Advantage Medicare Advantage

A long-standing and important question about Medicare Advantage (MA) is the extent to which outcomes for beneficiaries who enroll in participating private plans are due to selection or plan design and care management. Put simply, do the plans cause enrollees to be healthier or do they merely attract healthier enrollees? A recent NBER paper by Christopher Afendulis, Michael Chernew, and Daniel Kessler (ungated version here) exploited a clever design to address this question. The first thing to know is that MA plans are paid in a complicated way (PDF) only one feature of which you need to understand. Since 2001, plans in Metropolitan Statistical Areas (MSAs) with populations of 250,000 or greater receive a higher government payment rate. Higher payment allows those plans to offer more generous benefits and wider networks; in turn, they attract greater enrollment. However, it's a reasonable assumption that MA enrollees in areas just above the 250,000 cut-off are very similar to enrollees in areas just under the 250,000 cut-off. To the extent that is true, the difference in outcomes between the two sets of enrollees can be attributed to plan factors, not individual-level ones. The authors exploit this population-based discontinuity in payment to analyze MA outcomes, circumventing the confounding of selection for reasons just explained. This is known as a "regression discontinuity" design, though it is just a specific case of an instrumental variables (IV) set-up. The key is to focus analysis on a subsample around the discontinuity. In this case, the authors only included plans offered in MSAs with populations between 100,000 and 400,000. An indicator variable for above/below the 250,000 threshold serves as "the instrument," meaning the source of randomness. The assumption is that MA outcomes are related to this instrument only through its effect on MA enrollment, just as would be the case in random assignment in a randomized controlled trial (RCT). Now, this is not an RCT, so one could quibble with the instrument. To keep this post reasonable length, I will not do so. I actually think it's quite good for this purpose. I should note that the authors also control for some observable factors likely related to both MA enrollment and outcomes, like demographics, hospital market characteristics, number of beneficiaries and historical population. Using 2009 data, the authors found that

MA reduces the probability of [hospital] admission by 3.4 percentage points (standard error 1 percentage point). On a base admission rate of 16.9 percentage points [table 1], this amounts to a 22 percent reduction. Part of the effect of MA on admissions arises out of its impact on ACS [ambulatory-care sensitive, aka preventable] conditions; MA reduces the probability of an ACS admission by 0.69 percentage points (standard error 0.38 percentage points, significant at the 10 percent level). There is also a significant IV effect of MA on mortality of 0.7 percentage points (column (6) standard error 0.19 percentage points); on a base of 4.6 percentage points, this amounts to a 15 percent reduction.