Introduction
In causal inference, identifying the effect of a treatment on an outcome requires accounting for confounders—hidden variables that influence both. Traditional models often treat these confounders as static, independent variables. However, real-world scenarios, such as agricultural yields or economic trends, involve latent factors that evolve over time. This paper introduces a non-parametric framework designed to handle these stochastic confounders, providing a robust alternative to deep learning methods.
Modeling Temporal Confounders
The core innovation of this framework lies in extending Structural Causal Models (SCMs) to account for temporal dependencies. The authors model latent confounders as stochastic processes where the state at any given time $t$ is conditioned on the previous state at $t-1$.
By acknowledging that hidden variables have "memory," the framework captures the true dynamics of the system. This approach prevents the bias that occurs when time-varying factors, like soil fertility or market sentiment, are incorrectly assumed to be constant or independent across observations.
Kernel-Based Variational Inference
Rather than relying on deep neural networks, which often require extensive hyperparameter tuning and can yield inconsistent results in causal tasks, the authors utilize a kernel-based approach. They derive an Augmented Representer Theorem that functions within random input spaces, allowing for analytical solutions to complex non-linear functions in a Reproducing Kernel Hilbert Space (RKHS).
The model employs a Variational Auto-Encoder (VAE) logic, optimizing an Evidence Lower Bound (ELBO) to approximate the posterior of latent confounders. Crucially, this optimization is performed through kernel weights rather than neural network parameters, offering a more mathematically grounded and "elegant" solution for handling non-linearity without the instability of deep learning architectures.
Experimental Evaluation
The framework was tested against several benchmarks, including synthetic temporal datasets (TD2L, TD4L, TD6L), the Infant Health and Development Program (IHDP) dataset, and a real-world Gold-Oil price dataset.
The results highlight the importance of modeling temporal dependencies. On the complex TD6L dataset, the proposed model achieved an Average Treatment Effect (ATE) error of 0.237, vastly outperforming the CEVAE baseline, which scored 0.705. Furthermore, the model demonstrated faster convergence, reaching stable estimates with training sequences as short as $T=45$.
In real-world application, the model analyzed the causal link between oil and gold prices. It estimated that a rise in oil prices leads to a 0.77% increase in gold prices, a finding that closely aligns with established economic literature (e.g., Simáková 2011), which estimated a 0.64% increase using Granger causality.
Conclusion and Future Directions
This research demonstrates that kernel methods can effectively outperform deep learning in temporal causal inference tasks. While the current framework assumes a fixed one-step time-lag, it remains highly competitive even on static, i.i.d. data, suggesting the stochastic process assumption is a safe generalization. Future work will explore multi-step delays and the application of this framework to higher-dimensional proxy variables, addressing the computational scaling challenges inherent in kernel matrices for very large datasets.
Summary updated