This is a guest post by Adam A. Markovitz, BS and Andrew M. Ryan, PhD.
Our paper published on June 18th in Annals of Internal Medicine found that previous estimates of the effects of the Medicare Shared Savings Program (MSSP) have been overstated. After accounting for selective attrition of clinicians and beneficiaries from MSSP ACOs, the MSSP was associated with an increase in spending of $5 per quarter, statistically indistinguishable from zero.
In their response, McWilliams and colleagues raised a number of objections.
They claimed that ACOs have no incentive to avoid patients with higher risk. We disagree. If patients develop new health conditions while attributed to an ACO, the ACO is unable to include these conditions in patients’ risk scores. While designed to protect against upcoding, this provision creates an incentive for ACOs to prune patients who are decompensating. Second, as identified by David Dranove and colleagues in an essential paper on risk selection, even if risk adjustment is accurate, the outcomes of high risk patients will have higher variance. It is rational for risk-averse ACOs to avoid these patients.
McWilliams et al. also objected to our design and statistical specifications. One complaint concerns our decision to determine patients’ treatment status on the basis of actual CMS assignment. Instead of using actual CMS assignment, papers authored by McWilliams and colleagues have attempted to replicate CMS’ attribution algorithm to approximate ACO assignment. We believe that this approach has led these studies to miss the selective attrition observed in our paper. It is possible that the high risk patients that we observed exiting ACOs were never assigned to ACOs in this prior research.
While McWilliams et al. claim that our use of the true CMS assignment introduces “time-varying inconsistency in how utilization is used to define comparison groups,” this critique is unfounded. Using this approach, and otherwise replicating the preferred specification of McWilliams et al. (with market-year and ACO fixed effects), we found that the MSSP was associated with a significant reduction in spending. It should therefore be obvious that this analytic approach did not introduce bias toward the null.
We present a robust set of results showing evidence of selective attrition within ACOs. This includes evidence that
- higher spending patients and clinicians are more likely to exit ACOs;
- spending estimates attenuate as we progressively add fixed effects for markets, patients, and clinicians;
- ACOs are associated with our falsification outcome, hip fracture, in all standard adjusted specifications (including McWilliams et al.’s preferred specification);
- and the effects of ACOs decrease to approximately zero in our instrumental variables specification.
We also provide robust evidence for the validity of the instrumental variable, including well-balanced patient characteristics, comparable pre-period spending trends, and no association with hip fracture (our falsification outcome). The adjusted longitudinal model failed each of these tests.
McWilliams and colleagues object to this evidence for a variety of reasons, all of them unpersuasive. For instance, they argue that exit of high-risk beneficiaries is simply due to a glitch in the MSSP attribution methodology whereby sick patients are passively drawn from ACOs when they start to only receive care from non-ACO specialists. However, this would not explain our finding of pruning of physicians with high-cost patient panels.
But we also present a simple intention-to-treat analysis that demonstrates the strong influence of attrition bias on standard difference-in-differences estimates (see Supplement, Section H “Intention-to-treat analyses”). Based on actual CMS assignment, patient treatment status was “turned on” when they were first attributed to an MSSP ACO, remaining on for the duration of the study period. This model was not affected by choices related to patient attribution or changes in composition of providers within ACOs or physician groups. Estimates from this simple model from a balanced panel of beneficiaries with beneficiary fixed effects found small and non-significant effects of the MSSP (+$11 per quarter [95% CI, -$13 to $36]). This crucial validity check demonstrates that the effects of the MSSP disappear when patient composition is held constant and selective attrition is addressed.
We agree with McWilliams et al. that the effect of the MSSP is challenging to ascertain. Unlike hospital-based reforms — where hospitals exist before and after the reform is initiated and attribution is straightforward — evaluating the effects of unstable MSSP ACOs is more difficult. The failure of previous work to account for subtle compositional changes within ACOs has led researchers to miss an important source of bias.