Back to search

Georgia Institute of Technology

Transfer Learning for Robust Estimation of Heterogeneous Effects

Abstract

dc:description.abstract

This dissertation explores methodological advances at the intersection of optimization, machine learning, and causal inference, with applications in public health surveillance and personalized decision-making. The first part of the thesis addresses the challenge of timely and accurate outbreak detection during the COVID-19 pandemic. We develop a novel algorithm---Transfer Learning Random Forest (TLRF)---to estimate the instantaneous case growth rate in U.S.\ counties by adaptively selecting temporal fitting windows based on both local epidemiological conditions and demographic covariates. Leveraging transfer learning and nonparametric estimation, TLRF improves predictive performance in data-sparse regions and achieves substantial gains in early outbreak detection. Empirical validation shows that TLRF outperforms standard epidemiological baselines and could have enabled up to 224\% faster detection in real-world case studies, including one with the Colorado Department of Health and Environment. A publicly deployed tool based on TLRF provides actionable, localized outbreak alerts to policymakers. The second part of the thesis turns to the problem of subgroup discovery in heterogeneous treatment effect estimation. We propose a flexible, interpretable framework that combines orthogonalized causal effect estimation with kernel-based clustering to uncover latent subpopulations that respond differently to interventions. By learning a similarity kernel from debiased Conditional Average Treatment Effect (CATE) estimates and applying spectral clustering, our method identifies treatment-sensitive subgroups without relying on strong parametric assumptions. We show that this kernel-guided clustering acts as a form of regularization in the residual-on-residual regression framework. Through simulations and real-world applications in policy and healthcare, we demonstrate improved subgroup separation and interpretability compared to existing baselines. The third part of the thesis addresses the problem of transporting individualized treatment effect estimates across heterogeneous randomized controlled trials under covariate shift. Standard meta-analysis and network meta-analysis methods assume network connectivity and exchangeability of comparator arms across sites---conditions that are frequently violated when the target population lacks a concurrent treated arm or when baseline risk drifts across trial cohorts. We propose a placebo-anchored proxy-gold framework that treats source-trial outcomes as abundant but potentially miscalibrated proxy signals and uses scarce target placebo outcomes as high-fidelity calibration labels for baseline risk. A sparse linear correction anchors proxy outcome models to the target population, and the anchored models are embedded in a cross-fitted doubly robust learner to yield a Neyman-orthogonal CATE estimator for connected targets. For disconnected targets---where no treated outcomes are observed at the target site---the framework reduces to a principled screen-then-transport procedure guided by placebo-arm compatibility. Experiments on synthetic data and a semi-synthetic IHDP benchmark demonstrate that the proposed method is best or near-best across five evaluation metrics in connected settings, and competitive in disconnected settings with graceful degradation as structural assumptions weaken. Together, the three contributions demonstrate how statistical learning and causal reasoning can be combined to improve inference and decision-making in high-stakes, data-scarce environments.

Degree

thesis:*
Name thesis:degree_name
Operations Research, PhD
Department thesis:degree_department
Industrial and Systems Engineering
Grantor
Georgia Institute of Technology
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Wang, Zilong
Advisor dc:contributor.advisor
  • Ayer, Turgay
Committee members dc:contributor.committeemember
  • Yang, Shihao
  • Dey, Santanu
  • Goldsman, David
  • Jagpreet Chhatwal
ORCID person:identifier.orcid

Rights

Language dc:language.iso
English

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1853/82234

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Wang, Zilong. Transfer Learning for Robust Estimation of Heterogeneous Effects. Georgia Institute of Technology, 2026. https://hdl.handle.net/1853/82234