Factors explaining the use of health care services by the elderly

Transcription

1 University of Washington From the SelectedWorks of Paula Diehr August, 1984 Factors explaining the use of health care services by the elderly Paula Diehr, University of Washington Connie Evashwick Available at:

2 Factors Explaining the Use of Health Care Services by the Elderly Connie Evashwick, Genevieve Rowe, Paula Diehr, and Laurence Branch The Andersen model of health services utilization, which relates use of service to predisposing, enabling, and needfactors, has not often been applied to an elderly population. In this study, thefactors of the Andersen model were usedprospectively to predict utilization for a population sample of 1,317 elderly persons. Taken alone, the NEED construct was the most important single predictor of use of physician services, hospitalizations, ambulatory care, and home care. PREDIS- POSING factors were better predictors of the use of dental services. Some of the variables studied were not related to utilization in the direction that would have been predictedfrom previous studies on general populations. Multivariate analyses demonstrated that the three constructs should be applied simultaneously when predicting use of services. These findings can be applied to the specific task of planning services for older people. The dramatic changes occurring in the age structure of American society make providing services for the elderly a major challenge to health care professionals. In 1900, only 4 percent of the population was age 65 or over [1]. Today, about 11 percent of our population falls in that category [2], and current projections indicate that by the year 2000 as much as 15 percent of the nation's population will be age 65 or older [3]. Furthermore, the cohort of persons age 75 and over is the single fastest-growing age group [4]. The elderly are disproportionate users of the health care system. Connie Evashwick is Vice-President for Long-Term Care, Pacific Health Resources in Los Angeles; Genevieve Rowe is a Research Associate with Compass Consulting in Bellevue, Washington; and Paula Diehr is an Associate Professor in the Department of Biostatistics, University of Washington, Seattle. Address all communications and requests for reprints to Dr. Laurence Branch, Associate Professor and Director of Massachusetts Health Care Panel Study, Harvard Medical School, 643 Huntington Avenue, Boston, MA

3 358 Health Services Research 19:3 (August 1984) They average 6.1 visits per year to a physician, compared with 4.1 visits per year by adults age [5]. They have more hospital admissions and longer lengths of stay than any other age group [6]. Also, 5 percent of those age 65 or older reside in nursing homes, occupying 90 percent of all long-term care beds in the United States [7]. In terms of expenditures for health care, 30 percent of the monies spent on personal health care are by persons age 65 or older [8]. Eight percent of all personal health dollars spent by older people are for nursing home care. Social services for older adults have increased markedly in recent years, due to greater societal awareness of their needs and the resulting federal and state funding for an array of services. With the 1965 passage of Medicare, Medicaid, and the Older Americans Act, the federal government assumed the responsibility for funding health and social services for those individuals age 65 and over. Trends in the use of services and dollars spent indicate that older people are, indeed, receiving care. Lobbying groups for older Americans and those working in the field of gerontology continue to advocate the provision of improved health and social services. Current reductions in federal, state, and local budgets will result in the curtailment of public support for some of these services for older people. To achieve a balance between the demand for services and the resources available to provide them, efficiency is needed in the planning and development of health services. Despite years of discussion on issues pertaining to "long-term care," the methodology for determining the appropriate mix of services required by an older population is not well developed. Many of the data on the use of health services by older adults focus on the recipients of a specific service. For example, extensive information is available on the characteristics of nursing home residents and the recipients of Medicare-reimbursed home care, as well as on the diagnostic conditions of those age 65 and over admitted to a hospital. Numerous studies have analyzed the use of health services by a select subgroup, such as the participants in model projects for low-income public housing residents. Few studies, however, have focused on a random sample of persons age 65 and above to determine the range of services used over time and the respective characteristics of the users. Research on the utilization of health and social services by older people has identified a set of variables associated with use. However, few studies have been methodologically rigorous in examining the interrelationships of the independent variables; most stop with correlations of the independent and dependent variables. Nor have many

4 Elderly and Health Care 359 studies been undertaken to compare the relative importance of the independent variables in explaining the use of different types of health services. Leading gerontologists contend that the older population is highly heterogeneous. Similarly, researchers and providers of long-term care are adamant in pointing out that not all services are appropriate for all older people. A great deal of effort has been devoted to developing assessment instruments, which are used for individual care plans and reimbursement calculations - but not for projecting the demand for services. To plan services for a population, a methodology is required that takes into consideration characteristics of various subsamples and a range of service options. An array of raw data are available from census, health interview survey, state licensing, and vital statistics reporting, but these data may not be detailed enough to permit utilization projections at a local level. Before extensive efforts are made to gather additional data, it is imperative to know which data are most pertinent to the planning issues at hand Ṫhe purpose of the study reported in this article was to identify the factors important in explaining the use of various services by a random sample of older adults and to examine the interrelationships of these variables. By using multivariate analyses, the study enabled us to identify the factors most important in explaining and predicting the use of different services. The findings are intended to be useful to those planning and developing services for older adults and to those involved in forming institutional and public policies. Research on the health care delivery system has identified a number of factors related to the use of health services. Andersen and colleagues have formulated a model that incorporates many of these factors in specified relationships. We chose this model as the basis of our study. METHODS OF DATA COLLECTION In 1974, the Massachusetts Department of Public Health initiated a series of studies to identify services which could be provided by the Health Department to maximize the quality of life of older persons while minimizing costs. Particular emphasis was placed on determining the need for nursing home versus other services. The goal of the department was to foster independence and enable elderly persons to

5 360 Health Services Research 19:3 (August 1984) remain in their homes and communities rather than going into nursing homes. The Massachusetts Health Care Panel Study began in 1974, and the first follow-up survey was conducted in The survey instrument focused on the respondents' ability to perform the basic activities of daily living with or without assistance, current use of health and social services, and perceived need for social support and health care. The descriptive findings of the two surveys have been presented elsewhere [9-10]. The Andersen model of health services use is based on characteristics of the population at risk, the resources of the health care system, and utilization [11]. The population at risk is characterized by predisposing, enabling, and need factors. Utilization includes the type, site, purpose, and time interval of services used. Resources include the volume, distribution, and organization of health care providers and provider organizations. Length and time limitations of the questionnaire did not permit collection of data on all variables specified by the Andersen model. However, the variables included in the instrument were sufficient to permit the application of the model to predict and explain the use of health and related services by older adults. PREDISPOSING and ENABLING variables came primarily from standard questions about demographic and economic characteristics. NEED variables came from a modified Katz scale [12], a modified Rosow-Breslau scale [13], self-reported health status, and self-reported problematic physical conditions. The six measures of utilization were Physician Visits, Hospitalization, and Nursing Home, Dental, Ambulatory Care, and Home Care services. The latter two were comprised of several types of care. Measures of system resources were not included in this data set. The PREDISPOSING, ENABLING, and NEED variables were taken from the 1974 survey. They were used to predict the utilization variables from the 1976 follow-up survey. The variables used in this study and their operational definitions are given in Table 1. The respondents for the survey were drawn from a statewide area probability sample of households. Residents of nursing homes or other institutions were not eligible. A total of 8,614 households were selected. All individuals of age 65 or older were interviewed, which produced 1,625 respondents. The sample was reinterviewed 15 months later. Of the 1,625 original respondents, 1,466 participated in the second wave, including 1,317 who completed interviews. Details of the sampling and interview procedures have been presented elsewhere [10,14]. Table 2 presents the distribution of the sample respondents by age, sex, and race. These are similar to national figures for the elderly.

6 METHODS OF DATA ANALYSIS Elderly and Health Care 361 The purpose of the study summarized here was to determine the extent to which the utilization of different types of health care services (i.e., each of the six dependent variables) could be predicted by the three categories of independent variables specified by the Andersen model and to ascertain the relative importance of each independent variable in predicting the use of each service. This would enable those projecting the demand for services to be selective in the collection of data necessary for planning. In an earlier paper, the relationships between the dependent and independent variables were examined by use of multiple regression analysis performed on the cross-sectional data from the first survey [15]. Our study advances the previous work in two ways. First, independent variables from Time 1 are used to predict utilization variables from Time 2, thus emphasizing the predictive, rather than associative, use of available PREDISPOSING, ENABLING, and NEED data. Second, the analysis is extended by use of principal component analysis to create a single index to represent each construct in the Andersen model. This simplification allows us to look at the relationships among the three constructs in a more interpretable form. The analysis was conducted in four steps. First, the zero-order correlation coefficients between the independent and the dependent variables were calculated to explore the relationships between individual variables and utilization. Second, multiple regression analysis was used to predict each utilization measure as a function of all 20 variables. Third, the 20 variables were divided into subgroups representing NEED, PREDISPOSING, and ENABLING, and the multiple regression analysis was repeated using only one set of variables at a time. The R2 measures from each regression were compared to the total R2 obtained when all 20 variables were used and to the R2 obtained from each of the other two sets of independent variables. This showed the relative importance of each set of variables in explaining a given dependent variable. Finally, an index was created from each set of independent variables so that each construct was represented by a single variable rather than by six or seven separate variables. This made it easier to see the direction of the relationship between each construct and the utilization measures, and it simplified examination of the relationships among the constructs. The first principal component of each group of variables was chosen as the index. This step of the analysis is discussed in more detail later. We then examined the relationship of each of these indexes to the

7 Ca VV "04 o > > S C ~~~~~~~~~~~~~~~~~~~~~~~ v~~~~~~~~~~~~ S~~~~~~~ :3S< X 2 i E D 2 UZC4~~~~~~C~ N H~~~~~~~~~~~~~~~~~~ z ci) 0 D~~~~~~~ Cd4- C4-4~~~~~~~~~~~~~~~~~~~~- ~~~~~~~~z0 0 "0 0n-" u0 4 *~ ~Z ~"0~x IL) co5 Ui P~~~~~~~~~~~~

8 bo pq 0 11 $.4 &4 -~4.0 C14 4.) $- 0 4) S 0 l) 0 U D.D,=> ~~ce) o o O O ~ z G 11-4) HII 2 41) ) if 11 H DI -- -~4) o_ I" 4)V Q (D C) C.. UD1.D 2, U 4 0~~~~~~ O0 o o0o00o0 0 Z 0. *t. &4 O. i Oi>.0> t _ ~40 _; cl 0000 L-z ;e _ 4) Ut 4s.t Q Z,2 0~~~.E) cd :w :U2E > C4) O4)d C = CI Y n 0 C ~~~~v _ Ct cn dr v 4)4 Ci e4) ;;... " _C C v~ Fs.~..~. O 4)) _ CO C 1- cd C

9 364 Health Services Research 19:3 (August 1984) Table 2: Age, Sex, and Race of Sample Respondents Age: %o % % % Sex: Female 61.7% Male 38.3% Race: White 99.0% Nonwhite 1.0% six utilization variables, separately and together, using multiple regression. This revealed the effect of each construct on utilization while controlling for the other constructs. The results of the analyses were summarized to make a general statement about the relative effects of NEED, PREDISPOSING, and ENABLING factors on utilization of health services by the elderly. These analyses are available in greater detail from the authors [16]. FINDINGS UTILIZATION OF SERVICES The overall levels of service utilization are shown in Table 3. Eightytwo percent of the respondents reported seeing a doctor during the 15- month interval between the initial and follow-up surveys. This is consistent with the data of the National Health Interview Survey. Nearly one-third of the respondents had seen a dentist within the past 15 months. About one-fifth had been admitted to a hospital. These findings are also what would be expected based on national data. Very few persons reported using other types of health or healthrelated services. At most, 4 percent of the respondents had used any rehabilitation services, counseling services, speech therapy, hot meals, visiting nurse service, homemaker or home health aide, or other special care. The low utilization rates for such services are in themselves revealing. The small number of persons using these types of auxiliary and support services placed some limitations on further data analyses. Rather than analyzing each service separately, the services were combined into two variables: the Home Care variable, which included

10 Elderly and Health Care 365 Table 3: Utilization of Services Percentage of Respondents Using Services During Service Preceding 15 Months Physician services 82.4 Dental services 31.8 Hospital services 22.0 Nursing home services 1.1 Ambulatory care services Speech therapy 0.5 Rehabilitation therapy 2.8 Professional counseling 0.8 Home care services Nursing services 4.0 Hot meal service 1.2 Homemaker/Home health aide service 2.3 Special care services 4.0 visiting nurses, homemaker, home health aide, hot meals, and special care, and the Ambulatory Care variable, made up of rehabilitation therapy, speech therapy, and professional counseling. CORRELATIONS BETWEEN INDIVIDUAL VARIABLES AND UTILIZATION Table 4 presents the zero-order correlations between each PREDIS- POSING variable and each utilization variable. Rather than present numbers, the direction of each relationship is shown, and the statistical significance of the relationship is indicated by asterisks. For example, examining the variable AGE: older people used more Physician Services, more Hospital Services,.and more Ambulatory Care Services; significantly more Nursing Home and Home Care Services; and significantly fewer Dental Services. PREVENT (seeing physician on a regular basis rather than on a problem basis) is significantly related to more of the utilization variables than any of the other PREDISPOS- ING measures, and dental services is significantly related to more of the PREDISPOSING variables than any of the other utilization measures. Some discrepancies between these results and the Andersen model appear: those with more education used fewer physician and home care services and whites used fewer home care services. Table 5 presents similar information for the ENABLING variables. The Andersen model would predict that the more "able" would use more services, but in many cases a relationship is found in the

11 366 Health Services Research 19:3 (August 1984) Table 4: Relationships between PREDISPOSING Variables and Utilization Variables Nursing Ambulatory Home Physician Hospital Home Care Care Dental Services Services Services Services Services Services AGE Older use more more more* more morel less t SEX Females use more* less more more more more* RACE Whites use more more more more lesst less EDUC Higher use less* less more less less* morel ALONE Live alone use less more less less less more MARRIED Married use less more less less* less* more WDS Widowed use more more more more more lesst Divorced use more less more more more less Separated use less less less less less less PREVENT Have reg. visits use moret morel less* morel more morel *p <.05. tp..01. opposite direction. This is particularly true of TRANSPO, which is significantly related to all of the utilization variables except dental services, in the "wrong" direction. VAINS (Veterans Insurance) is not significantly related to any of the utilization variables, and hospital and nursing home services are not significantly related to any but the TRANSPO variable (in the wrong direction). These discrepancies might, of course, be caused by confounding of these ENABLING variables with age, health status, and income, suggesting that a multivariate analysis is needed to control for some of these other variables. Table 6 presents the information for the NEED variables. Here, nearly every variable is significantly related to utilization. The major discrepancy is for dental services, in which the more "needy" tended to have fewer services rather than more. As above, multivariate analysis is needed to remove the effect of possible confounding factors.

12 Elderly and Health Care 367 Table 5: Relationships between ENABLING Variables and Utilization Variables Nursing Ambulatory Home Physician Hospital Home Care Care Dental Services Services Services Services Services Services OCCUP White collar use less more more more less morel INCOME Higher use less less less less less$ morel MEDICAID With use morel more less more* moret lesst VAINS With use more more less less less less PRIVINS With use more more less more more less t OWNDOC Have use morel more more less less more TRANSPO With no problems use lesst less* less* lesst lesst more* *p.05. lp <.01. Table 6: Relationships between NEED Variables and Utilization Variables HLTHSTAT Poorer use PHYSACT Problem use STAIRS Problem use WALK'/2MIL Problem use PHYSCOND Have use ADLFUNC Poorer use tp <.01. Nursing Ambulatory Home Physician Hospital Home Care Care Dental Services Services Services Services Services Services morel morel morel morel morel lesst morel moret more morel morel less$ morel morel morel morel morel less$ morel morel morel morel morel lesst morel morel more morel morel less$ moret moret moret morel morel less

13 368 Health Services Research 19:3 (August 1984) MULTIPLE REGRESSION PREDICTING UTILIZATION FROM INDIVIDUAL VARIABLES Table 7 shows the R2, or proportion of variance reduction, when all 20 variables were included in a multiple regression equation to predict each of the utilization measures. Nearly 24 percent of the variability in Physician Services was explained, but only about 3 percent (not statistically significant) of the variability in Nursing Home Services was explained by the 20 variables. The last three lines in Table 7 show the R2 achieved when only variables from a given subset were used to predict utilization. For example, the PREDISPOSING variables taken alone explained 10.5 percent of the variability in Physician Services, compared to 6.3 percent explained by ENABLING variables and 16.5 percent explained by the NEED variables. No equation was statistically significant for Nursing Home Services, and only NEED variables were significantly related to Hospital and Ambulatory Care. In most cases, the NEED variables explained more of the variability than did the other two categories. The major exception was Dental Services, where the PREDISPOSING and ENABLING variables accounted for more of the variability than the NEED variables. The amount of variability accounted for by PREDISPOSING and ENABLING variables was about the same for all utilization measures except Physician Services, where the PREDISPOSING variables were substantially better predictors than the ENABLING variables. Table 8 shows the variables which were chosen by forward selec- Table 7: Squared Multiple Correlation Coefficients (R2) of Variables in the NEED, PREDISPOSING, and ENABLING Groups Predicting Six Utilization Variables Utilization Variables Independent Nursing Ambulato?y Home Variable Physician Hospital Home Care Care Dental Groups Services Services Services Services Services Services All Variables (N = 839) * *.1348*.1054* PREDISPOSING (N = 1199) * *.0865* ENABLING (N = 1065) * *.0712* NEED (N = 1069) *.0452* *.1190*.0277* *Significant, p <.001. N + changes because of missing data.

14 Elderly and Health Care 369 tion regression analysis with F-to-enter set at 4.0. This analysis was performed three times, once for the variables from each of the constructs. In most cases only one or two variables were chosen from each construct. Among the PREDISPOSING variables, Preventive, Age, and Marital Status were chosen more often than the others; Sex and Household Composition were never chosen. Among the ENABLING variables, the two variables chosen most often were Transportation and Medicaid. Veterans Insurance, Private Insurance, and Medicare were not chosen for any of the utilization measures. Among the NEED variables, Health Status was chosen most frequently, and the presence of a Physical Condition and Problems in Climbing Stairs were also chosen frequently. The Inability to Walk Half a Mile was chosen only as a predictor of Hospitalization. The directions of the relationships shown in Table 8 were quite similar to those shown in Table 6, which presented the univariate relationships of the PREDISPOSING and utilization variables. These findings were also generally consistent with what would be expected based on previous studies. However, older people had fewer Dental Services; the predisposition to see a physician on a regular basis rather than on a problem basis (PREVENT) was negatively associated with Nursing Home Services; and whites had fewer Home Care Services. Among the ENABLING and NEED variables, all relationships remained the same in the multiple regression results as in the correlation results. Again, people with Transportation problems used significantly more services, which does not agree with the Andersen model. The relationships among the independent variables and Dental utilization were consistently in the opposite direction from the relationships of the other utilization variables. This will be discussed in some detail later. As has been noted, some variables were significant predictors of utilization but not in the direction expected based on current theory. It seemed inappropriate to make statements about the relative strengths of constructs when some of the variables making up the constructs were not acting according to expectations. To address this problem, we reduced each of the three sets of independent variable constructs to a single index, making it easy to see the direction of any effects. The use of a single index rather than many individual variables also facilitated comparison of the constructs. The method of principal components [17] was used to determine a single "best" index from each set of variables. This index is a weighted sum of the individual variables, which has maximal correlation with the individual variables. Thus, if the variables have something in com-

15 370 Health Services Research 19:3 (August 1984) Table 8: Significant Variables from Stepwise Regression* Predicting Utilization Nursing Ambulatory Home Physician Hospital Home Care Care Dental Variabkls Services Servies Servie Srvices Sevices ervices PREDISPOSING: AGE (older) more less EDUC (more) more PREVENT (yes) more more less more MARRIED less less WDS less RACE (white) less ENABLING: INCOME (higher) MEDICAID (have) more more OCCUP (white collar) OWNDOC (have doctor) more TRANSPO (no problems) less less less less NEED: more more PHYSACT (problems) more STAIRS (problems) more more more WALK1/2MIL (problems) more HLTHSTAT (worse) more more more more less FUNCSTAT (poorer) more PHYSCOND (if any) more more more *F-to-enter was set at 4.0. mon (e.g., they are all measures of "need"), then the resulting index will also be a measure of "need," and will account for more of the information in the original set of variables than will any other such index Ṫhe Andersen model does not state that each of the three constructs should be reducible to a single dimension; however, this "best" index will be a good representation of the construct if the original

16 Elderly and Health Care 371 variables share the property of being related to the underlying construct. If some of the original variables are not highly correlated with the resulting index, this suggests that for some reason they are not highly correlated with the other variables in the construct. This does not invalidate the index but suggests that the "omitted" variables are not related to the construct as predicted. This could happen because the variable is more related to some other factor than to the construct, or because it represents a different dimension of the construct than the dimension being represented by the index. These points will be illustrated in the following description of the three indexes. The NEED index will be discussed first, since it is the easiest to understand. Principal component analysis of the six NEED variables resulted in the index: NEED =.50(PHYSACT) +.91(STAIRS) +.65(WALK1/2MIL) +.27(HLTHSTAT) +.29(ADLFUNC) +.42(PHYSCOND) This index accounted for 48 percent of the variability in the original six variables and, therefore, provided a reasonable summary of the information available on a person's self-reported health status and inferred need for care. As the signs of the coefficients of each variable indicate, the NEED index became larger for sick people and smaller for healthy people, and was thus a good representation of the construct NEED. Each of the individual variables was correlated approximately 0.7 with NEED, which was further validation that the index was a measure of need. Analysis of the ENABLING group was less satisfactory. This might have been expected since such diverse factors as income, insurance, type of occupation, having a regular doctor, and difficulty with transportation were all included in this group. The first principal component was: ENABLE = +.09(INCOME)-.75(MEDICAID)-.34{VAINS) +.77(PRIVINS) -.73(OCCUP) -.07(OWNDOC) -.34(TRANSPO) This index accounted for 24 percent of the variability in the seven ENABLING variables. It was not highly correlated with Veterans Insurance or with Having a Particular Physician. A reason for the former could be the small number (2.3 percent) of respondents who have veterans insurance. The relationship of OWNDOC to the other ENABLING variables is discussed later. Based on the signs of the statistically significant relationships, people with a low value on this

17 372 Health Services Research 19:3 (August 1984) index tended to have low income, to be on Medicaid, to have no private insurance, to be blue-collar workers, and to have difficulty in obtaining transportation. Except for the low correlation with the two variables mentioned above, this index had face validity as a summary of the ENABLING factors. Finally, the seven PREDISPOSING variables were analyzed. The resulting index was: PREDISPOSING =.1 (AGE) -.39(SEX) -.02(EDUC) +.05(PREVENT) -.71(MARRIED) +.70(WDS) +.16(RACE) +.64(ALONE) The index accounted for 33 percent of the variability in the eight PREDISPOSING variables. The signs of the coefficients show that a person who is older, female, not married, who is divorced, widowed or separated, and who is living alone, less educated, white, and seeing a physician only for problems would have higher values on this index. The signs of most of the variables were in the expected direction, suggesting that this was indeed an index of the predisposition to use services. Three of the variables were not significantly correlated with the new index (Race, Education and Prevention). The fact that Race is not represented by the index seems a small loss since virtually all of the respondents were white. The fact that Education and Prevention are not represented by the index is unsatisfying, especially since PRE- VENT was often the first PREDISPOSING variable chosen in stepwise regression. Thus, although the index had face validity as a measure of the construct PREDISPOSING, it did not offer a good representation of all of the variables in the group. It may be that predisposition (as operationalized here) is not a simple construct and cannot be represented by a single underlying dimension. The principal component analysis showed that two indexes would be required to account for 43 percent of the variability in the PREDISPOSING variables, and that a third index would bring this up to 57 percent. To explore the dimensionality of the PREDISPOSING construct further, the first three principal components, all of which had eigenvalues greater than one, were orthogonally rotated. The resulting factors did emphasize three different components of the PREDISPOS4 ING construct. The first factor had high loadings for ALONE, MAR- RIED, WDS, and SEX; the second factor had high loadings for AGE and EDUCATION; and the third factor had a high loading for PRE- VENT. Race never loaded on any of these factors, which-may be due

18 Elderly and Health Care 373 to the low number of nonwhites in the sample. Adding these additional factors would have defeated the objective of the index construction, which was to reduce the number of independent variables. For this reason, only the first principal component was used in the following analyses since it is a representative of PREDISPOSING factors and has some optimal properties. Table 9 shows the relationships among the three indexes for the 887 persons for whom complete information was available on all of the variables. All three indexes were significantly intercorrelated. ENA- BLING was negatively correlated with both PREDISPOSING and NEED, showing that subjects who were less "able" had more PREDIS- POSING factors and more NEED. Those subjects with more PRE- DISPOSING factors also had more NEED. These interrelationships suggest that multivariate analysis is necessary to identify individual contributions of the constructs. Table 10 shows the zero-order correlations between the three indexes and the six utilization variables. Physician Services, Hospital Services, and Ambulatory Care had significant positive correlations with the NEED index, but were not significantly correlated with the other indexes. Nursing Home Services had a significant positive correlation with both the PREDISPOSING and the NEED indexes. Home Care was related negatively to the ENABLING factor and positively to the NEED factor. Dental Services had a significant positive correlation with the ENABLING index and was negatively correlated with the NEED index. The NEED index was significantly related to all of the utilization measures: the "needy" had fewer dental services and more of the other five types of utilization. This is similar to the results found earlier for the regressions using the sets of variables (Table 8). The ENABLING index was not significantly related to Physician Services, Hospital Services, Nursing Home Admissions, or use of Ambulatory Care. The subjects who were the least "able" to obtain -services were significantly more likely to haye Home Care services. The more "able" subjects were significantly more likely to have Dental care. The PREDISPOSING index was significantly correlated only with Nursing Home Services, with those "predisposed" using more. Table 11 shows the results of stepwise regression using the three indexes in a single equation to predict the six dependent variables. The results are very similar to those of Table 10. For Hospital Services, Ambulatory Care, and Home Care, NEED was the only independent variable entering the equation. Both the NEED and PREDISPOSING indexes were important in explaining Nursing Home use, and the

19 374 Health Services Research 19:3 (August 1984) Table 9: Zero-Order Correlations between PREDISPOSING, ENABLING and NEED Indexes (N = 887) PREDISPOSIG ENABLING NEED PREDISPOSING 1.0 ENABLING -.30* 1.0 NEED.10* -.35* 1.0 *p >.067 is significantly different from zero. PREDISPOSING: High value means old, female, divorced, lives alone. ENABLING: High value means more able, e.g., higher income, more NEED: insurance. High value means more health problems and sicker. Table 10: Zero-Order Correlations between Three Indexes and Six Utilization Measures (N = 887) Nursing Ambulatory Home Physician Hospital Home Care Care Dental Services Srvices Services Services Sevices Services VDDPThZD1CQTW0 nn1 _ Al % ARA* A'%7 (91 - Al ENABLING ' NEED.349*.193*.Ot31 *.187* *p >.067 is significantly different from zero.. %/ *.274* -. VJ1.257* -.117* ENABLING index was significantly related to Dental Services. For Physician Services, NEED was significant, as in the zero-order correlation findings. However, the coefficient for the ENABLING index also became positive and significant after controlling for NEED, indicating that those persons who were more "able" did have more visits. The apparent negative relationship between Physician Services use and ENABLING found in Table 10 occurred because the "less able" were also "more needy." The other change from the simple correlations was that NEED was no longer a significant negative predictor of Dental Services use after controlling for ENABLING. The apparent negative relationship between NEED and Dental Services occurred here because the less needy were more able, and the more able used more dental services. The NEED and ENABLING indexes thus tended to show inverse relationships with the dependent variables when the relationships were examined independently. When examined simultaneously, however, the nature of the relationships became clearer. Thus, the multiple regression analysis of the indexes clarified some of the apparent discrepancies found in the other analyses.

20 Elderly and Health Care 375 Table 11: Significant Relationships from Stepwise Regression* Using Indexes to Predict Utilization Nursing Ambulatory Home Physician Hospital Home Care Care Dental Services Services Services Services Services Services PREDISPOSE more ENABLE more more NEED more more more more more *F-to-enter was set at 4.0. At this point in the analysis, the interactions of the three indexes were examined. The regression analysis using the three indexes to predict the six utilization measures was repeated and, after controlling for the main effects, the first-order interactions were allowed to enter the equation in a stepwise manner. The only significant interaction was that of ENABLE and NEED as a predictor of Dental Services. The direction of this relationship is consistent with what was previously discovered; those less needy and more able used more dental services. We were also interested in whether the relationships between the indexes and the measures of use would hold if the regression analysis was repeated; once, using those respondents age 65-74, and again, using those age 75 and older. The results were the same with respect to the direction of associations. Some relationships were no longer significant, but this was due to a large decrease in sample size and not to a change in the magnitude of the association. Several times during the analysis, questions arose regarding the appropriateness of the three groups of variables. The principal component analysis suggested more than one underlying dimension for both the ENABLING and PREDISPOSING constructs. To address this issue, factor analysis was performed using the variables from the three constructs combined into one group. If NEED, PREDISPOSING, and ENABLING are the underlying constructs being measured, we would expect the variables to factor into three groups, each with high loadings for the variables in one group. In fact, five factors were identified, which explained 56.6 percent of the total variation among all of the variables. The first factor clearly measured NEED. There were high positive loadings for all of the variables in this group. In addition, there was a high positive loading for Transportation. It seems that the Transportation variable measures problems in getting to a health care location in terms of functional status and not as an ENABLING characteristic.

21 376 Health Services Research 19:3 (August 1984) This variable was associated with the dependent variables in the same way as the NEED variables were in the correlation and regression analyses. The second factor consists of several of the PREDISPOSING variables: ALONE, MARRIED, WDS, and SEX. The third factor has high loadings for both PREDISPOSING and ENABLING variables that pertain to socioeconomic status (SES): EDUCATION, OCCUPATION, INCOME, and PRIVATE INSURANCE. HEALTH STATUS is also represented here. High values for this factor are obtained for those who are more educated, have higher incomes, have held white-collar jobs, have private health insurance, and enjoy better health status. The significant loading for HEALTH STATUS is not inconsistent with the accepted theory that those with higher SES have better health status. The fourth factor represents those who have a regular source of care (OWNDOC), have regular visits (PREVENT), and do not have a condition that bothers them (PHYSCOND). This seems to describe preventive behavior. RACE and VETERANS INSURANCE were omitted from all factors since they represented such small subpopulations. Clearly some variables belonged with a different group, an.d some groups should have been separated into two or even three groups. For this sample, the constructs apparently defining the population at risk are: "Need," "Living Situation," "Socioeconomic Status," "Preventive Behavior," and "Age." IMPORTANCE Due to the large sample size, the study produced many results which were statistically significant while having small R2 values. Since "statistical significance," "size of R2," and "importance" are often confused, we have constructed a simplified example. Table 12 shows the results of predicting Hospital Utilization from the presence of Physical Conditions (R2 =.03); Physician Services from Health Status (R2 =.11); and Dental Services from Usual Occupation (R2 =.05). All of these regressions were statistically significant, although the R2 values were small Ṫhe fourth column of Table 12 shows the average utilization rates obtained in the survey. The fifth column contains the estimated utilization rates for different levels of the independent variables. The final column presents 95 percent confidence intervals for differences in those levels.

22 Elderly and Health Care 377 The differences are impressive, despite the fact that the R2 is not large. For example, the hospital admission rate, which was 216 per thousand overall, was 280 per thousand for people with a physical problem, compared to only 130 per thousand for those without. The difference between these two groups is 150 per thousand and this difference has a 95 percent confidence interval from admissions. Thus, we are 95 percent confident that in the large population from which these data are a sample (i.e., persons over age 65 living in the community), the hospitalization rate for the one group is at least 100 admissions per thousand higher than for the other. For people in "excellent" health, the Physician Visit rate was estimated at 1.9, while for those in "poor" health, the rate was 6.6 visits. The 95 percent confidence interval for the difference between "excellent" and "poor" health is visits per year, again an important consideration for planning purposes. With respect to annual dental services, the mean is.32 per person; but it is.45 for white-collar workers versus.24 for blue-collar workers. The 95 percent confidence interval for the difference is dental visits per year. Thus, the difference between the occupation groups in total dental visits would seem to be very large. A low value of R2 shows that we cannot do a good job of predicting the utilization of an individual from these data. In aggregate, however, we can predict average utilization in population subgroups with considerable confidence. The factors identified in this analysis are thus important as well as statistically significant, despite the low R2 values obtained. DISCUSSION The purpose of the study described in this article was to try to identify the factors explaining utilization of health and social services by older people. This identification would provide information enabling us to develop a methodology that could be used in developing a rational base upon which to plan services for the older population. The findings on utilization are in themselves interesting: in general, the older population residing in the community made minimal use of any services other than physician care and hospital care. The simple correlation of the independent variables related to utilization showed several noteworthy patterns. Of the PREDISPOS- ING variables, which are basically demographic characteristics, no single factor was consistently a statistically significant predictor of utilization of the six categories of service.

23 378 Health Services Research 19:3 (August 1984) Most of the relationships tended to move in directions that would be expected based on the results of other studies. In general, older persons use more physician and home care services, have more hospital and nursing home admissions, and have fewer dental care visits; and those trends were exhibited by the population under study. Females in the adult population use more health services than males, and this relationship held here except for the use of hospital services. Differential use between males and females does, however, diminish with age and, indeed, sex was not an important predictive variable in the latter stages of the analysis. Except for nursing home care, persons living alone used fewer services. The explanation may be that persons who are healthier are able to continue living alone, while those in poorer health are not able to manage if they live by themselves, and thus live with others who are able to provide assistance. As would be expected, the level of education was positively correlated with use of dental care, but education was negatively correlated with use of four of the six service categories. The correlations of race with the utilization variables were inconsistent with what would be expected. However, this may have been due to the small number of blacks and other minorities in the sample. The correlation of the ENABLING variables and the six dependent variables would support the contention that financial barriers to access of care have largely been removed, except for dental care. However, the relationships between income, occupation, education, and the three types of insurance are fairly complex and should be examined further. Those who reported having a regular physician used services more, except for home care and ambulatory services. This may indicate that entry into the health care system still depends, to some extent, on having a personal physician to contact and that those who do not have a physician turn to alternative sources of care. Transportation was the only independent ENABLING variable that was consistent in showing a statistically significant correlation with the utilization variables. Those who said that transportation posed a problem for them also reported using more physician services, more ambulatory care, less dental care, more hospital and nursing home care, and more home care services. It is likely that those who use more services are sicker and thus find mobility and transportation more of a problem than those who are healthier. The factor analysis supported the idea that transportation problems could be considered a measure of need. Despite the fact that transportation is reported as a problem, the high service use by these respondents indicates that transportation does not present an insurmountable barrier to obtaining care.

24 Elderly and Health Care 379 It is worthwhile to look at the sample correlations from the perspective of the six utilization variables. One of the consistent trends was that of the direction of the relationships of independent variables with Dental Services-in the direction opposite that of their relationships with Physician Services use. For example, use of dental services was greater among younger persons, persons with higher education and/or higher income, and persons who had relatively good functional ability and self-perceived health status. Of the six services, all except dentistry were used more by people having poor health status, as indicated by each of the six independent variables. The dependent variable for use of dental services showed an inverse relationship with each of the NEED variables. This finding would be consistent with efforts to improve access to care. Changes in funding and coverage have concentrated on medical services and have included, to varying degrees, hospital, nursing home, and physician services, and ambulatory and home care services. Dental services have been excluded until recently, and this appears to be reflected in these data. Furthermore, dental care is often elective. Thus, despite the fact that many older persons have trouble with their teeth (false or natural) and gums, persons with other illnesses and limited resources would feel less need to seek dental care than other types of care. Multivariate analyses were used to clarify the findings of the univariate analyses. The results suggest that caution should be used in assessing the impact of one of the three constructs independent of the others. For instance, when only one factor at a time was considered, the results showed that high NEED was related to more utilization of most services but that it caused less demand for dental care; that PREDISPOSING factors were significantly related only to nursing home admissions; and that the ENABLING factors were highly related to dental use, but that the more able tended to use the other services less. The three variables considered simultaneously, however, showed that ENABLING factors were significantly and positively related to physician utilization and that NEED was neither significantly nor inversely related to dental use once the ENABLING factors were controlled. Thus, the inverse relationships between the NEED and ENA- BLING variables and the dependent variables, which were exhibited when the constructs were examined separately, became clarified when both types of variables were considered simultaneously. When all of the independent variables were used, either separately or as indexes, the R2 was generally low. While we have demonstrated that the findings are important in the prediction of the use of groups of people, the low R2 values demonstrate the large amount of variability

25 380 Health Services Research 19:3 (August 1984) still unexplained by the Andersen model with respect to the use of services by an individual. One possible reason for this lack of "fit" is that we have operationalized the model using only the variables available - which may or may not have been complete or optimal. Another possibility is that there exist major explanatory dimensions to the utilization of services which were not examined in this study. As noted, Andersen's model has three types of components: individual characteristics, system resources, and utilization measures. Due to the limitations of the data set, we examined -only the individual characteristics and utilization. The findings suggest that further analysis involving the system component of the model is warranted. A second consideration regarding the model is suggested. Andersen's framework was developed during the 1960s when access to services was one of the major issues of the health care delivery system. Major changes in the health care system, particularly Medicare, Medicaid, and the Older Americans Act, were designed to decrease financial barriers to access of care and to encourage the development of services to improve the availability of care. The elderly and poor, populations with particular disadvantages in access to care, thus were hypothesized to underutilize services. A model of utilization which emphasized demographic and economic characteristics was appropriate. Efforts to reduce discrimination in access focused on these groups. To the extent that barriers to access based on age and/or income were diminished between 1965 and 1975, and no longer differentiated the population, a model using these as predictive variables would not be successful in predicting utilization differences. Because recent changes in public policies and funding potentially could reinstate barriers of access, this model may be useful in explaining changes in utilization that occur during the next few years, The Andersen model may also be better in explaining service use for a total population than for a specific segment of the population. By selecting those who are old, have broad financial coverage, and have other similar characteristics (such as the high likelihood of having a chronic condition), the study may have reduced the variability that otherwise would be available for attribution. The relative importance of the three sets of variables is noteworthy. The PREDISPOSING and ENABLING variables are of minor predictive value compared to the NEED variables. Although Table 12 demonstrates that even "weak" variables can predict important differences in utilization, these findings support the position that planning future services based only on demographic and economic characteristics would produce a less refined prediction of service use than would estimating use based on the health status of the population. The utiliza-

26 Elderly and Health Care 381 Table 12: The Importance of (Apparently) Low R2 Values Utilization Measure Correlate R2 Mean Value Hospitalization PHYSCOND /thousand Physician visits EGFP Dental visits Usual Occupation Utilization Measure Estimate at Various Levels 95% Confidence Interval for Difference Hospitalization Yes: 280/thou. (Yes-No) (100,200) adm/thou No: 130/thou. Physician visits Excel: 1.9 (poor-exc) (3.9,5.6) visits Good: 3.5 Fair: 5.1 Poor: 6.6 Dental visits White col:.45 (white-blue) (.15,.27) visits Blue col:.24 tion of services appears to be related to physical and functional status far more than to economic status. Although this is well known to those who have been working in the field of gerontology and long-term care, it is not understood by many of the health care administrators and planners who are only recently or peripherally involved in services for the older population. Admittedly, health status is far harder to measure and estimate than are demographic and economic status. However, our findings suggest that the cost of collecting such data might be offset by the improved predictive power of the model. Those planning health services for older people would like to find direct formulas, certainly, to make it easier to predict the amount of services that will be demanded. The results of this study indicate that the application of utilization models as currently formulated is not simple; it does not produce clear-cut predictions of the use of services by the population. Insight to the relative importance of certain factors in explaining utilization of services on an aggregate basis may result, however, from refinement of these models. REFERENCES 1. U.S. Department of Commerce, Bureau of the Census. Historical Statistics of the United States: Part 1, Series Washington, D.C.: U.S. Government Printing Office, 1975, p U.S. Department of Commerce, Bureau of the Census. Statistical Abstracts of the United States, Washington, D.C.: U.S. Government Printing Office, 1978, Table No. 5, p. 8.

27 382 Health Services Research 19:3 (August 1984) 3. Statistical Abstracts, Table No. 5, p Robert Butler, M.D., Director, National Institute of Aging, in a speech at The University of Washington, Institute on Aging, March 29, U.S. National Center for Health Statistics. Current Estimatesfrom the Health Interview Survey: U.S DHEW/PHS, Series 10, No. 85. Rockville, MD, September 1973, Table No U.S. National Center for Health Statistics: Utilization of Short-Stay Hospitals: U.S DHEW/PHS, Series 13, No. 19. Rockville, MD, June 1975, Table B, p Hing, E., and A. Zappolo. A comparison of nursing home residents and discharges from the 1977 National Nursing Home Survey: United States. Advancedata/DHEW/NCHS (29):2, May 17, 1978, Table No Statistical Abstracts, Table No. 145, p Branch, L. G., and F. J. Fowler. The Health Care Needs of the Elderly and Chronically Disabled in Massachusetts. Boston: Center for Survey Research, Branch, L. G. Understanding the Health and Social Services Needs of Peopk over Age 65. Boston: Center for Survey Research, Andersen, R., and J. Newman. Societal and individual determinants of medical care utilization. Milbank Memonrial Fund Quarterly 51:95-124, Winter Katz, S., T. D. Downs, H. R. Cash, and R. C. Grotz. Progress in the development of the index of ADL. Gerontologist 10:20-30, Rosow, I., and N. Breslau. A Guttman Scale for the aged. Journal of Gerontology, 21:556-59, Branch, L. G., and F. J. Fowler. Methods and Technical Considerations to Accompany the Health Care Needs of the Elderly and Chronically Disabled in Massachusetts. Boston: Center for Survey Research, Branch, L. G., A. M. Jette, C. J. Evashwick, M. Polansky, G. Rowe, and P. Diehr. Toward understanding elders' health services utilization. Journal of Community Health 7:80-92, Rowe, G. Use of Health Services by the Elderly, Unpublished Master's Thesis. University of Washington, Department of Biostatistics, Seattle, Keinbaum, D., and L. Kupper. Applied Regression Analysis and Other Multivariate Methods. North Scituate, MA: Duxbury Press, 1978.