Printer Friendly

Evaluation of the HCFA model for the analysis of mortality following hospitalization.

From 1987 through 1990, the Health Care Financing Administration (HCFA) evaluated variations in the mortality rates experienced by patients admitted to hospitals participating in the Medicare program. This study was conducted to evaluate the adequacy of the model used for that purpose. Detailed clinical data were gathered on 42,773 patients admitted to 84 statistically selected hospitals. The effect of risk adjustment using the HCFA model, which is based on claims data, was compared to a risk-adjustment model based on physiologic and clinical data. Models that include claims data were markedly superior to those containing only demographic characteristics in predicting the probability of patient death, and the addition of clinical data resulted in further improvement. The correlation of ranks of hospitals based on a model that uses only the claims data and on one that uses, in addition, clinical data, was .91. As a screen for the identification of "high (mortality) outlier" hospitals, the claims model had moderate sensitivity (81 percent) and specificity (79 percent), a high negative predictive value (90 percent), and a low positive predictive value (64 percent) when compared to the clinical model. The two mortality models gave similar results when used to determine which structural characteristics of hospitals were related to mortality rates: hospitals with a higher proportion of registered nurses or board-certified physician specialists, or with a greater level of access to high-technology equipment had lower risk-adjusted mortality rates. These data suggest that the current claims-based risk-adjustment procedure may satisfactorily be used to characterize variations in mortality rates associated with hospitalization. The procedure could also be used as a basis for further epidemiological analyses of factors that affect the probability of patient death. However, it does not positively identify outlier hospitals as providers of problematic care.

In 1985, the Health Care Financing Administration (HCFA) began evaluating the variations in patient mortality rates among hospitals. The initial intent of the effort was to assist the peer review organizations (PROs) in targeting their activities. This objective was subsequently broadened to include providing information to Medicare beneficiaries, hospital administrators, physicians, and researchers about hospitals, both locally and nationally.

It was recognized from the outset that the results of the analyses could not provide definitive characterizations of differences among the hospitals regarding the effectiveness of care they provide because of the following limitations:

1. The data on which the analyses were based were obtained

from claims submitted by the hospitals to HCFA through

fiscal intermediaries but were not subjected to the quality

control standards usually applied to data used for epidemiologic

research (Hsia, Krushat, Fagan, et al. 1988).

2. Clearly the data submitted on the claims characterized only

in a limited way the condition of the patient at the time of

admission (severity of illness) (Eastaugh 1986; National

Association of Public Hospitals 1987; Dubois et al. 1987). If

hospitals differ substantially in the severity of illness of their

patients, differences in mortality rates of hospitalized

patients could be driven more by this factor than by differences

in the adequacy and appropriateness of care.

3. A particular specification of the analytical model (logistic

regression), of the outcome (death within 30 days of the last

admission of a given patient in the calendar year), and of

the description of the patient to define his or her contribution

to the risk of death (i.e., disease causing the admission,

comorbidities, prior history of hospitalization, demographics)

was chosen. However, the modeling of the probability of

death is still in an investigative stage, and no particular

model has as yet been universally accepted.

4. Although the adequacy and appropriateness of medical

interventions are properly assessed by the measurement of

mortality, morbidity, and disability rates and the expenditures

for health care that flow from them (Lohr 1988), only

mortality rates were examined.

5. Some controversy remains about the relevance of outcomes

of medical care - such as death - to the evaluation of quality

of care.

Consequently, HCFA took on several projects that examined the adequacy and usefulness of the model used to analyze the mortality rates of hospitalized patients. One project assessed the association of hospital structural characteristics (Hartz, Krakauer, Kuhn, et al. 1989) while another looked at the correlation of the incidence of problems identified by the HCFA Generic Quality Screens with mortality rates (Hartz et al. in press).

The objective of our study was to compare the results of risk adjustment achieved with data selected from the claims for payment submitted by hospitals to HCFA ("claims" model) with the risk adjustment achieved by means of detailed physiologic and clinical data ("clinical" model) obtained from the hospital records. Two issues are explored: (1) Does the claims model, with variables as defined in the Medicare Hospital Mortality Information from 1988, adequately characterize the overall influence of individual hospitals on the probability of patient death? and (2) Are estimates of the influence of hospital structural characteristics on patient mortality obtained with the claims model supported by the fuller clinical model?

METHODS

HOSPITALS AND PATIENTS SAMPLED

This study included 42,773 patients admitted in 1986 to 84 hospitals throughout the United States. As in the case of the Medicare Hospital Mortality Information releases, only the most recent discharges of patients in the calendar year were evaluated. The sample was designed to ensure wide representation of hospital size and of the differences between the observed mortality rates and the expected mortality rates. A hospital was not sampled if it had a mortality rate greater than 45 percent or had fewer than 50 eligible Medicare patients discharged in 1986. The remaining hospitals were grouped into three strata on the basis of number of Medicare cases (50-249, 250-699, and 700 +). Within each stratum, hospitals were divided into substrata (16 for the first stratum, 25 for the second, and 46 for the third) on the basis of the difference between the observed and predicted mortality rates. The numbers of substrata were determined to make them proportional to the information contributed by the typical hospital of each stratum (i.e., the sample contained more of the larger, more informative hospitals). One hospital was then randomly selected from each substratum. If the hospital chose not to participate, an alternative hospital was selected. For 84 of the 87 substrata it was possible to find a participating hospital. No institutions selected from large hospital chains chose to participate.

Medical records for all eligible 1986 Medicare discharges from the selected hospitals with fewer than 700 eligible discharges were abstracted, but only a random sample of 700 discharges were abstracted from the stratum that included hospitals with 700 or more discharges.

ANALYTIC PROCEDURE

Two models were used to obtain an adjusted mortality rate from each hospital: the claims model and the clinical model. The claims model contains patient descriptors obtained from the Social Security Administration and from the claims submitted to HCFA by the hospitals. This information includes dates of admission and discharge, one principal discharge diagnosis, up to four secondary diagnoses, age, sex, race, comorbidities, transfer status, date of death, and the prior hospitalization history of the patient (Health Care Financing Administration 1989).

The clinical model contains, in addition to the variables derived from the claims, detailed physiologic and clinical data abstracted from the hospital records. MediQual, Inc., under contract to HCFA, performed the abstraction using the MedisGroups[TM] system (Brewster, Carlin, Hyde, et al. 1985). In this method, approximately 400 items of clinical information pertaining to the patients history, physical examination, and laboratory and diagnostic tests performed within the first 48 hours of admission or prior to surgery are evaluated, and abnormal findings are recorded.

For the claims model, weighted logistic regression was performed on all 42,773 patients, using death within 30 days of admission as the dependent variable. The weighting factor was the ratio of the number of eligible discharges at the hospital treating the patient to the number of patients from that hospital whose records were abstracted. Thirty one of the claims-derived variables were identified as statistically significant predictors of patient death at the p < .01 level by means of backwards stepwise multiple logistic regression.

The more extensive array of claims and clinical variables in the clinical model necessitated a more elaborate procedure for the identification of those to be retained in the regression analyses. First, clinical findings were dropped if they were rarely abnormal (with approximately less than 90 occurrences in the sample). Next, cluster analysis was used to identify the diagnostic, claims, and clinical variables that tended to occur together in the study patients. For each of 12 clusters formed, backwards stepwise logistic regression was used to determine which of the variables were statistically significantly related (p < .005) to the probability of death within 30 days of admission. The selected variables were then included, along with demographic variables and variables pertaining to prior admissions in a final backwards stepwise logistic analysis. The final lists of patient-related variables, regression coefficients, and standard errors for both the claims model and the clinical model are given in the Appendix.

In both the claims and the clinical models, the measure of the hospital's influence on the probability of patient death was estimated by means of the regression coefficients of the indicator variables for the hospitals added to the patient level variables (the "Hospital Coefficients"). One hospital, the largest in the sample and one whose residual mortality (the difference between the observed and the predicted mortality rates) - as determined for the Medicare Hospital Mortality Information from 1988 - was very small (two deaths per thousand patients) was used as a reference throughout. Its regression coefficient, when it was estimated separately, was indistinguishable from 0: -0.038 [+ or -] 0.053 in the claims model and 0.003 [+ or -] 0.060 in the clinical model.

The ability of the models to predict death within 30 days of admission for the patients in the sample was measured by the overlap index (Hartz 1984). This index, derived from the Wilcoxon rank sum test, compares the predicted probabilities of death for patients who died with the probabilities of death of patients who did not die. It ranges from 0 (all patients who died had a higher predicted probability of death than all patients who did not die) to 1(patients who died were as likely to have the higher probability of death as those who did not). A second measure of the predictive power of the model is given by the fraction of pairs of patients - one who died and one who did not - in which the patient who died had the higher probability of death (concordant pairs) (.5 = random pairings, 1 = perfect pairings).

The adequacy of the risk adjustment by means of the claims model was assessed by comparing the regression coefficients estimated for the hospital indicator variables in both the claims and the clinical models. The measures used were (1) the correlations of hospital rank as determined by their regression and standardized regression coefficients (regression coefficients divided by their standard errors); (2) the proportions of hospitals estimated to have statistically significantly high regression coefficients by means of the claims model but not by the more complete clinical model; and (3) the changes in decile ranks of hospitals caused by the inclusion of clinical variables.

HOSPITAL CHARACTERISTICS

The data on hospital structural characteristics were obtained from the 1986 American Hospital Association (AHA) survey, and were available for 77 of the 84 hospitals in the study. Several hospital structural characteristics, as recorded in the AHA survey, had been previously found to be associated with mortality rates (Hartz, Krakauer, Kuhn, et al. 1989). These included (1) hospital finances, as indicated by occupancy rates and total payroll expenses per hospital bed; (2) hospital ownership, with hospitals put into three groups: for-profit, public, and not-for-profit; (3) the level of training of clinical staff, as indicated by the proportion of registered nurses and the proportion of all physicians in the hospital who were board-certified specialists. (Physicians in family medicine, general practice, emergency medicine, or general surgery were not counted among the specialists since they treat a very broad range of patients.) Teaching hospitals were defined by membership in the Council of Teaching Hospitals and osteopathic hospitals by membership in the American Osteopathic Hospital Association; (4) the sophistication and technical capability of the hospital, as indicated by the presence of a cardiac catheterization laboratory, an extracorporeal lithotripter, a magnetic resonance imaging facility, an open heart surgery facility, or organ transplant capability; and (5) hospital size, as indicated by the number of beds.

The association of hospital structural characteristics with mortality rates was evaluated as in the previous study (Hartz, Krakauer, Kuhn, et al. 1989). Following risk adjustment for patient characteristics as described earlier, a further adjustment was carried out in regressions at the hospital level for certain demographic and socioeconomic characteristics of the hospitals' patient populations not included in the claims and clinical models applied at the patient level. Of the variables tested, only the proportion of patients who were black had a statistically significant effect. With that factor adjusted for, standard linear regression procedures (Sweeney and Ulveling 1972) were used to calculate adjusted mortality rates for hospitals with or without specified structural characteristics, or, when those characteristics were represented by continuous variables (e.g., the number of hospital beds), adjusted mortality rates for hospitals in the upper and lower quartiles in terms of those variables (Hartz, Krakauer, Kuhn, et al. 1989).

RESULTS

Selected demographic characteristics and reasons for hospitalization (groupings of principal diagnoses) for patients admitted to the hospitals in this study, in comparison with all final Medicare discharges in 1986, are shown in Table 1. The sampling design resulted in a higher proportion of blacks (16 percent) in the sample than in the population from which it was drawn (8 percent). The groups were similar in other respects.
Table 1: Characteristics of the Populations in All Hospitals
and in the Hospitals Studied
 Percent in All Percent in
 Hospitals Hospitals
 (N = 5,585) Sampled
 (N = 84)
Demographic Characteristics
 Age 73.97 72.93
 Black 8.18 15.76
 Women 55.73 53.93
Disease Categories
(Causes of Admission)
 High Risk
 Cancer 3.51 4.21
 Cerebrovascular accidents 3.76 3.58
 Severe acute heart disease 3.99 3.99
 Severe chronic heart disease 4.84 4.48
 Gastroenteric catastrophes 0.83 0.79
 Pulmonary disease 5.35 5.60
 Renal disease 0.54 0.66
 Metabolic/endocrine disorders 1.68 1.88
 Sepsis 1.16 1.51
 Severe trauma 0.19 0.25
 Low Risk
 Low-risk heart disease 12.71 11.62
 Gastrointestinal disease 7.28 6.27


Tables 2 and 3 present assessments of the predictive power of the various models used to estimate mortality rates. In Table 2, two measures of goodness of fit are shown - the overlap index and, a related measure, the proportion of concordant pairs of patients. It is evident that the claims model is superior to a model using only demographic characteristics (age and sex) as predictive factors, and that the clinical model represents a further improvement.
Table 2: Evaluation of the Goodness-of-Fit of Regression
Models for the Probability of Death of Individual Patients
 Proportion of Overlap
 Concordant Pairs(*) Index([dagger])
Variables in Model
 Demographic only 0.64 0.72
 Claims data 0.84 0.33
 Clinical model 0.90 0.21
(*) Concordant pairs: in pairs consisting of one patient who
died and one who did not,
those pairs in which the patient who died had the higher
predicted probability of
dying. Range = 0.5-1.
([dagger]) Overlap index; measures the overlap in the predicted
probabilities of death among
patients who died and who did not die. Its range is 0 (perfect
prediction) to 1 (no
prediction).
Table 3: Comparison of Predicted and Observed Mortality
Rates of Patients Grouped by Hospital (Sample of 84 Hospitals)
 Mean Percentiles of the Sample
 (%) 95% 75% Median 25% 5%
Predicted Rates
 Published model(*) 13.2 18.7 14.6 12.9 11.5 9.5
 Claims model 13.2 18.3 14.8 13.0 11.4 9.2
 Clinical model 13.7 21.8 16.2 13.4 11.0 7.2
Observed Rates 15.1 29.9 18.9 14.0 10.1 6.1
(*) The published model is the model used in the Medicare
Hospital Mortality Information,
1988.


The stratification of hospitals in terms of the predicted probabilities of death of their patients is shown in Table 3. The claims model produces a twofold difference and the clinical model a threefold difference between the hospitals in the top 5 percentile and the bottom 95 percentile in terms of the predicted probability of death. The correlation of hospital rank based on the predicted mortality rates in the two models is .77.

The stratification of hospitals in terms of the predicted probabilities of death of their patients is shown in Table 3. The claims model produces a twofold difference and the clinical model a threefold difference between the hospitals in the top 5 percentile and the bottom 95 percentile in terms of the predicted probability of death. The correlation of hospital rank based on the predicted mortality rates in the two models is .77.

The correlations of ranks of hospitals produced by the model used in the Medicare Hospital Mortality Information (1988) (the published model), the claims model, and the clinical model applied to the sample in the present analyses are given in Table 4. The correlation of ranks based on the residual mortality rate obtained in the published model and the hospital regression coefficients (estimates of the hospital's contribution to the probability of patient death) obtained with the claims model is very high (.96), supporting the equivalence of the two approaches. The correlation of ranks based on the regression coefficients obtained with the claims and clinical models is .91, and .87 when based on the standardized regression coefficients. The latter quantity takes into account the uncertainty in the estimation of the regression coefficient.
Table 4: Correlations of Ranks of Hospitals
 Hospital Standardized
 Coefficient Hospital
 Coefficient
 Claims Model
Published Model(*)
 Residual 0.96 -
 Standardized residual - 0.92
 Clinical Model
Claims Model
 Hospital coefficient 0.91 -
 Standardized hospital coefficient - 0.87
Published Model
 Residual 0.88 -
 Standardized residual - 0.82
(*) The published model is the model used in the Medicare
Hospital Mortality Information, 1988.


In Table 5 we present the consistency in the identification of hospitals that are high outliers, that is, those hospitals whose regression coefficient is statistically significantly greater than 0 at the p < .05 level. Of the 33 hospitals that are high outliers according to the claims model, 12 (36 percent) are not considered high outliers by the clinical model, and, conversely, 5 of the hospitals not identified as high outliers by the claims model are so designated in the clinical model. If the clinical model is accepted as the reference standard, then the sensitivity of the claims model (the number of hospitals that are high outliers in both models divided by the number that are high outliers in the clinical model) is 81 percent (21 of 26). The specificity of the claims model (the number of hospitals that are not high outliers in the claims model and also are not high outliers in the clinical model, divided by the number that are not high outliers in the clinical model) is 79 percent (46 of 58). The positive predictive value (the number of hospitals that are high outliers in both models divided by the number that are high outliers in the claims model) is 64 percent (21 of 33), and the negative predictive value (the number of hospitals that are not high outliers in both models divided by the number of hospitals that are not high outliers in the clinical model) is 90 percent (46 of 51).
Table 5: Relationship of High Outlier Status under the
Claims and the Clinical Models
 Number of Hospitals
 Clinical Model
 Outlier(*) Not Outlier Total
Claims Model
 Outlier(*) 21 12 33
 Not outlier 5 46 51
 Total 26 58 84
(*) Outliers are hospitals whose standardized regression
coefficients were statistically significantly high at the
p < .05 level.


The change in the decile of the hospital rank that resulted from substitution of the clinical for the claims model is shown in Table 6. Only one of the hospitals in the extreme deciles (1, 2, 9, 10) changed two deciles or more, and a total of only eight hospitals (10 percent) moved three or more deciles.

[TABULAR DATA OMITTED]

A comparison of the structural characteristics of the hospitals in this study and those in the previous evaluation of 3,100 hospitals (Hartz, Krakauer, Kuhn, et al. 1989) is shown in Table 7. The principal differences in this study are a higher preponderance of public hospitals (42 percent versus 24 percent) and a higher proportion of black patients (16 percent versus 8 percent). In addition, whereas status as a for-profit hospital and as an osteopathic hospital were evaluated in the previous study, these types of hospital were insufficiently represented (seven and three cases, respectively) in this study. In addition, as noted earlier, no members of chain hospitals participated.

[TABULAR DATA OMITTED]

The influence of the hospital characteristics on the adjusted mortality rate (the sum of the overall average mortality rate and the residual mortality rate for that hospital) is shown in Table 8. The consistent effect of the addition of clinical data to the model is to blunt the influence of the structural characteristics of the hospitals on the adjusted mortality. However, the three structural characteristics that are statistically significant under the claims model (percentage of physicians who are board-certified specialists, percentage of nurses who are registered, and the high-technology index) remain statistically significant-or nearly so-under the clinical model. In addition, the trends observed in our sample of hospitals are quite similar to the trends observed in the 3,100 hospitals reported previously (Hartz, Krakauer, Kuhn, et al. 1989).

DISCUSSION

The validity of the results released by HCFA is best tested against a model that permits thorough risk adjustment and as direct an assessment of the hospital's contribution to the probability of patient death - and the uncertainty therein - as possible. Since the outcome of a hospitalization is highly dependent on the disease of the patient and its severity, the validity of the inference drawn from mortality rates about the performance of hospitals is highly dependent on the adequacy of the data and the model used for risk adjustment. Therefore, an intense effort was made to obtain a complete characterization of the patients, specifically of their condition at the time of admission, by abstracting detailed physiologic and clinical data from the hospital records.

The claims model and, even more so, the clinical model permit a substantial stratification of patients and hospitals in terms of the predicted probability of death within 30 days of hospitalization. The ranks of hospitals obtained using the claims model were highly correlated with those obtained using the clinical model (correlation coefficient = .9), and only one hospital in the extreme deciles of rank (1, 2, 9, 10) moved 2 deciles or more. We may, therefore, concluded that the claims model and, as a result, the model used in the Medicare Hospital Mortality Information releases, is quite useful in describing the distributions of mortality rates encountered at hospitals. Considered as a screening tool to identify "problem" hospitals (i.e., hospitals with significantly high mortality), the claims model has moderate sensitivity (81 percent) and specificity (79 percent). In addition, its negative predictive value is high (90 percent), implying that if a hospital is not an outlier under the claims model then it is unlikely to be an outlier under the clinical model. However, the positive predictive value was low (64 percent) in this study. This implies that, if the hospital is an outlier under the claims model, there remains a substantial probability (36 percent in this study) that it will not be an outlier under the clinical model. In contrast, the rankings of the hospitals, as indicated by the rank-order correlation coefficients and the limited movements spanning deciles of rank, were much more stable. This suggests that ranking hospitals according to adjusted mortality rather than defining a line beyond which they become "outliers" might be considered as an alternative format for presenting the data in the Medicare Hospital Mortality Information.

The trends in the influence of structural hospital characteristics were internally consistent in comparisons of results of the claims and the clinical models, when compared with a quite differently sampled set of hospitals reported on elsewhere (Hartz, Krakauer, Kuhn, et al. 1989). It is not surprising that hospitals with a higher proportion of physicians who are board-certified specialists, with a higher proportion of registered nurses, and hospitals that evidence technical sophistication through the possession of technologically advanced equipment and services have lower adjusted mortality rates. Indeed, the intuitive acceptability of such a result bolsters confidence that mortality rates associated with hospitalization, when properly adjusted for the risks contributed by the patients, are useful measures that may be applied in epidemiologic studies of factors that affect the outcomes of medical care.

The model taken as the reference standard in these analyses is burdened with substantial deficiencies. The limitations of data such as diagnosis codes reported on claims for reimbursement are well known (Hsia, Krushat, Fagan, et al. 1988). The rules for the abstraction of data from medical records in MedisGroups[TM] specify a 48-hour period following admission as the window for the identification of abnormalities, a period during which iatrogenic effects may come to the fore. In addition, as only abnormal findings are recorded, it is not clear whether the absence of an abnormality represents normality, the failure to seek the finding, or the failure to perform the test. The effectiveness of medical care is not adequately characterized by mortality rates alone. Other outcomes such as morbidity, disability and expenditures must also be evaluated, especially since a great deal of medical practice entails compromises and trade-offs among these outcomes. In addition, 30 days is probably too short a period of follow-up because medical interventions may have not only short-term but also long-term effects. It may well be that the variables included even in the clinical model do not sufficiently account for factors that affect the probability of death beyond the control of the provider of medical care. Additional factors that might be considered are socioeconomic and genetic characteristics of patients, and environmental factors.

The sampling process, designed to stratify hospitals by excess risk of death and to provide representative samples within the strata, did not fully succeed in the latter task. Thus, the investor-owned chains are not represented because they would not participate, and the larger teaching hospitals are overrepresented. As a consequence, there is some disparity in patient characteristics between those of the hospitals studied and those of all hospitals nationally, most notably in the percentage of patients who were black (Table 1). In addition, because the analyses were performed on hospitalizations that occurred in 1986 and because hospital practices are likely to have changed since then, the present applicability of some of the specific conclusions may be limited.

There is, therefore, considerable opportunity for the refinement and improvement of the methodology and models used to evaluate outcomes of hospitalizations. Nevertheless, it does appear that models based on the data obtained from the claims submitted by hospitals to HCFA do provide substantial information about the variations in the patterns of hospitalization outcomes.

Appendix
Variables in the Claims Model
 Regression Standard
Variable Coefficient(*) Error(*)
Intercept -3.728 0.276
Demographic
 (Age/65) -1.706 0.487
 Square of (Age/65) 2.087 0.217
 Sex -0.215 0.022
Disease Categories
 Cancer 1.443 0.043
 Cerebrovascular accidents 1.538 0.044
 Neurologic disorders -0.979 0.093
 Gastroenteric catastrophes 1.696 0.084
 Gastroenterologic disease -0.712 0.071
 Genitourinary disease -1.629 0.134
 Low-risk heart disease -0.942 0.056
 Severe chronic heart disease 0.707 0.044
 Severe acute heart disease 1.854 0.041
 Metabolic disturbances 0.956 0.064
 Orthopedic disorders -1.045 0.076
 Pulmonary disease 1.247 0.038
 Renal disease 1.795 0.094
 Sepsis 1.834 0.061
 Severe trauma 1.630 0.139
Comorbidities
 Diabetic disease 0.180 0.044
 Cancer 0.860 0.035
 Chronic heart disease 0.605 0.023
 Chronic liver disease 1.160 0.092
 Chronic renal disease 1.026 0.052
Prior Admissions
 High risk, one 0.644 0.029
 High risk, two 1.026 0.054
 High risk, three or more 1.557 0.089
 Low risk, two 0.265 0.080
 Intermediate risk, one 0.440 0.030
 Intermediate risk, two 0.752 0.053
 Intermediate risk, three or more 0.747 0.079
Transfers In -0.131 0.073
(*) Regression coefficients and standard errors given here are
for the population in the hospitals in the sample.
Variables in the Clinical Model
 Regression Standard
Variable Coefficient(*) Error(*)
Intercept 14.699 3.080
Demographic
 Square of (Age/65) 1.017 0.037
 Sex -0.247 0.025
Disease Categories
 Cancer 1.454 0.050
 Cerebrovascular accidents 1.357 0.053
 Neurologic disorders -0.568 0.096
 Gastroenteric catastrophes 1.145 0.097
 Gastroenterologic disease -0.382 0.074
 Genitourinary disease -1.256 0.138
 Low-risk heart disease -0.546 0.058
 Severe chronic heart disease 0.343 0.051
 Severe acute heart disease 1.571 0.048
 Pulmonary disease 0.469 0.046
 Sepsis 0.788 0.079
 Severe trauma 1.221 0.158
Comorbidities
 Cancer 0.872 0.043
 Chronic heart disease 0.670 0.027
 Chronic pulmonary disease -0.167 0.043
 Chronic renal disease 0.628 0.060
Prior Admissions
 High risk, one 0.437 0.033
 High risk, two 0.676 0.059
 High risk, three or more 1.258 0.098
 Low risk, two 0.369 0.086
 Intermediate risk, one 0.342 0.033
 Intermediate risk, two 0.677 0.059
 Intermediate risk, three or more 0.632 0.085
Clinical Findings
 Tumor: malignant -0.829 0.087
 Tumor: benign -2.303 0.602
 Disoriented X2 or X3 0.358 0.045
 Brain: high density 0.470 0.118
 Congestive heart failure 0.081 0.038
 Aorta: aneurysm 0.520 0.194
 Blood culture: positive -0.482 0.075
 Infiltrate 0.198 0.036
 Pleural effusion 0.312 0.041
 Pneumothorax -0.592 0.192
 Cyst -0.543 0.245
 Skin: infection -0.776 0.145
 Deficit: acute cranial nerve -0.336 0.110
 Klebsiella 0.030 0.115
 Coma/stupor 1.444 0.047
 Lethargy 0.417 0.041
 Murmur 0.184 0.047
 Gangrene (any site) 1.176 0.137
 Anuria-cathetered (< 400cc/day) 0.932 0.147
 History of metastasis 0.451 0.055
 History of seizures -0.426 0.079
 History of angina/myocardial
 infarction 0.176 0.034
 History of congestive
 heart failure 0.058 0.034
 History of asthma -0.286 0.085
 Hyperglycemia (> 149)([dagger]) 0.00076 0.00010
 Hyperkalemia (> 5.3) 0.127 0.027
 Alkalosis (pH > 7.45) 2.966 0.350
 Leukocytosis (> 10,000) 0.015 0.001
 Hypocalcemia (Ca < 8.0) -0.062 0.021
 Acidosis (pH < 7.35) -4.104 0.267
 Hypothermia (< 35.8 [degrees]
 Celsius) 0.199 0.020
 Bands (percent or absolute) 0.015 0.002
 Partial thromboplastin time
 (> 75.0) 0.011 0.002
 Platelets (< 100.0) 0.0019 0.0004
 Albumin (< 3.0) -0.390 0.019
 Serum glutamate oxaloacetate
 transaminase (> 100) 0.0010 0.0001
 Alkaline phosphatate (> 300) 0.00043 0.00009
 BUN (> 29) 0.0086 0.0005
 Total bilirubin (> 2.0) 0.096 0.009
 Partial oxygen pressure (< 75.0) -0.0055 0.0010
 Partial carbon dioxide pressure
 (> 45) 0.011 0.002
 Blood pressure systolic
 ([is less than or equal to]
 - 90 mm) 0.017 0.001
 Respirations (> 21/min) 0.027 0.001
(*) Regression coefficients and standard errors given here
are for the population in the hospitals in the sample.
([dagger]) The variables beginning with Hyperglycemia and the
variables for age were entered as continuous
variables. All others are binary (0,1).


Brewster, A. C., B. G. Karlin, L. A. Hyde, C. M. Jacobs, R. C. Bradbury, and M. C. Young. "MEDISGRPS: A Clinically Based Approach to Classifying Hospital Patients at Admission." Inquiry 22 (Winter 1985): 377-87. Dubois, R. W., J. H. Moxley, D. Draper, and R. H. Brook. "Hospital Inpatient Mortality: Is It a Predictor of Quality?" New England Journal of Medicine 317, no. 26 (1987): 1674-80. Eastaugh, S. R. "Hospital Quality Scorecards, Patient Severity, and the Emerging Value Shopper." Hospital & Health Services Administration 31, no. 6 (1986): 85-102. Hartz, A. J. "Overlap Index: An Alternative to Sensitivity and Specificity in Comparing the Utility of a Laboratory Test." Archives of Pathology Laboratory Medicine 108 January 1984): 65-67. Hartz, A. J., H. Krakauer, E. M. Kuhn, M. Young, S. J. Jacobsen, G. Gay, L. Muenz, M. Katzoff, R. C. Bailey, and A. A. Rimm. "Hospital Characteristics and Mortality Rates." New England Journal of Medicine 321, no. 25 (1989): 1720-24. Hartz, A. J., M. Gottlieb, E. M. Kuhn, and A. A. Rimm. Evaluation of the HCFA Hospital Mortality Studies. Health Services Research (in press). Hsia, D. C., W. M. Krushat, A. B. Fagan, J. A. Tebbut, and R. P. Kusserow. "Accuracy of Diagnostic Coding for Medicare Patients under the Prospective Payment System." New England Journal of Medicine 318, no. 6 (1988): 353-55. Health Care Financing Administration. Medicare Hospital Mortality Information - 1988. Washington, DC: U.S. Department of Health and Human Services, 1989. Lohr, K. N. "Outcome Measurement: Concepts and Questions." Inquiry 25, no. 1 (1988): 37-50. National Association of Public Hospitals. Statement of NAPH on the Release of 1987 Medicare Hospital Mortality Data. 1987. Sweeney, R. E., and E. F. Ulveling. "A Transformation for Simplifying the Interpretation of Coefficients of Binary Variables in Regression Analyses." Ammerican Statistician 26, no. 5 (1972): 30-32.
COPYRIGHT 1992 Health Research and Educational Trust
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1992 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:Health Care Financing Administration
Author:Krakkaier, Henry; Bailey, R. Clifton; Skellan, Kimberley J.; Stewart, John D.; Hartz, Arthur J.; Kuh
Publication:Health Services Research
Date:Aug 1, 1992
Words:5642
Previous Article:Factors affecting interstate use of inpatient care by Medicare beneficiaries.
Next Article:Determinants of need and unmet need among cancer patients residing at home.
Topics:

Terms of use | Copyright © 2016 Farlex, Inc. | Feedback | For webmasters