Cole, interaction factors and risk adjustment

We love Cole.

We loved and winced at the stories of faceplants and broken bones from naked mapping, sidewalk walking, dog walking, cat assaination attempts and any number of other things.  We love that Cole is making a good and honest go at the entire adulting thing.  It’s been a while since we’ve heard a classic Cole injury story.  The worst that has happened was the car in the cow field and that was not even his fault. Damn it Cole, I drafted this as the kids were getting ready for bed and you have to fall through the floor….

In last night’s post  John makes a very smart observation about his proclivity to self-injury:

This is more excitement in one day than I usually have in six months barring personal injuries, which, I’ve noticed, come less frequently now that I am sober.

This makes sense.  Alcohol does two things that leads to injuries.  It massively distorts our judgement and risk assessment ability and it wrecks our coordination.  Dumb ideas seem smart and simple physical feats become Olympian Gold medal floor routines with high degrees of difficulty after too many.

So how the hell am I going to tie this into health insurance and health economics?  Easy — this is a great launching point for a discussion on interaction factors for risk adjustment as well as the tail chasing nature of risk score maximization.

Most risk adjustment models will consider condition categories seperately.  They’ll try to estimate the average incremental cost of neurological conditions independently of the average incremental cost of cardiac conditions.  At the end of the model, a demographic weighing factor might be applied to bump up or down the added sums of each condition category that a person has.  So if Cardiology High is worth 1.32 points and Neurology Medium is worth 2.12 points, the individual  would have a pre-demographic adjusted risk score of 3.44 points.  And in most cases, that would be an accurate enough guesstimate of the projected incremental expense of everyone who has that combination of risk adjustment boxes checked off.

However there are some condition clusters which will interact with each other to significantly change the projected risk score.  An individual with Type 1 diabetes might ave a score of .72 if that is all that they have.  If they also have Neurology Medium, they would get .72 and 2.12 for a total score of 2.84.  However if they have both diabetes and Cardiology High an interaction factor might come into play.  Instead of the simple additive score of 2.04 points, the actuaries figure out that the cluster of people with this combination really cost on average 2.57 times as much as the typical covered life.

Interaction factors are a bit of a kludge to risk adjustment models.  They can  be very useful when there is good clinical support for why the costs are higher.  However they can also be used as a crutch to overfit a model to very messy data.

Going back to John, when he was drinking, he would have been categorized as having a substance abuse disorder which would have had a non-zero risk adjustment value. The problem from a risk adjustment perspective is he had no claims history to support that.

His drinking most likely contributed to quite a few of his skeletal injuries as well as excerbated any other problems he may have had.  If he was not insured through work, he would most likely have been covered by some risk adjusted program.  They would have had years of claims data of his klutz-dom and moments of fail.  Being a klutz is not a risk adjustable category.  It is not even diagnosible.  The insurer did not have a substance abuse diagnosis or claim until John decided to try to change and barged into an ER in Wheeling to get help.  Now he had a big cluster of claims with substance abuse diagnoses on them.  His insurer would have had John bucketed in the substance abuse category and received added revenue to cover John.

And then if John’s PCP is a decently diligent coder there would be a 50% chance that s/he added a substance use code after John’s annual PCP well visit in 2015.  That is a code that doctors sometimes add, sometimes don’t as diagnosis codes do not drive claims payments.  Additionally, in the opinion of the doctor, the former high risk adjustment value code might not be the appropriate code.  Instead either a low level code or a “history of” Z-code could be the medically appropriate code.  However, John’s insurer would have gotten stuck on the high risk adjustment value code and started to chase John’s PCP to use the same code in 2016 as was used in 2014 and 2015 to get the full risk adjustment value.

The risk adjustment is coming in late as the underlying root cause of some of John’s injuries was not producing compensating risk adjustment revenue when it was relevant to those injuries, and the insurer is chasing a code that is minimally relevant medically and expense wise as John’s been able to avoid Cole-like incidents.

UPDATE 1 From comments, the best illustration of sub-optimal decision making possible:

View post on imgur.com