def causalsens()

in causalml/metrics/sensitivity.py [0:0]

• 33 lines of code
• 3 McCabe index (conditional complexity)

    def causalsens(self):
        alpha_range = self.alpha_range
        confound = self.confound
        df = self.df
        X = df[self.inference_features].values
        p = df[self.p_col].values
        treatment = df[self.treatment_col].values
        y = df[self.outcome_col].values

        preds = self.get_prediction(X, p, treatment, y)

        sens_df = pd.DataFrame()

        sens = []
        for a in alpha_range:
            adj = confound(a, p, treatment)
            preds_adj = y - adj
            s_preds = self.get_prediction(X, p, treatment, preds_adj)
            ate, ate_lb, ate_ub = self.get_ate_ci(X, p, treatment, preds_adj)

            s_preds_residul = preds_adj - s_preds
            rsqs = a**2 * np.var(treatment) / np.var(s_preds_residul)

            sens.append([a, rsqs, ate, ate_lb, ate_ub])

        sens_df = pd.DataFrame(
            sens, columns=["alpha", "rsqs", "New ATE", "New ATE LB", "New ATE UB"]
        )

        rss = np.sum(np.square(y - preds))
        partial_rsqs = []
        for feature in self.sensitivity_features:
            df_new = df.copy()
            X_new = df_new[self.inference_features].drop(feature, axis=1).copy()
            y_new_preds = self.get_prediction(X_new, p, treatment, y)
            rss_new = np.sum(np.square(y - y_new_preds))
            partial_rsqs.append(((rss_new - rss) / rss))

        partial_rsqs_df = pd.DataFrame([self.sensitivity_features, partial_rsqs]).T
        partial_rsqs_df.columns = ["feature", "partial_rsqs"]

        return sens_df, partial_rsqs_df