FELISHA: An Intelligent Pipeline for Causal Inference Assisted by Large Language Models

The FELISHA project combines large language models (LLMs) with causal inference methods, builds an automated analysis pipeline based on the Petersen-van der Laan causal roadmap, implements statistical computing via R language bridging, and aims to lower the professional threshold for causal inference, providing intelligent causal analysis tools for fields such as social sciences, medicine, and business.

In the fields of data science and statistics, causal inference is a core method to understand causal relationships between variables, aiming to answer the question of 'how Y would change if X were altered'. However, causal inference analysis is highly professional and complex, and researchers face the following challenges:

• Understanding complex statistical theories (e.g., potential outcome framework, structural causal models)
• Choosing appropriate identification strategies (e.g., instrumental variables, regression discontinuity)
• Correctly implementing statistical estimation methods
• Interpreting results and evaluating the rationality of assumptions These thresholds limit the application of causal inference in broader fields.

Based on the Petersen-van der Laan Causal Roadmap

FELISHA takes this roadmap as its theoretical foundation, dividing the analysis into stages such as causal problem definition, model construction, identification strategy selection, statistical estimation, and sensitivity analysis, guiding users to complete the analysis systematically.

LLM-Assisted Intelligent Process

In the project, LLMs play roles in natural language understanding (converting research questions into causal queries), method recommendation, R code generation, result interpretation, hypothesis testing, etc.

R Language Bridging

Seamlessly integrate R's statistical computing capabilities (e.g., packages like MatchIt, AIPW) through the R bridging layer, enabling data conversion, package management, and result integration between Python and R.

Autonomous Master Control Mode

Supports full-process automation from problem understanding to result generation. Users only need to provide data and questions; the system requests confirmation at key nodes and finally generates a complete report.

Core Components

1. Causal Graph Engine: Constructs causal graphs, supports d-separation testing and backdoor criterion identification
2. LLM Interface Layer: Encapsulates interactions with LLM providers such as OpenAI and Anthropic
3. R Integration Layer: Implements Python-R bridging via rpy2
4. Workflow Engine: Manages the state and execution order of the analysis process
5. Report Generator: Compiles analysis results into structured documents

Typical Usage Process

1. Data upload (supports formats like CSV, RData)
2. Describe research questions in natural language
3. Automatically/interactively construct causal graphs
4. System recommends identification strategies and estimation methods
5. Automatically generate and execute R code
6. View statistical results and visual charts
7. Export reports containing the complete process

Social Science Research

Helps researchers in fields like economics and sociology clarify causal hypotheses, implement matching/weighting methods, and evaluate the robustness of results.

Medicine and Public Health

Supports causal effect estimation in observational studies, survival data processing, and mediation analysis to understand mechanisms of action.

Business Analysis

Used for marketing campaign effect evaluation, product revision impact analysis, and identification of driving factors behind user behavior.

Advantages

• Lower threshold: Non-professionals can perform standardized causal analysis
• Quality assurance: Systematic processes reduce common errors
• Educational value: Demonstrates decision-making processes to aid causal inference learning
• Reproducibility: Generated code and reports support analysis reproducibility

Limitations

• Hypothesis dependence: Results rely on untestable assumptions
• Complex scenarios: Automatic methods lack flexibility in highly complex designs
• Domain knowledge: Cannot replace experts' understanding of causal mechanisms
• LLM limitations: May produce hallucinations; key decisions require manual verification

Future Outlook

1. Integrate cutting-edge methods such as Bayesian causal inference and causal machine learning
2. Enhance interactive causal graph editing and result visualization
3. Support team collaboration to share models and analysis results
4. Customize and optimize for disciplines like economics and medicine

Conclusion

FELISHA combines the natural language capabilities of LLMs with rigorous statistical methods, providing an intelligent auxiliary tool for causal inference. It lowers the entry threshold, promotes the standardization and popularization of causal inference, and is an open-source project worthy of attention for researchers without a statistical background.