Fixed effects and high-dimensional fixed effects in Python¶
Short answer. sp.feols("y ~ x | unit + year", data=df, cluster="unit")
absorbs any number of fixed effects, clusters standard errors, and reports
within R². It is the Python counterpart of Stata's reghdfe and R's
fixest::feols, and on StatsPAI's cross-language test suite its coefficients
and standard errors agree with both to within 1e-6 (sp.parity_status("feols")).
For classical one-way panels with random effects and a Hausman test, use
sp.panel.
When this applies¶
- What FE buys you: control for anything constant within a unit (or within a year, an industry-year, …). Identification comes from variation within those groups.
- Data: a long panel or any data with group identifiers. Singletons are
dropped as in
reghdfe; collinear regressors are dropped and reported. - Careful with staggered treatment dummies: a TWFE coefficient on a treatment that switches on at different dates is not an average treatment effect when effects vary. See difference-in-differences.
Two-way fixed effects with clustered SEs¶
import statspai as sp
df = sp.load_dataset("castle_doctrine") # 50 states x 11 years, real data
twfe = sp.feols("l_homicide ~ post | sid + year", data=df, cluster="sid")
wtd = sp.feols("l_homicide ~ post | sid + year", data=df, weights="popwt",
vcov={"CRV1": "sid"})
print(sp.regtable([twfe, wtd], model_labels=["TWFE", "population-weighted"]))
TWFE population-weighted
post 0.069 0.076**
(0.056) (0.033)
Sid FE Yes Yes
Year FE Yes Yes
Cluster SE sid sid
N 550 550
The same two specifications in Stata 18 (reghdfe l_homicide post, absorb(sid
year) vce(cluster sid), with and without [aw=popwt]) give identical
coefficients and standard errors (0.069398 / 0.055860 and 0.075533 /
0.033194); see the castle-doctrine replication.
While fitting the weighted model StatsPAI also warns that the weights concentrate on a few states (weight-effective number of clusters 22.6 of 50), so cluster-robust intervals may be too short. Warnings like this are written to the result's diagnostics rather than swallowed.
More than two dimensions, and other options¶
sp.feols("y ~ x | firm + year + industry", data=df, cluster="firm") # three sets of FE
sp.feols("y ~ x | firm + year", data=df, vcov={"CRV1": "firm + year"}) # two-way clustering
sp.fepois("count ~ x | firm + year", df) # Poisson (PPML) with FE
sp.panel(df, "y ~ x", entity="firm", time="year", method="re") # random effects
Interacted fixed effects such as industry^year currently go through the
optional pyfixest backend (pip install "statspai[fixest]").
The optional Rust kernels (pip install statspai-hdfe, installed
automatically where a wheel exists) speed up demeaning on large panels;
sp.fast.backend_info() says whether they are in use. Timings against
fixest and pyfixest are in the performance guide.
Stata and R equivalents¶
| Task | Stata | R | StatsPAI |
|---|---|---|---|
| Absorb FE, cluster | reghdfe y x, absorb(id year) vce(cluster id) |
feols(y ~ x \| id + year, cluster = ~id) |
sp.feols("y ~ x \| id + year", data=df, cluster="id") |
| Poisson with FE | ppmlhdfe |
fepois() |
sp.fepois() / sp.ppmlhdfe() |
| FE vs RE | xtreg, fe / xtreg, re + hausman |
plm + phtest |
sp.panel() |
| Translate a command | — | — | sp.from_stata("reghdfe y x, absorb(id year)") |
Other Python options¶
pyfixest mirrors fixest closely and is a strong choice if HDFE regression
is all you need. linearmodels offers PanelOLS (up to two effects),
AbsorbingLS and random effects. See the
library comparison.
Tested with StatsPAI 1.43.1 on 2026-11-12. The code and numbers on this page
are run by tests/test_docs_task_pages.py.
Report a problem.