Setar multiple inheritance

linear_model.py

diff --git a/statsmodels/regression/linear_model.py b/statsmodels/regression/linear_model.py
index e4be23e..d7cf538 100644
--- a/statsmodels/regression/linear_model.py
+++ b/statsmodels/regression/linear_model.py
@@ -90,7 +90,7 @@ class RegressionModel(base.LikelihoodModel):
         self.df_resid = self.nobs - self.rank
         self.df_model = float(rank(self.exog) - self.k_constant)
 
-    def fit(self, method="pinv", **kwargs):
+    def _fit(self, method="pinv", **kwargs):
         """
         Full fit of the model.
 
@@ -143,6 +143,38 @@ class RegressionModel(base.LikelihoodModel):
             beta = np.linalg.solve(R, effects)
 
             # no upper triangular solve routine in numpy/scipy?
+        return beta
+
+    def fit(self, method="pinv", **kwargs):
+        """
+        Full fit of the model.
+
+        The results include an estimate of covariance matrix, (whitened)
+        residuals and an estimate of scale.
+
+        Parameters
+        ----------
+        method : str
+            Can be "pinv", "qr".  "pinv" uses the Moore-Penrose pseudoinverse
+            to solve the least squares problem. "qr" uses the QR
+            factorization.
+
+        Returns
+        -------
+        A RegressionResults class instance.
+
+        See Also
+        ---------
+        regression.RegressionResults
+
+        Notes
+        -----
+        The fit method uses the pseudoinverse of the design/exogenous variables
+        to solve the least squares minimization.
+        """

setar_model.py

class SETAR(OLS, tsbase.TimeSeriesModel):
    def __init__(self, endog, order, ar_order,
                 delay=None, thresholds=None, trend='c',
                 min_regime_frac=0.1, max_delay=None, threshold_grid_size=100,
                 dates=None, freq=None, missing='none'):

        if delay is not None and delay < 1 or delay > ar_order:
            raise ValueError('Delay parameter must be greater than zero'
                             ' and less than ar_order. Got %d.' % delay)

        # Unsure of statistical properties if length of sample changes when
        # estimating hyperparameters, which happens if delay can be greater
        # than ar_order, so that the number of initial observations changes
        if delay is None and max_delay > ar_order:
            raise ValueError('Maximum delay for grid search must not be '
                             ' greater than the autoregressive order.')

        if delay is None and thresholds is not None:
            raise ValueError('Thresholds cannot be specified without delay'
                             ' parameter.')

        if thresholds is not None and not len(thresholds) + 1 == order:
            raise ValueError('Number of thresholds must match'
                             ' the order of the SETAR model')

        # "Immutable" properties
        self.nobs_initial = ar_order
        self.nobs = endog.shape[0] - ar_order

        self.order = order
        self.ar_order = ar_order
        self.k_trend = int(trend == 'c')
        self.min_regime_frac = min_regime_frac
        self.min_regime_num = np.ceil(min_regime_frac * self.nobs)
        self.max_delay = max_delay if max_delay is not None else ar_order
        self.threshold_grid_size = threshold_grid_size

        # "Flexible" properties
        self.delay = delay
        self.thresholds = thresholds
        if self.thresholds:
            self.thresholds = np.sort(self.thresholds)

        # Estimation properties
        self.nobs_regimes = None
        self.objectives = {}
        self.ar1_resids = None

        # Make a copy of original datasets
        orig_endog = endog
        orig_exog = lagmat(orig_endog, ar_order)

        # Trends
        if self.k_trend:
            orig_exog = add_constant(orig_exog)

        # Create datasets / complete initialization
        endog = orig_endog[self.nobs_initial:]
        exog = orig_exog[self.nobs_initial:]
        super(SETAR, self).__init__(endog, exog,
                                    hasconst=self.k_trend, missing=missing)

        # Overwrite originals
        self.data.orig_endog = orig_endog
        self.data.orig_exog = orig_exog
    def fit(self):
        """
        Fits SETAR() model using arranged autoregression.
    
        Returns
        -------
        statsmodels.tsa.arima_model.SETARResults class
    
        See also
        --------
        statsmodels.regression.linear_model.OLS : this estimates each regime
        SETARResults : results class returned by fit
    
        """
    
        if self.delay is None or self.thresholds is None:
            self.delay, self.thresholds = self.select_hyperparameters()
    
        self.endog, self.exog, self.nobs_regimes = self.build_datasets(
            self.delay, self.thresholds
        )
        self.initialize()
    
        beta = self._fit()
        lfit = SETARResults(
            self, beta, normalized_cov_params=self.normalized_cov_params
        )
    
        return lfit
    
class SETARResults(OLSResults, tsbase.TimeSeriesModelResults):
    def __init__(self, model, params, normalized_cov_params=None, scale=1.):
        super(SETARResults, self).__init__(model, params, normalized_cov_params,
                scale)

tsa_model.py

diff --git a/statsmodels/tsa/base/tsa_model.py b/statsmodels/tsa/base/tsa_model.py
index 5f8c6b1..a4e964f 100644
--- a/statsmodels/tsa/base/tsa_model.py
+++ b/statsmodels/tsa/base/tsa_model.py
@@ -36,8 +36,8 @@ class TimeSeriesModel(base.LikelihoodModel):
                           "extra_params" : _missing_param_doc,
                           "extra_sections" : ""}
 
-    def __init__(self, endog, exog=None, dates=None, freq=None, missing='none'):
-        super(TimeSeriesModel, self).__init__(endog, exog, missing=missing)
+    def __init__(self, endog, exog=None, dates=None, freq=None, missing='none', **kwargs):
+        super(TimeSeriesModel, self).__init__(endog, exog, missing=missing, **kwargs)
         self._init_dates(dates, freq)
 
     def _init_dates(self, dates, freq):

For some reason it is not letting me comment on the Gist. However, instead of changing RegressionModel, why not just write a wrapper function around that fit. For example, you can have a _fit function as part of setar that is a wrapper around the OLS.fit(). Then you can call it withing the setar.fit().