Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments

Technical Summary: Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments

Problem Statement
The paper addresses the challenge of estimating causal effects when treatments are unstructured, high-dimensional objects, such as texts. In traditional causal inference with text, researchers must learn a low-dimensional representation of confounding features (e.g., tone, length, specific topics) from the data to adjust for them. Existing methods, such as those using BERT embeddings or topic models, attempt to learn these representations directly from the treatment texts. However, this process often struggles to disentangle the specific treatment feature of interest (e.g., military background) from other correlated confounding features, leading to biased estimates and violations of the overlap assumption. Furthermore, learning these representations from data is computationally intensive and may not yield accurate causal estimates, particularly when confounding is strong.

Methodology: GenAI-Powered Inference (GPI)
The authors propose the GenAI-Powered Inference (GPI) methodology, which leverages deep generative models, specifically Large Language Models (LLMs), to enhance causal inference. The core innovation lies in utilizing the true internal representation of generated texts directly from the LLM, rather than learning a representation from the data.

1. Experimental Design:

   • Researchers use an LLM to generate treatment objects (texts) based on specific prompts. Alternatively, existing texts can be "reused" by instructing the LLM to reproduce them exactly.
   • The LLM is required to operate in a deterministic decoding mode (e.g., greedy search with temperature set to zero) to ensure the output is a deterministic function of the internal representation.
   • The internal representation ($R_i$) of the generated text is extracted directly from the LLM's hidden layers.

2. Key Assumptions:

   • Separability (Assumption 5): The treatment feature ($T$) and confounding features ($U$) must be separable. Specifically, $T$ cannot be a deterministic function of $U$, and $U$ cannot be a deterministic function of $T$. This implies that one can intervene on the treatment feature without altering the confounding features.
   • Deterministic Decoding (Assumption 6): The generative model's output is a deterministic function of its internal representation, ensuring that the treatment object depends on the prompt only through this representation.

3. Identification and Estimation:

   • Nonparametric Identification: Under the separability and deterministic decoding assumptions, the authors prove the existence of a deconfounder function $f(R_i)$ that summarizes the confounding features. This deconfounder is a lower-dimensional representation of the internal state $R_i$ that satisfies conditional independence: $Y_i \perp R_i \mid T_i, f(R_i)$.
   • Estimation Strategy: The authors employ a neural network architecture based on TarNet to simultaneously estimate the deconfounder and the conditional outcome function.
   • Double Machine Learning (DML): To obtain asymptotically valid confidence intervals, the method applies DML. Crucially, the propensity score is modeled as a function of the estimated deconfounder, not the raw internal representation. This avoids the violation of the overlap assumption that would occur if the propensity score were modeled directly on the high-dimensional internal representation (where the treatment feature is a deterministic function).
   • Perceived Treatments: The methodology is extended to settings where respondents perceive treatment features differently using an instrumental variables approach, where the actual treatment feature serves as an instrument for the perceived feature.

Key Contributions

• Utilization of True Internal Representations: Unlike prior work that learns representations from data (e.g., via BERT embeddings), GPI extracts the true vectorized representation directly from open-source LLMs. This eliminates the need to learn causal representations from the data, improving both accuracy and computational efficiency.
• Formalization of Separability: The paper formally establishes the "separability" assumption as a necessary condition for nonparametric identification in this context, linking it to the concept of disentanglement and the overlap assumption.
• Diagnostic Tools: The authors propose diagnostic tools, including the analysis of propensity score distributions and the Independence-of-Support Score (IOSS), to detect potential violations of the separability assumption.
• Efficient Estimation: By using the internal representation and a specific neural network architecture, the method avoids the "curse of dimensionality" and overlap violations common in existing causal representation learning methods.

Results
The authors evaluate GPI through simulation studies based on the candidate profile experiment (Fong and Grimmer, 2016) and an empirical application using the same dataset.

• Simulation Studies:

  • Under the separability assumption, GPI (using both new generated texts and text reuse) demonstrates significantly smaller bias and root mean squared error (RMSE) compared to state-of-the-art methods (Outcome Model with BERT and DML with BERT).
  • GPI maintains proper 95% confidence interval coverage across weak, moderate, and strong confounding scenarios. In contrast, BERT-based methods suffer from severe undercoverage or breakdown under strong confounding.
  • Computational Efficiency: GPI is more than ten times faster than BERT-based estimators.
  • Text Reuse: The "text reuse" approach (regenerating existing texts) yields lower variance (RMSE) than generating new texts, likely due to the consistency of the internal representations.
  • Violation of Assumptions: When the separability assumption is violated (i.e., treatment and confounding features are entangled), all methods, including GPI, perform poorly, highlighting the importance of the assumption.

• Empirical Application:

  • Applied to the candidate profile experiment, GPI estimates a positive and statistically significant effect of military background on voter evaluations (ATE $\approx$ 4.85).
  • Competing BERT-based methods produced conflicting results: one yielded a negative significant effect, and the other an unreasonably large positive effect (ATE $\approx$ 45.7), likely due to overlap violations.
  • The IOSS for GPI was 0.10, indicating better support overlap compared to 0.41 for BERT-based methods.

Significance and Claims
The paper claims that GPI significantly enhances the validity of causal inference with unstructured treatments by leveraging the capabilities of GenAI. The primary significance lies in the ability to:

1. Disentangle treatment features from confounding features using the true internal representations of LLMs.
2. Avoid the need to learn causal representations from data, thereby reducing estimation error and computational cost.
3. Formalize the conditions (separability) required for valid identification, providing a theoretical foundation for using GenAI in causal inference.
4. Extend causal inference to perceived treatment features via instrumental variables.

The authors note that while the focus is on text, the methodology is applicable to images and potentially videos, provided the internal representations can be effectively managed. They emphasize that the method relies on the availability of open-source LLMs that allow access to internal representations and deterministic decoding.