Skip to content

CONCEPT Cited by 1 source

Market-mediated long-term effects

Definition

In a multi-sided marketplace (riders + drivers, buyers + sellers, hosts + guests), an intervention on one side propagates through shared market state — prices, wait times, inventory, match rate, incentives — to affect the other side, and those affected users' future behaviour then feeds back into market state again. The resulting long-term effects route through the market, not directly through the targeted users. Lyft's 2026-03-25 post names these "market-mediated long-term effects."

Contrast with direct long-term effects: if Lyft raises driver incentives, drivers who receive them may drive more next week and return more often — that's a direct effect on the same drivers. But extra supply in the market also means shorter rider wait times, lower surge, better rider experience → riders more likely to return; and extra supply means each driver is less busy → some drivers less likely to return. Those are market-mediated effects operating through different users than the one holding the intervention.

Why market mediation breaks naive A/B testing

A standard user-split A/B assigns users randomly to treatment or control and assumes SUTVA (Stable Unit Treatment Value Assumption): a user's outcome depends only on their own treatment, not on others'. Market mediation violates SUTVA by construction — treatment and control users share the same market and therefore affect each other through shared supply/demand, pricing, queues. Any estimator that assumes treatment and control are independent underestimates the market-mediated effect and biases the direct estimate toward zero (if the mediation is counterbalancing) or exaggerates it (if reinforcing).

Why market mediation is the hard part of LTE measurement

Direct long-term effects can be recovered with user-split experiments run for long enough windows (at a cost — drift, attrition, opportunity cost, business pressure). Market-mediated long-term effects fundamentally require a whole-market shock — user splits cannot see them because the market response itself is shared. That constraint drives the push to region-split experiments (applying the shock to entire markets and using other markets as control) or observational approaches with an explicit mediation model — e.g. surrogacy.

Examples from the Lyft post

  • Driver incentive increase — week 1: supply up (direct effect on drivers), surge down + wait time down (market state), rider experience up → rider retention up (market-mediated to riders). But also: each driver less busy → some driver retention down (market-mediated back to drivers). The net multi-week effect is the composition of all four channels.
  • Rider price increase — week 1: demand down (direct effect on riders), supply/demand ratio up → each driver busier → driver earnings up → driver retention up (market-mediated to drivers). Rider retention may fall from price + from missing rides during temporary demand-supply rebalancing.

Seen in

  • Lyft — Beyond A/B Testing (2026-03-25) — canonical wiki framing. The post's entire motivating argument is that estimating market-mediated long-term effects is the central methodology problem for resource-allocation decisions in Lyft's marketplace, and that classical user-split A/B cannot do it. Lyft's answer is the two-step surrogacy framework, with region-split experiments as the end-to-end ground truth.
Last updated · 319 distilled / 1,201 read