What Actuarial Modeling Reveals About Value-Based Care in Oncology

At the Community Oncology Alliance Payer Exchange & Innovation Summit in September, a panel of statistical experts posed a question that cut to the core of value-based care (VBC): When a VBC model produces results, is it measuring actual clinical value or simply reflecting the mechanics of how those results were calculated?

“The central question that we’re going to tackle today is, when we see a value-based care model and a value-based care result, are we measuring value? Or are we just measuring the methodology of the model?” asked moderator Lalan Wilfong, MD, chief medical officer of Navista, Cardinal Health’s oncology practice alliance.

Wilfong posed that question to Andrew Mackenzie, FSA, CERA, MAAA, a health economist and strategic adviser, and Jason Altieri, ASA, MAAA, an actuary and health care data analytics consultant at Milliman. Both have spent their careers building and evaluating the risk contracts that underpin oncology VBC, and both opened with a similar caveat: The concept works, but the execution frequently doesn’t.

“I certainly believe that value-based care can and does work in oncology,” Mackenzie said. “But I think most of the time it doesn’t. Most of the time, the result that is being paid out is biased and definitely not representative of the actual outcome.” Every measured result, he explained, combines 2 things: real changes in the system and measurement error. That error has 2 sources: structural bias built into a contract’s design and ordinary statistical variance. Larger samples shrink the confidence interval around an estimate, but a poorly designed model can produce a wide or skewed interval, regardless of sample size.

The distinction matters more in oncology than in almost any other area of health care. Population health models built for primary care benefit from large, relatively predictable patient pools. Oncology does not have that luxury.

“It’s a lot easier to predict total population spend than it is to predict oncology spend because you have a much more standard distribution of costs and a lot more volume of patients,” Mackenzie said. “In oncology, you have a small sample size and a huge amount of variance at the individual level.” He pointed to chimeric antigen receptor T-cell therapies as an example: A handful of million-dollar claims sit alongside a much tighter cluster of routine treatment costs, producing a distribution that is difficult for any model to fit clearly.

The Missing Variable Problem

Altieri illustrated how incomplete data compound that variance, drawing on his and Wilfong’s experience with breast cancer benchmarking under the Oncology Care Model (OCM). Cost was highly predictable once a patient’s HER2 status was known, he said. Patients fell into a “trinodal distribution” tied almost perfectly to that biomarker. The problem is that payers rarely have access to it.

“It’s very hard to build a model that’s dependent on biomarker results when that’s not data that [are] commonly shared nowadays,” Altieri said. Wilfong added that in the OCM era, a practice’s mix of HER2-positive and HER2-negative patients was largely a matter of chance, and small panels made that randomness look like performance. Today, with far more treatment branches in breast cancer alone, that problem has only multiplied.

Mackenzie framed this as a trade-off between sophistication and transparency. A more complex model, potentially even one driven by artificial intelligence, might price risk more accurately, “but would be much harder to explain what the results are and make connection points in terms of proving what’s driving value or not,” he said.

Altieri raised a related danger: overfitting. A model tightly tuned to the previous year’s cohort and treatment patterns may fail to generalize once new therapies enter the market. “The goal, really, is [to] only put in those kinds of necessary components that give you a meaningful improvement in narrowing that range of outcomes,” he said, and to drop variables—such as general comorbidity flags that matter in primary care but add little signal in oncology-specific spend—that look important but aren’t.

Mackenzie outlined 3 ways of establishing that would have happened without an intervention: the so-called counterfactual, randomized controls, which VBC rarely has; historical controls, vulnerable to shifting treatment patterns; and concurrent controls, which depend on finding a comparable population and reliable data to match against it. Rapid therapeutic change, Altieri noted, is precisely why oncology has lagged behind fields such as orthopedics in adopting bundled payment models. Hip and knee replacements follow well-defined episodes; oncology treatment paths are open-ended and multimodal, making historical benchmarks unreliable almost as soon as they’re set.

What Happens When Incentives Go Sideways?

The panel spent significant time on how VBC contracts can be gamed, sometimes unintentionally. Carving novel therapies out of a model entirely, Altieri warned, creates a perverse incentive to use those same therapies as a tool to exclude patients who require expensive treatment rather than to deliver appropriate care. Coding is another lever. Mackenzie pointed to the well-publicized settlement involving The Villages Health System Medicare Advantage (MA) risk-adjustment practices as an example from primary care,¹ noting that Medicare’s Hierarchical Condition Category model—the basis for MA revenue—explains only about 14% of cost variation in the broader population, and even less within a narrow disease population such as oncology. Altieri described how the OCM and the Enhancing Oncology Model rewarded documented comorbidities, creating pressure to “code better” rather than lower the total cost of care, a dynamic Wilfong summarized bluntly: “Instead of you actually doing something to lower the total cost of care, you’re just gaming the system by coding better.”

A subtler form of gamification involves shifting drug costs between Medicare Part B and Part D, particularly in MA arrangements where a plan may be at risk for one but not the other. “You can show that you saved a ton of money on a patient by shifting that money into a bucket that isn’t getting counted,” Altieri said, without actually improving total cost of care.

Mackenzie added that even in fully at-risk arrangements, Part B and D liabilities behave so differently, especially post Inflation Reduction Act, that shell-game dynamics persist. “What objectives are we actually achieving here?” he asked. “And if we’re just playing shell games of who’s paying for this—is CMS paying for it, is a drug company paying for it, is a provider paying for it—that’s not really sending our energy and our effort where it matters most, which is to the patient.”

Evaluating a Contract Before Signing It

Altieri and Mackenzie offered concrete guidance for practices weighing a VBC proposal. Altieri urged practices to secure actuarial expertise of their own, since payers routinely have staff dedicated to contract design, and practices often don’t. Beyond that, he recommended scrutinizing data access and audit rights, patient volume—small case counts make performance largely a matter of chance—and, critically, running simulations before signing.

“Look at the incentives in the program and really think through what you would do to improve care [and] reduce cost in this model,” he said, rather than signing first and searching for savings afterward. He also flagged benchmarking design as a frequent blind spot: Practices that are already efficient should seek models that compare them with the broader market rather than with their own historical baseline, which effectively penalizes prior improvement.

From the payer side, Mackenzie emphasized disease-specific design, prospective stress testing, and transparency about how confidence intervals and unintended consequences were evaluated. He described a recurring failure mode in which pricing, ongoing reporting, and final settlement each rely on different data sets and teams, eroding trust between payer and practice. “I think trust is paramount,” he said, adding that it can only be built through demonstrated rigor.

Social Determinants: Risk-Adjust or Stratify?

Audience questions pushed the discussion further. Barbara McAneny, MD, CEO of New Mexico Oncology Hematology Consultants, noted that outcomes can vary by as much as 20% based on zip code, and that community practices absorb real costs coordinating care for patients facing transportation and resource barriers, costs that current models largely ignore. Cleo Ryals, PhD, senior director of community oncology research partnership at Flatiron Health, responded that the accepted methodological answer isn’t risk-adjusting for socioeconomic status, because it functions as a mediator of outcomes, and doing so can penalize practices serving marginalized communities. The recommended approach, she said, is stratification, comparing practices against peers treating similarly situated patients, rather than adjusting the target itself.

The panel closed without finding a fix for oncology VBC’s underlying tension, only a shared framework for interrogating it. Sustainable models require asking not just what a result shows, but what design produced it and whether that design was built to measure value or simply to produce a number.

Reference

1. The Villages Health System LLC agrees to $541.5M settlement to resolve False Claims Act allegations. News release. US Department of Justice. August 26, 2026. Accessed September 25, 2026. https://www.justice.gov/opa/pr/villages-health-system-llc-agrees-5415m-settlement-resolve-false-claims-act-allegations