Institutes outside Greece
Estimation of treatment effects in matching marketplaces under interference
Abstract
dc:descriptionMarketplace companies routinely use randomized experiments to make operational decisions. These experiments can suffer from interference: treated and untreated units interact in ways that affect their outcome. In this thesis, we focus on mitigating interference bias in experiments conducted in matching marketplaces, where the platform actively matches supply and demand units via a linear programming algorithm. In the first part of this study, we examine pricing interventions, where a platform experiments with new base pricing levels at the marketplace level. In a matching marketplace, this type of experiment raises a crucial design question: should the platform match treated and untreated units differently because they paid different prices? We find that standard estimation techniques are biased, but the sign of this bias depends strongly on this design choice. Bias can be reduced using the "shadow price" estimator, which relies on the optimal dual solution of the platform’s supply-demand matching problem --- particularly when the platform ignores pricing differences at matching time. We validate our findings both theoretically, in a fluid limit setting, and numerically, in a finite-sample setting. In the second part of this study, we focus on marketplace experiments that are used to determine whether a specific demand-boosting intervention should be rolled out. Since rollouts can be global (affecting all users) or partial (targeting specific users), it is essential to measure the intervention’s effect on different sub-populations, a process known as heterogeneous treatment effect analysis. We first establish that heterogeneous treatment effects (HTEs) may not be uniquely defined in the presence of marketplace interference. We then distinguish two cases: when the HTE on a subset of users is uniquely defined, we show that the shadow price estimator can provably reduce bias from the standard estimator. When the HTE is not uniquely defined, we recast the problem of estimating treatment effects for each user group as a decision-making problem --- determining whether or not to roll out the treatment to each group. We use the shadow price estimator to design a robust decision rule, which we demonstrate to be more effective and less conservative than a comparable decision rule based on the standard estimator. Finally, we verify our theoretical results through numerical experiments on synthetic data. In the third part of this study, we introduce a principled method that leverages information from a single experiment to derive tight bounds on treatment effects, both for global treatment effects (GTEs) and heterogeneous treatment effects (HTEs). Our theoretical analysis demonstrates that the newly introduced "Bounds" estimator (BE) achieves comparable bias to the Shadow Price (SP) and standard approaches. Empirically, we show that while the "Bounds" estimator may exhibit higher bias in certain regions of the parameter space, it also exhibits substantially lower variance, leading to an overall lower root mean squared error (RMSE). Our analysis expands the set of available estimators for practitioners, who can more effectively trade off estimation bias and variance based on their needs.
Degree
thesis:*- Grantor dc:publisher
- Institutes outside Greece
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Karakolios, Kleanthis
- Καρακόλιος, Κλεάνθης
Subjects
dc:subject × 48- Αιτιακή συμπερασματολογία
- Τυχαιοποιημένα πειράματα
- Δοκιμή Α/Β
- Αποτελέσματα παρεμβάσεων
- Πλατφόρμες αγοράς
- Αγορές αντιστοίχισης
- Πειραματικός σχεδιασμός
- Παρεμβολή
- Καθολικές επιδράσεις παρέμβασης
- Ετερογενείς επιδράσεις παρέμβασης
- Εκτίμηση αποτελεσμάτων παρέμβασης
- Παρεμβάσεις τιμολόγησης
- Πειραματισμός
- Αμφίπλευρες αγορές
- Αντισταθμιστική σχέση μεροληψίας και διακύμανσης
- Μεροληψία λόγω αλληλεπίδρασης
- Causal inference
- Randomized experiments
- A/B testing
- Treatment effects
- Marketplaces
- Matching markets
- Experimental design
- Interference
- Global treatment effects
- Heterogeneous treatment effects
- Treatment effect estimation
- Pricing Interventions
- Experimentation
- Two-sided markets
- Bias-variance tradeoff
- Interference bias
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Αυτοματισμοί και Συστήματα ελέγχου
- Οικονομικά και Επιχειρήσεις
- Κοινωνικές Επιστήμες
- Διοικητική Επιστήμη και Επιχειρησιακή Έρευνα
- Στατιστικές, Πιθανότητες και Αβεβαιότητα
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
- Automation and control systems
- Economics and Business
- Social Sciences
- Management Science and Operations Research
- Statistics, Probability and Uncertainty
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59163