📊 statistics

GenAI Powered Dynamic Causal Inference with Unstructured Data

This paper introduces a novel statistical framework leveraging generative AI to enable dynamic causal inference on unstructured data, allowing researchers to estimate how the position of treatment features within sequences like text and video affects outcomes while providing valid confidence intervals.

Original authors: Kentaro Nakamura, Kosuke Imai

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kentaro Nakamura, Kosuke Imai

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out exactly which sentence in a speech convinced a listener to change their mind.

In the past, researchers treated a whole speech (or a photo, or a video) as one single, static block. They would ask, "Did this speech work?" But they couldn't answer, "Did the third sentence work better than the first one?" or "Does the order of the arguments matter?"

This paper, written by Kentaro Nakamura and Kosuke Imai, introduces a new way to answer those questions. They call it GenAI-Powered Dynamic Causal Inference. Here is how it works, broken down into simple concepts and analogies.

1. The Problem: The "Static Photo" vs. The "Movie"

Most old methods treat a piece of text like a photograph. You look at the whole picture and decide if it's "good" or "bad." But real life is more like a movie. Scenes play out one after another. What happens in Scene 3 depends on what happened in Scene 1 and 2.

If you are watching a movie and a character says something shocking, your reaction depends on whether they said it at the beginning (when you were just getting to know them) or at the end (when you already knew their backstory). Old methods couldn't measure this "timing" effect. They just looked at the whole movie and gave it one score.

2. The Solution: The "Magic Translator" (GenAI)

The authors use Generative AI (GenAI) as a special tool. Think of GenAI as a super-smart translator that doesn't just read words; it understands the deep "vibe" or "meaning" behind them.

When you feed a sentence into this AI, it doesn't just see the letters; it converts the sentence into a hidden code (called an "internal representation"). This code captures the meaning, the tone, and the context of that specific sentence.

The authors use this AI to do two things:

Translate the "Noise": Every sentence has the main point (the treatment) and a bunch of other details (the "confounders," like the font, the grammar, or the surrounding words). The AI helps separate the main point from the noise.
Create a "Deconfounder": Imagine you are trying to see if a specific ingredient makes a cake taste better. But every time you add that ingredient, the baker also changes the oven temperature. You can't tell what did it! The AI acts like a magic filter that mathematically "holds the oven temperature constant" so you can see the true effect of just that one ingredient, even though you can't physically stop the baker from changing it.

3. The Method: Rewinding and Fast-Forwarding

The researchers want to know: "What would have happened if we moved this specific sentence to a different spot?"

Since they can't actually go back in time and re-record the speech, they use a statistical trick called Stochastic Intervention.

The Analogy: Imagine you have a deck of cards representing different speech orders. Instead of forcing a specific order (which might be impossible or unrealistic), they gently "nudge" the deck. They ask, "If we slightly increased the chance of this sentence appearing here instead of there, how would the audience's reaction change?"

They use a Neural Network (a type of computer brain) to learn the rules of this game. It looks at the "magic translator" codes from the AI, figures out the hidden patterns, and calculates the cause-and-effect relationship for every single segment of the text.

4. The Real-World Test: The Hong Kong Protests

To prove this works, they tested it on a real experiment about the Hong Kong protests.

The Setup: People were shown short texts (vignettes) arguing why the US should support protesters. These texts had multiple sentences.
The Twist: The key argument (about US legal commitments) appeared in different spots in different texts. Sometimes it was the first sentence; sometimes it was the last.
The Old Way: Would just say, "The text worked."
The New Way: The authors' method found that timing matters. The argument had a much stronger effect when it appeared early in the text compared to when it appeared later. The "magic filter" successfully isolated the effect of when the argument was heard, separate from what the argument was.

5. The Guarantee: It's Not Just a Guess

The paper doesn't just say, "It looks like it works." They did two things to be sure:

Simulations: They created fake data where they knew the answer beforehand. Their method correctly found the answer every time, like a student who gets 100% on a practice test.
Math Proofs: They used heavy-duty math to prove that as you get more data, their method becomes more accurate and their confidence intervals (the "margin of error") are trustworthy.

Summary

In short, this paper gives researchers a new microscope for unstructured data. Instead of looking at a whole text as a blurry blob, they can now zoom in on specific sentences, understand their hidden context using AI, and measure exactly how the order and timing of information changes people's minds. It turns a static "did it work?" question into a dynamic "how and when did it work?" answer.

Technical Summary: GenAI Powered Dynamic Causal Inference with Unstructured Data

Problem Statement

Existing methods for estimating causal effects from unstructured data (text, images, video) typically treat these objects as single, static observations. This approach overlooks the inherent sequential nature of such data, where features unfold over time (e.g., sentence by sentence in text, frame by frame in video). In dynamic settings, the appearance of later features often depends on earlier ones, creating complex confounding structures. Standard adjustment methods fail here because confounding features are high-dimensional, unknown, and evolve over time. Furthermore, traditional causal inference frameworks struggle with the "positivity" assumption in high-dimensional dynamic settings, where prior sequences may nearly determine subsequent treatment assignments, making deterministic interventions infeasible.

Methodology

The authors propose a Dynamic GenAI-Powered Inference (GPI) framework that leverages generative artificial intelligence (GenAI) to handle dynamic interventions in unstructured data. The methodology consists of four main components:

1. Conceptual Framework and Assumptions

Sequential Structure: Unstructured data $X_i$ is modeled as an ordered sequence of segments $X_{i1}, \dots, X_{iS_i}$ .
Treatment and Confounders: For each segment, a deterministic function maps the raw data to a binary treatment feature $W_{is}$ and a low-dimensional vector of unobserved confounding features $U_{is}$ .
Sequential Separability: A key assumption stating that confounding features at a specific segment are not functions of the treatment feature at that same segment, though treatment features may depend on confounders.
Deep Generative Models: The framework utilizes internal representations ( $R_{is}$ ) from a deep generative model (e.g., LLMs). It assumes a "Factorized Deterministic Decoding" where the output segment is a deterministic function of the model's hidden state. This allows the unobserved confounders to be represented as functions of these internal latent states.

2. Identification via Stochastic Interventions

To address the violation of the positivity assumption common in dynamic settings, the authors adopt a dynamic stochastic intervention framework (building on Kennedy, 2019).

Instead of forcing a deterministic treatment, the intervention modifies the probability of treatment assignment at each step $s$ by an incremental parameter $\delta_s$ .
The target estimand $\Psi(\delta)$ is the average potential outcome under this stochastic intervention, where treatment probabilities are shifted relative to the observed data-generating process.
Nonparametric Identification: The authors prove that by adjusting for a "deconfounder" (a low-dimensional summary of the internal representations $R_{is}$ ), the causal quantity can be nonparametrically identified. Crucially, the deconfounder need not recover the true confounders exactly; it only needs to satisfy a mean-independence condition relative to the outcome.

3. Estimation Strategy

The estimation procedure combines neural networks with semiparametric inference:

Neural Network Architecture: A shared neural network architecture is trained to simultaneously learn:
1. The deconfounder function $f(R_{is}, s)$ , which maps internal representations to a low-dimensional space.
2. The outcome regression model $\mu$ , predicting the outcome given the treatment history and deconfounders.
Semiparametric Inference: To ensure valid inference, the authors derive an influence function for the target parameter. This function accounts for the estimation errors in the nuisance parameters (deconfounder, propensity scores, and outcome models).
Cross-Fitting: A $K$ -fold cross-fitting procedure is employed to estimate the nuisance functions and the target parameter, ensuring asymptotic normality and valid confidence intervals even when flexible machine learning models are used.

4. Theoretical Guarantees

The paper establishes that under regularity conditions (including bounded overlap and Lipschitz continuity of nuisance functions), the proposed estimator is:

Consistent: It converges to the true causal effect.
Asymptotically Normal: The estimator follows a Gaussian distribution, allowing for the construction of valid confidence intervals.
Rate of Convergence: The nuisance functions (deconfounder, propensity scores, etc.) must be estimated at a rate of $N^{-1/4}$ , which is achievable with standard deep neural networks.

Key Results

Simulation Studies

The authors conducted numerical simulations based on the motivating example (Hong Kong protests) to evaluate finite-sample performance.

Accuracy: The proposed estimator accurately recovered the target causal effects across various sample sizes ( $N \in \{2000, \dots, 5000\}$ ) and intervention parameters ( $\delta$ ).
Coverage: The 95% confidence intervals achieved coverage close to the nominal level (0.95) in most scenarios. The authors note slight undercoverage in some cases where the data-generating process was not Lipschitz continuous, a necessary condition for the simulation design.

Empirical Application

The method was applied to a randomized experiment by Fong and Grimmer (2023) regarding U.S. support for Hong Kong protesters.

Setup: Text vignettes were composed of multiple argument components. The treatment of interest (U.S. commitment to Hong Kong) appeared at different positions within the text.
Finding: The study revealed that the causal effect of the treatment feature is heterogeneous based on its position. Interventions applied to the first sentence significantly shifted the average counterfactual response, whereas the effect was substantially attenuated when the same feature appeared in the second sentence.
Validation: This result aligns with existing survey research on order effects (e.g., Krosnick and Alwin, 1987) but demonstrates the ability of the proposed method to detect such dynamics in unstructured data where traditional static methods would miss them.

Significance and Claims

The paper claims to make three primary contributions to the literature:

Dynamic Causal Inference for Unstructured Data: It extends the GPI framework (Imai and Nakamura, 2024, 2025) from static to dynamic settings, enabling researchers to estimate causal effects of sequences of treatment features and their specific positions within text or video.
Integration of GenAI and Causal Inference: It demonstrates how internal representations from generative AI models can serve as valid deconfounders for high-dimensional, evolving confounding structures without requiring manual feature engineering.
Methodological Rigor: It provides a semiparametric estimation framework with formal proofs of consistency and asymptotic normality, offering valid confidence intervals for dynamic interventions under stochastic settings.

The authors conclude that while the current application focuses on text, the methodology is extensible to other modalities like video and audio, and could be further developed for settings with sequential outcomes. The work highlights that ignoring the sequential structure of unstructured data can lead to misleading causal conclusions, and that accounting for the timing and placement of features is critical for accurate inference.