CMS has finally released the long-awaited specifications for target price computation for the BPCI Advanced program. This document provides general guidance into the process used to determine these targets, but many details will not be clear until the actual targets are released.
Our goal in this analysis will be to discover the underlying factors that will affect financial success in this program. This is important in designing longer-term strategies and committing resources to these strategies, with some assurance that the strategies will create consistent and predictable positive financial results. In the BPCI program targets were based on historical costs, and participants who reduced their costs could be confident of financial success. In CJR the initial targets were largely (66%) based on historical costs but transitioned to regional rates in the later years, so participants with initial success could turn to losses later in the participation period depending on whether their costs were above or below the regional average (this transition was a primary reason that many participants in the voluntary MSAs dropped out of the program). And in the Oncology Care Model targets are based entirely on the results of a regression model calculated on various patient characteristics, so there’s no assurance that decreasing costs will lead to financial success, or that success in one period will be repeated. Since the BPCI Advanced program is voluntary, applicants will consider the potential for financial success in their participation decisions, and understanding the bases for target development will facilitate these decisions.
We’ll be discussing the hospital targets; the physician targets incorporate multiple hospital targets and have similar derivations. Here's what we know at this point.
Objectives of the BPCI Advanced Targets
First, CMS has moved from the relatively straightforward historically-based targets in BPCI to a far more complex regression-based approach. This approach was adopted for several reasons stated in the CMS specifications, the first being to encourage both high and low cost providers to participate. In BPCI high-cost providers had the greatest initial opportunity for success, whereas in CJR lower-cost providers would end up being the most successful. CMS hopes to avoid the conflicting nature of these processes through the new target setting approach. (In this document “cost” will refer to the cost to the Medicare program; i.e., payments made to providers, not to costs incurred by individual providers.)
The second stated objective is to reward participants’ cost reductions that occur over time. Reducing cost is the only way that these initiatives generate "savings" for the Medicare program, since "savings" don't result from simply being efficient. For the BPCI Advanced program to create savings, there must be incentives for already-efficient hospitals to become even more efficient.
The third objective is to compensate for differences in patient mix. In the BPCI and CJR programs, the only accommodation for differences in patient mix is the classification of the DRG initiating the episode, plus the existence of fracture diagnoses in major joint replacement episodes. Since DRGs are stratified by the presence of complications in the diagnoses of the inpatient medical record, some complicating factors were already included in the BPCI target price calculation. However, CMS believes that this new process will be more effective at adjusting for patient complications that result in different episode costs.
The fourth objective is to compensate for differences in the cost trends among different types of providers. These changes occur for several reasons; first because Medicare payment rates for services change over time, and also because the utilization of services may also change due to different care management techniques or changes in medical technology. Both of these changes were accommodated in the BPCI and CJR programs through the use of "trend factors" that adjusted target prices by changes in the average episode costs of each episode type in hospitals throughout the country. However, this new adjustment is apparently designed to accommodate differences in episode costs that occur within certain groups of hospitals; for example at academic medical centers, and to apply those differences to that specific group of participants.
The last objective is to achieve all of the above objectives while still maintaining quality.
Begin with the Baseline
The BPCI Advanced target-setting process begins with claims data containing the actual historical cost to Medicare of each episode type at each hospital, obtained from the “baseline” period of 2013-2016. In this sense, it mirrors the process initially used in the BPCI and CJR programs. However, these amounts are further adjusted through myriad of regression-based factors that attempt to accommodate the issues described above. The first "adjustment" is the use of standardized costs rather than actual provider payments for these calculations. Standardized costs remove the geographic wage adjustments that are applied to all Medicare payments, allowing comparison of episode costs throughout the country on a consistent basis. Standardized costs were used in the CJR program but not in BPCI. These historical costs are then updated to current year payment rates.
Patient severity adjustment
At this point the regression models kick in. The initial adjustment is for patient severity, which utilizes hierarchal condition categories (HCCs) used to compute payment rates for Medicare Advantage and ACOs along with other factors including patient demographics, the DRG or APC that initiated the episode, previous institutional status, as well as other episode-specific adjusters including fracture status for major joint replacement episodes. These factors were all dumped into a regression model that predicts their effect on episode costs, and the resulting regression coefficients will be used in target computation.
Interestingly, the use of the historical baseline costs in BPCI already includes some aspects of case mix. Hospitals with many complex patients whose costs of treatment were correspondingly high would have those higher costs included in the baseline, which would result in a higher target. However, higher baseline costs could also result from inefficient care, not from patient complexity, so CMS elected to use the patient severity regression process to attempt to avoid building inefficiency into the targets.
Hospital characteristics adjustment
The next set of adjustments also results from regression models and incorporates factors related to the hospital at which the episode was initiated, and its peer group. These factors include the hospital's status as an academic medical center, an urban/rural facility, or a safety net hospital. The calculation also adjusts for hospital bed size and the census division in which the hospital is located.
Episode target calculations
The above calculations are used to compute the initial baseline target amounts for each of the 39 individual episode families. From the results of the previous step, CMS filters the data down to include only participants with sufficient episode volume to receive targets. For each remaining participant, CMS will compute a “predicted” episode cost for each individual episode at each provider using the above factors for each episode. This represents the amount that the regression model predicts that each individual episode should cost the Medicare program. The first term in the calculation is based on the patient case mix regression, which is next multiplied by a regression-derived factor that’s derived from peer group data.
Finally, the actual episode cost is divided by the predicted cost, resulting in a ratio that indicates if the hospital’s actual cost in each episode family exceeds or is lower than the amount predicted by the model.
Note the numerator and denominator in this ratio – a hospital whose actual costs exceed the predicted costs will have a high efficiency ratio. This seems counterintuitive, but it's how CMS chose to compute it.
Next CMS breaks down the target computation into three pieces:
- Establishing an episode cost for each hospital and episode family, blending the national average cost with the efficiency ratio of each hospital (SBS);
- Adjusting that cost by the patient case mix of each hospital (PCMA); and
- Further adjusting for hospital characteristics and trending that cost to the performance time period (PAT).
Standardized baseline spending (SBS)
At this point, the calculation becomes even more complicated (if that's possible). Here CMS will compute an amount representing each hospital’s adjusted baseline cost for each episode family. It does this by computing the average episode cost for each episode family across all hospitals (confusingly referred to in the CMS documentation as the “Dollar Amount”), and then multiplying that amount by the hospital-specific efficiency ratio computed above.
This is an interim calculation step and may appear to be misleadingly meaningful since it appears that the average national episode cost is a critical factor in the target development. However, this amount is DIVIDED by that factor in the next step, which eliminates it from the calculation; therefore the average national episode cost is not a final component of the target.
The SBS is calculated using the patient case mix during the baseline period, so it won’t change over time (until a new baseline is introduced).
Patient Case Mix Adjustment (PCMA)
The next factor to be computed is the PCMA. This factor adjusts for the patient case mix during the relevant baseline or performance time period. In computing the preliminary targets the patient case mix during the baseline will be used, but the case mix during the actual performance period will be used for the final targets. This adjustment is computed by dividing the hospital’s case mix adjusted episode cost by the national average predicted episode cost (again the “Dollar Amount”). This predicted cost does NOT incorporate the peer group adjustments that were used in the calculation of the efficiency ratio – it only uses the patient characteristics.
This factor adjusts for the hospital’s patient case mix. Presumably hospitals having a more complex case mix than the national average will have a factor greater than 1, whereas hospitals with less complex patients will have a factor less than 1. Since virtually all hospitals believe “their patients are sicker”, viewing this factor may a sobering moment for many hospitals who were hoping for an increase in episode targets to accommodate their complex patient mix.
As noted earlier, the predicted hospital cost will be based on the data available when the target is computed, so this factor will change across performance periods.
Peer Adjusted Trend (PAT) factor
This factor is computed separately for each calendar quarter and hospital, and is the output of a regression of peer group characteristics, time trends, and their interactions. It accommodates differences in hospital characteristics (as opposed to patient characteristics, which are factored in the PCMA), and also trends the Medicare payment amounts to the performance period. Because so many factors are conflated into this metric it will be difficult for an individual hospital to evaluate its effect on payment relative to other hospitals. Unless a hospital changes its classification within the regression groups the hospital-specific component of this factor should not change over time. It will change each quarter to compensate for changes in episode costs and Medicare payment rates.
Hospital Benchmark Price
These three factors are multiplied together to get the hospital benchmark price (HBP). The same process will be used to compute the targets for each performance period, using patient-specific data for that period.
and the final target equation reduces to:
with the average efficiency ratio equal to the average of each of the episode values of:
These are the only factors that drive the target. Note the absence of the national average episode cost.
Note that hospitals that are already efficient relative to their patient mix and peers will have lower targets than those that are inefficient. In other words, currently-inefficient hospitals aren’t disadvantaged by the target-setting process and still have an opportunity for success. This may induce more cost-inefficient hospitals to participate in BPCI Advanced since they may have greater opportunities for cost reductions.
Also note that the benchmark price isn’t final – it will be adjusted for actual patient characteristics occurring during the performance period, and the PCMA factor will be recomputed using those characteristics for final reconciliation. Ideally this adjustment will be neutral, since targets should be adjusted commensurately to follow the costs associated with patient complexity. The extent to which this actually occurs is a function of the accuracy of the regression models and remains to be seen.
The above process is summarized in the flowchart below.
A simpler diagram of the HBP calculation was prepared by Jessica Walradt for Northwestern Medicine and is below:
Evaluating the targets
This analysis of the target computation process shows that the ultimate driver of financial success in BPCI Advanced will be the ability of hospitals to reduce episode costs below baseline levels. Therefore, the first step in episode evaluation is not a review of the targets, but rather evaluation of the clinical composition of episodes and the hospital’s opportunity to reduce those cost components. Evaluation of clinical opportunities is beyond the scope of this paper, but will be addressed in a subsequent article. Analysis of targets should follow the opportunity assessment and identify areas in which aberrations in the targets make a previously identified opportunity less attractive.
What We Still Don’t Know
There are many details of the target-setting process that have yet to be revealed. Some of these are the following:
- What factors will be included in the regression? The CMS document mentions some general categories, but we don't know the specifics. For example, it mentions that HCCs will be used, but not if they'll be used as counts (patients having 1-3 HCCs, 4-6, etc.) or actual HCC scores. It also mentions that specific factors may be used for specific episode types; for example considering fracture status in major joint replacement, but we don't have any other examples.
- How much does each factor contribute to the target? Might a change in HCC score cause the target for particular episode to double, or might it only increase by a few percent?
- How much will targets change based on patient characteristics? This is significant in understanding the stability of targets over time. For hospital participating in CHF episodes, can the target expected to vary by large amounts or small amounts across quarters?
- How well do regression equations measure actual costs? This issue is discussed below.
- What are the regression coefficients? Knowing the regression coefficients will help to determine the degree to which each regression component will affect the target.
- What regression data will be provided to the hospitals, and when? Providing data on the patient characteristics used in the regression will assist hospitals in understanding changes in targets.
What This Means to BPCI-A Participants
Now that we've seen how CMS will compute the targets, what does this process mean to BPCI Advanced participants? Will this new target methodology be an advantage or disadvantage to them? In our opinion there are four major areas of concern: inconsistency, lack of transparency, prediction ability and potential bias.
Target Inconsistency
Participants in these types of programs always want to know their financial targets at the beginning of the performance period. In the BPCI program, the targets were retrospective and not finalized until the last reconciliation, which occurred more than a year after the performance period ended. This retrospectivity was unpopular with participants, and to mollify those participants CMS established "prospective" targets for the CJR program, which are set in advance and don't change throughout the reconciliation period. The BPCI Advanced program attempts to follow this pattern of "prospective" targets and will announce “preliminary targets” in May 2018, but these targets will be revised to utilize the actual case mix in the performance period. Ideally this change will compensate for a difference in patient mix between the baseline and performance period, and will adjust for differences in the cost of care between these mixes of patients. The extent to which this occurs is yet to be determined. Meanwhile, participants who wish to know the exact value of their targets in advance will continue to be dissatisfied.
Lack of transparency
In the BPCI and CJR programs, it was relatively simple to understand the relationship between the participant’s historical baseline cost (in BPCI) or regional average cost (in later periods of CJR) and their target in the performance period. However, as described above the target-setting process in BPCI advanced is extremely complex and relies on data elements that may not be available to participants. Explaining changes in targets across time periods may be problematic. Therefore, creating a level of confidence necessary to obtain buy-in for these programs may be more difficult than in the previous programs.
Prediction Ability of the Risk Adjusters
The ability of the regression model to adjust targets for differences in patient characteristics will be a key factor in target accuracy. In general, there should be a strong correlation between the predicted episode costs and the actual costs. A weak correlation would show that the model does not actually account for variations in episode costs that are caused by different patient characteristics. In the calculation example in the CMS Target Price Specifications shown in the accompanying graph, predicted cost shows a strong correlation with actual cost. The degree to which this occurs in the actual data remains to be seen.
Statistical aberrations
Regression models are built from large populations. While their statistical stability is strong in larger groups, it weakens as the size of the group diminishes. Therefore, there is some concern about the applicability of these models to participants having a small number of episodes.
Summary and Conclusions
The BPCI and CJR targets were developed through a relatively simple process in which the paths to success, and opportunities for achieving success, were clear. This simplicity was often criticized for not incorporating such factors as patient mix and hospital characteristics. In BPCI Advanced it appears that CMS tried to cover all bases, and perhaps included many factors primarily to be able to address critics by saying "It's in there". Hopefully the balance that they hope to create with this methodology will achieve its goal, creating the potential for success for broader group of participants.