Evidence-based Medicine (EBM) is the conscientious, explicit and judicious use of current best evidence in making decisions about the care of individual patient. Billions of dollars are spent annually on the conduct of randomized clinical trials (RCT), one type of experiments widely regarded as yielding the most valuable evidence. Yet, the number of studies is growing exponentially, and most experiment results are published only as text-based articles in the medical journal, causing difficulties for both practitioners and researchers in searching, synthesizing, and ultimately, translating the best available evidence to the patient care. To address the problem, I aim to develop a unified information extraction framework for medical evidence, and build novel computational approaches based upon it to make evidence from research more accessible in Evidence-based Medicine. In this dissertation, I (i) present a unified conceptual model and coordinated workflow for evidence representation, (ii) develop open-source NLP tools for supporting EBM tasks (evidence extraction, retrieval, and synthesis), (iii) develop a medical evidence base to cater various information needs, and (iv) present a new machine reading comprehension model for answering clinical questions.

No description available.

This dissertation consists of three essays in health economics concerned with measuring the determinants of health care resource utilization and health. In the first chapter, I study entry barriers in healthcare provider markets. In the U.S., proponents of regulatory entry barriers called CON programs claim that they reduce waste by limiting "unnecessary" entry. I examine CON programs in the dialysis industry, where their effects on market structure, access, health, costs, and welfare are poorly understood, and where patients are sensitive to access and quality. I combine quasi-experimental policy variation in low population areas with a structural model of patient preferences to find that marginal entrants improved access significantly, reduced hospitalization rates, and generated for patients the utility value of traveling 275-344 fewer miles per month; but there is evidence that they contributed even more to fixed costs. Using policy variation throughout North Carolina, I also find evidence that the NC dialysis CON program created a mechanism through which incumbents could block potential entrants by expanding in tandem with their local patient populations. Taken together, my findings suggest that stronger regulatory entry barriers in low population areas may raise total welfare at patients' expense---but they also amplify concerns that CON programs dampen competition statewide. In the second chapter, I study an empirical framework commonly used in health economics research to measure the impact of an event over time using observational data: the event study. Dating back to at least Snow (1855), event studies have been used in health economics research to study mortality, health care utilization, health insurance enrollment, provider competition, and much more. Under no anticipation and parallel trends assumptions, difference-in-differences are known to identify the event's average treatment effect on the treated when units experience one event at most. In this paper, I introduce a new event study framework to accommodate settings where units may experience multiple events. I introduce a matching estimator which consistently and transparently estimates the average treatment effect on the treated of a single event under generalizations of the conventional no anticipation and parallel trends assumptions. I show that the matching estimator is equivalent to a weighted least squares estimator for a particular set of weights. I also introduce a parallel pre-trends test which can be used to scrutinize these assumptions in the usual sense. Finally, I demonstrate in a series of Monte Carlo simulations that the estimator and parallel pre-trends test work well for a wide range of treatment effects, including dynamic, non-stationary, and history-dependent treatment effects. In the third chapter, I study when and why emergency departments initiate ambulance diversions, and what happens to diverted patients. Efficiently distributing scarce healthcare resources among patients with time sensitive healthcare needs and uncertain arrival rates is a hard problem. When an emergency department gets too full, ED managers sometimes request that incoming ambulances reroute their patients to alternative destinations. While such ambulance diversions can sometimes help an overcrowded ED manage its caseload, it can also harm incoming patients and reduce systemwide EMS responsiveness. In detailed administrative records cataloging when, where, and why diversions occur, as well as who got diverted, I document that diversions commonly last exactly 1 hour, approximately 4 hours, and exactly 8 hours (indicating that managerial frictions may directly affect ED availability); that diverted patients have different characteristics than non-diverted patients (including potentially more severe symptoms); and that diverted patients spend 65% longer on the road to the hospital than non-diverted patients. I also find that diversions often occur not only because of crowdedness, but also because of hospital systems failures. I identify directions for future research.