At Lyft, we run various randomized experiments to tackle different measurement needs. User-split experiments account for 90% of the randomized studies due to the higher power and fit for most use…

LyftEng's platform is a central hub for engineering insights and technology updates from Lyft's engineering team. Through articles, tech talks, and open-source contributions, LyftEng offers insights into engineering challenges, innovation projects, and best practices in software development. Readers can learn about Lyft's engineering culture, technology stack, and contributions to the broader engineering community.

Lyft Engineering

Interference bias, common in user-split experiments in multi-sided marketplaces like Lyft, can be mitigated using Marketplace Marginal Values (MMV). These values help adjust treatment effects by accounting for resource contention. While alternatives like cluster, time-split, and region-split randomization exist, each has limitations in power, cost, and applicability. MMV provides a powerful solution to interference, especially in network-effect-heavy environments, by leveraging dual values from optimization problems to correct estimated outcomes. Despite its benefits, MMV isn't universal and can't apply to all experimental scenarios.

Using Marketplace Marginal Values to Address Interference Bias

Productionalizing MMV in an Experimentation Platform