Panel data¶

statspai.panel — fixed/random effects, high-dimensional fixed effects (reghdfe-style), interactive fixed effects (Bai 2009), panel FGLS, panel limited-dependent models, panel unit-root tests, and panel utilities.

The exhaustive auto-generated listing is under Full API reference → panel; this page is the guided tour.

The dispatcher — `sp.panel`¶

sp.panel is a single entry point that dispatches to the estimator named by method=:

import statspai as sp

# Fixed effects (within estimator), clustered SEs:
r = sp.panel(df, "y ~ x1 + x2", entity="firm", time="year",
             method="fe", cluster="firm")
print(r.summary())

r = sp.panel(df, "y ~ x1 + x2", entity="firm", time="year", method="re")   # random effects
r = sp.panel(df, "y ~ x1 + x2", entity="firm", time="year", method="pooled")

`method=`	Estimator
`"fe"`	One- or two-way fixed effects (within)
`"re"`	Random effects (GLS)
`"pooled"`	Pooled OLS
`"between"`	Between estimator
`"fd"`	First differences

High-dimensional fixed effects — `sp.hdfe_ols` / `sp.absorb_ols`¶

For many fixed-effect dimensions (firm × year × worker …), use the reghdfe-style absorbing estimators, which demean iteratively instead of building dummies:

# Absorb firm and year FE; cluster by firm.
r = sp.hdfe_ols("y ~ x1 + x2 | firm + year", data=df, cluster="firm")

# Wild cluster bootstrap for few-cluster inference:
r = sp.hdfe_ols("y ~ x1 | firm", data=df, cluster="firm",
                wild=True, wild_n_boot=999, wild_seed=0)

sp.Absorber exposes the reusable demeaning operator, and sp.demean returns the within-transformed array plus the singleton-keep mask, for callers that want to drive the HDFE machinery directly. Singletons are dropped by default (drop_singletons=True).

Performance

The HDFE path is backed by a Rust demeaning kernel when available and transparently falls back to NumPy / pyfixest otherwise — see GPU acceleration and the sp.fast.* namespace.

Interactive fixed effects — `sp.interactive_fe`¶

Bai (2009) interactive fixed effects allow a factor structure (unobserved common shocks with heterogeneous loadings) instead of additive two-way FE:

r = sp.interactive_fe(df, y="y", x=["x1", "x2"], id="firm", time="year",
                      n_factors=2)

Other panel models¶

# Feasible GLS for panels with heteroskedastic / correlated errors:
r = sp.panel_fgls(df, "y ~ x1 + x2", entity="firm", time="year")

# Panel limited-dependent models:
r = sp.panel_logit(df, "d ~ x1 + x2", entity="firm", time="year")
r = sp.panel_probit(df, "d ~ x1 + x2", entity="firm", time="year")

# Panel unit-root tests (Im–Pesaran–Shin default; also LLC, Fisher):
ur = sp.panel_unitroot(df, variable="gdp", id="country", time="year", test="ips")

Comparing specifications — `sp.panel_compare`¶

Estimate the same model under several methods and line them up (a quick way to see how sensitive a coefficient is to the FE/RE choice):

tbl = sp.panel_compare(df, "y ~ x1 + x2", entity="firm", time="year",
                       methods=["fe", "re", "pooled"], cluster="firm")
tbl   # one column per method

A Hausman-style contrast between FE and RE is the classic use: if the FE and RE coefficients diverge materially, the random-effects orthogonality assumption is suspect and FE is the safer choice.

Utilities¶

sp.balance_panel(df, entity, time) — keep only units observed in every period (balanced panel).
sp.demean(...) / sp.Absorber(...) — low-level within-transform building blocks shared by the HDFE estimators.

Panel data¶

The dispatcher — sp.panel¶

High-dimensional fixed effects — sp.hdfe_ols / sp.absorb_ols¶

Interactive fixed effects — sp.interactive_fe¶