Statistical analysis of interdependence of country health resource variables, with special regard to manpower-related ones*

Bulletin of the World Health Organization, 59 (1): 129-141 (1981) Statistical analysis of interdependence of country health resource variables, with special regard to manpower-related ones* T. FOLOP1 & W. A. REINKE 2 The analysis reported here was the latest in a series of efforts to clarify the relative importance of the health system and ofsocioeconomicfactors to a nation's general level of health. The study has also quantified national and regional deviations from the general pattern as a basisfor selective investigation of the effects ofplanned interventions. The analysis was unusually comprehensive in that it included 131 WHOMemberStates. As in a number ofother studies, socioeconomicfactors werefound to accountfor much of the national variation in life expectancy. Inclusion of health resource variables added a special lagged effect which ultimately accounted for 90% of life expectancy variation. Evidence was also obtained that socioeconomic factors may operate partially through the development of health resources. It appears, therefore, that though socioeconomicfactors are necessarily linked to health improvement, they are not sufficient in the absence of corresponding development ofa viable health services infrastructure. Residual deviationsfrom the general pattern varied systematically by WHO region in 30% ofthe cases. Most notably, in theAfrican Region the numberofphysiciansis well below even the modest level expected on the basis of the socioeconomic situation in the region. There is, however, considerable variation within individual countries, and it was notpossible to find any significant relationship between the WHO manpower developmentprogramme and the national health resourceparameters. It is therefore concluded that statistical analysis is of limited applicability in this field. Since its inception over 30 years ago, the World Health Organization (WHO) has cooperated with its Member States in their efforts to improve the health of their populations by addressing, inter alia, both qualitative and quantitative issues of health man- power development (HMD). However, there has been little systematic effort to document the relationship between enunciated policy, objectives, programme setting and implementation, and ultimate outcome at country level. There seems therefore to be a need for such a study, since analysis of the reasons for success or failure in the past may be helpful in planning for the future. The methods selected for the analysis include docu- ment analysis, literature research, expert opinion study, country case studies, and statistical analysis. * This study was carried out as part of an analytical review of the World Health Organization's Health Manpower Development Programme, 1948-1978. 1 Director, Division of Health Manpower Development, World Health Organization, 1211 Geneva 27, Switzerland. 2 Professor, International Health and Biostatistics, School of Hygiene and Public Health, Johns Hopkins University, Baltimore, Maryland, USA. The present report highlights some of the results of the statistical analysis. The fundamental issues that had to be studied were the interrelationships between socioeconomic development, health services and manpower develop- ment, and health status improvements. It was hoped that the results of this study could then be used to provide some indication of the extent to which WHO/ HMD programmes have promoted health services and manpower development and thus health status improvements in Member States. Some of these issues have already been intensively studied but the results have been mixed because of the lack of clarity in the relationships of interest, the number of inter- dependent factors involved, and the lack of reliable data. An extensive bibliography compiled at Johns Hopkins University (1) served only to catalogue investigative approaches and unresolved issues. More recently, Barlow (2) presented a general model of health and development which depicted schematically the possible relationships between income, education, fertility, nutrition, and health, and centred on the way 4040 -129- 130 T. FIJLOP & W. A. REINKE in which health is affected by the other four factors. Barlow further classified analysis techniques accord- ing to whether the investigation was at the macro- or micro-level and whether a single-equation model or a set of simultaneous equations was employed. Of the five macro-level studies cited, all were of the single- equation type. Two of these (3, 4) involved health, education, and income only, while the other three (5, 6, 7) also included nutrition. Conclusions reached from large-scale studies vary substantially. Arriaga & Davis (8), Preston (7), and Stolnitz (9) are persuaded that public health activities deserve major credit for measurable improvements in health status. Fuchs (10), FQh5p (11), Krishnan (3), Scepin (12, 13), Smucker (14), Venediktov (15) and others, however, stress the importance of socio- economic improvements. In an earlier WHO study (16), simple correlation analysis of data from 68 countries yielded high associations between life expectancy and both per capita income and health resource factors such as medical density and bed/ population ratios. However, these investigators cautioned about the quality of the data and apparent differences in relationships between rich and poor countries. Examining evidence from a historical perspective, Grosse & Perry" and Preston (7) have suggested that health conditions, e.g., environmental sanitation, may be increasingly important deter- minants of life expectancy in place of socioeconomic factors, which were formerly dominant. In view of the pattern of ambiguous and conflicting findings, it is understandable that some studies have been limited to selected populations for which data of uniform quality were available on a few variables rel- evant to rather narrowly defined relationships. While this may improve the validity of the studies, it is difficult to generalize from their results. Krishnan's study (3) in India tested a model that paralleled closely the one employed here. Changes in crude death rates between 1951 and 1961 were investi- gated as a function ofper capita income, literacy rate, number of doctors per capita, number of hospital beds per capita, and per capita state expenditure on health services. Since the study was confined to 11 Indian states, however, its statistical meaning and general value have been questioned (2). Analysis of the relationship between education and health has been hampered in many countries by lack of meaningful data. Auster et al. (17) conducted one of the more definitive studies establishing such a relationship, but attention was limited to statewide indicators in the United States. Examinations of the relationship between fertility a GROSSE, R. N. & PERRY, B. H. Correlates of life expectancy in less developed countries. University ofMichigan, unpublished docu- ment, 1979. and health are likewise rare. One example is provided by Heller & Drake (18), who compared morbidity in children and their weight/height ratios with birth order and the number of young children in the family. While the analysis dealt with an important issue, its scope was extremely narrow. In general, review of the literature underscores a continuing dilemma. Large-scale, comprehensive studies have to deal with non-specific indicators of variable quality from heterogeneous populations. On the other hand, more circumscribed investigations are likely to produce results of limited applicability. THE STATISTICAL ANALYSIS Aims and objectives In order to study the development of health services, and the consequent effect on health status, in Member States, it was necessary to clarify basic relationships at national level between socioeconomic factors, health sector variables, and health status by quantitative as well as qualitative analysis. The basic pattern emerging from such a study could then be interpreted selectively according to the conditions in each particular Member State. The study also tried to analyse changes in these relationships with time. It was expected that careful analysis of the basic relationships and their dynamics would provide a framework for assessing the effects of specific inter- ventions. This analysis, some of the results of which are published here, is intended to establish points of reference on which to base assessment of future achievements. Selection of data base The relationships between the three basic components of the study are postulated in the following, very much simplified, way: socioeconomic factors health resources > health status This does not take into account the fact that these are all, in fact, two-way relationships. It is assumed that the countries with more favour- able socioeconomic conditions, including relatively high gross national product per capita, are likely to allocate more substantial material and personnel resources to the health sector, and that these resources will contribute to improvement in indices of health status, e.g., life expectancy. These indices are also influenced directly by socioeconomic factors. INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES 131 It was then necessary to choose the variables to represent the three components of the model, bearing in mind that values for each indicator and, where appropriate, changes with time, should be available for as many of the WHO Member States as possible. More than thirty potential measures were considered, but some of these were not available for a number of countries, while in other cases the data were inaccurate or inconsistent. For example, under- reporting of infant mortality was common. The following twelve variablesb were finally selected for analysis. Socioeconomic variables: GNP -per capita gross national product PE -percentage of children of primary school age in education TOFE -ratio of percentage of total children in primary education to percentage of female children in primary education PUTE -pupil/teacher ratio in primary education Health resource variables: HLTH -health employees per 1000 population MD -absolute number of physicians DR -physicians per 10 000 population NMW -nurses, midwives per physician BED -hospital beds per physician Health status variables: CBR -crude birth rate CHL -percentage of population aged under 14 years LE -life expectancy at birth Three of the four measures of socioeconomic status are concerned with educational opportunities. In addition to the overall measure of participation in education (PE), relative opportunity for females (TOFE) was noted. Pupil/teacher ratio (PUTE) was intended as a very crude measure of the quality of education. Four of the five health resource indicators employed were related to manpower, a matter of particular concern to this study. The number of health employees per 1000 population (HLTH) was used as an overall indicator of manpower availability. As an important category of worker, numbers of physicians were expressed both in absolute terms (MD) and relative to population size (DR). Nurses and midwives were related to the number of physicians (NMW), b Main sources of data: volumes of World health statistics annual, Geneva, World Health Organization, 1951-79; World atlas of the child, Washington, DC, The World Bank, 1979; World development report, 1979, Washington, DC, The World Bank, August 1979. rather than to the general population, for two reasons. First, countries with a relative abundance of one category of health worker tend to have above-average personnel/population ratios in other categories as well. To put all manpower indices on the same population base, therefore, was likely to introduce confusing correlations and redundancy into the analysis. The second reason for relating nurses and midwives to physicians was to obtain a measure of manpower balance. For similar reasons number of hospital beds (BED) was related to number of physicians rather than to total population. In effect, the BED variable was intended to provide a very crude indication of the health infrastructure in which personnel were employed, just as PUTE gave a semi- qualitative indication of the educational system. The last three variables in the list are indices of health status. The crude birth rate (CBR) and the relative age of the population (CHL) are frequently used as indicators of societal well-being, while life expectancy (LE) is a more direct measure of this condition. In view of the interest in relating the study findings to WHO programmes throughout the Organization's history, the two dates, 1950 and 1975, were chosen for analysis. In the case of PE, for example, the rate of change between the two years (DPE) was defined as the percentage difference between the current value (CPE) and the initial value (IPE), i.e., DPE= (CPE-IPE)x 100 IPE Other change variables were defined similarly. Where data were not available for the exact years, 1950 and 1975, measurements for the nearest available year were recorded. To illustrate the procedure, suppose that for country X, IPE = 46 and CPE = 77, then DPE = (77-46) x 100 = 67, 46 indicating that the percentage of children in primary education increased by 67% between 1950 and 1975. Periodic measurements were available for only six of the twelve basic variables (PE, DR, MD, BED, HL, LE), which each generated three secondary statistics (C, I, D), so that a total of 24 variables was available for analysis. Twenty-one of the 152 Member States ofWHO had five or more of these observations missing, and were excluded from the analysis so that the final data set consisted of 24 observations in 131 Member States, i.e., 3144 data points. 132 T. FJLOP & W. A. REINKE Of these, 38 observations were not available and had to be estimated. In all cases reasonable approxi- mations could be made, and the number of estimates was considered small enough (1.2%) not to distort the analysis. Analytical model The initial analysis consisted of two parts. At the first stage, socioeconomic factors were related to health resources as the dependent variables. The second-stage analysis related socioeconomic factors and health resources to health status variables. At each stage three sets of analyses were conducted. The first set dealt with initial conditions, the second with current conditions, and the third set related initial states and measures of change in the independent vari- ables to the current status of the dependent variables, thus testing for the presence of a time lag. It was hoped that a synthesis of the results obtained would be meaningful in itself and could then be related descrip- tively to HMD activities. Multiple linear regression analysis was conducted according to the foregoing format. Findings indicated that certain of the variables were generally of little or no consequence, and in particular, the only change variable that showed promise related to the number of hospital beds per physician. It also became apparent that some of the significant relationships could be further improved through logarithmic transformations (LN). The pruning and refinement of variables produced the modified format outlined in Table 1, which served as the basis for the final analysis reported here. REGRESSION ANALYSIS FINDINGS Resource predictions The extent to which socioeconomic (represented by exog(enous)) factors can explain variation in health resources is outlined in Table 2, using standardized beta-coefficients as indicators of relative import- ance.C Predictability is good with respect to health workers per 1000 population and physicians per 10 000 population, regardless of the time frame used. c The usual regression coefficients give an estimate of the amount of change in the dependent variable produced by unit change in each of the independent variables. However, the meaning of "unit of change" is not uniform, e.g., a one-dollar increase in per capita GNP has different connotations from a l1o increase in primary school enrolment. To facilitate comparability, beta-coefficients measure effects in standard deviation units. This conversion procedure also produces comparable differences between "observed" and "expected" values of the dependent variables. To further facilitate comparability, the original regression coefficients have been compared with their standard errors to produce t-statistics. Levels of significance of the t-values are given in Tables 2 and 3. Table 1. Indices used in final analysis Category Single Initial Change Currentmeasure value value Socioeconomic LN GNP IPE CPE PUTE Health resources LN HLTH LN IDR DBED LN CDR LN NMW LN IBED LN CBED Health status CBR ILE CLE Table 2. Summary of regression analysis findings for resource variables Beta coefficients Function LN GNP PUTE PE R2 LN HLTH = f(EXOGO) 0.50a -.017b 0.26a 0.69 LN HLTH = f(EXOGd) 0.58a -0.21b 0.17C 0.67 LN IDR = f(EXOGO) 0.40a -0.07 0.51a 0.79 LN CDR = f(EXOGO) 0.51a -0.15a 0.37a 0.84 LN CDR = f(EXOGd 0.59a -0.21a 0.27a 0.83 LN NMW = f(EXOGO) -0.00 0.10 -0.25 0.07 LN NMW = f(EXOGd) -0.09 0.13 -0.13 0.06 LN IBED = f(EXOGO) -0.02 0.29b -0.25C 0.21 LN CBED = f(EXOGo ) -0.20 0.27b -0.20 0.32 LN CBED = f(EXOGd. -0.26C 0.30b -0.13 0.31 a p<Q.001 b p< 0.01 c P<0.05 Coefficients of determination (R 2 ) indicate that 67-840/o of resource variation is explained, with per capita GNP the principal explanatory factor, and percentage of children in primary education also quite significant. Pupil/teacher ratios are generally less important, but they do provide some evidence that large numbers of teachers and large numbers of health workers tend to go together. Socioeconomic factors are less successful in predicting numbers of nurses and midwives and hospital beds per physician, producing R2 values between 0.06 and 0.32. The only relationship that is uniformly significant is that between the pupil/ teacher ratio and the bed/physician ratio. Health status predictions From Table 3 it can be seen that socioeconomic factors, notably per capita GNP, can explain more INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES Table 3. Summary of regression analysis findings for health status variables Beta coefficients Function LN GNP PUTE PE LN HLTH LN DR LN NMW LN BED DBED CBR R2 CBR = f(EXOGO) -0.388 0.18b -0.38a 0.67 CBR = f(EXOGd) -0.60a 0.20b -0.08 0.61 CBR = f(EXOGO, RESO) -0.13 0.16c -0.14 -0.10 -0.51a -0.12C -0.09 0.00 0.71 CBR = f(EXOGl, RES) -0.08 0.10 0.15C 0.04 -1.108 -0.08 -0.36a 0.71 ILE = f(EXOGO) 0.318a 0.20a 0.51a 0.82 CLE = f(EXOGO) 0.26a -0.198 0.58a 0.83 CLE= f(EX0Gd) 0.53a -0.25a 0.23a 0.71 CLE = f(EXOGO, RES0) 0.02 - 0.14a 0.318 -0.00 0.51a 0.02 -0.04 _0..1 1 b 0.88 CLE = f(EXOGt, RESt) 0.05 -0.10 0.00 -0.08 0.938 -0.01 0.11 0.81 CLE = f(EXOGO, CBR) 0.14c _0.13b 0.46a _0.31 a 0.86 CLE = f(EXOGt, CBR) 0.24a -0.15b 0.19a -0.48a 0.80 CLE = f(EXOGO, RESO, CBR) -0.01 -0.10C 0.28a -0.02 0.38a -0.01 -0.06 -0.11a -0.25a 0.90 CLE = f(EXOGt, RESt, CBR) 0.02 -0.06 0.06 -0.06 0.50b -0.04 -0.03 _0.39a 0.85 a p< 0.001 b p< 0.01 c P< 0.05 than 607o of the variation in crude birth rate. All of the regression coefficients are in the expected direction. Adding resource variables to the analysis increases the R2 value to 0.71, and, moreover, causes the physician and bed variables to replaceper capita GNP as the dominant factors. However, regression analysis is notoriously weak in distinguishing the effects of correlated independent variables, and the above results must be viewed with caution. Table 4. Final regression equations for analysis of interdependence of health resource variables: t-values are given in parentheses under partial regression coefficients 1. LN CDR = -3.165 + 0.571 LN GNP - 0.020 PUTE + 0.015 IPE (R2 = 0.84) (8.65) (3.52) (6.95) 2. LN CBED = 2.803 - 0.144 LN GNP + 0.024 PUTE - 0.006 IPE (R2 = 0.32) (1.60) (3.00) (1.85) 3. LN HLTH = -0.116 + 0.457 LN GNP - 0.019 PUTE + 0.009 IPE (R2 = 0.69) (5.99) (2.79) (3.50) 4. LN NMW = 1.059 - 0.002 LN GNP + 0.009 PUTE - 0.007 IPE (R2 = 0.07) (0.00) (0.91) (1.92) 5. CBR = 51.615 - 1.183 LN GNP + 0.178 PUTE - 0.050 IPE (R2 = 0.71) (1.26) (2.55) (1.56) - 4.366 LN IDR - 1.227 LN IBED + 0.001 DBED - 0.986 LN HLTH (4.02) (1.27) (0.06) (1.09) - 1.467 LN NMW (2.04) 6. CLE = 64.741 - 0.078 LN GNP - 0.116 PUTE + 0.097 IPE (R2 = 0.90) (0.13) (2.66) (4.99) + 3.275 LN IDR - 0.890 LN IBED - 0.027 DBED - 0.251 LN HLTH (4.66) (1.51) (3.44) (0.45) - 0.097 LN NMW - 0.252 CBR (0.22) (4.57) 133 T. FJLOP & W. A. REINKE As regards life expectancy, each of the socio- economic factors is highly significant, and together they explain 80% of its variation. Inclusion of the resource variables produces little increase in values of R 2, but once again the doctor/population ratio becomes the outstanding explanatory factor in place of the socioeconomic variables. Roughly similar results are obtained when the crude birth rate enters the equations in place of health resource variables. However, the crude birth rate does not appear to be as dominant a force as the doctor/population ratio. Inclusion of both health resources and crude birth rate in the analysis yields R 2 values as high as 0.90. The persistence of the pattern of shift in importance away from socioeconomic variables tends to increase its credibility. As noted earlier, the set of R2 values associated with a particular dependent variable tend to be similar. It is interesting to observe, however, that without exception the highest R2 in a set is that which relates the current status of the dependent variable to the initial state of the independent variables. Thus there is some evidence in support of a lagged response, and the investigations outlined below are limited to equations within each set that incorporate these lags. Synthesis offindings Of the 23 regression equations derived for the six dependent variables (Tables 2 and 3), attention ultimately focused on the six given in Table 4. Since these are prediction equations, the original (non- standardized) regression coefficients are listed, along with their t-values. Table 5 presents the simple correlation matrix in order to: (1) permit comparison of two-way associations with those derived after taking other factors into account; and (2) display areas of high correlation to emphasize the need for scepti- cism when considering the corresponding partial regression coefficients. Table 5. Simple correlation matrix for the final data set Socioeconomic factors Resource dependent variables PUTE IPE LN CDR LN CBED LN HLTH LN NMW Socioeconomic LN GNP -0.61 0.76 0.88 -0.52 0.80 -0.25 PUTE -0.48 -0.64 0.49 -0.60 0.22 IPE 0.82 -0.48 0.72 -0.30 Resource LN CDR -0.72 0.84 -0.41 LN CBED -0.44 0.54 LN HLTH -0.16 Resource independent variables Health status LN IDR LN IBED DBED LN HLTH LN NMW CBR CLE Socioeconomic LN GNP 0.83 -0.38 -0.31 0.80 -0.25 -0.77 0.81 PUTE -0.56 0.42 0.19 -0.60 0.22 0.59 -0.63 IPE 0.84 -0.40 -0.24 0.72 -0.30 -0.75 0.87 Resource LN IDR -0.58 -0.16 0.78 -0.43 -0.80 0.91 LN IBED . -0.14 -0.34 0.57 0.34 -0.50 DBED -0.21 0.04 0.22 -0.30 LN HLTH -0.16 -0.75 0.76 LN NMW 0.17 -0.35 Health status CBR -0.85 134 INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES To illustrate the comparison of two-way and multi- variate associations, consider the life expectancy dependent variable (CLE). IPE, LN IDR, andCBR all show high simple correlations with CLE, and the associations are retained in multiple regression equation 6. Independent variables LN HLTH and LN GNP also show a high simple association with CLE, but they lose importance in the multiple relationship. Note, however, that there is a high correlation among the independent variables themselves. More surpris- ingly, the simple correlation of DBED with CLE is slight, and it is not closely associated with any of the other independent variables, yet after adjustment for them, it is found, from equation 6, to be important. Taken together, the six regression equations of Table 4 produce the fairly straightforward set of relationships shown in Fig. 1. Socioeconomic factors show a strong association with two of the health man- power indicators, especially the number of physicians per 10 000 population. These socioeconomic factors are also directly associated with crude birth rate and life expectancy. When the health resource conditions (DR and BED) are combined with the socioeconomic factors, the correlation with crude birth rate and life expectancy increases. Similarly, crude birth rate combined with socioeconomic variables explains somewhat more of the variation in life expectancy than socioeconomic factors alone. Finally, inclusion of both crude birth rate and the resource variables DR and BED with socioeconomic variables gives the best correlation with life expectancy. We emphasize again, however, that the small improvement in R2 for life expectancy is not as striking as the shift in importance from socioeconomic factors to health resource vari- ables as these are introduced into the equations. There is some evidence, therefore, that socioeconomic con- ditions affect life expectancy through their influence on the health infrastructure, notably the number of physicians and hospital beds. -080 R2 0.60 < R2 < 0.80 ......- R2 0.60 Fig. 1. Diagrammatic representation of the relationships between socioeconomic factors, health resources, and health status. The ultimate outcome variable in the foregoing analysis was life expectancy, and it was associated with the largest value of R2 (0.90). It is of interest to examine more carefully the implications of selected regression coefficients, emphasizing again that they must be treated with scepticism. To make the investi- gation meaningful it is necessary to reconvert LN IDR to original rates, i.e., numbers of physicians per 10 000 population. The three independent variables of special interest are IPE, LN IDR, and CBR. Infor- mation concerning them is summarized in Table 6. The regression coefficient for IPE suggests that each 1 No increase in school enrolment could be expected to increase life expectancy by 0.0974 years. Thus, an increase of 10.3% in enrolment would be required to add a full year to life expectancy. For an "average" country, i.e., one with approximately 72% Table 6. Implications of regression coefficients in relation to life expectancy Change needed to increase life expectancy by 1 year Regression Variable Beta coefficient Mean Absolute value % of mean % children in primary 0.28 0.0974 72.1 10.3 14.3education (IPE) Physicians per 10 000 0.38 3.2747 1.64a 0.58a 35.4 population (IDR) Crude birth rate (CBR) -0.25 -0.2516 37.0 -4.0 -10.8 a Number converted from logarithm. 135 T. FOLOP & W. A. REINKE of eligible children in primary education, the required increase would be 14.3% beyond the present level. Corresponding calculations, with similar interpret- ations, are provided for the other variables of interest. Although the physician/population ratio was found to be statistically highly significant, the practical implication is that an average increase of 35% in the ratio is necessary to achieve a one-year increase in life expectancy. This is not likely to happen while other factors, notably GNP, remain constant. In practice, as well as statistically, it is difficult to separate socioeconomic factors from the impact of health resources development. By comparison, Table 6 suggests that a one-year improvement in life expectancy can be achieved through a reduction in crude birth rate by 4 per 1000, or about 11% of the average. ANALYSIS OF RESIDUALS It is well known that direct comparisons of health resource levels in different countries can be misleading because of local conditions. It would be preferable to make the comparisons after adjusting for these differ- ences. The regression results are encouraging in this regard since the six equations of Table 4 account for most of the national differences, even if little can, of course, be inferred concerning cause and effect. They provide individualized expectations of health re- sources and health status which can then be compared with actual conditions. Residual findings (actual minus expected) can perhaps indicate whether discrep- ancies are purely random or whether systematic patterns exist which might be related to special needs or interventions, e.g., WHO health manpower devel- opment programmes, not factored into the basic equations, because of inherent difficulties in quantifying them. Residuals by WHO Region Residuals for each of the six dependent variables were compiled for each country by WHO Region. The resulting average residuals are recorded in Table 7. The corresponding standard error of the calculated average was also determined in each case, and the t- Table 7. Average residuals for each WHO Region, measured as deviations from expected values in standardized form of regression equations Dependent variable No. of Region countries LN HLTH LN CDR LN NMW LN CBED CBR CLE Mean residuals Africa 41 0.06 -0.20 0.37 0.57 0.01 -0.09 Americas 26 -0.11 0.18 -0.56 -0.45 0.24 0.07 Eastern Mediterranean 17 -0.17 0.12 -0.41 -0.59 0.36 0.02 Europe 29 0.15 0.06 0.03 0.17 -0.39 0.01 South-East Asia 8 0.10 0.17 0.04 -0.75 -0.08 0.01 Western Pacific 10 -0.19 -0.16 0.50 -0.05 -0.08 0.14 Total 131 t-values Africa 0.69 -3.56a 2.68b 5.42a 0.07 -1.22 Americas -0.98 2.58b -3.20C -3.38c 2.71b 1.18 Eastern Mediterranean -1.27 1.36 -1.92 -3.63c 3.24c 0.20 Europe 1.43 0.88 0.20 1.34 -4.52a 0.13 South-East Asia 0.54 1.32 0.14 -3.16b -0.50 0.09 Western Pacific -1.08 -1.42 1.78 -0.26 -0.58 1.43 a p<0.001 b P< 0.05 C p< 0.01 136 INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES statistics evaluated are also shown in Table 7, along with their levels of significance. Residuals were deter- mined from the standardized form of the regression equations. Thus, for example, a residual of 0.5 is a measure of relative discrepancy that has essentially the same meaning when applied to any variable. The t-values, of course, are also subject to uniform interpretation. No significant region-wide departures from expec- tation were found with regard to total health man- power, although strong imbalances were noted within categories. The African Region as a whole has a severe shortage of physicians, even after adjustment for the unfavourable socioeconomic status of many coun- tries. However, there is a relative abundance of nurses and midwives and of hospital beds in comparison with physicians. The reverse pattern is found in the Region of the Americas and the Eastern Mediterranean Region, although the situation is not as extreme as in Africa. The only other health resource finding of significance relates to the South-East Asia Region, where the number of hospital beds does not match the number of doctors, which is higher than expected. Separate global analysis of residuals reveals that those for NMW and CBED are positively correlated and both are negatively associated with CDR residuals. It seems that the number of hospital beds is statistically more closely associated with the number of nurses than with the number of doctors. These findings, coupled with the non-significance of HLTH residuals in Table 7, suggests that the number of health workers per 1000 population tends to follow a standard pattern in relation to development, but that the composition of the manpower pool varies from country to country. Crude birth rate residuals show distinct regional patterns. Rates in the Region of the Americas and the Eastern Mediterranean Region were substantially higher than expected, whereas those in Europe were below expectations. Clearly, cultural factors are important in fertility, and these were not included in the regression equations. The equation involving life expectancy accounted for as much as 9007 of its variation, and there was no systematic pattern for the small residual. It is note- worthy, however, that residuals from earlier equa- tions involving only socioeconomic factors were positive for the Region of the Americas and the European Region, indicating an unusually high life expectancy, while those for the African and Eastern Mediterranean Regions were significantly negative. Only when health resources availability was taken into account were these systematic patterns lost. This finding provides further evidence of the differential effects of health resources and socioeconomic factors, in spite of their joint correlation. Individual residuals While examination of residuals by WHO Region has been rewarding, individual discrepancies can easily become buried in averages. Two approaches have been used to investigate the individual variations. First, the ten most positive and negative residuals for each variable have been highlighted in Table 8. Second, in view of our interest in health manpower programmes, all 131 countries have been listed in Table 9, according to magnitude of LN HLTH residual. Corresponding residuals for the CDR variable have been recorded in the adjacent column. In many respects the countries conform to the regional findings outlined above. In addition, several countries were found to be exceptional in other regards. The South-East Asia Region in particular is a region of wide individual differences which tend to cancel each other in overall averages. We were especially interested in trying to relate these exceptional cases to particular health manpower development initiatives. The number ofWHO fellow- ship awards to the country was taken as the index of HMD activity, and was adjusted to allow for popu- lation size and number of years of membership in WHO. The indices were then compared by means of t- tests for the exceptional cases in terms of total health manpower and physicians, listed in Table 8. While average results were in the expected direction, the variability in the indices was so large that they were not statistically significant. This suggests that this type of statistical analysis is of limited applicability in such cases, and more qualitative approaches, such as docu- ment analysis, expert opinion study, and detailed country case-studies are likely to be more useful. CONCLUSIONS An in-depth statistical analysis of interdependence of some indices characterizing national socioecon- omic conditions, material and personnel health resources, and health status was carried out, using data from 131 Members States of WHO. The aim of this study was to complement an analytical review of the World Health Organization's HMD programme for the years 1948-78. The analysis was limited both by the lack of reliable data and by the difficulties inherent in trying to quantify HMD programmes in general. A total of 24 variables were analysed for each of the 131 Member States. The statistical analysis of the 3144 data points provided some interesting results: (a) Between 67%o and 84% of the variation in the total number of health workers and in the number of 137 T. FOLOP & W. A. REINKE Table 8. Individual countries or areas with the most extreme residuals for each dependent variable Most negative Most positive LN HLTH Malaysia Burma Guyana Mongolia Panama Sri Lanka Australia Congo Nepal USSR Somalia Albania Switzerland Swaziland German Democratic Sweden Republic Central African Republic Gambia Bangladesh Pakistan LN CDR Thailand Mongolia Libyan Arab Pakistan Jamahiriya Malawi India Gabon Burma Malaysia Nicaragua Luxembourg Argentina Uganda USSR France Bulgaria Indonesia Bolivia Papua New Guinea Albania CBR Most negative Most positive .N NMW Spain United Republic of Cameroon Colombia Nigeria Algeria Sweden Nepal Finland Ecuador Indonesia Pakistan Gabon Mexico Trinidad & Tobago Belgium Lao People's Democratic Republic Iran Guyana Uruguay Central African Republic LN CBED Pakistan Zaire Mexico Gabon Paraguay Guyana Afghanistan Malawi Jordan Uganda Syrian Arab Republic United Republic of Cameroon Nicaragua Burundi Peru Lesotho Saudi Arabia Luxembourg Venezuela Zambia CLE Kuwait South Africa Libyan Arab Jamahiriya Iraq Nicaragua Honduras Dominican Republic Lebanon Mongolia Swaziland Gabon Bolivia South Africa Guinea-Bissau Congo United Republic of Cameroon Mauritania Gambia Argentina Senegal Nepal Chad Belgium Haiti Spain Bulgaria Mozambique Greece Afghanistan Poland Jamaica Albania Guyana Comoros El Salvador Sri Lanka Sudan Costa Rica Morocco Mexico 138 INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES Table 9. Country or area residuals arranged by WHO region and magnitude of LN HLTH residual LN LN LN LN Country or area HLTH CDR Country or area HLTH CDR Africa Central African Republic -0.78 -0.04 Cape Verde -0.62 -0.11 Burundi -0.55 -0.48 Niger -0.49 -0.42 Uganda -0.40 -0.63 Lesotho -0.39 -0.53 Guinea-Bissau -0.36 0.20 Rwanda -0.20 -0.44 Ghana -0.17 -0.51 Ethiopia -0.15 -0.58 Togo -0.15 -0.13 Nigeria -0.11 -0.30 Zaire -0.08 -0.41 Malawi -0.06 -0.91 Gabon -0.04 -0.77 Upper Volta -0.04 -0.24 United Republic of Cameroon 0.0 -0.40 Ivory Coast 0.02 -0.43 United Republic of Tanzania 0.08 0.07 Kenya 0.09 -0.34 Mali 0.09 -0.31 Chad 0.12 -0.06 Madagascar 0.13 0.38 Mozambique 0.15 0.26 Angola 0.17 0.05 Zambia 0.17 -0.06 Mauritania 0.18 -0.10 Mauritius 0.26 0.13 Benin 0.31 -0.21 Guinea 0.36 -0.23 Liberia 0.36 -0.04 Senegal 0.40 -0.06 Sierra Leone 0.45 -0.05 Botswana 0.51 0.07 South Africa 0.58 0.01 Comoros 0.72 0.27 Gambia 0.73 -0.06 Swaziland 0.85 -0.04 Congo 0.90 0.06 Americas Guyana -1.88 -0.44 Panama -1.51 0.11 Cuba -0.72 0.26 Brazil -0.41 0.03 Haiti -0.32 0.17 Nicaragua -0.27 0.70 Uruguay -0.22 0.21 Trinidad & Tobago -0.18 -0.41 Mexico -0.12 0.38 Bolivia -0.04 0.55 El Salvador 0.04 0.12 Venezuela 0.07 0.19 Argentina 0.08 0.63 Ecuador 0.09 0.34 Dominican Republic 0.10 0.44 Paraguay 0.10 0.51 Canada 0.11 -0.26 Chile 0.14 -0.11 Honduras 0.14 0.36 Colombia 0.16 0.41 Costa Rica 0.16 0.14 Peru 0.18 0.40 Guatemala 0.23 0.24 Jamaica 0.31 -0.03 United States 0.38 -0.32 Barbados 0.64 0.16 Eastern Mediterranean Somalia -1.08 0.34 Afghanistan -0.76 0.03 Lebanon -0.51 0.06 Jordan Saudi Arabia Kuwait Syrian Arab Republic Iraq Israel Democratic Yemen Libyan Arab Jamahiriya Egypt Tunisia Iran Cyprus Sudan Pakistan Europe Switzerland German Democratic Republic Algeria Belgium Italy Morocco Spain Federal Republic of Germany France Austria Netherlands Greece Turkey Norway Luxembourg Portugal Yugoslavia Denmark Poland Finland Ireland United Kingdom Bulgaria Hungary Czechoslovakia Romania Iceland Sweden Albania USSR South-East Asia Nepal Bangladesh Thailand Indonesia India Sri Lanka Mongolia Burma Western Pacific Malaysia Australia Korea Singapore Philippines Japan Fiji Papua New Guinea New Zealand Lao People's Democratic Republic -0.50 0.31 -0.50 0.04 -0.49 -0.91 -0.44 0.21 -0.33 -0.09 -0.29 0.43 -0.19 -0.55 0.02 0.13 0.05 0.34 0.06 0.02 0.25 0.15 0.42 0.27 0.67 0.19 0.73 1.05 -0.93 0.10 -0.92 -0.01 -0.70 -0.27 -0.68 0.06 -0.54 0.19 -0.43 -0.54 -0.33 0.23 -0.15 -0.31 -0.12 -0.61 -0.06 0.08 -0.05 -0.03 0.01 0.48 0.01 0.37 0.02 -0.32 0.08 -0.65 0.21 0.10 0.21 0.27 0.23 -0.30 0.32 0.17 0.50 -0.10 0.51 -0.06 0.52 0.05 0.63 0.56 0.65 0.38 0.68 0.41 0.70 0.30 0.72 -0.09 0.74 -0.18 0.85 0.52 0.87 0.62 -1.11 -0.32 -0.78 0.27 -0.65 -0.95 -0.43 -0.61 0.43 0.89 1.00 0.20 1.17 1.12 1.19 0.75 -1.95 -0.67 -1.13 -0.13 -0.46 0.02 -0.08 -0.32 -0.05 0.02 -0.03 -0.19 0.16 0.16 0.42 -0.59 0.56 -0.08 0.70 0.17 139 140 T. FOLOP & W. A. REINKE physicians is explained by socioeconomic factors, mainly theper capita GNP and, to a lesser extent, the percentage of children of primary school age in education. (b) Socioeconomic factors also explain more than 607o of the variation in crude birth rate and above 80% of life expectancy differences. (c) When resource variables are included in the analysis, they are found to be more important than socioeconomic factors in explaining variations in health status indices. (d) There seems to be a time lag in the response, i.e., the highest correlations are invariably observed between the current value of the dependent variable and the initial state of the independent variable (e.g., between present life expectancy and initial resource variables). (e) Analysis of regression coefficients shows that, statistically, a 35% increase in the physician/ popu- lation ratio would increase life expectancy by one year. (/) The study of residuals shows that the human health resource variables tend to follow a standard pattern in relation to socioeconomic development at the regional level, i.e., in the majority (78%) of cases, with the notable exception of Africa, they are at the expected level. However, the analysis of individual Member States reveals some deviations from this pattern. The analysis has proved that the statistical approach to the problem in question is not appropriate and can only be complementary to other qualitative methods. It does, however, reveal some interesting statistical relationships as described above and can detect the countries that show a marked deviation in their personnel health resources (and health status) indices from what one would expect on the basis of their socioeconomic situation. The explanation for these phenomena should be sought by other methods such as document analysis, expert opinion study, country case-studies, etc., and this is now being done in the analytical review of WHO's HMD programme. RESUME ANALYSE STATISTIQUE DE L'INTERDEPENDANCE DES VARIABLES RELATIVES AUX RESSOURCES DES PAYS EN MATIERE DE SANTE, ET PLUS PARTICULIEREMENT DE CELLES QUI TOUCHENT AU PERSONNEL L'analyse qui est rapportee ici est la derniere d'une serie de tentatives visant a clarifier l'importance relative du systeme de sante et des facteurs socio-economiques pour la sante des populations. L'etude a 6galement quantifiM les ecarts nationaux et r6gionaux par rapport au modele general, afin que les resultats servent de base pour une ttude selective des effets des interventions de programmation A l'kchelon national et par l'intermediaire de l'OMS. L'analyse ttait exceptionnellement complete puisqu'elle portait sur 131 Etats Membres de I'OMS. Comme dans un certain nombre d'autres etudes, on a observe que les facteurs socio-economiques jouent un grand role dans la variation nationale de l'esperance de vie. L'inclusion des variables relatives aux ressources en matitre de sante a fourni un effet sptcial retarde supplementaire, qui en definitive intervenait pour 90/. dans la variation d'esperance de vie. Ce qui est potentiellement plus important, on a obtenu la preuve du fait que les facteurs socio-economiques peuvent optrer dans une certaine mesure par leur effet de facilitation du deve- loppement des ressources de sante. II apparait donc que si les facteurs socio-6conomiques sont des conditions necessaires de I'amelioration de la sante, ils ne suffisent pas en l'absence du developpement correspondant d'une infrastructure viable des services de sante. Les ecarts r6siduels par rapport au modele general variaient systematiquement selon la R6gion de l'OMS dans 30% des cas. II faut noter tout particulierement que dans la Region africaine, le nombre de medecins est tres inferieur meme au niveau modeste auquel on pouvait s'attendre sur la base de la situation socio-economique dans cette Rtgion. Neanmoins, il y a une variation notable a l'interieur des pays consideres individuellement et il n'a pas ete possible de mettre en evidence une relation significative entre le pro- gramme OMS de developpement des personnels et les para- metres relatifs aux ressources de sante nationales. II en a W conclu que l'analyse statistique ne trouve qu'une application limitee dans ce domaine. REFERENCES 1. Department of International Health. Health and development: an annotated, indexed bibliography. Baltimore, Johns Hopkins University, 1972. 2. BARLOW, R. Research in human capital and develop- ment, 1: 45-75 (1979). 3. KRISHNAN, P. Social science and medicine, 9: 475-479 (1975). 4. MALENBAUM, W. In: Klarman, H. E., ed., Empirical studies in health economics, Baltimore, The Johns Hopkins University Press, 1970, pp. 31-54. INTERDEPENDENCE OF HEALTH RESOURCE VARIABLES 141 5. CORREA, H. Population, health, nutrition, anddevelop- ment, Lexington, Massachusetts, D. C. Heath, 1975. 6. GALENSON, W. & PYATr, G. The quality of labor and economic development in certain countries. Geneva, International Labour Office, 1964. 7. PRESTON, S. H. Mortality patterns in national popu- lations. New York, Academic Press, 1976. 8. ARRIAGA, E. E. & DAVIS, K. Demography, 6: 223-242 (1969). 9. STOLNITZ, G. J. The population debate: dimensions and perspectives. In: Papers of the World Population Conference, Bucharest, 1974, Population studies No. 57, Vol. 1, 1975. 10. FUCHS, V. R. Milbank memorial fund quarterly, 7: 153-182 (1979). 11. FOLOP, T. Egdszegijgyi szervezestan. Medicina, Buda- pest, 1978. 12. kCEPIN, 0. P. Problemi zdravoohranenija razvivasu- scihsja stran. Moscow, Medicina, 1976. 13. SCEPIN, 0. P. Aktualnyj problemi zarubetnovo zdravoohranenija. Moscow, Tsoliuv, 1978. 14. SMUCKER, C. M. Socio-economic and demographic correlates of infant and child mortality in India. Unpub- lished doctoral dissertation, University of Michigan, 1975. 15. VENEDIKTOV, D. D. Metdunarodnyji problemi zdravoohranenija. Moscow, Medicina, 1976. 16. GILLIAND, P. & GALLAND, R. World health statistics report, 30: 227-238 (1977). 17. AUSTER, R. ET AL. Journal of human resources, 4: 411-436 (1969). 18. HELLER, P. S. & DRAKE, W. D. Malnutrition, child morbidity, and thefamily decision process. Center for Research on Economic Development, University of Michigan, Ann Arbor, 1976 (Discussion paper No. 58).