Navigating the Bias-Variance Trade-off in Sequential Causal Inference: An Empirical Study of Functional Features and Representation Balancing
Year: 2026
Type: preprint
Abstract: Sequential causal inference faces the dual challenge of time-varying confounding and complex longitudinal dynamics. While deep learning models, particularly those integrating Functional Principal Component Analysis (FPCA) and domain adversarial balancing (DANN), have been proposed to estimate heterogeneous treatment effects, their true empirical advantage over classical machine learning baselines remains underexplored. In this paper, we conduct a rigorous empirical study across four synthetic scenarios (varying confounding strength, non-linearity, and temporal dependencies) with 10 independent random seeds. Our findings reveal three crucial insights: (1) Classical tree-based meta-learners (e.g., Causal Forests and X-Learners) consistently outperform complex sequential deep learning models in Precision in Estimation of Heterogeneous Effect (PEHE) under standard tabular settings. (2) Integrating FPCA into a GRU backbone yields a substantial improvement in deep models within highly non-linear regimes (reducing PEHE from 0.157 to 0.121). (3) Aggressive adversarial representation balancing (DANN) achieves perfect covariate overlap (as demonstrated via t-SNE) but induces a severe loss of prognostic information, exacerbating the Bias-Variance trade-off. We provide an informal theoretical interpretation of this phenomenon and suggest that deep sequential models are not a silver bullet for causal inference.
Source: SocArXiv (OSF Preprints)
Author: Shunxin Yao
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235411138
