Introduction Estimation

Meeting 10: Matching

Gumilang Aryo Sahadewo

Universitas Gadjah Mada

November 12, 2019

Introduction Estimation

Motivation

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

Recall the two potential outcomes:

(
Y1i Di = 1
Yi =
Y0i Di = 0
Yi = Y0i + (Y1i − Y0i ) Di

The causal effect of a program is (Y1i − Y0i )

What is the problem with estimating the causal effect?
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

Recall the two potential outcomes:

(
Y1i Di = 1
Yi =
Y0i Di = 0
Yi = Y0i + (Y1i − Y0i ) Di

The causal effect of a program is (Y1i − Y0i )

What is the problem with estimating the causal effect?
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

Recall the two potential outcomes:

(
Y1i Di = 1
Yi =
Y0i Di = 0
Yi = Y0i + (Y1i − Y0i ) Di

The causal effect of a program is (Y1i − Y0i )

What is the problem with estimating the causal effect?
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Motivation

The Problem of Counterfactual

Potential Outcome Framework

The fundamental problem of evaluation is no counterfactual

ATT = E [Y1i | Di = 1] − E [Y0i | Di = 1]
ATNT = E [Y1i | Di = 0] − E [Y0i | Di = 0]
ATE = E [Y1i − Yoi ] = P (D = 1) · ATT + P (D = 0) · ATNT
The methods that we’ve studied so far seek to construct a
valid comparison group
Matching is another method that applies statistical techniques
to construct a comparison group
Matching identifies average unobserved counterfactuals
Introduction Estimation

Matching Methods

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

The identifying assumption is selection on observables:

(Y (0) , Y (1)) ⊥ D | X

This is equivalent to:

Pr (D = 1 | Y (0) , Y (1) , X ) = Pr (D = 1 | X )

E (D = 1 | Y (0) , Y (1) , X ) = E (D = 1 | X )
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

The identifying assumption is selection on observables:

(Y (0) , Y (1)) ⊥ D | X

This is equivalent to:

Pr (D = 1 | Y (0) , Y (1) , X ) = Pr (D = 1 | X )

E (D = 1 | Y (0) , Y (1) , X ) = E (D = 1 | X )
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Differences between treatment and comparison group are

captured in X
ATT : Y0i ⊥ D | X → E [Y0i | D = 1, X ] = E [Y0i | D = 0, X ]
ATNT : Y1i ⊥ D | X → E [Y1i | D = 1, X ] =
E [Y1i | D = 0, X ]
ATE : Y0 , Y1 ⊥ D | X
What would be the threat to identification?
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Differences between treatment and comparison group are

captured in X
ATT : Y0i ⊥ D | X → E [Y0i | D = 1, X ] = E [Y0i | D = 0, X ]
ATNT : Y1i ⊥ D | X → E [Y1i | D = 1, X ] =
E [Y1i | D = 0, X ]
ATE : Y0 , Y1 ⊥ D | X
What would be the threat to identification?
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Differences between treatment and comparison group are

captured in X
ATT : Y0i ⊥ D | X → E [Y0i | D = 1, X ] = E [Y0i | D = 0, X ]
ATNT : Y1i ⊥ D | X → E [Y1i | D = 1, X ] =
E [Y1i | D = 0, X ]
ATE : Y0 , Y1 ⊥ D | X
What would be the threat to identification?
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Differences between treatment and comparison group are

captured in X
ATT : Y0i ⊥ D | X → E [Y0i | D = 1, X ] = E [Y0i | D = 0, X ]
ATNT : Y1i ⊥ D | X → E [Y1i | D = 1, X ] =
E [Y1i | D = 0, X ]
ATE : Y0 , Y1 ⊥ D | X
What would be the threat to identification?
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Differences between treatment and comparison group are

captured in X
ATT : Y0i ⊥ D | X → E [Y0i | D = 1, X ] = E [Y0i | D = 0, X ]
ATNT : Y1i ⊥ D | X → E [Y1i | D = 1, X ] =
E [Y1i | D = 0, X ]
ATE : Y0 , Y1 ⊥ D | X
What would be the threat to identification?
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Just like in the standard OLS framework, differences between

treatment and comparison group are not captured in X
Differences between the two groups are on the unobservable
characteristics
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables

Just like in the standard OLS framework, differences between

treatment and comparison group are not captured in X
Differences between the two groups are on the unobservable
characteristics
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Common Support

We observe individuals in the treatment and non-treatment

unit with the same characteristics
ATT : P (D = 1 | X ) < 1
ATNT : 0 < P (D = 1 | X )
ATE : 0 < P (D = 1 | X ) < 1
Imagine a scenario:
When X = x , then Pr (D = 1 | X = x ) = 1.
We won’t observe individuals in the control group with X = x
Thus, there we cannot obtain a valid comparison group
Introduction Estimation

Matching Methods

Matching Methods
Selection on Observables & Common Support

If the assumptions hold, we can use the observed average

outcome of the non-treatment units to estimate the
counterfactual outcome
Introduction Estimation

Finding a match

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Finding a match

Finding a match

The goal of matching is to approximate the characteristics

that explain individual’s decision to enroll
This procedure requires a large data set. Why?
Introduction Estimation

Finding a match

Finding a match

The goal of matching is to approximate the characteristics

that explain individual’s decision to enroll
This procedure requires a large data set. Why?
Introduction Estimation

Finding a match

Finding a match

Treated Untreated
Months unemployed Poor Months unemployed Poor
5 1 2 1
10 0 12 1
3 0 8 1
20 0 14 0
2 1 4 0
8 1 6 1
6 1 1 1
Introduction Estimation

Finding a match

Finding a match

Treated Untreated
Months unemployed Poor Months unemployed Poor
5 1 2 1
10 0 12 1
3 0 8 1
20 0 14 0
2 1 4 0
8 1 6 1
6 1 1 1
Introduction Estimation

Finding a match

Finding a match

Treated Untreated
Months unemployed Poor Months unemployed Poor
5 1 2 1
10 0 12 1
3 0 8 1
20 0 14 0
2 1 4 0
8 1 6 1
6 1 1 1
Introduction Estimation

Finding a match

Finding a match

Treated Untreated
Months unemployed Poor Months unemployed Poor
5 1 2 1
10 0 12 1
3 0 8 1
20 0 14 0
2 1 4 0
8 1 6 1
6 1 1 1
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Finding a match

The problem of matching

It is difficult to identify a match for each of the units in the

treatment group
The list of observed characteristics is large
Each characteristics takes on many values
We can easily run into the curse of dimensionality
Dilemma:
Limit the set of observed characteristics, but...
Increase the number of observed characteristics, but...
Introduction Estimation

Propensity Score Matching

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Propensity Score Matching

Idea

A solution to the curse of dimensionality problem is the

propensity score matching
The method computes the probability that the unit is enrolled
in the program using the observable characteristics
We do this for treatment and non-treatment units
Note that we only use the baseline or pre-treatment
observable characteristics
The propensity score is between 0 and 1
Introduction Estimation

Propensity Score Matching

Idea

A solution to the curse of dimensionality problem is the

propensity score matching
The method computes the probability that the unit is enrolled
in the program using the observable characteristics
We do this for treatment and non-treatment units
Note that we only use the baseline or pre-treatment
observable characteristics
The propensity score is between 0 and 1
Introduction Estimation

Propensity Score Matching

Idea

A solution to the curse of dimensionality problem is the

propensity score matching
The method computes the probability that the unit is enrolled
in the program using the observable characteristics
We do this for treatment and non-treatment units
Note that we only use the baseline or pre-treatment
observable characteristics
The propensity score is between 0 and 1
Introduction Estimation

Propensity Score Matching

Idea

A solution to the curse of dimensionality problem is the

propensity score matching
The method computes the probability that the unit is enrolled
in the program using the observable characteristics
We do this for treatment and non-treatment units
Note that we only use the baseline or pre-treatment
observable characteristics
The propensity score is between 0 and 1
Introduction Estimation

Propensity Score Matching

Idea

A solution to the curse of dimensionality problem is the

propensity score matching
The method computes the probability that the unit is enrolled
in the program using the observable characteristics
We do this for treatment and non-treatment units
Note that we only use the baseline or pre-treatment
observable characteristics
The propensity score is between 0 and 1
Introduction Estimation

Propensity Score Matching

Idea

The propensity score is:

e (x) = P (D = 1 | X = x)

The score is used to make this assumption:

X ⊥ D | e (X )

Combining:

(Y1 , Y0 ) ⊥ D | X and 0 < e (x) < 1

(Y1 , Y0 ) ⊥ D | e (X )
Introduction Estimation

Propensity Score Matching

Idea

The propensity score is:

e (x) = P (D = 1 | X = x)

The score is used to make this assumption:

X ⊥ D | e (X )

Combining:

(Y1 , Y0 ) ⊥ D | X and 0 < e (x) < 1

(Y1 , Y0 ) ⊥ D | e (X )
Introduction Estimation

Propensity Score Matching

Idea

The propensity score is:

e (x) = P (D = 1 | X = x)

The score is used to make this assumption:

X ⊥ D | e (X )

Combining:

(Y1 , Y0 ) ⊥ D | X and 0 < e (x) < 1

(Y1 , Y0 ) ⊥ D | e (X )
Introduction Estimation

Propensity Score Matching

Idea

Match treatment and non-treatment units with the closest

propensity score.
The matched non-treatment units become the comparison
group
The average difference in outcomes between the treatment
and the matched comparison is the estimate of the impact
The propensity score matching mimics a randomized
experiment
Treatment and comparison units have similar propensities.
Introduction Estimation

Propensity Score Matching

Idea

Match treatment and non-treatment units with the closest

propensity score.
The matched non-treatment units become the comparison
group
The average difference in outcomes between the treatment
and the matched comparison is the estimate of the impact
The propensity score matching mimics a randomized
experiment
Treatment and comparison units have similar propensities.
Introduction Estimation

Propensity Score Matching

Idea

Match treatment and non-treatment units with the closest

propensity score.
The matched non-treatment units become the comparison
group
The average difference in outcomes between the treatment
and the matched comparison is the estimate of the impact
The propensity score matching mimics a randomized
experiment
Treatment and comparison units have similar propensities.
Introduction Estimation

Propensity Score Matching

Idea

Match treatment and non-treatment units with the closest

propensity score.
The matched non-treatment units become the comparison
group
The average difference in outcomes between the treatment
and the matched comparison is the estimate of the impact
The propensity score matching mimics a randomized
experiment
Treatment and comparison units have similar propensities.
Introduction Estimation

Propensity Score Matching

Idea

Match treatment and non-treatment units with the closest

propensity score.
The matched non-treatment units become the comparison
group
The average difference in outcomes between the treatment
and the matched comparison is the estimate of the impact
The propensity score matching mimics a randomized
experiment
Treatment and comparison units have similar propensities.
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Find representative surveys to identify the treatment and

non-treatment units
Pool the sample and estimate the probability that each
individual receives the treatment (based on observable
characteristics)
It is important to include relevant variables to avoid a biased
estimate
theory and previous empirical findings
formal statistical tests
Always remember the tradeoff:
Small set of characteristics: selection on observable assumption
Large set of characteristics: problems of common support
Introduction Estimation

Propensity Score Matching

Steps to PSM

Obtain the propensity scores

Restrict the sample to units with a common support
For each enrolled unit, locate a subgroup of non-treated units
with similar propensity scores
Test whether the means for the treatment and non-treated
units are statistically different
Introduction Estimation

Propensity Score Matching

Steps to PSM

Obtain the propensity scores

Restrict the sample to units with a common support
For each enrolled unit, locate a subgroup of non-treated units
with similar propensity scores
Test whether the means for the treatment and non-treated
units are statistically different
Introduction Estimation

Propensity Score Matching

Steps to PSM

Obtain the propensity scores

Restrict the sample to units with a common support
For each enrolled unit, locate a subgroup of non-treated units
with similar propensity scores
Test whether the means for the treatment and non-treated
units are statistically different
Introduction Estimation

Propensity Score Matching

Steps to PSM

Obtain the propensity scores

Restrict the sample to units with a common support
For each enrolled unit, locate a subgroup of non-treated units
with similar propensity scores
Test whether the means for the treatment and non-treated
units are statistically different
Introduction Estimation

Propensity Score Matching

Steps to PSM

The measure of the impact is the difference between the

outcomes of the treatment and the matched comparison.
Introduction Estimation

Estimation

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Estimation

Matching Strategy and ATT

The matching strategy is:

Pair each treatment unit i with one or more comparable
non-treated units
Associate the outcome Yiobs a matched outcome Ŷi (0) given
the weighted outcomes of its neighbors:

wij Yjobs
X
Ŷi (0) =
j∈C (i)

C (i) is the set of neighbors with DP

= 0 of the treated subject i
wij is the weight of non-treated j, j∈C(i) wij = 1
Introduction Estimation

Estimation

Matching Strategy and ATT

The matching strategy is:

Pair each treatment unit i with one or more comparable
non-treated units
Associate the outcome Yiobs a matched outcome Ŷi (0) given
the weighted outcomes of its neighbors:

wij Yjobs
X
Ŷi (0) =
j∈C (i)

C (i) is the set of neighbors with DP

= 0 of the treated subject i
wij is the weight of non-treated j, j∈C(i) wij = 1
Introduction Estimation

Estimation

Matching Strategy and ATT

The matching strategy is:

Pair each treatment unit i with one or more comparable
non-treated units
Associate the outcome Yiobs a matched outcome Ŷi (0) given
the weighted outcomes of its neighbors:

wij Yjobs
X
Ŷi (0) =
j∈C (i)

C (i) is the set of neighbors with DP

= 0 of the treated subject i
wij is the weight of non-treated j, j∈C(i) wij = 1
Introduction Estimation

Estimation

Matching Strategy and ATT

The matching strategy is:

Pair each treatment unit i with one or more comparable
non-treated units
Associate the outcome Yiobs a matched outcome Ŷi (0) given
the weighted outcomes of its neighbors:

wij Yjobs
X
Ŷi (0) =
j∈C (i)

C (i) is the set of neighbors with DP

= 0 of the treated subject i
wij is the weight of non-treated j, j∈C(i) wij = 1
Introduction Estimation

Estimation

Matching Strategy and ATT

The matching strategy is:

Pair each treatment unit i with one or more comparable
non-treated units
Associate the outcome Yiobs a matched outcome Ŷi (0) given
the weighted outcomes of its neighbors:

wij Yjobs
X
Ŷi (0) =
j∈C (i)

C (i) is the set of neighbors with DP

= 0 of the treated subject i
wij is the weight of non-treated j, j∈C(i) wij = 1
Introduction Estimation

Estimation

Matching Strategy and ATT

The ATT:
E [Yi (1) − Yi (0) | Di = 1]
is estimated as:

ˆ = 1
X h i
ATT T
Yiobs − Ŷi (0)
N i:D =1
i

N T is the number of matched treated in the sample

Introduction Estimation

Estimation

Matching Strategy and ATT

The ATT:
E [Yi (1) − Yi (0) | Di = 1]
is estimated as:

ˆ = 1
X h i
ATT T
Yiobs − Ŷi (0)
N i:D =1
i

N T is the number of matched treated in the sample

Introduction Estimation

Matching Methods

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Matching Strategy and ATT

One-to-one matching is still the most desirable

But it is difficult to observe two units with the same
propensity scores
Matching methods have been developed to deal with this
problem:
Nearest-neighbor matching
Radius matching
Kernel matching
Stratification matching
Introduction Estimation

Matching Methods

Nearest Neighbor Matching

The absolute difference between the estimated propensity

scores of treatment and control groups are minimized
The group of control individuals are selected such that:

C (Pi ) = minj |Pi − Pj |

where:
Pi is the estimated propensity score for treated individuals i
Pj is the estimated propensity score for the control individuals j
Introduction Estimation

Matching Methods

Nearest Neighbor Matching

The absolute difference between the estimated propensity

scores of treatment and control groups are minimized
The group of control individuals are selected such that:

C (Pi ) = minj |Pi − Pj |

where:
Pi is the estimated propensity score for treated individuals i
Pj is the estimated propensity score for the control individuals j
Introduction Estimation

Matching Methods

Nearest Neighbor Matching

The absolute difference between the estimated propensity

scores of treatment and control groups are minimized
The group of control individuals are selected such that:

C (Pi ) = minj |Pi − Pj |

where:
Pi is the estimated propensity score for treated individuals i
Pj is the estimated propensity score for the control individuals j
Introduction Estimation

Matching Methods

Nearest Neighbor Matching

The absolute difference between the estimated propensity

scores of treatment and control groups are minimized
The group of control individuals are selected such that:

C (Pi ) = minj |Pi − Pj |

where:
Pi is the estimated propensity score for treated individuals i
Pj is the estimated propensity score for the control individuals j
Introduction Estimation

Matching Methods

Radius and Kernel Matching

Radius: each individual in the treatment group is matched

with individuals in the control group whose scores are within a
predefined interval of the treatment individuals’ propensity
score.
Kernel: each individual in the treatment group is matched
with the weighted average of control individuals’ outcomes.
Introduction Estimation

Matching Methods

Radius and Kernel Matching

Radius: each individual in the treatment group is matched

with individuals in the control group whose scores are within a
predefined interval of the treatment individuals’ propensity
score.
Kernel: each individual in the treatment group is matched
with the weighted average of control individuals’ outcomes.
Introduction Estimation

Syntax

1 Introduction
Motivation
Matching Methods
Finding a match
Propensity Score Matching

2 Estimation
Estimation
Matching Methods
Syntax
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

PSCORE

Stata command pscore calculates propensity scores

pscore also tests the balancing hypothesis through this
algorithm:
Split the sample in k equally spaced intervals of e (X )
Within each interval test that the average e (X ) of treated and
untreated do not differ
If the test fails, split the interval and test again
Continue until, in all intervals, the average e (X ) of treated
and untreated units do not differ
Within each interval, test that the means of each characteristic
do not differ between treated and untreated
Introduction Estimation

Syntax

Estimation

Use STATA package psmatch2 and pstest for estimation and

balance check