Better Fit, Same Game? Risk Adjustment, Coding Incentives, and the Limits of Prediction
本文评估了Franklin模型在医疗保险风险调整中的表现，指出其预测改进主要集中在低中成本人群，对高成本人群改善有限，且可能加剧编码激励问题，对减少逆向选择或促进高需求患者竞争的效果存疑。
Risk adjustment is not a prediction contest. It is an incentive design tool. Franklin represents a meaningful technical advance in predicting health care spending using the same underlying inputs as the CMS-HCC model.1 But the relevant policy question is not whether model fit improves. It is whether and how using the model will change what insurers do. The success criterion for risk adjustment is not higher R2. It is whether investing in high-need patients becomes a dominant strategy rather than a fiscal liability. Risk adjustment was developed to address a structural problem in prospective payment systems: plans that became attractive to very sick patients were financially penalized for doing so. The canonical example is early AIDS care. A plan that built superior systems for high-cost, high-need patients risked attracting those patients disproportionately and incurring losses. Risk adjustment was intended to remove that penalty. A simple test of success is whether plans advertise themselves as the best place for patients with the most complex needs—for example, those using ventilators or living with severe disability. By that standard, risk adjustment has not yet succeeded. Franklin improves predictive performance along several dimensions. Using age, sex, and diagnosis codes, it extracts more information from diagnostic data than the CMS-HCC model, particularly among beneficiaries with zero or one Hierarchical Condition Category (HCC). It substantially increases model fit when evaluated using log-transformed costs and reduces dispersion in predictive ratios. These are real and important achievements. However, the magnitude of improvement depends on how model fit is evaluated. The authors write that Franklin produces “three times better model fit,” as reflected by an improvement in R2 where the dependent variable is log costs (0.44 vs. 0.15). When the dependent variable is dollars, as opposed to the log of dollars, the improvement is much smaller (R2: 0.121 vs. 0.110), with only modest changes in mean absolute error. This pattern reflects the model's training objective. Franklin minimizes prediction error in log of costs, which weights percentage errors equally across the cost distribution. Improvements are therefore concentrated in the low and middle of the distribution. Gains in predicting the highest-cost cases—where spending is most concentrated and where selection incentives are strongest—are more limited. Payment and incentive design operate in levels rather than logs. To the extent that improvements in prediction are concentrated among the lowest-cost beneficiaries, the implications for selection may be limited, as differences in spending at the very low end of the distribution are small in absolute terms and therefore generate relatively weak financial incentives. To the extent that better prediction is most consequential for changing insurer behavior at the upper tail of the cost distribution, the more modest dollar R2 improvement is the more policy-relevant benchmark. This distinction matters for understanding how improved prediction affects insurer behavior. Risk adjustment operates at the intersection of three incentive problems: reducing selection against sicker patients, encouraging investment in care for high-need populations, and limiting incentives for excessive coding. Diagnosis-based risk adjustment has partially mitigated selection. It has not clearly induced aggressive competition to serve the sickest patients. It has, however, intensified coding. Any evaluation of Franklin should consider all three domains. Early discussions of selection sometimes focused on overt strategies—such as locating enrollment offices in inaccessible settings or targeting marketing to unusually healthy populations. In practice, such forms of selection are neither necessary nor typical. It is also important to recognize that selection does not arise solely from plan behavior. Beneficiaries face real tradeoffs when choosing between Medicare Advantage and Traditional Medicare. For many beneficiaries, MA offers lower premiums and additional benefits, but with more limited provider networks and greater use of prior authorization and other utilization management tools. Even in the absence of any plan effort to attract or deter high-need enrollees, these tradeoffs generate selection: beneficiaries expecting to use relatively little care will be more likely to enroll in MA, while those anticipating greater use of services will typically prefer the flexibility of Traditional Medicare. Similarly, beneficiaries with high care needs may be more likely to disenroll from MA over time. At the same time, plans do influence selection through design choices that differentially attract or deter higher-need beneficiaries. These may include network composition, marketing strategies, benefit design, and utilization management policies such as prior authorization. Such mechanisms can shape enrollment patterns without explicit targeting, and their effects may be only partially mitigated by improvements in risk adjustment. To date, there is limited evidence that improvements in risk adjustment alone have led plans to actively compete to attract the highest-need patients, although the growth of Special Needs Plans suggests that more targeted models of care can emerge when payment systems are sufficiently aligned with the needs of specific high-cost populations. The institutional context in which risk adjustment operates further shapes its consequences. In Medicare Advantage, differential coding raises total federal spending. Differential coding increased Medicare Advantage payments by tens of billions of dollars annually compared to the counterfactual of identical coding in Medicare Advantage and Traditional Medicare. While earlier analyses suggested figures on the order of $40–$50 billion per year, more recent MedPAC estimates incorporating the transition to V28 are substantially lower—approximately $22 billion in 2026 [1]. Even at these levels, coding incentives represent a significant fiscal issue for the federal budget, as well as raising concerns about billions of dollars being spent by Medicare Advantage insurers in socially unproductive efforts at finding and documenting diagnoses. In contrast, in budget-neutral environments such as many Medicaid programs and ACA marketplaces, payment is zero-sum, and increased coding redistributes payments across plans but does not increase total payments. The same model can therefore have very different welfare implications depending on the payment system in which it is embedded. Franklin's implications for coding incentives are ambiguous. The model's use of a broader set of diagnoses and more flexible functional forms may make it more difficult to identify a small number of “lever codes”—individual diagnosis codes whose addition substantially raises a beneficiary's risk score, and therefore the plan's payment—that have characterized coding optimization under HCCs. However, it may also increase the returns to more comprehensive coding and patient profiling. Plans are likely to respond not by abandoning coding strategies, but by investing more heavily in capturing complete diagnostic profiles and in analytic approaches to maximizing risk scores under the new model. The CMS-HCC model has led to the growth of an industry of vendors focused on risk score optimization—chart reviews, retrospective coding programs, and vendor analytics—precisely because payment depends on documented diagnoses. Franklin does not change this basic architecture; it changes the optimization target. Plans that have invested in comprehensive coding infrastructure under HCCs are well-positioned to redirect those investments under Franklin. The consulting and analytic services that optimize coding under the current model will adapt, not disappear. Evaluated across the three domains that matter for risk adjustment design, Franklin's performance is mixed. On selection, it offers modest improvement: better prediction across the low and middle of the distribution, but limited gains in the high-cost tail where adverse selection incentives are strongest. On investment in high-need care—the domain where risk adjustment has been least successful—the evidence is absent; improved statistical fit does not by itself change the financial calculus that has made attracting and retaining the sickest beneficiaries a fiscal liability under the current system. On coding, the model shifts the optimization landscape rather than shrinking it. The net effect on insurer behavior remains uncertain. If Franklin or similar models were considered for implementation, several governance issues would warrant attention. Policymakers would need mechanisms to limit aggregate coding intensity in Medicare's treasury-exposed environment, as well as careful modeling of behavioral responses, including changes in coding investment and enrollment strategies. The governance challenge is not merely theoretical. The primary existing mechanism for offsetting differential coding in Medicare Advantage—the coding intensity adjustment—has been set at 5.9% in every year since 2018, despite strong evidence that it should have been larger than 5.9% in 2018, and strong evidence that the gap between MA and fee-for-service coding has widened since then. If the policy infrastructure for managing coding intensity under the current model has proven resistant to accurate calibration, the prospect of recalibrating it for a more complex model warrants serious attention before implementation. Improvements in prediction should arguably be complemented by other policy tools. One such approach is to reduce plans' exposure to very high-cost beneficiaries through reinsurance. While models such as Franklin improve prediction across the low and middle of the cost distribution, reinsurance directly targets the upper tail, where spending is most concentrated and where selection incentives are strongest [2]. If reinsurance were implemented using Medicare Advantage encounter data, an additional potential benefit would be improved data quality. Reinsurance and risk-adjustment address different aspects of the incentive and equity problems and may be most effective when used together. Risk adjustment was intended to make compassion financially sustainable. Franklin may reduce statistical mispricing. But the decisive test is not R2. The decisive test is whether insurers compete to serve the sickest patients—not merely to document diagnoses more effectively. The author has nothing to report. The author declares no conflicts of interest. The author has nothing to report.