statspai-analysis

StatsPAI: agent-native causal inference, paper-ready

StatsPAI is one import statspai as sp over 1,500+ registered functions for causal inference and applied econometrics, numerically aligned with Stata / R reference implementations, with self-describing schemas and result objects that export to LaTeX / Word / Excel / BibTeX. This skill drives it through the canonical pipeline of an applied empirical paper, in one of three modes (default applied econ; epidemiology; ML-causal), and leaves paper-ready artifacts on disk at every step.

• Source: https://github.com/brycewang-stanford/StatsPAI
• Install: pip install "statspai[fixest,plotting]" (default pipeline); add neural for Dragonnet / TARNet / CEVAE, text for text-as-treatment. The bare install cannot run sp.feols (any y ~ x | fe regression) or make figures — see references/operating-loop.md for the extras matrix.
• Verified: validate_api_claims.py checks this skill against the installed package (statspai skill validate). Every sp.* name resolves, every sp.*(...) call in a code block binds to the real signature (for a function that takes **kwargs, to its agent schema or forwarding target), and the result attributes on the gate's smoke-fit list exist. With --run the pipeline files are also executed end to end on synthetic data, which is what catches a call that binds but is rejected at run time. There is no version stamp to go stale.
• Is this copy current? An installed skill is a plain copy and does not follow pip install -U statspai. statspai skill status compares it with the skill packaged in the installed version; if it reports stale, say so and suggest statspai skill install --force before relying on a call shape written here.
• Paper: Wang & Rozelle (2026), Journal of Open Source Software 11(125), 10604, https://doi.org/10.21105/joss.10604.

Operating loop (every request)

1. Route the mode first from the user's words — default applied econ, Mode A epi, Mode B ML-causal, or an export-only path. Do not run the whole paper pipeline for "make Table 1" or "export this regression". A request for one estimate, one check or one table takes the short path in references/quick-path.md.
2. Freeze the contract before estimating: y, treatment / exposure, unit / time ids, estimand, design, required artifacts, install extras. Infer a missing field only when the column names make it obvious; otherwise return a short blocking checklist instead of inventing columns.
3. Pick the estimator with the tables, not from memory: sp.route('did', design='staggered', timing_random='no', covariates='none') (also iv, rd, matching, ml_causal, qte, dynamic_panel) returns the registered call, an example, why, and the assumptions it adds; sp.decision_guide(family) lists the questions. sp.recommend(df, ...) routes from the data itself.
4. Start from the smallest verified call shape: the skeleton in references/verified-skeleton.md, then the section snippet. For anything unfamiliar, sp.describe_function(name) / sp.function_schema(name) before writing code; sp.search_functions("weak instrument robust CI") to find the name.
5. Widen one block at a time: data contract → plan → diagnostic figure → main estimate → robustness / export. After each block read the result's violations(), next_steps(), degradations and model_info before passing it downstream.
6. Gate the answer on artifacts: the final message lists the files produced, the identifying assumption, the estimator class, and any failed or skipped gate. Never say "paper-ready" when an export or a required diagnostic did not actually run.
7. Turn failures into bounded corrections: a StatsPAIError carries code, recovery_hint, diagnostics and alternative_functions; fix the smallest wrong rule and continue from the last verified artifact.