Some health care is expensive but does little to improve health. Eliminating this low-value health care spending while preserving care that improves health outcomes is a key goal of health policy. Prior authorization, where an insurer reviews and approves care before it happens, is intended to target reductions in health care use towards unnecessary care (Brot-Goldberg et al., 2023, Dillender, 2018, Eliason et al., 2021). In practice, doing so is difficult. Medical decision-making is complex and requires time and expertise to distinguish between necessary and unnecessary care. When conducted manually, prior authorization is costly to insurers, difficult to do in the timely manner needed for clinical decision-making, and susceptible to human error (McKinsey & Company, 2021).

Insurers increasingly use algorithms to predict patients’ needed level of care and deny coverage beyond what they determine is necessary (Ross and Herman, 2023a). Algorithms may improve prior authorization by quickly processing large amounts of relevant information and making recommendations based on historical data. This may allow insurers to precisely target reductions in health care use. However, relying on an algorithm may miss important details that are not observable to the model, leading to denials of needed care. As a result, policymakers, the press, and the courts have scrutinized insurers’ use of algorithms (Centers for Medicare and Medicaid Services, 2023, Chu and Nadler, 2023, Ross and Herman, 2023a).

Predictive algorithms increasingly guide decision-making in a wide array of settings—not only in health care but also in criminal justice and child protective services (Ludwig et al., 2024). For example, in the criminal justice setting, judges use predictive algorithms during sentencing or pre-trial detention hearings to assess the risk of re-offending (Albright, 2014, Sloan et al., 2023, Stevenson, 2018). In that setting, the goal of algorithmic decision-making is to target pre-trial detention or longer sentences to a subset of defendants with a high risk of reoffending. In some ways, it is analogous to the goal of algorithms used by health insurers: to reduce unnecessary health care use while continuing to cover a targeted subset of patients who need more intense care. However, in the criminal justice setting – and in many other settings where predictive algorithms are used – the algorithm is designed to support a single human decision-maker. The situation is more complicated when algorithms are used in prior authorization by health insurers. In the health care setting, algorithms are used to support the decision-making of one agent—the insurer—in monitoring another agent—the provider. Both agents influence the final decision of how much care to provide but have very different objectives. While the insurer may be trying to reduce unnecessary care, the provider may be trying to induce demand.

I study a large health insurer’s adoption of a predictive algorithm in its prior authorization process. I examine how health care use changes after the adoption of algorithmic decision-making. I also examine whether health outcomes changed to determine whether the reductions in health care use are sufficiently well-targeted to minimize adverse consequences for patients. I study these questions in the post-acute care setting, where patients receive rehabilitation services after a hospitalization. Post-acute care can be delivered at a skilled nursing facility (SNF), at an inpatient rehabilitation facility (IRF), or through a home health agency. There is substantial ‘waste’ in the US post-acute care system, where expensive care is provided that does not improve patient outcomes (Chandra et al., 2013, Doyle et al., 2017, McGarry et al., 2021, Regenbogen et al., 2019). One reason for this may be the incentives that post-acute care providers face. For example, Medicare pays SNFs on a per diem basis at rates that are well above their marginal costs (Medicare Payment Advisory Commission, 2024). This creates an incentive to induce demand by unnecessarily extending stays. As a result, reducing unnecessary post-acute care has been a broad focus throughout the health care system (Barnett et al., 2019, Biniek et al., 2019, Huckfeldt et al., 2017, McWilliams et al., 2017). One strategy to reduce post-acute care use has been the widespread adoption of prior authorization by Medicare Advantage (MA) plans (Kaiser Family Foundation, 2024b). Insurers that use prior authorization often contract with an outside firm – including NaviHealth, CareCentrix, and myNexus – to directly manage patients’ post-acute care and share in any savings. These firms all use predictive algorithms to aid their decision-making (Chu and Nadler, 2023).

I leverage the partnership of one of these companies, NaviHealth, with a large MA insurer to evaluate the causal effect of algorithmic decision-making on post-acute care use and health outcomes. This publicly announced partnership with Blue Cross Blue Shield of Michigan (BCBS MI) began on June 1, 2019. Using a difference-in-differences design and administrative data, I compare health care use and outcomes for BCBS MI enrollees before and after this change, using traditional Medicare (TM) beneficiaries in Michigan as my control group.

I find that the NaviHealth partnership led to an immediate and sustained decline in SNF length of stay. The average length of stay declined by 2.3 days, a 13% decline relative to the pre-period mean of 18.4 days. This overall effect was driven, in part, by large declines in longer SNF stays. For example, SNF stays over 30 days, which comprised 12.8% of stays in the pre-period, declined by 7.1 percentage points, a decline of 56%. Declines in length of stay were larger for patients admitted to for-profit SNFs compared to non-profit SNFs, suggesting that the NaviHealth algorithm may successfully identify and reduce induced demand by SNFs. Despite concerns about algorithmic bias or discrimination by decision-makers who exercise discretion in how to use the algorithm (Albright, 2014, Davenport, 2023, Obermeyer et al., 2019), I find that the effect of the NaviHealth partnership on SNF length of stay was similar across a number of patient subgroups, including for white and black patients. In contrast to these substantial effects on the intensive margin of post-acute care use, I find no change in the extensive margin of post-acute care (where patients are discharged). For example, I find no effect on the probability of discharge to a SNF. This suggests that the NaviHealth algorithm is most effective at reducing health care use once patients are admitted to a SNF. The lack of observed effects on the extensive margin may be related to what BCBS MI was doing to limit post-acute care use before the adoption of algorithmic decision-making, including their previous prior authorization system. Despite reductions in SNF length of stay, I do not observe any change in patient outcomes, as measured by 90-day readmissions and mortality. This suggests that the care that was denied was not sufficiently high-value to affect these important – but difficult to influence – outcomes.

These results significantly contribute to the literature in two ways. First, my work adds to the literature on the use of managed care techniques to reduce health care use (Afendulis et al., 2017, Agafiev Macambira et al., 2022, Baker et al., 2020, Currie and Fahr, 2005, Curto et al., 2019, Duggan et al., 2018, Jung et al., 2024, Kuziemko et al., 2018, Layton et al., 2022). Recent work has focused on the effects of prior authorization, demonstrating that it can reduce health care use and target these reductions at unnecessary care (Brot-Goldberg et al., 2023, Dillender, 2018, Eliason et al., 2021). My work builds on this line of research by investigating how insurers use technology in prior authorization, showing that this matters for how much health care use is screened out—and the value of the health care that is screened out. Second, my research contributes to the growing literature on the use of algorithms to aid human decision-making by evaluating a high-stakes, real-world application (Abaluck and Gruber, 2016, Agarwal et al., 2023, Albright, 2014, Angelova et al., 2023, Bundorf et al., 2024, Grimon and Mills, 2022, Gruber et al., 2020, Kleinberg et al., 2018, Sloan et al., 2023, Stevenson and Doleac, 2022). Within this literature, my work is unique in investigating the role of algorithmic decision-making when one agent uses it to monitor another agent with a different objective. My research demonstrates that using an algorithm in this way can reduce the use of relatively low-value care.