Claims and evidence¶
This page lists what StatsPAI says about itself on its website, README, PyPI page, documentation and MCP listing, and how each statement can be checked. A statement that cannot be checked here should not appear elsewhere. If you find one, please open an issue.
Last reviewed: 2026-11-12, StatsPAI 1.43.1. Counts are regenerated at each release; the commands give the live values.
What StatsPAI is¶
StatsPAI is an open-source, agent-native Python toolkit for econometrics and causal inference, validation-tiered against Stata and R.
| Claim | Evidence | Re-check |
|---|---|---|
| Open source, MIT licence | LICENSE; no GPL / AGPL runtime dependencies (project rule) |
pip show statspai → License-Expression: MIT |
| Python 3.10–3.13 | requires-python = ">=3.10"; CI matrix |
python -c "import statspai" on each version |
| One import covers the API | every public function is sp.<name> |
sp.list_functions(kind="function") |
| 1,680 registered functions, 92 submodules (1.43.1) | the function registry | python scripts/registry_stats.py |
| Agent-native: every public function has a machine-readable schema | schemas/functions.json, sp.function_schema(name) |
sp.function_schema("feols") |
| Agent-native: an MCP server ships with the package | statspai-mcp console script, server.json |
statspai-mcp --version; statspai mcp doctor |
| Structured results | every result class has to_dict, violations, next_steps, result_card, cite |
scripts/result_protocol_audit.py --check |
| Peer-reviewed software paper | Wang and Rozelle (2026), Journal of Open Source Software 11(125), 10604, doi:10.21105/joss.10604 | sp.citation() |
Numerical evidence¶
"Validation-tiered" means each function carries a grade saying what its numbers have been checked against. Only the first row is a claim of agreement with Stata or R.
| Grade | Meaning | Functions (1.43.1) |
|---|---|---|
bit-exact (T2) |
matches a named R / Stata reference on identical data, estimate and SE relative error ≤ 1e-6 | 501 |
aligned |
matches within a documented, looser tolerance | 75 |
analytical-only |
recovers a known true value on a deterministic design | 395 |
external-replication |
reproduces numbers from a published paper | 57 |
unverified |
registered API with no numerical evidence attached yet | 652 |
Re-check with sp.parity_summary(), or for one function
sp.parity_status("feols") (reference, tolerance, test file). The full table
is the parity matrix; interval coverage and speed, including
failures, are in the evidence scorecard.
Wording rule. Say "matches Stata / R" only for a function graded
bit-exact or aligned, and name the reference. Do not describe the whole
package as "matching Stata and R".
Agent integration¶
| Claim | Evidence | Re-check |
|---|---|---|
| Works with Claude Code | tested 2026-11-12: tools listed and ivreg called from a Claude Code session |
AI agents and MCP |
| Works with any stdio MCP client | tested 2026-11-12 against StatsPAI 1.43.0 from PyPI in a fresh environment: initialize, tools/list, tools/call |
same page |
| Claude Desktop, Cursor, Windsurf, VS Code, Codex | statspai mcp config generates their entries; not re-tested end to end on that date |
statspai mcp config --client <name> |
Performance¶
| Claim | Evidence | Re-check |
|---|---|---|
HDFE timings relative to fixest and pyfixest |
performance guide, with hardware, versions and load average recorded | python benchmarks/run_all.py |
No other speed claim is made. "Fastest" is not claimed.
Claims we do not make¶
- "The first" or "the only" agent-native econometrics library: not established, so not claimed.
- "Best" Python econometrics library: the library comparison explains where other packages are the better choice.
- "Replaces Stata / R" without qualification: StatsPAI offers a replacement path for many workflows, with the evidence grades above.
- User counts, ratings or endorsements.