Zubin Jelveh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zubin Jelveh has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
Bruce Kogut
2 shared publications- Tin Nguyen
University of Maryland, College Park
2 shared publications Aaron Chalfin
Affiliations Department of Criminology, University of Pennsylvania, Philadelphia, PA, United States of America, National Bureau of Economic Research, Cambridge, MA, United States of America
1 shared publication- Andrea Tentner
Department of Systems Biology, Harvard Medical School
1 shared publication Benjamin Hansen
Affiliations National Bureau of Economic Research, Cambridge, MA, United States of America, Department of Economics, University of Oregon, Eugene, OR, United States of America
1 shared publication- Donald Braman
The George Washington University Law School
1 shared publication
Latest Publications
- Open Access
- 15 April 2026
- Published by Springer Nature in Journal of Quantitative Criminology
- p. 1-42
Predicting behavioral health-involved police encounters: A machine learning approach
Objectives: We evaluate the predictive accuracy of machine learning algorithms that forecast individual-level risk of behavioral health-involved encounters with police (BHIP) among community members with prior contact with first responders. Methods: We linked arrest and ambulance data from one city (May 2016-October 2017) for 199,679 unique individuals. The first 12 months provided predictive features; the subsequent 6 months served as the outcome window. We compared machine learning models utilizing 283 features to a baseline Prior High-User (PHU) approach based solely on prior BHIP counts. Results: Machine learning achieved positive predictive values (PPV) of 50.8%, 20.2%, and 7.5% when identifying the top 0.1% (n=199), 1% (n=1,996), and 5% (n=9,983) highest-risk individuals, representing improvements of 20%, 57%, and 44% over PHU. Critically, 28% of individuals in the ML-identified top 1% had no prior BHIP during the identification period, enabling true prevention. ML models demonstrated superior sensitivity, identifying 26.7% of all future BHIP events when targeting the top 1%, versus 17.1% for PHU. A simplified four-feature model using ambulance data maintained comparable performance. ML’s improved PPV increases statistical power for intervention studies: a 200-person trial using ML-identified participants could detect a 38% reduction in BHIP, while achieving equivalent power with PHU would require 310 participants (55% larger sample). Conclusions: Machine learning substantially improves identification of individuals at BHIP risk, including those without prior events. This enhanced PPV can enable more efficient targeting of preventive services and make rigorous intervention research more feasible by increasing statistical power. Simplified models using only medical-services data suggest practical implementation pathways, though findings should be validated in other jurisdictions to ensure generalizability.
- 15 April 2026
- Published by Springer Nature in Journal of Quantitative Criminology
- p. 1-42
Open AccessPredicting behavioral health-involved police encounters: A machine learning approach
Objectives: We evaluate the predictive accuracy of machine learning algorithms that forecast individual-level risk of behavioral health-involved encounters with police (BHIP) among community members with prior contact with first responders. Methods: We linked arrest and ambulance data from one city (May 2016-October 2017) for 199,679 unique individuals. The first 12 months provided predictive features; the subsequent 6 months served as the outcome window. We compared machine learning models utilizing 283 features to a baseline Prior High-User (PHU) approach based solely on prior BHIP counts. Results: Machine learning achieved positive predictive values (PPV) of 50.8%, 20.2%, and 7.5% when identifying the top 0.1% (n=199), 1% (n=1,996), and 5% (n=9,983) highest-risk individuals, representing improvements of 20%, 57%, and 44% over PHU. Critically, 28% of individuals in the ML-identified top 1% had no prior BHIP during the identification period, enabling true prevention. ML models demonstrated superior sensitivity, identifying 26.7% of all future BHIP events when targeting the top 1%, versus 17.1% for PHU. A simplified four-feature model using ambulance data maintained comparable performance. ML’s improved PPV increases statistical power for intervention studies: a 200-person trial using ML-identified participants could detect a 38% reduction in BHIP, while achieving equivalent power with PHU would require 310 participants (55% larger sample). Conclusions: Machine learning substantially improves identification of individuals at BHIP risk, including those without prior events. This enhanced PPV can enable more efficient targeting of preventive services and make rigorous intervention research more feasible by increasing statistical power. Simplified models using only medical-services data suggest practical implementation pathways, though findings should be validated in other jurisdictions to ensure generalizability.
- Open Access
- 10 March 2026
- Published by Elsevier in SSRN Electronic Journal
The SPIRIT of Science: Institutional Responses to the P-Hacking Problem in Academic Clinical Trials
All researchers face pressure to demonstrate significance, but whether that pressure distorts findings may depend on how observable their analytic choices are. We examine this in clinical trials, where academic and industry sponsors face similar evidentiary benchmarks but differ in documentation practices and oversight contexts. Using p-value distributions from 10,596 completed randomized clinical trials registered on ClinicalTrials.gov, we test for threshold-driven distortions around p = 0.05. Academicsponsored trials exhibit an excess of primary-outcome p-values just below 0.05, while industry-sponsored trials show little or no such pattern; secondary outcomes show no comparable distortion for either sponsor type. Internal consistency checks between reported p-values and 95% confidence intervals replicate the same sponsor gap, and a validation exercise using p-values extracted from PubMed Central trial publications yields similar directional differences. The academic-industry gap persists after adjusting for trial characteristics, suggesting that organizational setting itself may matter. We examine the 2013 SPIRIT statement (Standard Protocol Items: Recommendations for Interventional Trials), which standardized protocol documentation for clinical trials. Academic trials initiated after SPIRIT show weaker evidence of threshold-adjacent bunching than those initiated earlier, whereas trials initiated earlier but reported later show no comparable change-a pattern suggestive of protocol design and documentation as a relevant margin.
- 10 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessThe SPIRIT of Science: Institutional Responses to the P-Hacking Problem in Academic Clinical Trials
All researchers face pressure to demonstrate significance, but whether that pressure distorts findings may depend on how observable their analytic choices are. We examine this in clinical trials, where academic and industry sponsors face similar evidentiary benchmarks but differ in documentation practices and oversight contexts. Using p-value distributions from 10,596 completed randomized clinical trials registered on ClinicalTrials.gov, we test for threshold-driven distortions around p = 0.05. Academicsponsored trials exhibit an excess of primary-outcome p-values just below 0.05, while industry-sponsored trials show little or no such pattern; secondary outcomes show no comparable distortion for either sponsor type. Internal consistency checks between reported p-values and 95% confidence intervals replicate the same sponsor gap, and a validation exercise using p-values extracted from PubMed Central trial publications yields similar directional differences. The academic-industry gap persists after adjusting for trial characteristics, suggesting that organizational setting itself may matter. We examine the 2013 SPIRIT statement (Standard Protocol Items: Recommendations for Interventional Trials), which standardized protocol documentation for clinical trials. Academic trials initiated after SPIRIT show weaker evidence of threshold-adjacent bunching than those initiated earlier, whereas trials initiated earlier but reported later show no comparable change-a pattern suggestive of protocol design and documentation as a relevant margin.
- Open Access
- 15 October 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society
- Vol. 8 (2) , 1875-1887
Effort-aware Fairness: Incorporating a Philosophy-informed, Human-centered Notion of Effort into Algorithmic Fairness Metrics
Although popularized AI fairness metrics, e.g., demographic parity, have uncovered bias in AI-assisted decision-making outcomes, they do not consider how much effort one has spent to get to where one is today in the input feature space. However, the notion of effort is important in how Philosophy and humans understand fairness. We propose a philosophy-informed approach to conceptualize and evaluate Effort-aware Fairness (EaF), grounded in the concept of Force, which represents the temporal trajectory of predictive features coupled with inertia. Besides theoretical formulation, our empirical contributions include: (1) a pre-registered human subjects experiment, which shows that for both stages of the (individual) fairness evaluation process, people consider the temporal trajectory of a predictive feature more than its aggregate value; (2) pipelines to compute Effort-aware Individual/Group Fairness in the criminal justice and personal finance contexts. Our work may enable AI model auditors to uncover and potentially correct unfair decisions against individuals who have spent significant efforts to improve but are still stuck with systemic disadvantages outside their control.
- 15 October 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society
- Vol. 8 (2) , 1875-1887
Open AccessEffort-aware Fairness: Incorporating a Philosophy-informed, Human-centered Notion of Effort into Algorithmic Fairness Metrics
Although popularized AI fairness metrics, e.g., demographic parity, have uncovered bias in AI-assisted decision-making outcomes, they do not consider how much effort one has spent to get to where one is today in the input feature space. However, the notion of effort is important in how Philosophy and humans understand fairness. We propose a philosophy-informed approach to conceptualize and evaluate Effort-aware Fairness (EaF), grounded in the concept of Force, which represents the temporal trajectory of predictive features coupled with inertia. Besides theoretical formulation, our empirical contributions include: (1) a pre-registered human subjects experiment, which shows that for both stages of the (individual) fairness evaluation process, people consider the temporal trajectory of a predictive feature more than its aggregate value; (2) pipelines to compute Effort-aware Individual/Group Fairness in the criminal justice and personal finance contexts. Our work may enable AI model auditors to uncover and potentially correct unfair decisions against individuals who have spent significant efforts to improve but are still stuck with systemic disadvantages outside their control.
- Open Access
- 16 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 354-358
Which Demographic Features Are Relevant for Individual Fairness Evaluation of U.S. Recidivism Risk Assessment Tools?
- 16 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 354-358
Open AccessWhich Demographic Features Are Relevant for Individual Fairness Evaluation of U.S. Recidivism Risk Assessment Tools?
- 29 April 2024
- Published by Oxford University Press (OUP) in The Economic Journal
- Vol. 134 (662) , 2439-2469
Political Language in Economics
Does academic writing in economics reflect the political orientation of economists? We use machine learning to measure partisanship in academic economics articles. We predict the observed political behaviour of a subset of economists using phrases from their academic articles, show good out-of-sample predictive accuracy and then predict partisanship for all economists. We then use these predictions to examine patterns of political language in economics. We estimate journal-specific effects on predicted ideology, controlling for author and year fixed effects, that accord with existing survey-based measures. We show considerable sorting of economists into fields of research by predicted partisanship. We also show that partisanship is detectable even within fields, even across those estimating the same theoretical parameter. Using policy-relevant parameters collected from previous meta-analyses, we then show that imputed partisanship is correlated with estimated parameters, such that the implied policy prescription is consistent with partisan leaning. For example, we find that going from the most left-wing authored estimate of the taxable top income elasticity to the most right-wing authored estimate decreases the optimal tax rate from 77% to 60%.
- 29 April 2024
- Published by Oxford University Press (OUP) in The Economic Journal
- Vol. 134 (662) , 2439-2469
Political Language in Economics
Does academic writing in economics reflect the political orientation of economists? We use machine learning to measure partisanship in academic economics articles. We predict the observed political behaviour of a subset of economists using phrases from their academic articles, show good out-of-sample predictive accuracy and then predict partisanship for all economists. We then use these predictions to examine patterns of political language in economics. We estimate journal-specific effects on predicted ideology, controlling for author and year fixed effects, that accord with existing survey-based measures. We show considerable sorting of economists into fields of research by predicted partisanship. We also show that partisanship is detectable even within fields, even across those estimating the same theoretical parameter. Using policy-relevant parameters collected from previous meta-analyses, we then show that imputed partisanship is correlated with estimated parameters, such that the implied policy prescription is consistent with partisan leaning. For example, we find that going from the most left-wing authored estimate of the taxable top income elasticity to the most right-wing authored estimate decreases the optimal tax rate from 77% to 60%.
- Open Access
- 4 April 2023
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 18 (4) , e0283811
No ground truth? No problem: Improving administrative data linking using active learning and a little bit of guile
While linking records across large administrative datasets [“big data”] has the potential to revolutionize empirical social science research, many administrative data files do not have common identifiers and are thus not designed to be linked to others. To address this problem, researchers have developed probabilistic record linkage algorithms which use statistical patterns in identifying characteristics to perform linking tasks. Naturally, the accuracy of a candidate linking algorithm can be substantially improved when an algorithm has access to “ground-truth” examples—matches which can be validated using institutional knowledge or auxiliary data. Unfortunately, the cost of obtaining these examples is typically high, often requiring a researcher to manually review pairs of records in order to make an informed judgement about whether they are a match. When a pool of ground-truth information is unavailable, researchers can use “active learning” algorithms for linking, which ask the user to provide ground-truth information for select candidate pairs. In this paper, we investigate the value of providing ground-truth examples via active learning for linking performance. We confirm popular intuition that data linking can be dramatically improved with the availability of ground truth examples. But critically, in many real-world applications, only a relatively small number of tactically-selected ground-truth examples are needed to obtain most of the achievable gains. With a modest investment in ground truth, researchers can approximate the performance of a supervised learning algorithm that has access to a large database of ground truth examples using a readily available off-the-shelf tool.
- 4 April 2023
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 18 (4) , e0283811
Open AccessNo ground truth? No problem: Improving administrative data linking using active learning and a little bit of guile
While linking records across large administrative datasets [“big data”] has the potential to revolutionize empirical social science research, many administrative data files do not have common identifiers and are thus not designed to be linked to others. To address this problem, researchers have developed probabilistic record linkage algorithms which use statistical patterns in identifying characteristics to perform linking tasks. Naturally, the accuracy of a candidate linking algorithm can be substantially improved when an algorithm has access to “ground-truth” examples—matches which can be validated using institutional knowledge or auxiliary data. Unfortunately, the cost of obtaining these examples is typically high, often requiring a researcher to manually review pairs of records in order to make an informed judgement about whether they are a match. When a pool of ground-truth information is unavailable, researchers can use “active learning” algorithms for linking, which ask the user to provide ground-truth information for select candidate pairs. In this paper, we investigate the value of providing ground-truth examples via active learning for linking performance. We confirm popular intuition that data linking can be dramatically improved with the availability of ground truth examples. But critically, in many real-world applications, only a relatively small number of tactically-selected ground-truth examples are needed to obtain most of the achievable gains. With a modest investment in ground truth, researchers can approximate the performance of a supervised learning algorithm that has access to a large database of ground truth examples using a readily available off-the-shelf tool.