🌍 Geo-Lift Driver Incentives: Cluster Randomization Click to visit Live App:
Figure 1: Placebo Test Validation
Figure 2: Difference-in-Differences Analysis (Parallel Trends)
📌 Project Overview
This project demonstrates how to run experimentation in a Two-Sided Marketplace when individual user randomization is impossible due to Network Interference.
The Problem: Why not User-Level A/B?
If we give a "Bonus" to Driver A in a city but not Driver B:
Driver A works more and takes all the rides.
Driver B gets fewer rides because of Driver A.
The Control Group (Driver B) is negatively impacted, making the Treatment look better than it actually is (Spillover Bias).
The Solution: Geo-Lift (Cluster Randomization)
We randomize by City.
Treatment: Tallinn, Vilnius (Drivers get Bonus)
Control: Riga, Tartu (No Bonus)
Method: We use Difference-in-Differences (DiD) to calculate the lift. We compare the change in Driver Hours in Treatment cities vs. the change in Control cities.
🛠️ Tech Stack
Language: Python 3.9+
Backend: FastAPI (Simulation of 4 cities & DiD Calc)
Frontend: Streamlit + Plotly (Map visualizations)
Analysis: Causal Inference (DiD Estimator)
🚀 How to Run Locally
- Setup Environment
python -m venv venv
.\venv\Scripts\activate
source venv/bin/activate
- Install Dependencies
pip install -r requirements.txt
- Run the Backend (Port 8001)
uvicorn backend.main:app --reload --port 8001
- Run the Frontend
streamlit run frontend/app.py
📊 Key Concepts
Parallel Trends Assumption: The assumption that without the bonus, Treatment and Control cities would have moved in parallel. Validated by the pre-experiment data in the line chart.
DiD Estimator: (Treatment_Post - Treatment_Pre) - (Control_Post - Control_Pre).