Experimentation Platforms and Learning Treatment Effects in Panels
Experiments in brick-and-mortar retail are contaminated for myriad reasons. Pragmatic inference in such settings is more akin to learning from observational data, as opposed to the typical setup one might consider for a carefully designed randomized experiment. So motivated, we consider the problem of causal inference in panels with general intervention patterns that may depend on the historical data. We provide a novel, near-complete solution to this problem that allows for rate-optimal recovery of treatment effects. Our work generalizes the outcome model of the difference-in-difference paradigm and expands the applicability of the synthetic-control paradigm. In doing so, we provide a novel de-biasing analysis that addresses the low-rank matrix regression with non-random intervention patterns and noise; a non-trivial feature of independent interest. Our algorithms form the core of a new testing platform we co-developed with a USD 100B drink company, which increased revenue by millions of dollars monthly in Mexico alone.
Tianyi Peng is a Ph.D. student at MIT. He is advised by Vivek Farias, and also mentored by Andrew Li. He is broadly interested in developing algorithms for learning and inference in large-scale dynamic decision-making systems. In particular, he is interested in developing next-generation experimentation platforms, which provide scalable, low-cost solutions for discovering beneficial strategies/policies. In translating these ideas, he is engaged with Anheuser-Busch InBev, Takeda Pharmaceuticals, TikTok, and Liberty Mutual. His work has been recognized as a finalist for the MSOM Student Paper Competition (2022), and has won the INFORMS Daniel H. Wagner Prize (2022), Applied Probability Society Best Student Paper Prize (2022), Jeff McGill Student Paper Award (2022) and the best thesis award at Tsinghua where he graduated with the 2017 Yao Class.
]]>Experiments in brick-and-mortar retail are contaminated for myriad reasons. Pragmatic inference in such settings is more akin to learning from observational data, as opposed to the typical setup one might consider for a carefully designed randomized experiment. So motivated, we consider the problem of causal inference in panels with general intervention patterns that may depend on the historical data. We provide a novel, near-complete solution to this problem that allows for rate-optimal recovery of treatment effects. Our work generalizes the outcome model of the difference-in-difference paradigm and expands the applicability of the synthetic-control paradigm. In doing so, we provide a novel de-biasing analysis that addresses the low-rank matrix regression with non-random intervention patterns and noise; a non-trivial feature of independent interest. Our algorithms form the core of a new testing platform we co-developed with a USD 100B drink company, which increased revenue by millions of dollars monthly in Mexico alone.
]]>