ResearchResearch

External Cohort Reweighting via ESS-Maximization

A novel statistical framework for borrowing information from external cohorts to augment inference in a target study. The core contribution is an optimization-based reweighting scheme that identifies importance weights for the external cohort by maximizing effective sample size (ESS) subject to covariate balance constraints, producing weights that make external data as informative as possible for inference in the target population while explicitly controlling for distributional shift between populations.

Weighted Multi-Study Analysis Package

An R package for causal inference across multiple observational studies, each containing multiple groups, using a unified balancing weight framework. The package implements three weighting approaches (including the novel FLEXOR method that maximizes effective sample size through iterative optimization) to create covariate-balanced pseudo-populations that enable valid estimation of group-specific potential outcome features (means, medians, standard deviations) under study-specific unconfoundedness.

A Causal Simulation Study of Care Management Targeting

Care management programs increasingly rely on prediction algorithms to decide which patients receive limited healthcare resources. Because those algorithms are trained to predict future utilization or cost, patients whose utilization is suppressed by structural barriers may be systematically underidentified. Using linked EHR and participant survey data from the NIH All of Us Research Program, this study treats three modifiable design choices (selection threshold, feature set, and training objective) as interventions on the allocation system causal graph and evaluates their effects on allocation equity. Training objective variation produced the strongest and most consistent equity effects, while threshold adjustment produced almost none, and an access-independent benchmark built from self-reported health status revealed disparities that utilization-based metrics could not detect.

Lottery, Need-Based, and Learned Targeting in the Oregon Health Insurance Experiment

Asks a question the Oregon Health Insurance Experiment lottery did not: if a program can only offer Medicaid to half its waiting list, who should receive the offer? Replicates the lottery effect on medical debt with four estimators (weighted OLS, 2SLS, g-computation, and AIPW), tests for heterogeneous effects with a cross-fitted causal forest, and compares five allocation rules under a fixed budget using doubly robust policy evaluation. Winning the lottery reduced medical debt by 5.2 percentage points, targeting improved on the lottery by about one percentage point at best, and a simple need-based rule captured most of that gain without any machine learning. The rule that averted the most debt also concentrated offers on particular groups and nearly excluded non-English speakers.

Interrupted Time Series Analysis

Analyzes the causal impact of the COVID-19 pandemic on hospital-onset MRSA infection rates using a 10-year national panel dataset spanning 2015–2024. Identification uses interrupted time series with state fixed effects to estimate pandemic disruption to infection control trends, controlling for pre-existing state-level trajectories. The framework was extended to evaluate state-level policy and demographic moderators of pandemic impact.

An Event-Study Difference-in-Differences Analysis

Estimates the causal effect of ACA Medicaid expansion on insurance coverage among low-income adults using individual-level ACS PUMS microdata spanning 2010–2023. Compares 2014 expansion states against never-expanding states via a two-way fixed effects event-study DiD design (fixest::feols), recovering year-by-year treatment effect estimates and state-level counterfactuals. Finds that expansion states achieved persistently lower uninsured rates post-2014, with an estimated 35,000 low-income adults left uninsured in Texas alone due to non-expansion.

Population Health Analytics Platform

Interactive population health analytics platform built on NHANES data. Presents large-scale health and behavioral survey data through polished, navigable visualizations covering health behaviors, chronic conditions, and social determinants of health.