CMS has decreed that Accountable Care Organizations (ACOs) need to be on track to assume risk.

According to the old rules, an ACO could stay in the one-sided risk model as long as they wanted.

Upside only? Most ACOs took that deal. Of the 125 ACOs that renewed in 2016, only 12 took on the potential for downside risk.

According to the new rules, there are only two tracks — BASIC and ENHANCED. And both tracks lead to two-sided risk. That means that in a few years, even the most risk-averse ACOs will have to pay CMS back if their beneficiaries’ healthcare costs go up.

ACOs have until February 22, 2019 to decide what risk track they want to be on.

For most ACOs, the decision will be simple — whatever option delays the onset of two-sided risk the longest is the option they will go for. And it’s easy enough to understand why:

We learned from Daniel Kahneman and Amos Tversky that loss aversion is stronger than greed. Most people would say it’s better not to lose $5 than it is to find $5.

In that same vein, the boards and leadership teams of most ACOs will say it’s better to avoid paying back shared losses (no matter how unlikely shared losses actually are) than it is to have a chance to earn more shared savings.

Still, there will be ACOs that want to make the smartest decision possible. For those ACOs, I want to help.

If I’m going to help, I need to understand ACOs at a much deeper level. And that understanding starts with an assumption:

Assumption: Not all ACOs are created equal.

If not all ACOs are created equal, what makes them different? Which ones are similar enough to be considered “peers”?

The ACO datasets published by CMS are an embarrassment of riches, and I spent a long time exploring before I found an answer I was satisfied with.

To give you an idea how comprehensive these datasets are, here is my working list of things I want to explore later:

  • What ACOs are driving patients to lower levels of care over time (fewer hospital inpatients, more primary care), as opposed to benefiting from structural advantages (like their underlying population getting healthier).
  • What it means to care for patients with lower socioeconomic status (in terms of cost and quality)
  • What impact ACO quality scores have on cost performance
  • Does competition save beneficiaries money? (AKA, do ACOs with dominant market share have worse cost performance?)

Every one of those might make for an interesting analysis down the road. AND every one of those analyses will be more meaningful if we can compare peers vs. peers, as opposed to all vs. all.

Keeping it super simple — How big is a ‘big’ ACO?

ACOs come in wildly different sizes, and any serious attempt at peer grouping them will start with size.

There are three straightforward measures we could use to measure the absolute size of an ACO:

  • How many assigned beneficiaries they care for
  • How many providers work for them
  • How much healthcare their assigned beneficiaries use (total healthcare expenses)

In my first attempt at peer grouping, I gave these three data points to the algorithm… and got back garbage.

Lesson learned — Normalize before you cluster if you want to pass muster

The smallest ACOs have fewer than 2,000 people they care for. The biggest have over 100,000. Healthcare expenses for the smallest ACOs might be less than $15 million. Expenses for the biggest are over $1 billion.

Here’s the problem — when you use a clustering algorithm, it measures how far apart any two data points are from each other. To the algorithm, the smallest ACO’s expenses at $15 million were “bigger” than the biggest ACO’s patient base of 100,000.

That means the algorithm let the differences in expenses swamp the differences in every other data point. Not ideal!

I wanted to weight patients, providers, and dollars equally, so I normalized everything on a scale from zero to one.

With the normalization problem solved, I arrived at the next perplexing problem of peer grouping:

Lesson learned — The analyst has UNLIMITED POWER to decide how many clusters there are

When you use a clustering algorithm, you have to tell it in advance how many clusters you want there to be.

That puts a LOT of power into the hands of the person doing the analysis (in this case, me). I just outlined four questions about ACOs I’d like to answer.

And for every question I tackle, I’ll want to use the peer groups I’m developing now.

With that in mind, I spent a long time with the dataset, tried a lot of different combinations and permutations, and eventually came out with peer groups I was satisfied with.

Here’s what that looked like:

The four sizes of ACOs — Small, Medium, Large, and Mega

According to this analyst, there are four sizes of ACOs in this country.

How did I arrive at the number four?

Aside from spending a long time getting comfortable with the dataset, at the end of the day four groups just looked right.

To get more specific, here is what I noticed:

  • When there were three groups, it seemed like half of the ‘Large’ ACOs were more similar to ‘Medium’ ACOs than they were to the largest ACOs (in other words, the dividing line between medium and large didn’t seem to be in the right place)
  • When there were five groups, the algorithm split into ‘large ACOs with a high provider-to-beneficiary ratio’ and ‘large ACOs with a low provider-to-beneficiary ratio’. That distinction didn’t seem especially useful.
  • When there were four groups, the differences were all about size, and the small, medium, and large groups seemed as homogeneous as it would be possible to get.

So how big is a ‘large’ ACO? Here is a table of the 25-75th percentile values for each size.

ACO SizeBeneficiariesProvidersExpenditures
Small6,800 – 12,30080 – 300$69M – $127M
Medium14,400 – 27,300680 – 1,180$149M – $279M
Large40,600 – 59,5001,440 – 2,350$430M – $585M
Mega74,600 – 106,5003,460 – 5,020$730M – $1.1B

Seeing the groups in this table makes me glad I separated them.

Here you can see that large ACOs are five times as big as small ACOs, and mega ACOs are double the size of even large ACOs.

Without peer groups, any analysis of ACOs would be comparing wildly different entities.

In the coming weeks, expect me to start tackling questions of ACO performance.

The point of the Medicare Shared Savings Program is to save the government money. Who is successful at that? How much credit do the ACOs deserve? Are they just riding a wave of lower cost driven by smarter incentives from payers? Are they leveraging community resources and actually keeping their populations healthier?

These are the questions I’m interested in answering. Not that they will be easy to answer. But if we can answer them, it might be possible to take what’s working and replicate it elsewhere. That would be a good thing.