Saturday, June 12, 2010

Drilling Down Medicare Variation and Dartmouth Atlas

Before taking off to Africa, former fellow student Aaron Holman sent me this editorial commenting on a new analysis of Medicare survey  data - to examine similarities and differences with the results of the mighty Dartmouth Atlas. The editorial is written by the boss of Aaron's (as well as my) former academic department, Health Policy and Management at the Harvard School of Public Health, Arnold Epstein. He does a fantastic job of putting the new analysis by Zuckerman et al. into the context of clinical practice (starting off by comparing clinical practice in two places where he trained), policy analysis, the old Dartmouth findings, and whether this study changes anything.

What does the new analysis say? In a nutshell, Zuckerman and colleagues found that some of the variation in Medicare spending across U.S. regions that the Dartmouth Atlas found can be explained by other factors: demographic factors, (self-assessed) baseline health, other health variables such as new diagnoses or death during the year of observation, income and insurance type as well as supply of health resources.
Although "only" 33% of the 52% differences between the highest and the lowest spending quintiles remain unexplained, I am amazed by how much demographic variables (especially race and ethnicity, see table 1) contribute to the difference in expenditures - they are the second highest coefficient in the model (see figure 1). Obviously, health status and new diagnoses are important, although the exact conditions that account for the differences remain unclear. The authors point out that death of beneficiaries during year of observation shows a huge relative difference between the spending quintiles although the absolute values (and differences) are relatively low.
Interestingly, also the annual family income and dual-eligibility status (for additional Medicaid insurance vs. supplemental private coverage) also plays a huge role as does the supply of health resources -- especially if you look at quintiles 4 and 5 (versus 1) in figure 1.
The only beef or question I have with this study is whether there was an interaction assessed between the independent variables? In other words, the presence of a significant interaction term might indicate that the effect of one or more explanatory variables on spending is different at different values of another explanatory variable. This should be based on a priori hypotheses, e.g. a recent cancer of heart failure diagnosis might explain the death of the beneficiary, low socio-economic status might be associated with lower baseline health status, and so forth.

The Dartmouth Atlas has recently come under critique (see recent New York Times story and blog post by Jeff Levin-Scherz). There are two main problems with their data: 1) there were few multivariate analyses available to assess if other factors might explain part of the variation; and 2) how quality correlates with with costs (in other words, the value provided to Medicare/payors and patients). The latter has been assessed, but I am not aware of any data that does so in a multivariate fashion, taking into account other possible factors. This will certainly be an issue in the near future as everybody is citing the Darthmous Atlas variation in Medicare. We should not forget that it was the decade-long research of the Dartmouth Atlas team who brought these variations in health care to our minds and into our discussion.


Post a Comment