Title: Methodological Advances in Experimental Design
About the Talk: The speaker will discuss a number of methodological advances in experimental design, and how the design stage of a study can be productively used to resolve problems before data is even collected. He will discuss three such cases: First, he will briefly examine how best to assign treatment when the goal is transporting causal effects to a known population. It turns out that improved generalization just requires a simple modification to the typical rerandomization objective (paper: https://arxiv.org/abs/2009.03860). Second, the speaker will look at how a design aimed at improved estimation of heterogeneous treatment effects works. This delves a bit into algorithmic CS theory, showing a connection between designs for good HTE estimation and a ubiquitous graph cutting problem, MAXCUT (paper: https://arxiv.org/abs/2010.11332). The bulk of the talk will be around a procedure for online (i.e. sequential) assignment of units to treatment, such as in a survey experiment. Specifically, extensions of the state-of-the-art for online discrepancy minimization from the CS literature would be introduced, as well as the procedures to accommodate multiple treatments and non-uniform treatment probabilities. This procedure works quite well, has provable robustness properties and is even sometimes competitive with _offline_ allocation procedures which are far slower than the proposed algorithm (paper: https://arxiv.org/abs/2203.02025).
About the Speaker: Drew Dimmery received his PhD in Politics from New York University in 2016 with a dissertation on methods in causal inference and machine learning, after which he worked for four years on the Adaptive Experimentation team in Facebook’s Core Data Science, developing and implementing methods and tools to improve experimentation at scale. Between 2021 and 2023, he was the Scientific Coordinator at the University of Vienna’s Data Science Research Network, supporting third mission activities of the University alongside his research. He has published in top machine learning journals such as ICML, KDD and AISTATS, as well as general-interest journals such as Science and Nature. Methodologically, his focus of study is on methods for experimentation and causal machine learning. Substantively, he focuses largely on the internet and social media.