STATISTICAL METHOD
A number of statistical techniques are available to analyze the relationships
among multiple variables. Multiple regression and discriminant
analysis are recommended when the nature of the research situation is
essentially exploratory (Goldstein and Dillon 1978; Klecka 1980). The
results obtained using multiple regression and discriminant analysis
are essentially comparable (Cleary and Angel 1984).
In their study, Herzlinger and Krasker used multiple regression
analysis, the most commonly used multivariate technique in the hospital
performance literature. Multiple regression was appropriate as they
hypothesized independent-dependent variable relationships. Independent
variables included a number of measures thought to influence
hospital performance. For-profit or not-for-profit status (e.g., control
status) was included as an independent variable. Several dependent
variables that reflected aspects of hospital performance (e.g., patient
revenues, patient days generated) were used as dependent variables.
Discriminant analysis is much less commonly used in health services
research than is multiple regression. Discriminant analysis allows
one to distinguish between two or more mutually exclusive groups on
the basis of a collection of discriminating (independent) variables that
measure characteristics on which the groups are expected to differ. The task is to describe the characteristic differences between or among the
groups.
In our studies, we have used discriminant analysis to focus not on
the exploration of independent-dependent variable relationships for
for-profit and not-for-profit hospitals, as did Herzlinger and Krasker,
but instead on the exploration of characteristics that discriminate
between for-profit and not-for-profit hospitals. The assumptions of
discriminant analysis (a number of mutually exclusive groups measured
at the nominal level) appear to fit the characteristics of this
problem more appropriately than those of multiple regression. In this
study, we used discriminant models with discriminating variables comparable
to Herzlinger and Krasker's independent variables. The
dichotomous dependent variable we used was control status, that is,
for-profit or not-for-profit status.
Three discriminant models were estimated. In the first model,
independent variables similar to those used by Herzlinger and Krasker
were employed. Lack of comparable financial and case-mix information
in our data set required that we substitute proxy measures for four
constructs; all other constructs were measured as in the Herzlinger and
Krasker study to maintain comparability of findings (Table 1).
We found that several of the independent variables used by
Herzlinger and Krasker were highly correlated with one another. Such
high intercorrelations (multicollinearity) can cause difficulties in model
estimation and interpretation, as mentioned by Reinhardt (1987). The
second model, therefore, included a subset of independent variables used
by Herzlinger and Krasker, modified to reduce multicollinearity
(or redundancy) among the discriminating variables (Table 1).
The third model included some new discriminating variables not
used in the Herzlinger and Krasker study (Table 1). To promote comparability
with Herzlinger and Krasker, this set of variables was similar
to those found in the first two analyses with, however, two important
differences. First, this set of variables was checked for multicollinearity
and was found to present low intercorrelations among the measures.
Second, the new discriminating variables were chosen to reflect work
previously done in the area of hospital performance, particularly studies
reported in For-Profit Enterprise in Health Care (Gray 1986), while not
straying too far from Herzlinger and Kraskeres construct formulation.
Each discriminant analysis was performed using three data sets.
The first analysis included all U.S. community hospitals belonging to
the 14 systems considered by Herzlinger and Krasker. In our analysis,
there were usable data for 563 of the 725 hospitals in the group. Our
second data set included all (N = 1,083) system-affiliated hospitals for which complete information was available in the AHA data set. Our
final data set included the 3,783 community hospitals in the AHA data
set for which data were complete.