Skip to content

PyFixest 0.70.0

Latest

Choose a tag to compare

@s3alfisc s3alfisc released this 08 Nov 19:36
· 2 commits to master since this release
829940f

PyFixest 0.70.0

TL;DR

  • Varying slopes. | f1[X2] estimates a separate slope on X2 for each
    level of f1, in feols(), fepois() and feglm().
  • Formulas are parsed by formulaic. The fitted model spec is reused for
    predict() and fixef(), so new data is encoded exactly like the estimation
    data, and predict(newdata=...) now works with i(), poly() and
    interacted fixed effects. Along with it come new formula spellings:
    • Fixed-effect interactions are written f1:f2.
  • New IV syntax. Instrumental variables are written in brackets,
    Y ~ X1 + [X2 ~ Z1], which also accepts interacted endogenous terms such as
    X2:Z2. The old Y ~ X1 | X2 ~ Z1 syntax still works, but it emits a
    deprecation warning and will be removed in a future release.
  • New default demeaner. LSMR from the within package with an adaptive
    preconditioner.
  • Typed results. Fitted models expose their state as documented components
    such as fit.coeftable, fit.variance_covariance and fit.model.
  • Also new: anytime-valid inference (SAVI), Wor‹◊ tables from etable(),
    weights and offsets in feglm(), clustering on three or more variables, and
    Python 3.15.
  • Many correctness fixes, several of which change results: ritest()
    p-values, i() reference levels, weighted RMSE, the IV first stage,
    decompose() and quantreg's cfm2 process. See Bug fixes.
  • pyfixest needs Python 3.11+ and newer formulaic and
    pandas; model names in multiple estimation changed; fixef() returns a
    DataFrame; most private attributes such as fit._se are gone; f1^f2 and
    Y ~ X1 | X2 ~ Z1 are deprecated. See Breaking changes
    and Migrating Fitted-Model Attributes.