ACOs suffer astonishingly high turnover rates among their doctors and patients; their patients are unusually healthy; and those unusually healthy ACO patients constitute about 5 percent of each ACO doctor’s panel of patients. These facts appear in three recent reports: CMS’s final evaluation of the Pioneer ACO program, and two papers published in Health Affairs by John Hsu et al.
Each of these facts – high turnover, healthier patients, and few ACO patients in each physician’s panel – poses problems that cannot be solved without a substantial redefinition of the ACO. How are doctors supposed to influence the health and cost of patients they see only sporadically or not at all? How are ACO doctors supposed to lower costs if their sickest and most costly patients are not in the ACO? How are ACOs supposed to alter physician behavior when their physicians see fewer than 100 ACO patients out of a typical panel of 1,500 to 2,000 patients?
This post is the first of a three-part series in which I discuss the documents I mentioned above – the final evaluation of the Pioneer program and the two papers by Hsu et al. In this essay I will review the findings of those documents regarding turnover, biased selection, and numbers of ACO patients per doctor. In the second installment I’ll discuss the implications of these findings for ACOs and for MACRA’s “alternative payment model” program. In the third installment I’ll ask whether the final evaluation of the Pioneer ACO program sheds any light on why the program failed to work as advertised.
The ACO revolving door
The final evaluation of the Pioneer program, published quietly last December, indicates the 23 Pioneer ACOs that participated in that program over the three-year period 2012-2014 lost two-thirds of their doctors and patients during that time.  Here is how L&M Policy Research, the firm that wrote the evaluation, described physician churn: “Looking across 23 ACOs in all three performance years …, 34 percent of Pioneer ACO providers (11,777 of 34,882) were affiliated in all three years.” (p. 25) And here is how L&M described patient churn: “Looking across all three performance years at the 23 Pioneer ACOs… , only 30 percent of aligned beneficiaries were aligned in all three years (352,421 of 1,173,843)….” (p. 29).
In two papers published over the last year in Health Affairs, one published last April and the other in March 2016, John Hsu et al. reported high turnover among doctors and patients in the ACO run by Partners HealthCare, a Boston hospital-clinic chain. That ACO was the second largest of the 32 Pioneer ACOs.  In the April 2017 paper , Hsu et al. reported that only 52 percent of the 748 primary care physicians listed as participants in Partners HealthCare’s ACO during the 2012-2014 period were affiliated during all three years. Oddly enough, 13 percent of those 748 doctors did not have any Medicare recipients assigned to them at all and, therefore, had no ACO patients to lose through the ACO revolving door.
In their March 2016 paper , Hsu et al. reported, “In 2014 …, only 45 percent of the beneficiaries [in the Partners ACO] had been aligned with the ACO since 2012….” (p. 425)
These churn rates should be no surprise to anyone who was paying attention to the Physician Group Practice (PGP) demonstration, a test of the ACO concept conducted by CMS over the period 2005-2010. CMS assigned Medicare recipients to the PGPs using the same plurality-of-primary-care-visit method they employed to assign recipients to the Pioneer ACOs. According to the final evaluation of the PGP demo, the PGPs lost 37 percent of their assigned patients over the first three years (see Table 11-2a p. 222). “PGPs generally retained approximately 70 percent of their assigned beneficiaries from one year to the next,” the report stated, “and … PGPs generally retained approximately 40 percent of their assigned beneficiaries after five years.” (p. 221).
CMS shunts sicker patients away from ACOs
In its first evaluation of the Pioneer program released in May 2015 (with a March 2015 date on it) in which L&M Policy evaluated the program’s first two years, L&M suggested that CMS’s method of assigning Medicare recipients to ACOs could result in favorable selection, that is, the assignment of healthier patients to ACOs. In this final evaluation, L&M did not mince words: They clearly stated that CMS’s assignment algorithm causes highly favorable selection.
To measure the degree of favorable selection, L&M calculated what they called a “spillover group” of patients for each ACO. This group consisted of Medicare recipients within the ACO’s market area who had at least one primary care visit with an ACO doctor during the year in question but did not have enough contact with the ACO’s primary care doctors to be assigned to the ACO by CMS’s plurality-of-primary-care-visit algorithm. These “spillover groups” turned out to be much sicker and more expensive than the groups CMS assigned to the ACOs. Here is how L&M put it (note that L&M uses CMS’s trendy word “aligned” as a substitute for “assigned”): “Aligned beneficiaries tended to have … substantially lower spending compared to those not aligned but receiving at least one qualified service from an ACO provider during a performance year (spillover group).” (pp. ix-x)
Data reported by L&M indicates the “spillover” patients were about 1.6 times sicker and more expensive than the assigned patients. Here are more quotes from the final evaluation: “The average PY1 [performance year 1] expenditures of these two populations differed significantly: $11,605 per aligned beneficiary compared to $18,992 per spillover beneficiary.”(p. 32) “In PY2, for example, [spillover patients] were more expensive than the beneficiaries aligned with ACOs in PY2 – $19,313 per beneficiary compared to $11,768.” (p. 34) “The spillover populations had higher proportions of beneficiaries with dual eligible status, six or more chronic conditions, or more inpatient stays than the aligned populations….” (p. 34)
Finally, it’s important to note that the spillover and assigned groups tended to stay separate over the three years of the Pioneer demonstration. As L&M put it, “[T]here was consistency over time for the two groups: aligned beneficiaries tended to be re- aligned in the following year, and spillover beneficiaries tended to remain not aligned with the ACO.” (pp. 34-35) 
ACO doctors see few ACO patients
Hsu et al.’s April 2017 paper reported that a total of 748 primary care doctors participated in Partners’ ACO during at least one of the three years (2012, 2013, or 2014). Those 748 doctors had an average of 91 ACO patients assigned to them during those three years out of an average panel size of 1,700. “This means that ACO beneficiaries accounted for less than 5 percent of the median physician’s patient panel,” the authors concluded (p. 644). (I calculate the percent to be 5.4 percent.)
To make matters worse, the sick patients that did get assigned to Partners’ ACO were not evenly distributed among the ACO doctors. A few doctors got far more than their share of sick patients. Hsu et al. illustrated how badly skewed the distribution of sicker patients was with this ominous remark: “ACOs’ ability to deliberately select participating physicians year to year … creates a relatively simple mechanism to ‘game’ the risk pool. For example, in our sample, dropping the twenty-two primary care physicians (top 5 percent) with the most high spending beneficiaries (spending more than $81,000) would reduce the mean Medicare ACO spending per beneficiary by 17 percent ….” (p. 646).
Finally, I remind you that the figures I’m reporting here are for the second-largest ACO among the 32 Pioneer ACOs. It’s possible that in smaller ACOs beneficiaries assigned to the ACO account for even smaller percentages of physician panels.
A restatement of the exceedingly obvious
Is it any wonder that ACOs are failing to cut Medicare’s costs, or that when ACO intervention costs are added, ACOs are probably raising total spending? Is it any wonder that ACOs are having, at best, only minor and mixed effects on quality?
I hereby announce the obvious: ACOs cannot possibly work as long as they must labor under the three handicaps I am discussing here – high doctor and patient turnover, limited ability to focus on the sickest patients because CMS is shunting sicker patients away, and few ACO patients per ACO doctor. For the ACO proponents who should have thought about these handicaps before they climbed on the ACO band wagon a decade ago, let me spell it out clearly.
The revolving door problem: It is neither logical nor fair to “hold doctors accountable” for populations of patients that include large numbers of phantom patients (patients doctors never see but are nevertheless “accountable” for) and patients who see numerous other doctors outside the ACO.
The biased selection problem: ACOs probably cannot lower total spending even if they were to receive a random selection of the Medicare population, but we can be certain they cannot do that if CMS continues to assign ACOs a disproportionately healthy population. Most readers of this blog have probably heard of the 20-80 problem – the sickest 20 percent of the population accounts for 80 percent of total medical costs. If it’s possible for ACOs to reduce net costs for at least a portion of their assigned patients, it will be for their sicker patients. But by sending ACOs disproportionately healthy Medicare beneficiaries, CMS is ensuring that a very difficult assignment under the best of circumstances is even more difficult.
The small-patient-pool problem. No one knows what tactics ACOs use that will allegedly make their doctors better doctors, but whatever it is, those tactics are unlikely to be effective when they’re applied to just 5 percent of the patients seen by the average ACO doctor. Even if ACOs could apply their magic to all 1,700 of each doctor’s panel, in other words, even if ACOs were just staff model HMOs with a new name, we would still have no evidence or reason to think they would lower costs or improve quality on balance. After all, if HMOs of any stripe had worked, we would not now be discussing ACOs and “medical homes” and other managed care fads invented decades after the HMO was unleashed on the populace. But when that ACO magic is applied to such a tiny percent of each doctor’s patients, it becomes even more difficult to imagine how ACOs are supposed to lower costs or improve quality.
In my next post I will ask whether these three problems – high turnover, biased selection, and small ACO patient pools – are fixable without a substantial redefinition of the ACO. My answer will be no. I’ll argue that ACO advocates can either redefine the ACO to look like staff-model HMOs, or they can radically redefine the ACO so that it focuses on a few clearly defined chronic diseases or subsets of patients. I’ll recommend the latter. I’ll also argue that even if ACO proponents wanted to choose the first option and risk another patient and doctor backlash against HMOs dressed up as ACOs, there’s no reason to believe that the HMOs in ACO drag would work any better today than they did in the closing decades of the last century.
 Thirty-two groups signed up for the Pioneer program and participated in the first year, 2012, but only 23 participated during all three years.
 Partners HealthCare was the second largest of the 32 ACOs if we use number of attributed Medicare beneficiaries as our measure of size. Partners ranked lower if we use number of participating doctors as our measure. The CMS final evaluation presents data on the number of physicians and patients in each of the 32 ACOs that participated for at least one year in the Pioneer program (see Table 30 p.99 and Table 32 p. 102). This data is presented for each of the three years (2012, 2013, and 2014) and for “any year.” The “any year” data show how many doctors and patients passed through the ACO revolving door over the three-year period. Partners was the second largest of the 32 ACOs as measured by total assigned patients in 2014 (77,135) and in any year (98,196). Heritage California ACO was the largest (96,617 in 2014 and 162,264 in any year).
 Deep within the final evaluation of the Physician Group Practice demo published in 2012, one can also find evidence that the plurality-of-visit method CMS uses guarantees that ACOs will get healthier patients. According to that evaluation, the average risk score for the Medicare recipients assigned to the ten PGPs in year one was 0.921 (1.0 equaled average risk). Interestingly, the report demonstrated that these risk scores would have risen to approximately average (that is, 1.0) if the method of assignment had been merely “one or more visits,” and would have fallen to 0.898 if “a majority of visits” had been used (see tables at page 222 of the final report ). In other words, the more patient loyalty CMS’s assignment algorithm requires before a patient can be assigned to an AC), the greater the favorable selection enjoyed by the ACO. The biased selection problem caused by the plurality-of-visit assignment method is worse for ACOs enrolling non-Medicare populations under 65, because the proportion of that population that fails to seek any medical attention in the course of a year is much higher than the proportion of those over 65.