Advancements in Management Science: Applications to Online Retail, Healthcare, and Non-Profit Fundraising

Abstract

Management science is an evolving-field that requires novel models and algorithms, combining methods from statistics, optimization, and machine learning. This thesis presents advancements in management science across three domains: revenue management, healthcare, and non-profit funding platforms. The chapters in this thesis develop rigorous algorithms and techniques which are relevant in practice, and present data-driven insights into each of the application areas.

Chapter 2 studies a personalized dynamic pricing problem commonly faced by online retailers. Customers arrive sequentially to the selling platform, and for each arrival the seller must make an immediate pricing decision for that customer. The seller aims to learn the demand as a function of price and customer covariates through price experimentation, while simultaneously earning as much total revenue as possible. Previous work on this topic have adopted a classical online learning setup, where the retailer begins the selling horizon with no information about the problem and gains all knowledge about the demand function from the online selling phase. However, this assumption is often not true in practice. Many retailers already possess some information about their product's demand from market research or previous sales data, and not utilizing this information is clearly suboptimal. The chapter develops a novel framework that allows the seller to incorporate historical data on pricing decisions and realized demand, and moreover enables one to study the effect that certain characteristics of this historical dataset have on online selling performance. Using this framework, a dynamic pricing algorithm is proposed which effectively uses both historical and real time data, and achieves provably optimal performance. Furthermore, a new distance measure is developed to quantify how close the historical pricing decisions are to being optimal. Using this distance measure, the chapter shows a surprising inverse relationship between this measure and the achievable online performance.

Chapter 3 focuses on applying causal inference techniques to study the treatment efficacy of different antibiotics on patients with urinary tract infection. Up to 50% of women will experience a urinary tract infection (UTI) in their lifetime, making it the third most common indication for antibiotic treatment in the United States. Though national treatment guidelines encourage using one of three antibiotics as the first-line treatment, other second-line and alternative antibiotics are still commonly prescribed in practice. Studies on the efficacy of first-line versus second-line and alternative antibiotics for UTI are limited and dated. The chapter presents a retrospective cohort study using the claims database from Independence Blue Cross to determine the relative efficacy and adverse event rates between different categories of antibiotics. By combining causal inference techniques with automated feature extraction using the Observational Medical Outcomes Partnership (OMOP) common data model, evidence is found which supports the use of guideline-recommended first-line treatments for uncomplicated UTI. Specifically, the rate of treatment efficacy is higher for first-line antibiotics relative to alternatives. Surprisingly, the analysis also finds evidence which supports increased efficacy of first line agents relative to second-line antibiotics, which are of broader spectrum, albeit the effect difference is smaller compared to the comparison between first-line antibiotics and alternatives. This large-scale cohort study which includes a comprehensive collection of covariates provides much-needed evidence to support the continued recommendation of first-line drugs for the treatment of UTI. The chapter also suggests the feasibility for performing complex causal inference analyses using automated feature engineering packages for OMOP-formatted datasets.

Chapter 4 studies an online matching problem where sequentially arriving donors must be matched to projects needing funding on peer-to-peer philanthropic crowdfunding platforms such as DonorsChoose.org. Empirical studies have shown that (i) donors have heterogeneous preferences over the projects, and (ii) many return to make more than one donation. Facing such donors, the platform’s aim is to match each donor to one of their preferred projects so as to maximize the total donation without over-funding any projects and without knowing the arrival pattern. Previous work in the literature have not studied the effect of returning donors on algorithm performance. The chapter shows an upper bound on the best achievable worst-case performance of any online algorithm which reveals the relationship between donor return rate and algorithm performance. Furthermore, numerical analysis shows that a simple known algorithm achieves a performance that improves with the number of returning donors without differentiating between the original and return donors. The algorithm is intuitive and straightforward to implement, and the results shed light on the practical value that returning traffic can bring for fundraising platforms.