3. Policy Evaluation Designs

Difference-in-Differences

Use a comparison group’s change to estimate the treated group’s missing untreated trend

Difference-in-Differences

DID subtracts a counterfactual change

Clearborough introduces a clean-air zone; Westborough does not.

Mean pollution, μg/m³BeforeAfterChange
Clearborough4230−12
Westborough3834−4

The DID estimate is

ATT^=(12)(4)=8 μg/m³.\widehat{ATT}=(-12)-(-4)=-8\text{ μg/m³}.

Equivalently, Westborough’s −4 change predicts a no-policy Clearborough value of 424=3842-4=38; observed pollution is 30, eight units lower.

Py

Construct the DID counterfactual

Idle

The last lines turn a hidden trend violation into an explicit sensitivity parameter.

For two periods,

E[Y1(0)Y0(0)G=1]=E[Y1(0)Y0(0)G=0].E[Y_1(0)-Y_0(0)\mid G=1] =E[Y_1(0)-Y_0(0)\mid G=0].

The assumption allows different baseline levels. It concerns the treated group’s unobserved untreated change after policy, so it cannot be proved by pre-policy data.

Pre-trends, institutional history and alternative comparison groups can make it more or less credible.

Regression form reproduces the table

Yit=α+βGi+λPostt+τ(Gi×Postt)+εit.Y_{it}=\alpha+\beta G_i+\lambda Post_t +\tau(G_i\times Post_t)+\varepsilon_{it}.

In a 2×2 saturated model, τ^\hat\tau equals the four-cell DID. Controls or additional periods change the estimand unless specified carefully.

Four threats, four diagnostics

ThreatEvidence that addresses it
differential untreated trendmultiple pre-periods, institutional argument, sensitivity bounds
anticipationpolicy-announcement timeline and alternative treatment date
concurrent interventionpolicy inventory and unaffected outcomes
spillovergeography/market links and wider treatment definition
composition changestable units or repeated-cross-section composition audit

A placebo that does not target the principal threat is decorative robustness.

Repeated cross-section versus panel

Panel DID follows the same units. Repeated cross-sections compare population means sampled each period. The latter requires stable population composition or conditional assumptions; migration caused by the policy can change who is observed.

Inference should reflect treatment assignment and serial correlation. A policy assigned to municipalities usually requires municipality-level dependence, not resident-level independence.

Classic illustration, modern reading

Card and Krueger’s minimum-wage study compares New Jersey and Pennsylvania fast-food employment before and after a policy change. Use it to identify the four-cell logic, then ask how measurement, comparison choice, one treated policy and later evidence affect the claim.

Quick check

The treated and comparison groups have identical pre-policy levels but different pre-policy trends. Is DID credible?

Answer
Identical levels do not supply parallel trends. The diverging pre-policy paths undermine the comparison unless a credible conditional or structural explanation supports the post-policy counterfactual, with sensitivity analysis.

Next: Staggered DID and Event Studies

Copyright © 2026