Brooke Kenney is a data analyst working with MVC

Chad Ellimoottil

Chad Ellimoottil is a Postdoctoral Fellow working with MVC.

“Why does the MVC risk-adjustment model differ from the one that Medicare uses?” This is a question that we have heard multiple times from hospitals. To address this question, we embarked on a series of analyses quite like peeling back the layers of an onion. The highlights of our journey are detailed below. In the end, we found three areas of discordance and we even made some changes to our model to make it better.

But first, what is risk-adjustment? Hospitals treat a variety of patients whose cost may be influenced by the severity of their physical condition. Since many healthcare programs link quality of care to cost, it is necessary to account for serving costly patients in a fair, standardized manner. Risk-adjustment (RA), therefore, is a statistical method that helps “level the playing field” by accounting for differences in case-mix.

Difference #1

The Cohort: We compared the risk adjustment model used by MVC and CMS for two conditions; AMI and Joint Replacement. The key differences in the cohorts are that MVC includes statewide BCBSM and Medicare patients, resulting in a younger population. The CMS cohort contains nationwide patients age 65 or older and applies several exclusions to patients with: HMO insurance, less than 12 months of enrollment data, specific discharge disposition categories, and others. When the observed episode costs of the MVC cohort were run through the CMS model, and vice versa, there was minimal impact on hospital ranking for total episode cost.

Difference #2

The Statistical Model: As it turns out, the patients we serve in the great state of Michigan are variable and complex. They require a regression analysis to explore the relationship between patient demographics and comorbidities on cost. In addition, there are a number of regression techniques developed to study a particular cohort. Perhaps the different models employed in our two RA methods are contributing to total episode cost differences? MVC uses a 2-step linear regression model with Poisson distribution to calculate an observed to expected ratio for each hospital. This ratio is multiplied by the state mean expected cost to produce a risk-adjusted mean cost for each hospital. CMS uses a random-effects model to calculate a predicted to expected ratio for each hospital. This ratio is multiplied by the national mean cost to estimate the risk-adjusted mean cost for each hospital.  To test for differences, we calculated the correlation of hospital ranking by holding the independent variables constant (first with MVC variables and next with CMS variables) and applied them to the two models. There was a strong correlation between the MVC and CMS hospitals, leading us to conclude that the model differences do not matter much.

Difference #3

The Comorbidities:  With a couple key layers removed from our onion, the investigation of cost differences has been narrowed down to the independent variables. This required a performance evaluation on the two types of comorbidity definitions; Elixhauser (MVC method) and Condition Category (CMS method). For example, does one method capture more comorbidities in the data than another? Does this capture noticeably impact the adjusted cost? In our effort to determine a preferred method, we found that CCs captured more ICD9 codes than Elixhauser by sheer volume of co-morbid categories, 180 to 29, respectively. After our comparisons of the MVC and CMS risk-adjustment models we found that CCs outperformed models with Elixhauser and we will be changing to Condition Categories to define comorbidities.


Based on these analyses, we will be incorporating multiple new changes to the risk adjustment model:

  • We will start using Condition Categories instead of Elixhauser categories
  • We will use a 6-month look back in the claims data from the time of the index-admission to identify comorbidities as well as the “present on admission” indicator to ensure that we are capturing comorbidities in the robust manner possible
  • We will also add professional claims to this loopback period in addition to the facility claims.

These risk-adjustment changes have occurred as of Feb 8th, 2016!  We call these changes “Update 1”. We are working on “Update 2”, which includes an analysis of our condition-specific risk-adjusters.

Last but certainly not least, we value feedback from our MVC participants on how we can customize our models to better define the patients we serve throughout the state. In fact, a team within the Coordinating Center is currently conducting site visits at all hospitals just to hear from you. We look forward to what 2016 has in store for MVC.