You perform one-way fixed-effects estimation by specifying the FIXONE option in the MODEL statement. The error structure for the one-way fixed-effects model is
where the are nonrandom parameters that are restricted to sum to 0, and the
are iid with zero mean and variance
.
The fixed-effects model can be estimated by ordinary least squares (OLS), treating the as coefficients on dummy variables that identify the cross sections. However, when N is large, you might want to estimate only
and not
.
Let . The matrix
represents the within transformation, the conversion of the raw data to deviations from a cross section’s mean. Let
and
. The within estimator of
is
The previous estimation does not involve the intercept term because is the same whether or not the intercept
is included in the model.
Standard errors, t statistics, and fit statistics such as mean square error (MSE) are all equivalent to those obtained from OLS regression of on
. The only exception is the error degrees of freedom, which equals
to account for the tacit estimation of the N fixed effects.
Each fixed effect is estimated as
where and
are cross-sectional means.
The fixed-effects model is parameterized so that the intercept is the fixed effect for the last cross section. That is,
Fixed effects are by default not displayed as part of the regression, but you can obtain them by specifying the PRINTFIXED option in the MODEL statement. In models that have an intercept, the printed fixed effects are the deviations . To display the untransformed fixed effects, specify both the NOINT and PRINTFIXED options.
Variance estimates of ,
, and
are obtained by the delta method.
The FIXONETIME option works similarly, except that the data are grouped by time period instead of by cross section.