remove var_scaling_factors from aggregation

Author: SvenKlaassen

doubleml/double_ml.py

@@ -537,7 +537,7 @@ def fit(self, n_jobs_cv=None, store_predictions=True, external_predictions=None,
                 self._fit_sensitivity_elements(nuisance_predictions)
 
         # aggregated parameter estimates and standard errors from repeated cross-fitting
-        self.coef, self.se = _aggregate_coefs_and_ses(self._all_coef, self._all_se, self._var_scaling_factors)
+        self.coef, self.se = _aggregate_coefs_and_ses(self._all_coef, self._all_se)
 
         # validate sensitivity elements (e.g., re-estimate nu2 if negative)
         self._validate_sensitivity_elements()
@@ -1392,7 +1392,7 @@ def _est_causal_pars_and_se(self):
                 self._all_se[self._i_treat, self._i_rep], self._var_scaling_factors[self._i_treat] = self._se_causal_pars()
 
         # aggregated parameter estimates and standard errors from repeated cross-fitting
-        self.coef, self.se = _aggregate_coefs_and_ses(self._all_coef, self._all_se, self._var_scaling_factors)
+        self.coef, self.se = _aggregate_coefs_and_ses(self._all_coef, self._all_se)
 
     # Score estimation and elements
     @abstractmethod

doubleml/double_ml_framework.py

@@ -310,7 +310,7 @@ def __add__(self, other):
             # compute standard errors (Uses factor 1/n for scaling!)
             sigma2_hat = np.divide(np.mean(np.square(scaled_psi), axis=0), var_scaling_factors.reshape(-1, 1))
             all_ses = np.sqrt(sigma2_hat)
-            thetas, ses = _aggregate_coefs_and_ses(all_thetas, all_ses, var_scaling_factors)
+            thetas, ses = _aggregate_coefs_and_ses(all_thetas, all_ses)
 
             doubleml_dict = {
                 "thetas": thetas,
@@ -358,7 +358,7 @@ def __sub__(self, other):
             # compute standard errors
             sigma2_hat = np.divide(np.mean(np.square(scaled_psi), axis=0), var_scaling_factors.reshape(-1, 1))
             all_ses = np.sqrt(sigma2_hat)
-            thetas, ses = _aggregate_coefs_and_ses(all_thetas, all_ses, var_scaling_factors)
+            thetas, ses = _aggregate_coefs_and_ses(all_thetas, all_ses)
 
             doubleml_dict = {
                 "thetas": thetas,
@@ -507,8 +507,8 @@ def _calc_sensitivity_analysis(self, cf_y, cf_d, rho, level):
                 all_sigma_upper[i_theta, i_rep] = np.sqrt(sigma2_upper_hat)
 
         # aggregate coefs and ses over n_rep
-        theta_lower, sigma_lower = _aggregate_coefs_and_ses(all_theta_lower, all_sigma_lower, self._var_scaling_factors)
-        theta_upper, sigma_upper = _aggregate_coefs_and_ses(all_theta_upper, all_sigma_upper, self._var_scaling_factors)
+        theta_lower, sigma_lower = _aggregate_coefs_and_ses(all_theta_lower, all_sigma_lower)
+        theta_upper, sigma_upper = _aggregate_coefs_and_ses(all_theta_upper, all_sigma_upper)
 
         # per repetition confidence intervals
         quant = norm.ppf(level)

doubleml/tests/_utils.py

@@ -101,7 +101,6 @@ def generate_dml_dict(psi_a, psi_b):
     thetas, ses = _aggregate_coefs_and_ses(
         all_coefs=all_thetas,
         all_ses=all_ses,
-        var_scaling_factors=var_scaling_factors,
     )
     scaled_psi = psi_b / np.mean(psi_a, axis=0)
 

doubleml/tests/_utils_doubleml_sensitivity_manual.py

@@ -31,8 +31,8 @@ def doubleml_sensitivity_manual(sensitivity_elements, all_coefs, psi, psi_deriv,
     all_sigma_lower = np.transpose(np.sqrt(np.divide(np.mean(np.square(psi_lower), axis=0), var_scaling_factor)))
     all_sigma_upper = np.transpose(np.sqrt(np.divide(np.mean(np.square(psi_upper), axis=0), var_scaling_factor)))
 
-    theta_lower, sigma_lower = _aggregate_coefs_and_ses(all_theta_lower, all_sigma_lower, var_scaling_factor)
-    theta_upper, sigma_upper = _aggregate_coefs_and_ses(all_theta_upper, all_sigma_upper, var_scaling_factor)
+    theta_lower, sigma_lower = _aggregate_coefs_and_ses(all_theta_lower, all_sigma_lower)
+    theta_upper, sigma_upper = _aggregate_coefs_and_ses(all_theta_upper, all_sigma_upper)
 
     quant = norm.ppf(level)
 

doubleml/utils/_estimation.py

@@ -239,7 +239,7 @@ def abs_ipw_score(theta):
     return ipw_est
 
 
-def _aggregate_coefs_and_ses(all_coefs, all_ses, var_scaling_factors):
+def _aggregate_coefs_and_ses(all_coefs, all_ses):
     # already expects equally scaled variances over all repetitions
     # aggregation is done over dimension 1, such that the coefs and ses have to be of shape (n_coefs, n_rep)
     coefs = np.median(all_coefs, 1)

doubleml/utils/tests/test_var_est_and_aggregation.py

@@ -54,14 +54,12 @@ def test_var_est_and_aggr_fixture(n_rep, n_coefs, level):
     theta, se = _aggregate_coefs_and_ses(
         all_coefs=all_thetas,
         all_ses=all_ses,
-        var_scaling_factors=all_var_scaling_factors[:, 0],
     )
 
     # with n_rep
     theta_2, se_2 = _aggregate_coefs_and_ses(
         all_coefs=all_thetas,
         all_ses=all_ses,
-        var_scaling_factors=all_var_scaling_factors,
     )
 
     result_dict = {