In our previous article on this topic we discussed ways in which randomness in patient selection causes variations in average episode costs. Even for higher-volume DRGs, these variations can create significant differences in the settlement amounts during CMS reconciliations. In this article we describe a simulation methodology that allows estimating the extent to which these variations will occur. This will assist participating organizations in setting proper expectations and planning for expected cash flows resulting from BPCI quarterly reconciliations.
Historical Average Episode Cost Variations
Taking a concept from Steve Wiggins’ excellent presentation at the National Bundled Payment Summit, we built a simulation model that utilizes the historical variation in quarterly average episode costs to estimate the potential for future reconciliation surpluses and deficits. The parameters for this model were developed from the historical data shown below. This table shows the average cost (to Medicare) for a selected bundled payment DRG throughout the historical periods for which completed (fully paid) claims were available.
In this example, the average quarterly episode costs for this DRG vary from a high of $42,694 to a low of $31,011, with a mean of $36,355 and a standard deviation of $3,009. These costs are graphed below along with a trendline for the periods.
The exponential trendline shows the costs are rising slightly each quarter at a rate of approximately .34% each quarter. This overall trend is important in evaluating the effectiveness of care management programs that may have been implemented, because the significant quarterly variation in cost makes it difficult to assess the underlying trends.
Simulation
To estimate the probability of future quarterly reconciliation surpluses and deficits due to random variation, we constructed a simulation model using the mean and standard deviation of the cost distribution above. We assume that these costs will be normally distributed as the central limit theorem suggests because they are averages of averages. We then created a series of 10,000 normally distributed random numbers that reporsent average episode costs using the mean and standard deviation parameters above, and charted the results.
This distribution shown below obviously follows a typical normal distribution pattern because it was derived from normally-distributed random numbers. Each bar shows the number of times an average episode cost (within a specific band of costs) occurred out of the total of 10,000 samples.
But our goal isn't to estimate the probability of a specific average episode cost; it's to estimate the probability of costs exceeding a certain threshold, with that threshold being the target amount. For simplicity, we assumed a target amount at the historical average cost. Therefore our analysis is simply to determine the probability that future quarterly average episode costs will exceed the average by an amount deemed to be unacceptable. For purposes of our example, we assume that a “significant” probability of loss of $5,000 or more per episode would be deemed unacceptable, meaning that we need to estimate the probability of the quarterly average episode cost exceeding $41,355 and decide whether that probability is “significant”. (Alternatively we could pick a percentage limit and determine the dollar amount that matches that percentage limit. We use that approach later in this article.)
For this estimate we need a cumulative loss probability curve, which is shown below. This curve shows the probability that a quarterly average episode cost will exceed the amount selected on the curve. There is obviously a 50% probability that the quarterly cost will exceed the average of $36,355, but what is the probability that they will exceed the $41,355 threshold established above? From the graph below, it appears that this probability is about 7%. If management believes that this is an acceptable risk threshold, it could continue to participate in this DRG. Otherwise, the variation would be beyond the acceptable risk, and the organization would decline to participate in this DRG.
Factors contributing to risk
In bundled payment arrangements, several factors contributed to participation risk:
· The inherent variation in the episode cost of the DRG-defined episodes. In general, surgical DRGs have lower coefficients of variation than medical DRGs.
· Number of episodes – DRGs having high number of episodes have lower variations in average episode cost than DRGs with fewer episodes.
· Length of the reconciliation period – Average episode cost computed on a quarterly basis will have lower variation than those computed on a monthly basis, which is one reason why quarterly variations are generally preferable. Annual average episode costs have approximately half variation of quarterly average episode costs because costs move closer to the average over longer periods of time.
· Participation in multiple episode families can reduce risk because costs across episode families rarely vary in the same direction of same time. In our article on Episode Choices and Bundled Payment Risk we show that participating in a larger number of episode families generally decreases overall cost variation.
Performing a comprehensive risk assessment
The oversimplified example above shows the analysis of BPCI participation risk for a single DRG, assume the target rate at the average historical cost. In actual practice this will not occur; the target will start at 2 to 3% lower than the historical cost and will be adjusted for Medicare payment rate changes and national productivity trends. In addition, BPCI participants should expect an overall downward trend in costs as their care management programs take effect; otherwise they should not be participating in the BPCI program. Also, participation in a single DRG is not possible; the minimum level participation is an episode family that includes multiple DRGs. Finally, many organizations participate in multiple episode families.
A comprehensive risk assessment incorporates these factors into the simulation process and estimate the variation in final reconciliation cash flow incorporating all of the above factors. In this way a participating organization can anticipate the variation in reconciliation amounts due to random patient selection, create reasonable expectations among physicians and the management team, and estimate cash flow reserves necessary for ongoing participation in BPCI program.
The BPCI 360 analytics tool set includes an Episode Risk Estimator that allows assessing the statistical risk associated with participating in individual episode families as well as combinations of families. It computes statistics for monthly episode costs for selected DRG groups and analyzes the distribution of costs across quarters. It then computes the maximum expected loss (at the 95th percentile limit) in a quarter as a dollar amount and also as a percentage of total episode costs. The example below shows these statistics for the combination of congestive heart failure, CABG, major joint replacement and stroke episode families. For this hospital using the selected historical periods as the basis, the average quarterly episode cost would be about $8 million, and the maximum loss (using the average cost as a surrogate for the target) would be about $880,000, or about 11% of the total cost. Using this tool, applicants to the BPCI program can evaluate the effect of participating in different episode families, and understand the contribution of each family to the total risk of participation.
