| Abstract [eng] |
A comprehensive simulation-based analysis of causal inference under interference is conducted for shared-inventory marketplace experiments, focusing on the estimation of average treatment effects (ATE) under varying interaction dynamics and randomization schemes. Using the FINN.no slate dataset, three outcome-generating models are examined: a baseline independent click model restricted to first clicks, an independent click model allowing repeated interactions, and a repeated clicks model. Together, these models capture increasing levels of user interaction intensity, competition, and spillover effects. Bernoulli and cluster randomization schemes are evaluated across multiple treatment probabilities, and exposure mapping is employed to separate treatment effects into direct and interference components at the user level. The results show that more aggressive interaction dynamics reduce causal efficiency, as intensified competition diminishes realized treatment effects. Bernoulli randomization systematically overestimates treatment effects at low treatment probabilities, while cluster randomization substantially reduces this inflation but results in downward bias at higher treatment probabilities. Exposure mapping further reveals strong negative competitive spillovers and demonstrates that total treatment effects are substantially smaller than direct effects. This study contributes to the field by showing that ignoring interference in shared-inventory marketplaces can lead to treatment effect estimates that are highly precise yet systematically biased. Appropriate randomization strategies and explicit accounting for exposure are therefore essential for obtaining credible and actionable causal insights to support product development and user experience improvements in modern digital experimentation. |
