How Much Is Postacute Care Use Affected by Its Availability?

Abstract

Objective

To assess the relative impact of clinical factors versus nonclinical factors—such as postacute care (PAC) supply—in determining whether patients receive care from skilled nursing facilities (SNFs) or inpatient rehabilitation facilities (IRFs) after discharge from acute care.

Data Sources and Study Setting

Medicare acute hospital, IRF, and SNF claims provided data on PAC choices; predictors of site of PAC chosen were generated from Medicare claims, provider of services, enrollment file, and Area Resource File data.

Study Design

We used multinomial logit models to predict PAC use by elderly patients after hospitalizations for stroke, hip fractures, or lower extremity joint replacements.

Data Collection/Extraction Methods

A file was constructed linking acute and postacute utilization data for all medicare patients hospitalized in 1999.

Principal Findings

PAC availability is a more powerful predictor of PAC use than the clinical characteristics in many of our models. The effects of distance to providers and supply of providers are particularly clear in the choice between IRF and SNF care. The farther away the nearest IRF is, and the closer the nearest SNF is, the less likely a patient is to go to an IRF. Similarly, the fewer IRFs, and the more SNFs, there are in the patient's area the less likely the patient is to go to an IRF. In addition, if the hospital from which the patient is discharged has a related IRF or a related SNF the patient is more likely to go there.

Conclusions

We find that the availability of PAC is a major determinant of whether patients use such care and which type of PAC facility they use. Further research is needed in order to evaluate whether these findings indicate that a greater supply of PAC leads to both higher use of institutional care and better outcomes—or whether it leads to unwarranted expenditures of resources and delays in returning patients to their homes.

Postacute care (PAC) was the fastest growing sector of the Medicare program throughout the early to mid-1990s. A number of factors including payment incentives, advances in drug treatments and surgical techniques, and improvements in outpatient care contributed to shorter lengths of stay in acute care hospitals and corresponding increases in PAC use. As more hospitalized patients transfer to PAC, the need to better understand the factors driving such transfers is growing.

Patients can access PAC services in many settings including skilled nursing facilities (SNFs), inpatient rehabilitation facilities (IRFs), and patients' homes with services from home health agencies (HHAs).1 IRFs provide intensive rehabilitation (three or more hours a day of therapy) in an inpatient setting. SNFs can also provide inpatient rehabilitation under the Medicare benefit, although it is generally less intensive than that provided in an IRF (Gage 1999). Home health care agencies provide therapy, nursing care, and assistance from home health aides.

In many instances, referrals to these settings are made in the absence of clear clinical criteria that would identify the best PAC setting for maximizing outcomes. Although studies have explored variations in outcomes across settings for stroke and hip fracture patients, there is a dearth of research that explains which patients are most appropriate for each PAC setting (Kane 1997; Kramer et al. 1997; Kane et al. 2000). Thus patients and doctors must weigh a range of clinical and nonclinical factors—such as the perceived quality of care delivered by a PAC provider and its convenience—when making these decisions.

In addition, admissions to PAC are often guided by a hospital discharge planner and PAC providers play a role in deciding which patients to accept. Although Medicare PAC eligibility criteria are codified in regulations, as a practical matter PAC providers, physicians, and hospital discharge planners have discretion in interpreting these guidelines. In fact, researchers examining PAC have observed tremendous variation in utilization rates, geographically and by type of discharging hospital (Benjamin 1986; Neu, Harrison, and Heilbrunn 1989; Swan and Benjamin 1990; Kenney and Dubay 1992; Kane et al. 1996; Schore 1996; Cohen and Tumlinson 1997; Kane, Lin, and Blewett 2002; MedPC 2003).

All of this suggests that a variety of nonclinical factors are likely to affect where patients go for PAC. Previous research has noted the importance of the supply or availability of PAC in an area on rates of use (Neu, Harrison and Heilbrunn 1989; Swan and Benjamin 1990; Kenney and Dubay 1992; Kane et al. 1996; Cohen and Tumlinson 1997; MedPAC 2003). This study develops more refined methods of measuring PAC availability and assesses the relative impact of clinical versus nonclinical factors, especially availability, in determining where patients go for PAC services.

Determinants of PAC Use

Researchers have found a number of patient-level, provider-specific, and area factors that affect the use of PAC and choice of PAC settings. Demographic and clinical factors including age, gender, race, marital status, functional status, history of disability, medical condition, and comorbidities influence the sites to which patients are discharged (Neu, Harrison, and Heilbrunn 1989; Manton et al. 1993; Steiner and Neu 1993; Blewett, Kane, and Finch 1995; Lee, Huber, and Stason 1997; Kane et al. 1998; Liu, Wissoker, and Rimes 1998; Gage 1999; Bronskill, Normand, and McNeil 2002; Finlayson 2002; McCall et al. 2003; MedPAC 2003). Use of PAC is generally positively associated with age and negatively associated with being married, presumably because patients' spouses often serve as informal caregivers (Kane et al. 1994; Liu, Wissoker, and Rimes 1998; Gage 1999; Shatto 2002). Primary and comorbid diagnoses affect decision making with respect to patient suitability for one site of PAC over another. For example, researchers have found that use of PAC was highest among people with Alzheimer's and Parkinson's, diseases that require a high level of clinical monitoring and assistance (Liu, Wissoker, and Rimes 1998). Living alone and functional dependency at discharge from inpatient care were also significant predictors of PAC (Kane et al. 1996; McCall et al. 2003).

Additional factors that influence use of PAC include hospital-level predictors such as the volume of Medicare patients served, hospital size, percent low-income patients, ownership, and status as a teaching hospital (Neu, Harrison, and Heilbrunn 1989; Steiner and Neu 1993; Blewett, Kane and Finch 1995; Bronskill, Normand, and McNeil 2002). Although the effects of these characteristics depend on the condition studied and the patient variables included in the analysis, more than one study found that discharge from teaching hospitals and hospitals with high-Medicare volume was associated with greater use of PAC.

Researchers have also identified a number of area-level predictors of PAC use. For example, researchers have found that higher-income communities have higher utilization rates of SNF and home health care (Neu, Harrison, and Heilbrunn 1989).

Finally, prior research has noted the influence of the supply of PAC on utilization, a finding consistent with research on use of other types of care (Gatsonis et al. 1995; Kane et al. 1996; Pritchard et al. 1998; Fisher et al. 2000). A positive correlation was found between the home health use and the number of home health agencies in an area, and a negative correlation was found between home health use and the number of nursing home beds per capita in some studies (Swan and Benjamin 1990; Kenney and Dubay 1992; Liu, Wissoker, and Rimes 1998; MedPAC 2003). Characteristics of discharging hospitals that may affect the ease of referrals to PAC, including ownership of a PAC facility, can boost PAC use (Young 1997; MedPAC 2003).

Although research has noted the effects of PAC supply on use, relatively little attention has been paid to the measurement of PAC supply. Prior studies have relied on simple counts of PAC providers and/or counts of PAC beds within geopolitical boundaries, such as counties or metropolitan statistical areas (MSAs), which may not capture the variation in accessibility or availability of PAC within these areas. In this study, we developed a more detailed and comprehensive approach to measuring PAC supply, and we determined which factors most affected the use of PAC services by Medicare beneficiaries in 1999.

Conceptual Framework

We conceptualized the decision to use PAC as a joint decision made by a hospitalized patient, his/her family, and his/her physician(s), and influenced by discharge planners at the acute care hospital and admission staff at PAC sites. Clinicians involved in the decision consider medical and rehabilitation needs when referring some types of patients to PAC, but clinical evidence is not available for all patient types. For those patients falling into “gray areas” in which there are no clinical norms, patient preferences, local practice patterns, PAC availability, and psychosocial factors play stronger roles. Thus, patient and family preferences and circumstances—such as whether or not patients have caregivers at home or are eligible for Medicaid-covered custodial nursing home care—are likely to influence the decision. In addition, factors such as the experience of the discharge planning staff and the financial pressure on the hospital to discharge the patient quickly may affect PAC use.

Finally, the overall attractiveness of the PAC options in the area and the availability of facilities willing and able to accept the patient come into play.2 Some areas have many IRFs competing to admit patients, while others have few. Similarly, there are areas in which SNF beds are rarely vacant and others in which SNFs actively market their services to discharge planners. Hospitals with IRF and/or SNF subproviders might find it easier to place their patients in those related facilities. Patient and family preferences for receiving care close to home can also affect PAC use.

Drawing on this framework, our overall analytic approach was to define relatively clinically homogenous populations that had high rates of PAC use and then build models using the factors hypothesized to influence whether they used an institutional PAC and if so, of what type.

Methods

Data Sources

We linked administrative data from a 100 percent sample of Medicare acute hospital, IRF, and SNF claims so we could observe choices of institutional PAC by our sample patients. We then drew on Medicare claims data, provider of services file data, enrollment file data, and data from the Area Resource File in generating predictors of site of PAC chosen.

Population Studied

We examined the use of PAC by three groups of Medicare patients discharged from acute care hospitals in 1999. We chose 1999 both because of data availability and because it is the only recent year during which no new PAC payment systems were implemented. We focused on the three largest patient groups using PAC: stroke patients; hip fracture patients; and lower extremity joint replacement patients. These conditions account for approximately 7 percent of Medicare acute discharges and one-quarter of discharges to PAC. Hip fracture was defined using an acute inpatient principal diagnosis of “fractures of the neck of the femur” (diagnosis codes 820.xx): hip fracture patients whose fractures could be because of bone metastases or who suffered major trauma to a site other than a lower extremity were excluded from the sample. Stroke was defined as intracerebral hemorrhage (431.xx), occlusion and sterosis of precerebral arteries with infarction (433.x1), occlusion of cerebral arteries with infarction (434.x1), and acute but ill-defined cerebrovascular disease (436.xx). Lower extremity joint replacement was defined using the Diagnosis Related Groups for joint replacement procedures (209, 471) excluding patients classified as hip fracture and those with reattachment procedures (84.26, 84.27, and 84.28).

We excluded certain groups of patients from our analyses. Patients who died in the hospital or within 30 days of discharge were dropped since their use of PAC was effectively truncated, as were patients for whom we did not have complete claims data.3 We restricted our sample of discharges to a beneficiary's first discharge for any given condition during 1999. Finally, we excluded patients who were residents of nursing homes at the time of their admission to acute care, since we hypothesized that these patients would most likely return to the nursing home after discharge from acute care without considering other PAC alternatives.4

Measures

Our dependent variable was the first PAC site used after discharge from an acute care hospital. We considered PAC use to be IRF or SNF care that began within 30 days of discharge from acute care and was covered by Medicare.5 We focused on use of institutional PAC because we were unable to distinguish patients returning to their homes from those sent to receive custodial nursing home care—that is, we did not have data on nursing home stays not paid for by Medicare. We grouped care delivered in swing beds with SNF care. Each of these types of care was defined using Medicare provider numbers and/or claim types. Patients who were readmitted to the hospital during the 30-day window were kept in the sample but acute care was not counted as a PAC site. Although Medicare rules allow SNF patients to delay entry for more than 30 days after their acute discharge, this did not affect our analyses: 97.3 percent of SNF patients in our sample began SNF care within 30 days of discharge if they used it at all.

We assembled, and included as independent variables in our models, a wide array of indicators of clinical, individual, discharging hospital, and PAC supply factors that might affect PAC choices.

Individual Predictors

We identified a number of patient-level characteristics hypothesized to affect use of PAC care and type of PAC used. To allow for nonlinear effects of age on PAC use in our models we classified patients into 3-year age bands. We also included gender, race, and place of residence (defined as an MSA, an area adjacent to an MSA, or rural area/not adjacent to an MSA using the county classification developed by the U.S. Department of Agriculture) in our analyses. All of these patient-level predictors were created using fields on the inpatient claims. In addition, we used the Medicare Denominator file to create indicators for whether patients were receiving Medicaid at the time of their acute admission or within 4 months of discharge.

Clinical Predictors

To capture the complexity of patients at the time of hospital discharge, we included a large set of comorbidities and complications tailored to our stroke, hip fracture, and joint replacement patients. The comorbidities used in our analyses were the chronic conditions identified by Iezzoni et al. (1994) as conditions that are nearly always present prior to hospital admission and hence are extremely unlikely to represent complications arising during the hospitalization. They included primary cancer with poor prognosis, metastatic cancer, chronic pulmonary disease, coronary artery disease, congestive heart failure, peripheral vascular disease, severe chronic liver disease, and diabetes mellitus with and without end-organ damage, chronic renal failure, nutritional deficiencies, dementia, and functional impairment.

The second type of case-mix variable was complications that were likely to have arisen during the hospital. To develop this list, we adapted the list of complications developed by Iezzoni et al. (1994), keeping only the complications that were likely to have a continued effect after hospital discharge, and therefore could influence the choice of PAC site (e.g., we excluded transient metabolic derangements and side effects of medications). In addition, we augmented the list to include some important complications for the Medicare population that had been omitted from Iezzoni's list. The resulting list of complications included postoperative pulmonary compromise; postoperative gastrointestinal hemorrhage; cellulitis or decubitus ulcer; septicemia; pneumonia; mechanical complications because of a device, implant, or graft; shock or arrest in the hospital; postoperative acute myocardial infarction (AMI); postoperative cardiac abnormalities other than AMI; procedure-related perforation or laceration; venous thrombosis and pulmonary embolism; acute renal failure; miscellaneous complications; delirium; dementia; stroke (for hip fracture and joint replacement patients only); and hip fracture (for stroke and joint replacement patients only).

We also created condition-specific clinical variables. For hip fracture and joint replacement patients we created indicators of the type of replacement the patient received. Hip fracture patients were classified as having surgery to pin their hip (i.e., no hip replacement), a total replacement, a partial replacement, and/or a revision of a previous joint replacement. We also coded the location of the fracture. For joint replacement patients, we coded the type of replacement (total, partial, revision), whether they were for knee or hip, and whether multiple replacements were conducted. For stroke patients we created indicators for the type of stroke. Finally, we created indicators for any use of an intensive care unit during the acute stay and the number of days spent in that unit.

Characteristics of Discharging Hospitals

Patterns of care and approaches to discharge planning in the acute care hospital can influence PAC use. Accordingly, we included a number of covariates to capture the orientation of acute care hospitals. They include size (average daily census or ADC), teaching status (resident to ADC ratio), ownership status (government, private nonprofit, or for-profit), Medicare patient percentage, case-mix index of the hospital, and low-income patient percentage. These measures were created using cost report and provider of service data available from the Centers for Medicare and Medicaid Services (CMS) website. In addition, we created variables that indicate whether the discharging hospital had a related SNF, IRF, or HHA subprovider listed on its cost report.

PAC Availability

We defined availability from a patient-specific perspective based on how close IRFs and SNFs were to patients' homes and how many of each type of facility were within reasonable distances of patients' homes. To construct our measures, we used patient and provider zip code information to measure the distance traveled from patients' residences to IRFs and SNFs. We used geocoding software to calculate distances from the midpoint of each beneficiary's zip code to the midpoint of the closest provider zip code. In addition, we considered the supply of formal substitutes and complements for formal SNF and IRF care. Specifically, we looked at the per-elderly supply of nursing home beds and the number of home health agencies in patients' areas of residence. Unfortunately, we had no data on patients' access to informal or family caregivers.

We created two measures of the availability of PAC. The first captures the distance from the patient to the closest provider (separate measures are created for closest IRF and closest SNF). Both the distance to the closest and the distance squared are included, since the effects of distance on PAC choice are likely diminishing.6 These variables measure how accessible the provider type is in terms of proximity. The second measure includes the number of PAC providers of each type within a given radius around the patient's home. We calculated these radii by condition and area type, and defined the radii using the 90th percentile of the distance traveled to that type of provider by beneficiaries living in that type of area; the 90th percentile was chosen since it reflected a generous definition of the market area, but was not biased by the care patterns of patients who might be receiving care far from home because of holidays or other reasons. We also created indicators for areas without any of a given type of provider as the lack of providers would have a strong negative effect on the use of that type of PAC.

Our measures of the “supply” of HHA care differed from that used for other PAC locations because HHA markets cannot be defined by patient travel patterns. Instead, we used patient claims data to determine which areas were served by which agencies. HHAs serving five or more residents within a given county and located in the same state or an adjacent state as those beneficiaries were counted as serving that county.7,8

Statistical Analysis

We identified hospitalized hip fracture, stroke, and lower extremity joint replacement patients and examined how each group's sociodemographic and clinical characteristics varied by PAC site used. We also examined how PAC use varied by characteristics of the discharging acute hospital and the supply of PAC care. We then fit multinomial logistic regression models of the form:

$$ln\hspace{0.17em}{\Omega }_{m|b}\left(X\right)\hspace{0.17em}=\hspace{0.17em}ln\frac{\text{Pr}(y\hspace{0.17em}=\hspace{0.17em}m|x)}{\text{Pr}(y\hspace{0.17em}=\hspace{0.17em}b|x)}\hspace{0.17em}=\hspace{0.17em}x{\beta }_{m|b}$$

(where b was the comparison group, no Medicare-covered institutional care) to assess the patient characteristics that predicted use of SNF or IRF care after discharge from acute care in a multivariate framework.9 We also fit “two-level” logistic regression models in which the first-level model predicted use of SNF or IRF care versus no Medicare-paid institutional care and the second-level model predicted use of IRF versus SNF care conditional on the use of institutional care. The fit and predictions from these models were virtually identical to those from the multinomial logit models, so we present only the multinomials.

Finally, we assessed the relative importance of clinical factors versus PAC supply factors in the choice of PAC site by simulating how much each set of factors changed the predicted probabilities of using IRF or SNF care. To look at the effect of supply factors on PAC use, we computed standardized predictions holding clinical factors constant at their means across all of our observations and predicting the probabilities of using IRF and SNF care for each observation (Lane and Nelder 1982). The resulting distributions of predicted rates of use demonstrate the extent to which supply factors shift patients across PAC sites when clinical factors are held constant. We then computed the same set of predictions holding the supply factors constant at their means but reflecting the effects of the full-observed range of values for the set of clinical variables. We compared the predicted distributions of probabilities of using IRF care, SNF care, or neither under these two scenarios to see which factors most affected the variability in PAC site used.

Results

Table 1 presents selected descriptive statistics for our three patient groups in 1999 overall and by type of PAC accessed. For all three conditions, SNF patients tend to be older and are more likely to be female than IRF patients. Patients not using Medicare-paid institutional care are, on average, younger. The hip fracture and stroke SNF patients have greater numbers of comorbidities and complications. In contrast, the hip fracture and stroke IRF patients have fewer comorbidities than the average patient in those groups, including lower rates of coronary artery disease, nutritional deficiencies, cellulitis or decubitis ulcer, and dementia (not shown in tables). Joint replacement patients, however, have similar levels of comorbidities in both IRFs and SNFs. The percentage of dual eligibles in IRFs is lower, and the proportion of Medicaid recipients who do not receive Medicare-paid institutional care is relatively high. There is a striking relationship between use of PAC and the availability of PAC, which is explored further below.

Table 1.

Sample Means by Condition, First Site of PAC

	Hip Fracture				Stroke				Lower Extremity Joint Replacement
	Overall Mean (SD)	No Medicare Institutional Care	SNF	IRF	Overall Mean (SD)	No Medicare Institutional Care	SNF	IRF	Overall Mean (SD)	No Medicare Institutional Care	SNF	IRF
Number of observations	106,570	5,898	67,476	25,407	149,091	45,845	46,998	33,059	151,168	22,300	51,650	44,650
Selected patient characteristics
Age (years)	82.6	80.3	83.6	81.3	78.9	76.8	81.5	77.8	75.0	72.8	76.4	75.5
	(7.5)	(8.4)	(7.2)	(7.2)	(7.6)	(7.3)	(7.5)	(7.1)	(6.0)	(5.3)	(6.2)	(6.0)
Female (%)	77.8	69.9	78.6	78.1	58.9	52.0	64.5	57.2	64.9	50.1	71.8	70.2
Medicaid coverage (%)	16.1	18.5	16.6	14.1	18.1	13.8	21.6	16.6	7.9	3.6	9.6	9.1
Any complications (%)(of 17 included in models)	14.7	15.4	16.1	11.9	12.0	7.9	18.3	10.5	9.8	7.8	11.2	9.9
Any comorbidities (%)(of 13 included in models)	61.6	60.8	64.1	56.2	75.3	67.0	80.6	79.6	36.2	30.6	38.8	38.6
Supply measures
Discharging acute has IRF subprovider	38.1%	39.1%	31.1%	58.4%	37.6%	37.1%	27.8%	55.5%	43.4%	46.4%	28.4%	61.7%
	(48.6)	(48.8)	(46.3)	(49.3)	(48.4)	(48.3)	(44.8)	(49.7)	(49.6)	(49.9)	(45.1)	(48.6)
Discharging acute has SNF subprovider	55.3%	54.1%	57.7%	52.8%	53.8%	53.2%	56.4%	54.0%	55.7%	49.3%	64.5%	53.4%
	(49.7)	(49.8)	(49.4)	(49.9)	(49.9)	(49.9)	(49.6)	(49.8)	(49.7)	(50.0)	(47.9)	(49.9)
Discharging acute has HHA subprovider	55.6%	57.1%	56.1%	53.8%	54.6%	54.7%	55.3%	53.4%	53.4%	49.9%	56.4%	52.4%
	(49.7)	(49.5)	(49.6)	(49.9)	(49.8)	(49.8)	(49.7)	(49.9)	(49.9)	(50.0)	(49.6)	(49.9)
Number of IRFs in radius around residence	8.8	7.8	8.5	10.0	9.9	9.6	9.7	10.2	10.8	9.5	10.7	11.8
	(9.6)	(8.7)	(9.5)	(10.0)	(10.3)	(10.1)	(10.3)	(10.4)	(11.9)	(11.3)	(12.2)	(11.9)
Number of SNFs in radius around residence	34.5	29.0	34.1	37.7	31.7	30.0	31.1	33.2	43.0	39.4	45.7	43.3
	(38.9)	(33.6)	(37.9)	(42.4)	(37.3)	(35.6)	(36.1)	(38.9)	(45.3)	(43.7)	(46.9)	(44.6)
Number of HHAs serving county of residence	57.3	44.5	54.6	70.3	59.4	56.7	54.6	67.1	52.2	33.2	52.6	63.6
	(81.6)	(71.5)	(79.5)	(89.7)	(86.3)	(84.6)	(80.4)	(94.0)	(77.8)	(60.3)	(80.8)	(83.0)
Nearest rehab (miles)	15.5	19.2	17.0	9.9	15.6	16.8	17.5	11.5	16.7	20.2	19.0	11.7
	(24.6)	(31.6)	(24.5)	(22.2)	(27.0)	(31.6)	(26.1)	(22.8)	(23.9)	(27.4)	(22.8)	(23.7)
Nearest SNF (miles)	2.0	2.6	2.0	1.8	2.1	2.2	2.0	2.0	2.3	2.8	2.1	1.9
	(4.7)	(6.3)	(4.6)	(4.4)	(4.5)	(4.9)	(4.5)	(4.3)	(5.1)	(6.3)	(4.8)	(4.8)
Selected other characteristics
Discharged from for-profit hospital (%)	13.3	10.8	11.9	17.7	13.2	13.1	12.8	14.1	12.2	9.6	11.2	15.8
Nursing home beds per 100 residents age 85+ in county	45.4	47.2	45.4	45.8	45.7	45.9	45.9	46.1	46.1	48.4	47.0	45.7

Note: Standard deviation for continuous variables are in parantheses.

SNF, skilled nursing facilities; IRF, inpatient rehabilitation facilities; HHA, home health agencies; SD, standard deviation.

As seen in the mean distances to nearest provider in Table 1, patients frequently use PAC providers that are far from their homes. Table 2 describes the distribution of distances, in miles, to the nearest IRF provider by condition and area type. The median hip fracture, joint replacement, or stroke patient in an MSA lives approximately five miles from the nearest IRF. Patients must travel farther for IRF care when they live outside of a MSA. The median distance from patients' places of residence to the nearest SNF provider, across all areas and all conditions, is always equal to zero.10 However, the distance to the nearest SNF provider does vary considerably: the top 10 percent of rural patients not living adjacent to an MSA have to travel over 12 miles to an SNF. The distances that some patients have to travel to reach the closest IRF are significantly greater, exceeding 70 miles for the most remote decile of patients, and even within MSAs patients regularly receive IRF care more than 20 miles from their homes. Table 2 also shows the distribution of the average number of providers within the radii defined by the 90th percentiles of distance traveled.

Table 2.

Distance to Nearest and Number of Providers in Radius around Patients' Residences by Condition and Area Type

	A. Distance in Miles to Nearest Provider						B. Number of Providers within Radius
	Hip Fracture		Joint Replacement		Stroke		Hip Fracture	Joint Replacement	Stroke
	Median	90th Percentile	Median	90th Percentile	Median	90th Percentile	Mean (SD)	Mean (SD)	Mean (SD)
IRF
MSA	4.5	19.3	5.0	20.3	4.4	19.6	6.4 (8.7)	7.0 (9.1)	6.5 (8.7)
MSA adjacent	25.9	43.8	25.9	44.8	26.0	43.8	6.9 (6.8)	6.9 (8.8)	8.6 (7.8)
Non-MSA	37.2	75.1	38.9	77.8	36.1	71.6	15.4 (16.4)	27.6 (23.6)	18.7 (18.3)
SNF
MSA	0.0	4.5	0.0	5.2	0.0	4.7	28.8 (37.5)	28.3 (37.1)	26.3 (35.2)
MSA adjacent	0.0	10.4	0.0	10.6	0.0	10.5	23.3 (23.0)	40.4 (32.7)	21.2 (21.7)
Non-MSA	0.0	12.9	0.0	13.9	0.0	12.6	42.6 (51.6)	108.0 (97.3)	34.1 (45.7)

SNF, skilled nursing facilities; IRF, inpatient rehabilitation facilities; SD, standard deviation; MSA, metropolitan statistical area.

These relationships generally held when we fit multinominal logistic regressions for choices between PAC sites for the hip, stroke, and joint replacement samples, and additional use patterns emerged. Online-only Appendix 1 presents the results from these logistic regressions (please see http://www.blackwellpublishing.com/products/journals/suppmat/HESR/HESR00366/HESR00366sm.htm).10 The first column shows the predictors of hip fracture patients using IRF care. The second column shows the factors affecting patients' use of SNFs (versus no Medicare-paid institutional care). A positive coefficient in the IRF column here generally indicates that patients with that characteristic are more likely to be discharged to an IRF than a noninstitutional setting, and a positive coefficient in the second column indicates that patients with that characteristic are more likely to go to an SNF. However, because the signs and magnitudes of the effects are difficult to interpret from the multinomial logit regression output, and because virtually all of the effects are highly significant given our sample size, we provide estimates of the marginal effects of these factors below.

The effects of PAC supply factors are strong and similar across conditions. Patients discharged from hospitals with IRF or SNF subproviders are more likely to go to them and less likely to go without institutional care. If all the hip fracture patients in our sample were discharged from a hospital with a related IRF, 34 percent of them would be expected to get IRF care; if none of them were, we predict that only 17 percent would get IRF care. (The corresponding figures for stroke patients are 30 and 17, and 41 and 21 for joint replacement patients.) In addition, hip fracture and stroke patients are less likely to seek IRF care if their discharging hospital has a related SNF; for hip fracture patients having a related IRF reduces the probability of using an SNF by 16 percent. Hip fracture and stroke patients are also less likely to get IRF care if they are discharged from a hospital with a related HHA.

The supply of IRFs relative to SNFs and the distance to each type of care are major determinants of which PAC site is used. The greater the number of IRFs in a patient's area, the more likely s/he is to seek IRF care. Conversely, the greater the number of SNFs in a patient's area, the less likely s/he is to go to an IRF. A one standard deviation increase in the number of SNFs in an area increases the probability that a hip fracture patient will use an SNF by 8.8 percent, and reduces the probability of IRF use by 21.4 percent. Interestingly, for all three conditions, those patients without IRFs in their area are less likely to use institutional care of either type. Distance to the nearest provider of each type is also important for all three types of patients. As distance to the nearest IRF increases, patients are less likely to seek out IRF services and as the distance to the nearest SNF increases they are more likely to seek IRF care; a one standard deviation increase in the distance to an IRF reduces the predicted probability of IRF use in our hip fracture model by a third and increases the probability of SNF use by 11.5 percent. The more nursing home beds in the county, normalized by the number of persons in the county over age 85, the more likely patients were to use IRFs or SNFs, although the significance of this relationship varied across the conditions.

Demographic, clinical, and other hospital and area characteristics remain important in the multivariate analyses. For example, all but two of the seventeen complications in the model were significant in either the IRF or the SNF branch. We have summarized the significance of these factors in Appendix 1.

Our simulations show the combined effects of the supply factors in the models. Table 3 shows the predicted proportion of patients not using Medicare institutional care, and the predicted proportions using IRFs and SNFs, under three different scenarios. The first sets of rows, labeled “A,” under each condition show the effects of supply factors on the range of predicted probabilities of using each care type. As described above, these were computed fixing all of the nonsupply factors, i.e. the sociodemographic, clinical, and hospital characteristics (other than ownership of a PAC provider) at their averages and then re-predicting PAC use for each patient. The range of predicted probabilities in these rows thus reflects only the effects of variation in PAC supply across the country. It shows that a hip fracture patient with average sociodemographic, clinical, and discharging hospital characteristics who lives in an area that puts him/her in the bottom 10th percentile with respect to IRF use—e.g., an area where there are many SNFs nearby but few IRFs—would have an 8.5 percent chance of going to an IRF, whereas one living in an area at the 90th percentile would have a 42.4 percent chance of going to an IRF. Holding nonsupply factors fixed, the interquartile range of the probability of getting IRF care is 20.7 percent, of getting SNF care is 18.9 percent, and of getting no institutional PAC is 4.4 percent.

Table 3.

Predicted Rates of PAC Use by Site

	Mean	10th Percentile	25th Percentile	Median	75th Percentile	90th Percentile	Interquartile Range
Hip
A. Predictions allowing only supply factors to vary
No Medicare-paid institutional PAC (%)	11.9	8.2	9.4	11.4	13.8	16.6	4.4
IRF (%)	23.0	8.5	12.9	19.4	33.6	42.4	20.7
SNF (%)	65.1	46.6	56.2	67.5	75.0	79.6	18.9
B. Predictions allowing only clinical factors to vary
No Medicare-paid institutional PAC (%)	13.1	7.0	8.2	10.4	16.2	23.1	8.0
IRF (%)	21.3	8.8	16.7	22.7	27.1	30.5	10.5
SNF (%)	65.6	50.9	59.0	66.7	73.8	79.3	14.8
Stroke
A. Predictions allowing only supply factors to vary
No Medicare-paid institutional PAC (%)	47.2	43.0	44.6	46.8	49.7	52.0	5.1
IRF (%)	21.9	12.9	15.6	18.8	30.3	34.0	14.8
SNF (%)	30.8	20.3	25.3	30.9	36.8	40.8	11.6
B. Predictions allowing only clinical factors to vary
No Medicare-paid institutional PAC (%)	47.7	25.5	35.3	50.0	60.6	68.3	25.3
IRF (%)	20.9	10.9	15.7	18.2	25.7	35.0	10.1
SNF (%)	31.4	14.0	19.6	28.7	39.9	53.1	20.3
Lower Extremity Joint Replacement
A. Predictions allowing only supply factors to vary
No Medicare-paid institutional PAC (%)	34.8	25.4	28.9	33.5	40.5	46.4	11.7
IRF (%)	30.5	13.5	19.0	27.1	43.4	50.4	24.4
SNF (%)	34.6	15.2	25.6	32.4	47.3	54.6	21.7
B. Predictions allowing only clinical factors to vary
No Medicare-paid institutional PAC (%)	38.1	17.7	26.0	37.8	49.6	60.1	23.6
IRF (%)	28.0	19.3	24.2	28.4	31.5	34.2	7.3
SNF (%)	33.9	20.0	26.1	33.6	41.5	48.0	15.4

SNF, skilled nursing facilities; IRF, inpatient rehabilitation facilities; PAC, postacute care.

The second sets of rows, labeled “B” under each condition, present the opposite scenarios. In these simulations the clinical complications, comorbidities, and condition-specific covariates vary as they do in the sample, while the other factors in the model (sociodemographic, hospital, and supply) are fixed at their averages. Looking again at the IRF row for hip fracture patients, a patient at the 10th percentile of likelihood of going to an IRF based on his/her complications, comorbidities, and type of fracture would have an 8.8 percent chance of going to an IRF and a 30.5 percent chance at the 90th percentile. (Given the relationships between IRF use and clinical factors described above, hip fracture patients falling at the lower end of the distribution in terms of rates of IRF use include patients with Medicaid coverage and those with complications and/or comorbidities.)

Comparison of the interquartile ranges of the predictions holding the nonsupply versus the nonclinical factors fixed shows the relative effects of those factors on the odds of use of each PAC location. These comparisons reveal that, for each condition, IRF use is the most affected by variation in factors related to the availability of PAC. Holding clinical factors constant, the probability of IRF use varies more than 20 percent from 12.9 percent at the 25th percentile to 33.6 percent at the 75th for hip fracture patients; the interquartile range for stroke patients is nearly 15 percent. For joint replacement patients variation in supply factors shifts the probability of going to an IRF from 19 percent at the 25th percentile to 43.4 percent at the 75th percentile. This effect is more than three times as large as the 7.3 percent shift for joint replacement patients because of complications, comorbidities, and the type of replacement surgery performed. The probability of not using Medicare-covered IRF or SNF care, on the other hand, is more affected by variation in clinical factors for each condition (e.g., 25.3 percent versus 5.1 percent for stroke). SNF utilization shows more variation across conditions, with supply factors affecting the use of SNF care for hip fracture (18.9 percent versus 14.8 percent) and joint replacement (21.7 percent versus 15.4 percent) patients more than the clinical ones.

Discussion

The availability of PAC is a major determinant of whether the three types of patients examined—those with hip fracture, stroke, or lower extremity joint replacement—use PAC care and which type of facility they use. The effects of distance to providers and supply of providers are particularly clear in the choice between IRF and SNF care. The farther away the nearest IRF is, the less likely a patient is to go to an IRF. The farther away the nearest SNF is, the more likely the patient is to go to an IRF. Similarly, the more IRFs there are in the patient's area the more likely the patient is to go to one and the more SNFs there are the less likely the patient is to go to an IRF. In addition, if the hospital from which the patient is discharged has a related IRF subprovider the patient is likely to go to an IRF; and if the discharging hospital has a related SNF subprovider the patient is more likely to go to SNF.

Our simulations demonstrate the importance of the clinical characteristics in the model relative to the PAC availability measures. While the clinical characteristics were generally more important determinants of whether a patient used an SNF or IRF, the availability measures were more important determinants of which PAC site was used. This suggests that clinical judgments about whether a patient will benefit from PAC play a role in the decision to use it, but that factors such as ease of referrals and accessibility of providers take precedence when choosing sites of care.

The major limitation of this study is that there could be other, unmeasured factors that are affecting the choice of PAC site. In particular, we are unable to observe whether patients used non-Medicare nursing home care after their acute stay.11 Thus, we are unable to distinguish those patients going to nursing homes (paid for by Medicaid or the patients themselves) from those patients returning to their homes. In addition, there may be other aspects of PAC supply—e.g., the number of unoccupied nursing home beds—that affect PAC use. Clinical factors than cannot be measured in discharge data, such as level of functioning, and sociodemographic factors, such as availability of caregivers, also affect PAC choices (Inouye et al. 2003). In addition, there could be important aspects of patient behavior or demand that affect the use of PAC, and that may even affect the supply of PAC in an area. Overall, our models did not explain much of the variation in PAC use. Nonetheless, they did include numerous patient and PAC supply factors that affected the choice of initial site of PAC.

The relationships we found were largely consistent across the three different conditions we examined, which were chosen to be representative of major types of PAC patients. Conditions that are treated predominantly with one type of PAC, however, would likely be less affected by PAC supply. It is also possible that these patterns could have changed since 1999 with the implementation of the home health and IRF prospective payment systems, but in other ongoing work we have not discovered major changes in the use of PAC for these three conditions.

While some might conclude that the evidence of higher utilization of services in areas with a greater supply of services is inefficient, there is little evidence-based research about PAC from which inferences can be drawn about the appropriate level of PAC. There is some evidence that aggressive postacute rehabilitation produces better functional outcomes for stroke but not for hip fracture, so it is noteworthy that PAC supply factors shifted use least for stroke patients (Kane et al. 1996; 1998; 2000; Kramer et al. 1997; Deutsch 2003). Still, predicted IRF use in our models varied tremendously across areas with different levels of PAC supply for all three conditions. More research is needed to evaluate whether these findings indicate that a greater supply of PAC leads to both greater use of institutional care and better outcomes, or whether it leads to unwarranted expenditures of resources and delays in returning patients to their homes.