Printer Friendly

Predicting hospital choice for rural Medicare beneficiaries: the role of severity of illness.

Previous research has confirmed that desirable hospital attributes as well as increased distance, or travel time, have an impact on hospital choice. These studies have become increasingly sophisticated in modeling choice. This study adds to the existing literature by estimating the effect of both hospital and individual characteristics on hospital choice, using McFadden's conditional logit model. Some patient characteristics have not previously been accounted for in this type of analysis. In particular, the effect of a patient's complexity of illness (as measured by Disease Staging) on the choice of hospital is taken into account. The data consist of over 12,000 Medicare discharges in three overlapping rural market areas during 1986. The hospital choice set was aggregated into seven groups of urban and rural hospitals. Results indicate that rural Medicare beneficiaries tend to choose hospitals with a large scope of service and with teaching activity over those with a lower scope of service and no teaching activity, holding other factors constant. Distance is a deterrent to hospital choice, especially for older Medicare beneficiaries. The more complex cases tend to choose larger urban and rural hospitals over small rural hospitals more often than less complex cases do.

The factors that influence a patient's choice of hospital are of interest to hospitals, insurance companies, and nearly all levels of government. Hospitals have been increasingly concerned with competition and often must decide whether to add or remove certain beds and whether to alter their scope or intensity of services. This is particularly true for rural hospitals, which have been experiencing closure as well as downscaling of service provision in recent years. The type of model estimated here may help predict the short-term effect of such decisions.

We analyzed the hospital choice of rural Medicare beneficiaries, explicitly accounting for the existence and characteristics of their alternative choices. With this analysis, the probability that an individual with given characteristics would choose one type of rural hospital over another, or would choose an urban alternative, can be predicted. We used our results to estimate the distance that Medicare beneficiaries would travel for services if a large rural hospital in western Minnesota were to close. These simulations were presented as part of a larger study (Adams, Wright, and Robbins 1989).

Although significant research has been done on the aggregate demand of health care services, comparatively little work has been published on the choice of provider types. In the past, data sets that allowed for the analysis of hospital choice were difficult to assemble and manipulate. Measuring hospital service communities based on people's actual hospital preferences, rather than on assumptions, apparently began in the early 1940s with studies by Lembcke (Griffith 1978; Lembcke 1952). As noted by Griffith, assembly of the needed data by individual hospitals could take years. With the increased availability of automated patient data, patient origin studies have grown in number and sophistication. A recent study (Morrisey, Sloan, and Valvona 1988) found rural market areas to be larger than previously believed, partially because many rural residents choose a distant urban center hospital. Other analyses have become increasingly sophisticated in the estimation techniques used (Cohen and Lee 1985; Garnick, Luft, Robinson, et al. 1989; Luft, Garnick, Mark, et al. 1990; Dranove, White, and Wu 1989). None of these has focused solely on rural beneficiaries.

The overriding goal of our study was to analyze hospital choices made by rural Medicare beneficiaries during 1986 in the geographic areas surrounding a demonstration hospital. The premise of this demonstration was that certain larger, more sophisticated rural hospitals are more like urban hospitals in terms of case-mix complexity and, hence, in costs per case. The demonstration hospital argued that if it were to close, patients would seek care at nearby urban hospitals, which would cost the Medicare program more money. The present study was part of a larger evaluation of that demonstration (HCFA Contract No. 500-87-0028-6). Our market definition and analysis reflect this focus.

This study adds to the existing literature in several ways. First, a summary typology of hospital choices was derived to avoid estimation problems and to enhance the study's policy relevance. Both urban and rural hospital "types" were included in this choice set, thereby allowing analysis of choice among different types of rural hospitals. Second, the analysis tests the effect of the patient's severity of illness on hospital choice. This measure, based on SysteMetrics' Disease Staging, was derived for all Medicare discharges in the market area. To our knowledge, it is the first use of this measure in a hospital choice analysis. Finally, following the recent developments in this area (Garnick, Lichtenberg, Phibbs, et al. 1989), the model was estimated using maximum likelihood estimation to fit McFadden's conditional logit model.


The premise underlying the model is that individuals, through their physicians, choose hospitals on the basis of a vector of hospital attractiveness (size, scope of service, etc.), given their need (diagnosis) and enabling conditions. Although the choice of a hospital is strongly affected by the patient's physician, individuals have knowledge and preferences about alternative hospitals and can accept or reject physicians' decisions. Earlier research argued that hospital choice is inherently limited to the distance physicians are willing to travel (15-mile radius), while recent evidence has suggested that patients are either willing to travel farther, or physicians are willing to refer over longer distance (Morrisey, Sloan, and Valvona 1988). Regardless of the actual decision-making process, we hypothesized that the location of the individual's residence is a major determinant of the choice. That is, individuals with similar characteristics will choose the closer of two equally attractive hospitals (or hospital/physician bundles).

Although numerous multivariate techniques are available to analyze hospital choice, we present maximum-likelihood estimates of a multinomial logit model. This method for estimating the coefficients avoids the difficulties and bias inherent in methods of linearization. Two linearization methods were developed and used previously in the literature; they differed on how the reference choice was defined. One, developed by Nakanishi and Cooper (the geometric means approach), and another, developed by Berkson and Theil (the single reference hospital approach), were analyzed in a recent study (Garnick, Luft, Robinson, et al. 1989). The coefficients derived by the linearization methods were shown to be unstable, particularly as the number of zero flows between zip codes and hospitals increased.

We use a more general conditional logit model than the multinomial logic models used in earlier studies. In this more general model, characteristics of the choice itself are modeled as well as characteristics of the individuals who choose. A model allowing choice characteristics to affect choice probabilities was put forward by McFadden (1974). We assert that individuals choose from a set of alternatives in a way that will maximize utility based on (1) the attributes of that choice as they perceive them, and (2) their own individual characteristics. The general framework of the model is:


[Y*] = level of utility of tth individual making the jth choice

The observed variables are defined as:

[] = 1, if individual t makes the jth choice;

[i.e., [Y*] = MAX([Y*.sub.t1], [Y*.sub.t2], . . . [Y*])

[] = 0, otherwise,

and assume that:

[Y*] = [[beta]'] + [alpha.sub.j'Z.sub.] + [epsilon.sub.ij]


[Z.sub.t] = individual characteristics; and

[] = vector of attributes of jth choice as perceived by the tth



[Mathematical Expression Omitted]

is the probability that individual t makes choice j.

The maximum-likelihood estimates of the coefficients in this model can be obtained by several methods. We used the method of scoring as described by Maddala (1983, pp. 74-75). Note that the coefficients for the choice characteristics ([beta]s) are constant across choices, whereas the coefficients for individuals characteristics ([alpha]s) may differ for each choice. By convention, the [alpha] coefficients for the mth choice are all set to zero. Note also, this model assumes that the hospital choice of the ith individual is independent of the person's residence or particular origin. As such, it does not model any of the effects of community characteristics other than distance.


Basic search and location models assert that the consumer's incentive to search will increase in line with his or her increased perceptions of differences between seller price and quality. Although measures of hospital price and quality varied across earlier studies, most used some measure of time price in estimating the choice among alternative hospitals. Some studies measured time price using distance while others used more defined measures, including reservation wages as a measure of opportunity costs (Coffey 1983).

Time price as a determinant of demand for services has become more important with increased health insurance coverage. One study of physician visits showed that time price elasticity became stronger as the money price became concurrently less relevant to the full (money plus time) price (Acton 1976). We hypothesized that variation in time price is more important, for several reasons, to studies of Medicare beneficiaries than to research on the general population. First, hospital insurance coverage is uniform across the Medicare group. Furthermore, the hospital deductible and copayment do not vary by the hospital chosen; these are set nationally for Medicare beneficiaries. Finally, although physicians' charges do vary across areas (particularly urban/ rural) most beneficiaries (over 70 percent) carry supplemental policies for these out-of-pocket costs. Some policies even cover the balance billing of physicians who choose not to assign claims. We expect time price, like any other price, to influence behavior; the higher-priced alternatives are less likely to be chosen than a lower-priced, substitutable alternative.

Analogous to a hedonic-price approach, we also evaluated the values attached to various hospital characteristics. Several studies confirmed that hospitals with greater scope of service are preferred to those with less. We used a Guttman scale to measure scope-of-service capacity, and also tested for effects of hospital size and teaching status. To the extent that size is correlated with scope of service, there may not be a separate effect. Some elderly patients may prefer smaller hospitals, making the effect of size on hospitals ambiguous. To the extent that teaching status reflects advanced knowledge in treatment processes or greater intensity of service per episode, or both, we hypothesized that this variable would have a positive effect on hospital choice.

While the characteristics of a service clearly affect its definition and price, it is less clear how individuals' characteristics enter into a model of choice. Certainly, individual preferences are shaped by characteristics such as age, sex, educational level, and so on, and these interact with health status and service consumption. We did not deal with the complexity of translating health care status or needs into demand. Rather, we tested directly for differences in hospital choices attributable to the individual's age and severity of illness within diagnosis-related group (DRG). In a descriptive study (Hogan 1988), age has been shown to reduce the willingness to travel for hospital services. This is somewhat surprising, given that the oldest old may be more severely ill and perhaps more in need of specialized services. Yet their frailness and lack of transportation may be factors. Another study indicated that the elderly have somewhat stronger preferences for sub-urban and small-town hospitals than the nonelderly (Cohen and Lee 1985). We hypothesized that age would negatively affect the choice of urban and very large hospitals over rural, closer alternatives.

An important characteristics that we expected would affect hospital choice was the patient's severity or complexity of illness. Since those more severely ill are more likely to need, or want, treatment with advanced technology and by highly skilled personnel, we expected increased severity to be associated with an increased probability of choosing a larger, more sophisticated hospital. These are largely urban hospitals; however, a recent study (Morrisey, Sloan, and Valvona 1988) indicated that urban hospitals also compete with nearby rural hospitals for less complex cases. Also, a major tenet of the demonstration hospital was that, as a larger, more sophisticated rural facility, it was playing a particular role in serving complex cases. An important part of our empirical investigation, then, lay in examining whether severity of illness predicted the choice of a larger rural over a smaller rural facility as well as the choice of an urban hospital over a rural hospital.

Finally, the patient's type of medical condition was expected to affect hospital choice. For example, surgical producers are often elective, allowing the patients more time to consider alternative therapies. Although some surgical procedures require advanced technology, and hence may lead to the choice of a larger, perhaps urban hospital, this may not always be the case. If the required expertise is available closer to home, individuals may prefer to stay in their community. From earlier work, the need for psychiatric treatment was less a deterrent to travel than the need for medical and surgical treatment (Cohen and Lee 1985).


The data used in this analysis were based primarily on the 100 percent MEDPAR files available from the Health Care Financing Administration (HCFA) for fiscal year 1986. The data from the MEDPAR file were used to identify each Medicare inpatient's age, DRG category, and zip code of residence. The zip codes defined as part of the market area were predominantly in Minnesota but also included some in North and South Dakota.

Hospital characteristics such as the region, teaching status, and other PPS hospital designations were derived from HCFA Provider of Service and Provider Specific files. In addition, data from the American Hospital Association's (AHA) 1986 Annual Survey were used to measure each hospital's acute care bed size and to derive a Guttman scale of the scope of service (p. 595). This derivation followed the established literature in this area, and its full description is provided elsewhere (Adams, Wright, and Robbins 1989).

Zip code statistics for hospitals and individuals were drawn from the HCFA data files. Distance was measured using the latitude and longitude of each zip code. The latitude and longitude of each zip code was obtained from commercially available data files (Geographic Data Technology), which had been merged with the MEDPAR and hospital data.


Various methods for defining a hospital market area have been developed. Some of these rely on spatial or geopolitical boundaries (counties, MSAs) while others rely on the analysis of price movements (Stigler and Sherwin 1985). While the latter method is founded on economic theory, its application is difficult even in unregulated, competitive industries. The precursors to some of the methods based on patient origin can be found in the earlier literature on service populations and areas (Griffith 1978; Lembcke 1952).

Since hospital care is highly subsidized and price information is often poorly measured, researchers have often used the "shipments approach" (here, patient flows) to define market areas (Garnick, Lichtenberg, Phibbs, et al. 1989; Morrisey, Sloan, and Valvona 1988) rather than using price information. Elzinga and Hogarty (1978) used an intuitive approach to determining a market area. Their notion was that a market was "self-contained" (and therefore defined) if producers inside the market "exported" relatively few services to consumers outside the market, and consumers inside the market "imported" relatively few services from producers outside the market. This method starts with a specific producer, or hospital. A somewhat different approach, based on an initial cluster of hospitals, was applied in a recent study of patient choice behavior among California hospitals (Garnick, Luft, Robinson, et al. 1989).

Our approach incorporates elements of both. We focus on three hospitals, located in Fergus Falls (demonstration hospital), Breckenridge, and Alexandria, Minnesota. We use the travel patterns of residents surrounding these three hospitals as the basis of our market derivation. We thereby define the market area based on observed consumer (patient) choices. We do not deal with supply-side issues, such as hospital pricing behavior or its relation to market concentration. Arguments about market definitions based on rural patient flows to urban areas have been important in recent antitrust hearings (Medicine and Health Perspectives October 16, 1989) and in recent theoretical discussions (Werden 1989; Luft et al. 1989). The issue is far from resolved.

To operationalize our definition, zip codes in Minnesota, North Dakota, and South Dakota were examined to determine if they had Medicare discharges from any of these three large (over 75 beds) rural hospitals. All zip codes with at least three were identified. Then, any other hospital that served at least three discharges from its set of zip codes was drawn into the market area. We modified this initial set of zip codes (and hospitals) to exclude extraneous travel patterns. An example was local admissions of residents form distant states, which most likely occurred while these individuals were traveling. Road maps were also used to check the contiguity of the zip codes included. Subsequently, several noncontiguous resident-origin zip codes were deleted. This resulted in 93 zip codes and 53 hospitals. The map of the market area shows its general shape, location, and distribution of urban and rural hospitals.

This definition of the market area captured a significant proportion (90-95 percent) of the total discharges from the three larger rural hospitals and majority of total market area discharges. The hospital in Fergus Falls served (at least three discharges from) 40 zip codes, the Breckenridge hospital served 32 zip codes, and the Alexandria hospital, 28. The hospitals overlapped considerably in their service areas. The total market area produced 12,266 Medicare discharges during 1986. We note that of these 12,266 discharges, a certain percentage (less than 10 percent) were most likely not unique patients, that is, they represented persons with repeat hospitalizations during the year.

An initial review of the travel patterns was informative. Of the 12,266 discharges, we found that 60 percent chose their closest hospital. This percentage was larger if the closest hospital was a large facility; 79 percent of the patients chose their closest if it was a large rural hospital, whereas only 54 percent chose their closest if it was smaller. These descriptive statistics suggested two things. First, the characteristics of rural hospitals affected choice, and second, even when a larger rural hospital was nearby, a significant percentage still traveled to an urban hospital. These patterns also suggested that the 53 hospitals could be collapsed into meaningful categories to ease the estimation process.


A major goal of the overall evaluation (Adams, Wright, and Robbins 1989) was to develop a model to predict where Medicare inpatients would go if the rural hospital that they had previously chosen were to close, that is, if it was eliminated from their choice set. An important policy question was whether these "displaced beneficiaries" would choose an urban over a rural hospital after the closure. A corollary question was: If they chose a rural hospital, what were its characteristics?

Although we could address some of these issues simply by analyzing a dichotomous choice, urban versus rural, we wanted more information from the model. That is, we were interested in characteristics other than the metropolitan location of the hospital chosen, particularly those characteristics that had implications for service scope (e.g., size). As noted, we were also interested in the choices of Medicare beneficiaries among alternative rural hospitals remaining in the choice set. The demonstration hospital argued that it, as well as other larger, sophisticated rural hospitals, were a special "type" that served as a substitute for urban hospitals.

We began, then, by developing a hospital typology for the multinomial choice model. By collapsing the 53 hospitals into categories, we reduced estimated problems (and costs) but maintained more choice distinctions than simply urban versus rural.

After an initial analysis of patient flows, we decided on the following typology and choice set:

Hosptype 1 = Distant urban hospital

Hosptype 2 = Nearest large urban hospital (200 + beds)

Hosptype 3 = Nearest small urban hospital (<200 beds)

Hosptype 4 = Nearest rural center (75 + beds)

Hosptype 5 = Next-nearest rural center (75 + beds)

Hosptype 6 = Nearest rural community (<75 beds)

Hosptype 7 = Next-nearest rural community (<75 beds).

For each zip code we determined the hospital fitting each of these categories and tagged it for analysis. The nearest and next-nearest hospitals were those in the zip codes that were nearest and next-nearest to the zip code of residence for each discharge. For each discharge, only one hospital met the definition for Type 2, 4, 5, 6, or 7. With reference to the map, (Figure 1), the large and small rural hospitals are split into nearest and next-nearest.

There was more than one type 1 hospital, however, and possibly more than one Type 3 hospital for each discharge. Specifically, Type 1 included a set of large, urban hospitals in Minneapolis and some hospitals associated with the Mayo Clinic in Rochester to which a small number of beneficiaries traveled. These hospitals were, on average, over 100 miles away from the zip codes in the predominantly rural, three-hospital market area. It could be argued that these hospitals should not be included; the patient flows to each of these urban hospitals were small. Yet these flows were of interest since these individuals traveled "outside" their central market area for care and, we hypothesized, were more severely ill. The set of hospitals in Type 3 were all located in the Fargo or Moorhead area; there were more than one in the Fargo zip code and Moorhead's zip code was only four miles away. In parts of the analysis we treated these two hospital types (1 and 3) as if they represented a single hospital. For example, we used a discharge-weighted mean latitude and longitude to derive the distance measure.

Although collapsing the 53 hospitals into only seven hospital types might not have "accounted" for all discharges, 97 percent (of an original 12,645) were allocated. Table 1 shows the actual distribution of the
Table 1: Total Discharges within Market Area Using Each
Hospital Type, by Age Category, 1986
 Percent Patient Age
 Discharges of Total <75 230 85+
HOSPTYPE 1 858 7.0 600 230 28
HOSPTYPE 2 1,304 10.6 715 451 138
HOSPTYPE 3 1,325 10.8 669 457 199
HOSPTYPE 4 3,353 27.3 1,364 1,319 670
HOSPTYPE 5 241 2.0 115 80 46
HOSPYPE 6 4,566 37.2 1,658 1,824 1,084
HOSPTYPE 7 619 5.0 246 235 138
 ______ _____ _____ _____ _____
Total 12,266 100.0* 5,367 4,596 2,303
* Percent rounded.

total (12,266) discharges allocated. Although this typology was derived for this particular area and its travel patterns, it should be applicable to most rural areas; categories could be collapsed where necessary. Our market area is affected by the "draw" of the highly recognized expertise of personnel at the Mayo Clinic. Other rural market areas may not exhibit choices as far away as these, but a recent study indicated that many included a "distant urban center" (Morrisey, Sloan, and Valvona 1988).

The data in Table 1 highlight the role of urban hospitals in this "rural" market area. Of the more than 12,000 patients allocated among the seven types, over 28 percent received services at urban hospitals: 10.6 percent went to a nearby urban hospital of 200 + beds, 10.8 percent went to a smaller nearby urban hospital, and almost 7 percent traveled relatively far to obtain services at a very distant urban hospital.


The independent variables used in the analysis comprise three types: hospital characteristics (or attributes) that do not vary across patients, patient characteristics that do not vary across hospitals, and distance variables that vary with both individuals and hospitals. We describe our methods for deriving these variables in this section. The variable means and standard deviations are presented in Table 2. The means of the hospital characteristics are for the set of hospital falling into each type; the mean distance variable was based on all combinations of patients and hospital types and the means for patients were based on all discharges.

We used three hospital attributes in the analysis: the size of the hospital, its Guttan scale of service capacity, and whether it had teaching activity or not. Both the size and the scope of services have been found to affect a patient's choice of hospital positively and significantly; Cohen and Lee (1985) previously found teaching status to be insignificant. Bed size is measured by the number of acute care beds. Since the focus is on inpatient services reimbursed under Medicare's prospective payment system, we omitted long-term care beds.

The Guttman scale is based on the principle of cumulative scaling, which takes advantage of the tendency of hospitals to acquire service capabilities in a predictable sequence. Scales of this type were developed and compared to other measures in earlier studies (Edwards, Miller, and Schumacher 1972; Cohen and Lee 1985). The Guttman scale starts with a list of significant services and specialized units, and searches for those items which in a consistent sequence, eliminating items that always occur jointly. Each hospital is scored by its highest-ranking (least common) service.

A scale of 17 specific items was used for the hospitals in our market area; details are presented elsewhere (Adam, Wright, and Robbins 1989). The scales were developed iteratively, using a SAS program to include and exclude services with the goal of representing diversified frequency of occurrence as well as internal consistency.

Data used in deriving this scale were self-reported in the 1986 Annual American Hospital Association Survey. Services ranged in prevalence from emergency services, offered by 98 percent of the hospitals, to specialized intensive care units, ICUs, reported by only 9 percent. As the means in Table 2 show, the urban hospitals had higher Guttman scales than the rural facilities did, as would be expected. The larger rural hospitals also tended to have a higher scope of service as measured by this scale. As shown, the urban hospitals were also more likely to have some level of teaching activity. Since size and Guttman scale tend to be correlated ([R.sup.2] = .55), it was unclear whether each would have an independent effect on hospital choice. We tested for the residual effect on Guttman scale and size in earlier specifications.


The measure of distance we used was miles "as the crow flies" from the centroid of each zip code of residence to that of the hospital. Although travel time might have been a more desirable measure, the required data were not readily available. Road distance might even have been preferred, but it also has drawbacks. While it is generally longer than straight-line distance (Williams et al. 1983), it does not account for traffic congestion or other barriers to travel. In any case, straight-line distance is a good approximation to road distance in some rural areas. Review of the road configuration in our market area showed flat terrain, with a major diagonal highway connecting two of the large rural hospitals and three major urban areas. In particular instances, road maps were used to compare "crow-fly" miles to road mileage and the distances were found to differ by 1-5 miles; some hand corrections were made. In estimating the effect of distance on choice, we tried distance, long of distance, and distance squared as independent variables.

Two patient characteristics used were age and DRG category at discharge. We tested for nonlinearities in the age effect by defining dummy variables for age categories. As shown in Table 2, the mean age was 76 years for beneficiaries discharged from the 53 hospitals in the market area. Three DRG group indicators were defined for medical, surgical, and psychiatric DRGs. These groupings followed those presented in the Federal Register (September 1, 1987). This variable simply reflected whether the treatment was medical, surgical, or psychiatric in nature. The majority of discharges in the market area (69 percent) were medical in nature.

The measure of severity used in this study was recently developed under contract to the Perspective Payment Assessment Commission (ProPAC) (Houchens, Conklin, and Briscoe 1989). A fuller discussion of the methods used is provided in the Appendix. This measure reflects both the severity of the DRG and the severity of cases within the DRG. It is derived from the Disease Staging methodology, which classified patients based on the etiology and stage of disease progression. This methodology first identifies a person's disease status in terms of four general stages: (1) initial onset of disease, (2) problems limited to an organ or systems but not progressed enough to exhibit increased risk of complication, (3) multiple site involvement, and (4) death. Each stage reflects increased severity in terms of risk of death or residual impairment.

These general stages were then subdivided into further stages for each individual disease [e.g., appendicitis has two substages under Stage 3; septicemia (3.1) and shock (3.2)]. Substages were defined within over 400 diseases and crosswalked to the appropriate DRGs. These substages were further divided into more discrete cells based on the presence or absence of comorbidities. Over 21,000 such cells were formed using the 1986 discharge data. To compare stages, or substages, across DRGs, we needed some means of calibration; this calibration underlies the derivation of the severity measures.

Essentially, the calibration method consisted of deriving weights much like the relative "cost" weights for DRGs in HCFA's case-mix index (CMI). These weights were based on average impatient standardized charges. These dollar values were first assigned to each cell, using the 100 percent MEDPAR file and national averages. A within-DRG score was derived by dividing the individual cell scores by the average within the DRG. The combined score, used here, normalizes the individual cell scores across all DRGs, thus accounting for the relative severity of the DRG as well as that of discharges in the cells within each DRG. The strictly within-DRG score would be appropriate if we were testing the model on discharges specific to only one DRG. The means of the combined severity score for all Medicare beneficiaries in the market area was 1.14. The means of the strictly within-DRG as well as the combined measure for rural patients traveling to each of the seven "Hosptypes" are presented in Table 3.
Table 3: Measures of Patient Relative Severity for Rural
 Discharges from Each HOSPTYPE
 Combined Within-DRG
 Severity Score Severity Score
HOSPTYPE 1 1.654 1.014
HOSPTYPE 2 1.317 0.996
HOSPTYPE 3 1.325 1.000
HOSPTYPE 4 1.144 0.978
HOSPTYPE 5 1.271 0.999
HOSPTYPE 6 1.024 0.978
HOSPTYPE 7 1.060 0.999


Our results were generally consistent with findings of earlier studies, and yet they provide new insight on the behavioral determinants of rural patient travel and hospital choice. In particular, the severity measure was found to affect the choice of hospital type. Important patterns by age were also found. Overall, the equations have fairly good predictive power. We present two measures, the pseudo [R.sub.2] and the percent of discharges correctly classified. The value of the [R.sub.2] is in line with that obtained in other studies of hospital choice ranging from .23 to .44 in the regressions reported. (Other regressions not reported here are available from the authors upon request.) The percent correctly classified is a little over 60 percent in almost all equations. This latter measure reflects the percentage correctly allocated when discharges are assigned to the Hosptype with the highest predicted probability. Although this measure is well accepted in the literature, a better prediction is derived when the percentage going to each Hosptype is based on the predicted partial probabilities of the individual going to each Hosptype. That is how the equation was used in our simulations (Adams, Wright, and Robbins 1989).


The findings in Tables 4-7 confirm the strong deterrent of increased distance, or increased time price, on the choice of a hospital among alternatives. If we consider two hospitals equally attractive in size, scope of service, and teaching attributes, the odds of the rural Medicare beneficiary going to one ten miles farther away are approximately 50 percent lower (using a coefficient of -0.0619 on non-log distance from regressions not reported here). Thus, Medicare beneficiaries located in remote rural areas strongly prefer to travel shorter distances to obtain care, all other factors held constant.

The results on distance are fairly stable across specifications. When the relationship of distance to hospital choice is modeled by the log of distance, the odds are nearly proportional to the inverse of distance. Although the coefficient, -1.04 presented in Table 4, is somewhat what smaller (in absolute magnitude) than that reported earlier (Garnick, Luft, Robinson, et al. 1989), these cannot be strictly compared.


All hospital attributes included in the equations were significant and positive in their effect on individual hospital choice. The specific coefficients estimated are shown in several alternative specifications in Tables 4-7. The estimated impact of each variable on the odds of hospital choice can be found by exponentiating the estimated coefficient. The coefficient on size, for example, is fairly stable across specifications and indicates that an increase in size of ten beds increases the odds of choosing a hospital by about 1.7 percent. However, this interpretation only applies within hospital groups with the same bedsize restriction, because the dependent variable is partially based on bed size.

The effect of the Guttman scale on hospital choice indicates that beneficiaries attach a positive value to greater scope of service, holding all other factors constant. For two hospitals differing by 10 on the Guttman scale, the estimated odds of choosing the one with greater scope of service is 39 percent higher (based on a coefficient of .03). Differences on the Guttman scale of this magnitude are difficult to achieve and exist in our data set only between the smallest rural hospital and the larger urban ones. A difference in the Guttman scale of 1.00, which is more descriptive of the variation across the urban hospital types, increases the odds of choosing the more sophisticated hospital by an estimated 3 percent.

The results indicate that teaching hospitals are preferred to non-teaching hospitals. The magnitude of this effect is quite large, as measured in the tables, and remains positive and significant throughout specifications.


An important (but not surprising) finding was that the relative severity of the Medicare beneficiary's illness does affect hospital choice. All of the results shown indicate that more severely ill patients in rural areas choose urban over rural hospitals more often than less severely ill patients, holding constant the size, scope of service and teaching status. Each 1.0 unit increase in patient complexity relates to an increase in expected resource use equal to the average amount of resources used over all discharges (about $5,000). An increase of 1.0 in the complexity we found to nearly triple the odds of choosing Hosptype 1 over Hosptype 7 in the first regression run (not reported here) which omitted diagnostic category. The more severely ill in this rural area were also more likely to choose Hosptype 1 over Hosptype 4, although the difference was smaller; an increase of 1.0 in across-DRG severity approximately doubled the odds of choosing Hosptype 1 over Hosptype 4 in this first regression run.

The effect of higher severity on choices within the set of rural hospitals in this earlier regression was also somewhat interesting. These effects can be estimated from the differences between the coefficients. The more severely ill chose rural Hosptype 5 over Hosptype 4, or, more often than those less severely ill, traveled past the closest rural hospital with more than 75 beds. Similarly, the more severely ill tended to choose rural Hosptype 7, the farther-away small rural hospital, over Hosptype 6 more often than the less severely ill did. (This effect was not statistically significant.) These results suggest that the severity of illness of Medicare beneficiaries affects choices both within the set of rural hospitals and across the entire market area.

In the second regression run (not reported here), we accounted for the effect of different diagnostic categories on the choice of Medicare beneficiaries among the various hospital types while still using non-log distance. Although the DRGs could be categorized into many sub-groups, we estimated the difference in being in a medical or surgical versus being in a (the omitted) psychiatric diagnostic category. The results indicated that those with psychiatric diagnoses were more likely than those with medical diagnoses to choose urban hospitals over rural ones. This is indicated by the negative and significant coefficients for those in medical DRG categories in the regressions reported. One study asserted that increased distance may be a positive asset for those seeking mental health services, as it helps preserve anonymity (Dear 1977); another found that residents traveled farther for surgery and cardiac treatment than for obstetric care (Kane 1969). There were fewer differences across these DRG groupings in the types of rural hospitals chosen, although those with surgical DRGs were more likely than those with psychiatric diagnoses to choose the closest large rural hospital.

When the diagnostic category is included, complexity is no longer statistically significant in the odds of choosing Hosptype 4, 5, or 6 over Hosptype 7. That is, the complexity of illness of Medicare beneficiaries significantly increased the odds of choosing an urban over a rural hospital but not of choosing a larger rural over a smaller rural hospital, once the nature of the treatment was accounted for. The results in Table 4 indicate that the increase in the odds of choosing Hosptype 4 over 7 may be due to the need for surgical services, not necessarily to the severity of illness. Correlation of these two measures, however, makes it difficult to separate these effects.


The results by age and age categories indicate that the more elderly a rural patient is, the stronger their preference for rural over urban hospitals, independent of the time price or other factors affecting choice. In Table 4 the signs indicated that increased age reduced the odds of choosing almost any hospital type (relative to Hosptype 7) except Hosptypes 4 and 6. This suggests that older Medicare beneficiaries are more likely to choose a small rural or nearby large rural hospital over other types.

In other regressions not reported here we found significant differences across each age category (<75, 75-84, and 85+) in the odds of choosing Hosptype 1-3 over Hosptype 7. Younger beneficiaries were more likely than older ones to choose an urban hospital over a rural one. To test for the interactive effects of age with all other variables, we ran separate equations for each group.

These results are shown in Tables 5, 6, and 7. There are interesting differences in the relative values of the coefficients for these groups. For example, as seen in Table 7, those over age 85 perceive increased hospital size as a negative attribute. There are perhaps increased administrative hassles and less personal treatment in larger hospitals which the oldest old do not find desirable. The results are different with respect to the role of diagnosis and its severity on hospital choice for the three age groups. For those under 75, complexity of illness rather than diagnostic category (these variables were insignificant) influences hospital choice. Those with more complex illnesses in this group are more likely to choose an urban over a rural hospital, and to choose a large over a small rural hospital. Complexity also increases the probability of traveling past the closest large rural hospital, or of choosing Hosptype 5 over Hosptype 4, for those under age 75.

For those 75 to 84, increased complexity and medical diagnosis influence the choice of urban over rural hospitals, but not that of large over small rural hospitals, while for those 85 and over, complexity is not a factor. For both of these age groups, those with a medical diagnosis are less likely to seek access to a distant urban hospital over their closest small rural facility. Also, for both groups the probability of choosing the nearest large rural hospital over the nearest smaller facility is significantly increased when they have a surgical diagnosis. Finally, even within these age groups, the older the Medicare beneficiary, the less likely he or she is to choose an urban hospital over the closest small rural facility.


The estimation of this hospital choice model for Medicare beneficiaries leads to several applications and implications. One application, as noted, is the use of such an equation to predict rural Medicare patients' choice of hospitals as if the "choice set" has been reduced. Distances to the (hypothetically) remaining hospitals in the market area can be calculated and new choices predicted. This prediction relies on an assumption of irrelevant alternatives; problems with this assumption when non-distinct choices are made are somewhat moderated in our model by the use of choice-specific constants (Hausman and Wise 1978). The specifications presented here are applicable to rural closures and were used in our larger study to estimate the change in travel time and Medicare outlays if the demonstration hospital were to close (Adams, Wright, and Robbins 1989). Overall, we predicted that discharges from this hospital would average an additional 20 miles of travel, should the hospital close. This increase would average 27 miles for those under 85 versus 11 miles for those over 85.

Alternative specifications could be developed for other rural or urban hospital market areas. In urban areas this would most likely require less aggregation, or perhaps a different typology of the hospital choice set since urban residents tend overwhelmingly to choose urban hospitals (Morrisey et al. 1988). In either urban or rural settings, such predictions can only be taken as estimates of short-run effects since alternative hospitals can alter their service provisions to accommodate the displaced patients. Simulations of hospital closure and alternative choices could be refined further if dynamic adjustments were included.

Regardless of how the hospital choice equations are used, our results indicate that they are misspecified if relative severity and complexity of illness is not accounted for. Our results indicate that physicians and patients make decisions about hospital choice based not only on the severity of the DRG, but on severity within DRGs as well. The finding that the more complex rural cases travel to urban hospitals may have implications for payment policy, since the prospective payment system (PPS) does not currently adjust urban versus and payments for within-DRG severity. If urban hospitals end up serving a relatively more complex caseload (inclusive of the within-DRG measure), Medicare payments may not be adequate. On the other hand, the finding that rural hospitals are less likely to be used by these complex cases may be indicative of past payment policies. Lower standardized ' amounts for rural facilities may have perpetuated differences in the relative sophistication of personnel and services in rural versus urban areas.

Indeed, a recent study concluded that women in rural areas in need of obstetrical care view their local rural hospitals as "inferior goods" in the economic sense (Bronstein and Morrisey 1990). As incomes rise, more rural women "migrate" to urban hospitals for care. Some argue, however, that these and other findings are not indicative of competition between urban and rural hospitals, since the migration is "one way." Rather, they simply imply that rural hospitals do not provide the needed services (Werden 1989). The question for rural hospital administrators is whether it is the set of services, their quality, or their perceived quality that is shaping these decisions.

Currently, there is opportunity for such changes, since rural hospitals will receive increased payments under PPS as the standardized amounts are increased to equal those for "other urban" hospitals. The recently enacted Essential Access Community Hospital program will also offer funds for the reconfiguration of hospital services in rural areas. What rural hospitals can do to alter services or service quality, or both, seems an important question for their survival. They may, in light of these studies, consider the feasibility of converting to specialized uses (e.g., rehabilitation, psychiatric care, substance abuse) or updating service capacity in specific areas, such as obstetrical care, or both. An important question is how successful such diversification efforts have been as a survival strategy for other hospitals.

Before diversifying, a rural hospital should consider the characteristics of its service population and what this indicates about the services needed or demanded. Is the service population relatively young, so that improvement of obstetrical sophistication will attract new admissions? Or is the service population older, with more chronic care needs? Our analysis suggests that rural elderly choices are so influenced by age that the market might be viewed as segmented. Apparently, the provision of specialized services in large urban centers does not attract the oldest old. Those over 85 in our market area exhibit preferences for a greater scope of service but in a smaller, rural facility. Those over 85 also tend to seek surgical services closer to home. Conversely, those under 75 prefer a higher scope of service and size. This information will be informative as rural hospitals alter their service mix.

Our findings by age also have implications for the effects of closures. The analysis indicates that responsiveness by the elderly to increased (time) price is significant. When a rural hospital closes, the (time) average price is increased. Unless new services are provided, whether in the short- or long-run, in the same location, this (time) price has certain implications for consumer welfare and access. If lost services exist elsewhere, or a remaining hospital increases its supply of them, the consumer suffers only the loss from increased (time) price. But if lost services are not made available, consumer welfare may be further reduced. The analysis indicates that these effects may be greater for the older Medicare beneficiary than for the younger Medicare beneficiary.



The measure of patient severity was based on empirical analyses of patient-level standardized charges in the 1986 100 percent MEDPAR data (excluding statistical outliers). The methodology used to measure within-DRG severity consisted of the steps now explained.

The 1986 MEDPAR sample discharges were first subclassified within each DRG. Discharges were then subclassified by the Principal Disease Stage and the number of unrelated comorbidities (0, 1-2, 3+). See Figure A.1. The 1986 data filled 21,907 cells in this classification system, for an average of 47 cells per DRG. Cells were empty either because they represented an impossible combination, such as Principal Disease 616 (Appendicitis) within DRG 78 (Pulmonary Embolism), or because the cell represented a combination that seldom occurs in the Medicare population.

The objective was to derive a within-DRG complexity weight for each cell that was proportional to the mean of standardized charges in each cell. The simplest such measure would normalize the sample mean charge in each cell (excluding statistical outliers) by dividing each cell sample mean by its total DRG mean. This measure would thus measure the complexity of patients in each cell in terms of the average total resources required to treat patients in that cell relative to the average total resources required to treat patients in the DRG.

The main problem with this approach is that many cell sample means would be statistically unreliable since they were based on small numbers of observations. When there were too few observations, an alternative method was used. We implemented a simple solution that involved a compromise between using the (unreliable) cell mean and replacing the cell mean with the DRG mean. We used a weighted average of the cell mean and an estimate of the DRG mean as follows:

estimated cell mean = p * (simple cell mean) + (1 - p) * (DRG mean)

Here p has a value between zero and one. If the simple cell mean was highly reliable (i.e., had a low standard error), then p was assigned a value near 1 and the estimated cell mean was set nearly equal to the simple cell mean. If the simple cell mean was highly unreliable (i.e., had a high standard error), then p was assigned a value substantially less than 1 and the estimated cell mean was drawn toward the estimated DRG mean.

While the amount of shrinkage is a statistical issue, the direction of shrinkage (i.e., the choice of overall mean to shrink toward) may be decided largely on nonstatistical grounds. The best strategy, then, was to identify homogeneous groups of cells and to shrink their means toward a common value. All cells within a DRG would seem to form a homogeneous group (after all, they are "diagnosis related"). Moreover, the focus of this study was to develop measures within DRGs. Therefore, each cell mean was shrunk toward an estimate of its DRG mean.

The patient's within-DRG complexity index, then, was derived simply by dividing the cell mean by the DRG mean. Each patient's within-DRG index was multiplied by the DRG relative weight (based on standardized charges, as HCFA's relative weights), since these weights could be considered "between-DRG" complexity indexes. Therefore, our combined severity index measured the overall patient severity, taking into account both within-DRG complexity and between-DRG complexity.


Acton, J. P. "Demand for Health among the Urban Poor with Special Emphasis

on the Role of Time." In The Role of Health Insurance in the Health

Services Sector. Edited by R. N. Rosett. New York: National Bureau of

Economic Research, 1976. Adams, E. K., G.E. Wright, and J. Robbins. Final Report of Evaluation of Rural

Secondary Specialty Center Demonstration Project. HCFA Contract no

500-87-0028-6. Princeton, NJ: Mathematica Policy Research, Inc.,

June 1989. Bronstein, J. M., and M. A Morrisey. "Determinants of Rural Travel Distance

for Obstetrics Care." Medical Care 28, no. 8 (August 1990):

853-65. Coffey, R. M. "The Effect of Time Price on the Demand for Medical Care

Services." The Journal of Human Resources 18, no. 3, (1983): 407-24. Cohen, M. A., and H. L. Lee. "The Determinants of Spatial Distribution of

Hospital Utilization in a Region." Medical Care 23, no. 1 (1985): 27-38. Dear, M. "Locational Factors in the Demand for Mental Health Care." Economic

Geography 58, no. 1 (1977): 233-42. Dranove, D., W. D. White, and L. Wu. "An Analysis of Admissions Patterns

in Local Hospital Markets." Paper presented in Harvard/Massachusetts

Institute of Technology/Boston University Economics Seminar Series.

Boston, MA, February 1989. Edwards, M., J. D. Miller, and R. Schumacher. "Classification of Community

Hospitals by Scope of Service: Four Indexes." Health Services Research

7, no 2 (1972): 301-23. Elzinga, K., and T. Hogarty. "The Problem of Geographic Market Delineation

Revisited: The Case of Coal." The Antitrust Bulletin 23, no. 1 (1978):

1-18. Garnick, D. W., H. S. Luft, J. C. Robinson, and J. Tetreault. "Appropriate

Measures of Hospital Market Areas." Health Services Research 22, no. 1

(April 1988): 69-89. Garnick, D. W., E. Lichtenberg, A. S. Phibbs, H. S. Luft, D. J. Peltzman,

and S. J. McPhee. "The Sensitivity of Conditional Choice Models for

Hospital Care to Estimation Technique." Journal of Health Economics 8,

no. 1 (1989): 377-97. Griffith, J. R. Measuring Hospital Performance. Chicago, IL: Blue Cross Association,

1978. Hausman, J. A., and D. A. Wise. "A Conditional Probit Model for Qualitative

Choice: Discrete Decisions Recognizing Interdependence and Heterogeneous

Preferences." Econometrica 46, no. 2 (1978): 403-26. Hogan, C. "Patterns of Travel for Rural Individuals Hospitalized in New

York State: Relationships between Distance, Destination and Casemix."

The Journal of Rural Health 4, no. 2 (July 1988): 29-41. Houchens, R. L., J. E. Conklin, and W. W. Briscoe. Measures of Complexity of

Illness within DRGs. Final Report, Prospective Payment Contract no.

T-47540316; Task Order #9. Washington, DC: ProPAC, March 1989. Kane, R. L. "Determination of Health Care Priorities and Expectations

among Rural Consumers." Health Services Research 4 (Summer 1969):

142-51. Lembcke, P. "Measuring the Quality of Medical Care through Vital Statistics

Based on Hospital Service Areas; Comparative Study of Appendectory

Rates." American Journal of Public Health 42 (March 1952): 276-86. Luft, H. S., A. S. Phibbs, D. W. Garnick, and J. C. Robinson. "Rejoinder to

Dranove and Shanley." Journal of Health Economics 8, no. 1 (1989):

479-83. Luft, H. S., D. W. Garnick, D. H. Mark, D. J. Peltzman, A. S. Phibbs, E.

Lichtenberg, and S. J. McPhee. "Does Quality Influence Choice of

Hospital?" Journal of American Medical Association 263, no. 1 (6 June 1990):

2899-2906. Maddala, G. S. Limited-Dependent and Qualitative Variables in Econometrics. Econometric

Society Monographs, No. 3. Cambridge, United Kingdom:

Cambridge University Press, 1983. McFadden, D. "The Measurement of Urban Travel Demand." Journal of Public

Economics 3, no. 2 (1974): 303-28. Morrisey, M. A., F. A. Sloan, and J. Valvona. "Defining Geographic Markets

for Hospitals and the Extent of Market Concentration." Law and Contemporary

Problems 51, no. 2 (September 1988): 165-94. Stigler, G., and R. Sherwin. "The Extent of the Market." Journal of Law and

Economics 28, no. 2 (1985): 555-85. Werden, G. J. "The Limited Relevance of Patient Migration Data in Market

Delineation for Hospital Merger Cases." Journal of Health Economics 8,

no. 1 (1989): 363-76. Williams, A. P., W. B. Schwartz, J. P. Newhouse, and B. W. Bennet. "How

Many Miles to the Doctor?" New England Journal of Medicine 311, no. 12

(20 October 1983): 958-63.
COPYRIGHT 1991 Health Research and Educational Trust
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1991 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Adams, E. Kathleen; Houchens, Robert; Wright, George E.; Robbins, James
Publication:Health Services Research
Date:Dec 1, 1991
Previous Article:Responses to prospective payment by rural New Mexico hospitals.
Next Article:Uncompensated care before and after prospective payment: the role of hospital location and ownership.

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