- Email: [email protected]
Relationship between hospital structural level and length of stay outliers
Health Policy 68 (2004) 159–168
Relationship between hospital structural level and length of stay outliers Implications for hospital payment systems Francesc Cots a,∗ , Lluc Mercadé a , Xavier Castells a , Xavier Salvador b a
Health Services Research Unit, Municipal Institute of Health, IMAS, Hospital del Mar, Passeig Mar´ıtim 25-29 E-08003, Barcelona, Spain b Catalan Health Service, Barcelona, Spain Received 5 March 2003; received in revised form 16 September 2003; accepted 16 September 2003
Abstract Background: Hospital structural level has been suggested as a factor that could explain part of the resource use variation left unexplained by diagnosis related groups (DRGs). However, the relationship between hospital structural level and the presence of cases of extreme resource use (outliers) is not known. Some prospective payment systems pay these cases separately. Objectives: To analyze the relationship between different hospital structural levels, defined according to hospital size, teaching activity and location, and the presence of length of stay (LOS) outliers. Research design: A logit model was used to analyze the patient discharge records of the acute care public hospitals’ Minimum Data Set in Catalonia (Spain) in 1998. The final population contained 631,096 discharges grouped in 329 adjacent DRGs. Measures: LOS outliers were defined as cases with a LOS exceeding the geometric mean plus two standard deviations of all the stays in the same DRG. The 64 public hospitals of the Catalan health system were classified into large urban teaching hospitals, medium-sized teaching and community hospitals, and small community hospitals according to their structural complexity. The model also controlled for patient and health care process characteristics. Results: Outliers accounted for 4.5% of total discharges distributed as follows: large urban teaching hospitals (5.6%), medium-sized teaching and community hospitals (4.6%), small community hospitals (3.6%). The probability of a patient being an outlier was higher in hospitals with greater structural complexity: large urban teaching hospitals (OR = 1.59), medium teaching and community hospitals (OR = 1.30) and small community hospitals (OR = 1). Adjustment through the control variables reduced differences among hospitals: large urban teaching hospitals (OR = 1.32), medium-sized teaching and community hospitals (OR = 1.22), and small community hospitals (OR = 1), but the differences remained significant (P < 0.01). Conclusions: Hospital structural level influences the presence of outliers even when controlling for patient and process characteristics. Thus, some outliers are due to hospital structural level and are not justified by patient characteristics. © 2003 Elsevier Ireland Ltd. All rights reserved. Keywords: Outliers; Hospital structural level; Hospital payment system; Risk adjustment
1. Introduction ∗ Corresponding author. Tel.: +34-93-248-32-89; fax: +34-93-248-32-33. E-mail address: [email protected] (F. Cots).
One of the main features of patient classification systems based on case-mix techniques (for instance
0168-8510/$ – see front matter © 2003 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.healthpol.2003.09.004
160
F. Cots et al. / Health Policy 68 (2004) 159–168
diagnosis related groups (DRGs)) is that, by eliminating values too far from the core of the patient group, they exclude cases of extreme resource use (outliers) and pay these cases separately. Outliers can amount to 4.8% of total hospital discharges, 15.1% of total hospital stays and 17.9% of total cost [1]. There are two reasons for treating outliers separately from other discharges. The first reason is to prevent overvaluation of the final mean value (masking effect) due to the exaggeratedly high cost of a small number of patients. Patient cost distribution is heavily skewed to the right and consequently, as several authors have pointed out, cost function distribution is log normal [2]. The second reason is to [3,4] attempt to reduce the financial risk they represent to providers by paying outliers with an extra payment [5–7]. The statistical relationship associated with each DRG suggests that the factors leading to intra-group variability could also lead to the existence of outliers. Authors such as Calore and Lezzoni [8], Thomas and Ashcraft [9] and Söderlund et al. [10] have described some of the limitations of DRGs to explain cost and length of stay (LOS) variation and have analyzed various severity measurement indicators to improve their explanatory capacity. These authors report that DRGs fail to explain 83, 94.3, and 80.9% of cost variation in untrimmed data and 70, 93.1, and 76.5% of that in trimmed data, respectively. These results indicate that a high proportion of resource use variation remains unexplained; hospital structural complexity has been put forward as one of the factors that might explain variation in patient resource use [11,12]. Differences in structural complexity among hospitals mainly arise because hospitals within a national health system with universal coverage are distributed throughout a territory according to criteria aiming to provide equity of access and to maximize specialization. Within a hospital network, small community hospitals tend to be close to the population they serve. Patients requiring more specialized treatment than that provided by small community hospitals are referred to medium-sized teaching and community hospitals. Finally, large urban teaching hospitals centralize specialties and/or facilities with advanced technology and high economic cost and are used to cover the needs of the entire health care system. These charac-
teristics produce differences in cost among the three types of hospitals. The teaching and research activity of large urban teaching hospitals imply greater structural complexity than that found in small community hospitals. Because these centers are highly specialized, structural costs (i.e. fixed costs associated with the hospital structure) tend to be higher and facilities more expensive than those of other hospitals. The possible relationship between the percentage of outliers and hospital structural level would have clear implications for hospital financing needs. The existence of this relationship remains to be elucidated. Specifically, the question of whether differences in the potential number of outliers between hospitals in different structural levels are due to differences in the complexity of the case-mix treated in the different hospital levels or whether these differences are also due to the structural levels themselves needs to be addressed. If the presence of outliers is due to hospital structural level, independent of the effect produced by patient characteristics, then a separate outlier payment would not be appropriate. This outlier payment would be inappropriate because the difference in cost per case attributable to the hospital characteristics cannot explain such a large cost difference (of a case) in relation to its DRG cost pattern. Thus, the aim of this study was to analyze the relationship between the probability of a patient requiring extreme resource use and the structural level of the hospital in which the patient was treated. To answer this question, all the inpatient hospital discharges from the acute public hospitals of the Catalan Health Service were analyzed. The Catalan health service uses a hybrid hospital financing system: 65% of the payment weights the activity through hospital structure according to a statistical clustering method called grade of membership (GoM) [13], while 35% of the payment weights the activity through DRGs [14]. Thus, the system does not explicitly recognize the existence of outliers. However, since hospitals with higher structural level receive greater payments and since these hospitals may have more outliers, it could be argued that outliers are incorporated into the system as a structural phenomenon grouped together with many other characteristics.
F. Cots et al. / Health Policy 68 (2004) 159–168
2. Methods A total of 631,096 hospital discharges from the Catalan public hospital system in 1998, registered in the Minimum Data Set defined by the Catalan Health Service, and grouped according to DRGs (Health Care Financing Administration version 13) were analyzed. These discharges were the total hospital activity of the public health sector in Catalonia (Spain), which covers 100% of the population of 6 million inhabitants. The activity carried out by the private sector (which accounts for approximately 20% of total hospital activity) was not analyzed since data were unavailable.
161
Each patient was classified into a hospital structural level, according to the type of hospital in which treatment took place. Information on the patients’ area of residence and the location of the hospital in which they were treated was analyzed using three geographic areas concentric to the city of Barcelona (Fig. 1). All the variables used were obtained through, or created from, information from the Minimum Data Set (Table 1). 2.1. Model A model relating outliers to the structural level of the hospital in which they were treated was designed.
Fig. 1. Catalonia: the three geographic areas.
162
F. Cots et al. / Health Policy 68 (2004) 159–168
Table 1 Model variables Variable
Values
Valid casesa
Meana
S.D.a
Outliers
0 = Not outlier case 1 = Outlier case
631,096
0.04
0.21
Hospital structural level
1 = Small community hospitals 2 = Medium-sized teaching and community hospitals 3 = Large urban teaching hospitals
631,096
1.87
0.77
Comorbidities
Logarithm (Elixhauser Index)
631,096
0.24
0.47
Complications
0 = No complications 1 = Complications
631,096
0.09
0.28
Complexity A-DRG weightb
0 = Low (weight lower than 0.548) 1 = Medium (weight from 0.548 to 1.002) 2 = High (weight higher than 1.002)
631,096
1
0.7
Age
1 2 3 4 5 6
= Age = Age = Age = Age = Age = Age
630,851
2.94
1.2
Distance
1 2 3 4
= No distance = Medium distance = Long distance = Place of residence unknown
624,398
1.16
0.42
Re-adimssions
0 = No re-admission(s) 1 = Index admission 2 = Re-admission(s)
631,096
0.37
0.71
Admission and DRG type
0 1 2 3 4
627,224
1.66
1.33
Death
0 = No death 1 = Death 2 = Discharge status unknown
629,554
0.03
0.17
a b
from 0 to 17 years from 18 to 45 years from 46 to 65 years from 66 to 80 years greater than 80 years unknown
= Planned and surgical = Planned and non-surgical = Emergency and surgical = Emergency and non-surgical = Circumstance on admission unknown
Missing values not included. A-DRG, adjacent diagnosis related group.
This model adjusted for variables that a priori could explain and justify extreme intra-DRG variability, namely comorbidities, complications, DRG complexity, age, distance between patients’ place of residence and the hospital, re-admissions, whether admission was planned or through the emergency department, whether DRG type was medical or surgical, and inpatient death [15] (Table 1).
2.2. Design of variables 2.2.1. Outliers To determine the response variable, that is, whether a patient was an outlier or not, a threshold for each DRG was calculated. Because cost information was unavailable, LOS was used as resource use approximation. The LOS of patients falling outside this cut-off
F. Cots et al. / Health Policy 68 (2004) 159–168
point was considered abnormally long. The threshold was calculated by adding two standard deviations to the geometric mean of all stays within a DRG [16]. 2.2.2. Hospital structural level Corresponding to the official classification used in the Catalan health system, 64 public hospitals were classified into large urban teaching hospitals (6), medium-sized teaching and community hospitals (17), and small community hospitals (41). Firstly, large urban teaching hospitals include teaching hospitals with more than 400 beds and advanced technology that carry out highly specialized research. These hospitals serve the entire health system and treat patients with complex conditions who cannot be adequately treated in the medium-sized teaching and community hospitals. All the hospitals of this level are located in the Barcelona area. The second level is composed by medium teaching or community hospitals, which have between 200 and 400 beds and are able to treat most medical conditions. These hospitals dispose of considerable technology and most carry out some teaching activity. Finally, small community hospitals with less than 200 beds serve the needs of a specific neighborhood or community. Services in these hospitals are less intensive than those in other hospitals. 2.2.3. DRG complexity This variable was introduced into the model through adjacent-DRGs (A-DRGs) classified in three groups (low, medium and high complexity) according to their cost-weight. We used the original DRGs of the Health Care Financing Administration’s version 13 but regrouped the DRGs that identified complications and/or comorbidities and patient age. This modification reduced the 492 potential DRGs to 331 A-DRGs. The low complexity group included discharges with a weight lower than the first quartile of the 631,096 discharges, the medium complexity group comprised the discharges with a weight between the first and third quartile, and the high complexity group included the discharges with a weight higher than the third quartile. 2.2.4. Comorbidities This variable identified patients’ clinical characteristics other than the principal reason for admission that could influence LOS or costs. To capture this effect, an indicator following the grouping proposed by Elix-
163
hauser et al. [17] and using the secondary diagnoses available in the MDS was created. 2.2.5. Complications Diagnoses considered complications according to Elixhauser’s criteria [17] together with general complications were used to create a dummy variable that identified complications. 2.2.6. Age To analyze the effect of age, patient age was divided into five groups. 2.2.7. Distance This variable incorporated the distance between patients’ place of residence and the hospital and differentiated between patients residing in the second and third geographic area who traveled to Barcelona for treatment from the remaining patients. The distance (in a straight line) from the second region was a maximum of 100 km, and that from the third region was a maximum of 200 km. 2.2.8. Re-admissions Hospital admissions taking place within 90 days of a previous discharge were considered to be re-admissions. The 90-day period was chosen in order not to exclude possible re-admissions. This variable also identified the first admission in a series, which we call the index admission [15]. 2.2.9. Admission and DRG type This variable was created by combining information from the MDS on type of admission and DRG. Admissions were classified into emergency and planned. DRG type was identified through the definition of DRG type (medical, surgical or indeterminate) and was divided into surgical and non-surgical. 2.2.10. Death This variable was created by using information from the MDS on discharge status and identified inpatient deaths. 2.3. Analysis To determine the relationship between hospital structural level and the presence of outliers, as well
164
F. Cots et al. / Health Policy 68 (2004) 159–168
Table 2 Main characteristics of the population studied according to hospital structural level Hospital structural level
Admissions (n) Outliers (%) Length of stay, mean (S.D.) Days associated with outliers (%) Comorbidities, mean (S.D.) Complications (%) Complexity, A-DRG weighta , mean (S.D.) Age, mean (S.D.)
Total
Large urban teaching hospitals
Medium-sized teaching and community hospitals
Small community hospitals
152,526 5.6 8.67 (12.95) 23.83 0.58 (1.21) 13.6 1.16 (1.43) 51.50 (24.04)
244,389 4.6 6.94 (9.31) 20.10 0.44 (1.04) 8.0 0.88 (0.79) 48.55 (26.84)
234,181 3.6 5.83 (7.58) 16.86 0.42 (0.97) 6.3 0.79 (0.50) 51.96 (24.68)
631,096 4.5 6.95 (9.83) 20.22 0.47 (1.06) 8.7 0.91 (0.92) 50.53 (25.44)
Distanceb (%) No distance Medium distance Long distance
66.2 27.2 6.3
86.8 11.1 1.2
95.0 3.1 0.2
84.9 12.0 2.1
Re-admissions (%) No re-admission(s) Index re-admission Re-admission(s)
73.1 10.8 16.1
76.2 9.9 14.0
79.3 8.7 12.0
76.6 9.6 13.8
Admission and DRG typeb (%) Planned and surgical Planned and non-surgical Emergency and surgical Emergency and non-surgical Deathb (%)
31.1 19.5 11.4 37.6 4.0
30.7 11.0 10.4 47.4 3.1
36.4 9.3 8.0 45.3 2.6
32.9 12.4 9.8 44.3 3.1
a b
A-DRG, adjacent diagnosis related group. Missing values not shown.
as the extent to which potential differences among hospitals were due to the different levels themselves or to patient characteristics within each level, the effect of each variable was estimated through a logistic regression model and was measured through odds ratios (ORs). The analysis was divided into two parts: firstly, the ORs of each variable were estimated separately in a simple logistic regression model (non-adjusted OR) and secondly, the ORs of the same variables were estimated in a multiple logistic regression model including all the variables (adjusted OR).
3. Results Of the 631,096 discharges, 28,234 were identified as outliers, representing 4.5% of the total. Outliers accounted for 20.2% of total days of stay with a mean
LOS of 31.4 days, which was five times longer than that of inliers. Outliers accounted for 5.6% of discharges from large urban teaching hospitals, 4.6% of those from medium-sized teaching and community hospitals, and 3.6% of those from small community hospitals (Table 2). The non-adjusted ORs revealed that structural level directly influenced the probability of a patient being an outlier: the higher the structural level of the hospital, the greater the probability of outliers (Table 3). The small community hospitals were chosen as the reference category with an OR equal to 1. The medium-sized teaching and community hospitals showed an OR equal to 1.3, and large urban teaching hospitals had an OR equal to 1.59 (P < 0.01). The control variables in the non-adjusted model revealed that patients had a higher probability of being outliers when the comorbidities burden, A-DRG complexity, age, and distance between the hospital and the
F. Cots et al. / Health Policy 68 (2004) 159–168
165
Table 3 Logistic Regression Model: estimated probability of a patient being an outlier according to hospital structural level and to control variables n Hospital structural level Small community hospitals Medium-sized teaching and community hospitals Large urban teaching hospitals
234,181 244,389 152,526
Outliers (%) 3.6 4.6 5.6
Comorbidities
Non-adjusted ORa,b
Adjusted ORa,b
∗
∗
1 1.30∗ 1.59∗
1 1.22∗ 1.32∗
1.86∗
1.50∗ 1 1.89∗
Complications No Yes
575,933 55,163
4.0 9.0
1 2.35∗
A-DRGc complexity Low Medium High
∗
∗
156,645 319,240 155,211
3.8 4.6 5.0
1 1.21∗ 1.33∗
1 1.03 0.74∗
Age 0–17 years 18–45 years 46–65 years 66–80 years 80+ years
∗
∗
80,217 170,164 150,077 168,495 61,898
2.5 3.2 4.5 6.0 6.3
1 1.28∗ 1.81∗ 2.47∗ 2.62∗
1 1.21∗ 1.68∗ 2.00∗ 1.83∗
Distance No distance Medium distance Long distance
∗
∗
535,537 75,825 13,036
4.3 5.4 6.1
1 1.27∗ 1.42∗
1 1.34∗ 1.66∗
∗
∗
1 1.52∗ 1.26∗
Re-admissions No re-admission(s) Index admission Re-admission(s)
483,575 60,794 86,727
3.9 6.8 6.0
1 1.78∗ 1.57∗
Admission and DRG type Planned and surgical Planned and non surgical Emergency and surgical Emergency and non-surgical
∗
∗
207,729 78,466 61,643 279,386
2.5 3.9 8.3 5.3
1 1.56∗ 3.49∗ 2.15∗
1 1.08∗ 3.38∗ 1.72∗
∗
∗
1 2.62∗
1 1.73∗
Death No Yes a b c ∗
609,811 19,743
4.3 10.5
OR, odds ratio. Missing values included in the analysis but not shown. A-DRG, adjacent diagnosis related group. P < 0.01.
place of residence were increased. Increased probabilities also depended on the presence of complications, whether an admission was an index admission or a readmission, whether the admission was an emergency, and whether death occurred during hospitalization. The relationship between structural level and outliers was maintained when the model was adjusted but in both large urban teaching hospitals and
medium-sized teaching and community hospitals the estimated effects in the adjusted model were lower than those in the non-adjusted model. In the latter, the estimated OR for medium-sized teaching and community hospitals changed to 1.22 and that for large urban teaching hospitals to 1.32 (Table 3). Almost all of the remaining adjusted variables except that of distance showed a considerably lower
166
F. Cots et al. / Health Policy 68 (2004) 159–168
OR than in the non-adjusted model. In the variable of A-DRG complexity, the OR of medium complexity was significantly lower than in the non-adjusted model and the OR of high complexity changed sign. In the variables of admission and DRG type, the OR of non-surgical DRGs also showed a considerable decrease, as did that of the variable of death.
4. Discussion The results of this study demonstrate that a relationship exists between hospital structural level and LOS outliers and confirm the initial hypothesis that the percentage of outliers in public hospitals increases with hospital structural level. However, the influence of the control variables on the probability of a patient being an outlier also confirms that patient and health care characteristics should also be taken into account when analyzing the causes of differences in the presence of outliers among hospital structural levels. If these characteristics are not randomly distributed between the different hospital structural levels, then differences in the presence of outliers among hospital structural levels may correspond to differences in the patient or health care characteristics of each structural level. Possibly, there is an effect according to which hospitals in different structural levels attract patients with different probabilities of being outliers. In fact, our database confirms that patient complexity (measured by DRG) and severity (valued through several variables) are related to the structural level. The multiple model, which adjusts for the effect of hospital structural level through the control variables, can be used to determine the extent to which differences in the probability of a case being an outlier among hospital levels are due to the level itself or are due to characteristics of the patients or health care process. The difference in the OR of hospital structural complexity between the non-adjusted and the adjusted model gives the value of the probability of a patient in a hospital structural level being an outlier that is related to the characteristics of the patient or health care process. Indeed, the decreased probability associated with the different hospital structural levels in the adjusted model compared with the non-adjusted model indicates that hospitals with a higher structural level tend to treat a greater number of patients who, due to their
characteristics, have an increased probability of being outliers. However, when these patient and health care process characteristics were analyzed in the adjusted model, the differences in probabilities remained clearly significant. Thus, factors intrinsic to hospital structural level itself explain these remaining differences. Possibly, some residual case-mix differences were not captured through the control variables used, and thus some of the differences found might not be due to hospital structural level itself. However, the adjusted model takes into account a large number of factors, beyond the standard patient classification systems, such as the comorbidities indicator, recently developed by Elixhauser, and those variables than can be created from the administrative data available in a minimum data set. Thus, although a small part of outlier differences among structural levels cannot be well explained by the case-mix variables used, we believe that an adjusted relationship between structural hospital level and percentage of outliers has been demonstrated. The teaching and research activities normally carried out in hospitals with higher structural complexity might justify this remaining difference in the percentage of outliers since the cost associated with these activities implies greater hospital cost. However, teaching and research activity justify certain extra resource use, but do not justify differences in the number of outliers. In particular, these activities justify a more dispersed DRG resource use distribution due to the differences associated with the hospital type, which would lead to a higher cut-off point and consequently fewer cases would be detected as outliers. Concerning the behavior of the remaining variables in the adjusted model, the variables explaining patient seriousness (comorbidities, complications, age) within DRGs, show a very clear relationship with the probability of a patient being an outlier. These variables, together with the case-mix, incorporated into the analysis through the DRG-weight, could have potentially explained and justified the differences between levels and, thus, might have contradicted the original hypothesis. However, this was not the case. Although these variables showed considerable explanatory power, they did not explain the causes that, after adjustment, were clearly related to structural complexity. This was also true of the variables defining the health care process. Emergency admissions, re-admissions, inpatient deaths and the distance
F. Cots et al. / Health Policy 68 (2004) 159–168
between patients’ place of residence and the hospital added explanatory power, but did not refute the initial hypothesis. In the category of high complexity DRGs, the change in sign from the non-adjusted to the adjusted model shows that outliers in this category were also more frequently characterized by other model variables associated with a higher probability of being an outlier than were the outliers in lower complexity DRGs. A model without this variable was tested (results not shown) and no substantial change in any of the variables was observed. One limitation of this study is that variables related to the characteristics of hospital structural level, such as hospital size, and the patient-clinician ratio, among others, were not included in the analysis. These variables might have identified the structural factors that explain differences in the probability of patients being outliers. The limitations of DRGs as a patient classification system could also have marginally influenced the results. Underlying intra-DRG variability may not be homogeneously distributed between the different structural levels, at least there is no evidence that it is. However, because DRGs are almost universally used as a patient classification system, this limitation is insuperable. Finally, because socio-economic information from the hospital discharge registers is lacking, the effect of socio-economic status on the probability of a patient being an outlier could not be estimated. Socio-economic status might be another factor that could justify differences in the presence of outliers among different hospital structural levels. Some authors [18,19] have proposed that the absence of this information could be overcome by using census data on patients’ residence. A preliminary result of the present study using a socio-economic variable related to the patients’ city or town of residence indicated that an ecological approach did not provide a solution to the problem of absence of individual socio-economic information. Although it cannot be inferred that there are significant differences in efficiency among different hospital structural levels in overall production function, an actual difference in the number of outliers was detected. Bearing in mind that the adjusting variables largely captured the case-mix treated in each
167
hospital structural level, many outlier cases were associated with hospital structural level itself and were not a consequence of the complexity of the case-mix treated in each level. Thus, the implication of this result for health policy, and more specifically for the financing and management of public hospitals, is that the presence of outliers influences final hospital cost and consequently financing needs. The outliers in a given hospital can account for approximately 4.5% (from 3.6 to 5.6%, depending on the structural level) of all discharges and can represent 20.2% (from 16.9 to 23.8%, depending on the structural level) of the total cost. After adjusting for severity, there seems to be no reason why large urban teaching hospitals and medium-sized teaching and community hospitals should have greater numbers of outliers than small community hospitals. Therefore, payment of extremely expensive cases could be adjusted according to the mean incidence of these cases within the whole system after adjusting for severity. Otherwise, an insurance effect would be produced, which would offer incentives to hospitals with high structural levels to have a greater number of unjustified outliers, since there would be no clear incentives not to have them.
Acknowledgements This study was supported by the ‘Fundación de Investigaciones Sanitarias-Instituto de Salud Carlos III-Ministerio de Sanidad y Consumo’, project number 99/0687.
References [1] Cots F, Elvira D, Castells X, Dalmau E. Medicare’s DRGWeights in a European environment: the Spanish experience. Health Policy 2000;51:31–47. [2] Lichtig LK. Hospital information systems for casemix management. New York: Wiley, 1986. [3] Briggs A, Gray A. The distribution of health care costs and their statistical analysis for economic evaluation. Journal of Health Services Research and Policy 1998;3:233–45. [4] Diehr P, Yanez D, Ash A, Hornbrook M, Lin DY. Methods for analyzing health care utilization and costs. Annual Reviews of Public Health 1999;20:125–44. [5] Keeler EB, Carter GM, Trude S. Insurance aspects of DRG outlier payments. Journal of Health Economics 1988;7:193– 214.
168
F. Cots et al. / Health Policy 68 (2004) 159–168
[6] Carter GM, Rumpel JD. Payment rates for unusual Medicare hospital cases. Santa Mónica: RAND, 1992. [7] Carter GM, Farley DO. Assenting the FY 1989 change in Medicare PPS outlier policy. Health Care Financing Review 1992;14:69–82. [8] Calore KA, Iezzoni L. Disease staging and PMCs. Can they improve DRGs? Medical Care 1987;25:724–37. [9] Thomas JW, Ashcraft MLF. Measuring severity of illness: six severity systems and their ability to explain cost variations. Inquiry 1991;28:39–55. [10] Söderlund N, Gray A, Milne R, Raftery J. Case mix measurement in english hospitals: an evaluation of five methods for predicting resource use. Journal of Health Services Research and Policy 1996;1:10–9. [11] Finlayson G, Roos NP, Jacobs P, Watson D. Using the Manitoba hospital management information system: comparing average cost per weighted case and financial ratios of Manitoba hospitals. MCHP2001. Manitoba: Manitoba centre for health policy and evaliation, 2001. p. 1–149. [12] López Casasnovas G, Saez M. The impact of teaching status on average costs in Spanish hospitals. Health Economics 1999;8:641–51. [13] Ibern P, Vertrees JC, Manton KG, Woodbury MA. Hospital groups and case-mix measurement for resource allocation and
[14]
[15]
[16]
[17]
[18]
[19]
payment. In: Diagnosis related groups in Europe. New York: Springer-Verlag; 1993. Cots F, Castells X. Cómo pagamos a nuestros hospitales. La referencia de Cataluña y el contrapunto desde Andaluc´ıa. Gaceta Sanitaria 2001;15:172–81. Cots F, Mercadé L, Castells X, Torre P, Riu M. Risk adjustment: beyond patient’s classification systems. Gaceta Sanitaria 2001;15:423–31. Cots F, Castells X, Elvira D, Saez M. Relevance of outlier cases in case mix systems and evaluation of trimming methods. Health Care Management Science 2003;6:27– 35. Elixhauser A, Steiner C, Harris R, Coffey RM. Comorbidity measures for use with administrative data. Medical Care 1998;36:8–27. Hofer TP, Wolfe RA, Tedeschi PJ, McMahon LF, Griffith JR. Use of community versus individual socioeconomic data in predicting variation in hospital use. Health Service Research 1998;33:243–59. Krieger N. Overcoming the absence of socioeconomic data in medical records: validation an application of a censusbased methodology. American Journal of Public Health 1992;92:703–10.
Relationship between hospital structural level and length of stay outliers Implications for hospital payment systems Francesc Cots a,∗ , Lluc Mercadé a , Xavier Castells a , Xavier Salvador b a
Health Services Research Unit, Municipal Institute of Health, IMAS, Hospital del Mar, Passeig Mar´ıtim 25-29 E-08003, Barcelona, Spain b Catalan Health Service, Barcelona, Spain Received 5 March 2003; received in revised form 16 September 2003; accepted 16 September 2003
Abstract Background: Hospital structural level has been suggested as a factor that could explain part of the resource use variation left unexplained by diagnosis related groups (DRGs). However, the relationship between hospital structural level and the presence of cases of extreme resource use (outliers) is not known. Some prospective payment systems pay these cases separately. Objectives: To analyze the relationship between different hospital structural levels, defined according to hospital size, teaching activity and location, and the presence of length of stay (LOS) outliers. Research design: A logit model was used to analyze the patient discharge records of the acute care public hospitals’ Minimum Data Set in Catalonia (Spain) in 1998. The final population contained 631,096 discharges grouped in 329 adjacent DRGs. Measures: LOS outliers were defined as cases with a LOS exceeding the geometric mean plus two standard deviations of all the stays in the same DRG. The 64 public hospitals of the Catalan health system were classified into large urban teaching hospitals, medium-sized teaching and community hospitals, and small community hospitals according to their structural complexity. The model also controlled for patient and health care process characteristics. Results: Outliers accounted for 4.5% of total discharges distributed as follows: large urban teaching hospitals (5.6%), medium-sized teaching and community hospitals (4.6%), small community hospitals (3.6%). The probability of a patient being an outlier was higher in hospitals with greater structural complexity: large urban teaching hospitals (OR = 1.59), medium teaching and community hospitals (OR = 1.30) and small community hospitals (OR = 1). Adjustment through the control variables reduced differences among hospitals: large urban teaching hospitals (OR = 1.32), medium-sized teaching and community hospitals (OR = 1.22), and small community hospitals (OR = 1), but the differences remained significant (P < 0.01). Conclusions: Hospital structural level influences the presence of outliers even when controlling for patient and process characteristics. Thus, some outliers are due to hospital structural level and are not justified by patient characteristics. © 2003 Elsevier Ireland Ltd. All rights reserved. Keywords: Outliers; Hospital structural level; Hospital payment system; Risk adjustment
1. Introduction ∗ Corresponding author. Tel.: +34-93-248-32-89; fax: +34-93-248-32-33. E-mail address: [email protected] (F. Cots).
One of the main features of patient classification systems based on case-mix techniques (for instance
0168-8510/$ – see front matter © 2003 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.healthpol.2003.09.004
160
F. Cots et al. / Health Policy 68 (2004) 159–168
diagnosis related groups (DRGs)) is that, by eliminating values too far from the core of the patient group, they exclude cases of extreme resource use (outliers) and pay these cases separately. Outliers can amount to 4.8% of total hospital discharges, 15.1% of total hospital stays and 17.9% of total cost [1]. There are two reasons for treating outliers separately from other discharges. The first reason is to prevent overvaluation of the final mean value (masking effect) due to the exaggeratedly high cost of a small number of patients. Patient cost distribution is heavily skewed to the right and consequently, as several authors have pointed out, cost function distribution is log normal [2]. The second reason is to [3,4] attempt to reduce the financial risk they represent to providers by paying outliers with an extra payment [5–7]. The statistical relationship associated with each DRG suggests that the factors leading to intra-group variability could also lead to the existence of outliers. Authors such as Calore and Lezzoni [8], Thomas and Ashcraft [9] and Söderlund et al. [10] have described some of the limitations of DRGs to explain cost and length of stay (LOS) variation and have analyzed various severity measurement indicators to improve their explanatory capacity. These authors report that DRGs fail to explain 83, 94.3, and 80.9% of cost variation in untrimmed data and 70, 93.1, and 76.5% of that in trimmed data, respectively. These results indicate that a high proportion of resource use variation remains unexplained; hospital structural complexity has been put forward as one of the factors that might explain variation in patient resource use [11,12]. Differences in structural complexity among hospitals mainly arise because hospitals within a national health system with universal coverage are distributed throughout a territory according to criteria aiming to provide equity of access and to maximize specialization. Within a hospital network, small community hospitals tend to be close to the population they serve. Patients requiring more specialized treatment than that provided by small community hospitals are referred to medium-sized teaching and community hospitals. Finally, large urban teaching hospitals centralize specialties and/or facilities with advanced technology and high economic cost and are used to cover the needs of the entire health care system. These charac-
teristics produce differences in cost among the three types of hospitals. The teaching and research activity of large urban teaching hospitals imply greater structural complexity than that found in small community hospitals. Because these centers are highly specialized, structural costs (i.e. fixed costs associated with the hospital structure) tend to be higher and facilities more expensive than those of other hospitals. The possible relationship between the percentage of outliers and hospital structural level would have clear implications for hospital financing needs. The existence of this relationship remains to be elucidated. Specifically, the question of whether differences in the potential number of outliers between hospitals in different structural levels are due to differences in the complexity of the case-mix treated in the different hospital levels or whether these differences are also due to the structural levels themselves needs to be addressed. If the presence of outliers is due to hospital structural level, independent of the effect produced by patient characteristics, then a separate outlier payment would not be appropriate. This outlier payment would be inappropriate because the difference in cost per case attributable to the hospital characteristics cannot explain such a large cost difference (of a case) in relation to its DRG cost pattern. Thus, the aim of this study was to analyze the relationship between the probability of a patient requiring extreme resource use and the structural level of the hospital in which the patient was treated. To answer this question, all the inpatient hospital discharges from the acute public hospitals of the Catalan Health Service were analyzed. The Catalan health service uses a hybrid hospital financing system: 65% of the payment weights the activity through hospital structure according to a statistical clustering method called grade of membership (GoM) [13], while 35% of the payment weights the activity through DRGs [14]. Thus, the system does not explicitly recognize the existence of outliers. However, since hospitals with higher structural level receive greater payments and since these hospitals may have more outliers, it could be argued that outliers are incorporated into the system as a structural phenomenon grouped together with many other characteristics.
F. Cots et al. / Health Policy 68 (2004) 159–168
2. Methods A total of 631,096 hospital discharges from the Catalan public hospital system in 1998, registered in the Minimum Data Set defined by the Catalan Health Service, and grouped according to DRGs (Health Care Financing Administration version 13) were analyzed. These discharges were the total hospital activity of the public health sector in Catalonia (Spain), which covers 100% of the population of 6 million inhabitants. The activity carried out by the private sector (which accounts for approximately 20% of total hospital activity) was not analyzed since data were unavailable.
161
Each patient was classified into a hospital structural level, according to the type of hospital in which treatment took place. Information on the patients’ area of residence and the location of the hospital in which they were treated was analyzed using three geographic areas concentric to the city of Barcelona (Fig. 1). All the variables used were obtained through, or created from, information from the Minimum Data Set (Table 1). 2.1. Model A model relating outliers to the structural level of the hospital in which they were treated was designed.
Fig. 1. Catalonia: the three geographic areas.
162
F. Cots et al. / Health Policy 68 (2004) 159–168
Table 1 Model variables Variable
Values
Valid casesa
Meana
S.D.a
Outliers
0 = Not outlier case 1 = Outlier case
631,096
0.04
0.21
Hospital structural level
1 = Small community hospitals 2 = Medium-sized teaching and community hospitals 3 = Large urban teaching hospitals
631,096
1.87
0.77
Comorbidities
Logarithm (Elixhauser Index)
631,096
0.24
0.47
Complications
0 = No complications 1 = Complications
631,096
0.09
0.28
Complexity A-DRG weightb
0 = Low (weight lower than 0.548) 1 = Medium (weight from 0.548 to 1.002) 2 = High (weight higher than 1.002)
631,096
1
0.7
Age
1 2 3 4 5 6
= Age = Age = Age = Age = Age = Age
630,851
2.94
1.2
Distance
1 2 3 4
= No distance = Medium distance = Long distance = Place of residence unknown
624,398
1.16
0.42
Re-adimssions
0 = No re-admission(s) 1 = Index admission 2 = Re-admission(s)
631,096
0.37
0.71
Admission and DRG type
0 1 2 3 4
627,224
1.66
1.33
Death
0 = No death 1 = Death 2 = Discharge status unknown
629,554
0.03
0.17
a b
from 0 to 17 years from 18 to 45 years from 46 to 65 years from 66 to 80 years greater than 80 years unknown
= Planned and surgical = Planned and non-surgical = Emergency and surgical = Emergency and non-surgical = Circumstance on admission unknown
Missing values not included. A-DRG, adjacent diagnosis related group.
This model adjusted for variables that a priori could explain and justify extreme intra-DRG variability, namely comorbidities, complications, DRG complexity, age, distance between patients’ place of residence and the hospital, re-admissions, whether admission was planned or through the emergency department, whether DRG type was medical or surgical, and inpatient death [15] (Table 1).
2.2. Design of variables 2.2.1. Outliers To determine the response variable, that is, whether a patient was an outlier or not, a threshold for each DRG was calculated. Because cost information was unavailable, LOS was used as resource use approximation. The LOS of patients falling outside this cut-off
F. Cots et al. / Health Policy 68 (2004) 159–168
point was considered abnormally long. The threshold was calculated by adding two standard deviations to the geometric mean of all stays within a DRG [16]. 2.2.2. Hospital structural level Corresponding to the official classification used in the Catalan health system, 64 public hospitals were classified into large urban teaching hospitals (6), medium-sized teaching and community hospitals (17), and small community hospitals (41). Firstly, large urban teaching hospitals include teaching hospitals with more than 400 beds and advanced technology that carry out highly specialized research. These hospitals serve the entire health system and treat patients with complex conditions who cannot be adequately treated in the medium-sized teaching and community hospitals. All the hospitals of this level are located in the Barcelona area. The second level is composed by medium teaching or community hospitals, which have between 200 and 400 beds and are able to treat most medical conditions. These hospitals dispose of considerable technology and most carry out some teaching activity. Finally, small community hospitals with less than 200 beds serve the needs of a specific neighborhood or community. Services in these hospitals are less intensive than those in other hospitals. 2.2.3. DRG complexity This variable was introduced into the model through adjacent-DRGs (A-DRGs) classified in three groups (low, medium and high complexity) according to their cost-weight. We used the original DRGs of the Health Care Financing Administration’s version 13 but regrouped the DRGs that identified complications and/or comorbidities and patient age. This modification reduced the 492 potential DRGs to 331 A-DRGs. The low complexity group included discharges with a weight lower than the first quartile of the 631,096 discharges, the medium complexity group comprised the discharges with a weight between the first and third quartile, and the high complexity group included the discharges with a weight higher than the third quartile. 2.2.4. Comorbidities This variable identified patients’ clinical characteristics other than the principal reason for admission that could influence LOS or costs. To capture this effect, an indicator following the grouping proposed by Elix-
163
hauser et al. [17] and using the secondary diagnoses available in the MDS was created. 2.2.5. Complications Diagnoses considered complications according to Elixhauser’s criteria [17] together with general complications were used to create a dummy variable that identified complications. 2.2.6. Age To analyze the effect of age, patient age was divided into five groups. 2.2.7. Distance This variable incorporated the distance between patients’ place of residence and the hospital and differentiated between patients residing in the second and third geographic area who traveled to Barcelona for treatment from the remaining patients. The distance (in a straight line) from the second region was a maximum of 100 km, and that from the third region was a maximum of 200 km. 2.2.8. Re-admissions Hospital admissions taking place within 90 days of a previous discharge were considered to be re-admissions. The 90-day period was chosen in order not to exclude possible re-admissions. This variable also identified the first admission in a series, which we call the index admission [15]. 2.2.9. Admission and DRG type This variable was created by combining information from the MDS on type of admission and DRG. Admissions were classified into emergency and planned. DRG type was identified through the definition of DRG type (medical, surgical or indeterminate) and was divided into surgical and non-surgical. 2.2.10. Death This variable was created by using information from the MDS on discharge status and identified inpatient deaths. 2.3. Analysis To determine the relationship between hospital structural level and the presence of outliers, as well
164
F. Cots et al. / Health Policy 68 (2004) 159–168
Table 2 Main characteristics of the population studied according to hospital structural level Hospital structural level
Admissions (n) Outliers (%) Length of stay, mean (S.D.) Days associated with outliers (%) Comorbidities, mean (S.D.) Complications (%) Complexity, A-DRG weighta , mean (S.D.) Age, mean (S.D.)
Total
Large urban teaching hospitals
Medium-sized teaching and community hospitals
Small community hospitals
152,526 5.6 8.67 (12.95) 23.83 0.58 (1.21) 13.6 1.16 (1.43) 51.50 (24.04)
244,389 4.6 6.94 (9.31) 20.10 0.44 (1.04) 8.0 0.88 (0.79) 48.55 (26.84)
234,181 3.6 5.83 (7.58) 16.86 0.42 (0.97) 6.3 0.79 (0.50) 51.96 (24.68)
631,096 4.5 6.95 (9.83) 20.22 0.47 (1.06) 8.7 0.91 (0.92) 50.53 (25.44)
Distanceb (%) No distance Medium distance Long distance
66.2 27.2 6.3
86.8 11.1 1.2
95.0 3.1 0.2
84.9 12.0 2.1
Re-admissions (%) No re-admission(s) Index re-admission Re-admission(s)
73.1 10.8 16.1
76.2 9.9 14.0
79.3 8.7 12.0
76.6 9.6 13.8
Admission and DRG typeb (%) Planned and surgical Planned and non-surgical Emergency and surgical Emergency and non-surgical Deathb (%)
31.1 19.5 11.4 37.6 4.0
30.7 11.0 10.4 47.4 3.1
36.4 9.3 8.0 45.3 2.6
32.9 12.4 9.8 44.3 3.1
a b
A-DRG, adjacent diagnosis related group. Missing values not shown.
as the extent to which potential differences among hospitals were due to the different levels themselves or to patient characteristics within each level, the effect of each variable was estimated through a logistic regression model and was measured through odds ratios (ORs). The analysis was divided into two parts: firstly, the ORs of each variable were estimated separately in a simple logistic regression model (non-adjusted OR) and secondly, the ORs of the same variables were estimated in a multiple logistic regression model including all the variables (adjusted OR).
3. Results Of the 631,096 discharges, 28,234 were identified as outliers, representing 4.5% of the total. Outliers accounted for 20.2% of total days of stay with a mean
LOS of 31.4 days, which was five times longer than that of inliers. Outliers accounted for 5.6% of discharges from large urban teaching hospitals, 4.6% of those from medium-sized teaching and community hospitals, and 3.6% of those from small community hospitals (Table 2). The non-adjusted ORs revealed that structural level directly influenced the probability of a patient being an outlier: the higher the structural level of the hospital, the greater the probability of outliers (Table 3). The small community hospitals were chosen as the reference category with an OR equal to 1. The medium-sized teaching and community hospitals showed an OR equal to 1.3, and large urban teaching hospitals had an OR equal to 1.59 (P < 0.01). The control variables in the non-adjusted model revealed that patients had a higher probability of being outliers when the comorbidities burden, A-DRG complexity, age, and distance between the hospital and the
F. Cots et al. / Health Policy 68 (2004) 159–168
165
Table 3 Logistic Regression Model: estimated probability of a patient being an outlier according to hospital structural level and to control variables n Hospital structural level Small community hospitals Medium-sized teaching and community hospitals Large urban teaching hospitals
234,181 244,389 152,526
Outliers (%) 3.6 4.6 5.6
Comorbidities
Non-adjusted ORa,b
Adjusted ORa,b
∗
∗
1 1.30∗ 1.59∗
1 1.22∗ 1.32∗
1.86∗
1.50∗ 1 1.89∗
Complications No Yes
575,933 55,163
4.0 9.0
1 2.35∗
A-DRGc complexity Low Medium High
∗
∗
156,645 319,240 155,211
3.8 4.6 5.0
1 1.21∗ 1.33∗
1 1.03 0.74∗
Age 0–17 years 18–45 years 46–65 years 66–80 years 80+ years
∗
∗
80,217 170,164 150,077 168,495 61,898
2.5 3.2 4.5 6.0 6.3
1 1.28∗ 1.81∗ 2.47∗ 2.62∗
1 1.21∗ 1.68∗ 2.00∗ 1.83∗
Distance No distance Medium distance Long distance
∗
∗
535,537 75,825 13,036
4.3 5.4 6.1
1 1.27∗ 1.42∗
1 1.34∗ 1.66∗
∗
∗
1 1.52∗ 1.26∗
Re-admissions No re-admission(s) Index admission Re-admission(s)
483,575 60,794 86,727
3.9 6.8 6.0
1 1.78∗ 1.57∗
Admission and DRG type Planned and surgical Planned and non surgical Emergency and surgical Emergency and non-surgical
∗
∗
207,729 78,466 61,643 279,386
2.5 3.9 8.3 5.3
1 1.56∗ 3.49∗ 2.15∗
1 1.08∗ 3.38∗ 1.72∗
∗
∗
1 2.62∗
1 1.73∗
Death No Yes a b c ∗
609,811 19,743
4.3 10.5
OR, odds ratio. Missing values included in the analysis but not shown. A-DRG, adjacent diagnosis related group. P < 0.01.
place of residence were increased. Increased probabilities also depended on the presence of complications, whether an admission was an index admission or a readmission, whether the admission was an emergency, and whether death occurred during hospitalization. The relationship between structural level and outliers was maintained when the model was adjusted but in both large urban teaching hospitals and
medium-sized teaching and community hospitals the estimated effects in the adjusted model were lower than those in the non-adjusted model. In the latter, the estimated OR for medium-sized teaching and community hospitals changed to 1.22 and that for large urban teaching hospitals to 1.32 (Table 3). Almost all of the remaining adjusted variables except that of distance showed a considerably lower
166
F. Cots et al. / Health Policy 68 (2004) 159–168
OR than in the non-adjusted model. In the variable of A-DRG complexity, the OR of medium complexity was significantly lower than in the non-adjusted model and the OR of high complexity changed sign. In the variables of admission and DRG type, the OR of non-surgical DRGs also showed a considerable decrease, as did that of the variable of death.
4. Discussion The results of this study demonstrate that a relationship exists between hospital structural level and LOS outliers and confirm the initial hypothesis that the percentage of outliers in public hospitals increases with hospital structural level. However, the influence of the control variables on the probability of a patient being an outlier also confirms that patient and health care characteristics should also be taken into account when analyzing the causes of differences in the presence of outliers among hospital structural levels. If these characteristics are not randomly distributed between the different hospital structural levels, then differences in the presence of outliers among hospital structural levels may correspond to differences in the patient or health care characteristics of each structural level. Possibly, there is an effect according to which hospitals in different structural levels attract patients with different probabilities of being outliers. In fact, our database confirms that patient complexity (measured by DRG) and severity (valued through several variables) are related to the structural level. The multiple model, which adjusts for the effect of hospital structural level through the control variables, can be used to determine the extent to which differences in the probability of a case being an outlier among hospital levels are due to the level itself or are due to characteristics of the patients or health care process. The difference in the OR of hospital structural complexity between the non-adjusted and the adjusted model gives the value of the probability of a patient in a hospital structural level being an outlier that is related to the characteristics of the patient or health care process. Indeed, the decreased probability associated with the different hospital structural levels in the adjusted model compared with the non-adjusted model indicates that hospitals with a higher structural level tend to treat a greater number of patients who, due to their
characteristics, have an increased probability of being outliers. However, when these patient and health care process characteristics were analyzed in the adjusted model, the differences in probabilities remained clearly significant. Thus, factors intrinsic to hospital structural level itself explain these remaining differences. Possibly, some residual case-mix differences were not captured through the control variables used, and thus some of the differences found might not be due to hospital structural level itself. However, the adjusted model takes into account a large number of factors, beyond the standard patient classification systems, such as the comorbidities indicator, recently developed by Elixhauser, and those variables than can be created from the administrative data available in a minimum data set. Thus, although a small part of outlier differences among structural levels cannot be well explained by the case-mix variables used, we believe that an adjusted relationship between structural hospital level and percentage of outliers has been demonstrated. The teaching and research activities normally carried out in hospitals with higher structural complexity might justify this remaining difference in the percentage of outliers since the cost associated with these activities implies greater hospital cost. However, teaching and research activity justify certain extra resource use, but do not justify differences in the number of outliers. In particular, these activities justify a more dispersed DRG resource use distribution due to the differences associated with the hospital type, which would lead to a higher cut-off point and consequently fewer cases would be detected as outliers. Concerning the behavior of the remaining variables in the adjusted model, the variables explaining patient seriousness (comorbidities, complications, age) within DRGs, show a very clear relationship with the probability of a patient being an outlier. These variables, together with the case-mix, incorporated into the analysis through the DRG-weight, could have potentially explained and justified the differences between levels and, thus, might have contradicted the original hypothesis. However, this was not the case. Although these variables showed considerable explanatory power, they did not explain the causes that, after adjustment, were clearly related to structural complexity. This was also true of the variables defining the health care process. Emergency admissions, re-admissions, inpatient deaths and the distance
F. Cots et al. / Health Policy 68 (2004) 159–168
between patients’ place of residence and the hospital added explanatory power, but did not refute the initial hypothesis. In the category of high complexity DRGs, the change in sign from the non-adjusted to the adjusted model shows that outliers in this category were also more frequently characterized by other model variables associated with a higher probability of being an outlier than were the outliers in lower complexity DRGs. A model without this variable was tested (results not shown) and no substantial change in any of the variables was observed. One limitation of this study is that variables related to the characteristics of hospital structural level, such as hospital size, and the patient-clinician ratio, among others, were not included in the analysis. These variables might have identified the structural factors that explain differences in the probability of patients being outliers. The limitations of DRGs as a patient classification system could also have marginally influenced the results. Underlying intra-DRG variability may not be homogeneously distributed between the different structural levels, at least there is no evidence that it is. However, because DRGs are almost universally used as a patient classification system, this limitation is insuperable. Finally, because socio-economic information from the hospital discharge registers is lacking, the effect of socio-economic status on the probability of a patient being an outlier could not be estimated. Socio-economic status might be another factor that could justify differences in the presence of outliers among different hospital structural levels. Some authors [18,19] have proposed that the absence of this information could be overcome by using census data on patients’ residence. A preliminary result of the present study using a socio-economic variable related to the patients’ city or town of residence indicated that an ecological approach did not provide a solution to the problem of absence of individual socio-economic information. Although it cannot be inferred that there are significant differences in efficiency among different hospital structural levels in overall production function, an actual difference in the number of outliers was detected. Bearing in mind that the adjusting variables largely captured the case-mix treated in each
167
hospital structural level, many outlier cases were associated with hospital structural level itself and were not a consequence of the complexity of the case-mix treated in each level. Thus, the implication of this result for health policy, and more specifically for the financing and management of public hospitals, is that the presence of outliers influences final hospital cost and consequently financing needs. The outliers in a given hospital can account for approximately 4.5% (from 3.6 to 5.6%, depending on the structural level) of all discharges and can represent 20.2% (from 16.9 to 23.8%, depending on the structural level) of the total cost. After adjusting for severity, there seems to be no reason why large urban teaching hospitals and medium-sized teaching and community hospitals should have greater numbers of outliers than small community hospitals. Therefore, payment of extremely expensive cases could be adjusted according to the mean incidence of these cases within the whole system after adjusting for severity. Otherwise, an insurance effect would be produced, which would offer incentives to hospitals with high structural levels to have a greater number of unjustified outliers, since there would be no clear incentives not to have them.
Acknowledgements This study was supported by the ‘Fundación de Investigaciones Sanitarias-Instituto de Salud Carlos III-Ministerio de Sanidad y Consumo’, project number 99/0687.
References [1] Cots F, Elvira D, Castells X, Dalmau E. Medicare’s DRGWeights in a European environment: the Spanish experience. Health Policy 2000;51:31–47. [2] Lichtig LK. Hospital information systems for casemix management. New York: Wiley, 1986. [3] Briggs A, Gray A. The distribution of health care costs and their statistical analysis for economic evaluation. Journal of Health Services Research and Policy 1998;3:233–45. [4] Diehr P, Yanez D, Ash A, Hornbrook M, Lin DY. Methods for analyzing health care utilization and costs. Annual Reviews of Public Health 1999;20:125–44. [5] Keeler EB, Carter GM, Trude S. Insurance aspects of DRG outlier payments. Journal of Health Economics 1988;7:193– 214.
168
F. Cots et al. / Health Policy 68 (2004) 159–168
[6] Carter GM, Rumpel JD. Payment rates for unusual Medicare hospital cases. Santa Mónica: RAND, 1992. [7] Carter GM, Farley DO. Assenting the FY 1989 change in Medicare PPS outlier policy. Health Care Financing Review 1992;14:69–82. [8] Calore KA, Iezzoni L. Disease staging and PMCs. Can they improve DRGs? Medical Care 1987;25:724–37. [9] Thomas JW, Ashcraft MLF. Measuring severity of illness: six severity systems and their ability to explain cost variations. Inquiry 1991;28:39–55. [10] Söderlund N, Gray A, Milne R, Raftery J. Case mix measurement in english hospitals: an evaluation of five methods for predicting resource use. Journal of Health Services Research and Policy 1996;1:10–9. [11] Finlayson G, Roos NP, Jacobs P, Watson D. Using the Manitoba hospital management information system: comparing average cost per weighted case and financial ratios of Manitoba hospitals. MCHP2001. Manitoba: Manitoba centre for health policy and evaliation, 2001. p. 1–149. [12] López Casasnovas G, Saez M. The impact of teaching status on average costs in Spanish hospitals. Health Economics 1999;8:641–51. [13] Ibern P, Vertrees JC, Manton KG, Woodbury MA. Hospital groups and case-mix measurement for resource allocation and
[14]
[15]
[16]
[17]
[18]
[19]
payment. In: Diagnosis related groups in Europe. New York: Springer-Verlag; 1993. Cots F, Castells X. Cómo pagamos a nuestros hospitales. La referencia de Cataluña y el contrapunto desde Andaluc´ıa. Gaceta Sanitaria 2001;15:172–81. Cots F, Mercadé L, Castells X, Torre P, Riu M. Risk adjustment: beyond patient’s classification systems. Gaceta Sanitaria 2001;15:423–31. Cots F, Castells X, Elvira D, Saez M. Relevance of outlier cases in case mix systems and evaluation of trimming methods. Health Care Management Science 2003;6:27– 35. Elixhauser A, Steiner C, Harris R, Coffey RM. Comorbidity measures for use with administrative data. Medical Care 1998;36:8–27. Hofer TP, Wolfe RA, Tedeschi PJ, McMahon LF, Griffith JR. Use of community versus individual socioeconomic data in predicting variation in hospital use. Health Service Research 1998;33:243–59. Krieger N. Overcoming the absence of socioeconomic data in medical records: validation an application of a censusbased methodology. American Journal of Public Health 1992;92:703–10.