Disease based output measures for hospitals

By Qinghuan Luo, Economic Research Section

Health care plays a central role in the lives and well-being of Australians, and is a large and growing share of the Australian economy¹. Accurate measurement of health care output and productivity is important for policy makers and researchers to gain insights into key drivers for health spending growth, and how health care contributes to growth of the Australian economy. Rapid advances in medical technology and innovation generally lead to improved outcomes, e.g. new, less invasive procedures or more effective drugs. This poses particular challenges in measuring changes in output price and volume which take into account impacts of improvement in health care quality. Measurement of health care output is further complicated by a lack of market prices for much of the output, as a large proportion of health care in Australia is in the non-market sector.

As part of the ongoing development work on enhancing output measures for health care, the ABS aims to derive improved output volume measures that capture quality changes including those reflected in a shift or substitution across different types of treatments or different care settings². Fully accounting for quality changes, especially for those linked to treatment outcomes, requires patient level data that contain information on treatment outcome. Such data are generally not available. It is more feasible to construct partial disease based output measures which can take into account some aspects of such shift or substitution. The improved output measures can be used for deriving productivity estimates.

This paper focuses on constructing disease based output measures for hospital services in Australia, as part of the development work outlined in the recent discussion paper on enhancing health care output measure². Hospital services are an important area of the health care industry due to its large share of overall health expenditure. There is a strong interest in hospital measures in relation to the efficiency and productivity of this sector, as public hospital spending comprises a large proportion of governments overall spending. It is more feasible to construct a disease based output measure for hospitals than for other areas of the health industry, as hospital administrative data with clinical information are routinely collected by state/territory health authorities and these data can be potentially used for measuring hospital output. However, as episode level data are not readily available, this study focused on the use of aggregated data that are more easily accessible and suitable for regular production of statistics.

This paper presents the results of quantity indexes constructed for hospital output on a disease basis using the Australian Institute of Health and Welfare (AIHW) summary statistics of hospital admitted patient care expenditure.

This paper outlines the approach for constructing a disease based quantity index for measuring changes in real output of hospitals, and presents the results based on high level expenditure data sourced from Australian Institute of Health and Welfare (AIHW). It should be noted that the same approach discussed here could be used for constructing a disease based price index where unit cost is used as proxy price and output volume can be derived by deflation.

Generally, measurement of output on a disease basis requires data with expenditure or cost mapping to disease and patient characteristics, which, in the case of hospitals, can be produced from hospital administrative data at episode levels. However, such administrative data were not available for this analysis. This paper focuses specifically on output measures for hospital admitted patient care as this type of care is the dominant part of hospital services and both summary statistics of expenditure on diseases by demographics and separation statistics are available from AIHW.

Section 2 provides an overview of Australian hospital services, output measures and data sources. Section 3 discusses the method for constructing a quantity index for hospital output. Section 4 presents the results of quantity indexes for hospital admitted patient care.

Australian hospital services are provided by both public and private hospitals, with public hospitals being mainly funded by state/territory governments and the Australian Government. Treatments of private patients at private hospitals are partially funded by the Australian Government directly through Medicare, and indirectly through government rebates paid on private health insurance (as are private patients in public hospitals).

Australian hospital services can be largely categorised into three broad types of care: admitted patient care (AP), Emergency Department care (ED), and non-admitted patient care (NAP). Admitted patient care refers to hospital services provided to patients who undergo a hospital’s formal admission process. By contrast, non-admitted patient care refers to services provided to patients who do not undergo a hospital’s formal admission process. The majority of hospital spending is in admitted patient care (see Figure 1). This type of care is the primary focus of this paper. Admitted patient care comprises three subcomponents: admitted acute care, admitted subacute care and admitted non-acute care, which are all included in this study.

When discussing hospital output it is useful to consider health care output in the broader context of health care. Here output refers to output quantity or volume. It is conceptually preferable to measure health care output on a basis of complete treatment rather than specific types of goods and services going into the treatment (Triplett 2001; National Research Council 2010; Aizorbe and Highfill 2015; OECD, Eurostat and World Health Organisation 2017).

One advantage of redefining output based on treatment of disease is that substitutions across different types of service or health care providers, such as a shift to less invasive treatments, or from inpatient to outpatient care, can be accounted for (Aizorbe and Highfill 2015; Groshen et al 2017). A second advantage is that it is more meaningful to discuss quality changes, especially those reflected as improvement in treatment outcome, on a disease basis than focusing on specific type of goods and services (Hall and Highfill 2014). For example, measuring quality changes may require outcome information on treatment for specific types of disease. While there are good reasons for full disease based measurement, in practice, it is more realistic to measure output on a partial disease basis, limited to broader categories of providers, such as hospitals versus non-hospital providers⁴.

Based on this concept, it is desirable to measure hospital output based on treatment of disease across different types of care (admitted, non-admitted and ED). A disease based measure can capture shifts across different types of treatments for the same disease at hospitals, or shifts across different types of care, e.g. from inpatient to outpatient care (i.e. non-admitted patient care). Calculation of output measures on a disease basis across three different types of care can prove to be challenging in the Australian case due to considerable challenges in mapping non-admitted care and ED data to disease. The non-admitted care and ED data use different classifications and may not contain the same level of detailed clinical information as admitted care. Therefore this work focuses on admitted patient care.

In Australia, hospital data are routinely collected by state and territory health authorities. Relevant data sources include Independent Hospital Pricing Authority (IHPA) National Hospital Cost Data Collection (NHCDC) and AIHW National Hospital Morbidity Database (NHMD). NHCDC contains episode level costs and activity for covering all types of hospital care and is used for producing Australian Refined Diagnosis Related Groups (AR-DRG) cost weights. NHMD contains episode level records of admitted patient care at both public and private hospitals. Disease based analysis requires that cost of treatment be mapped to disease, which is usually done using the episode level data (e.g. AIHW 2017).

In this study, the summary statistics of hospital expenditure on admitted patient care was used, as the episode level data is not readily accessible. The summary data was produced by AIHW, with expenditure being mapped to disease coded in the International Classification of Diseases (ICD) at the Chapter level by age groups (5 year age groups). (ICD Chapters are the broadest categories of diseases.) The data includes both public and private hospitals, from 2004-5 to 2012-13. The data contains two sets of cost allocation to diagnoses. The first set is allocation of expenditure to principal diagnoses only. The second set is allocation of expenditure by splitting it across all cost-relevant diagnoses equally including principal diagnoses, as a way of including comorbidity (AIHW, 2017). Cost-relevant diagnoses are referred to diagnoses that contribute to the total treatment cost⁵. The data uses the Australian Modification of the 10th Revision of ICD (ICD-10-AM), 8th Edition and the data item description is provided in Table A1 in Appendix.

In addition, hospital separation statistics by principal diagnoses from the AIHW Principal Diagnosis Data Cubes was used for constructing quantity indexes of ICD Chapters by age group from the ICD subchapter level. The data was also used for updating the number of separations in the expenditure data to 2014-15.

As hospital quantity data, such as hospital separation statistics, are more detailed and timely than cost data, the quantity index approach is adopted. (Note that the terms of quantity index and volume index are used interchangeably.) Within this approach, an annual chained Laspeyres quantity index is constructed using numbers of episodes of care (i.e. separations) and input cost or expenditure weights in a reference period. The corresponding implicit price index can be derived indirectly from the nominal output divided by the quantity index.

In constructing a quantity index, stratification was used for controlling for heterogeneity in treatment of disease and in patient characteristics. In theory, episodes of treatment should be grouped by disease by patient characteristics at sufficiently detailed levels so that products in each group can be considered as being approximately similar. However, the level of detail where diseases and patient characteristics can be grouped is largely determined by data availability, as in the case considered here. In the expenditure data used here diseases are categorised broadly by ICD Chapters, with patient characteristics being differentiated by age groups. Note that age is one of the most important characteristics, e.g. patient's comorbidity tends to depend on age. Gender is also a relevant characteristic of patients but it was not included in the data. It is reasonable to assume that differentiation between genders can be partially reflected in some ICD codes (though there can still be a difference in comorbidity.)

Although data on average length of stay (ALOS)⁶ is available from the Principal Diagnosis Data Cubes, this is not used as a defining characteristic for grouping episodes. This means that different patient days for the same disease may be put in the same group. As a result, new or improved treatments or better care that reduce patient days and overall treatment cost can be reflected as price effects rather than reduction in volume (Schreyer 2010; Diewert 2017). A further reason for not grouping episodes by length of stay is that length of stay may implicitly depend on age groups and patient’s comorbidity (as noted above, comorbidity tends to depend on age). Patients in the older age groups or patients with comorbidities tend to require a longer stay in hospitals. However, average length of stay could be used as proxy weights, e.g. for ICD diseases at a lower level of detail when the cost information is not available. For example, average length of stay can be used as proxy weights for ICD subchapters by age group when cost weights are not available (see discussion below and footnote 7).

Quantity indexes by ICD Chapter were calculated as an annual chained Laspeyres index using two different methods. In both cases, the cost allocation to principal diagnosis provided in the expenditure data was used.

The first method uses the number of separations of each diagnosis-age group at the Chapter level as the lowest level indexes. There are about 350 diagnosis-age groups at this level. These unit value indexes are aggregated (across age groups) to the indexes for ICD Chapters and the total quantity index using expenditure weights.

The second method is to construct quantity indexes for ICD Chapters from diagnosis-age groups at a greater level of detail, i.e. the subchapter level. There are around 4000 diagnosis-age groups at this level. As subchapter expenditure weights are not available, average length of stay is used as proxy weights. The aggregation involves two stages. The first stage is to aggregate subchapter level groups to quantity indexes of Chapter for each age group using average length of stay as weights. The second stage is to aggregate these indexes to the Chapter indexes and total index using the expenditure weights as in the first method.

One complication in stratification by disease by patient characteristics is comorbidity. Although the data includes alternative allocation of cost to cost-relevant diagnoses, it assumes equal weights among the relevant diagnoses. Due to lack of well-accepted methods for including comorbidity, allocation of cost to principal diagnoses is assumed here.

It is worth commenting that AR-DRG can be an alternative option for grouping treatments of diseases. In Australia AR-DRG is used for grouping patients in admitted patient care. DRGs are structured by broad categories of medical, surgical, and other. The main advantage is that published DRG weights are available for admitted acute care. However, DRG groups patients based on their similar resource use, strictly speaking not by disease. For example, patients with different principal diagnoses but with similar resource use can appear under the same DRG. Treatment of the same type of disease may fall under different categories, such as medical versus surgical treatments. Based on these considerations, stratification by ICD of diagnoses is preferred over DRG. AR-DRG can be useful for differentiating patient severity in grouping episodes of care. However, this information is not included in the summary data used here.

Table 1 shows average annual growth rates of annual chained Laspeyres quantity indexes by ICD Chapter, calculated using the first method mentioned above, i.e. calculation from the diagnosis-age groups at the ICD Chapter level. A quantity index was calculated without differentiation between public and private hospitals. This allows a compositional shift between sectors. The aggregated quantity index generally grows at a slower rate than growth in total number of patient separations. The annual growth rate of the aggregated quantity index is about 3.4% over the period from 2004-05 to 2014-15, considerably faster than the annual growth rate of the Australian population (about 1.7%)⁸. This means that real expenditure on hospital admissions per Australian increased during this ten year period. This could be due to factors such as aging population and/or increase in prevalence, though population health data is needed for further analysis of these factors in driving volume growth – which is beyond the scope of this paper.

The aggregated quantity index exhibits a general growing trend, though the growth shows a brief slowdown in 2012-13. The slowdown can be attributed to the slower growth in public hospital admissions.

Figure 2 compares the aggregated quantity index (method 1) with the one calculated using the second method where quantity indexes by ICD Chapters and age groups were constructed from the numbers of separations by ICD subchapters and age groups, weighted by average length of stay (see Section 3). The quantity index using the second method trends similarly, but at a slightly lower growth rate, to that calculated using the first method. Figure 3 shows the price indexes derived by dividing the aggregated nominal expenditure (from the expenditure data) by the corresponding quantity indexes.

Figure 4 compares public versus private hospitals. The indexes were calculated using the first method as for Table 1. Private hospital output increased at a faster average annual rate than public hospitals, mainly driven by a strong growth in private hospital admissions. Growth in private health insurance membership could be a factor in driving the increased demand for private hospital services, as the private insurance hospital coverage, as percentage of population, increased from about 43.3% in 2004-05 to 47.3% in 2014-15⁹.

Figure 5 compares percentage movement in the aggregated hospital quantity index calculated by considering treatments at public and private hospitals to be two different products. This assumes no substitution between the two sectors. The result shows that while there are minor differences, the two indexes trended similarly. Note that relative weights between public and private hospital admitted patient care are based on the aggregated expenditure in the AIHW data. 

The disease based quantity index can be used in the National Accounts for estimating real output and productivity of hospital services. The quantity index can be used directly for estimating changes in output volume measure on the assumption that the index movement is representative of overall hospital output. Preferably, the quantity index can be used for constructing a deflator or implicit price index which is then used for deriving volume measure by deflation. For example, the quantity index can be used with nominal expenditure, say for admitted care, to create a price index, as in Figure 3 - which is then used as a proxy price measure for total output of hospitals to derive its volume measure.

As the expenditure data only has ICD Chapter level disease breakdowns, this could have some limitations on capturing a prevalence pattern shift, i.e. compositional changes between different diseases within ICD Chapters. For this reason, the second method as discussed above is preferred, although this could be improved using data with expenditure allocated to a greater level of detail in ICD.

Note there are some limitations for using numbers of separations as a basic unit of quantity. There are cases where multiple separations within a defined period (e.g. financial year) may actually form part of treatment for the same disease, where it should be counted as one treatment. There is a similar issue with readmission being accounted as a separate treatment, should be in some cases considered as part of the one treatment. Nonetheless, as the majority of separations are single episodes, use of numbers of separations as a counting unit can be considered a reasonable approximation.

As the present approach does not include non-admitted patient care, it does not capture substitutions between inpatient care and outpatient care. For example, due to advance in medical technology, a treatment that traditionally requires admitted care can now be done at a clinic in a hospital as outpatient. Future work could include both non-admitted patient care and ED presentations when relevant data become available. Non-admitted care data usually contain some clinical information that allows disease mapping. From 2011-12, hospital ED data collection started to include clinical information that can be used for disease mapping. As a result, in principle, it is possible to construct a disease based output measure. However, disease mapping for both non-admitted care and ED services remains challenging due to different classification systems being used in data collection. This could be an area for future development.

It is worth emphasising that although a disease based approach can capture some aspects of quality changes reflected as compositional shifts or substitutions, explicit quality adjustment is required to account for quality change reflected as a change in treatment outcomes. However, development of explicit quality adjustment is a major challenge for National Statistical Offices (NSOs), due to a lack of both a well-accepted conceptual framework and relevant data for measuring changes in treatment outcomes. For further discussion on explicit quality adjustment and the ABS preferred approach, see the recent discussion paper².

The key feature of the disease based approach used here is that products are grouped by diagnoses and patient characteristics. This has the advantage that the effects of substitutions across different types of treatments, driven by advancements in medical technology and improved care practice, are reflected in output measures.

As admitted care is the dominant component of hospital services in terms of hospital expenditure, the quantity index for admitted care can be used for constructing a deflator or implicit price index for hospital services. The price index can be used for estimating changes in real output of all hospital services. Future work could include non-admitted patient care and ED in the output quantity index.

As a disease based quantity index or price index has breakdowns by disease type, it is useful for disease based analysis of drivers for growth in health expenditure, such as identifying types of disease that contribute to the rising health expenditure, or determining whether growth in spending for a particular type of disease is due to an increase in the disease prevalence or in volume per patient.

The quantity index may be useful in enhancing productivity measures and providing further insights into the health care industry. The ABS currently does not measure productivity for the non-market sectors (including health care) and has a long-term goal to develop such measure for these non-market sectors. The ABS will continue to explore potential enhancements in this area.

Table A1 is summary of the AIHW expenditure data used in this analysis. The data is summary statistics of hospital expenditure on admitted patient care for both public and private hospitals, including psychiatric hospitals, in Australia from 2004-05 to 2012-13.

The data contains allocation of hospital expenditure on admitted patient care to each separation, with mapping to diagnoses, derived using the AIHW Hospital Morbidity Costing Model (HMCM) (AIHW 2017). The HMCM uses the Independent Hospital Pricing Authority (IHPA) AR-DRG weights and other relevant cost information or cost indicators (such as length of stay) from National Hospital Cost Data Collection (NHCDC) (IHPA 2017). The data uses ICD-10-AM, 8th Edition.

The AIHW data is suitable for measuring hospital output of hospital admitted patient care as it has expenditure mapped to disease, covering all types of admitted patient care separations, including acute, subacute and non-acute separations. By contrast, the AR-DRG weights published by IHPA cover admitted acute separations only, as they are used primarily for Activity Based Funding (ABF) hospital pricing.

The data contains two different allocations of cost to disease. The first one is allocation of cost to principal diagnosis only, without considering comorbidities. For example, if a separation has three diagnoses, the total cost is assigned only to the principal diagnosis. The second one is a split of cost evenly across all cost relevant diagnoses. For example, if a separation has three diagnoses, 1/3 of the cost is allocated to each diagnosis. The summary data contains aggregate of number of diagnoses in separation.