>>So this morning I’m going to do
a presentation on risk adjustment, risk adjustment using the CDPS model. So risk adjustment, why am I
talking about risk adjustment? So couple reasons. One is that risk adjustment is pretty essential
for effective managed care in Medicaid. So risk adjustment is very prevalent in states
that use a lot of Medicaid managed care. It’s being used in the Medicare program. It’s being used in state health exchanges. So it is something that’s good to know about. It’s also useful to have
algorithms, diagnostic algorithms, when you’re working with the Medicaid data. So if you’re doing a very
focused study on newborns and healthy mothers you might
not need risk adjustment, you kind of got a fairly homogeneous population. But if you’re looking at larger groups of people
who are disabled more generally or other types of groups, you might want to control for illness
severity and this gives you a way of doing that. And so it’s a useful way of kind of taking all
those ICD-9 codes and doing something manageable with it, and so that’s why
I wanted to talk about it. So this is what I’ll talk
about in the next half hour. We’re going to talk about what are the goals
of risk assessment and why do we do it. I’ll give you a brief history of it,
and then I’ll show you how it operates. And I’ll talk about other ways in
which risk adjustment might be helpful. So what is risk adjustment? So we think of it as two parts; one
being a health-based risk assessment. So that’s measuring illness burden at the
individual level using the data that you have. And so typically that will include
demographics and diagnoses, some models. We’ll also use pharmaceutical claims
to measure illness severity based on the pharmacotherapy you are receiving. And then some models will also include data
on say cognitive and functional limitations which will be important if you’re looking
at home and community based waiver programs, long-term care, things like that. And then risk adjustment is using
that data to adjust something. So these models were developed to adjust
capitated payments to health plans, enrolling Medicaid beneficiaries. It could be used to risk adjust other outcomes
like hospitalization rates and things like that. So it’s kind of an adjuster
for illness severity. And why is it necessary? So you’ll hear about the different
rules; the 2080, the 1060. These are some statistics from disabled Medicaid
beneficiaries; the top one percent accounting for thirty percent of the expenditure. So we know the distribution of healthcare
utilization expenditures is highly skewed with a long fat tail, they call it [inaudible]. So you have small numbers of people
accounting for large proportions of costs. And the problem with that is that if
you’re in a capitated payment system where you’re not risk adjusting, you’re
receiving the same payment per person, it gives you a huge incentive
to avoid sick people, right? You can do all the care management you want,
but really you’re going to getting hammered if you enroll people who are sick on average. And so what risk adjustment tries to do is to
minimize those incentives for cream skimming and to make equitable comparisons
across health plans; kind of level the playing field in a sense. And also then provide adequate
financing for programs that do enroll people who
are sicker than average. So if you think about where you are in your
regional area, there might be a hospital, an urban area that maybe
serves lower income residents, maybe a place that attracts sicker
individuals, you look at specialty hospitals, a children’s hospital, something like that. And if you have health plans associated with
those academic medical centers or other areas, they’re not going to do as well financially if they’re receiving the same
amount of payment per person. So it’s both to kind of minimize plays
active marketing to the healthy and kind of adequately finance those that
do attract a sicker population, so kind of both sides of it there. And this is kind of as I was talking about
some reasons why a particular health plan or network might have a sicker population. There is again -affiliated with an
academic medical center, cancer center, maybe places that are renowned for their
diabetes care might attract sicker groups or those types of groups and that
might result in higher spending. And it might have certain geographic regions,
it might have higher case mix enrollees, so urban areas might look different than
rural areas and suburban areas and so on. There are risk variations out there and
here are some factors that drive it. And the advantage of implementing an effective
risk adjustment system is you kind of use that health assessment to measure the
illness burden at the individual level and then you assemble individuals in groups
based on health plans in which they enroll. And then you have metrics to say that this
plan has a higher risk burden than this plan. And then you can use those metrics
to adjust capitated payments. So the health plans that enroll sicker
beneficiaries would get a higher capitated payment. So it would still be capitated, it would
still be one payment per person providing the incentives for the risk plan to
manage that care most effectively, but the payment would be higher if you
had sicker individuals enrolling in it. And we calculate those metrics and make those
adjustments using data and claims typically. And the most predominant source of
data there would be the diagnoses. And you saw the diagnostic field and perhaps
used it yesterday in the inpatient claims which have diagnoses on them and a lot of the
OT claims will have diagnoses on them as well. In the MAX [phonetic] the OT have two diagnoses
and the inpatient have I think up to nine, although all nine aren’t always filled. So this is the plan; we’re
going to promote efficiency and reduce the incentive for risk selection. Any questions on why we do risk adjustment
and why it’s necessary, or the benefits? So what are the key ingredients for
successful health-based payment? No cheating if you’re looking on the slides. Three things. [ Loud unidentified sound ] So [inaudible] is equitable data. What’s the second one? The second one… [ Loud unidentified sound ] Is equitable data. [ Laughter ] And you can guess the third one. [ Loud unidentified sound ] So data becomes a huge issue, right? So you’re basing all these models on the
diagnostic profile of the beneficiaries. You really need to have a fair or
equitable or equal data collection system. So if one plan isn’t collecting
diagnoses very well, they’re going to look very
healthy when that’s not true. If another plan invests a lot of money in making
sure everyone’s got all their diagnoses covered which plans do, then it’s not equitable,
especially when this was rolling out and they continued to look at changes over
time in health plans and tried to be cognizant. I know Kaiser had a big effort at
one point investing in coding, right? Training the providers to
code or the medical assistants and things like that and it still goes on. Hospitals invest a lot of money in
coding, it drives the GRG payments. So they are kind of doing that already and we
just need to be aware of that and make sure that some other health plans
aren’t disadvantaged because of it, but what ends up happening is everyone
really has an incentive to code. But what is happening is managed care is getting
better at coding now than [inaudible] service. So they’ve kind of surpassed
[inaudible] service in their coding. And if you look at Medicare Advantage, the payments for Medicare Advantage
plans have been going up and up and up. And because it’s linked to [inaudible] service,
they’re getting more and more payments, but it’s because they’re coding more. So eventually they’re going to need to fix that,
but so far they don’t want to hear about it. But Rick Kronick is trying to tell them that and
eventually maybe they will make an adjustment. Here’s a history of the risk adjustment. So these models were first
developed in the 1990s. They were developed for the Medicare
program; the ACGs and the DCGs. Medicare promoted these early, but
adopted the models fairly later. Medicaid started in 1997, and then Medicare
Part C started in 2004, and then Part D, the pharmacy part, started in 2006. So Medicaid got a big jump-start. There’s more managed care in Medicaid
overall than Medicare; Medicaid we saw was about fifty percent; I think
Medicare is still about under twenty. Here are different states and different
populations that they risk adjust. Some adjust basically most
people SSI, disabled, and TANF. Some have done just the disabled, and some have
done just the TANF, but most do them combined. And here are the models. And most of the places have used CDPS because
it is a Medicaid specific risk adjuster and for a long time we gave it out free, but there are a couple other
models that are used too. And the Medicaid Rx is our pharmacy-based
model, our NDC code-based model. And so I’ll talk about those a little bit. And risk adjustment in healthcare reform, risk adjustment is being used
in the state health exchanges. There is a risk adjustment model specific for
the health care exchange that you can download from CMS somewhere, I’m not sure exactly where,
but if you can’t find it you can e-mail me and I have it in my e-mail somewhere. And then Medicaid programs are
overwhelmingly used using managed care for the Medicaid expansion, and they’ll need
to be doing risk adjustment in there as well. There’s some other areas they’re look at this; some of the dual eligible pilot programs
are looking at risk adjustment models. There’s interest in trying
to risk adjust between home and community based waiver services
and long-term care services. There’s a huge gap in payments in those, so institutional long-term
care is a lot more expensive, and it’s driven by a day rate, you know? A nursing home day rate. It’s driven by functional limitations,
so it’s kind of a different animal. And then you’re looking at measuring things like
functional limitations and cognitive status. But managed long-term care is
becoming more and more of interest. And some places are getting additional data
through web portals, clinician assessments, things like that, a way to electronically
collect these self reported data on large numbers of people. Wisconsin I think is doing that. So there’ll be some innovations
coming hopefully, and some interested in risk adjustment and socioeconomic status. So looking at healthcare utilization;
how is that related to income? How is it related to race ethnicity? How is it related to English
language proficiency? So some kinds of interesting public health
issues associated with risk adjustment. So what so CDPS? So CDPS is one model, its one option
you can use, its Medicaid specific. Alex Hauser [phonetic] has a model, I don’t
know if it’s used for risk adjustment, but it’s used for health-based risk
assessment, and that’s available from the Agency for Healthcare Research and Quality, AHRQ. And then there are the DCG models, the ACG
models and so on, they’re all fairly similar; they all try to do the same thing. So CDPS looks at all the ICD-9
diagnoses, now we’re also using ICD-10, and there’s about seventeen thousand
ICD-9s and it looks at them all. It uses about six thousand and it maps
some of the fifty-eight disease categories. So these are within nineteen major
categories like cardiovascular, psychiatric, pulmonary; that would be a major category. And within those it has a hierarchy and it
assigns diagnoses to one of these places, and how we did that was with a lot of
regression analyses and a lot of clinical input. And a colleague and I would work together and
I would run the regressions and he would run it by his clinicians and we kind of
[inaudible] grouping diagnoses, mostly at the three digit level,
then into groups that had kind of similar cost coefficients and similar
clinical meaningfulness and clinical severity. And eventually, it took us
two years, we mapped them all. And so that’s CDPS. So it’s similar to kind of the HCC models,
which are built off the DxCGs that they use in Medicare for Part C and Part D. I think
Medicaid pays a little bit more attention to some of the more severe, but
less common diagnoses that you see in the Medicaid program among
the disabled in particular. And the HCCs put more effort into mapping out
different types of cardiovascular diseases, for example, because you have more
elderly and more common chronic diseases and cardiovascular would be a big part of that. So we have kind of less detail
there, but more consideration of other Medicaid related diagnoses. And then we have payment weights
that on Medicaid on the MAX data. And we have separate models for
disabled TANF adults and TANF children. And from yesterday; why would
we develop different models for the different aid categories?>>The cost?>>The cost. The cost of cost and the
utilization profile is very different. Remember the disabled had this giant
big column of all these different things and the adults and the children were small. And even the adults and the
children differed by a factor of two. So we have different models
for each of these groups. Here are the major CDPS categories
that we talked about. So these correspond pretty well to
the categories in the ICD-9 codebook. And so these are the major categories:
Cardiovascular, Psychiatric, and so on. And pregnancy is considered a
chronic disease in our model, but it’s important in the Medicaid program. And here’s an example of the
hierarchy within cardiovascular. So the diagnoses here would be very high things
like a heart transplant or a valve transplant or a valve malfunction, so things
that involve some major surgeries. Below that would be heart failures,
so kind of a severe chronic condition that requires ongoing management. Below that would be kind of a heart
attack which is also important, but a little bit more acute [phonetic]. And then below that would be hypertension. So those are just examples of
diagnoses within those categories. We kind of choose ones that were fairly
prevalent that kind of help anchor the category, but within these it’s kind of hierarchical. So if you have both hypertension and AMI; we’ll
just code you as having cardiovascular low and not low and extra low, or if you
have heart failure and hypertension, same thing, we’ll code you as heart failure. So the idea there is to reduce incentives
for upcoding and the effect of upcoding. So I was talking about how some plans might be
a little bit more aggressive in their coding than others which will drive up the risk
scores, this tries to mitigate some of that. So if someone does have heart disease of some
sort, we’ll look for the most expensive feature of that and not everything
that’s been checked off. So that tried to reduce the
incentives for upcoding.>>Over what time frame [inaudible]?>>We look over a year.>>A year prior?>>A year prior, yes. Well there are two kinds of
models; there are concurrent models and there are prospective models. So a concurrent model you look at
the diagnoses within a base year and you predict expenditures in that year. In a prospective model you
use the same diagnoses, but you predict expenditures for the next year. And there’s a little bit of a philosophical
debate about which one you want to do. And what it comes down to in concurrent models,
because the cost, the dependent variable, is in the same year, the more that you
have higher weights on the coefficient. So you pay more for heart
transplant, heart failure, and AMI, and you pay less for the intercept, the base
rate, so more dollars go into the diagnosis. Whereas, in the prospective model, when
you’re looking out into the future, fewer dollars go into the diagnosis. So a lot of the Medicaid plans have
been doing prospective risk adjustment. The heath care exchanges are
doing concurrent risk adjustment. As these things get implemented, there’s
often a delay in implementing it, so even if you do prospective, it’s hard to say,
“Well, I’ve got all the diagnoses for this year and I’m predicting expenditures for next year.” In reality, you have the diagnoses
from a year or so ago or two ago by the time you collect it all and analyze. And so some people say you
should just do concurrent because really you’re just taking the
diagnostic profile from two years ago that you’re hoping is pretty similar to
what’s going to happen this coming year and then you just want to predict
what expenditures are going to be in this coming year so you use concurrent. So you could use either one. Weights are additive across major categories, so
you could have say cardiovascular and diabetes and they would both count;
psychiatric and diabetes and so on. But you couldn’t have multiple psychiatric
or multiple cardiovascular categories. So in addition to these diagnostic based models, there are also pharmaceutical
based models that use NDC codes. Ours is called the Medicaid Rx model. I don’t know there’s like
a DCRx something like that. But the idea here is you take the NDC codes
and link them to the disease categories. And so in here you’re trying to group NDCs
into similar pharmacotherapy approaches. And so we combine -pharmaceutical
is used for cardiovascular disease, used for depression, used
for psychosis, and so on. And so it’s a little bit
different than diagnosis in that diagnoses you’re
identifying the clinical condition and with the NDC codes you’re
identifying the pharmacotherapy. But drugs can be used for different things,
so a little bit of a trade-off here. If you’re using an NDC, a
pharmaceutical identifier, it might not be exactly for
what you are identifying. So I’m trying to think of a
good example, but I don’t know. Bupropion is used for depression and smoking
cessation and so we try to limit the ones that are in that category to the doses that are
specific for depression, but it’s not always that clear-cut, so you might have some people
being prescribed that for smoking cessation that then [phonetic] get
identified as depression. Similarly, there are some seizure disorder
medications that can be used either for seizure disorder or bipolar disorder. So again, it’s a therapeutic class here,
and we find they’re pretty similar I think. We also have a combined model where we take
fifteen of the forty-five drug categories and combine then in a hierarchy
for CDPS to do a combined model. I think if you look at a TANF
population, the diagnoses and the pharmacotherapy do pretty
similar because there you’re looking at common chronic diseases for the most part
and a lot of that is treated by pharmacotherapy. And among the disabled population,
diagnoses do a little bit better because you have more specific illnesses
striving utilization in that group. So I think pharmacies are better. Having the pharmacy in there is
nice because you can identify things that aren’t always well diagnosed. So hypertension, for example, once
you’re started on say an ACE inhibitor, you might not get a hypertension diagnosis
every year even though you’re on the medication. The medication kind of keeps you in there. Mental illness is another area where
often the diagnoses are aren’t there; either they’re not added to the chart
because of stigma or for payment reasons; that they don’t pay for mental
health visits so they’re not going to code mental health diagnosis,
or if they’re getting prescriptions from a mental health specialist that’s outside
the main system like a carve-out or something like that, the data systems might
not be feeding each other right. So that’s an area in psychiatric where you find
that pharmacotherapy identifies more individuals than diagnoses so that’s where a kind
of combined model is kind of important. As you think about the different things
you’re looking at -like if you are looking at behavioral health you might
want to consider that as well. There are different weights. These are regression models. The algorithm creates indicator variables
that can be used as independent variables in a regression model where
the dependent variable is cost. We talked a little bit about
concurrent versus specific. You could also use the weights which are
the regression coefficients that come with these software packages to assign
a risk score or a severity score. Or if you have an [inaudible] you
can estimate your own coefficients; you’ll need pretty fairly
large sample sizes there. When we actually do risk adjustment that we
think about what’s in the benefit package. So is this an acute care HMO? That’s kind of what this is built for. But sometimes even those will carve-out pharmacy
coverage, and that will be given delivered by a pharmacy benefit management company,
so you’ll want to carve-out pharmacy from your payment rate, or they’ll
pay for [inaudible] separately, or they’ll pay for certain
pharmaceuticals separately. Well they have hepatitis C medications,
expensive ones; they’ve been paying for that separately in some cases and so on. So you might have a customized weight set. Then we talk about here the difference
between prospective and concurrent. I do have a slide on the weights. So this shows you kind of how
different categories might result in different risk scores. These are all normalized to
1.0 where that’s the base rate. So here maybe for a disabled beneficiary
1.0 could be say a thousand dollars. So you’re looking at paying two thousand
dollars per month extra for someone with cardiovascular very high or about
a thousand dollars extra for someone with psychiatric high, which
is mostly Schizophrenia. So that’s Schizophrenia and
then bipolar disorder, effective psychosis, and then depression. And so these are payment up and
beyond the base rate or the intercept and this is how you would calculate scores. So there’s an intercept like
any regression model when you go predictions there’s an intercept. We’ve got age and gender factors. So here it shows you for
example a fifty-year-old female with type 2 diabetes and hypertension. The intercept is 0.225. The demographic factor is 0.121. Then the type 2 diabetes would be 0.322. And 0.130 for hypertension. So you add them up and that person would have
a risk of about 0.8, which is below average. So this is disabled population. If all you have are diabetes and hypertension,
you’re not that sick relative to overall average in the diabetes population -I
mean the disabled population. So remember expenditures are skewed so it’s a
smaller number of people that are accounting for a large amount of expenditures. So the mean will also be skewed to the right. Then if you look at the same female with
bipolar disorder, you add in that 0.626 factor, and then they’re at a 1.4, so they’re
forty percent higher than the mean. And so if the base rate was
a thousand dollars per month, then for this individual the payment would
be fourteen hundred dollars per month. And so this is how we do it. You take the case mix score, calculate
the individual level, wrap it up, predict it at a group level, then
do a case mix score for each plan, and then you multiply that
by the base rate and so on. And then there is an issue about are
you calculating -is this budget neutral? And at the federal government level it’s
not budget neutral so they’re paying more and more each year for these
Medicare Advantage plans. Whereas, if you look within a state,
often what they do is they say, “Well this is the rate this
year, this is the base rate.” And then they divide that up among
health plans based on risk score. So they’re not paying more this year. If all the plans increase their coding
and everyone’s risk score goes up, they’re not paying any more,
it’s still budget neutral. They renormalize everybody to 1.0 before
they start distributing the payments. So if you don’t make it risk neutral then
you’re kind of facing cross-pressures over time. Most states do increase their
base rates each year, but then they can do it more purposely rather than letting the data decide
how much they’re going to spend. So the actuaries get involved in this
they make all kinds of adjustments, they might do partial capitation, they might
play around with risk corridors or reinsurance, there are different carve-out approaches. So if you get into risk adjustment, you start
at some point to get into the weeds and dealing with all these kinds of specific issues. And we usually give these to
the actuaries to figure out. It’s not very fun. So we kind of consult on the development
of the model and they do all the hard work of getting the data and calculating the
scores and communicating to the state and the plans and everything that’s going on. So here are some kind of things that
you’re hearing about in healthcare or heard about for a while; the chronic care model, complex case management,
accountable care organizations. We’re looking at primary care medical homes and now behavioral health
homes, integration and so on. And all of these things have some common
elements; we’re talking about team based care, we’re talking about care management,
and we’re talking about increased IT. And risk adjustment kind of fits
well I think into these concepts because if you have robust IT, that can
support the data collection you need to do risk adjustment well. And then if you do risk adjustment well,
it gives plans a lot more flexibility to how they invest those dollars. So instead of worrying about the primary
care provider billing for everything, they can maybe focus on team based care, maybe
you’ll use some lower cost providers in the mix to do some of the care management. They can do phone calls, e-mails, things
like that with a capitated payment. And they can bring in different types
of people and different types of care. They can have some nurse focus on diabetes care
and social workers focused on mental health and care transitions and things like that. So just to summarize, risk
adjustment kind of is necessary in this environment of managed
care and capitation. Some thought needs to be put in to have
it done well [inaudible] little payments. It does seem to move money around, I
don’t have any slides on that today, but we do see money moving and moving in
places where we expect people to be sicker. Data kind of remains a key challenge and
will, but it provides the opportunity or reduces the incentives for risk selection
and then also provides the opportunity for plans to develop kind of higher quality
models without kind of worrying about its financial impact quite as much.