Quasi-experimental design
A quasi-experimental design estimates the effect of a change without randomly assigning people or groups to receive it. It uses comparisons across groups, time periods or eligibility rules to separate the change's effect from other influences.
How it works
In business, quasi-experiments help evaluate pricing, training or policy changes when random assignment is impractical. Unlike a randomized A/B test, the company does not assign participants to treatment and control groups by chance.
The hard part is choosing a credible comparison. Groups might differ before the change, or face different conditions afterward. A quasi-experimental design must address those alternative explanations, not merely compare two outcomes.
Separating a training effect from general sales growth
A company introduces sales training in one region but not another. Management chooses the region rather than assigning it randomly. These illustrative monthly revenues show a difference-in-differences calculation:
| Region | Before training | After training | Change |
|---|---|---|---|
| Training introduced | $100,000 | $130,000 | +$30,000 |
| No training | $100,000 | $110,000 | +$10,000 |
Estimated training effect: ($130,000 − $100,000) − ($110,000 − $100,000) = $20,000 per month.
The estimate is $20,000, not the full $30,000 increase. It relies on both regions having followed parallel revenue trends without training, with no spillovers or other region-specific changes that distort the comparison. Equal starting revenue alone does not establish that.
For a startup, this distinction helps prevent treating market-wide growth as evidence that a new sales process works.
Questions people ask
What is the best example of a quasi-experiment?
There is no single best example. A staged policy rollout is a useful one: compare changes in locations that adopt the policy with changes in locations that have not adopted it. Its credibility depends on how comparable those locations are.
How do you tell if a study is quasi-experimental?
Look for an intervention, no random assignment, and a comparison strategy intended to estimate its effect. Simply observing that customers who use a feature spend more does not establish a quasi-experiment: those customers might already be more engaged.
