도구변수를 활용하여 인과효과를 추정하기

Simple Example of DoWhy using Instrumental Variables

작성자: 최보경
원문: Simple example on using Instrumental Variables method for estimation

Loading the dataset

import numpy as np
import pandas as pd
import patsy as ps

from statsmodels.sandbox.regression.gmm import IV2SLS
import os, sys
from dowhy import CausalModel

개인의 교육이 미래 소득에 주는 영향을 추정하기 위한, 가상의 데이터 셋을 생성합니다.

교란변수(unobserved confounder) : 개인의 능력(ability)
도구변수(instrumental variable) : 교육 보조(education_voucher)

n_points = 10000 # 데이터 개수
education_abilty = 1 # discrete
education_voucher = 2 # discrete
income_abilty = 2 # discrete
income_education = 4 # education 과 곱해지는 계수로, 찾고자 하는 인과효과 estimate 의 정답지


# confounder (unobserved)
ability = np.random.normal(0, 3, size=n_points)

# instrument
voucher = np.random.normal(2, 1, size=n_points)

# treatment
education = np.random.normal(5, 1, size=n_points) + education_abilty * ability +\
            education_voucher * voucher

# outcome
income = np.random.normal(10, 3, size=n_points) +\
         income_abilty * ability + income_education * education

# 가상의 dataset (exclude confounder `ability` which we assume to be unobserved)
data = np.stack([education, income, voucher]).T
df = pd.DataFrame(data, columns = ['education', 'income', 'voucher'])
df.head()

education

income

voucher

5.638493

29.349797

1.974566

8.522389

46.239001

1.949956

12.840748

64.198096

2.253389

2.954071

15.591053

0.939508

10.117519

51.873490

1.779209

Using DoWhy to estimate the causal effect of education on future income

4가지 절차를 따릅니다.

문제를 인과 그래프로 모델링합니다. (model)
관측된 변수로부터 인과추론이 될 수 있는지 확인합니다. (identify)
인과효과를 추정합니다. (estimate)
인과효과 추정치의 Robustness 를 확인합니다. (refute)

step 1 : model

model = CausalModel(
          data = df,
          treatment = 'education',
          outcome = 'income',
          common_causes = ['U'], # unobserved confounding 을 임의의 U 컬럼에 넣는 것
          instruments = ['voucher']
        )
model.view_model()

from IPython.display import Image, display
display(Image(filename="causal_model.png"))

step 2 : identify

identified_estimand = model.identify_effect(proceed_when_unidentifiable = True)
print(identified_estimand)

Estimand type: nonparametric-ate

### Estimand : 1
Estimand name: backdoor1 (Default)
Estimand expression:
     d                           
────────────(Expectation(income))
d[education]                     
Estimand assumption 1, Unconfoundedness: If U→{education} and U→income then P(income|education,,U) = P(income|education,)

### Estimand : 2
Estimand name: iv
Estimand expression:
Expectation(Derivative(income, [voucher])*Derivative([education], [voucher])**
(-1))
Estimand assumption 1, As-if-random: If U→→income then ¬(U →→{voucher})
Estimand assumption 2, Exclusion: If we remove {voucher}→{education}, then ¬({voucher}→income)

### Estimand : 3
Estimand name: frontdoor
No such variable found!

identify_effect()

Identified Estimand 를 리턴해주는 메소드로, Estimand 가 3가지 타입이 나옵니다.

Estimand 의 종류가 non-parametric ATE 일 때 Backdoor / IV / Frontdoor 와 같은 Identification methods 를 리턴합니다.

# step 3 : estimate
# step 2 에서 Estimand 가 3가지가 있는데, 그중에서 IV 를 선택함. 

estimate = model.estimate_effect(identified_estimand, 
                                 method_name = 'iv.instrumental_variable',
                                 test_significance = True)

print(estimate)

Estimand type: nonparametric-ate

### Estimand : 1
Estimand name: backdoor
Estimand expression:
     d
────────────(Expectation(income))
d[education]
Estimand assumption 1, Unconfoundedness: If U→{education} and U→income then P(income|education,,U) = P(income|education,)

### Estimand : 2
Estimand name: iv
Estimand expression:
Expectation(Derivative(income, [voucher])*Derivative([education], [voucher])**
(-1))
Estimand assumption 1, As-if-random: If U→→income then ¬(U →→{voucher})
Estimand assumption 2, Exclusion: If we remove {voucher}→{education}, then ¬({voucher}→income)

### Estimand : 3
Estimand name: frontdoor
No such variable found!

Step 3: Estimate

#Choose the second estimand: using IV
estimate = model.estimate_effect(identified_estimand,
        method_name="iv.instrumental_variable", test_significance=True)

print(estimate)

*** Causal Estimate ***

## Identified estimand
Estimand type: nonparametric-ate

### Estimand : 1
Estimand name: iv
Estimand expression:
Expectation(Derivative(income, [voucher])*Derivative([education], [voucher])**
(-1))
Estimand assumption 1, As-if-random: If U→→income then ¬(U →→{voucher})
Estimand assumption 2, Exclusion: If we remove {voucher}→{education}, then ¬({voucher}→income)

## Realized estimand
Realized estimand: Wald Estimator
Realized estimand type: nonparametric-ate
Estimand expression:

Expectation(Derivative(income, voucher))⋅Expectation(Derivative(education, vou

      -1
cher))
Estimand assumption 1, As-if-random: If U→→income then ¬(U →→{voucher})
Estimand assumption 2, Exclusion: If we remove {voucher}→{education}, then ¬({voucher}→income)
Estimand assumption 3, treatment_effect_homogeneity: Each unit's treatment ['education'] is affected in the same way by common causes of ['education'] and income
Estimand assumption 4, outcome_effect_homogeneity: Each unit's outcome income is affected in the same way by common causes of ['education'] and income

Target units: ate

## Estimate
Mean value: 3.984270555650861
p-value: [0, 0.001]

해석 방법 교육(education)을 한 단위 높일 때, 소득(income)이 4.01 포인트 가량 높아진다는 의미를 가지는 Estimate (추정치)를 보입니다.

추정된 인과효과는, p-value 가 0 ~ 0.001 사이로 통계적으로 유의합니다.

그 다음으로는, Placebo refutation test 를 통해서 이 추정치의 강건성을 검사합니다. 이 테스트에서는, 독립확률변수(X)가 도구변수와의 상관관계를 유지하면서, Treatment를 대체하도록 합니다. 따라서 진실된 인과효과는 0이 됩니다.

이 Estimator 가 올바르게 0에 가까운 값을 리턴하는지를 확인합니다.

step 4 : refute

ref = model.refute_estimate(identified_estimand, 
                            estimate,
                            method_name = "placebo_treatment_refuter", 
                            placebo_type = "permute")

print(ref)

Refute: Use a Placebo Treatment
Estimated effect:3.984270555650861
New effect:-0.004644847508369445
p value:0.43999999999999995

2SLS와 비교

2-stage linear square regression 동일한 로직이기 때문에 기존 패키지의 coefficient, p-value 와 동일한 결과를 보입니다.

income_vec, endog = ps.dmatrices("income ~ education", data=df)
exog = ps.dmatrix("voucher", data=df)

m = IV2SLS(income_vec, endog, exog).fit()
m.summary()

Dep. Variable:

income

R-squared:

0.890

Model:

IV2SLS

Adj. R-squared:

0.890

Method:

Two Stage

F-statistic:

1.373e+04

Least Squares

Prob (F-statistic):

0.00

Date:

Sun, 19 Dec 2021

Time:

09:48:50

No. Observations:

10000

Df Residuals:

9998

Df Model:

coef

std err

P>|t|

[0.025

0.975]

Intercept

10.2193

0.314

32.568

0.000

9.604

10.834

education

3.9779

0.034

117.195

0.000

3.911

4.044

Omnibus:

3.175

Durbin-Watson:

1.953

Prob(Omnibus):

0.204

Jarque-Bera (JB):

3.137

Skew:

0.040

Prob(JB):

0.208

Kurtosis:

3.032

Cond. No.

25.7

Previous멤버십 리워드 프로그램의 효과 추정하기 Next모델 결과 해석하기

Last updated 3 years ago