Prerequisites
Potential outcomes, difference-in-differences designs, clustered uncertainty and research ethics.
Use invented percentages of people with at least one event, counted once per period: program 30% to 20%, comparison 20% to 18%. The outcome and periods match. Difference in differences is (20 - 30) - (18 - 20) = -8 percentage points. A causal interpretation needs credible parallel untreated trends and other assumptions, not arithmetic alone.
Worked example
Both groups improved. Comparing only final values ignores their different starting points. Potential spillovers, changing composition, measurement shifts and concurrent programs can undermine the contrast. Use appropriate uncertainty estimation and participant protections.
Further inquiry
Write the group-time contrast and its counterfactual interpretation. Specify why parallel untreated trends, stable measurement and limited interference matter. Propose sensitivity checks for changing composition and spillovers, and explain why a pretrend check cannot prove the identifying assumption. Use fictional aggregates, not student health data.
Review criteria
- Keep percentage-point changes distinct from relative percentages.
- Separate identification assumptions from numerical calculation and uncertainty.
- Discuss privacy, essential access and limitations before any causal claim.