domains / families / research / 10-phack-polyglot

10-phack-polyglot

separate

Run the instrumented specification search in the user's own statistical language — Stata (reghdfe, ivreghdfe, rdrobust, did2s), R (fixest, rdrobust, did2s), Python (statsmodels, linearmodels) or StatsPAI — and bring the results back into the audit, null calibration and honest report. Exports the enumerated grid as a language-neutral specs table plus a generated runner, ingests the runner's ledger, replays the null draws in that language, and reports cross-language parity. Use when an analysis or an agent evaluation must happen in Stata, R or StatsPAI rather than the Python engine, when a p-hacked result was produced in one of those languages and needs auditing on its own footing, when reading Stata / R / StatsPAI code for search and disclosure signals, or when checking that the engine's numbers agree with a reference implementation.

When to use

• Running specification search in Stata, R, Python, or StatsPAI
• Auditing p-hacked results in non-Python languages
• Reading Stata, R, or StatsPAI code for search signals
• Checking engine numbers against reference implementations
• Exporting specification grids as language-neutral tables
• Replaying null draws in target languages
• Reporting cross-language parity

When not to use

• The runners reproduce the engine's *grid semantics* (which rows, which
• Percentile-based rules (winsorising, IQR and percentile trims) use each
• Null replay in Stata re-runs the full grid B times inside Stata; budget
• `rdrobust`'s `bwselect` is never used: bandwidths are resolved in Python

Skill page

Source