Statistics, experimentation, causal reasoning · insight-0027
Look for other events at the intervention date before attributing a break to your change
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- aviation, energy, public-sector
Summary: situation, action, reason, boundary
Situation
Fares were changed on a route group in March, and an interrupted time series shows a sharp drop in load factor in the same month.
Action
Build a timeline of known events around the intervention (competitor moves, calendar changes, system releases). Add a control series that was exposed to the other events but not to your change (controlled ITS), and test placebo break dates.
Reason
ITS handles slow confounders through the trend but has no defence against a confounder that changes at the same moment as the intervention.
Boundary
Any single-series before/after evaluation in operations, pricing or policy.
Symptom or risk
A coinciding event, such as a competitor adding capacity, a holiday shift or a data pipeline change, produces the break, and the team reverses a good decision.
Exceptions
Randomized rollouts with concurrent controls do not need this reconstruction.
Trade-offs
Suitable control series may be imperfect; their differences need justification.
Pitfalls
Using a control series that the intervention itself affects, such as a neighbouring route that absorbs diverted passengers.
Example
Synthetic airline data: competitor capacity on IST-ADB, IST-AYT and IST-FRA steps up 25% on 2024-03-01, so a March 2024 fare change on those routes is confounded by design.
Lessons and concepts
Claims and their sources
Interrupted time series is relatively robust to slowly changing confounders but events coinciding with the intervention can bias it; controlled ITS and other design extensions help.