The study introduces a novel method for estimating treatment effects from single-arm trials using a latent-variable model. It addresses the challenges of limited sample sizes, ethical concerns, and missing covariate observations. The proposed model outperforms existing methods in both direct treatment effect estimation and patient matching scenarios.
Randomized controlled trials (RCTs) are considered the gold standard for treatment effect estimation but have limitations such as cost and ethical constraints. Single-arm trials offer an alternative but require external control data. The study proposes a deep latent-variable model that can handle missing covariate observations by learning group-specific and shared latent representations.
The model uses amortized variational inference to learn latent spaces for patient matching or direct treatment effect estimation. Evaluation on benchmark datasets and real-world electronic health records shows improved performance compared to previous methods. The approach provides a principled way to handle tasks in treatment effect estimation from single-arm trials.
Key points include the importance of identifying predictive latent spaces, handling structured missingness patterns, and improving performance in both direct estimation and patient matching scenarios. The study highlights the significance of proper modeling of MNAR structures in improving performance across different variations of the proposed method.
إلى لغة أخرى
من محتوى المصدر
arxiv.org
استفسارات أعمق