P124: Robust Method for Longitudinal Data via Modeling Sequential Correlation
Poster Presenter
Ran Tu
Phd Student
Georgetown University United States
Objectives
The objective of this study is to propose a semiparametric model for longitudinal data, with sequential dependence fully implemented, and developing a robust causal inference method to compare group differences.
Method
We propose a semiparametric model that captures sequential dependence using B-splines for function approximation. We incorporate double robust estimator for causal inference. The performance is assessed by simulations, with comparisons to Generalized Estimating Equations and random effects models.
Results
Simulation studies show that the proposed method enhances parameter estimation efficiency, effectively accounting for sequential dependence, and improves function approximation. Compared to existing approaches, our model achieves lower bias and higher power across various parameter settings. The double robust estimator approach ensures greater robustness against misspecification, leading to precise treatment effect estimation.
Conclusion
By accounting for sequential dependencies and enhancing estimation robustness, this methodology improves longitudinal study designs in clinical trials, epidemiology, and real-world data applications. The proposed method exhibits inflated Type I error due to the initial spline-based estimation of the link function, introducing relative bias in parameter estimation. This bias is significant compared to the small sample variance. To address this, we adjusted the sample variance before computing the Wald statistic, controlling Type I error around 0.05 across various parameter settings. Having resolved this issue, future work will explore missing data mechanisms in causal inference with longitudinal data and validate the model’s performance on real-world longitudinal datasets.
