NUR 653 Module 2 Risk Stratification Paper Example

Stratifying a Population for Care Management Without Being Fooled by Regression to the Mean or a Biased Target

[Student Name]

Southern New Hampshire University

NUR 653: Population Care Management

Module Two Risk Stratification Paper

[Instructor Name]

[Date]

Stratifying a Population for Care Management Without Being Fooled by Regression to the Mean or a Biased Target

Population care management begins by deciding who should receive the most intensive help, since no organization can give everyone a nurse care manager. The obvious choice is to target patients who cost the most, on the logic that a small group accounts for a large share of spending and that helping them will save the most money. Mesa Valley Health Partners, a composite accountable care organization with 42,000 attributed patients, proposed exactly this: enroll the 840 patients in the top 2% of last year's costs in intensive care management and judge the program by the change in their costs. This paper examines the proposal. It argues that targeting by past cost invites two errors, mistaking regression to the mean for program success and building inequity into the target, and it recommends a stratification approach based on current clinical need and likelihood of benefit, with an enrollment design that allows fair evaluation.

Regression to the Mean in Mesa Valley's Own Data

Patients who are in the top tier of costs in one year are often there because of an unusual event: a heart attack, a surgery, a serious infection. The following year, many recover and their costs fall without any intervention, a statistical pattern called regression to the mean. To see how strong the effect would be, the population care manager examined Mesa Valley's claims from two earlier years, before any care management existed. Of the patients in the top 2% of costs in the first year, only 29% remained in the top 2% the next year, and their average costs fell by 46%. A program enrolling these patients and comparing their costs before and after would have appeared to cut costs almost in half while doing nothing at all.

What a Randomized Trial Showed

The danger is not hypothetical. Finkelstein et al. (2020) conducted a randomized controlled trial of the Camden Coalition's care management program for patients with high use of hospital care and complex medical and social needs, a program widely praised after before-and-after comparisons showed large drops in hospital use. In the trial, 800 hospitalized patients were randomly assigned to the program or usual care. Within 180 days, readmission rates were nearly identical in the two groups, about 62%. Both groups' hospital use fell sharply after enrollment, which is exactly what regression to the mean predicts. The authors concluded that the earlier apparent success reflected this statistical pattern and that rigorous evaluation is needed before such programs are expanded.

When the Target Itself Is Biased

The second error concerns what is being predicted. Obermeyer et al. (2019) examined a widely used commercial algorithm that selected patients for care management based on predicted future costs. Health systems had historically spent less on Black patients carrying the same disease burden, a legacy of unequal access, so the model read lower spending as better health and judged Black patients to be healthier than equally sick White patients. At any given risk score, Black patients had more chronic conditions. Using cost as a stand-in for need built existing inequities into the selection process. Mesa Valley's proposal to target the highest-cost patients would inherit the same problem. A check of last year's top 2% found that patients enrolled in Medicaid, whose use of specialty care is limited by access, were underrepresented relative to their burden of chronic disease.

The Limits of Risk Scores

Predictive models can help, but they have limits. Kansagara et al. (2011) systematically reviewed models predicting hospital readmission and found that most had only modest ability to distinguish patients who would be readmitted from those who would not, and that few included social or functional factors. A model that predicts poorly will enroll many patients who would not have been hospitalized and miss many who would. Scores are best used as one input alongside clinical information and the judgment of the primary care team, which often knows which patients are struggling in ways the data do not capture.

A Fairer Approach

Mesa Valley should stratify on current, measurable need and on the likelihood that care management can help. The proposed approach identifies patients with persistent rather than one-time high use, such as two or more hospitalizations in each of two consecutive years; patients with poorly controlled conditions that respond to management, such as an A1c above 9% or blood pressure above 160/100 on repeated readings; and patients with positive social needs screens combined with a chronic condition. Primary care teams may refer any patient they believe would benefit. A predictive score will be used to rank patients within these groups, not to define them. Because clinical measures such as A1c and blood pressure do not depend on how much care a patient has received, this approach reduces the bias built into cost-based targets.

Designing Enrollment for Fair Evaluation

The program has capacity for about 600 patients in its first year, but the proposed criteria identify about 1,400. Rather than choosing arbitrarily, Mesa Valley will enroll eligible patients in three waves, four months apart, assigned by random selection. Patients waiting for later waves receive usual care in the meantime and serve as a comparison group for those already enrolled. This staggered design is fair, since everyone eligible will eventually be offered the program, and it allows the organization to compare outcomes between enrolled and not-yet-enrolled patients who share the same starting point, so regression to the mean affects both groups equally.

The approach also asks what care management can realistically change for each group. Patients with persistent high use driven by uncontrolled chronic disease and social barriers are the most likely to benefit from a nurse who visits, coordinates and follows up. Patients whose high costs stem from a planned surgery, a new cancer diagnosis under active specialty treatment or end-stage illness already receiving palliative care need different support and are routed to those services instead. Matching patients to the kind of help they can use keeps the care managers' caseloads focused where they can make a difference.

Conclusion

Targeting last year's costliest patients seems efficient, but Mesa Valley's own data show that any program would appear to succeed with them, and national evidence shows that cost-based targets can encode inequity. Stratifying on persistent, measurable, modifiable need, using scores only to rank and team referral to catch what data miss and enrolling in randomized waves gives the organization a fairer way to choose patients and an honest way to learn whether its program works.

References

Finkelstein, A., Zhou, A., Taubman, S., & Doyle, J. (2020). Health care hotspotting: A randomized, controlled trial. New England Journal of Medicine, 382(2), 152-162. https://doi.org/10.1056/NEJMsa1906848

Kansagara, D., Englander, H., Salanitro, A., Kagen, D., Theobald, C., Freeman, M., & Kripalani, S. (2011). Risk prediction models for hospital readmission: A systematic review. JAMA, 306(15), 1688-1698. https://doi.org/10.1001/jama.2011.1515

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. https://doi.org/10.1126/science.aax2342