Events

This paper addresses a critical challenge in e-commerce experimentation: the bias in A/B testing caused by product stockouts. When products have limited inventory and may become unavailable during experiments, standard randomized controlled trial (RCT) estimators systematically overestimate treatment effects that are sign consistent across products. This occurs because stockouts create dynamic interference between treatment and control groups, violating the Stable Unit Treatment Value Assumption (SUTVA). We develop a fluid limit analysis to characterize this bias and propose a novel Stockout-Discounted (SD) estimator that accounts for demand transfer patterns when products become unavailable. By applying appropriate discount weights based on product substitution behaviors, our estimator reduces bias while maintaining equal or lower variance compared to standard RCT approaches. We establish theoretical guarantees for our approach under general choice models without assuming any specific underlying customer choice behavior. Our findings provide practical guidance for platforms seeking accurate treatment effect estimation in inventory-constrained environments, where traditional A/B testing methods fail to account for the dynamic nature of product availability.