synthetic control methods for comparative case
TRANSCRIPT
SYNTHETIC CONTROL METHODS FORCOMPARATIVE CASE STUDIES: ESTIMATING THEEFFECT OF CALIFORNIA�S TOBACCO CONTROL
PROGRAMProgram Evaluation Presentation
Alberto Abadie Alexis Diamond Jens Hainmueller
Andrés Castañeda
October 2009
Universidad del Rosario (Institute) Program Evaluation October 2009 1 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
One Slide Presentation
Motivation
California�s Background
Methodology
Implementation
Data and Sample
Estimation Steps
Tables and Figures
Universidad del Rosario (Institute) Program Evaluation October 2009 2 / 18
Motivation
Justify the synthetic control approach
Study the e¤ects of California�s Proposition 99.
Universidad del Rosario (Institute) Program Evaluation October 2009 3 / 18
Motivation
Justify the synthetic control approach
Study the e¤ects of California�s Proposition 99.
Universidad del Rosario (Institute) Program Evaluation October 2009 3 / 18
California�s Background
Washington 1893. Moral and Health
Proposition 99. 1988
Earmarked: $100 million State, $20 million research
Universidad del Rosario (Institute) Program Evaluation October 2009 4 / 18
Methodology1
Objective: Construct a Synthetic variable
Framework:
j + 1 Regions: 1 exposed to treatment and j controlsT0 Number of pre-intervention periods and 1 � T0 � TY Nit is the outcome that would be observed by region i in time t withno treatmentY Iit is outcome that would be observed by region i in time t withtreatment
Universidad del Rosario (Institute) Program Evaluation October 2009 5 / 18
Methodology2
A1: Intervention has no e¤ect on the outcome before the treatmentperiod, so Y Nit = Y
Iit
After the treatment period Y Iit � Y Nit = αit
Dit is an indicator if i is exposed to the treatment
Therefore we can write Yit = Y Nit + αitDit
Universidad del Rosario (Institute) Program Evaluation October 2009 6 / 18
Methodology2
A1: Intervention has no e¤ect on the outcome before the treatmentperiod, so Y Nit = Y
Iit
After the treatment period Y Iit � Y Nit = αit
Dit is an indicator if i is exposed to the treatment
Therefore we can write Yit = Y Nit + αitDit
Universidad del Rosario (Institute) Program Evaluation October 2009 6 / 18
Methodology2
A1: Intervention has no e¤ect on the outcome before the treatmentperiod, so Y Nit = Y
Iit
After the treatment period Y Iit � Y Nit = αit
Dit is an indicator if i is exposed to the treatment
Therefore we can write Yit = Y Nit + αitDit
Universidad del Rosario (Institute) Program Evaluation October 2009 6 / 18
Methodology2
A1: Intervention has no e¤ect on the outcome before the treatmentperiod, so Y Nit = Y
Iit
After the treatment period Y Iit � Y Nit = αit
Dit is an indicator if i is exposed to the treatment
Therefore we can write Yit = Y Nit + αitDit
Universidad del Rosario (Institute) Program Evaluation October 2009 6 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are ZiUnknown common factor is λt
Varying factor loadings µiIf λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are ZiUnknown common factor is λt
Varying factor loadings µiIf λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are Zi
Unknown common factor is λt
Varying factor loadings µiIf λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are ZiUnknown common factor is λt
Varying factor loadings µiIf λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are ZiUnknown common factor is λt
Varying factor loadings µi
If λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
Methodologyprocedure
The aim is to estimate each αit for all t > T0Hence Y I1t is observed, we need to estimate Y
N1t to get
α1t = Y1t � TN1t (1)
A2: Y Nit = δt + θtZi + λtµi + εit
Covariates are ZiUnknown common factor is λt
Varying factor loadings µiIf λt is constant we get dif in dif
Universidad del Rosario (Institute) Program Evaluation October 2009 7 / 18
MethodologyProcedure 2
Consider W = (w2, ...,wj+1)0 such that wj � 0 and ∑j+1
j=2 wj = 1
A3: we can chose�w �2 , ...,w
�j+1
�0such that
j+1
∑j=2w �j Y
Nj = Y N1 (2)
j+1
∑j=2w �j Z
Nj = ZN1 (3)
This suggest that equation (1) would be
α̂1t = Y1t �j+1
∑j=2w �j Y
Nj (4)
Universidad del Rosario (Institute) Program Evaluation October 2009 8 / 18
MethodologyProcedure 2
Consider W = (w2, ...,wj+1)0 such that wj � 0 and ∑j+1
j=2 wj = 1
A3: we can chose�w �2 , ...,w
�j+1
�0such that
j+1
∑j=2w �j Y
Nj = Y N1 (2)
j+1
∑j=2w �j Z
Nj = ZN1 (3)
This suggest that equation (1) would be
α̂1t = Y1t �j+1
∑j=2w �j Y
Nj (4)
Universidad del Rosario (Institute) Program Evaluation October 2009 8 / 18
MethodologyProcedure 2
Consider W = (w2, ...,wj+1)0 such that wj � 0 and ∑j+1
j=2 wj = 1
A3: we can chose�w �2 , ...,w
�j+1
�0such that
j+1
∑j=2w �j Y
Nj = Y N1 (2)
j+1
∑j=2w �j Z
Nj = ZN1 (3)
This suggest that equation (1) would be
α̂1t = Y1t �j+1
∑j=2w �j Y
Nj (4)
Universidad del Rosario (Institute) Program Evaluation October 2009 8 / 18
MethodologyProcedure 2
Consider W = (w2, ...,wj+1)0 such that wj � 0 and ∑j+1
j=2 wj = 1
A3: we can chose�w �2 , ...,w
�j+1
�0such that
j+1
∑j=2w �j Y
Nj = Y N1 (2)
j+1
∑j=2w �j Z
Nj = ZN1 (3)
This suggest that equation (1) would be
α̂1t = Y1t �j+1
∑j=2w �j Y
Nj (4)
Universidad del Rosario (Institute) Program Evaluation October 2009 8 / 18
Implementation
Let X1 be a vector of characteristics for the exposed region
And X0 is a matrix that contains the same variables for the untreatedregions
The idea is obtain the vector W � that minimize jjX1 � X0W �jj
In particular jjX1 � X0W jjv =q(X1 � X0W )0 V (X1 � X0W )
Universidad del Rosario (Institute) Program Evaluation October 2009 9 / 18
Implementation
Let X1 be a vector of characteristics for the exposed region
And X0 is a matrix that contains the same variables for the untreatedregions
The idea is obtain the vector W � that minimize jjX1 � X0W �jj
In particular jjX1 � X0W jjv =q(X1 � X0W )0 V (X1 � X0W )
Universidad del Rosario (Institute) Program Evaluation October 2009 9 / 18
Data and Sample1
Variable of interest: Annual per capita cigarette consumption at thestate level
Panel data for the period 1970 �2000
Proposition 99 (P.99) was passed in 1988
Synthetic California is meant to reproduce the consumption ofcigarettes that would have been observed without the treatment in1988
Discarding:
Large-scale tobacco controlTaxes by 50 cents
Universidad del Rosario (Institute) Program Evaluation October 2009 10 / 18
Data and Sample2
Average retail price of cigarettes
Per capital personal income (logged)
The percentage of population age 15 �24
Per capita beer consumption
Three year lagged smoking consumption (1975, 1980 and 1988)eamer�font themes�de�ne the use of fonts in a presentation
Universidad del Rosario (Institute) Program Evaluation October 2009 11 / 18
Estimation Steps
1 Using the techniques described above the synthetic California (SC) isconstructed
SC is the mirror of the predictors of cigarette consumption in Californiabefore the treatment
2 The e¤ect of P.99 is estimated as the di¤erence in cigaretteconsumption between California and SC after P.99 was passed
3 A series of placebo studies are performed
Universidad del Rosario (Institute) Program Evaluation October 2009 12 / 18
Before Results
Universidad del Rosario (Institute) Program Evaluation October 2009 13 / 18
Trend in per capital sales: California Vs United States
Universidad del Rosario (Institute) Program Evaluation October 2009 14 / 18
Trend in per capital sales: California Vs SyntheticCalifornia
Universidad del Rosario (Institute) Program Evaluation October 2009 15 / 18
Per capital sales gap: California Vs 38 Control States
Universidad del Rosario (Institute) Program Evaluation October 2009 16 / 18
Per capital sales gap: California Vs 19 Control States withmean square prediction error less than two timesCalifornia�s
Universidad del Rosario (Institute) Program Evaluation October 2009 17 / 18
Final Remarks
Cigarettes sales in California were about 26 packs lower than whatthey would have been in the absence of P.99
The Methods are consistent regardless of the number of availablecomparison units.
The probability of obtaining a post/pre-P.99 MSPE ratio as large asCalifornia�s is 0.026
Universidad del Rosario (Institute) Program Evaluation October 2009 18 / 18