MATCHING METHODS

Matching covariates treated control groups

to make them
qsmile

two types 1 outcome not available Matching followup

cost saving

2 outcome Available Matching I bias

treatment

q effect

outcome values not used in MAKING

Can be done multiple times and most

balanced

Mattea atop

4 key steps for Marlena

Closeness where an individual is a good oath

Implementing it

given

assess and cerate 0 until a good motel

outrone matched samples

Analysis q given

0 Before Class

SUTVA: Stable Unit Treatment Value Assumption. Outcome of one individual is not a ected by
treatment assignment of another individual.

Poorly treatments us

defined

Effects of causes us

Sutor

Yi Jill Yi o

ti Y 1 8 Troi'D

Improving water quality and its impact on hospitilization. SUTVA violation because

Hand Washing Practices to Reduce the risk of getting diarrhea.

SUTVA isnt the worst thing. Just a better design. Explicitly incorporate that into your research design.

Use an interaction term or smth to that nature.

E ects of Causes vs Causes of E ects (Next lecture)

Intro to Matching:

Non-Experimental Research Design.

Example: What is the e ect of a windmill on farmland values. Nick Pates, Mark.

Windmill are not randomly assigned across states

Using a Matching Estimator: Why?

Matching Estimator usually makes sense when there are a lot more potential control units/observations
than treatment units/observations.

You only want the controls that are similar to the treatment cases.

In Economics: You have a better chance of nding a good match.

Other good thing about matching is: Permits pre-outcome analysis.

Balance Table in RCT:

We have the treatment and control Means:

List all the covariates and their means. In RCTs you want them both to be similar to one another.

With pre outcome analysis, you can select a sample and construct a balance table. And just look if
both look similar.

RCTs assess balance in covariate means between treatment and control units/groups

Matching estimators require us to assess likeness of covariate distribution between treatment and
control.

Show a balance table. in RCTs

Matching show a table but show p-value/t-test

xx

k
Same mean but di erent distribution.
k

Examples of Matching Estimators:

1. Exact Matching on X1, X2, X3 .. Xn

2. One to One Matching based on a distance metric

3. 1 to n matching based on a distance metric

Distance Metric: Euclidean Distance. Pick the variable, say age, pick the groups that are closest in age.

Can also de ne them in terms of age and experience.

Propensity Score Matching can also do this

Mahalanobis Distance: Takes into account weights which down-weights potential control units when
distance is farther in one or more dimensions.

1 to n: Composite control unit. Becomes like a weighted average.

Control

treats

if

Protector

low probably I

being in treetet

Bad

potential cartons


mrs

Dykema 4W ATE US ATOT

Statistical Signi cance: *** p-value

Homework: bysort miss: sum names of variables that you want to summarize

will not include summarize that has a missing value*

another way: sum apgar if miss == 1

sum apgar if miss == 0

For variables you observe, compare the characteristics of missing values vs non-missing values.

Just generate a summary stats table for which you have observations on all the participants. Compare
the characteristics.

*** Approach for assessing a paper/ writing a referee report: Prepare and structure

• Provide a summary

• Note to self:

• Main empirical objective

• E ective defense of research design

• E ective implementation entation of research design

• Presentation of results

• Writing quality

Comments on Introduction, motivation, research design, results

Summary: Impact of remittances on Welfare in a particular region of Nepal

Dataset size?

Main objective: Causal impact of remittances on welfare

ATE, ATOT.

ALWAYS DO PRE OUTCOME ANALYSIS WHEN DOING MATCHING/PROPENSITY SCORE

Dont start a sub section with a table. Always start with text before a table.

If two samples are not the same, then we can assume that they are going to be di erent across
unobservables which we can not control for thereby leading OLS to be biased.

Laxmi's Paper:

Objective

Pre outcome analysis: Highlights that we are trying to approximate an experiment (i.e an RCT)

Unconfoundedness Assumption:

ftp.Ditpzzit Ei

Yi d

regress Di Lot 4 Zi t 2

ataman
D unexplained
m
Assumption:

variation
error term of the second equation is unrelated to the error term of the mDi
rst equation.

if this were not hold, then there is something in treatment group that is making them getting the

Lil Yi
o
treatment.

Dit ti
That something will go to the error term of the rst equation.

2nd Assumption in the imbens paper: Overlap Assumption

een E wi nien Pr Wit alien

Overlap assumption

redid freebut
 lil

f treat

Treatment group has prob from 0.2 to 1

Control group has prob from 0 to 1

fothoot

The predicted probabilities are gonna lie between 0 and 1

ATE, ATT, ATC

pulpit

Banc Pins Ei

Di Pot Binet

Lhs is an estimate of the probability of being in treatment based on observable characteristics

ftp.niitpiniitpsns
x1, x2, x3

Di

The Ols estimator is called the linear probability model

Discussion of LPM:

Di Rot Fini t É

prob Diii

Gedn y

Because LPM generates negative or greater than 1 probabilities, we often prefer to estimate the
propensity score using a logit or probit estimator (both of which arex non-linear estimator).

Predicted probability is not going to vary linearly with the covariates.

propensity score is just a predicted probability (can use both linear and non-linear probability models)

Matching: To recover causal e ects. Embedded within the procedure is a predictive objective.

time
Prediction excercise: We don't care about the right hand side of the equation. Or biasedness. All we
care about is do we get predictive values that are close to the true value.

NOA models to predict hurricanes.

If linear model gives better prediction,

How to assess Balance:

to look at the treatment and control group for all the covariates:

pairwise comparison

2nd way: Look at each one of covariates and their distribution

I L


Read through page

t E