Sensitivity Analysis: Lalonde Dataset

Tools for Causality
Grenoble, Sept 25 - 29, 2023
Philipp Bach, Sven Klaassen

Authors
Affiliation

Introduction

We illustrate the use of sensitivity analysis on the famous Lalonde data set. There are basically two different samples, one experimental and one non-experimental subset. […]. The samples have been used for sensitivity analysis in Imbens (2003) and, more recently, in Veitch and Zaveri (2020).

A nice feature about the Lalonde data set is that makes it possible to compare the results from an experimental setting to an observational setting. Hence, we can compare the robustness of the results from an observational causal study to an experimental study, which is very unlikely affected by unobserved confounding.

In this notebook, we replicate the sensitivity analysis in Imbens (2003) and Veitch and Zaveri (2020). There a confounding scenario is considered that would introduce a bias of $1000.1

Data

We organize our analysis along the lines of Imbens (2003) and Veitch and Zaveri (2020) who consider two different sample specifications:

  1. Experimental Sample - treated and control units randomly selected from the original experimental sample
  2. Lalonde Restricted Sample - selected sample of treated (148/185) and control (242/2490) units with individual earnings in 1975 and 1975 below $5000.

The original data files are available from Veitch and Zaveri (2020).

The data contains nine covariates

  • married
  • age
  • indicators for black and Hispanic,
  • years of education
  • (real) earnings in 1974 and 1975,
  • and indicators for positive earnings in 1974 and 1975

and the outcome (earnings in 1978). The treatment variable is an indicator for participation in the job training program.

1. Experimental Sample

import pandas as pd

# TODO: Fix Data loading (from URL?)
lalonde_exp = pd.read_csv('../../datasets/imbens-cleaned/imbens1.csv')
lalonde_exp.shape
(445, 11)

2. Restricted Sample

# TODO: Fix Data loading (from URL?)
lalonde_restricted = pd.read_csv('../../datasets/imbens-cleaned/imbens4.csv')
lalonde_restricted.shape
(390, 11)

Brief Summary Statistics

We mainly consider past earnings as a benchmark variable. To see the difference across the experimental and non-experimental samples, we plot the distribution of past earnings for the treated and control units.

# Experimental sample
lalonde_exp.groupby('treatment')['RE75'].plot.hist(bins = 20, alpha = 0.8, density = True, legend = True)
treatment
0.0    Axes(0.125,0.11;0.775x0.77)
1.0    Axes(0.125,0.11;0.775x0.77)
Name: RE75, dtype: object

# Experimental sample
lalonde_restricted.groupby('treatment')['RE75'].plot.hist(bins = 20, alpha = 0.8, density = True, legend = True)
treatment
0.0    Axes(0.125,0.11;0.775x0.77)
1.0    Axes(0.125,0.11;0.775x0.77)
Name: RE75, dtype: object

Causal Estimation: Partially Linear Regression

We will first estimate the Average Treatment Effect (ATE) of the training on earnings using a partially linear regression model.

import doubleml as dml

Data Backend

x_cols = lalonde_exp.columns.drop(['RE78', 'treatment']).to_list()

dml_data_exp = dml.DoubleMLData(lalonde_exp, y_col='RE78',
                          d_cols='treatment',
                          x_cols=x_cols)
dml_data_restricted = dml.DoubleMLData(lalonde_restricted, y_col='RE78',
                          d_cols='treatment',
                          x_cols=x_cols)

Initialize Learners and Causal Models

We use the same specification for the learners as in Veitch and Zaveri (2020).

from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier

N_EST = 100
MAX_DEPTH = 5

ml_m = RandomForestClassifier(
            random_state=42, n_estimators=N_EST, max_depth=MAX_DEPTH)

ml_g = RandomForestRegressor(
            random_state=42, n_estimators=N_EST, max_depth=MAX_DEPTH)
import numpy as np

np.random.seed(123)
dml_plr_exp = dml.DoubleMLPLR(dml_data_exp, ml_g, ml_m, n_folds = 10)
dml_plr_exp.fit()
<doubleml.double_ml_plr.DoubleMLPLR at 0x21ccb347fd0>
np.random.seed(123)
dml_plr_restricted = dml.DoubleMLPLR(dml_data_restricted, ml_g, ml_m, n_folds = 10)
dml_plr_restricted.fit()
<doubleml.double_ml_plr.DoubleMLPLR at 0x21ccb53d4d0>

Let us summarize the input in a table.

results_plr_df = pd.concat([dml_plr_exp.summary, dml_plr_restricted.summary], keys = ["Experimental", "Restricted"])
results_plr_df.round(3)
coef std err t P>|t| 2.5 % 97.5 %
Experimental treatment 1726.899 677.070 2.551 0.011 399.867 3053.932
Restricted treatment 2283.652 1029.706 2.218 0.027 265.466 4301.838
import matplotlib.pyplot as plt
import seaborn as sns
sns.set()
colors = sns.color_palette()
plt.rcParams['figure.figsize'] = 10., 7.5
sns.set(font_scale=1.5)
sns.set_style('whitegrid', {'axes.spines.top': False,
                            'axes.spines.bottom': False,
                            'axes.spines.left': False,
                            'axes.spines.right': False})

plt.figure(figsize=(10, 15))
errors = np.full((2, results_plr_df.shape[0]), np.nan)
errors[0, :] = results_plr_df['coef'] - results_plr_df['2.5 %']
errors[1, :] = results_plr_df['97.5 %'] - results_plr_df['coef']

plt.errorbar(['Experimental', 'Restricted'], results_plr_df.coef, fmt='o', yerr=errors)
plt.title('Lalonde Data: PLR Estimates')
plt.ylabel('Coefficients and 95%-CI')

_ = plt.xlabel('Data Set')

Sensitivity Analysis: Partially Linear Regression

Now let’s perform some sensitivity analysis for the different samples. We start with the experimental sample.

General setting: No particular \(H_0\) provided.

dml_plr_exp.sensitivity_analysis()
print(dml_plr_exp.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower       theta  theta upper     CI upper
treatment  169.541845  1309.011131  1726.89944  2144.787748  3235.412501

------------------ Robustness Values ------------------
           H_0     RV (%)   RVa (%)
treatment  0.0  11.820295  4.117965
dml_plr_restricted.sensitivity_analysis()
print(dml_plr_restricted.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower       theta  theta upper     CI upper
treatment -223.933549   1487.96296  2283.65209  3079.341221  4770.699752

------------------ Robustness Values ------------------
           H_0    RV (%)   RVa (%)
treatment  0.0  8.368444  2.189394

Benchmarking

Let us replicate the benchmark analysis from Imbens (2003) and Veitch and Zaveri (2020). In their analyses, they consider a confounding scenario that would introduce a bias of $1000. One possibility would be to consider a null hypothesis that corresponds to a value \(\hat{\theta}_0 - 1000\). However, as we are going to visualize our results in a contour plot, this enables us to consider various bias levels at once.

We consider the following benchmark scenarios

  • leaving out each of the covariates
  • leaving out the latest income lag variables together (\(RE75\) and \(pos75\))
  • as well as jointly all variables on preprogram earnings (\(RE74\), \(RE75\), \(pos74\) and \(pos75\))
benchmark_sets_single = [[var] for var in x_cols]
benchmark_sets_earnings = [['RE75', 'pos75'],
                          ['RE74', 'RE75', 'pos74', 'pos75']]
benchmark_scenarios_single = [var for var in x_cols]
benchmark_scenarios_earnings = ["recent earnings", 'preprogram earnings']
benchmark_sets = benchmark_sets_single + benchmark_sets_earnings
benchmark_scenarios = benchmark_scenarios_single + benchmark_scenarios_earnings
df_benchmark_exp = pd.DataFrame({'scenario': benchmark_scenarios,
                            'benchmarksets': benchmark_sets,
                            'cf_y': np.nan, 'cf_d': np.nan, 'rho': np.nan,
                            'delta_theta': np.nan})
df_benchmark_restricted = df_benchmark_exp.copy()

This can take a while as we have to estimate a new PLR for each benchmark scenario.

# Experimental data
for this_scenario in range(len(benchmark_scenarios)):
  bench_set = benchmark_sets[this_scenario]
  bench_results = dml_plr_exp.sensitivity_benchmark(bench_set)
  df_benchmark_exp.loc[this_scenario, 'cf_y'] = bench_results['cf_y'][0]
  df_benchmark_exp.loc[this_scenario, 'cf_d'] = bench_results['cf_d'][0]
  df_benchmark_exp.loc[this_scenario, 'rho'] = bench_results['rho'][0]
  df_benchmark_exp.loc[this_scenario, 'delta_theta'] = bench_results['delta_theta'][0]

df_benchmark_exp.round(3)
C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:5: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\2384365899.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`
scenario benchmarksets cf_y cf_d rho delta_theta
0 age [age] 0.000 0.000 -1.000 -61.852
1 education [education] 0.020 0.014 0.539 123.734
2 black [black] 0.017 0.000 -1.000 -64.900
3 hispanic [hispanic] 0.000 0.000 -1.000 -14.380
4 married [married] 0.000 0.004 -1.000 -112.151
5 RE74 [RE74] 0.000 0.000 -1.000 -289.242
6 RE75 [RE75] 0.007 0.016 -1.000 -172.036
7 pos74 [pos74] 0.003 0.001 -1.000 -90.068
8 pos75 [pos75] 0.003 0.000 -1.000 -106.454
9 recent earnings [RE75, pos75] 0.011 0.015 -0.515 -91.123
10 preprogram earnings [RE74, RE75, pos74, pos75] 0.000 0.036 -1.000 -78.475
# Restricted sample
# Numerical instabilities
for this_scenario in range(len(benchmark_scenarios)):
  bench_set = benchmark_sets[this_scenario]
  bench_results = dml_plr_restricted.sensitivity_benchmark(bench_set)
  df_benchmark_restricted.loc[this_scenario, 'cf_y'] = bench_results['cf_y'][0]
  df_benchmark_restricted.loc[this_scenario, 'cf_d'] = bench_results['cf_d'][0]
  df_benchmark_restricted.loc[this_scenario, 'rho'] = bench_results['rho'][0]
  df_benchmark_restricted.loc[this_scenario, 'delta_theta'] = bench_results['delta_theta'][0]

df_benchmark_restricted.round(3)
C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\anaconda3\Lib\site-packages\doubleml\_utils_checks.py:204: UserWarning:

Propensity predictions from learner RandomForestClassifier(max_depth=5, random_state=42) for ml_m are close to zero or one (eps=1e-12).

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:6: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:7: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:8: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`

C:\Users\PuD\AppData\Local\Temp\ipykernel_26336\1472702386.py:9: FutureWarning:

Series.__getitem__ treating keys as positions is deprecated. In a future version, integer keys will always be treated as labels (consistent with DataFrame behavior). To access a value by position, use `ser.iloc[pos]`
scenario benchmarksets cf_y cf_d rho delta_theta
0 age [age] 0.000 0.187 1.000 789.143
1 education [education] 0.055 0.014 0.146 103.249
2 black [black] 0.006 0.191 0.523 440.470
3 hispanic [hispanic] 0.000 0.004 -1.000 -37.113
4 married [married] 0.000 0.110 1.000 117.655
5 RE74 [RE74] 0.000 0.011 -1.000 -37.487
6 RE75 [RE75] 0.010 0.019 -0.970 -340.159
7 pos74 [pos74] 0.003 0.030 0.068 16.033
8 pos75 [pos75] 0.000 0.000 -1.000 -124.626
9 recent earnings [RE75, pos75] 0.018 0.036 -0.733 -475.483
10 preprogram earnings [RE74, RE75, pos74, pos75] 0.011 0.204 -0.841 -942.301

Visualization: Contour Plot

Let us visualize the results in a contour plot. Note that the results shown in the contour plot are conservative in the sense that the parameter \(\rho\) is set to a value of 1. Using lower values of \(\rho\) would lead to less conservative results. However it is not possible to combine multiple values of \(\rho\) in the same contour plot as it operates as a scaling factor for the bias.

Experimental sample
benchmarks_dict = {'cf_d' : df_benchmark_exp['cf_d'].astype(float).round(3).tolist(),
                   'cf_y' : df_benchmark_exp['cf_y'].astype(float).round(3).tolist(),
                   'name' : benchmark_scenarios}

contour_exp = dml_plr_exp.sensitivity_plot(grid_bounds = (0.2, 0.2), benchmarks = benchmarks_dict)
contour_exp.show()

print(df_benchmark_exp.round(3))
               scenario               benchmarksets   cf_y   cf_d    rho  \
0                   age                       [age]  0.000  0.000 -1.000   
1             education                 [education]  0.020  0.014  0.539   
2                 black                     [black]  0.017  0.000 -1.000   
3              hispanic                  [hispanic]  0.000  0.000 -1.000   
4               married                   [married]  0.000  0.004 -1.000   
5                  RE74                      [RE74]  0.000  0.000 -1.000   
6                  RE75                      [RE75]  0.007  0.016 -1.000   
7                 pos74                     [pos74]  0.003  0.001 -1.000   
8                 pos75                     [pos75]  0.003  0.000 -1.000   
9       recent earnings               [RE75, pos75]  0.011  0.015 -0.515   
10  preprogram earnings  [RE74, RE75, pos74, pos75]  0.000  0.036 -1.000   

    delta_theta  
0       -61.852  
1       123.734  
2       -64.900  
3       -14.380  
4      -112.151  
5      -289.242  
6      -172.036  
7       -90.068  
8      -106.454  
9       -91.123  
10      -78.475  
Restricted sample
benchmarks_dict = {'cf_d' : df_benchmark_restricted['cf_d'].astype(float).round(3).tolist(),
                   'cf_y' : df_benchmark_restricted['cf_y'].astype(float).round(3).tolist(),
                   'name' : benchmark_scenarios}

contour_restricted = dml_plr_restricted.sensitivity_plot(grid_bounds = (0.2, 0.2), benchmarks = benchmarks_dict)
contour_restricted.show()

print(df_benchmark_restricted.round(3))
               scenario               benchmarksets   cf_y   cf_d    rho  \
0                   age                       [age]  0.000  0.187  1.000   
1             education                 [education]  0.055  0.014  0.146   
2                 black                     [black]  0.006  0.191  0.523   
3              hispanic                  [hispanic]  0.000  0.004 -1.000   
4               married                   [married]  0.000  0.110  1.000   
5                  RE74                      [RE74]  0.000  0.011 -1.000   
6                  RE75                      [RE75]  0.010  0.019 -0.970   
7                 pos74                     [pos74]  0.003  0.030  0.068   
8                 pos75                     [pos75]  0.000  0.000 -1.000   
9       recent earnings               [RE75, pos75]  0.018  0.036 -0.733   
10  preprogram earnings  [RE74, RE75, pos74, pos75]  0.011  0.204 -0.841   

    delta_theta  
0       789.143  
1       103.249  
2       440.470  
3       -37.113  
4       117.655  
5       -37.487  
6      -340.159  
7        16.033  
8      -124.626  
9      -475.483  
10     -942.301  

Fixed value for bias

Let us compare the robustness of the analysis to a fixed value for the bias (\(=1000\$\)) by specifying an alternative \(H_0\).

# Restricted sample (observed data)
bias_value = 1000
H0 = dml_plr_restricted.coef - bias_value

dml_plr_restricted.sensitivity_analysis(null_hypothesis = H0)

print(dml_plr_restricted.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower       theta  theta upper     CI upper
treatment -223.933549   1487.96296  2283.65209  3079.341221  4770.699752

------------------ Robustness Values ------------------
                  H_0    RV (%)   RVa (%)
treatment  1283.65209  3.755653  0.000348
benchmark_setting_restricted = {'cf_d' : [df_benchmark_restricted.iloc[10]['cf_d'].astype(float).round(3)],
                     'cf_y' : [df_benchmark_restricted.iloc[10]['cf_y'].astype(float).round(3)],
                     'name' : [benchmark_scenarios[10]]}

contour_restricted_h0 = dml_plr_restricted.sensitivity_plot(grid_bounds = (0.25, 0.25),
  benchmarks = benchmark_setting_restricted)

contour_restricted_h0.show()
print(df_benchmark_restricted.iloc[10])
scenario                preprogram earnings
benchmarksets    [RE74, RE75, pos74, pos75]
cf_y                               0.010842
cf_d                               0.204396
rho                               -0.840951
delta_theta                     -942.301403
Name: 10, dtype: object
# Experimental sample (observed data)
H0 = dml_plr_exp.coef - bias_value

dml_plr_exp.sensitivity_analysis(null_hypothesis = H0)

print(dml_plr_exp.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower       theta  theta upper     CI upper
treatment  169.541845  1309.011131  1726.89944  2144.787748  3235.412501

------------------ Robustness Values ------------------
                 H_0    RV (%)   RVa (%)
treatment  726.89944  7.028315  0.000544
benchmark_setting_exp = {'cf_d' : [df_benchmark_exp.iloc[10]['cf_d'].astype(float).round(3)],
                     'cf_y' : [df_benchmark_exp.iloc[10]['cf_y'].astype(float).round(3)],
                     'name' : [benchmark_scenarios[10]]}

contour_exp_h0 = dml_plr_exp.sensitivity_plot(grid_bounds = (0.25, 0.25),
  benchmarks = benchmark_setting_exp)

contour_exp_h0.show()
print(df_benchmark_exp.iloc[10])
scenario                preprogram earnings
benchmarksets    [RE74, RE75, pos74, pos75]
cf_y                                    0.0
cf_d                               0.035982
rho                                    -1.0
delta_theta                      -78.475004
Name: 10, dtype: object

Causal Estimation: Interactive Regression Model

Alternatively, we can use an entirely nonparametric model to estimate the ATE or the Average Treatment Effect on the Treated (ATTE).

np.random.seed(123)
dml_irm_ate_exp = dml.DoubleMLIRM(dml_data_exp, ml_g, ml_m, n_folds = 10, score = 'ATE', trimming_threshold = 0.025)
dml_irm_ate_exp.fit()

# TODO: Diagnostics - Prop Score!
# TODO: Compare results for different learners (LightGBM, Lasso)
# TODO: Use IRM for ATTE only!
<doubleml.double_ml_irm.DoubleMLIRM at 0x21ccb629050>
np.random.seed(123)
dml_irm_ate_restricted = dml.DoubleMLIRM(dml_data_restricted, ml_g, ml_m, n_folds = 10, score = 'ATE', trimming_threshold = 0.025)
dml_irm_ate_restricted.fit()
<doubleml.double_ml_irm.DoubleMLIRM at 0x21ccbcb57d0>
results_irm_ate_df = pd.concat([dml_irm_ate_exp.summary, dml_irm_ate_restricted.summary], keys = ["Experimental", "Restricted"])
results_irm_ate_df.round(3)
coef std err t P>|t| 2.5 % 97.5 %
Experimental treatment 1676.152 694.596 2.413 0.016 314.769 3037.535
Restricted treatment 4173.361 669.070 6.238 0.000 2862.007 5484.715
import matplotlib.pyplot as plt
import seaborn as sns
sns.set()
colors = sns.color_palette()
plt.rcParams['figure.figsize'] = 10., 7.5
sns.set(font_scale=1.5)
sns.set_style('whitegrid', {'axes.spines.top': False,
                            'axes.spines.bottom': False,
                            'axes.spines.left': False,
                            'axes.spines.right': False})

plt.figure(figsize=(10, 15))
errors = np.full((2, results_irm_ate_df.shape[0]), np.nan)
errors[0, :] = results_irm_ate_df['coef'] - results_irm_ate_df['2.5 %']
errors[1, :] = results_irm_ate_df['97.5 %'] - results_irm_ate_df['coef']

plt.errorbar(['Experimental', 'Restricted'], results_irm_ate_df.coef, fmt='o', yerr=errors)
plt.title('Lalonde Data: IRM Estimates (ATE)')
plt.ylabel('Coefficients and 95%-CI')

_ = plt.xlabel('Data Set')

np.random.seed(123)
dml_irm_atte_exp = dml.DoubleMLIRM(dml_data_exp, ml_g, ml_m, n_folds = 10, score = 'ATTE', trimming_threshold = 0.025)
dml_irm_atte_exp.fit()
<doubleml.double_ml_irm.DoubleMLIRM at 0x21ccbcb7050>
np.random.seed(123)
dml_irm_atte_restricted = dml.DoubleMLIRM(dml_data_restricted, ml_g, ml_m, n_folds = 10, score = 'ATTE', trimming_threshold = 0.025)
dml_irm_atte_restricted.fit()
<doubleml.double_ml_irm.DoubleMLIRM at 0x21ccbcefc10>
results_irm_atte_df = pd.concat([dml_irm_atte_exp.summary, dml_irm_atte_restricted.summary], keys = ["Experimental", "Restricted"])
results_irm_atte_df.round(3)
coef std err t P>|t| 2.5 % 97.5 %
Experimental treatment 1790.802 738.426 2.425 0.015 343.515 3238.090
Restricted treatment 2752.281 825.913 3.332 0.001 1133.522 4371.041
import matplotlib.pyplot as plt
import seaborn as sns
sns.set()
colors = sns.color_palette()
plt.rcParams['figure.figsize'] = 10., 7.5
sns.set(font_scale=1.5)
sns.set_style('whitegrid', {'axes.spines.top': False,
                            'axes.spines.bottom': False,
                            'axes.spines.left': False,
                            'axes.spines.right': False})

plt.figure(figsize=(10, 15))
errors = np.full((2, results_irm_atte_df.shape[0]), np.nan)
errors[0, :] = results_irm_atte_df['coef'] - results_irm_atte_df['2.5 %']
errors[1, :] = results_irm_atte_df['97.5 %'] - results_irm_atte_df['coef']

plt.errorbar(['Experimental', 'Restricted'], results_irm_atte_df.coef, fmt='o', yerr=errors)
plt.title('Lalonde Data: IRM Estimates (ATTE)')
plt.ylabel('Coefficients and 95%-CI')

_ = plt.xlabel('Data Set')

Sensitivity Analysis: Interactive Regression Model

# ATE
dml_irm_ate_exp.sensitivity_analysis()
print(dml_irm_ate_exp.sensitivity_summary)

dml_irm_ate_restricted.sensitivity_analysis()
print(dml_irm_ate_restricted.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
            CI lower  theta lower        theta  theta upper     CI upper
treatment  81.115523  1245.293207  1676.151943   2107.01068  3230.468256

------------------ Robustness Values ------------------
           H_0     RV (%)   RVa (%)
treatment  0.0  11.168636  3.526529
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower        theta  theta upper     CI upper
treatment  1915.26637  2949.293327  4173.360686  5397.428046  6628.249376

------------------ Robustness Values ------------------
           H_0    RV (%)   RVa (%)
treatment  0.0  9.859796  7.408799
# ATTE
dml_irm_atte_exp.sensitivity_analysis()
print(dml_irm_atte_exp.sensitivity_summary)

dml_irm_atte_restricted.sensitivity_analysis()
print(dml_irm_atte_restricted.sensitivity_summary)
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower        theta  theta upper    CI upper
treatment  121.048443  1358.054458  1790.802261  2223.550064  3420.10178

------------------ Robustness Values ------------------
           H_0     RV (%)  RVa (%)
treatment  0.0  11.835759  3.77788
================== Sensitivity Analysis ==================

------------------ Scenario          ------------------
Significance Level: level=0.95
Sensitivity parameters: cf_y=0.03; cf_d=0.03, rho=1.0

------------------ Bounds with CI    ------------------
             CI lower  theta lower        theta  theta upper     CI upper
treatment  421.485548  1854.179304  2752.281361  3650.383418  5138.154461

------------------ Robustness Values ------------------
           H_0    RV (%)   RVa (%)
treatment  0.0  8.909182  4.121919

Benchmarking

We consider the benchmark set of all pre-treatment income variables.

# All pre-treatment income variables
benchmark_set = ['RE74', 'RE75', 'pos74', 'pos75']

bench_irm_ate_exp = dml_irm_ate_exp.sensitivity_benchmark(benchmark_set)
bench_irm_ate_restricted = dml_irm_ate_restricted.sensitivity_benchmark(benchmark_set)
print(bench_irm_ate_restricted)
bench_irm_atte_exp = dml_irm_atte_exp.sensitivity_benchmark(benchmark_set)
print(bench_irm_atte_exp)
bench_irm_atte_restricted = dml_irm_atte_restricted.sensitivity_benchmark(benchmark_set)
print(bench_irm_atte_restricted)
C:\Users\PuD\anaconda3\Lib\site-packages\doubleml\_utils_checks.py:204: UserWarning:

Propensity predictions from learner RandomForestClassifier(max_depth=5, random_state=42) for ml_m are close to zero or one (eps=1e-12).

C:\Users\PuD\anaconda3\Lib\site-packages\doubleml\_utils_checks.py:204: UserWarning:

Propensity predictions from learner RandomForestClassifier(max_depth=5, random_state=42) for ml_m are close to zero or one (eps=1e-12).
               cf_y  cf_d  rho  delta_theta
treatment  0.049217   0.0 -1.0  -719.317259
               cf_y      cf_d       rho  delta_theta
treatment  0.056706  0.160853  0.178883   225.272376
               cf_y      cf_d  rho  delta_theta
treatment  0.049217  0.121465 -1.0  -2167.62187
# Utils to get benchmark dict
def make_bench_dict(pd_bench, name):
  bench_dict = {'cf_d' : pd_bench['cf_d'].astype(float).round(3).tolist(),
                'cf_y' : pd_bench['cf_y'].astype(float).round(3).tolist(),
                'name' : [name]}
  return bench_dict
contour_ate_exp = dml_irm_ate_exp.sensitivity_plot(grid_bounds = (0.25, 0.25), benchmarks = make_bench_dict(bench_irm_ate_exp, 'Pre Income'))
contour_ate_exp.show()
print(bench_irm_ate_exp)