Matching & Subclassification
August - December 2025
Ananya Iyengar
(St. Stephen’s College) Quantitative Methods 2025 1 / 11
Selection Bias
What is a confounder?
1 Observable
2 Unobservable
In RCTs, randomisation creates covariate balance.
In observational studies, treatment assignment is a function of covariates.
(St. Stephen’s College) Quantitative Methods 2025 2 / 11
A Motivating Example
Did first class passengers have a lower probability of dying when the
Titanic sank?
D: First Class Seat, Y: Death
Women and children have a higher probability of being in first class.
Confounders: Gender and Age
What do we need to infer causal impact?
1 Given the confounder, potential outcomes are orthogonal to treatment.
That is, controlling for age and gender, whether or not a passenger was in
first class is as-good-as random.
2 This is called the Conditional Independence Assumption.
3 Moreover, what if only women and children were in first class? This will
not do: we need to covariate to exist for D=0 and D=1 i.e. Common
Support.
(From Mixtape.)
(St. Stephen’s College) Quantitative Methods 2025 3 / 11
Subclassification
1 Stratify the data (adult women, adult men, girl children, boy children).
2 Calculate the survival probabilities for each strata.
3 Calculate a weighted average of survival rates.
4 The weights are given by proportion of population in each strata.
Figure: The Average Treatment Effect (ATE) with subclassification
(St. Stephen’s College) Quantitative Methods 2025 4 / 11
Curse of Dimensionality
What if there are many covariates?
Stratification becomes complicated.
We may not have common support in each strata.
Need alternative ways of creating a counterfactual that controls for
confounders.
(St. Stephen’s College) Quantitative Methods 2025 5 / 11
Exact Matching
Find an untreated “twin” for every treated unit.
Suppose Age is the only confounder – find for every treated unit with a
particular age an untreated unit with the same age.
Can match with one or multiple untreated units.
Figure: Exact Matching: ATT
(St. Stephen’s College) Quantitative Methods 2025 6 / 11
Approximate Matching
What if there is no exact covariate match in the untreated group?
Use any distance metric d(.) to evaluate the nearest neighbour.
Maintain a cut-off of how close we want to distance to be.
Examples of d(.): Euclidean distance, Mahalanobis distance
(St. Stephen’s College) Quantitative Methods 2025 7 / 11
Propensity Scores
High dimensionality; how to create a counterfactual?
Revisit the Conditional Independence Assumption: covariates impact the
likelihood of treatment.
Propensity Score: Conditional likelihood of receiving treatment i.e.
p(D/X), where X is a vector of confounders.
Common support =⇒ 0 < p(D/X) < 1.
Compute the propensity score by regressing D on all Xs using
logit/probit (MLE methods).
Why not OLS?
(St. Stephen’s College) Quantitative Methods 2025 8 / 11
Given the propensity score:
1 Conditional p(x), treatment D is orthogonal to X.
2 Conditional on p(x), the distribution of X is the same for D = 1 and
D = 0.
(St. Stephen’s College) Quantitative Methods 2025 9 / 11
How do we use the propensity score?
1 Inverse Probability Weights: Weigh observations o the basis of the inverse
of their probability of being treated.
Figure: ATE with IPW
2 Approximate Matching on the Propensity Score: Using a distance metric
to evaluate the distance between propensity scores of treated and
untreated groups and create a counterfactual.
(St. Stephen’s College) Quantitative Methods 2025 10 / 11
Problems with using the propensity score
Discarding unmatched observations – loss of power, leaving out
information
After a point, trimming of units may increase covariate imbalance (King
& Nielsen, 2019)
Trade-off between balance and sample size.
(St. Stephen’s College) Quantitative Methods 2025 11 / 11