Learning to Care
Can Machine Learning Target Health Care Fraud? Evidence From Medicare Hospitalizations
Shubhranshu Shekhar, Jetson Leder-Luis & Leman Akoglu
Journal of Policy Analysis and Management, Winter 2026
Abstract:
The United States spends more than $4 trillion per year on health care, largely conducted by private providers and reimbursed by insurers. A major concern in this system is overbilling and fraud by hospitals, who face incentives to misreport their claims to receive higher payments. In this work, we develop novel machine learning tools to identify hospitals that overbill insurers, which can be used to guide investigations and auditing of suspicious hospitals for both public and private health insurance systems. Using large-scale claims data from Medicare, the US federal health insurance program for the elderly and disabled, we identify patterns consistent with fraud among inpatient hospitalizations. Our proposed approach for fraud detection is fully unsupervised, not relying on any labeled training data, and is explainable to end users, providing interpretations for which diagnosis, procedure, and billing codes lead to hospitals being labeled suspicious. Using newly collected data from the Department of Justice on hospitals facing anti-fraud lawsuits, and case studies of suspicious hospitals, we validate our approach and findings. Our method provides a nearly fivefold lift over random targeting of hospitals. We also perform a postanalysis to understand which hospital characteristics, not used for detection, are associated with suspiciousness.
Exploring the Early Effects of State Consumer Protection Policies on Medical Debt in Collections
Fredric Blavin et al.
Health Services Research, February 2026
Study Setting and Design: This study uses a quasi-experimental research design to estimate the impact of consumer protection laws implemented between 2020 and 2022 in Illinois, Maryland, New Mexico, and Oregon on the share of consumers with medical debt in collections. These laws primarily aim to protect consumers against medical debt by expanding access to hospital financial assistance. We use a synthetic control approach to estimate changes in medical debt following the implementation of policies in treatment states relative to changes in select control states. We also assess the effects of earlier policies implemented between 2013 and 2019 in Washington, Utah, and North Carolina.
Data Sources and Analytic Sample: This analysis relies on two extracts of credit bureau data from one of the country's three main credit bureau agencies. The first extract consists of random samples from June 2017 to June 2024 of approximately 125,000 consumers in each treatment state and 500,000 residents from the pool of 14 selected comparison states in each year. The second extract is based on a 2%–4% random sample of consumers in each year from 2011 to 2022.
Principal Findings: We did not observe a statistically significant reduction in medical debt associated with policies implemented in these states within the study timeframe. In most states in our primary analysis, point estimates of the treatment effects are near zero, and in nearly all state-years, we can only rule out declines in medical debt larger than 1–3 percentage points following policy implementation.
Raising the Stakes: Physician Facility Investments and Provider Agency
Elizabeth Munnich et al.
American Economic Review, February 2026, Pages 502-534
Abstract:
Principal-agent problems often extend beyond what can be directly addressed through conventional incentive arrangements. We examine a context where physicians are likely under-incentivized to minimize total medical costs until their private financial interests align with those of patients. Leveraging novel data on physician ownership of ambulatory surgery centers -- that is, same-day facilities -- we show that these equity holdings cause a substitution away from higher cost, rival settings that lowers Medicare spending by 10–40 percent per physician. We find no clear evidence of perverse behavior following these investments. Instead, our findings demonstrate how entrepreneurial activity can indirectly limit principal-agent problems and improve efficiency.
Helping addicts: When can trying to do good be dysfunctional?
Federico Guerrero et al.
Public Choice, January 2026, Pages 107-128
Abstract:
We integrate a dynamic model of addiction with a dynamic model of imperfect altruism to examine help for addicts. Help for the addict is a public good that reduces the rate of addiction, and addiction emits an externality that negatively impacts those who might help. Our model suggests help that reduces addiction for one person can dysfunctionally increase addiction for another who has different preferences. Although an increase in the number of available helpers reduces addiction when there is little congestion, it can create a free rider problem that dysfunctionally eliminates private help when congestion sufficiently reduces the marginal utility of helping. Increasing the government provision of help can dysfunctionally increase addiction when it crowds out private help provision that is more efficient. In general, our model indicates that effectively addressing addiction is context-dependent, which aligns with the literature on addiction treatment.
The Effects of M&A in Medicare Advantage: A Case Study of Humana-Arcadian Management Services
Jake Kramer
Review of Industrial Organization, January 2026, Pages 15-54
Abstract:
Under U.S. antitrust policy, enforcers must identify mergers that are potentially harmful and propose remedies where necessary. I estimate the price, plan star rating, variety, share, and cost effects of Humana’s acquisition of Arcadian Management Services (AMS) -- a Medicare Advantage HMO -- and the U.S. Department of Justice’s subsequent divestiture order to provide insight with respect to antitrust enforcers’ success. I find that Humana experienced cost synergies following the merger and passed them through to customers via lower premiums and higher star ratings. AMS plans were cancelled almost immediately after they were acquired, which reduced product variety. These plans were of a lower quality and beneficiaries in these plans were highly inertial, so AMS variety reductions are unlikely to have caused consumer welfare harms.
Switching Medicare Plans Outside Open Enrollment Was Increasingly Common, Especially Among Sicker Enrollees, 2015–22
Grace Mackleby, Angela Liu & Erin Trish
Health Affairs, February 2026, Pages 164-174
Abstract:
To mitigate adverse selection in Medicare Advantage (MA) and traditional Medicare plans, MA enrollees are only allowed to switch between MA plans or enroll in traditional Medicare during specific open enrollment periods. However, during the past decade, regulators relaxed these enrollment periods to facilitate switching throughout the year. Using administrative data from the period 2015–22, we documented that among incumbent MA enrollees, an increasing share of switching to other MA plans or traditional Medicare occurred outside the standard Medicare open enrollment period. Compared with Medicare open enrollment period switchers, alternative enrollment period switchers tended to have higher risk scores and hospitalization rates before switching. In addition, beneficiaries who switched to traditional Medicare or broader-network plans tended to have higher risk scores and hospitalization rates before switching than those who switched to other plans. Our results highlight a trade-off: Flexible enrollment periods may allow some beneficiaries to enroll in plans that better suit their needs but may also allow beneficiaries with higher medical costs to sort into certain MA plans or traditional Medicare, which potentially generates adverse selection if other selection mitigation policies are insufficient.
AI in the Lab: AlphaFold2’s Impacts on Human-Produced Knowledge
Jiarui (Jerry) Qian
University of Virginia Working Paper, January 2026
Abstract:
This paper evaluates how Artificial Intelligence (AI) shapes scientific knowledge production by human researchers, taking Google’s AlphaFold2 (AF2) as a case. AF2 predicts protein structures with near-experimental accuracy at a fraction of the cost. Exploiting variation in AF2’s prediction confidence in a difference-in-differences design, I find that experimental output for well-predicted proteins falls by about 35%. This decline is more pronounced for proteins that are costly to experiment on or have limited downstream demand. The reallocation of effort is also heterogeneous across experimental methods, concentrated in widely accessible X-ray crystallography while leaving the capital-intensive cryo–electron microscopy (Cryo-EM) method largely unaffected. Meanwhile, AF2 has little effect on the quality of execution but appears to alter incentives: the academic reward for working on experimentally unattempted but well-predicted structures largely disappears, as reflected in a sharp drop in the probability of publication in top-tier journals. Overall, AF2 shifts experimental effort toward not-well-predicted areas where human comparative advantage persists.
Medicare Part D and Hospital Admissions due to Antimicrobial Resistance
Ricardo Ang
Health Economics, forthcoming
Abstract:
Antimicrobial resistance (AMR) has been increasing rapidly in the United States despite government efforts to contain its spread. Both under-utilization and overuse of prescribed antimicrobials contribute to rising resistance. The introduction of Medicare Part D in 2006 expanded prescription drug coverage for the elderly, including coverage for antimicrobial medications. If cost barriers had previously led to under-utilization of prescriptions, then Medicare Part D could have mitigated AMR by improving access to antimicrobials. However, if Medicare Part D also encouraged excessive antibiotic use, it may have inadvertently contributed to greater resistance. Using data from the Nationwide Inpatient Sample (NIS), Healthcare Cost and Utilization Project (HCUP), Agency for Healthcare Research and Quality for January 2001 to September 2008, I estimate the causal impact of Medicare Part D on AMR-related hospital discharges using a difference-in-differences identification strategy. My findings suggest that Medicare Part D led to a slower increase in AMR-related inpatient discharges among the Medicare-eligible population.
Transactive Memory Systems and Hospital Trauma Team Performance: Shared Experience in Action Teams
Linda Argote et al.
Organization Science, January-February 2026, Pages 48-70
Abstract:
We identify and test a theoretical mechanism linking shared team experience and team performance: a transactive memory system (TMS). Our empirical context is the care of patients with severe acute traumatic injury in a hospital emergency department. We coded behavioral indicators of transactive memory from video recordings of trauma resuscitations in a hospital emergency department. We obtained objective measures of team performance -- patient lengths of stay in the intensive care unit and in the hospital -- from hospital records, as well as information about the experience of team members. Our results of analyzing data from 121 patients reveal that patients treated by trauma teams with strong TMS experience significantly shorter lengths of stay in the intensive care unit (ICU) and in the hospital than patients treated by trauma teams with weaker TMS. The magnitude of the effects was large: Increasing TMS by one standard deviation was associated with a reduction in hospital length of stay (LOS) of 3.3 days and a reduction in ICU LOS of 1.9 days. Experience working together predicted the strength of the team’s transactive memory over and above the effect of individual experience. Further, transactive memory mediated or explained the effect of team experience on team performance. These results were robust to multiple sensitivity analyses in which we varied our definition of team experience and our modeling approach and included controls for team, task, and context characteristics. We discuss the implications of these findings for strengthening TMS in trauma resuscitation teams and for theories on transactive memory and organizational learning.
Nursing Homes as Insurers? The Effect of Provider-Led Institutional Special Needs Plans
Amanda Chen et al.
Health Services Research, February 2026
Study Setting and Design: I-SNPs are a type of Medicare Advantage (MA) plan that allows insurers to differentiate their benefits exclusively for long-term residents in nursing homes. Since I-SNPs first became available in 2006, there has been growth in provider-led I-SNPs where nursing homes are financially integrated or partnered with an insurer to operate a plan for their own residents. We used a difference-in-differences design to estimate the effect of starting a provider-led I-SNP arrangement on several facility-level outcomes, including the share of a facility's long-stay residents who were enrolled in an I-SNP, hospitalizations, medication use, pressure ulcers, physical restraints, falls, and mortality.
Principal Findings: The start of a provider-led I-SNP arrangement led to a 17.0 percentage point (pp) increase (standard error [SE]: 0.006) in I-SNP enrollment among facility residents within 4 years relative to control nursing homes. We also estimate that the start of a provider-led I-SNP arrangement significantly decreased hospitalizations (−1.0 pp, SE: 0.002), increased the use of antipsychotic (0.4 pp, SE: 0.002) and hypnotic drugs (0.3 pp, SE: 0.001), and reporting of pressure ulcers (0.4 pp, SE: 0.002).
Conclusions: Provider-led I-SNPs allow nursing homes to bear financial risk for their residents. These results suggest that this form of risk bearing may successfully reduce utilization (e.g., hospitalizations), but with unclear implications for quality as increased use of sedating drugs and rates of pressure ulcers could either reflect poorer care or retention of sicker patients due to lower hospitalization rates.