Jiacheng Zou (MS&E): Sparse Models for Large Panel Data

Jiacheng Zou: Sparse Models for Large Panel Data 

with Applications in Healthcare and Economics

Abstract: In this talk, I will present papers [1, 2] on the application of sparse models for large panel data. Large panels present intricate hypotheses due to unit-level variable selection. Each unit can have varying support sets. At the same time, high unit counts necessitate false discovery control. [3] proposed Panel Multiple Testing that provides panel Family-Wise Error Rate (FWER) control, allows arbitrary cross-unit correlation, and is more powerful than Bonferroni by adapting to panel structure.
Papers [1, 2] identify testing problems of this setting, and use Panel Multiple Testing to recover the joint panel model with FWER control. Paper [1] proposes a novel method of edge identification with FWER control for the latent bipartite decision graph of a two-sided, money-free platform. First, each demand-side agent decision regression conducts variable selection of supplier fixed effects. Then Panel Multiple Testing identifies significant supplier effects. This principled approach controls for spurious edges. Using the graphs constructed, we analyze the effects of recent modifications in the national kidney allocation policy. Paper [2] models multiple change points detection in large time-series panels as a variable selection problem of the time periods. Each time period with significant structural breaks is chosen as change points by Panel Multiple Testing with FWER control, after unit-level LASSO screening. In extensive simulations with various signal-to-noise ratios, our method reduces false discoveries and boosts correct selections, lifting F1 score over 20% against leading benchmarks.

[1] "Learning Bipartite Decision Graph with Applications for National Kidney Allocation Policy Evaluations", Jiacheng Zou, Johan Ugander and Itai Ashlagi, 2023
[2] "Large Dimensional Change Point Detection with FWER Control as Automatic Stopping", Jiacheng Zou, Yang Fan and Markus Pelger, 2023
[3] "Inference for High-Dimensional Panel Data with Many Covariates", Markus Pelger and Jiacheng Zou, 2022