StatsPAI: A Python Toolkit Unifying Causal Inference and AI-Agent Interfaces

Forum topic · 小凯 · 2026-06-20

Summary

StatsPAI is an MIT-licensed Python package incubated under Stanford's REAP project, packaging 1,000+ functions across 23 causal-inference method families (DID, IV, RD, synthetic control, DML, causal forests, meta-learners, TMLE, neural causal, causal discovery) into a single `import`. Its design targets three pain points: the fragmentation between R and Stata workflows, the lack of reproducible publication-quality output, and the absence of machine-readable interfaces for LLM-based agents. Each registered function exposes a discoverable schema and an MCP-server scaffold (`sp.agent.mcp_server`), allowing external agents to invoke estimators through tool-calling. A validation-tier system (certified, validated, api_stable, experimental) signals numerical alignment with R/Stata references, with `xtabond` recently reaching machine-precision parity. All estimators return a unified `CausalResult` object whose `regtable` helper exports side-by-side tables to Word, Excel, and LaTeX using journal templates such as AER. The article weighs StatsPAI's advantages (unified API, modern ML-causal coverage, automated robustness tooling, free cost, agent-ready schema) against Stata's decades of production hardening and R's role as the first-mover for frontier econometric methods, while flagging limits in certified-function coverage, Python ecosystem maturity, and residual agent hallucination risk.

Overview

StatsPAI is a Python package from Stanford's REAP project, aiming to consolidate fragmented causal-inference tooling (Stata's proprietary environment, R's 20+ incompatible packages, Python's gap on DID/RD/DML) into one import. At v1.16.0 it ships 1,000+ registered functions across 23 method families, 310k lines of core code, and 156k lines of tests.

Key points

Agent-first interface layer: Every function exposes machine-readable metadata via sp.list_functions(), sp.describe_function(), and sp.function_schema() (OpenAI/Anthropic tool-calling format). A bundled MCP server scaffold (sp.agent.mcp_server) lets external LLMs invoke registered functions in natural language. The library treats agents as first-class users rather than retrofitting a human-oriented API.
Validation-tiered honesty: Each function carries a validation_status flag (certified, validated, api_stable, experimental). The xtabond (Arellano-Bond dynamic panel GMM) estimator was rewritten to achieve machine-precision parity with Stata before being promoted to certified.
Unified CausalResult object: All estimators return the same object with .summary(), .plot(), .to_latex(), .to_word(), .to_excel(), and .cite() methods. The sp.regtable helper merges results across DID, IV, RD, and DML into journal-styled tables (AER/QJE/Econometrica templates) in a single line.
Honest cross-tool comparison: The README concedes where Stata (40+ years of production validation, svy: survey support, reviewer trust) and R (frontier-method first-movers such as fixest, did2s, HonestDiD, plus ggplot2 and CRAN peer review) still lead.
Causal "full stack" coverage: Classical econometrics (DID variants, 2SLS/LIML/GMM, CCT/robust RD, ADH/ASCM/gsynth/SDID, PSM/CEM/entropy balancing), ML causal (DML, GRF causal forests, S/T/X/R/DR meta-learners, TMLE, DeepIV), neural causal (TARNet, CFRNet, DragonNet), and frontier methods (NOTEARS/PC/LiNGAM discovery, conformal causal inference, Mendelian randomization, Bayesian causal forests, policy learning). Each family ships with a method-selection decision tree, assumption checklist, and diagnostic flow rather than bare API docs.
Publication-grade output pipeline: sp.regtable replaces multi-step LaTeX formatting with .to_word()/.to_excel()/.to_latex(); handles mixed-magnitude formatting and automatic asterisk annotations.
Estimand-first workflow DSL: sp.causal_question(...) forces researchers to declare treatment, outcome, population, estimand, and design before estimation. The returned IdentificationPlan can be frozen into a pre-registration markdown file via plan.to_markdown().
Automated robustness "interrogation": sp.spec_curve() explores the full grid of control sets × subsamples × standard-error types; sp.oster_bounds() computes Oster (2019) selection bounds; sp.sensitivity_dashboard() summarizes every check on a single page.

Limitations

Certified coverage is still a minority of the 1,000+ functions; most sit at api_stable or validated.
Compared to Stata's 40+ years of production hardening, edge-case handling and large-sample performance are unproven.
Schema exposure does not eliminate LLM hallucination risk on method choice or assumption interpretation; sp.causal_question mitigates but does not replace human judgment on complex designs.
R remains the first-mover for frontier econometric methods, so StatsPAI must continuously track upstream packages to avoid a perpetual catch-up posture.

Takeaway

StatsPAI's strategic contribution is not any single estimator but its architectural bet: schema layers as system calls, validation tiers as a certification regime, and the unified CausalResult object as a data-exchange format. If certified coverage expands and adoption grows inside the LLM-agent ecosystem, it could become the default causal-inference standard library for Python + AI-agent settings rather than a replacement for Stata or R.

References

GitHub: https://github.com/brycewang-stanford/statspai
PyPI: https://pypi.org/project/StatsPAI/