End-To-End Causal Effect Estimation from Unstructured Natural Language Data

Summary

This work aims to the causal effect estimation from unstructured observational text data, proposing a pipeline based on LLM to extract the treatment, covariate and outcome from the unstructured data by LLM. It's a interesting exploration for causla inference and the pipeline process of the inference method is quite standard.

Strengths

(1) The problem solution for treatment effect estiamtion is quite novel. (2) The proposed LLM pipeline can address the limitation of the current effect estimation method, i.e., can not deal with the unstructured data.

Weaknesses

(1) some details are missing, leading to the poor readability. (2) Some equations are not correct (based on my understanding). (3) The presnetation of the whole methodology is not good enough, alought I am not an expert of LLM, I know how LLM works. However, i can not imagine how the proposed pipeline work and how it estimate ATE very clearly. (4) The evaluation part is kind of weak, given that baselines are too limited and simple.

Questions

1.there are some claims that are not well expalined, leading to confuse people. For exampel, "value may be recouped from outcomes that would otherwise be lost" this claim sounds good but i don't what does it mean here. 2.the overall idea of this work to utilize LLM to explore the causal effect estimation from unstructured data is good, i like it. But the writing can not match the good idea. For example, line 54-59, i have no idea about what the authors try to express, The transition from sentence to sentence is too stiff 3.The binary and discrete assumption about the outcome Y and covariates X is kind of too strict. Typically, one can assume the teatment assignment is binary but outcome and covariates are continous. It's not clear why this work adopt such assumption. if so, the applying scale of the proposed method would be limited. also, the assumption 4 is also too strong. 4.in the outcome predictor eq. (9), why it is P(Y=1|R_i,X_i,T_i) instead of P(Y=t|R_i,X_i,T_i)? I don't understand. When t=0, the outcome predictor still ouput the expectation over P(Y=1|xxx)? 5.the description how to extrct the covariates and their corresponding values is not clear. How to determine the feature values? For each report, the set of covarivates is the same, right? Can you present a specific example? what does each feature value mean in the real-world? The readability of the method section is poor. 6.In the evaluation table 1, the "IPW (Structured) " is a baseline or ground truth? there is no any discussion in the main body about this baseline and i am so confused about the discussion below because of it, given that the "IPW (Structured) " performs the best. 7.how does the ground truth of ate comes from? Can you clarify this point? by the way, i like the idea of this work, but the presentation of this paper is really poor. thus if the above questions can be addressed well, i will consider to increase the score.

Rating

4

Confidence

4

Soundness

3

Presentation

1

Contribution

3

Limitations

authors have adequately addressed the limitations.

Summary

The paper introduces a family of causal effect estimators named NATURAL, designed to use LLMs for mining causal effect estimates from observational text data. The authors address the challenge of automating data curation and using LLMs to impute missing information, presenting a method that conditions on structured variables to assist in computation of causal effect estimators. The authors develop six datasets (two synthetic and four real) to evaluate their method, and demonstrate that NATURAL estimators produce causal effect estimates within close range of ground truth from randomized trials.

Strengths

1. Clear and thoughtful use of LLMs to estimate causal effects from text data. The authors present a cool methodology, properly formalized, for using LLMs to extract and impute structured variables from observational text data. 2. Development and evaluation of six datasets, including real-world clinical trial data. 3. Strong results showing causal effect estimates are within a close range of ground truth values.

Weaknesses

1. The original claims made in the abstract+intro seems to be too grandiose, but in fact it is about the use of LLMs to classify variables of interest from the text. I would change the framing of the paper a bit to better reflect that. It would make the claims of the paper less objectionable and would better demonstrate the usefulness of NATURAL. 2. There is a vast literature in the social sciences on recovering interpretable variables directly from text for estimating causal effects from observational data. Most of these approaches rely on probabilistic approaches like topic models, but recently there’s growing interest in doing so automatically with LLMs. The authors should better address this literature, and compare their approach to relevant methods if possible. 3. The approach is really cool, but like any ML-pipeline approach it surely introduces errors at each step of the process (see for example Egami et al @ NeurIPS 2023). Each error will then bias the causal effect estimation. From the results of the paper I’m convinced empirically your approach works well, but would I be able to understand whether it produces accurate estimates on a new dataset? A deeper analysis and perhaps a way to propagate error/uncertainty to the estimation would be very helpful here.

Questions

1. Authors explain that as part of their process, they “filter for reports that are likely to conform to the experimental design”. Could this introduce bias? 2. All of the “classical strategies” used assume no-hidden confounding. Could you imagine how someone would use an LLM with a more complex identification strategy (say IV, or RDD)? You claim at the end of the intro that you can anticipate someone doing so, I wonder if you can elaborate.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have addressed the limitations of their work, acknowledging the dependency on the quality and calibration of LLMs and the potential inaccuracies in approximating the required conditional distributions. They emphasize that NATURAL estimators should not be used for high-stakes decision-making without experimental validation and stress the need for domain expert involvement to ensure causal assumptions are met.

Summary

Estimating causal effects is costly and time-consuming. The authors propose to use large language models (LLMs) to mine unstructured text data for causal effect estimation. This paper introduces NATURAL, a family of causal effect estimators using LLMs to process unstructured text data. This seems to be a good application of LLMs for the task of causality.

Strengths

- Determine the effects of treatments can be expensive and time-consuming so the paper introduces a novel task of effect estimation using unstructured natural language data. - This method is adaptable to various domains where textual data is abundant. - This seems to be a more practical application of LLM for the task of causality.

Weaknesses

- LLMs might inherit biases present in their training data or hallucinate. - Social media users might not reflect the demographics or behaviors of the broader population. - Finetuning the model to one domain may not lead to good generalisation in others, during fine-tuning does the LLM learn inductive biases regarding the "topic" or task?

Questions

- How do the authors make sure that all of the confounders are accounted ? I can imagine there to be many confounders mentioned in the text. - There could be a bias in social media text, I thought people would report negative cases more often than positive, did you observe any such biases?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Mentioned

Summary

The paper seeks to use LLMs for a workflow that extracts structured variables from free text, filters the dataset, and then computes a causal estimate using the imputed variables. The paper applies this methodology to two synthetic and four real-world datasets and shows that it produces estimates that are comparable to those from known ground truth or RCT estimates.

Strengths

This is a very ambitious paper that pulls together a complicated workflow for estimating causal effects from real-world natural language datasets. Just showing that this can be done is a significant contribution. The paper is clearly written and the main points are easy to follow. While I am somewhat skeptical of any practical applicability of this work, the paper is reasonable conservative in highlighting its limitations.

Weaknesses

The paper at almost every turn relies on prompting an LLM and implicitly assuming that the model returns data from a desired distribution. In general, this will introduce measurement error which could systematically bias the overall method's estimates. See for example [1] and [2] below. The synthetic data evaluation is fairly limited. It would be nice to at least understand how performance varies as sample size increases from a few hundred examples to many thousands (especially as the IPW oracle baselines are not particularly good). This would also allow you to stress-test the different assumptions being made. If you have synthetic data in which reports for certain patients are systematically truncated or censored, or where different groups write reports in different languages or styles, how does that affect the different steps (i-vi)? [1] Ashwin, Julian, Aditya Chhabra, and Vijayendra Rao. "Using Large Language Models for Qualitative Analysis can Introduce Serious Bias." Development Research (2023). [2] Egami, Naoki, et al. "Using imperfect surrogates for downstream inference: Design-based supervised learning for social science applications of large language models." Advances in Neural Information Processing Systems 36 (2024).

Questions

Line 267-268: What is the point of sampling a persona from Big Five personality traits? How does this simulate realism? Line 147-149 says "if all reports are the constant, empty string ... we have full access to the true observational joint density." Am I correct that this is a necessary assumption for the overall method to work in theory, but the actual implementation (steps (v) and (vi) in lines 206-220) would not work if all reports are the constant, empty string? This may be obvious, but it could be worth clarifying that when you say "we cannot formally guarantee that [our method] satisfies Assumptions 3 and 4" you mean the form of Assumption 4 that does not involve all reports being empty strings. In Tables 1 and 2, which LLMs are used? LLAMA2-70B? Overall, it would be helpful to more clearly label which LLMs are used when. In Line 204, the paper says that the GPT-4 API sufficed for filtering; was GPT-4 used anywhere else? How was the decision made of which LLM to use where? It might be helpful to have a figure or appendix that walks through every step (the paper does this for many steps, just not all in one place) in which an LLM is queried and discusses what assumptions are being made.

Rating

7

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

Overall, the paper does a good job of not overclaiming. However, see weakness 1 above. I would also suggest the authors discuss the challenges of using RCT estimates to validate observational studies -- there are many reasons why those estimates might not align (selection bias, dropout, etc).