I’m fascinated by churn and its implications right now. I want to propose the skeleton of a bare-bones churn prediction model. This will be incomplete and there will be some massive black boxes where an input and process will be assumed to have occurred without details.
Let us start with the simplest model with incredibly restrictive and unrealistic assumptions. We will build towards complexity and veracity.
Let us assume that each individual on Exchange in Time 0 will stay on Exchange in Time 1. Let us also assume that the initial selection in Time is “optimal”Let us also assume that there is only two companies so the decision is switch or stay. Let us also assume that there is a single product offered by each company. Let us assume the price offered for each product to each individual is constant and independent of the choices of other individuals in the market. Let’s assume cost sharing is a function of total medical spending. Let us also assume that the switch is frictionless. Finally, let us assume that each individual is 100% certain about what their future period health costs will be. The optimization problem we are trying to solve then is a cost minimization problem where the cost is premium plus cost sharing.
That gives us the following logic for each individual insured by Company A in Period 0:
IF Total Cost A1 is less than Total Cost B1 then stay else switch.
This would be a completely unsticky market where everyone is buying on price alone as there is no insurance functionality.
That is not what we see. This model is absurdly too simple. Let’s start peeling back some assumptions.
Let’s assume that people have a probability distribution of future costs. There is uncertainty. That uncertainty is a function of not knowing if you will be hit by a bus in the first day of Time 1, there is uncertainty about if and when you will get diagnosed with cancer. This is where the product changes from a discount club card to an insurance product as we begin to deal with future uncertainty.
So the decision process now becomes a bit more complex.
IF Premium A1+expected cost sharing A1 is less than Premium B1+expected cost sharing B1 then stay else switch
Now how do we figure out what future costs could look like and the individual probability distribution is appropriate.
My personal probability distribution is different than the probability distribution of a 23 year All American distance runner. It is certainly different than that of a 28 year old who has been having knock me up sex for the past three months. It is very different than the distribution for my parents.
We’ll make the assumption that people who were expensive in the past will probably be expensive in the future. But then we’ll need to make an assumption that there are differences in cost sharing. Some people have expensive chronic conditions while others just have cats that try to kill them on an ad-hoc basis. Now the equation starts to look like the following:
IF Premium A1 + Cost sharing (which is a function of previous Acute costs + another function of previous chronic condition costs times some multiplier for age)A1 is less than Premium b1 + Cost sharing (which is a function of previous Acute costs + another function of previous chronic condition costs times some multiplier for age)B1 then stay else switch
But switching is not frictionless. Medical records need to be transferred, authorizations and procedures need to be learned. The cost is a variable cost. An individual who has no claims in their entire membership span with Company A has an equal cost of learning Company A’s way of doing things and learning Company B’s way of doing things. An individual who sees a doctor every week with frequent prescriptions and at least annual in-patient admissions will have spent a lot of time and energy learning how to navigate Company A. Going forward the learning costs are far lower if that individual stays with Company A rather than switches to Company B. So far this is overwhelmingly a function of medical risk; individuals with higher medical risk will tend to have developped a stickiness/friction against sticking while minimal utilizers will not have any adhesion to Insurer A.
Networks also need to be compared as part of the frictional costs of transferring. Healthy people will have minimal relationships with their providers. They might have an annual primary care provider appointment, women might have an annual Ob/Gyn appointment and perhaps a few urgent care or physical rehab appointments every few years. There is a minimal relationship at the rendering provider level and the other common services (UC and PT) there is no interpersonal relationship established. However an individual with significant health concerns will have routines and relationships that are valuable to maintain. They’ll hopefully have a PCP that listens to them, they’ll have a cardiologist whose staff knows what a “personal” normal looks like, they’ll have the hematological clinic in-network and nearby. Those are very valuable items to someone managing complex conditions.
If the networks between A and B are identical, there is no network costs. However once some providers are in A but not in B, switching costs rise for people who currently see the in-A not B providers. These costs are very low for the very healthy and could be quite high for people with chronic conditions. Now we get an equation that looks like this:
IF Premium A1 + Cost sharing (which is a function of previous Acute costs + another function of previous chronic condition costs times some multiplier for age)A1 +Costs(Administrative switching as a function of age and health) +costs(network switching as a function of health, geography and probability of providers being in A not B) is less than Costs (B total) then stay in A else go to B.
Finally, still working with the assumption that the selection of Plan A in the first time period was optimal, there is an option value of the network that Plan A had assembled that is unique compared to Plan B. For instance a 29 year old woman who intended to get pregnant in Time Period 0, would see a great amount of value in having a leading Ob/Gyn hospital in network just in case her potential pregnancy went bad. Now if she had the baby and had her tubes tied at the same time in Time Period 0, the value of the Ob/Gyn hospital as a unique selling feature for Time Period 1 goes down dramatically. The same logic applies to a professional pitcher with a nasty curve ball. He wants the best Tommy John’s orthopedic surgeons in network during his playing career. The day after he retires, the option to see those specialists is far less valuable to him.
The same applies to cancer centers and any other high end acute specialty hospitals and specialists.
IF Premium A1 + Cost sharing (which is a function of previous Acute costs + another function of previous chronic condition costs times some multiplier for age)A1 +Costs(Administrative switching as a function of age and health) +costs(network switching as a function of health, geography and probability of providers being in A not B) is less than Costs (B total)+Net Option Value of Network of A then stay in A
From here aggregate cost profiles and estimate net losses. If there is perfect information about members covered by Insurer B, then a mirror image inflow model could be run as well.
I think this is a workable framework that is still absurdly simplistic as it assumes no leakage or inflows from other types of insurance. It assumes only two competitors, it assumes very good information at very low costs, it assumes clear plan designs so people can accurately project their future net costs. This is still absurdly simple.
But I think it is a start of something.
Chyron HR
NERRRRRRRRRRRRRD
Mike J
This is one reason why I hate insurance commercials. “Three out of four people who switched saved money!” And 99 out of 100 who weren’t going to save money if they switched, didn’t.
But yes, your framework seems like a reasonable start, if a bit simplistic. Happily, computers are powerful enough to game out scenarios with more than two competitors, varying percentages of well and sick, etc.
Richard Mayhew
@Mike J:
Agreed, this is a massive simplification… I was having an interesting conversation about what a low stickiness insurance market looks like, and making my thoughts explicit on churn was a needed step in thinking things through a bit more. Simple is good, and then it is time to layer complexity to better match reality.
And yes, once the logic is created, programming a simulation model to do this is an enjoyable week’s worth of work (and most of that time is getting the data sets to play nicely with each other).
@Chyron HR: It took you this long to figure that one out…. slow man, slow.
Cheap Jim, formerly Cheap Jim
OK, but where’s the spherical cow?
Gin & Tonic
I saw the title and only thought of Alex and “the old in-n-out.”
chopper
@Cheap Jim, formerly Cheap Jim:
you’ve had my mom’s hamburgers?
raven
@Gin & Tonic: I always get a shirt when I’m in LA.
eta That is the extent of my ability to comment on this post.
burnspbesq
OT;
immediate equalizer for Dortmund. Hang in there, Amir.
pseudonymous in nc
This is true, but I think in the exchange environment the stickiness of renewal is offset during the year by the strikes accumulated every time an appointment needs yet another PCP referral, or every time coverage for a procedure is routinely denied first time round until the provider sends in an extra lump of paperwork.
These are sometimes intangible when comparing total costs — if you assume that the individual’s time is free and peace of mind is worthless, which is mostly the case in the American Way of Healthcare — but I’m sure that plenty of people switch to objective worse plans because they take every slight and refusal and empathy-free inconvenience very personally.
Gin & Tonic
@raven: That comment was based on A Clockwork Orange.
Richard Mayhew
@pseudonymous in nc: Oh I completely agree with you. But I have no idea how to really model F-U-ism on consumer behavior. There is always an error term/fudge factor in a predictive model. I think a model like the one I am proposing would do very well with the corner cases (people with no claims and people with gigantic and recurring claims). I need to see how this model would do with people in the middle.
There are a lot of people on Exchange who will have 3 or fewer claims a year which includes their flu shot. Those individuals very lightly touch the medical system so they don’t get too frustrated with it.
raven
@Gin & Tonic: Yea, so? I free associate.
Mnemosyne
@pseudonymous in nc:
Saw my comments about my asshole medical group, did you? I don’t have to change insurance (it’s employer insurance anyway) but I will be finding a new PCP in a different medical group as soon as this current hurdle is dealt with.
And, yes, my current situation is why I laugh mordantly whenever anyone tries to tell me that all our problems will be solved if we just get rid of insurance companies. That will NOT solve the problem of for-profit providers steering patients to the facilities that they make a profit from, but “insurance companies suck!” sure makes an easy slogan.
Linnaeus
Feel the…churn?
pseudonymous in nc
@Richard Mayhew:
I suppose the beginnings of a model might be percentage of claims rejected or modified on first contact (and the amount involved) and then the number of customer service interactions and the amount of time spent on them. If you’re on the phone every month trying to work out why a pre-approval was denied or a claim only partly covered, then you’ll be looking for somewhere else once open enrollment starts, even if you’re likely to have to jump through the same hoops with a new insurer.
As Mnemosyne suggests, this can have more to do with providers’ revenue streams than insurance, e.g. a desire to get one extra MRI out of each referral to the imaging facility. But if a provider fucks you over, you can generally show up at a facility and wave a bill at a real person, whereas if you show up at your insurer’s corporate office, you’ll probably be arrested.
? Martin
Don’t underestimate the role of transactional costs. They turn out to be incredibly important in many networks.
How much effort is involved in switching (hours, hassle) and how does the user value that effort? If it’s reasonably easy, with obvious effort invested in making it easy and transparent for the consumer, the user is more likely to be willing to pay that cost. If it’s frustrating and burdensome, then they won’t.
When you see two Starbucks across the street from each other, that’s what you are seeing. A coffee has a value of maybe $3 and 5 minutes to acquire. What is the cost of a left hand turn at a difficult intersection (plus another to get back on track)? What if it was a right turn instead? What if the line is long and it’s 8 minutes instead of 5? If it were a $20 purchase, no problem, but $3?
The two stores solve two transactional costs – one the cost of the intersection (everyone can make a right turn instead of half making a left) and the other is the cost of the line (dividing the market). Everything around Amazon is centered on this – one-click paying, Prime shipping, the new Dash buttons, subscriptions for certain items (send me deodorant every 4 weeks), etc. We see this in my house – is it worth driving farther to the cheaper gas station, or will we burn the savings on the gas to get there and waste time on top of it?
Switching insurance has HUGE transactional costs. The actual cost of changing policies, the cost of learning the new network (which you’ve indicated) and so on. My mom does this as a volunteer for Medicare D coverage, and its really clear that the cost of simply understanding the policy differences so thoroughly overwhelms the user that they would rather stick with the more expensive policy.
If you minimize those transactional costs you’ll get more churn (the cost to switch goes down) but you get more aggregate efficiency in the system (people pay less on average). A huge number of startups are really focused on this, making their money through arbitrage. Services that will comparison shop insurance, banks, loans, hotels, investments, and so on according to a set of parameters that each person can provide, and they take a small cut of your savings.
Our society is going through a difficult transition from the tyranny of scarcity to the tyranny of abundance. Increasingly people aren’t lacking for options, its that they have too many of them and can’t optimize. That’s what Google does. It’s what Amazon does with their reviews. It’s what needs to happen with health insurance – though it’s a more complex problem, so it’ll take longer to solve.
Richard Mayhew
@? Martin: what you said much better than me
martian
@pseudonymous in nc: Exactly why I’m looking forward to switching after a mere week on my new plan. What was supposed to be a routine follow up with my opthamalogist uncovered a retinal tear. Doc wanted to fix immediately; insurance company started the paperwork hokey pokey. Also hitting a wall getting my daughter’s PT covered. Fucking HMO. And it was the only option on the exchange that had most of our specialists in network. Really hating on insurance companies today. At least, eavesdropping on the doctor’s clerk, I caught on to her habit of asking the insurance contact for a reference number for every call. Turns out to be a key thing in cutting down the confusion of multiple calls/multiple points of contact. Also, being able to provide the correct system code for the diagnosis myself has been important. Silver linings, I guess.
Mnemosyne
@martian:
Double check and make sure that the medical group you belong to is not adding its own chaos to the mix. I don’t want to get into the whole megillah again, but suffice to say that my insurance company was telling me one thing while the administrators for my HMO’s medical group were telling me something else.
martian
@Mnemosyne: Thank you. We did definitely have some crossed wires at the doctor’s office. People were not communicating. But the HMO was the biggest problem with the expanding list of hoops to be jumped. Then, after a pointed conversation with a rep about torn retinas leading to detached retinas and much higher costs and poorer outcomes, it was revealed from on high that my PCP should include a special code for an expedited referral to get things moving along. It’s how arcane all of this is that really gets to me. There’s no handbook. What if I didn’t have the wherewithal to push hard? What about other people who don’t? There shouldn’t be so many eddies, and rivulets, and side channels to explore when pursuing straightforward medical treatment. That’s the real dream of single payer, I think, at least for me – how much less complicated it seems.
Oh, and good luck with your biopsy. I’m sure it’s nerve wracking even when you aren’t really expecting bad news.
Mnemosyne
@martian:
Thanks! The odds of it actually being something serious are quite low, but our family hasn’t exactly had an excess of good luck with our health, so I’m extra nervous that we will manage to beat the odds once again and have it be something bad.
? Martin
@Richard Mayhew: And another way to look at it – which you are free to use – is that if transactional costs weren’t significant, particularly in aggregate, then voter ID laws and these anti-abortion laws requiring waiting periods, admonishments from your doctor, magic wand insertions and so on wouldn’t be as effective as they are. Small transactional costs add up in meaningful ways, but we routinely dismiss them as meaningful because they are individually small. You can slip them by people without them noticing. But some of the most successful businesses (such as Amazon, Apple with TouchID/ApplePay) have gotten there by eliminating small frictions that are almost subconscious to us.
cat
I am late to the party, but have you done PCA? As Martin said, consumer behavior can be rather irrational especially in situations where they have too much information, no expertise, and little time.
I admit, I haven’t paid much attention to your churn posts and so I don’t know what your end goal is. If it’s just to have an estimate of how much churn Mayhew Ins will have during the next open enrolment a Human derived model should be fine. If you want to predict if policy holder X is going to churn, I would think you would probably need to have a machine/human, ML, derived model using the vast swaths of data Mayhew Ins has about X.
The free 2 play games have a huge motivation to keep paying customers since the acquisition costs are so high and they are finding ML to be very helpful in saving customers they want to keep.