Causal Inference and Decision Making on Digital Platforms

Abstract

This thesis examines advanced methodological approaches for addressing complex inference and decision-making challenges in digital and economic environments characterized by uncertainty, measurement error, and endogeneity. The development of robust methodological solutions in this domain is increasingly critical as organizations rely on algorithmic decision-making for high-stakes business decisions, platforms design experimentation systems for continuous optimization, and researchers leverage AI-generated variables for econometric analyses. Traditional methods often fail in these contexts, potentially leading to biased inferences, suboptimal decisions, and misguided strategies that can significantly impact business outcomes and research validity. Through three interconnected essays, this thesis demonstrates novel frameworks that enhance both theoretical understanding and practical applications in causal inference and decision making for online and offline settings, providing essential methodological tools for researchers and practitioners navigating the complexities of modern digital economics. The first essay introduces the $\epsilon$-BanditIV algorithm to address endogeneity issues in online dynamic decision-making contexts. By incorporating instrumental variables into the Multi-Armed Bandit (MAB) framework, this approach achieves both optimal regret minimization and asymptotically consistent parameter estimation, establishing a methodological foundation for unbiased causal inference in adaptive experimentation settings with endogenous covariates. The second essay addresses the critical challenge of measurement error in machine learning (ML) or artificial intelligence (AI) generated regressors within partially linear models. By developing estimators that utilize Two-Stage Least Square (TSLS) and Generalized Method of Moments (GMM) under the Double Machine Learning (DML) framework, this work provides a robust solution for debiasing predictions, advancing causal inference where ML/AI tools are used to generate feature variables for econometric analysis. The third essay investigates decision-making processes in livestream e-commerce, focusing on hosts' product selection and presentation timing behaviors. Using structural models that combine online learning frameworks with survival analysis, this research reveals how livestream hosts balance exploration-exploitation trade-offs and optimize product transitions to maximize sales performance, providing both theoretical insights and practical recommendations for livestream commerce optimization. Together, these essays contribute to advancing methodological approaches across econometrics, machine learning, and information systems, offering frameworks that address fundamental challenges in causal inference and decision-making. The first two essays develop theoretical solutions to estimation challenges that threaten causal validity in both sequential learning environments and static models with ML/AI-generated variables. The third essay then demonstrates how structural modeling of online learning processes can reveal opportunities for improved decision-making in practice. This complementary relationship between theoretical advances in causal inference and empirical applications in decision optimization creates a cohesive contribution with significant implications for marketing, e-commerce, and digital platform design in environments characterized by dynamic interactions, endogeneity concerns, and imperfect information.