Abstract
Objectives:
The number of countries adopting per case hospital payment systems has been continuously increasing in recent years. Nonetheless, debates persist regarding their consequences for equity of access to services. This concern relates to the failure of diagnostic classifications properly to take into account patients’ care requirements, raising the threat of case selection (‘cream skimming’). We examine the heterogeneity of costs within diagnostic categories related to socioeconomic (SE) factors using length of stay (LOS) as a proxy measure of care needs and costs. We evaluate its consequences in terms of fairness in resource allocation between hospitals.
Methods:
We employ data on all discharges in 2002-03 from a sample of 60 Belgian hospitals (617,275 observations), measuring the association between LOS and SE factors using generalized linear models. We design a resource allocation formula based on the Belgian financing scheme, where non-medical activity is paid based on a normative number of in-patient days, and measure financial penalties and rewards according to whether payment is adjusted for the SE characteristics of patients or not.
Results:
Both patients’ SE status and hospitals’ area SE profile have a significant impact on LOS, which persists after controlling for detailed diagnostic and hospital characteristics. Hospitals treating low income patients are financially penalized as a result.
Conclusion:
SE factors are a predictor of in-patient LOS and should be taken into account in per case resource allocation among hospitals.
Introduction
The number of countries adopting per case prospective hospital payment systems has been continuously increasing since the early 1980s. In the USA, hospitals have been paid a fixed price per diagnosis related group (DRG) for Medicare patients since 1983. Similar arrangements have been adopted in Portugal, France, Canada, Australia, Belgium, and more recently in the UK. 1
Although widely used, much debate persists about the equity consequences of per case payments. Several authors have pointed to the failure of classification systems to accurately reflect patients’ needs, raising the threat of patient selection. 2 Most of this debate has centred on the failure to appropriately account for the severity of disease, pointing to the heterogeneity within DRGs. 3 However, classification systems have greatly improved over time, notably through the inclusion of more diagnostic categories and severity indicators (all patient DRGs and all patient refined DRGs–AP-DRGs and APR-DRGs, respectively). In this paper, we analyse resource allocation according to socio-economic (SE) factors, which has been largely neglected by researchers and policy-makers.
That underprivileged people have higher health care needs has long been recognized. Higher in-patient costs and lengths of stay4–7 are thus desirable from an equity perspective. However, per case payment may threaten equity due to unreimbursed SE heterogeneity. Under per case payment, the hospital receives an average amount for treating all patients within a given category; if categories do not distinguish the care requirements of poor and non-poor patients, hospitals attracting a large share of underprivileged patients are thus unfairly penalized, putting at risk financial solvency and quality of care. Also, unreimbursed SE heterogeneity may create an incentive to attract wealthier patients with lower expected costs and to avoid poorer ones (so called ‘cream skimming’).
We use the Belgium hospital payment system as an example of this debate. In Belgium, non-medical activity (hotel and nursing care) has been traditionally paid for on a per diem basis for all in-patient days. Since 1994, a normative number of in-patient days have been attributed to each admission, first by AP-DRG and then from 2001 by APR-DRG. These normative days correspond to the average length of stay (LOS) within the patient's APR-DRG and severity level. Hospitals are not paid the per diem for days beyond the norm; they make surpluses if the number of in-patient days fall short of the target. As a result, if the diagnostic classification system is limited in predicting differences in LOS between poor and wealthy patients, hospitals with a large proportion of underprivileged patients are unfairly penalized. Note that Belgian hospitals are mostly private, not-for-profit institutions and are free to use their surpluses at their own convenience. The financial situation is, therefore, of paramount importance, as it influences the capacity to invest and attract the best practitioners and, over the long run, the hospital's survival.
Regarding other hospital revenues, medical procedures are paid by fee-for-service; the per diem for non-medical activity results from a complex scoring system based on the structural characteristics of the hospital and from cost comparisons across hospital groups; finally, co-payments are very low. 8 Activity-based payments for in-patient days are thus the major incentive to reduce length of stay.
We seek new evidence on the association between LOS, as a proxy for costs, and SE factors, extending previous research. 9 The impact of SE factors is measured through individual and area-based socioeconomic status (SES). The impact of hospital SE profile is also evaluated to emphasize aggregate SE effects. We compare the effect of SE factors between the more and less detailed diagnostic classification system to highlight the heterogeneity within diagnostic categories and the specificity of SE factors as a LOS-predictor. Finally, we determine the resource allocation among hospitals based on the Belgian financing scheme and measure the equity consequences of excluding SE factors.
Methods
Sample
A random sample of 60 Belgian acute care hospitals was chosen for this study (out of a total of 116). Data on all in-patient discharges were available for these hospitals between June 2002 and June 2003, systematically collected by the Ministry of Health for financing, administrative and medical purposes. The database contained 617,275 overnight in-patient stays. Of these, 98,282 were excluded because they are not covered in the prospective payment system (hospitalizations in psychiatric and rehabilitation units). Then, data are merged with area-based characteristics by neighbourhood provided by the National Institute for Statistics for the year 2001. There are 19,781 neighbourhoods with an average population of 530 inhabitants (median 300 inhabitants) representing small and homogeneous areas. Unfortunately, the data for neighborhoods of fewer than 200 habitants had to be deleted for confidentiality reasons. This limitation reduced the sample to 443,448 observations.
Explanatory variables
LOS was modelled as influenced by socioeconomic and other factors grouped into patient and hospital characteristics (Table 1).
Characteristics of the sample
Indicates the correlation level between the variable and individual LOS. All correlation indices are statistically significant at 5%
CI — confidence interval; LOS — length of stay
Patient characteristics
Individual SES was characterized by four different variables corresponding to the social insurance categories that entitle people to reduced co-payments in Belgium. The first variable refers to patients who meet at least one of the following criteria: welfare recipient; basic income for elderly people recipient; or disability benefit recipients (WELFARE). Reduced co-payments are also attributed to other categories conditional on their deductible household gross income not exceeding €13,606: widowed; disabled; retired; or orphaned individuals; long-term unemployed; older than 50 years; and physically challenged people (LOW_INCOME). The third variable refers to isolated people (ISOLATION) and the last to people considered dependent according to a series of criteria, mainly mobility and self-care ability (DEPEND).
The use of area-based variables allows us to overcome insufficient information on individual SES. In particular, individual SE data are poor at discriminating between retired people whose non-salary income represents a substantial part of their income. Three types of area-based variables are included in the analysis:
Education: the percentage of people older than 64 years with a primary school diploma (P_PRIM65) and those with a high-school diploma (P_HIGH65);
Income: the percentage of people living on a disposable income lower than €20,000 per year (P_POOR);
Assets: the percentage of homeowners (P_OWNER), and the percentage of households without a car (P_NOCAR).
In addition to age and sex, we include the individual APR-DRGs split into four severity levels related to the extent of physiological impairment or loss of organ function (1,028 categories). 10 We compare the results between APR-DRGs and AP-DRGs, the less detailed classification without severity categories (607 categories).
Hospital characteristics
We create a dummy variable distinguishing the hospitals with a disproportionate share of underprivileged patients; i.e. those in the fourth quartile for the four SE individual variables (DIS_SHARE).
Health care use varies across hospitals because of differences in practice style, productivity, or efficiency. The only information available in our data concerns hospital size (more or fewer than 300 beds, SIZE). As an alternative to this limited information, we use hospital dummies (see Statistical analysis).
Statistical analysis
We work in logarithms in order to account for the right-skewed and non-negative distribution of LOS. In order to avoid heteroskedasticity related to re-transformation, we use a generalized linear model with a log-link function (between the linear predictor and the mean of the distribution function) and assume a gamma distribution of error terms. 11 (More detail is provided in Appendix 1, available online only at: http://www.jhsrp.rsmjournals.com/cgi/content/full/16/4/197/DC1).
In order to account for hospital characteristics, we first control for hospital size and the percentage of underprivileged patients (Model 1). However, these two variables are insufficient to reflect hospital practices; hence, our results on SE factors may be biased due to unobserved hospital heterogeneity. We thus estimate a second model including hospital-specific fixed effects; i.e. one dummy per hospital (Model 2). By doing so, the SIZE and DIS_SHARE variables cannot be estimated in this second model, due to multi-collinearity.
Model 2 is then estimated controlling for AP-DRG instead of APR-DRG, in order to detect how more accurate severity indicators affect the impact of SE factors (Model 3). Then, Model 2 is estimated for the sub-sample of people over 64 years (Model 4). Finally, in order to estimate the potential bias due to working on the sub-sample of patients for whom area-based information is available, we estimate Model 2 on the complete sample without the area-based variables (Model 5).
Hospital financing and unfair penalties
Goodness-of-fit criteria indicate that Model 2 performs better than does Model 1 (see Results), so it is used as the reference for the financing scheme. From the estimates of Model 2, we derive the payment of the 60 hospitals in the sample, attributing a predicted number of days to each in-patient stay according to the categories to which it belongs. Four separate payment systems are built: unadjusted, adjusted for AP-DRG without SE factors, adjusted for APR-DRG without SE factors, and adjusted for APR-DRG and SE factors. Unfair penalties are given by the difference between expected and observed LOS for SE categories when using models unadjusted for SE factors.
Results
Descriptive statistics in Table 1 show that LOS is greater for underprivileged categories, among patients coming from underprivileged areas and admitted to hospitals with a higher percentage of underprivileged patients. As expected, LOS increases with age and severity of disease.
Do SE factors influence LOS regardless of severity of disease?
The estimates for individual SE characteristics are significant and conform to intuition (Table 2). In Model 1, welfare recipients have a 15% longer stay on average, while low-income, isolated, and dependent patients experience a 6%, 9%, and 13% longer LOS, respectively. Results are very similar for Model 2.
Estimates from the generalized linear models (gamma distribution)
Significant at the 5% level,
significant at the 10% level. The estimates for AP-DRGs, APR-DRGs and hospital-specific fixed effects are not reported for brevity
DRG = diagnosis related groups; APR-DRG = all patient refined DRG; AP-DRG = all patient DRG
Patients coming from neighbourhoods with a higher percentage of homeowners and a lower percentage of households without a car have significantly shorter stays, on average. Patients from neighbourhoods with a higher percentage of poorer people also stay longer. Longer stays are also associated with a lower percentage of people with a high school diploma among the elderly. The percentage of elderly with a primary diploma in the area is never significant, probably due to the lack of variation across areas.
Comparing Models 2 and 3 shows that SES estimates are slightly lower controlling for APR-DRGs as compared to AP-DRGs, but they remain high and significant. In Model 4, we observe that the impact of low income and being a welfare recipient is lower among patients older than 64 years as compared to the whole sample; however, this result is likely to be artefactual, as income is mostly not observed among many retired people. In contrast, the two area-based education variables and the poverty factor have a higher impact than for the complete sample. Finally, Model 5 indicates that the impact of individual SE indicators is only slightly affected by working on the complete sample; that is, including patients for whom area-based variables are not available. Hence, the potential bias of working on a sub-sample is negligible.
What are the equity consequences when SE factors are not accounted for?
Results addressing this question are shown in Table 3. The case of low-income patients is the most notable. When no adjustment is made, these patients represent a loss of 3.47 days, on average. This represents on average 36% of the total payment for the days of care. The loss falls to 1.80 and 1.49 days when adjusting for AP- and APR-DRG, respectively, but still represents 19% of the total payment. In contrast, these patients represent a small gain when financing is adjusted for SE factors. Similar patterns are observed for all individual SE characteristics. Effects are much smaller for area-based categories, but their inclusion always reduces penalties (benefits) associated with patients from underprivileged (privileged) areas. Appendix 2 (available online only at: http://www.jhsrp.rsmjournals.com/cgi/content/full/16/4/197/DC1) shows how the adjustment for SE factors alters resource allocation amongst hospitals. We observe that losses (rewards) decrease for hospitals with a lower (larger) proportion of more privileged patients.
Incentives for selection according to individual socioeconomic status (SES) categories: per-patient under-financed (2) and over-financed (+) number of days, and percentage loss (-) or reward (+) as a proportion of the total payment for days of care
Notes: values have been obtained applying model, controlling for APR-DRG and setting coefficients for hospital dummies at the average value.
Privileged status means that the patient does = fall into any of the individual SES categories (welfare recipient, low income, isolated or dependent person)
DRG = diagnosis related group; AP-DRG = all patient DRG; APR-DRG = all patient refined DRG
Discussion
Using a large and recent database of discharges from Belgian hospitals, we observe that SES influences LOS (as a proxy for cost) regardless of patient severity. The impact of individual SES is greater among elderly patients except when income is the main criterion of SES. Area-based SE characteristics are also a predictor of LOS and can be used as a proxy for individual characteristics.
The association between deprivation and longer in-patient stays may have several interpretations. Lower use of primary care services may lead to poorer patients being admitted without a definite diagnosis, often through emergency departments, so that they require more tests and examinations. Poorer access to nursing homes, domiciliary care and informal care are also likely to prolong LOS. Physicians may be more reluctant to let more deprived patients go home and may apply more intensive treatments over a shorter period of time, due to expected lower compliance and more difficult rehabilitation. Lower literacy and communication barriers may create difficulties in the management of the treatment and the recovery process. We also expect elderly people to be more vulnerable to poverty and isolation, related to their poorer health and higher levels of dependency.
Our data also indicate that hospitals with an underprivileged patient profile have longer stays. Hence, SE characteristics may have an aggregate effect that goes beyond the sum of individual ones. Doctors, nurses and administrative staff may devote part of their time to dealing with social and financial problems instead of treating clinical issues, and these financial problems may hinder the capacity to innovate.
Finally, we show that unfair resource allocation is high when the payment system adjusts only for APR-DRGs, but decreases dramatically when APR-DRGs and SE factors are taken into account. The inclusion of SE factors reduces the financial penalty faced by hospitals treating a larger number of underprivileged patients and may discourage hospitals from improving their financial situation through ‘cream-skimming’ thereby improving equity of access to hospital services.
The limitations of our study are, first, the use of administrative data which, although widely used, are always less precise than specific surveys. It may well be that SE factors reflect differences in non-observed clinical characteristics of patients. Further research should investigate the impact of SES on LOS controlling for a wider range of clinical characteristics. Second, the use of area-based indicators to overcome the lack of specific individual data may over-estimate the impact of patient level SES due to contextual effects. People's health and needs, either rich or poor, may be affected by the living conditions in their area of residence. However, such contextual effects also reflect SE conditions that hospitals should not be held responsible for, and so can legitimately be taken into account in an analysis looking at the impact of socioeconomic factors on LOS and cost of treatment.
Footnotes
Acknowledgements
We are much grateful to David Lekeux and Sylvain Baillot for their help and comments at preliminary stages of the study. We would like to thank also the Belgian Health Care Knowledge Center (Federaal Kenniscentrum voor de Gezondheid, KCE) for financial support. Data availability and collaboration from the Intermutualist Agency (IMA) are also gratefully acknowledged.