3M Commentary

Predictive models and population health risk: Payment (part one)

January 25, 2016 / By Richard Fuller, MS, Norbert Goldfield, MD

In this blog we take a look at risk-adjustment models used in population health. There are two related but distinct uses for population risk models – payment and population health management.  We will be taking a look at each of these. Part one of this blog explores the application of risk models to payment.  Next month, we will analyze new trends in risk adjustment and population health management.

The most widely known application of a population risk-adjustment model in the U.S. is the Hierarchical Condition Category (HCC) model used to set Medicare Advantage (MA) payment rates. A slightly modified version of this model is used in most state health exchanges for commercial populations. The HCC risk-adjustment model uses observed enrollee risk-factors, typically diagnoses reported on claims, to predict resource consumption in a future period; this is then used to establish a prospective payment rate for enrollees. It is well-documented that risk factors obtained in the current period explain the spending pattern of that period far more accurately than for a predicted period. For example, the Society of Actuaries1 has published comparative studies that routinely report that the accuracy of payment-to-cost matching can be doubled using a concurrent model. This is no trivial matter as higher risk, in this case, the risk of incurring an actuarial loss over which you have little or no control, requires a higher premium. A concise objective of risk-adjustment for managed care plans can be summarized with the following statement, “the effectiveness of risk-adjustment can be measured by whether it increases plans’ willingness to provide coverage to high-needs patients, whether it reduces favorable selection and overpayment, and whether it decreases the incentive of high-risk enrollees to disenroll.”2 Payment accuracy is the pillar of this objective. So if the concurrent model is more accurate and is less likely to result in selection against enrollees, then why do we persist in using predictive models to set prospective payment rates?

The Medicare Payment Advisory Commission (MedPAC)3 offers three reasons to favor a predictive model:

1. The timing lag between coding a diagnosis and the diagnosis impacting a risk score makes plans become more engaged in managing enrollee health and costs as they will only realize increased payment for sicker enrollees in the future;

2. The lag between coding a diagnosis and the impact of a diagnosis on a risk-score means that they are less likely to upcode as any gains will appear in the future; and

3. There is less need to correct risk scores with new data as it is technically complex to develop a method for updating risk-scores on a timely basis.

Let’s examine each of these points sequentially:

The timing lag between coding a diagnosis and the diagnosis impacting a risk score makes plans become more engaged in managing enrollee health and costs as they will only realize increased payment for sicker enrollees in the future

First, we need to ask ourselves whether we accept that a risk score linked to that which has happened (concurrent) offers weaker incentives to manage enrollee health and costs than one linked to a future period (predictive). For both models this will largely be driven by how sensitive the risk-adjustment is to reported changes in chronic conditions and acute health events. However, what is suggested by the MedPAC comment is that delaying payment for health deterioration might result in better management in an attempt to stave off the associated costs. But is this credible? Take, for example, hospital admissions and ED visits for those who are chronically ill. These admissions and visits will drive up enrollee cost relative to the rate paid (i.e. reduce profitability) even if a higher severity for observed diagnoses is assigned. This happens in both concurrent and predictive models. However, there will be a greater divergence between the rate received and enrollee cost experienced in a predictive model since many individuals with chronic diseases cycle in and out of activity such as mental health crisis. This is the pattern of super-utilizers4.The prediction of next year’s resource use will not cover the cyclical cost of the current year.  Therefore, a predictive model will more likely view enrollees with high current year health costs as having the potential to be unprofitable – particularly if those costs are potentially cyclical or ongoing (see the next blog) – and subsequently treat them as “bad bets” for enrollment.

Also, what are we expecting plans to manage? For hospitals, we often make clear our view that not all complications are preventable – it is why we prefer to adjust prospective DRG payment rates at discharge using potentially preventable complications (PPCs) rather than denying all complication-related payments. Similarly, we agree that not all readmissions are preventable – so we remove acts of chance such as accidents from counting in readmission rates through potentially preventable readmissions (PPRs). Why should we then expect plans to take on the risk that an otherwise healthy patient will have a traumatic injury or new diagnosis and, as a result, cost money this year that will likely not turn into elevated payment next year? Or, likewise, to manage all aspects of care – such as mental health crisis – as part of their risk based upon a current year description of what might happen to average cost next year.  The use of a predictive model significantly increases these sources of additional risk and negatively impacts the underlying profit margin of select enrollees.

The lag between coding a diagnosis and the impact of a diagnosis on a risk-score means that they are less likely to upcode as any gains will appear in the future

Researchers have extensively documented challenges to the belief that use of a predictive model has removed the incentive to upcode. Study after study (as well as a CMS take back of funds) has detailed the rise of coding intensity in the Medicare Advantage program5, which relies on a predictive model.

There is less need to correct risk scores with new data as it is technically complex to develop a method for updating risk-scores on a timely basis.

MedPAC’s final point, which is that computing concurrent risk scores is problematic due to the timing with which data is received, is an artifact of the way in which the payment system and associated risk adjustment methodology are designed. New York State’s Medicaid program has used a concurrent model for risk adjustment, carrying over the prospective payment weight from base years into future years, with no sign of difficulty.

Before the advent of the HCC model in 2004 (and the PIP-DCG implementation in 2000) there was much discussion around the adequacy and impact of predictive models. Newhouse et al provide an assessment from this period of the impact upon potential selection practices and stinting on service provison6 that could result from the choice of risk-adjustment model. Tellingly, they highlight (without endorsing) that the mood of the time was to err on not paying for services provided lest stimulate supply push. Both risk-adjustment and available data were limited. In fact, the argument over what constitutes random and non-random variation in spending is itself outdated. We have evolved beyond this focus on efficient spending – that which generates value and improves outcomes quality.

We are urging those administering and, often, expanding managed care programs, to adopt an alternative approach to that of using predictive models to blindly curtail supply push. This is particularly important for Medicaid programs containing vulnerable populations, which are often subject to use restrictions, and those encouraging the adoption of smaller plans that find 75 percent of “random” spending to be too risky a proposition. Concurrent models, which offer far greater accuracy in matching payment to cost, can be built upon and augmented with quality outcomes-based targets.  Instead of trying to predict the future amidst uncontrollable costs, we should be offering an average rates based upon the present and transparently removing those costs that can and ought to be controlled. This can be achieved through rate-based adjustments to plans for the frequency and cost of poor outcomes defined by classification systems such as Potentially Preventable Events. For longer-term enrollees, we should adopt payment adjustments that track the average rate of enrollee health status change.7  Surely, concurrent payment models mixed with incentives to decrease potentially preventable events represent a better approach to establishing value than statistically-impoverished predictive payment models?

Richard Fuller, MS, is an economist with 3M Clinical and Economic Research.

Norbert Goldfield, MD, is medical director for 3M Clinical and Economic Research.


¹Winkelman R, Mehmud S. A Comparative Analysis of Claims-Based Tools for Health Risk Assessment.; 2007. https://www.soa.org/research/research-projects/health/hlth-risk-assement.aspx/risk-assessmentc.pdf&usg=AFQjCNHoXWPsxy0BLEPSxd-strIf-LViAA.

²Schone E, Brown R. Risk Adjustment: What Is the Current State of the Art, and How Can It Be Improved?; 2013. http://www.rwjf.org/en/library/research/2013/07/risk-adjustment—what-is-the-current-state-of-the-art-and-how-c.html.

³Medicare Payment Advisory Commission. Improving Risk Adjustment in the Medicare Program.; 2014. http://www.medpac.gov/documents/reports/jun14_ch02.pdf?sfvrsn=0.

4Johnson TL, Rinehart DJ, Durfee J, et al. For many patients who use large amounts of health care services, the need is intense yet temporary. Health Aff (Millwood). 2015;34(8):1312-1319. doi:10.1377/hlthaff.2014.1186.

5Kronick R, Welch WP. Measuring coding intensity in the Medicare Advantage program. Medicare Medicaid Res Rev. 2014;4(2). doi:10.5600/mmrr2014-004-02-a06.

6Newhouse JP, Buntin MB, Chapman JD. RISK ADJUSTMENT AND MEDICARE.; 1999. http://www.commonwealthfund.org/usr_doc/newhouse_riskadj_revised_232.pdf.

7Fuller RL, Goldfield NI, Averill RF, Eisenhandler J, Vertrees JC. Adjusting Medicaid Managed Care Payments for Changes in Health Status. Med Care Res Rev. September 2012. doi:10.1177/1077558712458540.