Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456

Jaehyun Song
Jaehyun SongAssistant Professor at Waseda University
Samii et al. (2016)
Retrospective Causal Inference with Machine
Learning Ensembles: An Application to
Anti-recidivism Policies in Colombia
2018/03/09
0
(Retrospective Intervention Effects; RIE)
• (outcome)
•
• (PSM)
•
(MSE)
2
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
•
• ( )
˙ ˙ ˙
3
˙ ?
• ATE (+ATT/ATC)
• ATE E[Y(Treated)]−E[Y(Controlled)]
• RIE E[Y(a)]−E[Y]
• RIE (a)
˙ ˙ ˙ ˙
⇒ ATE
• E[Y(a)] Matching
4
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
Y
Aj j
• 2×2×2
{A1,A2,A3}
W
U
• (A)
×
W //
**
%%

Y
A1
44
A2
99
...
??
U //
55
::
??
AJ
CC
5
Aji i j
• Aji = a a′
• τji(a,a′) =
Yi(a,A−j)−Yi(a′,A−j)
• j
( )
Q Aj
W //
**
%%

Y
A1
44
A2
99
...
??
U //
55
::
??
AJ
CC
5
: (W,A−j) OK
set.seed(19861008)
W - rnorm(1000, 0, 3)
W - rnorm(1000, 0, 3)
A1 - 1 + 2 * W + 3 * U + rnorm(1000, 0, 0.1)
A2 - 2 + 3 * W + 4 * U + rnorm(1000, 0, 0.1)
A3 - 3 + 4 * W + 5 * U + rnorm(1000, 0, 0.1)
Y - 5 + 1 * W + 2 * A1 + 3 * A2 + 4 * A3 + rnorm(1000, 0, 0.1)
6
A1 ( = 2) ...
 summary(lm(Y ~ A1))
Estimate Std. Error t value Pr(|t|)
(Intercept) 11.56375 0.37966 30.46 2e-16
A1 13.38875 0.03474 385.39 2e-16
...???
7
A2,3 ...
 summary(lm(Y ~ A1 + A2 + A3))
Estimate Std. Error t value Pr(|t|)
(Intercept) 3.02025 0.01212 249.219  2e-16
A1 -0.31520 0.04251 -7.414 2.62e-13
A2 2.71173 0.08256 32.844  2e-16
A3 5.62086 0.04165 134.946  2e-16
......???
8
W
 summary(lm(Y ~ A1 + W))
Estimate Std. Error t value Pr(|t|)
(Intercept) 12.359611 0.037851 326.5 2e-16
A1 12.664882 0.004146 3054.7 2e-16
W 4.682613 0.014818 316.0 2e-16
.........???
9
A2,3,W
 summary(lm(Y ~ A1 + A2 + A3 + W))
Estimate Std. Error t value Pr(|t|)
(Intercept) 5.00778 0.02246 223.00 2e-16
A1 2.00331 0.02937 68.22 2e-16
A2 3.02114 0.02755 109.66 2e-16
A3 3.98110 0.02286 174.17 2e-16
W 1.00791 0.01120 89.95 2e-16
!!
10
U
 summary(lm(Y ~ A1 + A2 + A3 + W + U))
Estimate Std. Error t value Pr(|t|)
(Intercept) 4.96969 0.12074 41.159  2e-16
A1 2.00807 0.03291 61.021  2e-16
A2 3.02668 0.03252 93.058  2e-16
A3 3.98850 0.03247 122.848  2e-16
W 0.95214 0.17403 5.471 5.66e-08
U -0.07346 0.22874 -0.321 0.748
...
( )
11
RIE (
)
Aj ⊥ (Yi(a,A−ji),Yi(a′
,A−ji))′
|(A−ji,W)
12
OLS Matching
1. (homogenous)
2.
3. DGP
•
Direct matching (Ho et al. 2007 )
• →
13
RIE
RIE
j (Aj) RIE
ψj = E[Y(aj,A−j)]
counterfactual
− E[Y]
observed
• Aj aj ( Aj )
14
1. A = a Y = Y(a)
• SUTVA
•
2. aj Y(aj,A−ji) ⊥ Aj|(W,A−j)
•
3. aj Pr[Aj = aj|W,A−j]  b (b )
• (overlap)
• overlap
15
RIE IPW
ψIPW
=
1
N
N
∑
i=1
I(aj)
ˆgj(aj|Wi,A−ji)
Yi −Y.
• I(aj): Aj = aj 1 Aj ̸= aj 0
• ˆgj(aj|Wi,A−ji): Pr[Aj = aj|Wi,A−ji] ⇐
⇒ Aj = aj
16
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
ˆgj(aj|Wi,A−ji)
•
• logistic KRLS BART
• v-fold Cross-validation
1.
2. (MSE)
⇒ Super Learner algorithm
17
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
Cross-validation
: (MSE)
ℓc
j =
1
N
N
∑
i=1
I(Aji = aj)− ˆg
c,v(i)
j (aj|Wi,A−ji)
2
• ˆgc,v(i): c index v v-fold CV
sub-sample (hold-out) index (v(i) i v)
⇒ Aj
19
Ensemble
(w)
(w1∗
j ,...,wC∗
j ) = arg min(w1∗
j ,...,wC∗
j )
1
N
N
∑
i=1
I(Aji = aj)−
C
∑
c=1
wc
j ˆg
c,v(i)
j (aj|Wi,A−ji)
2
,
C
∑
c=1
wc
j = 1, wc
j  0.
• MSE (ℓc
j ) w
20
Ensemble
w ⇒ Ensemble IPW
ˆgj(aj|Wi,A−ji) =
C
∑
c=1
wc∗
j ˆgc
j (aj|Wi,A−ji).
20
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
Y(0) = W1 +0.5(W1 −min(W1))2
+ε0,
Y(1) = W1 +0.75(W1 −min(W1))2
+0.75(W1 −min(W1))3
+ε1,
Pr[A = 1|W1] = logit−1
(−0.5+0.75W1 −0.5[W1 −mean(W1)]2
),
W,ε ∼ Normal(0,1). (1)
DGP
•
•
• A (0 → 1) Y(1) Y(0)+X
21
3 2 1 0 1 2
0.00.10.20.30.4
W1
Propensityscore
Propensity score over W1
21
3 2 1 0 1 2
050100150
W1
Y(1)(filled)andY(0)(hollow)
Potential outcomes over W1
21
3 2 1 0 1 2
050100150
W1
Treated(filled)andcontrol(hollow)outcomes
Observed data
(after treatment assignment)
21
(Nc = 5)
1. Logistic regression
2. t-regularized logistic regression
3. Kernal regularized least squares (KRLS)
4. Bayesian additive regression trees (BART)
5. v-support vector machine (SVM)
1. OLS
2. Naïve IPW: logistic regression
3. Matching: Mahalanobis
Normal(0,1) noise 0 ∼ 10
22
Number of noise covariates
Bias
0 5 10
0.050.050.100.150.20
OLS
Matching
Naive IPW
Ensemble IPW
• Matching Ensemble IPW
• Ensemble IPW Noise ( )
23
Number of noise covariates
S.E.
0 5 10
0.91.01.1
OLS
Matching
Naive IPW
Ensemble IPW
•
23
Number of noise covariates
RMSE
0 5 10
0.000.050.100.150.200.25
OLS
Matching
Naive IPW
Ensemble IPW
• Ensemble IPW
• Noise RMSE ×
23
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
1 ( )
• 0 ∼ 3
• 114
24
1 ( )
• {0,1}
• Employment:
• Security:
• Confidence:
• Depression:
• Excom.peers:
• Ties to commander:
24
1. WLS:
2. Naïve IPW: Logistic
3. Matching:
4. Ensemble IPW ←
25
•
• BART ...
26
( )
/11C2:5385,12945,556C51D19125183129453558491
27
( )
/11C2:5385,12945,556C51D19125183129453558491
28
Ensemble
29
RIE
C2:5385,12945,556C51D19125183129453558491
30
•
•
• HPC 20
•
31
Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456
1 of 46

Recommended

Budynas sm ch20Budynas sm ch20
Budynas sm ch20luisorellano125
613 views27 slides
Tugas modul 6Tugas modul 6
Tugas modul 6trinurpratiwi
298 views2 slides
EligheorEligheor
Eligheornickjeorly
186 views4 slides
BryanBryan
BryanJesthiger Cohil
363 views5 slides

More Related Content

What's hot(17)

Capítulo 03   materiaisCapítulo 03   materiais
Capítulo 03 materiais
Jhayson Carvalho378 views
Sriram1000991882-Report-FractureMechanicsSriram1000991882-Report-FractureMechanics
Sriram1000991882-Report-FractureMechanics
Sriram Sambasivam363 views
1515
15
Bhargav Dhavala56 views
EligheorEligheor
Eligheor
nickjeorly449 views
Budynas sm ch01Budynas sm ch01
Budynas sm ch01
luisorellano125838 views
Mecánica estática Mecánica estática
Mecánica estática
hbk1441194 views
Laporan pemodelan dan simulasiLaporan pemodelan dan simulasi
Laporan pemodelan dan simulasi
Irwansyah Hazniel1.8K views
Poster of surveying tasksPoster of surveying tasks
Poster of surveying tasks
Talha Hussain107 views
Capítulo 04   carga e análise de tensãoCapítulo 04   carga e análise de tensão
Capítulo 04 carga e análise de tensão
Jhayson Carvalho474 views
07.18.2013 - Michael Clemens07.18.2013 - Michael Clemens
07.18.2013 - Michael Clemens
AMDSeminarSeries355 views
Ejerciciooo3Ejerciciooo3
Ejerciciooo3
DanielRojas11800274 views
solucionario del capitulo 12 solucionario del capitulo 12
solucionario del capitulo 12
jasson silva18.5K views
01 deflectionof beams01 deflectionof beams
01 deflectionof beams
tareqsamar235 views
Graficas de movimientoGraficas de movimiento
Graficas de movimiento
davidinfanzon42 views
Engineering economy sample test #2 solutionEngineering economy sample test #2 solution
Engineering economy sample test #2 solution
Le Nguyen Truong Giang112 views

Similar to Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456

Ch02Ch02
Ch02Paralafakyou Mens
241 views32 slides

Similar to Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456(20)

Electronusa Mechanical SystemElectronusa Mechanical System
Electronusa Mechanical System
Electronusa Mechanical System305 views
Electronusa Mechanica SystemElectronusa Mechanica System
Electronusa Mechanica System
Electronusa Mechanical System172 views
Ch02Ch02
Ch02
Paralafakyou Mens241 views
Electronusa Mechanical SystemElectronusa Mechanical System
Electronusa Mechanical System
Electronusa Mechanical System263 views
Electronusa Mechanical SystemElectronusa Mechanical System
Electronusa Mechanical System
Electronusa Mechanical System238 views
Electronusa Mechanical SystemElectronusa Mechanical System
Electronusa Mechanical System
Electronusa Mechanical System327 views
Electronusa Mechanical SystemElectronusa Mechanical System
Electronusa Mechanical System
Electronusa Mechanical System292 views
Ee2365 nol part 2Ee2365 nol part 2
Ee2365 nol part 2
Arun Kumaar1.9K views
Chi2017 yamanaka novideoChi2017 yamanaka novideo
Chi2017 yamanaka novideo
Meiji University474 views
BENDING STRESS IN A BEAMSBENDING STRESS IN A BEAMS
BENDING STRESS IN A BEAMS
Vj NiroSh2.4K views
Linear programmingLinear programming
Linear programming
Reyna Marie Rosales580 views
Capítulo 02   considerações estatísticasCapítulo 02   considerações estatísticas
Capítulo 02 considerações estatísticas
Jhayson Carvalho243 views

Recently uploaded(20)

SOA PPT ON SEA TURTLES.pptxSOA PPT ON SEA TURTLES.pptx
SOA PPT ON SEA TURTLES.pptx
EuniceOseiYeboah7 views
Thanks Giving Encouragement Wednesday.pptxThanks Giving Encouragement Wednesday.pptx
Thanks Giving Encouragement Wednesday.pptx
FamilyWorshipCenterD8 views
Al Kindi.pptxAl Kindi.pptx
Al Kindi.pptx
MubbaraShahzadi5 views
CitSciOz MOUA Inspiring Change Through ArtCitSciOz MOUA Inspiring Change Through Art
CitSciOz MOUA Inspiring Change Through Art
Christian Bartens37 views
Prospectus (1).pdfProspectus (1).pdf
Prospectus (1).pdf
PancrazioScalambrino12 views

Review: Cyrus, Samii, Laura Paler, and Sarah Zukerman Daly. 2016. “Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia.” Political Analysis. 22 (4) pp. 434-456

  • 1. Samii et al. (2016) Retrospective Causal Inference with Machine Learning Ensembles: An Application to Anti-recidivism Policies in Colombia 2018/03/09 0
  • 2. (Retrospective Intervention Effects; RIE) • (outcome) • • (PSM) • (MSE) 2
  • 4. • • ( ) ˙ ˙ ˙ 3
  • 5. ˙ ? • ATE (+ATT/ATC) • ATE E[Y(Treated)]−E[Y(Controlled)] • RIE E[Y(a)]−E[Y] • RIE (a) ˙ ˙ ˙ ˙ ⇒ ATE • E[Y(a)] Matching 4
  • 7. Y Aj j • 2×2×2 {A1,A2,A3} W U • (A) × W // ** %% Y A1 44 A2 99 ... ?? U // 55 :: ?? AJ CC 5
  • 8. Aji i j • Aji = a a′ • τji(a,a′) = Yi(a,A−j)−Yi(a′,A−j) • j ( ) Q Aj W // ** %% Y A1 44 A2 99 ... ?? U // 55 :: ?? AJ CC 5
  • 9. : (W,A−j) OK set.seed(19861008) W - rnorm(1000, 0, 3) W - rnorm(1000, 0, 3) A1 - 1 + 2 * W + 3 * U + rnorm(1000, 0, 0.1) A2 - 2 + 3 * W + 4 * U + rnorm(1000, 0, 0.1) A3 - 3 + 4 * W + 5 * U + rnorm(1000, 0, 0.1) Y - 5 + 1 * W + 2 * A1 + 3 * A2 + 4 * A3 + rnorm(1000, 0, 0.1) 6
  • 10. A1 ( = 2) ... summary(lm(Y ~ A1)) Estimate Std. Error t value Pr(|t|) (Intercept) 11.56375 0.37966 30.46 2e-16 A1 13.38875 0.03474 385.39 2e-16 ...??? 7
  • 11. A2,3 ... summary(lm(Y ~ A1 + A2 + A3)) Estimate Std. Error t value Pr(|t|) (Intercept) 3.02025 0.01212 249.219 2e-16 A1 -0.31520 0.04251 -7.414 2.62e-13 A2 2.71173 0.08256 32.844 2e-16 A3 5.62086 0.04165 134.946 2e-16 ......??? 8
  • 12. W summary(lm(Y ~ A1 + W)) Estimate Std. Error t value Pr(|t|) (Intercept) 12.359611 0.037851 326.5 2e-16 A1 12.664882 0.004146 3054.7 2e-16 W 4.682613 0.014818 316.0 2e-16 .........??? 9
  • 13. A2,3,W summary(lm(Y ~ A1 + A2 + A3 + W)) Estimate Std. Error t value Pr(|t|) (Intercept) 5.00778 0.02246 223.00 2e-16 A1 2.00331 0.02937 68.22 2e-16 A2 3.02114 0.02755 109.66 2e-16 A3 3.98110 0.02286 174.17 2e-16 W 1.00791 0.01120 89.95 2e-16 !! 10
  • 14. U summary(lm(Y ~ A1 + A2 + A3 + W + U)) Estimate Std. Error t value Pr(|t|) (Intercept) 4.96969 0.12074 41.159 2e-16 A1 2.00807 0.03291 61.021 2e-16 A2 3.02668 0.03252 93.058 2e-16 A3 3.98850 0.03247 122.848 2e-16 W 0.95214 0.17403 5.471 5.66e-08 U -0.07346 0.22874 -0.321 0.748 ... ( ) 11
  • 15. RIE ( ) Aj ⊥ (Yi(a,A−ji),Yi(a′ ,A−ji))′ |(A−ji,W) 12
  • 16. OLS Matching 1. (homogenous) 2. 3. DGP • Direct matching (Ho et al. 2007 ) • → 13
  • 17. RIE
  • 18. RIE j (Aj) RIE ψj = E[Y(aj,A−j)] counterfactual − E[Y] observed • Aj aj ( Aj ) 14
  • 19. 1. A = a Y = Y(a) • SUTVA • 2. aj Y(aj,A−ji) ⊥ Aj|(W,A−j) • 3. aj Pr[Aj = aj|W,A−j] b (b ) • (overlap) • overlap 15
  • 20. RIE IPW ψIPW = 1 N N ∑ i=1 I(aj) ˆgj(aj|Wi,A−ji) Yi −Y. • I(aj): Aj = aj 1 Aj ̸= aj 0 • ˆgj(aj|Wi,A−ji): Pr[Aj = aj|Wi,A−ji] ⇐ ⇒ Aj = aj 16
  • 22. ˆgj(aj|Wi,A−ji) • • logistic KRLS BART • v-fold Cross-validation 1. 2. (MSE) ⇒ Super Learner algorithm 17
  • 24. Cross-validation : (MSE) ℓc j = 1 N N ∑ i=1 I(Aji = aj)− ˆg c,v(i) j (aj|Wi,A−ji) 2 • ˆgc,v(i): c index v v-fold CV sub-sample (hold-out) index (v(i) i v) ⇒ Aj 19
  • 25. Ensemble (w) (w1∗ j ,...,wC∗ j ) = arg min(w1∗ j ,...,wC∗ j ) 1 N N ∑ i=1 I(Aji = aj)− C ∑ c=1 wc j ˆg c,v(i) j (aj|Wi,A−ji) 2 , C ∑ c=1 wc j = 1, wc j 0. • MSE (ℓc j ) w 20
  • 26. Ensemble w ⇒ Ensemble IPW ˆgj(aj|Wi,A−ji) = C ∑ c=1 wc∗ j ˆgc j (aj|Wi,A−ji). 20
  • 28. Y(0) = W1 +0.5(W1 −min(W1))2 +ε0, Y(1) = W1 +0.75(W1 −min(W1))2 +0.75(W1 −min(W1))3 +ε1, Pr[A = 1|W1] = logit−1 (−0.5+0.75W1 −0.5[W1 −mean(W1)]2 ), W,ε ∼ Normal(0,1). (1) DGP • • • A (0 → 1) Y(1) Y(0)+X 21
  • 29. 3 2 1 0 1 2 0.00.10.20.30.4 W1 Propensityscore Propensity score over W1 21
  • 30. 3 2 1 0 1 2 050100150 W1 Y(1)(filled)andY(0)(hollow) Potential outcomes over W1 21
  • 31. 3 2 1 0 1 2 050100150 W1 Treated(filled)andcontrol(hollow)outcomes Observed data (after treatment assignment) 21
  • 32. (Nc = 5) 1. Logistic regression 2. t-regularized logistic regression 3. Kernal regularized least squares (KRLS) 4. Bayesian additive regression trees (BART) 5. v-support vector machine (SVM) 1. OLS 2. Naïve IPW: logistic regression 3. Matching: Mahalanobis Normal(0,1) noise 0 ∼ 10 22
  • 33. Number of noise covariates Bias 0 5 10 0.050.050.100.150.20 OLS Matching Naive IPW Ensemble IPW • Matching Ensemble IPW • Ensemble IPW Noise ( ) 23
  • 34. Number of noise covariates S.E. 0 5 10 0.91.01.1 OLS Matching Naive IPW Ensemble IPW • 23
  • 35. Number of noise covariates RMSE 0 5 10 0.000.050.100.150.200.25 OLS Matching Naive IPW Ensemble IPW • Ensemble IPW • Noise RMSE × 23
  • 37. 1 ( ) • 0 ∼ 3 • 114 24
  • 38. 1 ( ) • {0,1} • Employment: • Security: • Confidence: • Depression: • Excom.peers: • Ties to commander: 24
  • 39. 1. WLS: 2. Naïve IPW: Logistic 3. Matching: 4. Ensemble IPW ← 25