Treatment Effect Estimation With Potential Outcomes for a Single-Arm Trial Compared With Historical Controls: A Case Study of Locally Recurrent Head and Neck Squamous Cell Carcinoma Patients Treated With Adjuvant Nivolumab

Nusrat Harun; Rodney Sparapani; Purushottam W Laud; Siva Sivaganesan; Jennifer L Leddon; Sulsal Haque; Alice L Tang; Trisha M Wise-Draper

doi:10.1002/hed.28215

Treatment Effect Estimation With Potential Outcomes for a Single-Arm Trial Compared With Historical Controls: A Case Study of Locally Recurrent Head and Neck Squamous Cell Carcinoma Patients Treated With Adjuvant Nivolumab

Head Neck. 2025 Jun 17. doi: 10.1002/hed.28215. Online ahead of print.

Authors

Nusrat Harun¹, Rodney Sparapani², Purushottam W Laud², Siva Sivaganesan³, Jennifer L Leddon⁴, Sulsal Haque⁴, Alice L Tang⁵, Trisha M Wise-Draper⁴

Affiliations

¹ Cytel Inc, Cincinnati, Ohio, USA.
² Division of Biostatistics, Medical College of Wisconsin, Milwaukee, Wisconsin, USA.
³ Division of Statistics and Data Science, Department of Mathematical Sciences, University of Cincinnati, Cincinnati, Ohio, USA.
⁴ Division of Hematology/Oncology, Department of Internal Medicine, University of Cincinnati, Cincinnati, Ohio, USA.
⁵ Department of Otolaryngology-Head and Neck Surgery, University of Cincinnati, Cincinnati, Ohio, USA.

PMID: 40525328
DOI: 10.1002/hed.28215

Abstract

Background: Time/resource constraints might preclude a randomized controlled trial. Single-arm oncology trials with historical controls are an alternative. With causal inference, treatment effect estimates can be computed in the absence of randomization.

Methods: From a single-arm trial of 39 head and neck squamous cell carcinoma patients treated with adjuvant nivolumab, we compare 2-year disease-free survival (DFS) to untreated historical controls. We resort to the potential outcomes framework known as Rubin's causal model (RCM). For time-to-event outcomes, RCM relies upon survival analysis regression with baseline covariates. We contrast the average treatment effect (ATE) estimated by three survival methods: Cox proportional hazards (CPH) versus machine learning alternatives, random survival forests (RSF), and Bayesian Additive Regression Trees (BART).

Results: The ATE in favor of nivolumab: CPH 0.202 (0.098-0.306); RSF 0.159 (0.070-0.248); and BART 0.268 (0.126-0.406).

Conclusions: The uncertainty is considerable, yet all three methods show nivolumab is superior to control.

Keywords: BART; G‐computation; RCM; Rubin's causal model; average treatment effect; machine learning; potential outcomes; single‐arm trial; value function.