Selecting a Patient-Reported Endpoint With Longitudinal Models

Why Endpoint Choice Is a Modeling Decision

A patient-reported outcome (PRO) can be measured in several plausible ways: symptom intensity, symptom interference, physical function, or a broader health-status score. The choice affects more than the wording of a questionnaire. It determines what treatment benefit the trial estimates, how much measurement variability enters that estimate, and whether clinicians and patients can interpret the result.

A longitudinal model helps compare candidate endpoints over time, rather than selecting one from a single visit or a collection of separate tests. But a small standard error does not make an endpoint important to patients, and a familiar scale is not automatically sensitive to the treatment effect. The decision should combine the model-based treatment contrast with measurement quality, clinical meaning, and the practical ability to collect the endpoint consistently.

Modeling Candidate Endpoints Over Time

Consider a randomized trial comparing a new treatment with placebo. Patients complete a symptom diary and a function questionnaire at baseline and weeks 4, 8, and 12. For each candidate endpoint, define an improvement score so that a positive value means improvement. For example, symptom improvement can be baseline score minus follow-up score if lower symptom scores are better; function improvement can be follow-up score minus baseline score if higher function scores are better.

A repeated-measures model can estimate treatment contrasts at each post-baseline visit while accounting for correlation among observations from the same patient. One common specification for the improvement score \(Y_{ij}\) for patient \(i\) at visit \(j\) is:

Here, \(T_i=1\) for a patient assigned to treatment and \(T_i=0\) for placebo; \(B_i\) is the baseline endpoint score; \(\mu_j\) is the placebo mean improvement at visit \(j\) for a patient at the mean baseline value; and \(\theta_j\) is the treatment-minus-placebo contrast at that visit. The errors for one patient across visits have a covariance matrix \(\Sigma\). Thus, the model allows a patient's repeated measurements to be correlated; measurements from different patients are treated as independent.

This specification assumes a suitable mean model, an adequately modeled within-patient covariance, and no unmodeled dependence between patients. Likelihood-based inference also relies on the missing-data assumption used for the analysis. In a typical missing-at-random analysis, the probability of a missing score may depend on observed scores and other included information, but not on the unobserved score after conditioning on that information. Missingness caused by unobserved worsening can violate this assumption. The trial should define the estimand and handling of events such as treatment discontinuation before choosing an endpoint; the example here targets outcomes regardless of discontinuation, with observed follow-up analyzed under the stated missing-at-random assumption.

The visit-specific estimate \(\hat{\theta}_j\) is interpreted in the endpoint's units. A positive value indicates greater mean improvement with treatment, given the scoring convention above. The treatment effect can change over time, so the confirmatory-trial decision should identify a primary visit or a scientifically justified summary before looking at the results.

Why Measurement Variability Matters

When measurement variability is large relative to the treatment contrast, estimates are less precise. In a simple two-arm comparison with independent patients and a common outcome standard deviation \(\sigma\), the standard error of the difference in means is:

This follows because the variance of a difference between independent sample means is the sum of their variances, and the variance of each mean is \(\sigma^2/n\). The formula is a simplification: a longitudinal model may gain precision by using baseline and repeated observations, and its standard error depends on the covariance structure. Still, it shows why an unreliable or noisy endpoint can weaken evidence even when its scale seems clinically attractive.

Averaging repeated reports can reduce random measurement error, but only to the extent that errors are not strongly correlated. Suppose daily measurement errors have standard deviation \(\sigma_e\), and the correlation between any two days is \(\rho\). The variance of the mean of \(m\) daily errors is:

If daily errors are independent, \(\rho=0\), averaging reduces the variance by a factor of \(m\). If they are positively correlated, the reduction is smaller. This calculation concerns random error, not whether the questionnaire measures an outcome patients value or whether the recall period is appropriate.

Worked Comparison: Precision, Clinical Meaning, and Selection

Endpoint selection requires a clinical interpretation for each scale. A within-patient threshold for meaningful change is not automatically the right threshold for a between-arm mean difference. For this case, assume patient-anchor work and clinical discussion have justified trial-level thresholds of 1.0 pain point and 5.0 function points for the treatment-minus-placebo mean contrast. These are hypothetical decision benchmarks for this trial, not universal values.

Diagnostics, Alternatives, and Failure Modes

A model estimate is useful only if the outcome and its analysis behave plausibly. Before selecting an endpoint, examine the raw trajectories and model diagnostics, including:

• Response distributions. Strong floor or ceiling effects can prevent patients from showing further improvement, compress treatment contrasts, and make normal-based summaries misleading.
• Residual and covariance behavior. Check residual patterns by visit and treatment, influential observations, and whether the fitted within-patient covariance represents the observed correlations adequately. A poorly fitting covariance can distort standard errors and intervals.
• Missingness and completion. Compare completion over time and by treatment, baseline severity, and prior reported scores. Differential dropout or worsening-related nonresponse can make a missing-at-random analysis questionable; sensitivity analyses may then be needed.
• Measurement performance. Assess reliability, recall period, item interpretation, responsiveness, and burden. A high completion rate does not prove that a measure captures the outcome patients consider important.
• Time course. Check whether the estimated effect develops, plateaus, or diminishes in a clinically plausible pattern. Selecting the most favorable visit after inspecting results risks overstating evidence.

The main alternative is a single-visit analysis, such as baseline-adjusted analysis at week 12. It can be transparent when week 12 is the sole decision-relevant time, but it does not directly use intermediate outcomes to describe the trajectory. A repeated-measures model can estimate multiple visit-specific contrasts and account for their correlation; it does not remove the need to prespecify which contrast drives the endpoint decision.

A change-from-baseline analysis is another option, but an unadjusted change score can be less precise than an analysis that accounts for baseline and can be sensitive to baseline imbalance and measurement error. An area-under-the-curve summary can represent overall symptom burden, but it combines timing and magnitude according to its integration rule; an early benefit followed by loss of effect may be obscured by the same average as a steady, smaller benefit. A responder endpoint can be easier to communicate as the proportion meeting a threshold, but dichotomizing a continuous score discards information and makes results sensitive to the chosen cutoff. These approaches answer different questions; none is a substitute for defining the intended patient-relevant estimand.

A common mistake is to choose whichever candidate has the smallest p-value or the largest standardized effect. That reasoning can reward a scale with a convenient variance or a favorable chance finding, even if its interpretation is weak or its measurement properties are poor. The correction is to prespecify the intended treatment contrast, assess uncertainty, use clinically justified between-arm interpretation, and evaluate validity, reliability, burden, and sensitivity together. A statistically detectable difference is not automatically meaningful to patients, and a clinically attractive scale is not automatically capable of measuring a treatment effect reliably.