Episodes of care were constructed to understand the factors associated with the use of different PAC services during an episode. Transition patterns were analyzed to learn more about the proportion of patients discharged to each post-acute setting. DRG-specific episode patterns were also examined to identify variations in condition-specific discharge patterns. Key variables of interest in the episode analyses were utilization and payments by type of PAC service. Note that the Medicare payment amounts on each claim were used and these payment amounts reflect any payment adjustments made for outlier costs, or facility characteristics, such as indirect medical education (IME), disproportionate share payments (DSH), and local wage differences.
The PAC episode data were also case-mix adjusted using both the 3M APR-DRG grouping software and the MS-DRG software. Several revisions of the Medicare DRG system have been implemented to overcome the limitations of DRGs. These revisions include the refinement of the way principal diagnoses and procedures are stratified into categories based on the presence or absence of substantial complication or comorbidity (CC) in secondary diagnoses. In the March 2008 ASPE report, RTI used only APR-DRGs to case-mix adjust PAC episodes due to the timing of the analyses and the release of the MS-DRGs. In this study, both the APR-DRGs and the MS-DRGs were used for case-mix. The use of the APR-DRGs allows for comparison to the previous analyses and the use of the MS-DRGs reflects current Medicare policy.
Comorbid conditions were also measured using Hierarchical Condition Categories (HCCs). The HCCs were used in these analyses because they to provide a convenient method for collapsing ICD-9 codes into meaningful disease groupings to identify comorbid or complicating conditions. HCCs were assigned to the index acute admission claims to identify the presence or absence of a comorbid condition. The HCCs provided additional information on the effects comorbidity on service utilization. A description of the application of the APR-DRG, MS-DRG, and HCC software to the 2006 PAC episode files follows.
APR-DRG System. In the 1980s, CMS developed All-Patient DRGs (AP-DRGs) and expanded Medicare DRGs to include neonatal, obstetric, and other conditions typical to the under-65 population. This resulted in the development of almost 1,200 DRGs, yet patient severity of illness and mortality were not predicted and many secondary diagnoses were not included in the AP-DRG system. AP-DRGs formed the basis for APR-DRGs which were developed by 3M Health Information Systems in the early 1990s. APR-DRGs added severity of illness and risk-of-mortality subclasses for each base APR-DRG. In determining the severity level, 3M revised the CC list to accommodate the non-Medicare population. 3M also incorporated principal diagnosis, age, interactions of multiple secondary diagnoses, and combinations of non-operating procedures with principal diagnosis. The severity of illness and risk-of-mortality subclasses have levels of 1 to 4, indicating minor, moderate, major, and extreme, respectively. Based on these enhancements, APR-DRGs represented a significant improvement over previous severity-adjusting systems.
MS-DRG System. In FY2008 CMS adopted Medicare-Severity (MS) diagnosis-related groups (MS-DRGs) to account for differences in patient mix in the Medicare inpatient hospital payment system (Wynn and Scott, 2007). The grouping logic for the MS-DRG system is the same as the CMS-DRG logic. It collapses paired DRGs (DRGs distinguished by the presence or absence of CCs and/or age) into base DRGs and then splits the base DRGs into CC-severity levels. The general structure of the MS-DRG logic establishes three mutually exclusive, hierarchical severity levels for each base DRG: 1) with major CCs (MCCs), 2) with CCs, and 3) without CCs. However, severity levels are consolidated for a base DRG if the following criteria for a subgroup will not be satisfied:
- At least a 3.0 percent reduction in variance would result
- At least 5.0 percent of discharges in the MS-DRG would be assigned to the subgroup
- At least 500 discharges would be assigned to the CC or MCC subgroup
- Subgroups would have at least a 20.0 percent difference in average charges between them
- Subgroups would have at least a $4,000 difference in average charges between them
When the subgroups did not meet these criteria, the MCC and CC severity levels were collapsed in one of three ways: "With CC/MCC DRG" and "Without CC/MCC DRG" Or " with MCC DRG" and "without MCC DRG" (collapsing "no cc" and "cc" severity levels. Third, some base MS-DRGs were not subdivided at all because of insufficient differences between the subgroups (based on the previously listed criteria, such as 3.0 percent reduction in variance or a $4,000 difference in average charges between subgroups). This variable stratification complicates the definition of severity across DRGs. Within each DRG, a discharge is assigned to the highest severity level of any secondary diagnosis. There is no adjustment in the severity-level for additional factors or CCs, except that certain conditions with high-cost devices are assigned to a CC severity level.
In March 2007, RAND released a report evaluating the MS-DRG system and other severity-adjusted DRG systems that the agency was considering (Wynn and Scott, 2007). They found that, in comparison to the other severity-adjusted systems, the MS-DRGs have a much higher percentage of discharges assigned to the lowest severity level. For example, 60.0 percent of discharges are assigned to Severity Level 0 in the MS-DRG system, compared to only 20.0 percent in the APR-DRG system. Wynn and Scott cite several reasons for this, including the re-assessment of CC assignments, the collapsing of the no CC and CC severity levels in some DRGs, and no severity subgroups in 53 base DRGs. The researchers also found that the MS-DRGs explain 43.0 percent of the cost variation, which was a 9.0 percent improvement over the unadjusted CMS-DRGs.
Wynn and Scott (2007) note that although the underlying logic of the MS-DRG system uses standard severity levels (for which lower numbers indicate lower levels of severity), the criteria for establishing severity subgroups result in severity levels that vary by base DRG. Because the severity levels are often collapsed and the resulting subgroups depend on the particular DRG, the MS-DRG is a more complicated system to understand than the other severity-adjusted DRG systems. Wynn and Scott did, however, note that one major advantage of the MS-DRG system over other severity-adjusting systems is that the CC list and severity-level assignments reflect current Medicare data and the logic therefore likely reflects current patterns of care.
As noted previously, in the March 2008 ASPE report, RTI used only APR-DRGs to case-mix adjust PAC episodes due to the timing of the analyses and the release of the MS-DRGs. For the research described herein, both the APR-DRG and the MS-DRG grouper software packages were used to assign a severity-of-illness measure to the index acute hospitalization. Analyses of utilization, length of stay, and Medicare payments were performed by DRGs and by APR-DRGs and MS-DRGs to learn more about differences in post-acute service use by diagnosis and severity level. Severity adjustment using the APR-DRG system allows for comparison to similar analyses using 2005 Medicare claims data performed in the previous ASPE March 2008 project. Applying the APR-DRG system to the 2006 data also provides an opportunity to understand some of the differences between the APR-DRG and MS-DRG systems.
The DRG-specific nature of the MS-DRG grouping logic means that it is not possible to have direct comparisons between the APR-DRGs and the MS-DRGs. Direct comparisons between the two groupings were not made in this report; rather, we report utilization and payments separately by APR-DRG and MS-DRG in order to illustrate the different post-acute care patterns by each of the severity groupers.
Comorbidities. Comorbidities are conditions that exist at the same time as the primary condition in the same patient (CDC, 2008). For example, hypertension is a comorbidity of many conditions such as diabetes, ischemic heart disease, and end-stage renal disease. Many of the common comorbidities are also chronic diseases, including cardiovascular disease, cancer, and diabetes. These conditions are among the most prevalent, costly, and preventable of all health problems (CDC, 1999). Chronic illness accounts for 70.0 percent of deaths and over 75.0 percent of direct health care costs in the United States (Thrall, 2005). Because people with comorbid and chronic illnesses have greater health needs at any age, they account for a disproportionately high share of healthcare costs (Hoffman, 1996). About 20.0 percent of all Medicare beneficiaries have five or more chronic conditions, and chronic, comorbid conditions account for over two-thirds of Medicare spending (Berenson, 2004).
It is important to examine the impact of comorbid conditions on an acute event to understand the resource drivers in high-cost populations, such as the chronically ill. In order to examine the effect of chronic and complicating conditions on PAC utilization and costs, we examined the HCCs present in our sample of beneficiaries with an index acute hospital claim in 2006. As described in detail below, the HCC software is generally used for risk-adjustment purposes. The software generates a set of variables to indicate the presence of comorbid conditions and also generates a risk score that can be used in risk-adjustment models. For these analyses, we used only the variables indicating the presence of comorbid conditions.
In 2004, CMS began to use the CMS-HCC risk-adjusted payment approach, which uses diagnostic and demographic information on the claims to predict resource use (Pope et al. 2004; Noyes et al. 2006). This risk adjustment model uses a subset of International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis codes from the claims to place beneficiaries into 70 disease groups called HCCs. Each HCC includes conditions that are related clinically and have similar effects on costs. Several researchers have incorporated HCCs into models to represent comorbidities (Pope et al. 2004; Noyes et al. 2006; Ettner et al. 2001). The principal application of the condition categories (CCs) is to predict costs of Medicare Advantage plan enrollees. In that case the incremental cost of having a disease in each CC is determined statistically and the effect of the CCs and selected demographic factors are summed to create a predicted total. The same classes, with different incremental costs, are used for aged and disabled community dwelling, long-term institutionalized and ESRD beneficiaries.
The classification system was developed in a collaborative process with physicians and econometricians. The clinical foundation of small, clinically homogeneous groups of ICD-9-CM codes was merged with data-driven information to develop the larger CC groupings. Some of the CCs are themselves grouped into hierarchies of related conditions. During a year, a person may be diagnosed with lower and higher levels of severity of a condition. When the hierarchy is imposed, only the highest cost level is used in describing the person. Having been coded with a lower related CC also adds to the prediction. In the diabetes hierarchy, if a person has been coded with simple diabetes, diabetes with ophthalmologic manifestations and diabetes with renal manifestations, the only CC used would be the last in the group. Use of the hierarchy with the CCs is optional.
The classification system, with or without the hierarchy, has been used in other settings in which controlling for risk is important. For example, the hospital mortality ratings on the Medicare Compare web site are developed with sets of the CCs as part of the risk adjustment
The HCC system was created using the guidance of several principles (Pope et al. 2004), including:
- HCCs are clinically meaningful. They all relate to a reasonably well-specified disease or medical condition that defines the category.
- HCCs predict medical expenditures. Diagnoses in the same HCC are reasonably homogeneous with respect to their effect on both the current and future year's costs.
- HCCs have adequate sample sizes. Diagnostic categories that will affect payment have adequate sample sizes to permit stable expenditure estimates. Extremely rare diagnostic categories cannot reliably determine expected costs.
- HCCs use hierarchies. The most severe manifestation of a given disease process defines its impact on costs, so that related conditions are treated hierarchically, with more severe forms of a condition being flagged for a person (and less serious ones not being flagged).
- HCC system does not reward coding proliferation. The number of times that a particular code appears does not increase predicted costs.
- HCC classification system assigns all ICD-9 codes. There is exhaustive classification, because each diagnosis code contains relevant clinical information.
The CMS HCC software generates variables identifying the presence or absence of 189 conditions. For the purposes of these analyses we looked at the 70 most common conditions. Appendix A shows the 70 HCC groups that are used in this analysis.
It is important to note that hierarchies are imposed among related conditions, so that a person is only coded for the most severe manifestation among related diseases. Also, although HCCs reflect hierarchies among related disease categories, for unrelated diseases, HCCs are allowed to accumulate. For example, a beneficiary with heart disease, stroke, and diabetes will have at three separate HCCs coded (and their costs are predicted to reflect the increments for all three diseases). As Pope and colleagues note, the HCC model is more than simply additive because some disease combinations interact. For example, the presence of congestive heart failure (CHF) and chronic obstructive pulmonary disease (COPD) is likely to increase predicted costs by more than the sum of the separate increments for beneficiaries who have CHF or COPD alone. In this study we report on the total episode cost by the presence or absence of certain combinations of comorbidities that are often present together within the same individual (COPD and CHF; diabetes and renal failure).
In order to calculate the HCC categories as measures of comorbidity in our sample, we used the available CMS-HCC software originally developed by RTI researchers (reference 18 in Noyes; Pope et al. 2004). The goal in using the HCCs was to learn more about the extent of chronic and comorbid conditions for beneficiaries using PAC and to learn more about how the presence of chronic and comorbid conditions affects the use of services within episodes of care. The HCCs provide a convenient method of collapsing ICD-9 codes into meaningful disease groupings.
Though the HCCs are generally based on a year's worth of claims data for the purposes of risk adjustment, in this case, the HCCs are used for their ability to provide meaningful disease groupings for understanding chronic and comorbid conditions. HCCs were constructed by running the acute index admission diagnoses reported on the claims through the CMS HCC software. The program assigns individuals to up to 70 HCC groups based on diagnoses on the claims. We used the diagnosis codes on the index hospital admission claims to calculate the HCC indicators.
We performed two specific analyses using these HCC indicators of comorbidity. First, we examined prevalence and rank order for the 20 most frequent HCCs in our overall PAC sample. We present the prevalence and rank order of these 20 HCCs for the beneficiaries with index acute hospital admissions in five most common DRGs in our sample. These results are presented in Table 3-11. Next, we calculated the number of HCC indicators assigned to each beneficiary and categorized beneficiaries based on this number (i.e., 0, 1, 2, 3, 4, or 5 or more HCCs). We used these categories to examine index acute hospital admission and episode length of stay and payments for the beneficiaries with index acute hospital admissions in the five most common DRGs in our sample. These results are presented in Table 3-12.