Introduction to Difference-in-Differences
M Rahul
Institute of Economic Growth, Delhi
25 September 2025
The counterfactual
▶ How do we measure what would
have happened if the other
circumstance had prevailed?
Difference-in-Differences
▶ Quasi-experimental approach to estimate causal effects.
▶ Compare changes over time between treated and control groups.
▶ Useful when randomisation isn’t feasible but a clear policy change occurs.
The treatment is applied at a specific point in time
Parallel Trends Assumption
▶ In absence of treatment, treated and control groups follow the same trend.
▶ Core identifying assumption of DiD.
▶ Not directly testable, but we can use pre-trend plots etc.
Given parallel trends assumption, we can identify the Average Treatment Effect on the
Treated (ATT)
Two-Period, Two-Group Estimator
▶ ATT:
τ̂DiD = (ȲT ,post − ȲT ,pre ) − (ȲC ,post − ȲC ,pre )
▶ Difference of changes: outcome change in treated minus outcome change in
control.
Effect of minimum wage on employment (Card and
Krueger, 1994)
▶ On April 1st, 1992, New Jersey
raised minimum wage from $4.25
to $5.05.
▶ Surveyed 410 fast food
restaurants in New Jersey and
Pennsylvania, before after the
change.
▶ Pennsylvania restaurants: control
group.
Source: Card and Krueger (2000)
OLS Version of DiD
▶ Model:
Yit = α + δTreati + λPostt + τ (Treati × Postt ) + εit
▶ τ gives ATT under parallel trends.
▶ Easy to extend to multiple periods and covariates.
Yit = α + δTreati + λPostt + τ (Treati × Postt ) + εit
Before After After - Before
Control group α α+λ λ
Treatment group α+δ α+δ+λ+τ λ+τ
Treatment - Control δ δ+τ τ
Event Study Designs
▶ Leads and lags of treatment indicator.
Yit = αi + λt + −1 τ =−q γτ Dτ t + τ =0 δτ Dτ t + xit + εit
P Pm
▶ Graph dynamic treatment effects and check pre-trends.
▶ Useful to visualise timing and persistence of effects.
Covariates in DiD
▶ Can improve precision and reduce residual variance.
▶ Avoid “bad controls” affected by treatment.
Staggered Adoption (Multiple Periods)
▶ Groups receive treatment at different times (“rollout” designs).
▶ Most commonly used: Two Way Fixed Effects:
yit = α0 + δDit + Xit + αi + αt + ϵit
Problems with TWFE
TWFE estimator may be biased when effects vary over time or across groups.
“Forbidden comparisons”
New estimators: Callaway & Sant’Anna (2020), Sun & Abraham (2020), etc.
Key Takeaways
▶ DiD is powerful but hinges on parallel trends.
▶ Always check pre-trends, placebo tests, falsification exercises.
▶ Use updated methods for staggered adoption.
Thank you!
Garbage Incinerators and Home Prices
▶ A garbage incinerator was built in North Andover, Massachusetts
▶ The goal is to figure out what kind of impact the garbage incinerator had on
home prices over the course of two years.
▶ Data on prices of houses that sold in 1978 and another sample on those that sold
in 1981
