Choice of Hospital for Delivery: A Comparison of High-Risk and Low

Choice of Hospitalfor Delivery: A Comparisonof High-Risk andLow-Risk WomenCiaran S. Phibbs, David H. Mark, Harold S. Luft, Deborah J.Peltzman-Rennie, Deborah W Garnick, Erik Lichtenberg, andStephen J. McPhee

Objective. This artide tests whether or not the factors that affect hospital choice differfor selected subgroups of the population.Data Sources. 1985 California Office of Statewide Health Planning and Development(OSHPD) discharge abstracts and hospital financial data were used.Study Design. Models for hospital choice were estimated using McFadden's conditionallogit model. Separate models were estimated for high-risk and low-risk patients, and forhigh-risk and low-risk women covered either by private insurance or by CaliforniaMedicaid. The model induded independent variables to control for quality, price,ownership, and distance to the hospital.Data Extraction. Data covered all maternal deliveries in the San Francisco Bay Area in1985 (N - 61,436). ICD-9 codes were used to dassify patients as high-risk or low-risk.The expected payment code on the discharge abstract was used to identify insurancestatus.Principal Findings. The results strongly reject the hypothesis that high-risk and low-riskwomen have the same choice process. Hospital quality tended to be more important forhigh-risk than low-risk women. These results also reject the hypothesis that factorsinfluencing choice ofhospital are the same for women covered by private insurance as forthose covered by Medicaid. Further, high-risk women covered by Medicaid were lesslikely than high-risk women covered by private insurance to deliver in hospitals withnewborn intensive care units.Conclusions. The results show that the choice factors vary across several broadly definedsubgroups of patients with a specific condition. Thus, estimates aggregating all patientsmay be misleading. Specifically, such estimates will understate actual patient response toquality of care indicators, since patient sensitivity to quality of care varies with thepatients' risk status.

Keywords. Hospital choice, quality of care, maternity care, conditional choice models,socioeconomic factors

202 HSR: Health Services Research 28:2 (June 1993)

In choosing a hospital to enter for treatment, a number of factors mayenter an individual's thinking. For instance, which hospital, from amongthe possibilities, has the most up-to-date facilities? Where does my physi-cian suggest I go? Do I anticipate the need for special intensive services?Where will I feel most comfortable? Where do many of the people Iknow go? Will my insurance-coverage allow me to choose this hospital?How far away is the hospital? For each person, each of these factorsvaries in relative importance. Some people may prefer proxnimity whileothers may value backup intensive care services.

Several investigators have used linear versions of conditional-choicemodels to look at the factors influencing hospital admissions for allpatients (Folland 1983; Lee and Cohen 1985; Erickson and Finkler1985) or for aggregated subgroups (McGuirk and Porell 1984). In all ofthese models, distance from the patient's home to the hospital was themost important factor. Garnick, Lichtenberg, Phibbs, et al. (1989) have

This research was supported by Grant Number HS-05411 from the National Centerfor Health Services Research and Health Care Technology Assessment (NCHSR). Dr.Phibbs was supported, in part, by a postdoctoral fellowship from the National Centerfor Health Services Research Institutional National Research Award HS00026. Dr.Lichtenberg and Dr. Mark were supported, in part, by postdoctoral fellowships fromthe Pew Charitable Trust Health Policy Program T87-01210.Address correspondence and requests for reprints to Ciaran S. Phibbs, Ph.D., ResearchEconomist, Center for Health Care Evaluation, VA Medical Center (152), 795 WilowRoad, Menlo Park, CA 94025. Dr. Phibbs is Research Economist at the Center forHealth Care Evaluation, Palo Alto Veterans Affairs Medical Center, and AssistantProfessor, Departnent of Health Research and Policy, Stanford University School ofMedicine. At the time of this research, Dr. Phibbs was a postdoctoral fellow at theInstitute for Health Policy Studies, University of Califomia, San Francisco. David H.Mark, M.D., M.P.H. is Assistant Prosor in the Department of Family and Commu-nity Medicine, Medical College of Wisconsin, Milwaukee. At the time of this research,Dr. Mark was a postdoctoral fellow at the Institute for Health Policy Studies, Universityof California, San Francisco. Harold S. Luft is professor of Health Economics at theInstitute for Health Policy Studies, University of California, San Francisco. Deborah J.Peltzman-Rennie is a Programmer/Analyst at the Institute for Health Policy Studies,University of California, San Francisco. Deborah W. Garnick, Sc.D. is AssociateResearch Professor, Institute for Health Policy, Brandeis University, Waltham, MA. Atthe time of this research, Dr. Garnick was a researh associate at the Institute for HealthPolicy Studies, University of California, San Francisco. Erik Lichtenberg, Ph.D. isAssociate Professor, Department of Agricultural and Resource Economics, University ofMaryland, College Park, MD. At the time of this research, Dr. Lichtenberg was apostdoctoral fellow at the Institute for Health Policy Studies, University of California,San Francisco. Stephen J. McPhee, M.D. is Associate Professor, Department of Medi-cine and Institute for Health Policy Studies, University of California, San Francisco.This article, submitted to Healh Services Research on February 27, 1991, was revised andaccepted for publication on September 25, 1992.

Choice of Hospitalfor Delivery

shown that it is preferable to use maximum-likelihood estimators insteadof linear approximations. Luft, Garnick, Mark, et al. (1990) used thismethod to examine hospital choice for 12 diagnoses and surgical proce-dures. They found differences in the relative influence of the factorsaffecting hospital choice across different diagnoses and procedures. Thisfinding raises an obvious question: do the factors that influence hospitalchoice vary across subgroups of patients with the same diagnosis?

We used maternal delivery to test for differences in hospital choiceacross subgroups of patients with the same medical condition. Deliveryis a good choice for examining these differences because of the long timethat passes between first knowledge of the need for hospital services andthe date when provision is actually required. Expectant mothers havemore time to shop around for the hospital that best meets their needsthan do patients with most other conditions. The concerted efforts toregionalize perinatal care for high-risk women and infants (McCormick,Shapiro, and Starfield 1985) should also influence the observed hospitalchoices. Finally, the large number of deliveries yields a sufficient volumeof perinatal deaths (despite relatively low mortality rates) to make itfeasible to use perinatal mortality as a marker for hospital quality.

The primary purpose of this study was to examine factors that mayinfluence different subgroups of the general population in their choice ofa hospital for delivery. We specifically focused on the choices of mothersconsidered to be at high risk of facing serious complications for them-selves or their infants and of mothers considered to be at low risk. In thisarticle we used data from all hospitals in the greater San Francisco BayArea with delivery services in 1985. Hospital choice was modeled as afunction of four types of variables: quality, price and payment, hospitalownership, and geographic access. This model is estimated separatelyfor the high-risk and low-risk groups to compare differences in hospitalchoice by risk status. For each of these risk groups, we also looked at theeffects of payment source on choice by comparing the hospital choice ofwomen covered by California Medicaid (Medi-Cal) with that ofwomencovered by private insurance.

DATA

Hospital discharge abstracts for 1985 from the Office of StatewideHealth Planning and Development (OSHPD) were selected for allpatients for whom delivery was the listed type of admission. To reducethe data set to a manageable size, deliveries only from zip codes in thegreater San Francisco Bay Area were considered.1 The size of this area

203


(over 100 miles long and about 30 miles wide) and its number of hospi-tals were sufficient to mean that fewer than 1.5 percent of all patientsresiding in the region received hospital care outside of its boundaries.Patients hospitalized in federal, state, or HMO hospitals were excludedsince their choice of hospital was more limited than that of otherpatients. The hospital discharge data induded the zip code of patientresidence, which we used to detennine the geographical origin of thepatient. The discharge abstracts also contained data on maternal diagno-ses, procedures, age, race, hospital charge,2 and expected source ofpayment.

Data on hospital ownership, teaching status, and facilities wereobtained from the OSHPD Hospital Disdosure Report. Data on thelevel of newborn intensive care available at each hospital were obtainedfrom the Maternal and Child Health Section of the California Depart-ment of Health Services. Data on the Medi-Cal contract status wereobtained from a published report of the California Department ofHealth Services (1988). Latitude and longitude coordinates for residen-tial zip codes were obtained from commercially available geographicfiles. Exact coordinates for hospitals were determined from topographicmaps plotted by the U.S. Geological Survey. Data on risk-adjusted birthoutcomes were obtained from the California Maternal and Child HealthData Base (Rust, Rust, and Williams 1989).

METHODS

STATISTICAL MODEL

Qualitative-choice models are a dass of methods used to study situationsin which an individual chooses from among a set of alternatives. Thesemodels may be estimated using the conditional logit method developedby McFadden (1974). The attractiveness of hospital j to patient m inlocation i, Y*,O, is a linear function of observable attributes of the hospi-tal and patient, Xi, and random variables, e,,,,:

Y*!rn = V(Xon) + eonThe probability that hospitalj is chosen equals the probability that it ismore attractive than all other hospitals in the choice set. If the randomcomponents, eo,, have identical independent Weibull distributions, thenthe probability of observing the selection of hospital j by patient m inlocation i, P!nip7 is:

pjBm = {exp[V(Xi,n)])} / {k exp[V(Xon)]} k = 1, ... , N


where N is the number of hospitals. The advantage of this model is thatin generating the parameter estimates it explicitly considers the charac-teristics of the alternatives that were rejected as well as the one that waschosen. We have shown elsewhere that the maximum-likelihood methodof estimating the parameters for this particular type of problem is morerobust than the other commonly used methods (Garnick, Lichtenberg,Phibbs, et al. 1989).

A key assumption of the conditional logit model is the indepen-dence of irrelevant alternatives (IIA). This assumption implies that theaddition or subtraction of alternative choices will not affect the estimatedparameters. We tested for violation of this assumption using the auxil-iary regression method of McFadden (1987).3 This test is asymptoticallyequivalent to the test of Hausman and McFadden (1984). This test didnot reject the IIA assumption (X2 = 0.6, d.f. = 17).

Given that we specifically designate different levels of hospitals,based on the level of care provided, we also considered the possibility thata nested logit model should be used. For this application, a nested logitmodel would have the first level of analysis examine choice across group-ings of comparable hospitals and the second level of analysis examinechoice within each grouping. This was of specific concern for this analysis,since high-risk women might consider only hospitals that offered tertiaryservices. Again, the auxiliary regression to test failed to reject the assump-tion that the model was not nested (X2 = 0.4, d.f. = 16) (McFadden1987).

PATIENT SUBGROUPS

Deliveries were classified as high- or low-risk based on the AmericanAcademy of Pediatrics/American College of Obstetrics and GynecologyGuidelines for Perinatal Care (1988). These guidelines define the levelof care appropriate for different maternal conditions.5 We dassified ashigh-risk deliveries the 23,903 women who had at least one of the guide-line conditions for which it is recommended that delivery occur in ahospital with at least an intermediate-level neonatal intensive care unit(NICU). All deliveries not identified as high-risk were classified as low-risk. Our coding scheme was designed to ensure that we captured allhigh-risk women. With 38.9 percent of the cases dassified as high-risk,we almost certainly dassified many women who were actually of moder-ate or nominal risk as high-risk: this admittedly dilutes the observedeffects of risk status on choice of hospital.

Many of the conditions that dassify a pregnancy as high-risk, suchas maternal chronic disease, can be identified in advance. Others, such

205


as severe hypertensive disorders, surface later in pregnancy, but eventhen patients and physicians can consider switching hospitals. Condi-tions, such as pretern labor, that cannot always be predicted are thosethat the mother or her physician should immediately regard as unstable/precarious. The standard of care in these cases is to transfer or directlyrefer the high-risk mothers to a site that offers the appropriate level ofcare if the mother is stable enough to allow the move. Thus, it seemsplausible that joint patient-physician preferences about where to deliverwill be different for women identified as being high-risk, than for low-risk women-up to the point where transfer of the mother is consideredunsafe. Specifically, high-risk women and their physicians who knowtheir risk status, should be more willing than low-risk women and theirphysicians to travel farther, and should be more concerned about qualityand the services available to treat high-risk cases. To test for differencesin the choice process between high- and low-risk patients, the model wasestimated separately for each group (unrestricted models). These esti-mates were compared with the estimates of data pooled from bothgroups, which had forced the parameter estimates to be equal for bothgroups (restricted model). The difference between the log likelihood ofthe restricted model and the sum of the log likelihoods of the unrestrictedestimates is distributed x2 with k degrees of freedom, where k is thenumber of regressors in the model.

The demographic and socioeconomic characteristics of Medi-Cal-eligible women are quite different from those of the insured popula-tion. With respect to choice of hospital, their increased reliance on publictransportation is especially important. In addition, the willingness ofphysicians and hospitals to accept Medi-Cal patients may differ fromtheir willingness to treat insured patients. Therefore, the parameters ofthe choice model are likely to be different for these two patient groups.For this test, patients with an expected payment source of Medi-Calwere compared with patients whose expected source of payment wassome form of private insurance (commercial insurance, Blue Cross, BlueShield, non-Kaiser HMO or other prepaid plan, or Medicare).6 Patientsnot included in these two groups were exduded from the comparison. Tofocus the comparison on the differences between these two groups, thiscomparison was made separately for the high-risk patients and the low-risk patients. The unrestricted models were separate estimates for theMedi-Cal patients versus insured patients, and the restricted model wasfor all patients in these two groups.


VARIABLES IN MODEL

The dependent variable for our estimates was patient-hospital pairs.Since the data elements are identical for all patients in a zip code whochose the same hospital, we aggregated to the zip code-hospital level andweighted the estimates by the number of patients in each pair for compu-tational convenience. Thus, computationally, the dependent variablewas the number of patients from a given zip code who chose a givenhospital. The independent variables in the model included zip code-spe-cific patient characteristics and characteristics of each hospital. Thesevariables could be categorized into four general groups: quality, charges,ownership, and distance. Table 1, further on, shows the mean values ofthe variables used in the analysis: by hospital, for all patients, for thehigh-risk and low-risk patients, and for the low-risk and high-riskprivate-pay and Medi-Cal patients.

QUALITY

We used both structural and outcome variables to measure the quality ofcare. The level of neonatal intensive care available is a measure of therange of services the hospital is able to provide to infants. These unitsshould tend to attract more women who anticipate a high-risk delivery.Given that some complications cannot be predicted in advance, givingbirth at a hospital with a NICU assures the availability of any level ofcare that may be required. Therefore, the presence of a NICU may be afeature that attracts some women for low-risk deliveries. We used datafrom the California Department of Health Services to create categoricalvariables for the presence of level III (tertiary), high-level II (intermedi-ate), and low-level II NICUs. Teaching affiliation with a medical schoolmay also be interpreted as a marker of quality by some women. Otherlow-risk women may wish to avoid the "extra attention" commonly asso-ciated with the dinical training of house staff and medical students inteaching hospitals.

Hospital characteristics that can affect the birth experience mayinfluence the motheres choice. We included binary variables for thepresence of an alternative birth center and the provision of parent train-ing classes. By 1985, the rapid increase of cesarean section deliveries waswell documented, and many experts were questioning whether or not allof them were necessary (Bottoms, Rosen, and Sokol 1980; Fraser,Usher, McLean, et al. 1987; Gleicher 1984). On the other hand, it hasbeen shown that cesarean section delivery may improve outcomes, espe-cially for high-risk infants (Williams and Hawes 1979). We included thehospitals' cesarean section rates in the model to control for the fact that

207


this variable might affect choice of hospital.7 Cesarean sections wereidentified by a primary or secondary ICD-9 procedure code of 740-744or 749. Percentages of cesarean sections were based on all deliverypatients in each hospital, including those who lived outside the SanFrancisco Bay Area.

Although an expectant mother is obviously concerned about herown well-being, the rarity of maternal mortality makes it likely thatnewborn outcomes are predominant in her assessment of hospital qual-ity. For newborns, the dominant risk factor is birth weight. Since exactbirth weight data are not available on the OSHPD discharge data, weused a Z-score-based outcome measure derived from vital records dataas a hospital-level measure of quality (Williams 1979; Rust, Rust, andWilliams 1989). This measure adjusts for birth weight, sex, race, andplurality to calculate an expected perinatal mortality rate for each hospi-tal. Indirect standardization was used to compare the expected mortalityrate with the actual mortality rate. The Z-score was used to correct fordifferences in sample size across hospitals, which affects the level ofconfidence associated with observed outcomes. A Z-score of zero impliesthat the actual number of perinatal deaths equaled the expected numberof deaths. Z-scores greater than zero imply that actual deaths weregreater than expected deaths, with the converse true for Z-scores lessthan zero. This measure has been shown to be a good indicator of thequality of perinatal care independent of socioeconomic status (Williams1979). Since the Z-score takes the birth weight as it is given, and adjustsfor this risk accordingly, the potential for omitted-variable bias due to thecorrelation between lower birth weights and socioeconomic status isreduced. It has been shown that selective maternal referral of cases withcongenital anomalies introduces a bias in the Z-scores of some tertiaryhospitals in California (Rust, Rust, and Williams 1989). To eliminatethis bias, we used the Z-scores that Rust et al. calculated after deleting alldeaths due to congenital anomalies.8

CHARGES

Although insurance coverage reduces the effective price for hospitalservices, relative price has still been found to influence hospital choicefor some conditions (Luft, Garnick, Mark, et al. 1990). To capture thiseffect we used the hospital charge listed on the discharge abstract in themodel. To adjust for the effect of case mix on charges, we estimated aregression model where charges were dependent on mother's age,expected payment source, race, cesarean section, emergency roomadmission, whether the hospital included physician charges in its bill,


and whether the hospital had a Medi-Cal contract. This model wasestimated for low-risk deliveries only, to minimize the possibility thatunobserved case-mix difference might bias the results.9 We used theratio of actual charges divided by expected charges to control for theeffect of relative-price differences in the conditional choice model.

MEDI-CAL CONTRACTING

In 1982, California passed legislation switching Medi-Cal to a selectivecontracting program and enabling formation of preferred provider orga-nizations (PPOs).10 Both of these changes restricted the options of manypatients. To control for the hospital's eligibility for Medi-Cal births(about 20 percent of all births), our model included the number ofmonths in which each hospital had a Medi-Cal contract. Unfortunately,data on PPOs are not readily available; thus, their effect could not beexamined.

SOCIOECONOMIC STATUS

The socioeconomic status of patients may influence their choice of hospi-tal. Indigent patients may be concentrated at public institutions and alimited number of other hospitals due to the restrictions of Medi-Calselective contracting, patient reliance on public transportation, and theunwillingness of some hospitals to accept uninsured or underinsuredpatients. More affluent patients may prefer well-appointed hospitalsthat cater to their needs. To test the magnitude of such socioeconomicgrouping, we created a variable that combined the percent of each hospi-tal's patients who had private insurance with the percent of the popula-tion in each zip code with at least some college education." However,simply multiplying these variables would not have picked up the desiredeffect, since high education-high percent insured combinations wouldbe near one while low education-low percent insured combinationswould be near zero. Under a "likes attract" hypothesis, these two combi-nations need to be similar. Thus, the interaction term we induded isbased on the deviations from the mean value of each of the components.

This variable yields large positive values for the high-high and low-low groups. It yields large negative numbers for opposites (low-education zip code paired with a high percent-insured hospital and higheducation zip code paired with a low percent-insured hospital). A posi-tive coefficient can be interpreted as support for the hypothesis thatindividuals of high (low) socioeconomic status (SES) tend to concentratetheir admissions in hospitals that primarily serve individuals with good(poor) insurance coverage. The subsamples used in the comparison of

209


Medi-Cal and privately insured patients limit the variance of the SESinteraction variable just described; Medi-Cal patients tend to come fromlower-education zip codes, and private-paying patients tend to comefrom higher-education zip codes. Thus, any variance that remains is duemostly to differences in the percentage of each hospitas admissionscovered by private insurance. To avoid the reduction in the varianceinduced by the interaction term, we used the percentage of each hospi-tal's patients covered by private insurance for the models comparingMedi-Cal and privately insured women. Here, the "likes attract" hypoth-esis would imply a positive coefficient for the private insurance sub-sample and a negative coefficient for the Medi-Cal subsample.

OWNERSHIP

Ownership status may also influence choice of hospital. The strongestfactor is probably local government ownership, since these hospitals,which primarily serve indigent populations, often are less attractive toprivately insured patients for this reason. Hospitals operated by localhospital districts, while controlled by a publidy elected board, do nothave a special mission to serve the poor. It is unclear how this type ofownership influences choice of hospital. Many individuals wish to avoidfor-profit hospitals in the belief that profit making is inconsistent withhigh-quality care. Depending on one's beliefs, Catholic ownership maybe regarded as an attraction or, possibly, detraction. We created binaryvariables for each of these ownership categories. The excluded categoryis private, not-for-profit, non-Catholic hospitals.

DISTANCE

Hospital markets are relatively localized (Garnick et al. 1987). Thus,patients tend to prefer hospitals closer to home. Because neither roaddistance nor travel times were available, straight-line distance was calcu-lated from the location of the patients' zip code to the exact location ofeach hospital using latitude and longitude coordinates. To account forthe bottlenecks associated with travel in the San Francisco Bay Area, weincluded a variable indicating whether a bridge or tunnel would be usedon the most direct route between a zip code and a hospital.

RESULTS

The final data set included 61,436 deliveries. Of these, 37,583 (61.1percent) were classified as low-risk, and 23,903 (38.9 percent) were


classified as high-risk. Of the low-risk deliveries, 9,329 (24.9 percent)were covered by Medi-Cal and 23,500 (62.6 percent) were covered byprivate insurance. Medi-Cal covered 6,725 (28.1 percent) of the high-risk deliveries, and private insurance covered 14,987 (62.7 percent) ofthem. The remaining deliveries were self-pay or HMO, and were notinduded in the payer source comparisons. Table 1 shows the means ofthe variables in the model for all deliveries and for each subgroup.Among the data subsets shown on Table 1, most of the variables aresimilar for all of the groups. There are, however, some notable excep-tions. Compared to low-risk patients, many more high-risk patients useteaching hospitals and hospitals that provide level III newborn intensivecare. High-risk patients are also concentrated in hospitals with lowerrisk-adjusted mortality rates. Many fewer Medi-Cal patients cross atransportation bottleneck (bridge or tunnel) than private-pay patients. Amuch higher percentage of Medi-Cal patients use public and Catholichospitals. Surprisingly, there was no difference in the cesarean sectionrate between high-risk and low-risk women.

Table 1 also shows the means for the 48 San Francisco area hospi-tals. About half of the hospitals are private, not-for-profit hospitals.Most of the hospitals provide parent training classes and have an alter-native birth center. Slightly over one-third of the hospitals provide somelevel of newborn intensive care. Medi-Cal contracts were present at 27of the hospitals for all 12 months of 1985, and one had a contract forseven months. The remaining 20 hospitals had no Medi-Cal contract atany time in 1985.

Table 2 shows the results for the estimates for all deliveries and thehigh- and low-risk subsets. Quality, price, ownership, and distance allhad significant effects on the choice of hospital, as we have previouslyfound for other types of cases (Luft et al. 1990). Teaching hospitals andhospitals that provide newborn intensive care were preferred, withhigher levels of newborn intensive care preferred over lower levels. Bet-ter outcomes, as measured by the risk-adjusted Z-score for perinatalmortality, also increased the probability that a mother would choose todeliver at a particular hospital. Patients also preferred hospitals thatoffered parent training classes and alternative birth centers, and hospi-tals that had higher cesarean section rates. Holding other factors con-stant, patients preferred hospitals with lower charges. Having a contractwith the state Medi-Cal program reduced the probability that a hospitalwould be chosen. The SES interaction variable had the expected positiveeffect. Catholic hospitals were more attractive, as were public hospitals,than proprietary and district hospitals. Straight-line distance from thezip code of patient residence to the hospital had a very strong effect on

211

LoL --Q Mmo M o M C1 LOecl r< .4 ui C4a .<6 l~

<;ugeeseee eeee XOY~~~~~~~~.4.a0 < b N b b W @ e + o _ N b-N N Oe~~~~~~~4t~~~~~~O0e-eX ¢0-o+

3'6 "888 888 "637n

Y ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~~~( ._ O

*3CNO O O +- o C4 t6 0; '-) 0 D6r4CL

tD (I OOO0 l~ X 4 ;~ C-OLI'

m3 0 0 _leq% -e o _0oDCO . *{3~~~~~c V; %6RaEEeugcc r.. C6C 6Q '~ OL

4 t .E o +en m - C O c"'0O en I 4- f

r~~~~~-.o . CZ C-tO CO_ 5

WI 0

30 '-{4r 0LO 3qLO 00 cn CO 4m tD v

LO 10* 4oo4- O C 0o C14 0o5. 6c6&o eo -coI oen C14 C4C-

m U-o Lo C-

0l C14 WC C 11 CO 0,CO. en t - COCD10) C14010 C4e C10 F C4 -o

M W Rt i C4L

C~0LO O1CO- MCO

CO r_ 6a41 rcl 4el; - C- 400 LOCOC 0 C _-

Lo o LeS) co

04 C44 '-~ LO10- C,4 10 t.. C-

C, cn c- to 04

_Ct

1010 CS4 Ct4C C; 0 CO_ o

o C1 _ COCLo -_LOLo " m

CNO_

I LO tl-. c) CO _

ent ce)C C 00CO C,4 t-.10-COCO-010-C-4 0 CO0t- 10t- CO C-40T"o f)t,,0V

CO * 10- CO COCN 000CO 1 COCO O O Ce (7 _ _;C% _ _ I o10

-

10 C C LO CU e0 CO

C CO 1 - CO 100O

CO0 CO +a,_

c)

o 0) "-t'.-C 0C

LrC C C I

100CU_CUCCO 0

CM _ C4~ l--O~,CO

e- - CO - CooCC 1o - U CO

Le C0 CU CO COCO

100060--

_ _ C

.~

4)4. _

0 S Q ..oP.

"0 b,-

~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~

4.) r.

t~4 F% z 4 z

4)bO4)0

C)4)

0

vCXCCUCUgCU

0X

N

90CU4)

4)Q0"0

D)

4)

.5U0YC

U)

CU4)CU

BqU4*CU

UL)

$-4

0

"0

b._

$-4

04._

3

0

04

U)

0

$4

C4

0

0

0

CO)

(U4._

He

.U0~CU7U

co

IL)"0

N

CUOU4)

4)

2

C)

"0U)

0O

4)

4).5"0

05


the choice of hospital. The necessity of passing tirough a major trafficbottleneck (a bridge or tunnel) to reach a hospital also reduced theprobability that that hospital would be chosen.

The hypothesis that the high- and low-risk patients made the samehospital choices for delivery was rejected (p < .001 xI = 800, d.f. =17). The third and fifth columns of Table 2 show the estimates of theseparate models for the high- and low-risk groups. Low- and high-riskwomen placed about the same importance on the presence of intermedi-ate newborn intensive care, a hospitals Medi-Cal contract status, anddistance. High-risk women had stronger preferences for hospitals withbetter quality measures (lower risk-adjusted mortality rates, teachinghospitals, and hospitals with level III NICUs), and were more willing totravel through a traffic bottleneck to reach their preferred hospital.Women expecting high-risk deliveries were more willing to go to a publichospital, more likely to deliver at hospitals with high cesarean sectionrates, and more likely to avoid a proprietary hospital than were womenexpecting low-risk deliveries. Surprisingly, women in the high-risk cate-gory seemed more sensitive to charges than those in the low-risk group.

Table 3 shows the estimates that compare the low-risk Medi-Caland private insurance subgroups. The hypothesis that low-risk patientscovered by Medi-Cal and patients covered by private insurance madethe same choices was rejected (p < .001 x2 = 4,963, d.f. = 17). Medi-Cal patients were less likely than women with private insurance tochoose a teaching hospital, a hospital with a NICU (any level), a hospitalwith a high cesarean section rate, a hospital with a higher risk-adjustedmortality rate, a hospital with an alternative birth center, a proprietaryhospital, or a hospital that was more difficult to reach. They were morelikely to choose a hospital with a Medi-Cal contract, a hospital withhigher charges, a hospital that offers parent training dasses, or a publichospital. The percentage of a hospital's admissions covered by privateinsurance has a significant positive effect for the private-pay patientsand a significant negative effect for the Medi-Cal patients.

The estimates that compare the high-risk Medi-Cal and privateinsurance patients are shown in Table 4. The hypothesis that these high-risk subgroups have the same choice model was rejected (p < .001 x2 =

3,407, d.f. = 17). In general, although the magnitude of the differencesvaried, the direction of the differences between the two subgroups wasthe same as that of the differences between the low-risk subgroups. Theexceptions were (1) the increased aversion of privately insured high-riskpatients to proprietary hospitals-now equal to that of Medi-Calpatients; (2) the stronger aversion of patients with private insurance tohospitals with high Z-scores, instead of a weaker aversion; and (3) the

CE 0 WO0)U Cf Om C- 0 C14- M£ocr

C1 -Lboc

t4 to Lceo .O C,4 -

LO . * . *. . . . *

* (M m c o eLo

V-4M ~ 0~'4C4-Lei (0C-C0 4 - C -4(0

1 LO Cf 1

04 (1 tDco r- I& Co . -4 C-4

cn C1* -o- 44- "t

* O(0 ~00ocn'-~o -4 (0 Co00 LO eC,0

Co ' U, , 0t 0 Co Cs U, t

Co U, fCo ' - o 0

- 4 U C 0I r0C

6o- - o+46666

L

01ol- -O 01 LO0 0

0 +U: e o o o co . .10eb07_ - a"_s:tV-

*

-C. *r

.'

.

(o - U, - e (0 U,- 0 C C4 Co rb 0-- 1-4 0C mI

-4 1_

C%N ce) -4 -O4 C'4

o" C"4 C"4 c& co

!! V- Lir LO V) c_ ; C '-.; C

U,)6;

C14 co U, oco 07) C-4 en 0'C-

Co 0D ceX 0

0 - en LO+b

Cf C C

c

o 6_ 6otU + o

. . . .

e

C141

Cs4 t 0C -- UO -C - C

t- o14*

ol

m ld,-;666 o--4_I

0

U,r

*~~ ~ ~

4)0

v 04)~~~~~~~~~~~~~~~~r

z 4 )z ,a

4-6

0

3

Q

0

a)

C4-

U)

0.a*C

N

4)J

4)

0

0

4i

w

"0.U4)

0

0

540

s0,

gU544)4)

QCbo

"0d

C-l tD O 000O

-C - N1.4 1- 0C,4 O Ct _ U- 00 t_

_ _-_ I-_I I _

cl4 el;4C N.- U

00N U-C c4 ) C-4 a)

_-

O CL4O)0 C4 N CO4toa_

O. 00 .~ .~ . . N. 0 N. .

70 Ve0 -O 00-0 000

Co 000 Ob -f uSeCNo

sP-+ Ct LOCM LOC

6o q o Cc1i-0-00-

t_; cl ro _4 4 ci c4 C; CoCN IN C C

L 0000Ce4u) Cl t 0C - t-

oo"04 " LO 0 co U" 04

000 0 " 0 000)Lo

O O L0 O 0O O O_4 t- U: t-,QCl

cl CS + _; O 06 Lr Cli 1_ fCe)CN C14 C1

oC) _0 O C O o)oCO <:e-- OUCI04N. 000 C" C -

- 04- t- N -O " 0664~CS!66 6~ 6

oD C) 00 1-N. 0Qo LO U") tN t- C) t

O oC zCO Cf LO

- o* Co

+ N.tCQ). 00)0oX 0- 0U" 14 t- Aco o 00N- - co t

Ue) 0 0. . 04 t. C t

m-010-0 04 N. Cue: 00oo4 r-csn

LO C4 en t)

- e I-

- -11 N. N.U"Ce"-i0

-

m) O) ce) -4 oce) rN 0)0000t_)iO 00 Q 0) ¢

Ot- 4a)O 6_

_ *n) 00 Ce4 orC o

0))0 CN CE"

e s _

C13~~~~~~~~~~~~~~~~~~~~~~~~~~~~~C-

~r

co

0~~~0

W

0~~~~~~~0b

.0 0CU.b0

V

C's

ri

~0

*0

bi)

0T

Co

HO

0

CU

V

CU

VV

V

E

0

.)"0

V

0

V

0

V

"0

CU

CE

.5CU

"0

._5V

CU

CU.5


finding that Medi-Cal patients are now more likely, instead of less likely,to select a teaching hospital.

DISCUSSION

The results presented here dearly show that the relative influence ofvarious factors affecting patients' choice of hospitals differs for varioussubgroups of patients with one dinical condition in common: maternaldelivery. For almost all of the variables in the model, the effects took thesame direction but differed in magnitude. We chose to look at maternaldelivery since it was relatively simple to identify the high-risk patients,and the volume of cases made it more likely that we would find statisti-cally significant differences. Further, the relatively long period ofadvanced notice of the need for hospitalization before delivery increasesthe ability of the patient to shop around. Additional studies must bedone to determine whether differences in choices among subgroups ofpatients will also be observed for other diagnoses and procedures wherepatients have less advance notice of the need to be hospitalized.

We have previously shown that differences exist in hospital choicesamong patients with different conditions (Luft, Garnick, Mark, et al.1990). The results presented here indicate that the estimates of patients'choice of hospital vary across identifiable subgroups within a definedgroup of patients. Researchers will need to be sensitive to these potentialdifferences in applying these models. Specifically, when interested inparticular subgroups of patients, researchers will have to estimate themodels separately for those subgroups. Given that the number ofpatients in any one zip code who choose a particular hospital for treat-ment dedines rapidly as the subgroup under study is narrowed, suchanalysis will increase the number of zip code-hospital pairs with zeropatients. This implies that such estimates will have to use maximum-likelihood estimates, instead of linear-approximation techniques, toobtain stable parameter estimates (Garnick, Lichtenberg, Phibbs, et al.1989).

The comparison between the high- and low-risk cases shows thatthose factors likely to have a strong effect on expected outcome weremore influential in the choice of hospital by high-risk women (and/ortheir physicians). This is consistent with the hypothesis that the sensitiv-ity of demand to quality varies direcdy with the probability of perinataldeath: the greater the risk of perinatal mortality, the more the patient isconcerned about the expected quality of care. If this finding holds forother diagnoses, then estimates that aggregate all patients with a given

217


diagnosis will not accurately measure the quality elasticity of demandunless they include some measure of case mix (i.e., the risk of mortalityof each patient).

Most of the results presented here were consistent with expecta-tions. The increased willingness of high-risk women to give birth inhospitals with higher cesarean section rates is consistent, given that forhigh-risk cases cesarean deliveries have been shown to be associated withimproved outcomes (Williams and Hawes 1979). Although low-riskwomen do not ignore quality, they tend to place less importance on theoutcome-oriented quality factors we induded, which is consistent withtheir low-risk status. These mothers seem to place more emphasis onfactors perceived to influence the "ambiance" of the birth experience.Specifically, they are less likely to choose a tertiary teaching hospital,and thus can avoid the extra "attention" of house officers and medicalstudents.

Caution must be applied in interpreting some of the differences inthe estimated parameters between women covered by Medi-Cal andthose covered by private insurance. Although the coefficients for theprivately insured patients almost certainly reflect the preferences ofthose patients and their physicians, we cannot draw the same condu-sions for women covered by Medi-Cal. The 'choice" estimates for theMedi-Cal sample were also influenced by access restrictions and factorsassociated with serving large Medi-Cal populations. This may provide apartial explanation for the results that show Medi-Cal patients with astronger preference for higher-charge hospitals.12 Similarly, whatappears to be a stronger aversion to cesarean sections by Medi-Calmothers may be due not to patient preference but to the fact thatresource-constrained public hospitals simply peiforn fewer cesareansections. 13

A more disturbing finding is that high-risk women covered byMedi-Cal were more likely to choose hospitals with worse perinataloutcomes (higher Z-scores), and were less likely to deliver at hospitalsproviding any level of specialized care for newborns. These results raisethe possibility that high-risk women covered by Medi-Cal face barriersto care at appropriate facilities. We cannot identify the cause of thesedifferences with our model, but this matter merits further investigation.

Most of the observed differences between the Medi-Cal and pri-vately insured groups have more standard explanations. For exaample,Medi-Cal patients were more likely to be admitted to a public hospital orto a hospital with a Medi-Cal contract, and were less likely to be admit-ted to a proprietary hospital. As our "likes attract" hypothesis predicts,they were more likely to be admitted to a hospital with a larger propor-


tion ofuninsured patients. Although the distance parameters were aboutthe same, the much larger estimate for the bridge-tunnel variable indi-cates that geographic access is a much more important issue for the poor.It is also the case that in the Bay Area these bridge-tunnel bottlenecksusually signify county boundaries. Since much of the public transit inthe Bay Area is county-based, this finding may be capturing more thanjust a traffic bottleneck.

Although we did observe significant differences in hospital choicebetween high-risk and low-risk women, our estimates -due to our broaddefinition of high-risk- probably understate the true effect of risk onhospital choice. We expect that we would find a risk gradient if werefined our definition of high-risk into several groups of increasing risk.We chose not to do so in this analysis since our primary purpose was tosee if there were differences in the choice decision across subgroups ofpatients, not to estimate precisely the magnitude of the effect of risk onhospital choice. Further, such an analysis would have been very difficultto undertake with the data we used, since ICD-9-CM codes indicateonly the presence of a diagnosis, not its severity. In spite of these limita-tions, our results demonstrate that hospital choice is sensitive to patientrisk and insurance status. Increased sensitivity to risk factors amonghigh-risk women suggests that referral patterns are at least somewhatrational. The different results for Medi-Cal women reflect the seg-mented medical markets that influence choices for privately versus pub-licly insured women.

ACKNOWLEDGMENTS

We thank Ronald Williams and two anonymous reviewers for helpfulcomments on an earlier draft of this artide. Any errors or omissionsremain our own.

NOTES

1. For more detail on how these zip codes were selected, see Garnick, Lichten-berg, Phibbs, et al. (1989).

2. For some hospitals, this data field includes some physician charges. Thedata indicate which hospitals also bill for physician services. We controlledfor this in our regression model to calculate risk-adjusted charges.

3. The prediction errors of the original model are used as the dependentvariable in a model where the independent variables are the weighted

219


average of each independent variable in the original model. The weightsare the estimated probabilities from the original model.

4. This result is for the model comparing all high-risk deliveries with all low-risk deliveries. The statistic reported below for the nested logit model is alsofor this data set.

5. The Appendix available from the authors shows these guidelines and liststhe ICD-9 codes that were used to classify women as high-risk.

6. Medicare was grouped with private insurance because it is a relatively"good" payer for delivery. Further, Medicare obstetrics patients are veryrare, that is, only those women covered under the disability provisions ofMedicare, and so forth.

7. We considered, and rejected, the possibility that cesarean section might beendogenous-that it is a function of some of the other independent vari-ables and hospital choice. While a hospital's cesarean section rate may beinfluenced by the preferences of the women who deliver there, it is likelythat most of the variance in cesarean section rates across hospitals is due todifferences in the practice styles of the physicians on staff at each hospital.Because of this variation in practice styles, significant variability exists inthe cesarean section rates across all of the other factors controlled for in theanalysis.

8. Although there is a known volume-outcome relationship for births, wecould not indude any direct measures of number of hospital beds or patientvolume in the model because these variables were highly collinear withseveral of the other variables in the model. Furthermore, we were attempt-ing to explain choice of hospital based on variations in outcomes, not thereasons for variations in outcomes. In addition, the number of patientschoosing a hospital from any one zip code is a small fraction of totalvolume.

9. For more details on the case-mix adjustments, see Garnick, Lichtenberg,Phibbs, et al. (1989) or Luft, Garnick, Mark, et al. (1990). Copies of theseregression-results are available from the authors on request.

10. Under selective contracting, Medicaid patients were restricted to receivinghospital care at contract hospitals, except on an emergency basis. By 1985,all of the areas in our study had been incorporated into the selective con-tracting program.

11. More than 12 years of education is used as a proxy for higher income.Educational level and income are very highly correlated, especially whenaggregated to the zip code level.

12. A referee also noted that, since Medi-Cal patients do not pay any coinsur-ance, this result is consistent with economic theory.

13. A similar, but different explanation would attribute this difference to thefact that physicians who treat a larger number of Medi-Cal patients are lesslikely to let economics influence their medical decision making. Thus, theyare not swayed to perform more cesarean sections by the higher reimburse-ment for operative deliveries. It may also be the case that some cesareansections are encouraged by private physicians to end a long labor, while inhospitals using rotating physicians the responsibility for delivering the childis simply passed to the next-shift physician (Fraser et al. 1987).


REFERENCES

American Academy of Pediatrics Committee on Fetus and Newborn and theAmerican College of Obstetrics and Gynecology Committee on Obstet-rics. Matrnal and Fetal Medicine. Guidelines for Perinatal Care. 2d ed. ElkGrove, IL: American Academy of Pediatrics, 1988.

Bottoms, S. F., M. G. Rosen, and R. J. Sokol. 'The Increase in the CesareanBirth Rate." New EnglandJournal ofMedicine 302, no. 10 (1980): 559-63.

California Department of Health Services Seletive Provider Contracting Program, AnnualReport, 1988. Sacramento: California Department of Health Services,Health and Welfare Agency, 1988.

Erickson, G. M., and S. A. Finkler. "Determinants of Market Share for aHospital's Services." Medical Care 23, no. 8 (1985): 1003-18.

Folland, S. T. 'Predicting Hospital Market Shares." Inquiry 20, no. 1 (1983):34-44.

Fraser, W., R. H. Usher, F. H. McLean, C. Bossenberry, M. E. Thomson,M. S. Kramer, L. P. Smith, and H. Power. "Temporal Variation in RatesofCesarean Section for Dystocia: Does 'Convenience' Play a Role?" Ameri-can Joural of Obsttrics and Gynecology 156, no. 2 (1987): 300-304.

Garnick, D. W., E. Lichtenberg, C. S. Phibbs, H. S. Luft, D. J. Peltzman, andS. J. McPhee. 'The Sensitivity of Conditional Choice Models for HospitalCare to Estimation Technique."Journal ofHealth Economics 8, no. 4 (1989):377-97.

Garnick, D. W., H. S. Luft, J. C. Robinson, and J. Tetreault. 'AppropriateMeasures of Hospital Market Areas." Health Services Research 22, no. 1(1987): 69-89.

Gleicher, N. 'Cesarean Section Rates in the United States: The Short-TermFailure of the National Consensus Development Conference in 1980."Journal of the American Medical Association 252, no. 23 (1984): 3273-76.

Hausman, J., and D. McFadden. "A Specification Test for the MultinomialLogit Model." Econometrica 52, no. 5 (1984): 1219-40.

Lee, H. L., and M. A. Cohen. "A Multinomial Logit Model for the SpatialDistribution of Hospital Utilization."Journal ofBusiness and Economic Statis-tics 3, no. 2 (1985): 159-68.

Luft, H. S., D. W. Garnick, D. H. Mark, D. J. Peltzman, C. S. Phibbs, E.Lichtenberg, and S. J. McPhee. "Does Quality Influence Choice of Hos-pital?" Journal of the American Medical Association 263, no. 21 (1990):2899-2906.

McCormick, M. C., S. Shapiro, and B. H. Starfield. "The Regionalization ofPerinatal Services. Summary of the Evaluation of a National Demonstra-tion Programn."Journal of the American Medical Association 253, no. 6 (1985):799-804.

McFadden, D. "Conditional Logit Analysis of Qualitative Cllbice Behavior." InFrontiers in Econometrics. Edited by P. Zarembka. New York: AcademicPress, 1974.

. "Regression-Based Specification Tests for the Multinomial LogitModel."Journal of Econometrics 34, no. 1/2 (1987): 63-82.

McGuirk, M., and F. W. Porell. "Spatial Patterns of Hospital Utilization: TheImpact of Distance and Time." Inquiry 21, no. 1 (1984): 84-95.

221


Rust, K. J., F. P. Rust, and R. L. Williams. 1981-1985 Materal and ChildHealth Data Base Descriptive Narrative. Monograph, Community and Orga-nization Research Institute, University of California, Santa Barbara,1989.

Williams, R. L. "Measuring the Effectiveness of Perinatal Care." Medical Care17, no. 2 (1979): 95-109.

Williams, R. L., and W. E. Hawes. 'Cesarean Section, Fetal Monitoring, andPerinatal Mortality in California." American Jiounal of Public Healh 69,no. 9 (1979): 864-70.

Documents

Choice of Hospital for Delivery: A Comparison of High-Risk and Low