Course outline

Switchback Experiments

By the end of this lesson, you should be able to: recognise when a user-level split measures the wrong quantity rather than a noisy version of the right one, design a switchback that survives time-of-day, handle carryover, and analyse one at the right unit.

A change you cannot split by user

Meridian smooths surge pricing. The planted effect is a 6.2% reduction in ETA.

Split riders 50/50 and measure. The result:

DesignMeasuredVersus truth
Switchback, whole city on and off+7.04%114%
User-level A/B+3.47%56%
Planted truth+6.20%

The A/B test found a bit more than half the effect. The natural read is that it's underpowered, so run it bigger:

RidersMeasured95% CI
5,000+3.83%[+2.63%, +5.03%]
20,000+3.78%[+3.18%, +4.39%]
80,000+3.47%[+3.17%, +3.78%]
400,000+3.64%[+3.34%, +3.95%]

The interval tightens around 3.47%, not around 6.20%. More data buys precision on the wrong number.

That's the distinction the whole lesson turns on. The A/B test isn't noisy, it's biased, and no sample size fixes a bias.

Why the control group isn't untreated

Surge smoothing works by changing how drivers are allocated. Smooth it for treated riders and drivers reposition. Those repositioned drivers then pick up control riders too.

So control improves. The measured gap is the difference between two improved groups rather than between treated and untreated, and it comes out at roughly half the truth.

This is interference, and it's structural rather than accidental: any change that acts on a shared resource violates the assumption that one user's assignment doesn't affect another's outcome. Supply allocation, pricing, ranking inventory, matching, and notification volume all qualify.

The fix is to stop trying to have treated and untreated users at the same time. Turn the whole city on and off, alternating over time, and compare the on windows to the off windows. Now there's no untreated rider sharing a driver pool with a treated one, because in any given hour everyone is on the same setting.

Two panels. Left, a step line of ETA in minutes across the first 72 hours of the experiment with alternating shaded indigo blocks marking treated windows, showing a strong repeating daily cycle much larger than the difference between shaded and unshaded stretches. Right, two bars, a teal switchback at plus 7.04% and an amber user-level A/B at plus 3.47%, against a dashed rose line marking the planted 6.2%.
Left: the daily cycle dwarfs the effect, which is what the design has to survive. Right: only one of the two designs is measuring the change.