Social science
How to check parallel trends in difference-in-differences
Plot pre-treatment outcomes, then inspect the size and intervals of pre-treatment estimates. The post-treatment counterfactual trend also needs a defensible comparison group and institutional evidence.
Gather unit identifiers, periods, first treatment dates, the outcome definition, and sample rules. Record announcement dates, implementation dates, and concurrent shocks. Changes in sample composition can alter the trend plot.
Parallel trends concerns an untreated counterfactual after treatment begins. Pre-treatment estimates help flag issues. Report wide intervals or limited power when they matter.
How to do it
Define treatment timing and comparison units
Mark the first treated period for each unit. State whether comparisons use never-treated or not-yet-treated units or another design. For staggered adoption, record cohorts and valid comparison periods.
Plot observed outcomes first
Report group means and sample counts by calendar time. Explain weighting, adjustments, and sample restrictions. Look for missing periods, outliers, and abrupt changes in sample composition.
Inspect pre-treatment estimates and intervals
State the reference period and plot estimates with uncertainty. Consider joint tests and substantively meaningful departures. If announcement precedes implementation, examine anticipation.
State assumptions and rival explanations
Check differential shocks and coverage. With staggered timing, choose an estimator appropriate to groups and periods, and explain effect aggregation and sensitivity checks.
Worked example
Synthetic example with a policy starting at period 0. Negative values denote pre-treatment periods. The entries are group outcome means.
| Period | Treated mean | Comparison mean |
|---|---|---|
| −3 | 48 | 46 |
| −2 | 50 | 48 |
| −1 | 53 | 51 |
| 0 | 59 | 53 |
The raw gap is 2 in each pre-treatment period and 6 at period 0. Formal inference still needs the sample definition, standard errors, anticipation assessment, and identification assumptions.
Common questions
Do insignificant pre-treatment coefficients prove parallel trends?
Read interval width and substantively plausible departures alongside institutional evidence. A test may have little power to detect departures.
What if I have one pre-treatment period?
Report baseline levels and design evidence. A single observation per group cannot establish a pre-treatment trajectory. Seek historical data or a design with a justified alternative.
Verified sources
The official sources below were checked on 2026-11-13. Access conditions and update dates are rechecked before publication.
RigorMed workspace
Audit my policy evaluation design
Bring authorized group and period summaries, treatment dates, and outcomes. Draft a timeline, pre-trend checks, and unresolved assumptions about comparison units.
