LinearRegression

Guide & worked example · Source

LinearRegression(solver='torch', device=None)

Least squares with optional additive fixed effects.

X must include a constant if desired. weights are strictly positive analytic/WLS weights, not frequency counts. classical uses weighted SSE divided by residual degrees of freedom; HC0/HC1 use the weighted sandwich. HC1 counts the exact rank of the absorbed effects by default.

Absorbed columns (including an intercept) retain zero coefficient/SE slots. predict(X) returns the slope component only, not FE level predictions. fitted_values_ includes the fitted effects for the training observations.

Attributes

Name	Description
device
params
solver

Methods

Name	Description
fit	Fit; optionally override absorbed rank for large multi-way FE designs.
predict	Slope-only predictions; additive FE contributions are not included.
summary	Print a summary of the model results.
to	Move all model parameters to specified device.

fit

LinearRegression.fit(
    X,
    y,
    se=None,
    fe=None,
    weights=None,
    *,
    tol=1e-10,
    maxiter=100000,
    df_absorbed=None,
)

Fit; optionally override absorbed rank for large multi-way FE designs.

Rank-deficient non-absorbed columns raise instead of returning arbitrary coefficients. Failed absorption raises rather than returning a partial fit.

predict

LinearRegression.predict(X)

Slope-only predictions; additive FE contributions are not included.

summary

LinearRegression.summary()

Print a summary of the model results.

to

LinearRegression.to(device)

Move all model parameters to specified device.

Parameters

Name	Type	Description	Default
device	torch.device or str	Target device (‘cuda’, ‘cpu’, etc.)	required

Returns

Name	Type	Description
	BaseEstimator	Self for method chaining