Conference Agenda

Patient reported outcomes (PROs) are collected on a daily basis in clinical trials to measure patients’ quality of life, e.g. symptoms. Often these data are reported in a small-range ordinal scale and analyzed without considering their longitudinal aspect. The emergence of electronic data collection methods for home-based measurements has enabled routine, daily capture of various symptom scores, highlighting the need to develop statistical methods for the analysis of these intensive ordinal longitudinal data. Of interest are both their mean structure over time and variability which is known to be linked to disease progression and to be affected by treatments. To model the dynamics of ordinal PROs, we propose a location-scale latent process model which includes two types of variability across patients: individual underlying level flexibly modelled over time (e.g., with splines) using random effects and covariates, and individual short-term variability with a variance of the error which is expressed as a linear structure of covariates such as treatment arm and a patient-specific random intercept. The model is estimated in a Maximum Likelihood framework with an interface in R. The high-dimensional intractable integrals in the optimization are approximated using the Quasi-Monte Carlo method. The estimation procedure is validated by a simulation study, and we apply the methodology to data from two clinical trials, one in asthma and one in COPD, to evaluate the effect of treatment on dynamics of the respiratory symptoms and their variability.

Background / Introduction: Incorporating external data, such as external controls, holds the promise of improving the efficiency of traditional randomized controlled trials especially when treating rare diseases or diseases with unmet needs.

Methods: In the first part of the talk, we describe novel weighting estimators grounded in causal inference. From a trial design perspective, operating characteristics including Type I error and power are particularly important and results will be presented. In the latter part, we describe proper estimation and inference of long-term treatment effect during the open-label extension phase in the absence of placebo-controlled patients. Within the framework of causal inference, we propose several difference-in-differences type methods and a synthetic control method for the combination of randomized controlled trials and external controls.

Results: For the first part, our proposed weighting estimators achieve significant power gain, while maintaining Type I error close to the nominal value of 0.05 under the simulation settings investigated. For the latter part, our realistic simulation studies demonstrate the desirable performance of the proposed estimators in a variety of practical scenarios. In particular, difference-in-differences type methods outperform synthetic control method and are the recommended methods of choice in scenarios similar to the ones we have investigated.

Conclusion: In both studies, we assessed in our realistic simulation studies representing a variety of practical scenarios and provided an application through a phase III clinical trial in rare disease.

References

Zhou, X., Pang, H., Drake, C., & Zhu, J. (2024). Causal estimators for incorporating external controls in randomized trials with longitudinal outcomes. Journal of the Royal Statistical Society Series A, qnae075.

Zhou, X., Pang, H., Drake, C., Burger, H. U., & Zhu, J. (2024). Estimating treatment effect in randomized trial after control to treatment crossover using external controls. Journal of biopharmaceutical statistics, 34(6), 893–921.

Background

Recruiting patients for clinical trials in rare neurodegenerative diseases is challenging, because of ethical concerns surrounding placebo enrollment and difficulties in determining optimal sample size for statistical significance. Innovative statistical techniques such as prediction powered inference for clinical trials (PPCT) (Poulet et al., 2025) can improve trial statistical power and reduce sample size requirements. Amyotrophic Lateral Sclerosis (ALS) clinical trials are driven by two key outcomes: survival as the primary measure and the ALSFRS-R score as the secondary measure. This study aims to apply PPCT to estimate the average treatment effect, first by using the classical PPCT estimator for continuous outcomes and then by adapting the method for survival outcomes.

Methods

Natural history data from various observational datasets were used to train a progression model capturing the natural disease development in untreated patients (digital placebo twins). A joint model was applied to account for ALS disease progression features, which include both longitudinal and event-based data (Disease Course Mapping). Then, we applied this model to predict the disease progression of clinical trial patients. The PPCT method incorporated these predictions into trial analyses. PPCT first compares treated patients with their digital placebo twins, to estimate the average treatment effect (continuous outcome) or the hazard rate estimator (survival outcomes). To ensure robustness, it accounts for prediction errors and placebo effects by comparing the predicted placebo progression to the observed one, thereby debiasing the final estimations. We adapted PPCT to survival outcomes, by applying the general PPCT estimator formula to the hazard rate estimator.

Results

By applying PPCT to the Kaplan-Meier estimator, we enhanced the statistical power of the log-rank test. We further validated the methodology by using PPCT on the treatment effect estimator, which compares the mean ALSFRS-R score progression between the placebo and treatment groups. We observed a significant reduction in its variance with narrower confidence intervals. To assess the accuracy of our predictions, we analyzed the correlation between the predicted and observed progression (R²). Higher R² values were associated with narrower confidence intervals.

Discussion

Our study underscores the importance of leveraging high-quality observational data to accurately train the joint spatiotemporal model and reach high statistical power for both average treatment effect estimation and log-rank test. The application of PPCT methodology and its adaptation to survival outcomes is a relevant method either to reduce the sample size requirement (pre-trial) or to improve the statistical analysis of the clinical trial results (post-trial).

Background: Traditionally, a clinical trial with multiple treatment arms is analysed as if the arms are completely unrelated to each other. However, the trial design may lead to some arms having shared features. A common example is the study of several doses of the same drug, where the trial has a titration phase, so that the doses are the same at the first few visits. The idea of the proposal is to account for this feature in the analysis, following the basic principle of analysing a trial according to its design.

Methods: Technically, it is very easy to account for this feature in the analysis and this makes the results more coherent. The approach is illustrated with results from a depression trial of adjunctive brexpiprazole in doses of 1 and 3 mg/day compared to placebo for patients with inadequate response to antidepressants. The benefits are also documented with simulations, covering both analysis of a full trial and performance in case the trial design calls for an interim analysis.

Results: While the precision improvement in placebo comparisons at the final visit is only small, the comparison of the active arms becomes more precise, and it also improves precision of treatment effects at the earlier visits. A further advantage is that interim analyses become more precise because the method better utilizes the measurements at early time points, which are present for patients ongoing at the time of doing the interim analysis. As this is suggested as a drug development approach, we also show how it fits into the estimand framework.

Conclusion: Considering the simplicity of the approach, it is beneficial for the primary analysis, even though the precision improvement is only small. The benefits for secondary analyses and interim analyses are more important, substantiating the relevance of the suggestion.