12 responder difference.
Check the bridge assumptions. The 0.75 retention factor must be supported for the drug, dose range, population, and endpoint in question; the week-12 placebo distribution and common standard deviation must be appropriate; and the responder definition must be clinically meaningful. Here, those inputs are stipulated as hypothetical model predictions, not universal properties of treatment effects.
The predicted week-12 mean difference is \(0.75\times6=4.5\) points. With a placebo mean of 18, the predicted dose mean is \(18+4.5=22.5\) points. The placebo response probability remains \(1-\Phi((22-18)/12)\approx0.3694\). For the dose, the standardized threshold is \((22-22.5)/12=-0.04167\), giving a response probability of \(1-\Phi(-0.04167)=\Phi(0.04167)\approx0.5166\). The implied difference is approximately \(0.516607-0.369441=0.147166\), or 14.72 percentage points.
This calculation depends first on the longitudinal model’s estimate of waning from week 8 to week 12, and then on the distributional mapping from the week-12 mean to a responder probability. Each link should be reported and assessed separately. If there is no credible evidence for the retention factor, the week-12 prediction is not established by the week-8 result alone.
Evidence, Diagnostics, and Alternatives
A transparent dose rationale should make the bridge auditable. At minimum, describe the source data, target population, endpoint definitions, assessment times, model form, assumptions, and uncertainty. Where possible, use data in which both endpoint definitions were collected in the same randomized participants. A paired design helps estimate their relationship directly, but it does not guarantee transportability to a different population, dose, visit, or trial conduct.
- Check endpoint alignment. Confirm scoring direction, units, baseline definition, visit window, responder threshold, and handling of discontinuation and rescue treatment. A threshold change is not just a change of label.
- Check the shape of the relationship. Plot paired measurements and model residuals; look for curvature, unequal residual spread, outliers, floor or ceiling effects, and different relationships across treatment groups.
- Check predictions, not only fit. Assess calibration and prediction error in data not used to fit the bridge, if available. A high correlation can coexist with biased predictions at the clinically important threshold.
- Check the tails and thresholds. Responder rates depend on the distribution near and beyond the cutoff. Compare model predictions with observed response proportions and examine sensitivity to plausible thresholds and distributional assumptions.
- Carry uncertainty through the translation. Uncertainty in the Phase 2 effect, crosswalk parameters, variability, threshold, and time course all contribute to uncertainty in the predicted Phase 3 effect. A single point prediction can conceal a weak bridge.
The main alternatives have different evidentiary strengths. Analyzing a common endpoint collected in both phases avoids some scale-conversion assumptions, although differences in population, timing, or trial conduct can remain. A paired-data crosswalk provides a direct empirical link between two measures but depends on the model and its transportability. A distributional model can translate continuous effects into responder probabilities, but requires assumptions about outcome distributions and threshold meaning. Standardizing effects by a standard deviation may help compare magnitudes across scales; it does not establish a conversion to a clinical threshold or prove that endpoints measure the same construct.
| Approach | What it supports | Key limitation |
|---|---|---|
| Common endpoint in both phases | Direct comparison on the shared measure | Does not remove differences in population, visit, or estimand |
| Paired-data crosswalk | Estimated relationship between two measures | Can fail under nonlinearity, group-specific relationships, or poor transportability |
| Distributional responder model | Predicted probability of crossing a defined threshold | Depends on the modeled distribution, variability, and threshold |
| Standardized effect comparison | Scale-free summary of relative effect magnitude | Does not establish endpoint equivalence or a responder-rate difference |
A bridge is weak when its essential assumptions cannot be checked, when the target endpoint is not measured in the source data, or when the data do not cover the relevant doses, populations, or time points. It is also limited by measurement error: noisy or inconsistently administered endpoints can weaken the estimated relationship. In such cases, describe the target effect as uncertain rather than presenting a precise translation as direct evidence.
A Common Mistake
Wrong reasoning: “The Phase 2 mean improved by six points, and Phase 3 calls a patient a responder after a six-point improvement, so the Phase 2 result implies a high Phase 3 response rate.” This confuses a group mean difference with an individual threshold and ignores the control distribution, outcome variability, and assessment time.
Correction: Specify the arm-specific outcome distributions and threshold, or use a validated empirical model to estimate response probabilities. If the visit differs, separately justify the time-course link. State which estimates are observed and which are model-based, and report uncertainty and diagnostic evidence for each bridge.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why can a six-point mean difference on a continuous scale not, by itself, determine a responder-rate difference?
- A paired endpoint sample has a strong correlation. What additional checks are needed before using its crosswalk in Phase 3?
- Phase 2 measures an endpoint at week 8; Phase 3 uses a different endpoint at week 12. What are the two distinct links the rationale must support?
- Why is a standardized effect not a substitute for a bridge to a clinical responder threshold?
Solutions. (1) The response rate depends on the threshold’s position relative to the arm-specific outcome distributions, including their variability and shape, as well as the control rate. The mean difference alone does not specify these quantities. (2) Check linearity or other model form, residual variability, treatment-group interactions, calibration near the target threshold, and performance in independent or held-out data; also assess whether the population, dose, and visit are transportable. (3) The argument needs evidence that the effect carries from week 8 to week 12, and a separate relationship connecting the Phase 2 endpoint to the Phase 3 endpoint. (4) Standardization rescales an effect by variability; it does not identify an individual-level threshold probability or establish that the measures have the same clinical meaning.