Observational Methods
Diff-in-Diff, IVs & Natural Experiments
Diff-in-diff: subtract twice
A feature launched only in region A. Comparing A-after to A-before confounds with time trends; comparing A-after to B-after confounds with region differences. Difference-in-differences subtracts both: (A_after − A_before) − (B_after − B_before) — the change in treated minus the change in control.
What it removes: any time-invariant difference between regions (A was always richer) and any common shock hitting both (the holiday week). What survives is the treatment effect — if the load-bearing assumption holds: parallel trends. Absent treatment, the regions would have moved together. Checkable in pre-periods (were they moving together before?), never provable for the post-period — the eternal honesty tax.
insight
Diff-in-diff is the workhorse of policy and product rollouts because it kills two confounders at once — the permanent difference and the shared trend. Its weakness is always the same single word: 'parallel'.