Generalized Control Functions via Variational Decoupling.

Aahlad Manas Puli, Rajesh RanganatharXiv 2019 · OpenReview

Abstract

Causal effect estimation relies on separating the variation in the outcome into parts due to the treatment and due to the confounders. To achieve this separation, practitioners often use external sources of randomness that only influence the treatment called instrumental variables (IVs). We study variables constructed from treatment and IV that help estimate effects, called control functions. We characterize general control functions for effect estimation in a meta-identification result. Then, we show that structural assumptions on the treatment process allow the construction of general control functions, thereby guaranteeing identification. To construct general control functions and estimate effects, we develop the general control function method (GCFN). GCFN's first stage called variational decoupling (VDE) constructs general control functions by recovering the residual variation in the treatment given the IV. Using VDE's control function, GCFN's second stage estimates effects via regression. Further, we develop semi-supervised GCFN to construct general control functions using subsets of data that have both IV and confounders observed as supervision; this needs no structural treatment process assumptions. We evaluate GCFN on low and high dimensional simulated data and on recovering the causal effect of slave export on modern community trust.

Introduction

Estimating causal effects from observational data is a fundamental challenge in science and policy. The primary obstacle is confounding—unobserved variables that influence both the treatment and the outcome. While Instrumental Variables (IVs) provide a path to identification by offering external randomness, traditional methods like Two-Stage Least Squares (2SLS) or standard Control Function (CFN) methods often rely on restrictive structural assumptions, such as linearity or additive noise. When these assumptions are violated—for instance, in the presence of multiplicative confounding—classical estimators often fail to provide unbiased results.

Causal graph illustrating the relationships between the instrument (ε), treatment (t), outcome (y), and hidden confounder (z).

The GCFN Framework

This paper introduces the General Control Function Method (GCFN), a flexible framework designed to handle complex, non-linear interactions. GCFN operates in two distinct stages:

  1. Variational Decoupling (VDE): This stage learns a "general control function" (GCF) denoted as $\hat{z}$. Using an autoencoder-inspired architecture, the model learns to map the treatment and IV to a latent space that can reconstruct the treatment. A critical component is a mutual information penalty that ensures $\hat{z}$ is independent of the IV, effectively "trapping" the confounding variation within the latent variable.
  2. Outcome Modeling: The learned GCF $\hat{z}$ serves as a proxy for the hidden confounder. The causal effect is then estimated by regressing the outcome on the treatment and $\hat{z}$.

Meta-Identification and Semi-Supervision

The authors provide a meta-identification result (Theorem 1), proving that any control function that can reconstruct the treatment while remaining jointly independent of the IV guarantees identification of the causal effect. Furthermore, the paper introduces a Semi-Supervised GCFN variant. This approach utilizes small subsets of data where confounders are observed to guide the learning of the control function, significantly relaxing the need for structural assumptions about the treatment process.

Experimental Results

GCFN was evaluated against several benchmarks, including 2SLS, standard CFN, and deep learning methods like DeepIV and DeepGMM.

  • Robustness to Non-Linearity: In simulations featuring multiplicative interactions in the treatment and outcome processes, GCFN demonstrated superior performance. While traditional methods like CFN and 2SLS produced high RMSE (0.58 and 0.55 respectively), GCFN achieved a significantly lower RMSE of 0.09.
  • High-Dimensional Data: On benchmarks using MNIST images as proxies for instruments or confounders, GCFN performed competitively with specialized deep learning estimators.
  • Real-World Application: The method successfully replicated the causal findings of Nunn and Wantchekon (2011) regarding the impact of historical slave exports on modern community trust in Africa.
RMSE comparison in a multiplicative setting, showing GCFN's robustness to confounding strength compared to CFN and DeepIV.

Conclusion

GCFN offers a robust alternative to traditional IV methods by shifting the focus from outcome-process assumptions to the treatment process. By leveraging variational decoupling and independence penalties, it provides a flexible tool for causal inference in complex, non-linear environments. While it requires a strong instrument and some structural knowledge of the treatment process, its ability to handle non-additive confounding represents a significant advancement in the field.

Summary updated

Abstract

Causal effect estimation relies on separating the variation in the outcome into parts due to the treatment and due to the confounders. To achieve this separation, practitioners often use external sources of randomness that only influence the treatment called instrumental variables (IVs). We study variables constructed from treatment and IV that help estimate effects, called control functions. We characterize general control functions for effect estimation in a meta-identification result. Then, we show that structural assumptions on the treatment process allow the construction of general control functions, thereby guaranteeing identification. To construct general control functions and estimate effects, we develop the general control function method (GCFN). GCFN's first stage called variational decoupling (VDE) constructs general control functions by recovering the residual variation in the treatment given the IV. Using VDE's control function, GCFN's second stage estimates effects via regression. Further, we develop semi-supervised GCFN to construct general control functions using subsets of data that have both IV and confounders observed as supervision; this needs no structural treatment process assumptions. We evaluate GCFN on low and high dimensional simulated data and on recovering the causal effect of slave export on modern community trust.
BibTeX
@misc{puli2019generalized,
  author = {Aahlad Manas Puli and Rajesh Ranganath},
  title = {{Generalized Control Functions via Variational Decoupling.}},
  year = {2019},
  eprint = {1907.03451},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.03451},
}
Figure