Researchers Propose SplitGraph Design For Spillovers
At a Harvard Data Science Initiative seminar, Panos Toulis and collaborators present a network-aware "SplitGraph" experimental design and Fisher randomization-based analysis to improve power for estimating spillover effects in interconnected settings. They apply the approach to a tax audit experiment covering roughly 0.5 million firms and 8.7 million transactions, highlighting a tuning parameter that prioritizes direct versus spillover estimands.
Key Points
- 1Introduce SplitGraph network-aware randomization that uses connectivity to assign treatments and improve power.
- 2Demonstrate Fisher randomization tests provide finite-sample, model-free inference under complex interference.
- 3Advise defining causal estimand upfront because optimal randomization depends on whether estimating direct or spillover.
Scoring Rationale
Significant methodological advance applied to a large real-world audit experiment; limitation is seminar-based single-source exposition without broader replication.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
