Big Data in Context

Overview

Coursework for BIO 511 (Big Data) at Arizona State University, centered on a final research paper analyzing the COMPAS proprietary risk-assessment algorithm used in criminal sentencing. The paper argues that competing definitions of algorithmic fairness, calibration versus error-rate parity, are mathematically irreconcilable when group base rates differ, and that trade-secret law has filled the federal regulatory vacuum around such tools. Supporting deliverables include a project proposal and five individual reading responses on data ethics, fairness, and governance. **Highlight:** ProPublica's analysis of 7,000+ Broward County defendants found Black defendants nearly twice as likely to be falsely labeled high-risk

Key Achievements

• Wrote a final analytical paper, 'Encoded Injustice,' framing the 2016 ProPublica COMPAS audit and the Wisconsin Supreme Court ruling in State v. Loomis as a conflict between commercial accuracy and constitutional due process
• Argued the fairness 'impossibility result': no algorithm can satisfy both calibration and error-rate parity when recidivism base rates differ between groups without perfect prediction, drawing on Castelnovo et al. (2025)
• Connected the case to broader systemic bias via Obermeyer et al. (2019) on a healthcare risk algorithm and Ruha Benjamin's 'New Jim Code,' and contrasted COMPAS opacity with biobanking consent norms
• Authored a structured project proposal defining the case study, central analytical question, source set, and methodological approach
• Completed five individual reading responses spanning data ethics in modeling practice (Barocas & Boyd, Leonelli), GenBank's open-data history (Strasser), algorithmic calibration bias, and the FAIR vs. CARE data-governance frameworks

Approach

Used a case-study method tracing how distinct actors, Northpointe, the courts, ProPublica, and the defense bar, applied incompatible evaluation criteria to the same tool. Grounded the analysis in the peer-reviewed fairness literature and legal precedent to show the trade-off between fairness metrics is a normative rather than technical choice, then extended it into the regulatory ecosystem of trade-secret protection and emerging transparency mandates.

Tools & Technologies

Results

Deliverables are written documents rather than computed metrics; the core conclusion is presented in docs/Elsaady_BIO511_FinalProject.docx, which contends that algorithmic fairness in sentencing cannot be resolved by statistical refinement and requires democratic, normative judgment.