Improving Equity in Decision Support Systems

Open Access
- Author:
- Newton, Rob
- Graduate Program:
- Industrial Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- March 13, 2023
- Committee Members:
- Steven Landry, Program Head/Chair
Jose Ventura, Major Field Member
Soundar Kumara, Major Field Member
Paul Griffin, Chair & Dissertation Advisor
Terry Harrison, Outside Unit & Field Member - Keywords:
- Industrial Engineering
Operations Research
Equity
Decision Support Systems
Opioid Settlement
Facility Location
Deep Neural Network
Human Capital - Abstract:
- This dissertation introduces several methods to improve equitability among stakeholders in decision support systems. Each model provides decision makers frameworks with which to consider both resulting overall impacts as well as distributions across subgroups. Whether allocating funds or services, we seek to provide a foundation for decision makers to navigate their problem sets to prioritize accurate and fair allocations inclusive of all stakeholders. In 2021, four major pharmaceutical manufacturers and distributors reached a proposed settlement agreement with 46 state Attorneys General of $26 billion to address their liabilities in fueling the US opioid epidemic. It raises important questions about abatement conceptualization and measurement for allocating settlement funds among sub-state entities. We outline the political economy tensions undergirding the settlement and allocation, introduce an abatement conceptual framework, describe how an abatement formula was developed for Pennsylvania to allocate settlement funds, and summarize considerations for future settlement allocation efforts. We, then, evaluated the weighted combination of metrics agreed upon by Pennsylvania stakeholders to allocate settlement funds along with three other strategies from the literature that optimize allocations based on social welfare functions. We present these strategies and use them to compute new allocations for comparison. Specifically, we contrast Pennsylvania's strategy with strategies that 1) minimize total deviation, 2) minimize the worst-case (minimax) regret, and 3) balance efficiency with equity using alpha-fairness. While the Pennsylvania allocation is noteworthy in that all parties agreed to it, an allocation based on relative regret may be a more fairly perceived allocation---whether using minimax regret or blending efficiency using proportional fairness. We, next, offer a model to site additional locations to address inequity found from minimizing mean absolute deviation informed by locations recommended as maximal covering location problem solutions---to maximize the number of emergencies within the overall average distance. We used these solutions to recommend additional locations to decision makers, and provide an example solution using 2019 data from the United States Internal Revenue Service and San Francisco Fire Department that significantly increases the access to emergency rooms for low income ZIP code tabulation areas while reducing overall average distance to an emergency room and the mean absolute deviation. Finally, we discuss a model to support how Air Force Special Operations Command screens and selects candidates to lead operational squadrons each year. Currently, the command’s senior leaders manually score each eligible officer’s record. Senior leaders then evaluate officers who score above a given level, discussing suitability and fit for unit leadership. We developed a deep neural network to assist record scoring in a bias-aware process, utilizing an existing personnel database, generating scores for officer records. While demographics were unbalanced (predominantly white males), our network demonstrated 94% accuracy for candidate selection when compared to the 2020 results, and no significance of gender and race factors, suggesting the approach assist the board during the initial sorting phase. Such a model could free senior leaders from spending valuable time reviewing hundreds of records, instead spending collective time applying knowledge of candidates and squadrons to ensure the command selects high caliber leaders.