0% found this document useful (0 votes)
9 views45 pages

Understanding Matching in Statistics

Matching
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
9 views45 pages

Understanding Matching in Statistics

Matching
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Lecture 8.

Matching
MGT 2002 Statistics II

Feng-An Yang1

1 Departmentof Agricultural Economics


National Taiwan University

1 / 45
Outline

Introduction

Conditional Independence Assumption

Matching

Common Support Assumption

Practical Issues in Matching

Propensity Score Matching

Application

Summary

2 / 45
Introduction

▶ Our goal is to estimate the Average Treatment Effect (ATE)

ATE = E [Y1i ] − E [Y0i ]

▶ However, with observational data, we can only observe


outcomes for one potential state—treated or untreated—not
both
E [Y1i |Ti = 1] − E [Y0i |Ti = 0]
▶ As a result, the estimated effect may differ from the true ATE
due to non-random treatment assignment

▶ This is where the independence assumption becomes crucial

3 / 45
Introduction

Independence Assumption
Treatment assignment is independent of potential outcomes:
(Y1i , Y0i ) ⊥⊥ Ti
▶ This implies that the treatment assignment is random
▶ Individuals who received and did not receive the treatment
are, on average, similar in all respects that affect the outcome
▶ Individuals who received the treatment would have
experienced the same outcome in the absence of treatment as
those who did not receive it, and vice versa:
▶ E [Y1i |Ti = 1] = E [Y1i |Ti = 0] = E [Y1i ]
▶ E [Y0i |Ti = 1] = E [Y0i |Ti = 0] = E [Y0i ]
▶ If this assumption holds, we can identify the causal effect

E [Y1i |Ti = 1] − E [Y0i |Ti = 0] = E [Y1i ] − E [Y0i ] = ATE

4 / 45
Introduction

▶ Is this assumption likely to hold in real-world settings?


▶ Let’s revisit the example of the relationship between health
insurance and doctor visits
▶ Do you believe that:
▶ E [Y1i |with insurance] = E [Y1i |without insurance]
▶ E [Y0i |with insurance] = E [Y0i |without insurance]
▶ If the independence assumption holds, then individual
characteristics should, on average, be similar between
individuals with and without health insurance
▶ This condition is known as covariate balance

5 / 45
Introduction
▶ Covariates are the observed characteristics of individuals that
may influence both treatment assignment and the outcome
▶ Covariate balance means that the distribution of these
covariates is similar between the treatment group and the
control group
▶ In practice, we usually check the sample means between the
two groups to investigate the covariate balance
▶ This can be done by conducting mean comparison t-tests
between the treatment and control groups for each covariate
▶ If the independence assumption holds, we would expect:

Insured individuals Uninsured individuals


Covariate
(Treatment) (Control)
Age (years) 44.8 45.1
Income (USD/year) 49,700 49,200
Pre-existing condition 0.59 0.58
Education (years) 14.9 15.1

6 / 45
Introduction
▶ In observational settings, the independence assumption often
does not hold
▶ Individuals may self-select into the treatment group based on
factors related to the outcome
▶ Those who choose to purchase health insurance may be
systematically different from those who do not
▶ For example, they may have worse health, higher income, or
greater risk aversion

Insured individuals Uninsured individuals


Covariate
(Treatment) (Control)
Age (years) 45.2 39.8
Income (USD/year) 52,000 38,500
Pre-existing condition 0.62 0.35
Education (years) 15.1 12.3

7 / 45
Introduction

▶ Given the differences in observed characteristics (covariates)


between the treatment and control groups, what if we could
account for these differences directly?

▶ Then, any remaining differences in outcomes may be more


credibly attributed to the treatment effect

▶ In other words, we aim to “replicate” a randomized


experiment by making the treatment and control groups
comparable on observed variables

▶ Here, we would need to make a stronger assumption

8 / 45
Conditional Independence Assumption
Conditional Independence Assumption (CIA)
Treatment assignment is independent of potential outcomes,
conditional on observed covariates

(Y1i , Y0i ) ⊥
⊥ Ti |Xi

▶ Treatment assignment is as good as random once we control


for observed covariates Xi
▶ Also Known as the Selection on Observables assumption
▶ The expected outcome individuals would have experience
with/without the treatment is the same between those who
received the treatment and those who did not for each value
of Xi
E [Y1i |Ti = 1, Xi ] = E [Y1i |Ti = 0, Xi ]
E [Y0i |Ti = 1, Xi ] = E [Y0i |Ti = 0, Xi ]
9 / 45
Conditional Independence Assumption

▶ Recall our example: estimating the effect of health insurance


on the number of doctor visits
▶ Individuals who purchase insurance are older, sicker, more
educated, and have higher income
▶ Under the CIA, after controlling for observed characteristics
like age, income, health status, and education, the decision to
purchase insurance is as good as random
▶ Among individuals with the same observed covariates, whether
someone has insurance or not is independent of their potential
doctor visits under both insurance and no insurance
▶ So we can compare insured and uninsured individuals with
similar covariates to estimate the effect of health insurance

10 / 45
Conditional Independence Assumption

treatment effect = E [Y1i |Ti = 1|Xi ] − E [Y0i |Ti = 0|Xi ]


= E [Y1i − Y0i |Ti = |Xi ]
+ (E [Y0i |Ti = 1, Xi ] − E [Y0i |Ti = 0, Xi ])
= E [Y1i − Y0i |Ti = 1|Xi ] = ATT

▶ Under the CIA, we can estimate the Average Treatment effect


on the Treated (ATT)
▶ Recall that ATT refers to the average effect of the treatment
on those who actually received it
▶ This is often of practical and policy interest because many
real-world policy questions focus on how effective a program
or intervention was for its actual participants
▶ Did health insurance improve doctor visits for people who
chose to enroll?
▶ Did a job training program increase earnings for those who
attended?
11 / 45
Matching

ATT = E [Y1i − Y0i |Ti = 1|Xi ]

▶ With observational data, we observe Y1i for treated units, but


we do not observe their counterfactual Y0i
▶ What can we do?
▶ What if we impute the missing potential outcome for each
treated unit using information from one or more control units
with the closest values of X ?
▶ By conditioning on X , we can approximate E [Y0i |Ti = 1|Xi ]
▶ This method is called Matching

12 / 45
Matching

Let’s illustrate matching using a hypothetical example


▶ Outcome variable: Number of doctor visits
▶ Treatment variable: Having health insurance
▶ Covariates: age and self-reported health status (1-10)
▶ Treatment group: 5 individuals with health insurance
▶ Control group: 10 individuals without insurance
▶ Suppose that age and health status are the only observable
factors that affect both health insurance coverage and the
number of doctor visits

13 / 45
Matching
Treatment Control
ID Age Health Visits ID Age Health Visits
T1 52 4 8 C1 31 8 2
T2 55 3 9 C2 53 3 7
T3 50 5 7 C3 30 8 2
T4 48 4 6 C4 48 4 5
T5 53 3 9 C5 53 3 6
C6 55 3 5
C7 29 9 1
C8 50 5 4
C9 32 7 3
C10 52 4 6
51.6 3.8 7.8 43.3 5.4 4.1
▶ Individuals with health insurance are, on average, older, sicker,
and see doctors more often than those without insurance
▶ Without accounting for age and health status, the estimated
effect of insurance on the number of doctor visits is
3.7(= 7.8 − 4.1)
14 / 45
Matching

Treatment Matched Control


ID Age Health Visits ID Age Health Visits
T1 52 4 8 C10 52 4 6
T2 55 3 9 C6 55 3 6
T3 50 5 7 C8 50 5 4
T4 48 4 6 C4 48 4 5
T5 53 3 9 C2,C5 53 3 6.5
51.6 3.8 7.8 51.6 3.8 5.5

▶ After matching, individuals with health insurance now have


the same average age and health status as those without
insurance
▶ Conditional on age and health status, the estimated effect of
insurance on the number of doctor visits is 2.3(= 7.8 − 5.5)

15 / 45
What We Learn from This Matching Exercise?

▶ In observational settings, treatment and control groups may


differ substantially in observed characteristics (e.g., age,
health status), which can bias causal estimates

▶ Matching helps us construct a more comparable control group


by pairing treated individuals with similar individuals from the
control group based on observed covariates

▶ Here, we implemented exact matching, which only works


when there are control units with identical covariate values —
a condition that often requires a rich dataset

▶ Implicit in our example, we see another important assumption


needed for the matching procedure to work

16 / 45
Common Support Assumption

▶ Matching relies on the availability of similar individuals across


treatment and control groups

▶ What if some treated individuals have no comparable


counterparts in the control group?

▶ Some treated individuals will have no comparable matches,


making it impossible to estimate their counterfactual
outcomes

▶ This brings us to an important requirement for matching to


work

17 / 45
Common Support Assumption

Common Support Assumption


There must be sufficient overlap in the distribution of observed
covariates between treated and control groups

0 < P (T = 1|X ) < 1

▶ For any combination of covariates X , there must be both


treated and control individuals
▶ This assumption ensures that each treated unit can be
matched to a control unit (and vice versa)
▶ If some treated individuals have covariate values not observed
in the control group, we cannot estimate their counterfactual
outcomes

18 / 45
Common Support Assumption

Density

0 0.2 0.4 0.6 0.8


Propensity Score
Treatment Control

19 / 45
Common Support Assumption

1
0.8
Common Support Region
Density

0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Propensity Score

Treatment Control

20 / 45
Common Support Assumption

▶ If only a few units fall outside the common support, we can


restrict the sample to the region of common support (also
known as trimming)
▶ However, this means the estimated treatment effect will only
generalize to treated individuals within the common support,
not to the entire treated population

▶ If a large proportion of treated individuals lack comparable


counterparts, matching may not be appropriate, and
alternative methods should be considered

▶ Therefore, it is crucial to check for common support before


implementing matching procedures

21 / 45
Summary

▶ The matching method offers a way to approximate the missing


counterfactual outcome by using control individuals who share
similar observed characteristics with treated individuals
▶ The conditional independence assumption (CIA) implies that,
once we control for observed covariates, treatment assignment
is as good as random conditional on observables
▶ Under the CIA, control individuals with the same covariate
values as treated individuals can serve as valid counterfactuals
▶ The common support assumption ensures that for each
treated individual, there exists at least one comparable control
unit with the same covariate values
▶ However, in practice, there are important considerations and
limitations when implementing matching methods

22 / 45
Regression vs Matching under the CIA

▶ So far, we’ve focused on matching—but why not just use a


regression model that controls for observable covariates?

▶ For a regression model to yield an unbiased estimate of the


treatment effect, it implicitly assumes correct model
specification (i.e., the functional form is correctly specified
and no key variables are omitted)

▶ Matching has an advantage in this regard: it is a


non-parametric method that directly compares the outcomes
of treated individuals with matched controls, and is therefore
less reliant on functional form assumptions

23 / 45
Regression vs Matching under the CIA
Regression
▶ Uses all available data to estimate the relationship between
treatment and outcome, controlling for covariates
▶ Implicitly imposes a functional form (e.g., linearity)
▶ Can extrapolate beyond the region of common support
▶ Can be sensitive to model specification

Matching
▶ Constructs treatment effect by comparing treated units with
control units that have similar covariate values
▶ No functional form assumption
▶ Avoids extrapolation by restricting comparisons to comparable
observations
▶ Reduces bias from model misspecification, but may suffer
from lower efficiency and limited matches
24 / 45
Practical Issues in Matching
Continuous variables
▶ It is often difficult to find a control unit with exactly the same
value on continuous covariates (e.g., height, weight)
▶ Possible solutions include:
▶ Categorizing continuous variables
▶ Matching on the closest value (e.g., nearest neighbor)
▶ These approaches are more flexible but may result in lower
matching quality

Sample Size
▶ In small samples, it may be difficult to find suitable matches
for treated units
▶ Fewer observations reduce the likelihood of finding
comparable controls

25 / 45
Practical Issues in Matching

Common Support
▶ For matching to work, each treated individual must have at
least one comparable control individuals
▶ Lack of common support can lead to loss of generalizability or
bias

Number of Covariates
▶ As the number of covariates increases, it becomes harder to
find exact or close matches
▶ This is known as the “curse of dimensionality”

26 / 45
Propensity Score Matching
How to deal with the curse of dimensionality?
▶ Matching on many covariates directly becomes increasingly
difficult as the number of covariates grows
▶ Can we reduce the dimensionality of the matching problem
without losing key information?

Propensity Score Matching


▶ Rosenbaum and Rubin (1983) proposed using the propensity
score as a balancing score
▶ Matching on the propensity score allows us to achieve
covariate balance in expectation without matching on every
covariate directly
▶ This reduces the problem to one dimension, making matching
more feasible and efficient

27 / 45
Propensity Score Matching

Propensity Score
The propensity score is the conditional probability of receiving
treatment given a set of observed covariates

P(Ti = 1|Xi ) ≡ P(Xi ) ≡ Pi

▶ The central idea is to condense all the information contained


in Xi into a single scalar: the propensity score
▶ Propensity scores are estimated using a binary choice model
(since treatment status is binary: treated vs. not treated)
▶ Common models: logit or probit
▶ These models are used when the dependent variable is binary
▶ Details of these models are beyond the scope of this course

28 / 45
Propensity Score Matching

Conditional Independence Assumption (CIA)


If treatment assignment Ti is independent of potential outcomes
(Y1i , Y0i ) conditional on observed covariates Xi , then Ti is
independent of potential outcomes conditional on the propensity
score Pi

(Y1i , Y0i ) ⊥
⊥ Ti |Xi ⇒ (Y1i , Y0i ) ⊥
⊥ Ti |P(Xi )

▶ This implies that conditioning on the single index—the


propensity score P(Xi )—is sufficient to establish independence
between treatment assignment and the potential outcomes
▶ It addresses the dimensionality problem by reducing the
high-dimensional covariate space to a single scalar value

29 / 45
Propensity Score Matching

Validity of Conditional Independence Assumption


▶ After matching on the propensity score, we should check
whether the distribution of covariates is similar between
treated and matched control groups
▶ As discussed earlier, in practice, we usually check the sample
means between the two groups to investigate the covariate
balance
▶ Insignificant differences in covariates between the treatment
group and the matched control group provide indirect evidence
that the conditional independence assumption is likely to hold
▶ Note that we can only provide indirect evidence, as identifying
assumptions are not directly testable

30 / 45
Propensity Score Matching

Common Support Assumption


There must be sufficient overlap in the distribution of propensity
score between treated and control groups

0 < P (T = 1|X ) < 1

▶ For each value of propensity score among treated individuals,


there exists at least one comparable control invididuals with a
similar value of the propensity score
▶ This assumption ensures that each treated unit can be
matched to a control unit (and vice versa)

31 / 45
Propensity Score Matching

▶ There are several ways to estimate treatment effects with


propensity scores
▶ Matching on the Propensity Score
▶ Stratification
▶ Inverse Probability Weighting
▶ Covariate Adjustment

▶ Here, I will only cover matching on the propensity score


▶ Once the propensity scores are estimated, treated and control
individuals with similar scores can be matched
▶ The next step involves selecting a matching algorithm to
identify suitable matches

32 / 45
Common Matching Algorithms
Nearest Neighbor Matching (will only cover this)
▶ Each treated unit is matched to the control unit with the
closest propensity score

Caliper Matching
▶ Matches are only made if the difference in scores is within a
pre-specified range

Radius Matching
▶ Treated units are matched to all controls within a certain
radius of the propensity score

Kernel or Local Linear Matching


▶ Weighted average of all control units, with weights decreasing
with distance in the propensity score
33 / 45
Nearest Neighbor Matching

▶ For each treated unit, find one (or more) control unit(s) with
the closest propensity score
▶ These matched control units are used to estimate the
counterfactual outcome for each treated unit
▶ Average Treatment Effect on the Treated (ATT)

[= 1
X 
ATT Yi − Yj(i)
NT
i∈{Ti =1}

▶ Yi is the outcome for treated unit i


▶ Yj(i) is the outcome for the nearest matched control unit to i
▶ The calculation of standard error of the ATT is more
complicated and is omitted here

34 / 45
The Effect of Exercise on BMI

▶ Obesity and overweight are major public health challenges


linked to chronic conditions such as diabetes, heart disease,
and certain cancers

▶ Regular physical activity is widely recommended as a way to


reduce or maintain Body Mass Index (BMI)

▶ Understanding the causal effect of exercise on BMI can help


inform public health campaigns and school and workplace
fitness initiatives

▶ Here, we will use real-life data to investigate the effect of


exercise on BMI

35 / 45
The Effect of Exercise on BMI
▶ Research question: Does regular exercise reduce BMI?
▶ Treatment: Whether or not an individual participate any
physical activity
▶ Outcome: BMI
▶ Data: 1990 Behavioral Risk Factor Surveillance System

Challenge
▶ People who exercise regularly may differ in many ways from
those who don’t (e.g., education, health)
▶ These differences make it challenging to identify the causal
effect of exercise on BMI
▶ To address this issue, we control for relevant observable
covariates in the analysis
▶ For simplicity, the analysis will include the following
covariates: Age, Male, White, Married, College, and Employed
36 / 45
Sample Means by Treatment Status Before Matching
Variable Treatment Control Difference
Age 43.315 49.226 -5.911***
Male 0.443 0.410 0.032***
White 0.861 0.790 0.071***
Married 0.572 0.546 0.026***
College 0.516 0.299 0.217***
Employed 0.658 0.564 0.095***
Observations 55,287 23,402

▶ Individuals who exercise are systematically and significantly


different from those who do not exercise
▶ For example, individuals who exercise tend to be younger and
more educated
▶ These differences motivate the use of Propensity Score
Matching (PSM), as it provides a way to adjust for observable
differences between individuals who do and do not exercise
37 / 45
Nearest Neighbor Matching

▶ We implement one-to-one nearest neighbor matching

▶ For each individual who exercises, we find one individual from


the non-exercising group with the closest propensity score
▶ Matching is performed with replacement

▶ Before presenting the estimated treatment effects, we first


assess the validity of key identifying assumptions:
▶ Conditional Independence Assumption (CIA)
▶ Common Support Assumption

38 / 45
Sample Means by Treatment Status After Matching
Variable Treatment Control Difference
Age 43.315 43.301 0.014
Male 0.443 0.439 0.003
White 0.861 0.863 -0.003
Married 0.572 0.572 0.000
College 0.516 0.514 0.002
Employed 0.658 0.660 -0.001
Observations 55,287 55,287
▶ After matching, the differences in covariates between the two
groups become very small and statistically insignificant
▶ This indicates that covariate balance is achieved and suggests
that the Conditional Independence Assumption (CIA) likely
holds
▶ Individuals who exercise now have similar observed
characteristics to their matched counterparts who do not
exercise
39 / 45
Common Support Assumption

4
Density

Density
2

0
.2 .4 .6 .8 1 .2 .4 .6 .8 1
Propensity Score Propensity score

Treatment Control Treatment Control

▶ There is sufficient overlap in the density distribution of the


propensity score between individuals who exercise and those
who do not
▶ This suggests that the common support assumption likely
holds

40 / 45
Estimation results

Model 1 Model 2 Model 3


ATT -1.106*** -0.755*** -0.056
(0.048) (0.049) (0.230)
Standard errors in parenthesis

▶ Model 1: BMIi = α + θExercisei + ϵi


▶ This is equivalent to calculating the mean difference in BMI
between individuals who exercise and those who do not
▶ Model 2: BMIi = α + θExercisei + Xi β + ϵi
▶ Xi includes Age, Male, White, Married, College, and Employed
▶ Model 3: One-to-One nearest neighbor matching
▶ This represents the mean difference in BMI between individuals
who exercise and their matched counterparts who do not
exercise
▶ The ATT represents the estimated effect of exercise on BMI

41 / 45
Estimation results
Model 1 Model 2 Model 3
ATT -1.106*** -0.755*** -0.056
(0.048) (0.049) (0.230)
Standard errors in parenthesis
▶ In Model 1, the estimated ATT suggests that individuals who
exercise have a 1.106 lower BMI, on average
▶ In Model 2, after controlling for observable covariates, the
estimated effect drops to -0.755, indicating that part of the
initial effect was driven by differences in observed
characteristics
▶ In Model 3, using one-to-one nearest neighbor matching, the
estimated effect further reduces to -0.056 and becomes
statistically insignificant
▶ This substantial reduction suggests that much of the
difference in BMI is explained by observable characteristics,
rather than the causal impact of exercise itself
42 / 45
Key Takeaways from Matching Methods
▶ In observational studies, random assignment of treatment is
typically not feasible
▶ A simple comparison of average outcomes between treated
and control groups is likely to suffer from selection bias
▶ Matching helps mitigate this bias by constructing the
counterfactual outcome for each treated unit using control
units with similar observed characteristics
▶ The Conditional Independence Assumption is essential—it
states that, conditional on observed covariates, treatment
assignment is independent of potential outcomes and
effectively random
▶ The Common Support Assumption requires sufficient overlap
in the distribution of covariates between treated and control
groups, ensuring that each treated unit has comparable
control units for matching

43 / 45
Key Takeaways from Matching Methods

▶ Exact Matching
▶ Matches treated and control units with exactly the same
values of observed covariates
▶ Conceptually simple and intuitive
▶ Often impractical with many covariates or continuous variables
(curse of dimensionality)

▶ Propensity Score Matching (PSM)


▶ Matches units based on their estimated probability of receiving
treatment, given observed covariates
▶ Reduces multidimensional covariate space to a single score
▶ More feasible with high-dimensional data

44 / 45
Limitations and When to Be Cautious

▶ Matching can only account for observed confounders


▶ It relies heavily on the Conditional Independence Assumption
▶ This is a very strong assumption, and in many real-world
settings, unobserved confounders that affect both the
outcome and the treatment may still exist
▶ It is often difficult to convincingly argue that all relevant
confounders have been observed and properly controlled for
▶ Therefore, matching should be seen as a last resort, used
when better research designs are not available

45 / 45

You might also like