experiments.py 文件源码

def weak_instruments(self, n_sims=20):

        np.random.seed(1692)

        model = feedforward.FeedForwardModel(19, 1, dense_size=60, n_dense_layers=2)

        treatment_effects = []
        ols_betas, ols_ses = [], []
        old_corrs, new_corrs = [], []
        for _ in xrange(n_sims):
            df = self.treatment_gen.simulate_data(False)

            X = np.hstack((self.x, df['new_treat'].values[:, None]))
            Z = np.hstack((self.x, df['instrument'].values[:, None]))

            ols_beta, ols_se = self.fit_ols(df['treatment_effect'], X)
            ols_betas.append(ols_beta)
            ols_ses.append(ols_se)

            old_corr = df[['instrument', 'new_treat']].corr().values[0, 1]
            new_instrument, new_corr = model.fit_instruments(X, Z, df['treatment_effect'].values, batchsize=128)
            new_corrs.append(new_corr)
            old_corrs.append(old_corr)

            Z2 = Z.copy()
            Z2[:, -1] = new_instrument[:, 0]

            iv = IV2SLS(df['treatment_effect'].values.flatten(), add_constant(X), add_constant(Z2))

            model.reset_params()

        if new_corr:
            logger.info("Old corr: %.2f, New corr: %.2f", np.mean(old_corrs), np.mean(new_corrs))
        logger.info("Treatment effect (OLS): %.3f (%.4f)", np.mean(ols_betas), np.mean(ols_ses))
        logger.info("Treatment effect: %.3f (%.4f)", np.mean(treatment_effects), np.std(treatment_effects))