Parametric and Nonparametric Learners for Adaptive Experiments
TU Dortmund · submitted, under review
My dissertation studies adaptive experimentation from both parametric and nonparametric perspectives, with a focus on sequential treatment allocation under complex dependence structures.
My doctoral research topic is adaptive experimentation, or sequential decision-making. Broadly, I study how classical online learning frameworks, developed under idealized independence assumptions, need to be adjusted when units of observation are not isolated from one another. In many economic and social settings, the outcome of one unit depends not only on the treatment it receives, but also on the treatments assigned to others with whom it interacts. This phenomenon is known as interference.
My work develops a unified theoretical framework for this problem, proposes a family of algorithms — both parametric and nonparametric — tailored to different interference structures, and connects the classical bandit setting to the interference setting.
A working paper based on this research is currently being finalized for submission. Details will be shared here upon publication.
We propose weighted bootstrap confidence intervals for the bias-corrected local linear estimator, building on the bias-correction approach of Calonico, Cattaneo, and Farrell (2018) and Cheng and Chen (2019), and study pointwise performance.
Joint work with Professor Carsten Jentsch; an extension of my master’s thesis.
German Statistical Week, 2021 (online), University of Kiel. Nearest Neighbor Matching: Does the M-out-of-N Bootstrap Work When the Naive Bootstrap Fails?
German Probability and Statistics Days (GPSD), 2018, University of Freiburg. Modeling and Prediction of Dynamic Networks Using Binary Autoregressive Time Series Processes.