A couple of weeks ago, the Centers for Medicare & Medicaid Services (CMS) released the 2014 physician Medicare utilization data file. This is the third file released in as many years and contains a large portion of the utilization detail for 923,813 individual National Provider Identifier (NPI) codes, the majority of which are for physicians, NPs, and PAs.
In general, the file contains information on the procedures providers report to Medicare, along with the number of unique beneficiaries, the number and identification of services provided to those beneficiaries (for both drug and non-drug related services), and the payments made to those providers. In total, this file accounts for $90,012,301,857.01 in payments and 2,617,809,025 procedures. And this does not represent all payments and all services for all Medicare beneficiaries, as this data only represents traditional Medicare services and not Medicare Advantage (because for some reason they are not required to provide data to the public sector). So even though our tax dollars are used to support Medicare Advantage plans, they are not subject to the same level of accountability as traditional Medicare.
So much for transparency.
Also, the database excludes entries for which a procedure for a given provider was reported on fewer than 10 beneficiaries in order to protect privacy.
Of particular interest is that the 2014 file contains Hierarchical Condition Category (HCC) codes. HCC codes were introduced by CMS in 2004 for the purpose of risk-adjusting beneficiaries with the scores being correlated to the diagnoses codes assigned to each beneficiary. In general, HCC codes are used by Medicare to negotiate payment rates for Medicare Advantage (MA) plans, which is the same as risk-adjusting capitation payments. CMS’s risk adjustment model measures the disease burden for 70 HCC categories and then creates a risk value, which for the 2014 database ranges from 0.2713 to 10.9674. Interestingly, that highest value is associate to a nurse practitioner.
This risk-adjustment factor is primarily used within the MA model to determine, or even predict, the Medicare cost for caring for each group of beneficiaries. The system is actually prospective, whereby the scores for the prior year are used to predict or estimate the expenditures for the next year. Because of this, a cottage industry has been developed that focuses on helping physicians improve their diagnosis coding. In fact, I know several “HCC consultants” who are hired by MA to assist their physicians with improving (hopefully, that’s all) their ICD coding, which in turn increases the average HCC risk scores and subsequently increases the payment from Medicare to the MA organization. As such, there is a general consensus that since MA plans are paid based on diagnoses and Part B providers are paid based on procedures, MA patients are better coded than traditional Medicare patients.
It is important to note that risk scores are based on expenditures or paid claims in the traditional Medicare program even though they are used to factor payments for MA patients. There is a lot more involved in creating these risk adjustment factors – far more than can be discussed here – but suffice it to say, the primary purpose is to be able to estimate or predict future Medicare payments. And that’s a shame, because it just doesn’t work, at least based on my simple analysis.
In the 2014 data dump, as mentioned above, for every provider Medicare released the number of unique Medicare beneficiaries, the number of procedures reported, and the amount paid for each. As such, it is pretty easy to calculate a couple of important ratios: procedures per unique beneficiary and payments per unique beneficiary. The latter is watched closely by media outlets, whistleblower wannabes, and other watchdog organizations. The idea is that the higher this value, the more money a given provider is “taking” from the Medicare trust fund, and the more likely the provider is doing something wrong (although this is a bad idea based on fatally flawed logic).
One might expect that as the paid amount per unique beneficiary increases, so would the HCC score, and as the HCC score changes, the amount paid per unique beneficiary changes in the same direction, and with a similar magnitude. In essence, a reasonable person would expect that there would be a pretty decent correlation between paid ratios and HCC scores. But that is simply not the case. I ran a couple of very simple statistical tests to determine direction and magnitude of the relationship between two variables: in this case, HCC scores and paid amount per unique beneficiary. I ran the test two ways: first on the full database of some 877,000 line items (NPI codes) and second with a sample of 1,000 lines. The reason for the latter is that sometimes, if the database is very large, the test can predict significance when it doesn’t really exist.
The results of the correlation analysis for both the universe and the sample were quite close: 0.050 and 0.035, respectively. It doesn’t really matter, however, since the values are both so low as to indicate that there isn’t any relationship between the two variables. I also conducted several regression analyses using different methods to ensure that I didn’t ignore some important parameter. The reason I did this was to get what is called an R-squared value, also called a coefficient of determination. In general, the R-squared value is used to measure how close the data is to the fitted regression line. Another way to interpret this is to estimate how well the response variable (paid amount) can be predicted by the predictor variable.
In this case, the best R-squared value was 1.8 percent, meaning that only 1.8 percent of the paid amount can be predicted or explained by the HCC score. In essence, this means that nearly 98 percent of the paid amount can be explained by something other than that score (something that we know nothing about).
Now, I am going to admit that I am not the leading expert in HCC- or Medicare-adjusted risk scores, but the literature is pretty clear, and it’s well-supported by the consultants with whom I spoke; the risk scores are primarily designed to predict how much Medicare is going to spend next year on those beneficiaries enrolled in MA plans.
And if that is all they were designed to do, they do a pretty bad job of it. So, one question remains: if the HCC risk score doesn’t adequately predict payments, then what does?
And that’s the world according to Frank.