Risk Adjustment:Vox::Narrow Networks:538

There are two recent health policy articles by interested lay expertise sites that have me scratching my head. In Vox’s case, I am seeing a conclusion without context. For 538, I can not figure out the model that leads to a core assumption. These sites’ jobs are to inform the public and in these cases I think they can do a better job of their job.

Let’s start with Vox as Sarah Kliff looks at a Society of Actuaries analysis of risk scores on the individual market. She draws a very strong conclusion.

Between 2014 and 2015, SOA finds that Obamacare’s average risk scores went up by 5 percent. This means that the overall pool of people on the marketplace were sicker in 2015 than 2014. You can see the data here, in a table from the report.

The Society of Actuaries draws a much weaker conclusion

Risk measures published in the CCIIO release show that the average measure of risk increased from 2014 to 2015. Increased risk scores may be a combination of identification through better coding as well as a measure of the actual population health….

The program is still too immature to draw conclusive inferences about the future of the pool or marketplaces.

Vox saw a number that went up (bad) and wrote a story with no context.

As an insurance professional who spent half of today working on risk adjustment optimization problems, there are several critical pieces of context that need to be addressed before deciding whether or not the real increase in reported risk scores is an actual story or an artifact of Red Queen Race incentives. We know the following about 2014 and 2015 Exchanges.
1) They are a new program
2) There are a lot of new insurers in a revamped market segment
3) There is very thin claims histories
4) Significant 2014 marketshare went to technically naive insurers
5) The population was sick
6) Significant population moves occurred
7) Significant gaps in medical history were common for a large proportion of covered lives.

With this context, I have no idea how to evaluate whether or not an increase in reported risk scores is an actual increase in pool illness burden or merely an artifact of carriers getting better at data mining historical data for probable diagnosis codes and then getting their PCP’s to code those high value diagnoses on a processed claim. Co-ops that were surprised by high cost members in 2014 may have aggressively chased for treated but non-coded diagnoseses in 2015. We don’t know that. I strongly suspect better code chasing is a significant element of the increase in risk scores. I can’t tell you how much though. If we wanted to get a good clue on that we either need to look at some harder to game numbers like age/gender breakdowns or the number of kidney transplants per 1,000,000 covered months or the number of Hep-C cures dispensed. Are those numbers changing up or down? Or more basically, is the Per Member Per Month costs dramatically changing? None of these numbers in and of themselves are dispositive, but triangulation plus industry specific knowledge allows for a coherent story and vision to be built.

If they are going up, they support a contention that the risk score is mostly reflecting actual pool health risk. If they are flat or declining, then the story is better coding efficiency.

Now 538’s Tim Mullaney looks at the problem with broad network, high reimbursement Exchange carriers. He quotes an analyst that is offering an implicit model that I can’t figure out:

More importantly, middle-class customers who are used to group insurance are often much less willing to accept narrow networks than poorer people who are glad for any coverage, said Les Funtleyder, a health care fund manager at E Squared Capital. That could put insurers in a catch-22: Given the prices individuals are willing to pay, insurers can’t afford to offer networks that are broad enough to attract healthy customers. And if insurers can’t lure the young, healthy consumers who haven’t yet signed up for Obamacare, they can’t turn a profit. “It’s just math,” Funtleyder said. “If you give people more benefits, it costs more.”

From my point of view, there is a model here that I think is anchored in Employer Sponsored Insurance land (where costs are extremely well hidden) that expects magical pixie dust to allow for plans that don’t say no to do so cheaply.

What insurers need are lots of low cost people. Preferably those people are low costs because they are intrinsically healthy. They could be low cost because they just don’t get problems addressed. We need low cost people to cover the people with a double organ transplant scheduled tonight. The last buyer for a policy is fundamentally indifferent between the peace of mind of paying their monthly premium as an almost guaranteed net loss and the fear of getting hit with a quarter million dollar claim. This model assumes that the last person to make the buy and the first person to make the no-buy decision are both very healthy with very few anticipated medical expenses. The low cost buyer is mainly concerned about price not network for a given actuarial value because in the vast majority of cases, they are either not touching the network at all, or they are superficially touching it for perhaps a well visit, perhaps a flu shot, and perhaps an urgent care visit. They don’t care that they don’t have the option value of going to a high end specialty hospital for cancer as they don’t think they’ll get cancer as a 26 year old triathlete. They don’t care that care coordination/case management runs people through the ringer to get an MRI authorized as they most likely will not need an MRI. They care that the premium competes for scarce dollars and that the incremental spread between the monthly premium and the mandate penalty is more than their expected value plus a risk aversion multiplier.

Healthy people who anticipate that they will remain healthy are the least likely people to buy on network as a core selling point. They buy on price. Networks matter for sick individuals. 538 is allowing for a very unusual model to be advanced and there is no commentary to either support why this implicit model makes sense or that there is a problem with the statement.