M1-084M1: Mathematics & Statistics Fundamentals实验设计 (A/B)Hard
⚡Flashcard Drill➔Mastery:
Explain Switchback Experiments and Their Applications in Marketplace and Two-Sided Delivery Platforms
🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
📐 Mathematical Definition
⚡ One-Sentence Core Answer
A switchback experiment alternates treatment and control policies sequentially over time intervals across geographical regions, overcoming two-sided marketplace interference where users and drivers share a single supply pool.
📌 Key Takeaways
- •Core Motivation: In ride-sharing (Uber, Lyft, DoorDash), treating passenger A cannibalizes driver supply for passenger B (violating SUTVA).
- •Design Structure: Divides physical space into clusters (cities/zones) and time into windows (e.g. 30-minute intervals); each window randomly switches between Treatment and Control.
- •Washout Periods: Discards data from the first 5-10 minutes of each window to allow supply-demand system dynamics to reset.
📐 Mathematical Derivations
Unit of randomization is the Cluster-Time Window rather than individual user . The regression model is: , where is region fixed effect and is time-of-day/day-of-week fixed effect. Standard errors must be clustered at the Cluster-Time level (or Region level via Block Bootstrap) to account for strong temporal autocorrelation across adjacent time intervals: . Naive user-level standard error calculation results in false positive rates exceeding 50% due to unmodeled intra-window clustering.
🏭 Production Trade-offs
Switchback design trade-offs: (1) Window length: Short windows (15 min) increase effective sample count but suffer from severe carryover bias (drivers dispatched under Treatment algorithm carry over into Control window). Long windows (2 hours) eliminate carryover but reduce sample size and power. (2) Washout periods: Discarding the initial 10 minutes of each window cleans carryover effects at the expense of lost data throughput.
⚠️ Common Interview Pitfalls
- ✕Evaluating user-level p-values directly in a switchback experiment (massive Type I error inflation; must aggregate to the window level).
- ✕Ignoring carryover effects from past windows without incorporating washout buffers.
🎯 Interviewer Follow-up Inquiries
- ?How does carryover bias distort treatment effect estimation in switchback designs?
- ?Why must standard errors in switchback experiments be clustered at the geographic region level?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
📚