My Starter Project on the Lyft Rider Data Science Team

This post describes a Lyft data scientist's starter project using the Rider Experience Score (RES) tool to estimate long-term causal effects of rider experiences on retention without relying on A/B tests.

• •RES quantifies how specific rider experiences (e.g., low ETA, early driver arrival) impact long-term ride counts, addressing the limitation that A/B tests typically run only 2–6 weeks
• •Naive difference-in-means estimation on observational data produces biased results due to confounders such as geographic region (Simpson's Paradox illustrated)
• •RES employs the Augmented Inverse Propensity Score Weighting (AIPW) estimator, a double ML method with Neyman Orthogonality and doubly robust statistical properties
• •Three XGBoost/LightGBM models estimate propensity scores and potential outcome functions for treatment and control, combined with cross-fitting for unbiased ATE estimation