TechPrufer
Get started
Lyft
Lyft logoLyft · Feature trail

Quasi-Experimental Causal Inference Platform

2 posts · Feb 2026 – Jul 2026

What this trail covers

Lyft developed and continues to enhance its Quasi-Experimental Causal Inference Platform, which leverages Augmented Inverse Propensity Weighting (AIPW) models to estimate Average Treatment Effects (ATE) in scenarios where randomized experiments are not feasible. The platform emphasizes rigorous validation through mandatory confounder management, including pre-defined and customizable confounder sets with leakage prevention, and advanced diagnostics. Key features include propensity score correction, outcome reweighting to address downsampling bias, and a diagnostic scorecard with common support and covariate balance checks. The platform's performance is validated against A/B tests.

All Lyft trails
  1. Beyond A/B Testing: Using Surrogacy and Region-Splits to Measure Long-Term Effects in MarketplacesThis post introduces a two-step methodology for estimating market-mediated long-term effects in a marketplace. Step 1 models the impact of policy changes on negative user experiences using residualized regressions, controlling for time-of-week patterns and market information. This is validated with switch-back experiments. Step 2 models the impact of negative user experiences on future outcomes using observational, double-robust causal inference (AIPW), combining propensity and outcome models. This is validated with user-split experiments. The overall approach combines these estimates with direct long-term effects and verifies them using region-split experiments optimized by a forward selection algorithm.
  2. Trusting the Untestable: Validation and Diagnostics for the Doubly Robust ModelsThis post details the implementation and validation of Doubly Robust Models (specifically AIPW) within Lyft's Quasi-Experimental Causal Inference Platform. It introduces critical validation mechanisms including rigorous confounder management (pre-defined and customizable sets with leakage prevention), propensity score correction and outcome reweighting to address downsampling bias, and a diagnostic scorecard featuring common support (propensity overlap) and covariate balance checks. The post also describes empirical validation of the AIPW estimates against A/B tests using weekly ride challenges.