model.py 文件源码

def _slow_path(self):
        """Frisch-Waugh-Lovell implementation, works for all scenarios"""
        has_effect = self.entity_effects or self.time_effects or self.other_effects
        w = self.weights.values2d
        root_w = np.sqrt(w)

        y = root_w * self.dependent.values2d
        x = root_w * self.exog.values2d
        if not has_effect:
            ybar = root_w @ lstsq(root_w, y)[0]
            return y, x, ybar, 0, 0

        drop_first = self._constant
        d = []
        if self.entity_effects:
            d.append(self.dependent.dummies('entity', drop_first=drop_first).values)
            drop_first = True
        if self.time_effects:
            d.append(self.dependent.dummies('time', drop_first=drop_first).values)
            drop_first = True
        if self.other_effects:
            oe = self._other_effect_cats.dataframe
            for c in oe:
                dummies = pd.get_dummies(oe[c], drop_first=drop_first).astype(np.float64)
                d.append(dummies.values)
                drop_first = True

        d = np.column_stack(d)
        wd = root_w * d
        if self.has_constant:
            wd -= root_w * (w.T @ d / w.sum())
            z = np.ones_like(root_w)
            d -= z * (z.T @ d / z.sum())

        x_mean = np.linalg.lstsq(wd, x)[0]
        y_mean = np.linalg.lstsq(wd, y)[0]

        # Save fitted unweighted effects to use in eps calculation
        x_effects = d @ x_mean
        y_effects = d @ y_mean

        # Purge fitted, weighted values
        x = x - wd @ x_mean
        y = y - wd @ y_mean

        ybar = root_w @ lstsq(root_w, y)[0]
        return y, x, ybar, y_effects, x_effects