New Causal State-Space Model Improves Longitudinal Treatment Effect Estimation

This paper presents a new approach to causal inference, specifically for estimating how individual treatments affect outcomes over time using observational data. Current methods often face a dilemma: they try to make data representations independent of treatment assignment, which can inadvertently remove crucial information needed for accurate outcome prediction. The authors formalize this problem and propose two new models, CSSD and CSSPD. CSSD adapts selective State Space Models with a direct multi-step decoder to predict all future outcomes simultaneously, avoiding cumulative errors. CSSPD enhances CSSD by incorporating Contrastive Predictive Coding and Local Information Maximisation. This augmentation helps maintain temporal predictability within the data representation and recovers local covariate details that might otherwise be lost. Experimental results on medical and cancer simulation datasets show that CSSPD significantly outperforms existing methods like the Causal Transformer, particularly for longer prediction horizons and across various confounding strengths, while maintaining efficient computational costs. This work offers a structured solution to a fundamental challenge in causal inference.