Title: Applications of Optimization and Causal Inference

Date: Wednesday, May 6, 2026

Time: 3:00 pm – 5:00 pm ET

Location: Remote on Zoom — https://gatech.zoom.us/j/93388746171?pwd=IknDRQHq8fIL8QJi7YQTV9kveeYnxp.1 

Meeting ID: 933 8874 6171

Passcode: 734561

Committee:

Dr. Turgay Ayer (Advisor), Virginia C. and Joseph C. Mello Endowed Chair, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Shihao Yang, Harold E. Smalley Early Career Professor, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Santanu Dey, Anderson-Interface Chair and Professor, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. David Goldsman, Coca-Cola Foundation Professor, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Jagpreet Chhatwal, Director, Center for Health Technology Assessment, Mass General Brigham; Associate Professor, Harvard Medical School

Abstract:

This thesis develops methods for causal learning under

data scarcity and distributional shift, combining transfer learning,

kernel learning, and doubly robust semi-parametric estimation in public health 

and personalized medicine.

The first part develops Transfer Learning Random Forests (TLRF) for

county-level epidemic growth-rate estimation and outbreak prioritization

under fixed public-health capacity. Its core technical step is a

proxy-identification argument that recasts latent

growth-rate estimation as conditional mean learning on observable

log-difference outcomes, followed by adaptive spatiotemporal transfer

via random-forest weighting. In a retrospective comparison against

historical decisions by the Colorado Department of Public Health and

Environment, TLRF yields substantially higher investigation positive

predictive value (including a 224% relative improvement as a point

estimate under the same decision budget).

The second part develops a causal-kernel subgroup discovery framework:

Neyman-orthogonalized CATE estimation and kernel learning under the

Robinson decomposition are followed by kernelized convex clustering to

identify subpopulations defined by treatment responsiveness rather than

covariate similarity alone. This component turns treatment-effect

heterogeneity into decision-ready subgroup structure to support

hypothesis generation, follow-up trial enrichment, and targeted

intervention design.

The third part develops a placebo-anchored transfer framework for

transporting individualized treatment effects across heterogeneous

trials under covariate shift, including settings where the target

population has limited or no treated outcomes. A sparsity-driven,

LASSO-based proxy-gold transfer correction is integrated with

cross-fitted doubly robust estimation to support target-population

effect learning under explicit transport assumptions. This provides a

practical evidence-synthesis pathway for target-population decision

support when standard IPD meta-analysis assumptions are not fully met.

Across these three settings, the unifying theme is principled

information borrowing: semiparametric identification, orthogonalized

causal kernel learning, and transfer-based doubly robust inference for

decision-making when target data alone are insufficient.