Why Endpoint Harmonization Matters
A treatment effect can be transported across regions only if its outcome has a sufficiently consistent meaning. Two trials may both report “response,” “symptom improvement,” or “disease control,” while using different thresholds, instruments, or assessment times. The labels alone do not establish that the outcomes are comparable. A difference attributed to population transport may instead reflect how the endpoint was defined or measured.
Endpoint harmonization makes the outcome used for comparison explicit and, when justified, translates measurements to a common definition or scale. It is not a numerical adjustment to perform automatically. The choice must be supported by clinical meaning, measurement properties, and the data available. As established in the earlier material on aligning estimands across trials and regions, the target outcome and time point are part of the estimand, not merely details of data collection.
A useful distinction is between the clinical construct of interest (such as pain severity), the measurement (such as a questionnaire score), and the endpoint rule applied to that measurement (such as a decrease of at least 30% by week 12). Differences at any of these levels can change the treatment contrast. A shared construct is necessary for meaningful comparison, but it is not sufficient: instruments can cover different symptoms, thresholds can classify patients differently, and assessment times can capture different stages of response.
When Is a Mapping Defensible?
Let \(Y^*\) denote the clinical outcome on a common target scale and let \(Y_r\) be the observed endpoint in region \(r\). A measurement model describes how the recorded endpoint relates to that target, for example \(Y_r=g_r(Y^*,X)+\epsilon_r\), where \(X\) may include baseline characteristics and \(\epsilon_r\) represents measurement error. Harmonization requires enough evidence to identify or approximate the mapping \(g_r\). It also requires that the mapping be appropriate for the populations and treatment groups being compared.
A mapping may be supported by a validation study in which both instruments are administered to the same participants, by established scoring rules, or by a well-justified measurement model. The evidence should cover the relevant outcome range and clinical groups. If the relationship differs by region, baseline severity, or treatment, a single conversion may distort the treatment contrast. Where no credible mapping exists, the correct conclusion may be that the effects are not directly comparable on the available endpoint data.
For a continuous endpoint, a simple linear conversion illustrates the key requirement. Suppose a validated relationship is \(Y_A=a+bY_B\), with the same \(a\) and \(b\) applying to both treatment groups and relevant patients. Then the treatment contrast on scale A is \(b\) times the contrast on scale B: the additive offset cancels. If the relationship is nonlinear, patient-level distributions and the mapping itself generally matter; multiplying a mean contrast by a constant is no longer generally valid.
This result assumes the linear mapping holds across both treatment groups, is measured on the same target construct, and does not vary with treatment assignment or region in an unaccounted-for way. A high correlation between instruments is not enough: two instruments can rank patients similarly while differing systematically in scale, thresholds, or responsiveness.
Worked Example: Converting a Validated Continuous Scale
Suppose a hypothetical dual-instrument validation study supports the relation \(Y_A=12+10Y_B\) across the relevant symptom range, with no evidence that the relation differs by treatment group. In a regional trial, the mean improvement is 2.4 points on instrument B in the treatment group and 1.6 points in the control group. Determine the treatment contrast on instrument A.
The observed contrast on B is \(2.4-1.6=0.8\) points. Applying the validated conversion to each group gives \(Y_A\) means of \(12+10(2.4)=36\) and \(12+10(1.6)=28\), respectively. Thus the contrast is \(36-28=8\) points on A. Equivalently, the intercept cancels and \(10(0.8)=8\).
The harmonized effect is therefore an 8-point greater mean improvement on scale A in the treatment group. This conversion is justified here only because the mapping is assumed to be validated for the relevant range and common to both groups. Without that evidence, the numerical conversion would not establish clinical comparability.
Endpoint Thresholds Can Change the Effect
A continuous outcome may be converted into a binary responder endpoint by applying a cutoff. That step loses information and makes the treatment effect depend on the cutoff. Even when two regions assess the same underlying construct, different response definitions can produce different risks and risk differences. The comparison must therefore name the response rule, not just the word “responder.”
Worked Example: Different Response Cutoffs
Consider hypothetical treatment and control groups with normally distributed improvement \(X\), measured in standardized units. Assume \(X\mid T=1\sim N(1,1)\) and \(X\mid T=0\sim N(0,1)\). A regional definition counts a patient as a responder when improvement is at least 0.5 units. Calculate the risk difference, then compare it with a definition requiring at least 1 unit.
Under the 0.5-unit cutoff, treatment response probability is \(P(X\geq0.5\mid T=1)=1-\Phi((0.5-1)/1)=\Phi(0.5)\). Control response probability is \(P(X\geq0.5\mid T=0)=1-\Phi(0.5)=\Phi(-0.5)\). Thus the risk difference is calculated from the unrounded probabilities:
The risk difference is approximately 0.3829, or 38.29 percentage points. It is not obtained by subtracting prematurely rounded probabilities. Under the 1-unit cutoff, the response probabilities are \(\Phi(0)=0.500000\) for treatment and \(\Phi(-1)=0.158655\) for control. The risk difference is \(0.500000-0.158655=0.341345\), or 34.13 percentage points.
The underlying distributions are unchanged, yet the risk difference falls from about 38.29 to 34.13 percentage points when the cutoff changes. This calculation relies on the stated normal distributions and common variance of 1; actual endpoint analysis would require those assumptions to be assessed or replaced with an appropriate model. The result shows why effects based on different response thresholds should not be treated as interchangeable.
Assessment Schedules and Target Time
An endpoint measured at week 8 is not automatically the same endpoint as one measured at week 12. If the treatment effect changes over time, a difference in assessment time changes the estimand. A common approach is to estimate effects at a shared target time using longitudinal data, but this requires a model for the time course and adequate observations near that time. A week-8 result should not simply be relabeled as a week-12 result.
Worked Example: Aligning Effects to a Common Visit
In a hypothetical study, the treatment-minus-control mean improvement is estimated as 4 points at week 8 and 6 points at week 12 on the same validated scale. Region A reports the week-12 contrast; Region B reports the week-8 contrast. For illustration, suppose the longitudinal model between weeks 8 and 12 is linear, with the contrast increasing by a constant amount per week. Estimate the week-12 contrast from Region B's week-8 result, using the week-12 data to specify the slope.
The slope of the treatment contrast is \((6-4)/(12-8)=0.5\) improvement points per week. Moving the Region B contrast forward four weeks gives \(4+4(0.5)=6\) points at week 12. On these assumptions, the two estimates align at a 6-point treatment benefit at week 12.
The calculation is an illustration, not a general correction formula. It assumes a linear treatment contrast over weeks 8 to 12 and comparable measurement at each visit. If the response plateaus, changes direction, or has a different trajectory in a region, this extrapolation can be biased. A suitable analysis would use the repeated measurements and report uncertainty in the estimated week-12 contrast.
A Practical Harmonization Workflow
A defensible assessment begins with the clinical question, rather than with an available conversion formula. It should make clear what outcome and time point are being compared, what information supports the measurement mapping, and how conclusions change under plausible alternatives.
Name the construct, scale or response rule, population, and assessment time that define the comparison of interest.
Record instrument items and scoring, thresholds, units, visit timing, and any assessment window that differs across regions.
Check whether validation or paired measurements support conversion across the relevant range and whether the mapping plausibly applies across treatment groups and populations.
Use a justified conversion, common endpoint definition, or longitudinal model; then calculate or standardize the treatment effect for the target population as appropriate.
Compare reasonable thresholds, mappings, or time-course assumptions and state whether the transport conclusion depends on the choice.
Diagnostics, Alternatives, and Limitations
Diagnostics should address whether the proposed mapping works where it is needed, not merely whether a model fits on average. For paired instruments, inspect plots of one score against the other, residual patterns across the score range, and subgroup differences in calibration. For thresholds, tabulate classification agreement and examine how many patients change responder status under alternative rules. For time alignment, inspect observed trajectories, visit coverage near the target time, and uncertainty in model-based predictions. Sparse data near the target visit make extrapolation more dependent on assumptions.
When direct conversion is well supported, mapping to a common scale or endpoint is often interpretable. Another option is to analyze a common subset of items or a common visit, if doing so preserves the clinically relevant question and does not create a selected comparison. A measurement model can represent item-level data when the instruments differ in content or scale, but it brings its own assumptions about how items measure the construct. If none of these approaches is defensible, report the region-specific effects separately and state that direct endpoint-based comparison is limited.
Harmonization does not remove other reasons that effects may differ across regions. A common outcome definition is one requirement for comparability, not evidence by itself that treatment effects are exchangeable or that population differences have been handled. Conversely, apparent agreement after conversion does not prove that the mapping is correct. Uncertainty from calibration, endpoint measurement, and the treatment-effect estimate should be considered together when the data permit.
Common Pitfall
Check Your Understanding
Answer each question, then compare with the solutions.
- A validated mapping is \(Y_A=5+2Y_B\) in treatment and control groups. If the mean contrast on B is 3 points, what is the contrast on A, and what assumption makes this conversion valid?
- Why does a high correlation between two symptom scales not, by itself, prove that their treatment effects are directly comparable?
- Using the normal distributions in the threshold example, what is the treatment response risk difference when the response cutoff is 1 unit?
- A regional trial measured an endpoint at week 8, while the target estimand is at week 12. Name one approach to alignment and one important assumption or limitation.
- If no credible mapping is available between two endpoint instruments, what is a defensible reporting choice?
Solutions. 1. The contrast on A is \(2(3)=6\) points; the additive offset cancels, and the slope must apply consistently to both groups and relevant patients. 2. Correlation measures co-variation, not agreement in units, calibration, thresholds, or responsiveness; a high correlation can coexist with systematic scale differences. 3. Treatment probability is \(\Phi(0)=0.500000\), control probability is \(\Phi(-1)=0.158655\), and the difference is \(0.341345\), or 34.13 percentage points. 4. Estimate the week-12 effect from repeated measurements with a longitudinal model; this depends on the trajectory model and sufficient information to support prediction at the target time. 5. Report region-specific effects and explain that direct endpoint-based comparison is not justified by the available measurement evidence.