Title: Estimation of Treatment Effects in Matching Marketplaces under Interference

Date: Thursday, April 24th, 2025

Time: 11:00 am - 1:00 pm (EDT) 

Location: Groseclose 402

MS Teams: link

Meeting ID: 269 191 089 348

Passcode: Wx7DK9FU

Kleanthis Karakolios

Machine Learning PhD Student

School of Electrical and Computer Engineering

Georgia Institute of Technology

Committee

1 Dr. Arthur Delarue (ISyE, Advisor)

2 Dr. Xin Chen (ISyE)

3 Dr. Juba Ziani (ISyE)

4 Dr. Joel Sokol (ISyE)

5 Dr. Andre Calmon (Scheller CoB)

Abstract

This thesis develops methodological tools to mitigate interference bias in randomized experiments conducted within matching marketplaces. Interference occurs when treated and untreated units interact, violating standard causal assumptions and introducing bias into treatment effect estimates. The work addresses three key experimental settings: pricing interventions, heterogeneous treatment effect (HTE) estimation, and treatment effect bounding. First, the thesis analyzes pricing experiments and demonstrates that standard estimators are systematically biased, with the direction and magnitude of bias depending on whether treated and untreated units are matched differently. A shadow price-based estimator, which leverages dual variables from the platform’s matching algorithm, is shown to significantly reduce this bias. Second, in the context of HTE estimation, the thesis distinguishes between settings where the effect is uniquely defined and those where it is not. In the latter case, the problem is recast as a policy decision task, and shadow price-guided decision rules are proposed to improve rollout strategies. Finally, the thesis introduces a bounding framework that derives tight upper and lower bounds on treatment effects using a single experiment. The resulting point estimator, based on these bounds, achieves substantially lower variance and reduced root mean squared error relative to existing methods, offering a practical tool for managing the bias-variance trade-off in experimental evaluation.