How to Construct a COA Endpoint Under Draft PFDD Guidance 4

In modern clinical development, teams frequently assume that demonstrating an instrument is psychometrically sound marks the conclusion of outcome strategy. Under the Food and Drug Administration's (FDA) draft guidance Patient-Focused Drug Development: Incorporating Clinical Outcome Assessments Into Endpoints for Regulatory Decision-Making (PFDD Guidance 4, April 2023), that assumption stops one step too early. Establishing that a Patient-Reported Outcome (PRO), Clinician-Reported Outcome (ClinRO), Observer-Reported Outcome (ObsRO), or Performance Outcome (PerfO) is fit-for-purpose under final October 2025 PFDD Guidance 3 is necessary, not the endpoint. Sponsors still have to convert COA scores into an estimand-aligned trial endpoint, prespecify how they will interpret clinically meaningful change, and read trial results against that plan without treating statistical significance as patient benefit. Completing the plan is not labeling, not COA qualification, and not FDA acceptance of the endpoint.

For clinical development, biostatistics, and outcome-assessment teams who already have a fit-for-purpose COA, the draft's recommended operator sequence is:

• Treat Guidance 4 strictly as draft guidance: The Federal Register notice of availability is 88 FR 20530 (6 April 2023), docket FDA-2023-D-0026. The series sits in FDA's PFDD program under the 21st Century Cures Act and FDARA Title I (PDUFA VI). The landing page still marks the document 'Not for implementation. Contains non-binding recommendations.' Under 21 CFR 10.115, draft Level 1 guidances are labeled draft, are nonbinding, and are implemented after public comment and issuance of a final version. An interpretation plan supports a regulatory discussion; it does not guarantee endpoint clearance, formal COA qualification, or labeling.

• Lock the endpoint definition before collection starts: The protocol and statistical analysis plan (SAP) should define the COA type, named instrument, specific score, missing-item rules, timing and baseline (not screening scores), role (primary, secondary, or other), and a specific trial objective.

• Default to baseline-adjusted continuous follow-up models: The draft's usual recommended analysis is a comparison of randomized groups on the follow-up score or a prespecified summary, after adjusting for baseline. Cite FDA's May 2023 final covariate-adjustment guidance for that recommendation; Guidance 4 still footnotes the May 2021 draft and was not itself revised. Dichotomize only when the question is the probability of a clearly defined important health state, and then prespecify one independently derived threshold.

• Derive interpretation thresholds from independent external anchors: Meaningful score difference (MSD) and meaningful score regions (MSR) should be derived from independent anchors (and other well-justified methods) before the registration trial. Distribution-based statistics do not consider the patient voice and are insufficient as the sole basis for MSD.

• Apply a range for continuous effect interpretation vs. a single threshold for dichotomized endpoints: For continuous endpoints, compare the average treatment effect and its confidence interval to a prespecified MSD or MSR range. For categorized responder endpoints, one independent threshold defines the endpoint. Empirical cumulative distribution function (eCDF) plots can supplement interpretation; they do not replace the prespecified group comparison.

Clinical trial sponsors often conflate instrument validation with endpoint validity. While qualitative concept elicitation under PFDD Guidance 2 establishes what matters to patients, and psychometric evaluation under finalized PFDD Guidance 3 confirms that an instrument is fit-for-purpose, neither activity defines the trial endpoint itself. Guidance 4 addresses an entirely distinct scientific question: how to operationalize COA scores into an endpoint capable of supporting valid statistical inferences regarding how patients feel, function, or survive.

Draft Guidance 4 states that a fit-for-purpose COA is necessary for a strong endpoint rationale, but it is not sufficient. Footnote 12 is the illustration: a symptom-intensity COA that is fit-for-purpose can still yield a weak endpoint if the endpoint averages intensity over seven days when patient input and the product's mechanism identify worst intensity as the relevant experience.

Similarly, selecting whether an outcome should be captured by a patient, clinician, observer, or performance measure—a decision addressed during early development in COA modality selection—does not determine whether the score should be analyzed as a continuous mean change, a time-to-event milestone, or a categorized responder rate. Table 1 of final Guidance 3 provides a suggested format for communicating the COA-based endpoint approach (meaningful aspect of health, concept of interest, COA type, name and score, and the endpoint). Table 2 of that same guidance is the eight-component fit-for-purpose rationale, not the endpoint-construction table. Guidance 3 still refers construction and interpretation of the endpoint to Guidance 4.

Regulatory review of COA endpoints frequently uncovers fatal ambiguities in the protocol or statistical analysis plan. A COA-based endpoint cannot simply cite an instrument name. The protocol and statistical analysis plan should lock the endpoint definition before collection starts.

According to draft Guidance 4, sponsors should clearly describe the COA-based endpoint, including:

• Assessment Classification: Explicit designation of the COA type (PRO, ClinRO, ObsRO, or PerfO) and the named instrument used to measure the concept of interest.

• Targeted Metric: The exact score utilized for primary inference (e.g., total composite score, a specific validated subscale score, or an individual core symptom item).

• Item-Level Missingness Algorithm: Prespecified rules and scientific justification for handling missing item responses within an administration session when computing scale scores, adhering strictly to the instrument's validated scoring manual.

• Timing and Longitudinal Aggregation: The precise assessment visit schedule, permissible completion windows, recall periods, and the mathematical formula used to combine multi-day diary entries (e.g., weekly mean of the worst daily pain scores with a minimum requirement of four valid daily entries).

• Rigorous Baseline Definition: A clear mathematical calculation of baseline. Crucially, The draft recommends that screening COA scores used to confirm severity for enrollment should not be used as the baseline value. Using screening scores can inflate apparent treatment responses through regression to the mean. A later pre-randomization assessment should serve as baseline; when multiple pre-randomization measurements are taken, the protocol should define how baseline is calculated.

• Regulatory Hierarchy and Role: The endpoint's role (primary, secondary, or other), the corresponding specific trial objective, and the intended indication.

• Context-Specific Fit-for-Purpose Justification: Evidentiary rationale explaining why the instrument and endpoint are valid in the specific enrolled population, trial design, and disease stage.

• Patient and Caregiver Importance: Direct empirical or qualitative evidence establishing that the score captures an aspect of health that patients or their caregivers recognize as important in daily life.

Sponsors frequently attempt to justify an endpoint definition by citing its acceptance in a previously approved drug application. Draft Guidance 4 explicitly rejects this defense: prior regulatory precedent does not constitute adequate justification. Context of use inevitably differs across trial populations, concomitant standard of care changes over time, and regulatory science continually evolves. An endpoint accepted a decade ago under the 2009 PRO guidance may fail modern evidentiary expectations under the PFDD framework.

A single COA instrument can generate multiple distinct endpoint constructions: continuous score changes, time-to-event outcomes, responder proportions, or composite indices. The protocol should select an architecture that directly answers the trial's core scientific question without unnecessarily sacrificing statistical efficiency.

For continuous or ordinal COA scores, draft Guidance 4 firmly recommends evaluating the follow-up COA score at a predefined time point or a summary over a predefined post-baseline period, with randomized arms compared using linear regression or ANCOVA models adjusting for baseline score. As finalized in FDA's May 2023 guidance Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products, baseline adjustment significantly enhances statistical precision, reduces variance, and protects Type I error control in parallel-group trials.

The draft also flags important challenges in utilizing percent change from baseline as a default primary endpoint construction. While popular in clinical literature, percent change introduces severe statistical pathologies: it is mathematically undefined when baseline scores equal zero, highly unstable when baseline values are small, inherently asymmetric (a 50% reduction requires a 100% increase to reverse), and frequently violates normality assumptions. The draft's illustration: two patients with equal-and-opposite absolute changes of +5 and −5 produce an average change of 0 on the original scale, but an average percent change of +25% that appears to show improvement. Percent change can therefore misstate the direction of the average effect relative to the original metric.

Transforming continuous COA scores into binary 'responder' endpoints (e.g., proportion of patients achieving a ≥10-point improvement) is widely requested by clinical teams seeking straightforward interpretations. However, draft Guidance 4 highlights that dichotomizing continuous data discards critical variation, substantially reduces statistical power, and requires larger sample sizes.

Dichotomization is methodologically justified only when the primary scientific objective is estimating the probability of entering a clearly defined, clinically recognizable health state. Permissible examples cited by FDA include:

• Complete Symptom Resolution: Measuring the proportion of patients achieving absolute freedom from pain or nausea in acute settings (such as acute migraine therapy).

• Clinically Validated Categorical Thresholds: Investigator global assessment of acne lesions as clear or almost clear, the draft's illustration citing FDA's May 2018 acne-effectiveness guidance—not a transferable cutoff for other indications.

When dichotomization is used, the score threshold should be prespecified and derived from data other than the registration trial used to demonstrate effectiveness. Using that trial both to discover the cutoff and to demonstrate effectiveness is circular. Sponsors should also explore treatment effects over a range of thresholds.

When a pathology impacts multiple distinct domains of feeling and functioning, sponsors should make an explicit trade-off among three structural endpoint models:

• Separate Endpoints Across Domains: Testing individual symptoms or functional impacts as distinct endpoints (e.g., co-primary endpoints or a primary endpoint supported by ranked secondary endpoints). This preserves clinical transparency but increases sample size demands and requires formal multiplicity error control under FDA's October 2022 Multiple Endpoints in Clinical Trials guidance.

• Multi-Component Endpoints: Combining two or more components within a patient by a prespecified algorithm—scores from different COAs, subscales of one COA, or, only for a composite-indicator measurement model, individual items. While this avoids multiplicity penalties and captures disease heterogeneity, it introduces dilution: averaging across affected and unaffected domains can obscure meaningful improvements in dominant symptoms. All multi-component endpoints rest on some implicit or explicit weighting scheme; the draft says sponsors should make those weights explicit and justify them. If a combined effect is found, component-level analyses may be needed so that benefit is not driven by trivial items while important functions are unchanged.

• Personalized Endpoints (e.g., Most Bothersome Symptom, Goal Attainment Scaling): Allowing individual patients to designate their primary baseline symptom or functional goal for prospective tracking. While personalized endpoints maximize patient-centric relevance, they present formidable methodological challenges: patient priorities shift over time, selected goals may not align with the drug's mechanism, and deterioration in unselected domains can go unrecorded. The draft says the process to construct a personalized endpoint should be standardized, and the same full set of relevant assessments should still be collected for all patients so individual symptoms and functions can be compared.

Draft Guidance 4 aligns COA collection and analysis with the ICH E9(R1) estimand framework. In trials utilizing COAs, statistical failures frequently stem from conflating item non-response, missed clinical visits, and intercurrent events into a generic 'missing data' bucket.

Trial teams should keep three distinct methodological boundaries:

• Item-Level Non-Response: When a patient skips individual questions or tasks, missingness should be handled with the instrument scoring algorithm. The draft does not replace that algorithm with an ad hoc trial-level imputation model for incomplete items.

• Missing Assessment Visits: When an entire scheduled assessment is missed (due to missed study visits, technical transmission failures, or administrative oversight), this represents true missing data. The statistical analysis plan should propose methods that account for missingness with respect to a particular estimand. Guidance 4 notes that repeated COA scores can still enter a longitudinal mixed-effects or generalized-estimating-equations model when the contrast is at a fixed time point; sensitivity analyses belong with the same estimand, as detailed in missing data sensitivity frameworks.

• Intercurrent Events: Post-randomization events that affect the ability to observe an endpoint or the interpretation of that endpoint—including death, rescue medication, treatment discontinuation, or new assistive-device use—are not missing data. They are intercurrent events governed by ICH E9(R1). Potential intercurrent events and the strategy for each belong in the SAP under ICH E9(R1). This article uses that alignment so the COA endpoint remains reconstructable; it does not retell the published estimand-specification article.

The draft warns that analyses of subgroups defined by post-baseline events (for example, patients alive and on treatment) are likely to be misleading because randomization is broken. If the clinical question is treatment effect regardless of adherence, the protocol should plan to continue follow-up and administer the COA after discontinuation of treatment.

The draft recommends collecting only necessary COAs, reducing burden, counseling patients, providing reminders, notifying the site of missed assessments, and collecting reasons for missingness to inform sensitivity analyses.

Statistical significance does not, by itself, indicate whether a detected effect corresponds to a clinically meaningful treatment effect. The draft's job in Section III is to interpret that magnitude in terms of how patients feel or function. To bridge this divide, draft Guidance 4 formalizes two core interpretability concepts: Meaningful Score Difference (MSD) and Meaningful Score Regions (MSR). Understanding their derivation is essential for trial planning.

• Meaningful Score Difference (MSD): The magnitude of difference between two scores on a COA scale that patients, caregivers, or clinicians judge to represent an important, noticeable change in health status. While frequently established via within-patient change over time (improvement or deterioration), MSD can also be derived via cross-sectional methods such as vignette judgments or Idio Scale Judgment.

• Meaningful Score Regions (MSR): A qualitative mapping that partitions the continuous score continuum into distinct clinical categories representing qualitatively different patient experiences (e.g., categorizing an abstract 0–100 functional disability score into 'no impairment,' 'mild functional limitation,' 'moderate limitation,' and 'severe dependence'). Techniques for deriving MSRs include bookmarking, item-response theory (IRT) mapping, and standard-setting methodologies.

As documented in a 2025 review in the Journal of Patient-Reported Outcomes, draft Guidance 4's introduction of MSD and MSR generated extensive public docket commentary (including submissions from PhRMA and BIO on Docket FDA-2023-D-0026) raising concerns that introducing new terms alongside established concepts—such as Minimal Clinically Important Difference (MCID), Minimally Important Change (MIC), and responder thresholds—risks regulatory confusion. The 2025 Journal of Patient-Reported Outcomes review notes that Guidance 4 remains draft, that docket comments worried MSD and MSR would add confusion beside MCID, MID, MIC, and responder language, and that the final document may change terms. Use the draft's terms as the operating vocabulary of this still-draft document, define them on first use, and do not treat them as a finalized FDA glossary or as a labeling guarantee.

FDA explicitly requires anchor-based methods as the primary evidentiary foundation for establishing MSD and MSR. An anchor is an external, clinically interpretable variable—such as a Patient Global Impression of Severity (PGIS) or Patient Global Impression of Change (PGIC)—that shares an established conceptual relationship with the COA.

Draft Guidance 4 establishes strict criteria for valid anchors:

• Conceptual Congruence: The draft prefers an anchor that matches or includes the COA's concept of interest (for example, a global pain-severity rating to interpret a bodily-pain subscale, rather than a distal health-utility index).

• Multiple Anchor Triangulation: The draft recommends multiple anchors and, citing Coon and Cook 2018, relating COA differences to those anchors rather than relying on a single convenient scale.

• Mitigating Recall Bias: While widely used, PGIC ratings suffer from significant retrospective recall errors and present-state bias (patients often rate their 'change' based entirely on how bad they feel today, regardless of baseline severity). Guidance 4 recommends evaluating longitudinal shifts in PGIS categories (e.g., moving from 'moderate' to 'mild' severity) alongside PGIC assessments to cross-validate results.

• Distribution-Based Methods Are Supportive Only: Distribution-based statistics (such as 0.5 standard deviations, standard errors of measurement [SEM], or Cohen's d effect sizes) measure only the statistical noise and precision of the scale. They contain no input from patients regarding whether a score shift matters in daily life. Guidance 4, like the still-listed 2009 PRO guidance, states that distribution-based methods are insufficient as the sole basis for an MSD or responder definition; they can describe measurement variability.

The draft recommends deriving interpretation thresholds before the start of the registration trial, using independent Phase 2 data, longitudinal observational studies, or dedicated qualitative anchor studies. If a sponsor relies on published literature thresholds, it should explain why those MSDs or MSRs can be generalized, covering comparability of disease, population, standard of care, location, calendar time, COA version, endpoints, and follow-up. For transformed scales (for example, mapping Likert responses to 0–100), the draft says to report the threshold on both the transformed and raw scales; the MSD should be at least a one-category change on at least one raw item.

Once an empirical interpretation framework has been established, the statistical analysis plan should say how trial results will be evaluated against those thresholds. Draft Guidance 4 establishes a fundamental operational dichotomy based on the endpoint's mathematical construction:

When the primary endpoint is a continuous baseline-adjusted mean score, sponsors should not force the data into a binary responder analysis. Instead, the analysis should directly compare the estimated between-group treatment effect (and its 95% confidence interval) against the prespecified range of plausible MSD estimates:

• Definitive Demonstration of Benefit: If values of the treatment effect that are consistent with the observed data (reflected by the 95% confidence interval) lie entirely above the largest MSD estimate, the draft says this strongly suggests that the average effect corresponds to a difference most patients would consider meaningful.

• Ambiguous or Inconclusive Benefit: If the confidence interval overlaps the MSD range, the draft's illustration is that the comparison does not clearly correspond to a meaningful overall difference, even though a small portion of patients might still experience a treatment effect they regard as meaningful.

• Baseline-Dependent Heterogeneity: If empirical derivation demonstrates that MSD varies based on baseline impairment (e.g., patients with severe baseline disability require a larger absolute point reduction to notice improvement than mildly affected patients), the SAP should compare the treatment effect to baseline-specific MSD values rather than a single constant.

To substantiate continuous findings, protocols should provide supplementary empirical cumulative distribution function (eCDF) and probability density (ePDF) plots showing the full distribution of within-patient score changes across treatment arms. As established in FDA's predecessor 2009 PRO guidance (which remains active guidance until the PFDD series is finalized), eCDF curves annotated with the prespecified MSD threshold visually demonstrate the proportion of patients in each randomized group who achieved meaningful benefit, preventing extreme outliers from distorting the interpretation of group means.

When an endpoint categorizes patients into binary responder states (e.g., proportion achieving ≥2-category improvement on an MSR scale), a range cannot be used during primary hypothesis testing. The protocol should prespecify a single, exact threshold derived strictly from external historical data.

During submission, regulatory reviewers will inspect:

• Pre-Randomization Lock: Confirmation that the single cutoff was locked in the protocol and SAP before unblinding.

• Complete Independence: Proof that the cutoff was not calibrated or optimized using unblinded registration trial data.

• Threshold Sensitivity Analyses: Analyses exploring treatment effects over a range of thresholds, so the observed benefit is not an artifact of a single cutoff.

Locking the endpoint definition, choosing a construction that matches the scientific question, keeping missing items, missing visits, and intercurrent events on the same estimand, and deriving an independent MSD or MSR range before the registration trial gives a sponsor a discussion package. It does not finalize Guidance 4, qualify the COA, or guarantee labeling.