A difference-in-differences readout shows a big effect, but the pre-period trends were already diverging — diagnose it.
You ran a difference-in-differences analysis on a single-city pricing launch. The estimate is large and significant, but before trusting it you plot the pre-launch weeks — the treatment city was already pulling ahead of control before week 0.
Constraint: judge only from the pre-period trend shown; do not assume the launch worked.
Weekly revenue per user (₽), launch at week 0
week treatment control
-4 100 100
-3 104 99
-2 109 98
-1 114 97
1 128 96
2 133 95Diagnose what is broken and state what you would report.
Parallel trends is violated — treatment was already diverging upward from control before week 0. Diff-in-diff charges that pre-existing gap to the launch, so most of the 'effect' is bias, not impact. Report that the estimate is not credible — you need a control with matching pre-trends before claiming any effect.
- ✗Reading a diverging pre-trend as treatment momentum
- ✗Believing more control units fix a parallel-trends violation
- ✗Reporting the diff-in-diff estimate despite a failed pre-trend check
- →What method could still recover an effect when pre-trends are not parallel?
- →How would an event-study plot quantify the pre-trend violation?
Split the gap into two parts. Before launch (weeks −4…−1) treatment rises 100 → 114 while control falls 100 → 97 — they already diverge by roughly 4–5 ₽ per week. So the two series were not moving in parallel before any treatment. Diff-in-diff takes the difference of differences and charges the whole post-launch gap to the treatment, but part of that gap is just the continuation of a pre-existing trend — bias from an unobserved factor that separates the cities.
Contribution to the "effect" (treatment − control):
week −4: 0 ← started level
week −1: 17 ← gap already exists pre-launch (this is NOT the effect)
week 2: 38 ← naive DiD counts almost all of this as effectWhat to report: the estimate is not credible — the parallel-trends assumption is broken. Do not report the number as a causal effect. You need a control with a matching pre-trend (matched cities or a synthetic control), or an event-study showing the gap opens exactly at week 0, not before.
