Personalized Public Policy Analysis in Social Sciences Using Causal-Graphical Normalizing Flows

Sourabh Balgi, José M. Peña, Adel DaoudAAAI 2022 · DOI

Abstract

Traditional public policy analysis often relies on population-level averages, which can lead to inefficient "one-size-fits-all" interventions. This paper presents Causal-Graphical Normalizing Flows (c-GNF), a deep learning framework designed for Personalized Public Policy Analysis (P3A). By integrating the flexibility of normalizing flows with the structural constraints of causal directed acyclic graphs (DAGs), c-GNF enables individual-level counterfactual reasoning. Our results demonstrate that c-GNF accurately predicts individual responses to interventions in complex social systems, outperforming traditional statistical methods when functional forms are unknown.

Introduction

In the social sciences, the Average Causal Effect (ACE) is the standard metric for evaluating policy impact. However, just as personalized medicine tailors treatments to individual biology, P3A seeks to tailor social interventions to individual circumstances. The primary challenge lies in estimating counterfactuals: determining what would have happened to a specific individual under a different policy regime. Traditional methods, such as Inverse Probability Weighting (IPW), often struggle with individual-level predictions and require rigid assumptions about the mathematical relationships between variables. Mis-specification of these functional forms can lead to significant bias.

Methodology: Causal-Graphical Normalizing Flows

The c-GNF framework addresses these challenges by combining two powerful modeling paradigms: Normalizing Flows (NFs) and Causal Directed Acyclic Graphs (DAGs).

Causal mechanisms and the c-GNF architecture for counterfactual inference.

Normalizing Flows utilize invertible deep neural networks to transform simple probability distributions into complex ones that represent observed data. This invertibility is crucial for "Abduction"—the process of inferring an individual's unique, unobserved characteristics (the "individual essence") from their observed outcomes. By embedding a domain-expert-defined DAG into the network architecture, the model ensures that all learned transformations respect the underlying causal structure.

Key technical innovations include:

  • Encapsulated Structural Causal Models (SCM): The model captures complex social dynamics without requiring explicit mathematical equations.
  • Gaussian Dequantization: A novel technique to handle discrete social science variables (e.g., binary graduation status) within the continuous framework of normalizing flows.
  • Three-Step Counterfactual Inference: Following Pearl’s logic, the model performs Abduction (inferring latent noise), Action (intervening on a treatment variable), and Prediction (calculating the counterfactual outcome).

Experimental Setup

To evaluate c-GNF, we utilized a simulated 2-wave social system model reflecting real-world research on neighborhood environments and high school graduation. The simulation included continuous covariates (parental income), binary treatments (neighborhood exposure), and continuous outcomes (test scores). Simulation was necessary because true counterfactuals are unobservable in real-world data; it allowed us to measure the model's accuracy against a known "ground truth."

Results and Discussion

Our evaluation focused on both population-level accuracy (ACE) and individual-level personalization.

Comparison of causal effect estimation methods across different sample sizes.

As shown in Figure 5, c-GNF (green) demonstrates high stability and accuracy across varying sample sizes. While traditional methods like IPW and Regression-with-Residuals (RWR) perform well when the true model is known, they fail significantly under model misspecification. In contrast, c-GNF's use of deep neural networks allows it to learn the "functional form" directly from data, making it highly robust.

The model achieved between 91% and 99.6% accuracy in identifying optimal individual-level treatments. Furthermore, c-GNF provided precise estimates that clearly distinguished treatment effects, reducing the ambiguity often found in traditional confidence intervals that overlap with the "zero-effect" line.

Conclusion

c-GNF represents a significant step toward personalized social interventions. By leveraging the universal approximation capabilities of neural networks within a causal framework, it moves beyond the limitations of average-based policy analysis. Future work will focus on relaxing the requirement for a pre-defined DAG and applying the framework to large-scale observational datasets, such as those tracking the long-term impacts of international economic programs.

Summary updated

BibTeX
@inproceedings{balgi2022personalized,
  author = {Sourabh Balgi and José M. Peña and Adel Daoud},
  title = {{Personalized Public Policy Analysis in Social Sciences Using Causal-Graphical Normalizing Flows}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21437},
}
Figure