Causal Inference and Market Experiment Design with MLSYNTH

Introduction

Jared Amani Greathouse
- PhD Candidate, Public Policy, Georgia State University
- Advisor: Jason Coupet
- Specialization: Econometrics and Causal Inference
- Focus: Synthetic Control Methods, High-Dimensional Panel Data, Machine Learning for Treatment Effects

My Work

• Developed mlsynth, the largest Python SCM package to date
  
• Suite of tools for policy evaluation using panel data
  
• Supports dozens of estimators: vanilla SCM, cluster SCM, factor models, PDA, proximal SCM, and more
  
• Handles complex, high-dimensional causal inference settings

What Makes MLSYNTH Special

• Consolidates decades of non-experimental SCM and panel methods from multiple software ecosystems into Python with one singular syntax
  
• Easy installation, automation, and scaling
  
• Supports a broad set of estimators for diverse research questions
  
• MAREX: Market Experiment Design feature for selecting treated and control units under constraints
  
• Decision-making tool: explore multiple scenarios with penalties, clusters, costs, budgets—overnight simulations possible

Challenges of Standard Experiments

• Randomized trials are often impractical at scale
  • High coordination and cost
    
  • Ethical or logistical constraints
    
• Simple randomization ignores pre-existing heterogeneity
  
• Cannot easily accommodate budgets, costs, or cluster structures
  
• Slow to iterate: each new design requires manual planning

Why Prefer MAREX

• Optimizes both treated and control selection simultaneously
  
• Enforces cluster structure, budgets, and cardinality constraints
  
• Penalizes units far from cluster means to ensure representativeness
  
• Allows overnight simulation of multiple experimental scenarios
  
• Treats synthetic control as a planning tool, not just an estimator

MAREX vs. Standard Experiment

Feature	Standard Experiment	MAREX
Validity in small/clustered marketing tests	Often compromised	Maintains validity via synthetic control optimization
Budget-aware	Sometimes requires ad hoc adjustments	Built-in into optimization
Cluster-aware	Possible but may reduce power	Enforced and balanced by design
Ethical / feasibility constraints	Manual adjustments	Integrated constraints
Scenario testing / iteration	Limited by randomization and manual setup	Overnight simulations with multiple configurations
Representativeness of treated/control	Needs careful design, can fail	Ensured via cluster-weighted synthetic controls

Takeaway

• MAREX is a planning tool for experiments where randomization is expensive, infeasible, or restricted
  
• Makes representative, feasible, and comparable treatment assignments
  
• Analysts can simulate, optimize, and justify experimental decisions before implementation
  
• Bridges econometrics theory and practical business experiment design