Ph.D. Dissertation
Abstract. While there is a large body of literature examining how physicians respond to
financial incentives in the context of Medicare and private payers, evidence on this issue for
Medicaid is scant. Since Medicaid patients typically constitute only a small fraction
of a physician’s patients, evidence in other settings may not apply to Medicaid. On
this front, I study how care provision responds to Florida Medicaid’s 2017 payment
reform, which transitions from a fee-for-service to a prospective payment system for
outpatient services. This transition creates procedure-specific payment shocks. Using
procedure-level policy exposure measures, I find evidence that physicians reduce the
use of procedures that are expected ex ante to be more likely to receive no payment
under the new system. Additionally, the effects are concentrated on patients without
co-morbidities and are observed only in facilities that are more dependent on
Medicaid revenues. These findings imply that physicians do respond to financial incentives
for Medicaid services. Thus, similar reforms hold out the promise to improve
cost-efficiency in health care for Medicaid patients.
Abstract. Although the Affordable Care Act’s expansion of Medicaid reduced the cost of opioid
addiction treatment, it also made opioid prescriptions more accessible, potentially leading to
higher rates of opioid addiction and death. This study examines how the expansion affected
drug overdose mortality rates. Using a difference-in-differences framework, this study finds
that the expansion increased drug overdose mortality rates by 1.208 per 100,000 people at the
county and quarter level, a 30.3% increase compared to the average mortality rate in the
expansion counties before the expansion. Findings also suggest that the expansion fuelled the
prevalence of illicitly manufactured fentanyl, a synthetic opioid, which largely accounts for the
estimated effects. Moreover, the effects on mortality were lower in expansion counties with
greater increases in insurance or opioid prescribing rates after the expansion, and more
prominent in expansion counties with higher drug overdose mortality rates before the
expansion. Taken together, these findings are consistent with the argument that stringent
restrictions on prescription opioids alongside the expansion of Medicaid resulted in more
people turning to illicitly manufactured opioids, increasing overdose deaths.
Abstract. The “gateway” hypothesis contends that marijuana use increases people’s risk of progressing
to use illicit hard drugs (e.g., cocaine, heroin). As some United States (U.S.) states have
legalized recreational marijuana in recent years and many are considering decriminalizing
marijuana, it is crucial to investigate whether such legislation will subsequently fuel hard drug
use. Although a large body of medical and economic literature has examined the “gateway”
hypothesis, existing studies tend not to differentiate between correlation and causation. This
study contributes to the literature by using a bivariate survival panel model that controls for
confounding variables. Using the National Longitudinal Survey of Youth 1997, I found strong
evidence of “gateway” effects among youth in the U.S. Furthermore, the effects are more
pronounced among those who first used marijuana before the age of 18, as well as those who
used marijuana more frequently. Moreover, the effects are lower in African Americans and
become less potent as people age. These results inform the current debate over the potential of
marijuana use during adolescence to further hard drug involvement and highlight the
importance of postponing the onset and reducing the frequency of marijuana use.
M.S.Thesis
Abstract. Visual sleep scoring of Polysomnograms (PSG) by an expert is a time-consuming process. Although a number of automatic sleep scoring methods have been proposed in the literature, most of them are based on supervised algorithms. That is, labels in their training data assigned by an expert are required. In this thesis, we propose an unsupervised method for wake/sleep scoring without labels a priori. Features based on temporal and spectral analysis are extracted from a single channel of EEG. Principal
Component Analysis (PCA) is used to reduce the number of features while identifying patterns in the data. The Gustafson–Kessel algorithm is used for clustering analysis, and sleep scoring is done by retrieving one characteristic feature of wake: the alpha rhythm. Sixteen subjects from the MIT-BIH Polysomnographic Database were tested by this method. Compared to actual stage scoring, 14 have scoring accuracy above 75%, and the average accuracy is 79.35%.
Capstone Project
Abstract. Malaria is a severe and often fatal disease. Conventional malaria detections are performed by microscopists who analyze microscopic blood smear images in laboratory settings, which requires human expertise and large investments. These resources may be inadequate in developing counties, where malaria is more predominant. Deep learning models thus may play a role in facilitating malaria detection and reducing healthcare costs. In this study, I build a Convolutional Neural Network (CNN) based neural network algorithm to classify images of blood cells into being parasitized or uninfected. The algorithm achieves high overall accuracy, about 98%. Other CNN models with similar structures also perform very well with similar performance. Thus, the model performance is robust to CNNs with varying configurations. As such, deep learning techniques show potential in achieving high accuracy even being applied to the healthcare setting fully automatically. Further tuning of the model may obtain even better performance. As smartphones are widely used, apps based on deep learning could be developed so that malaria detection can be widely conducted even with smartphones, increasing the cost-efficiency and the number of tests. However, the algorithm shows limitations as certain uninfected cells are wrongly classified while it fails to detect certain parasitized cells. This provides a caveat in applying the algorithm to actual healthcare settings and points to a direction for future improvement.