Optimize Intervention Allocation in A/B Testing

This project is to gauge the individual treatment effect in randomized controlled trials (RCT) using Meta-Learner, a machine learning framework. And then provide advice for next round intervention allocation based on the pattern we found in individual effects. You can read my full description in thesis.pdf.

Affiliation: Institute for Social and Economic Research and Policy, Columbia University.

Keywords: Randomized Controlled Trails (RCT), Meta-Leaner

Software:

• Python : econml, linearmodels, statsmodels sklearn

Reference

• Paper Introduction to Meta-Learner
• Computation about meta-learner: EconML

Data and Scope

Note: Data sets are not posted due to privacy and intellectual property concerns. Special thanks to Dr. Don Green grants me the access to the data files.

• Exploratory analysis on three text-messaging voter mobilization experiments conducted before 2016 general election.

  Experiment	Treatment Group	Control Group	Script
  One Arizona Campaign	200,442	50,187	code/One Arizona Experiment.ipynb
  NextGen Climate Campaign	94,257	94,229	code/NextGen Climate Experiment.ipynb
  Vote.org	905,396	301,920	code/Vote Org Experiment.ipynb
  Pooled	1,019,697	446,336	code/Pooled Analysis.ipynb

• Estimate individual treatment by regression with interaction term

• Estimate individual treatment by meta-learner approach

Exploratory analysis

Regression

Meta-Learner