Analyzing Colombian wage structure

ANALYZING COLOMBIAN WAGE STRUCTURE by Gary S. Fields Series: Studies in Employment and Rural Development No. 46 Division: Employment and Rural Development Department: Development Economics Development Policy Staff International Bank for Reconstruction and Development This paper was prepared as a contribution to Research Project 67148 on Urban Labor Markets in Latin America. The views expressed represent those of the author and not necessarily those of the Bank. This report may not be published nor may it be quoted as representing the views of the Bank and its affiliated organizations. Washington, D.C., May 1978 "Analyzing Colombian Wage Structure": Summary This paper examines the determinants of wage structure in Colombia. The work is carried out at two levels: household and interindustry. Several data sources are used, all pertaining to the period from 1967 to 1973. Twelve main hypotheses and numerous subsidiary relationships are investigated empirically. Summary Table 1 sketches the data. Sum- mary Table 2 outlines the main results. *Partial support for this research was received from the World Bank and from the Economic Growth Center, Yale University. Financial assistance from these institutions and research assistance of Judith Oder are grate- fully acknowledged. Helpful comments on the review draft were received from Albert Berry, Richard Webb, Rakesh Mohan, Mark Leiserson, Alan Gutman, and Eric Hanushek. SUMMARY TABLE 1. EMPIRICAL RESULTS PRESENTED IN THIS PAPER AND DATA SETS USED. Survey of Urban Family Census of Survey of Occu- Population Cen- Census of Income and Expenditure Manufacturing pational & Ceo- sus of 1973, 4% Manufacuring, (PRESFAM) conducted by conducted by graphic Mobili- Public use sample, conducted by CEDE in 4 major cities, DANE in larger ty, Conducted by nationwide DANE in larger 1967-68 firms, 1967 CEDE in 4 urban firms, 1968 areas, 1967 (n - 877 (n - 20 two- (n 860,000 indi- (n - 88 three- manufacturing (n - 2,949 digit manufac- viduals, various digit manufac- Topic workers) families) turing industries) (n - 331 workers) subsamples used) turing industries) Individual and industry characteristics as deter- minant of incomes in X X urban Colombia Individual and family background characteristics as determinants of income X in urban Colombia Personal and regional effects on income structure in Colombia A. Urban study X B. National study X C. Comparison of urban and rural X areas Determinants of inter- sectoral wage structure X X SUMMARY TABLE 2 ANALYZING COLOMBIAN WAGE STPUCTURE: SUMMARY OF HYP9THESIS TESTS. Evidence at the Evidence at the Mousehold Level Interindustrv Level Hypothesized Direction of Simple Multivariate Simple Multivariate Variable Effect Tabulations Analysis Tabulations Analysis Education + + + + + ** Experience + + + + + Sex + for males + + + + Characteris- tics of workers in sectors + + + + Not identifiable qocio-economic background + + + Data not available Value added ** per worker + + + + + Capital per ** worker + + + + ** Firm size + for large firms + + + + Concentration + + Data not + + available Foreign invest- ment + + + + + Region Variable Variable Variable Data not available Urban/rural Differ Differ Differ Data not available + In multivariate analysis, effect is stativtically significant and quantita- tively large. + In multivariate analysis, effect is statistically significant and quantita- tively small. + In multivariate analysis, effect is not statistically significint. I. Introduction With the rising concern for income distribution in less developed countries (LDC's), it is incumbent upon students of economic development to gain a clear understanding of the determinants of incomes at the individual and household levels. Research into the causes of inequality in LDC's has demonstrated in the countries studied that variations in labor income account for a larger fraction of total income inequality than all other income sources combined.1 Partly, this is because labor's functional share is higher than any other, partly because most individuals or families receive most or all of their income from the work they do. For a further understanding of inequality and poverty, therefore, we must understand why some persons receive higher wages (or self-employment income) from the labor market than do others. Economic theory does not yet offer a comprehensive explanation for wage diversity. There are, however, partial explanations based on con- siderations of labor demand,2 labor supply, technological variability,4 and institutional influences.5 . While some of these various strands 1These findings have been reported for Taiwan (Fei-Kuo-Ranis, 1978), Pakistan (Ayub, 1977), and Colombia (Fields, 1977a).This liter- ature is summarized in Fields (1977b). 2 See Welch (1970) and Johnson (1970). 3There is an enormous literature in the human capital tradition re- lating market wages to education, training, parental time inputs, and other forms of human investment. Mincer (1974) covers the theoretical aspects of these relationships, Harbison (1973) and Blaug (1973) discuss the issues ald present empirical eviaence related to LDU's, and Blaug (1976) presents an extensive annotated bibliography of the relevant literature. 4This problem has been analyzed with considerable skill by Stiglitz; see, for example, Stiglitz (1974, 1976). 5See, for example, Berg (1969) and Reynolds (1969). Also, the pre- dictions of market theories and institutional theories are compared and contrasted by Heady (1976). of analysis have been synthesized,1 the construction of a unified theory of the determinants of wages- and size distribution of income has so far eluded economists. The formulation of such a theory merits high priority. In the meantime, empirical research must be carried out without the aid of a single formal theory. In what follows, I adopt an eclectic approach to the wage determination process, exploring both economic and econometric issues related to wage determination. The study will be conducted both at the microeconomic level to explain differences in earned income among individual workers and at the sectoral level to explain differences in average wages across industries. Policy concerns motivate this research. To design policies for reducing poverty, we must understand what causes people to have low incomes. Different policies are in*order if the cause of low incomes is lack of skills rather than lack of demand for skilled people. Know- ledge of the determinants of wage structures help suggest which areas of change are likely to be most fruitful in raising the incomes of the poor. The empirical analysis covers wage structure in Colombia, for which particularly rich and comprehensive data are available. Although the analysis is limited to one particular LDC, and to one point in time be- cause of data limitations, it is hoped that the general methodological procedure will be indicative of the manner in which this problem might be approached in othor countries. One limitation of the Colombian data should be mentioned before proceeding. The various surveys and censuses generally did not distin- guish total income from labor earnings or earnings per week from the An interesting example is Rosenzweig's (1977) integration of farm house- holds' labor supply decisions with off-farm labor demand conditions. -3- hourly wage. This paper is about the structure of rewards received in the labor market, which for shorthand I refer to as "wage structure." All the hypotheses are framed in terms of wage structure and this termin- ology is sometimes used in the empirical sections even when the research is based on income or earnings. In Colombia, as in other LDC's, wage determination has been studied in either of two ways. Some authors give principal attention to the char- acteristics of industries and firms. In this type of study, factors such as the value added per worker, largeness and foreign ownership of the firm, and presence or absence of a labor union are thought to be the principal determinants of wages.1 Other studies, more in the human capital tradition, give greater weight to the characteristics of workers. This type of research looks to an individual's education, labor market experi- ence, and similar personal characteristics in attempting to explain his or her income.2 An important feature of these two types of studies is that they have been made to integrate these two types of studies in the 3 4 United States. and in a few less developed countries. The present paper seeks to provide an integrated analysis of the determinants of wage struc- ture in Colombia, paying particular attention to methodological issues, which other research studies have sometimes overlooked. Previous studies of Colombia have demonstrated that a number of characteristics of individuals and firms are associated with wages. 1See, for instance, Nelson et al. (1971), Sanjin6s (1975), Heady (1976), Berry and Urrutia (1976), and Fields and Marulanda (1976). This literature is reviewed in Section IV.A. 2Among the studies from this perspective are Schultz (1968), Selowsky (1969), Gonz9lez (1971), Musgrove (1974), Urrutia (1974), Kugler (1975), Fields (1976), and Fields and Schultz (1977). For further review of the earnings function literature, see Section III.A. 3See Weiss (1966), Stafford (1968), Wachtel and Betsey (1972), and Kalachek and Raines (1976). 4Especially noteworthy are the unpublished studies done at the World Bank by Sabot (Tanzania), Mazumdar (Malaysia), Webb (Peru), and Altimir and Piflera (Chile and Peru). -4- The variables considered in some of the available studies for Colombia are catalogued in Table 1. These and other variables are considered in greater detail below. TABLE 1. SCOPE OF STUDIES OF COLOMBIAN WAGE STRUCTURE Found to be Associated with Explanatory Factor Wages in Studies By: Education Schultz (1968), Selowsky (1969), Gonzalez (1971), Prieto (1971), Musgrove (1974), Urrutia (1974), Kugler (1975), Fields Age or Experience (1976), Fields and Schultz (1977), Fields, (1977a). Sex Schultz (1968), Berry and Urrutia (1976), Fields (1976), Fields and Schultz (1977) Socio-economic Background Parra (1973), Urrutia (1974), Kugler (1975), Fields (1976) Value Added per Worker Nelsonet al. (1971), Urrutia (1968), Sanjines (1975), Heady (1976), Fields Capital per Worker and Marulanda (1976 ) Size of Firm Nelson et al. (1971), Fields and Marulanda (1976) Foreign Ownership Dfaz-Alejandro (1974), Fields and Marulanda (1976) Product Market Variables Heady (1976) Region or Location Prieto (1971), Musgrove (1974), Berry and Urrutia (1976), Fields and Schultz (1977), Fields (1977a) Before accepting the simple tabular evidence, two caveats must be borne in mind. One is that most of these studies , especially those con- ducted before 1976, are univariate rather than multivariate. There is therefore the possibility that a factor which appears to be correlated with wages may make no independent contribution beyond that provided by some other variable. For example, capital intensity of a firm may contribute no additional explanatory power beyond what is learned by considering value added per worker, or vice versa. The other difficulty arises even in multivariate studies. Suppose the true process of wage determination is of the form (1) W - a + a1 2 2 +3 3 + e, where W is the wage or some transform thereof, X1 is a vector of character- istics of the worker or a group of workers, X2 a vector of characteris- tics of the firm or industry in which they are employed, X a vector of characteristics of other workers employed in the same firm or industry, and e an error term, assumed to be uncorrelated with X1 X2 and X If we then run the simpler regressions (2) W= a' + X+ , (3) W= a" + " X2 +e and (4) W =al"' + ' I3 + C the models are misspecified and the estimated regression coefficients a1' 82, and 83 will be biased unless X1, 2, and X3 are orthogonal, i.e., statis- tically independent of one another.1 However, the likelihood is that X1. X2,'and X3 are positively correlated, which means that the true effects of some of the elements of X1' X2, and X3 will be smaller than the effects estimated from (2), (3) or (4), and may well fail conventional tests of statistical significance. Specifically, the characteristics of the firm in which the individual is employed may make little additional contribution to the explanation of wages once the worker's own characteristics and the characteristics of his co-workers are taken into account, or vice versa. The major empirical tasks before us, then, are to explain why each See Goldberger (1964, pp. 194-196). -6- individual or industry characteristic would be expected to have an inde- 22ndent association with wage determination (i.e., why 8V 82 and 83 might all be expected to be significantly different from zero) and to gauge the relative importance of these two sets of characteristics in accounting for wage differentials in the Colombian labor market. Both the correlates of individuals' incomes and the correlates of average sectoral wages are considered. -7- II . notheses about the Determinants of Wage Differentials This section presents a number of hypotheses about the correlates of wage differentials in Colombia. The hypothesized relationships are best thought of in a multiple correlation sense, i.e., how much of a association there is between income (or wage) and the several measured characteristics of individuals and their jobs. Because a multiple regression framework is used to measure these correlations, the hypotheses and empirical results are expressed in terms of dependent and independent variables. However, unidirectional causality is not necessarily implied, especially in the intersectoral analysis, nor should it be, since many of the variables are jointly determined. The reader is ad- vised to interpret the results accordingly.1 Besides the numerous hypotheses about the determinants of wage differ- entials which are tested in this paper, the effects of many other possible influences on wages remain untested, mainly for lack of data. These fall into two groups: real differences and compensating differentials. Unioni- zation is perhaps the most important real effect that cannot be tested. Among the compensating differentials which cannot be considered are higher pay for unpleasant or unsafe work, compensation for intermittent employ- ment, and higher nominal incomes in higher price areas. Hypothesis 1. Workers with more education, and sectors with larger proportions of these workers, have higher wages ceteris paribus. Hypothesis 2. Workers with more experience, and sectors with larger proportions of these workers, have higher wages ceteris paribus. The evidence from Colombia summarized in Table 1 establishes that IIn the present state of our knowledge, the structural model determin- ing wages, employment, and other sector-level variables cannot be written down formally nor could such a model be estimated statistically, both be- cause of lack of identification (in the econometric sense) and because of lack of.requisite data. -8- incomes and earnings are positively-related to an individual's education, labor market experience, and other characteristics. This conforms with studies from all over the world.1 The principal notion underlying the majority of these studies is that of investment in human capital. That is, those persons with higher levels of education and more years of experience are presumed to embody more productive capacity, and for this reason, they receive higher wages.2 Notwithstanding the prominence . of the human capital model in the literature, there are other reasons for expecting that persons with more education or experience might receive higher wages.3 One possibility is that society might place higher value on highly-educated or highly-experienced workers independently of their productivity. More likely is the phenomenon of "screening," where employers prefer to hire workers with more education or 4 experience, because they believe that these workers will be more productive. Education or experience are thus used as devices for selecting potentially- better workers from a large pool of applicants. This model applies especially. where wages are set above market-clearing levels for institutional reasons. If an employer is obligated to pay higher-than-market-clearing wages, any differential in productivity between workers with more education and those with less would lead employers to prefer the better-educated, even if the productivity differential were due to inherent differences among individuals and not to the process of schooling itself. 1For a summary of this research, see Psacharopoulos (1973). 2For forceful statements of the human capital view, see Becker (1964) and Mincer (1974). For an analysis of the role of experience and on-the-job training, see Rosen (1972). A thoughtful essay on the puzzling economic value of education is that of Blaug (1972). For the same ideas in a broader context, see Blaug (1973). 4- See Fields (1972), Spence (1973), Stiglitz (1975), Thurow (1975) and the citations therein. -9- Hypothesis 3. Male workers, and sectors with larger liroportions of men, have higher wages ceteris paribus. An undisputed fact is that Colombian men earn more than women. Un- doubtedly, some fraction of this differential is due to discrimination, which may take the form of unequal pay for equal work or the systematic confinement of women to certain low-paying occupations and economic sectors. Some other part of the income differential is due to cultural differences in Colombian society. Colombian women differ from men, having less educa- tion, job experience, and permanence within the labor force, and greater child care and other home responsibilities. In the microeconomic analysis, either cultural differences or labor market discrimination would be expected to result in an earnings function for women which lies below that for men (although they do not necessarily have the same shape). In the sectoral analysis, a relation between sex composition of the labor force and average wage would be encountered for either of two reasons: (i) women get paid less than men in the same in- dustry, or (ii) if s.x is used as a means of allocating men and women to different industries, and if women are crowded into the lower paying in- dustries, women's wages would be further depressed in the crowded industries. Hypothesis 4. Otherwise identical workers receive higher wages if they work near other high wage workers. There are reasons to expect that individual workers of any given sex with any given level of education and experience may receive higher wages if they work alongside other workers who themselves are more educated or experienced or in industries which are predominantly male. In the context of a neoclassical production function, the higher wage would reflect the -10- use of more complementary factors of production and hence rewards to greater factor productivity. A more institutional approach would regard the higher wage as being the benefit to the individual worker of member- ship in a high wage "job cluster."1 In an interindustry regression, it is impossible to distinguish whether high wage industries pay higher wages to all workers or just to those with more education and experience. However, in a microeconomic regression, these two motivations are readily identified, since the individual's own characteristics and the average characteristics of the sector's labor force are entered as separate sets of variables. Hypothesis 5. Workers from higher socio-economic backgrounds have higher wages ceteris paribus. It is frequently argued that the socio-economic status of an individual tends to be correlated with that of his parents. Some theorists see this intergenerational association as natural and expected, reflecting higher income parents' willingness and/or ability to buy more education and in other ways invest more in so-called "child quality ".2 Others see this association in more sinister terms: as a consequence of a stratified society 1Institutional arguments of this sort are presented in detail in Doeringer and Piore (1971). The term "job cluster" is from Dunlop (1957). 2A particularly clear statement of this view is found in Becker (1967). The "child quality" terminology is borrowed from the "new home economics," one good summary of which is found in Nerlove (1974). *-11- in which those at the top rig the system to perpetuate their favored status within and across generations. Interpretations may differ, but intergenerational transmission of economic position is an unmistakable fact in Colombia as elsewhere. (Remember, though, that the hypothesis is posed ceteris paribus, so it must be tested in a multivariate context.) Hypothesis 6. Sectors with higher value added per worker and the workers in those sectors have higher wages ceteris paribus. Hypothesis 7. Sectors with more capital per worker and the workers in those sectors have higher wages ceteris paribus. Elementary economic theory would predict that when equilibrating forces in labor markets are freely-functioning the forces of competition would produce wage equality for comparable workers. Since all firms would pay the same wage but differ with respect to 'productivity' and 'capital intensity,' these latter variables would have no significant relation with (i.e., be statisti- cally independent of) wage rates. However, when there are multiple technolo- gies, value-added per worker (so called 'productivity') and capital per worker (so-called "capital-intensity') would differ across firms.2 LFor examples of the less sanguine view, see Bowles and Gintis -(t197'5) and Carnoy. (1974). 2Firms that use relatively more capital relative to labor to produce any given amount of product would have higher 'productivity' and higher ?capital-intensity.' -12- Since we do observe wage differentials and these differentials are associated with value added and capital per worker in the multivariate analy- sis, the simple textbook theory cannot suffice. One reason may be that the equilibrating forces are not free to operate. A considerable amount of labor mobility might be impeded by restrictions on entry, lack of information, or costs of movement, for example. Wages may be prevented from falling due to institutional rigidities caused by labor unions, minimum wage legislation, government wage policy, and the like. The higher wage would induce firms to move up their labor demand curves and employ fewer workers. There would then arise a correlation between the wage in an economic sector and the value added per employed worker, which is what we are measuring by 'productivity'. Another possibility is that the equilibrating forces do operate but the association between wages, productivity, and capital-intensity reflects other economic motivations of firms not captured in standard textbook-level theory. Firms may benefit by paying higher wages, for instance, by reducing labor 1 turnover costs or by raising worker efficiency. Those industries with highly capital-intensive,interdependent technologies might benefit the most from high wage policies. 2 1For analyses of models of this sort, see Stiglitz (1974, 1976). 2To illustrate the point, in the automobile assembly lines, any damage to the machinery or underutilization of it due to absenteeism becomes extremely costly. To avoid these unfortunate events, automobile firms might raise their wages to assure themselves of a sufficient number of reliable workers. In this case, greater value added and capital-intensity provide the economic rationale for higher wages. -13- A related argument has to do with labor unions. While very little is known about the Colombian labor movement, it is clear that unions in some firms or industries are more powerful than in others. Some estimates place the wage effect of unions at approximately ten to twenty percent.1 Unions in the United States possess greater negotiating power in highly- profitable industries, apparently because the cost of a strike is higher when more profits are foregone.2 Insofar as profits are related to value added and capital-intensity (a not unreasonable assumption in Colombia) , labor unions may be providing an additional impetus for higher wages to be paid in the high 'productivity', highly capital-intensive sectors. Monopsony may also be a part of the explanation. Consider two industries which have identical labor supply functions up to some point. After that point, assume industry A faces a steeper function than indus- try B. A would employ fewer workers and pay higher wages than B. On the assumption that the value of the marginal product of labor and value added per worker are positively correlated, a positive relation between wage and value added per worker would arise due to monopsony effects. Note that all of the above arguments pertain to wage differentials among otherwise identical workers. The argument is that 'produrtivity' and 'capital intensity' are significantly related to wages even after standardizing for worker quality. In a microeconomic earnings function, worker quality is- controlled for by variables like years of education and experience. If statistically significant effects of industry variables are found in the multivariate analysis, these may legitimately be inter- preted as confirming the microeconomic versions of Hypotheses 6 and 7. 1The standard reference on labor unions in Colombia is Urrutia (1969). A more recent study with better quantitative information is that of Tenjo (1975). 2The theoretical arguments and empirical evidence are presented in Perry (1966). -14- In the study of intersectoral wage structure, heterogeneity of the labor force poses a more formidable problem of interpretation. The diffi- culty is that the variables are denominated in non-standard units of 'labor' and thus are not very well-measured. There then arises the possi- bility of a spurious association between average wages in an economic sector and value added and capital intensity per worker. The problem may be illustrated with reference to labor unions. Suppose that workers differ in skill and are paid on a piece work basis. When a union raises wages, the higher wages serve to attract a larger pool of better-qualified work- ers, and the employer can then choose the best workers from among them. Though the firm may pay higher wages per hour, it also receives more units of output per hour. Hence, higher 'productivity' in one sector as compared with another may be the consequence of higher wages achieved by unions and not the cause of the higher wages.1 Therefore, while an association be- tween wages, value added per worker, and capital per worker would be con- sistent with the view that workers in high productivity sectors are reward- ed by higher wages, perhaps with their unions inducing the firms to share a part of their profits, it would also be consistent with the view that higher wages alter the skill mix but leave labor's share relatively un- changed. Simultaneity is clearly a problem in interpreting these asso- ciations. The absence of data for Colombia on union membership for 1967 do not permit us to test among the alternatives mentioned. However, it is interesting to note that in the United States, where this type of information is available, Weiss (1966) and Ashenfelter and Johnson (1972) observed a 20% wage differen- tial between union and non-union workers. They also found that unionizdd firms attract workers with more education and more experience. After adjusting the wage differential for these differences, Weiss found that unionized workers received wages similar to those received by comparable workers elsewhere and Ashenfelter and Johnson found that the union effect was not significantly different from zero. Thus, it may be concluded that one important effect of unions is to reallocate more productive workers to firms or industries which are forced by union pressure to pay higher wages. -15- In summary, we have considered five reasons why firms with higher 'productivity' or greater capital-intensity might pay higher wages: lack of equilibration in labor markets, firms' responses to a more complex set of economic forces than are usually considered, the impact of labor unions, monopsony elements in the labor market, and the lack of standardi- 1 zation for labor quality, particularly in capital-intensive processes. Care should be exercised when reading the results below not to impose a more structural interpretation than the model warrants. Hypothesis 8. Sectors with proportionally more large firms and the workers in those sectors have higher wages ceteris paribus. The evidence that large firms in Colombia pay higher wages (e.g., Nelson, Schultz, and Slighton (1971, Chapter 5)) is mutatis mutandis, not ceteris paribus. It may be hypothesized, however, that the effect of the size of firm variable remains even after controlling for the influence of other variables, in particular 'productivity' and capital-intensity in the industry and the workers' personal characteristics. There are four , reasons for this hypothesis. The first reason is the simple technological point that there tends to be greater interdependence in production among workers in large firms compared to small firms. With this greater interdependence comes the need for a more reliable work force, which is obtained through higher wages. The argument here is closely related to that made earlier concerning the hypothesized relationship between capital-intensity and wages. 1One nagging doubt which cannot be rejected out of hand is the possi- bility that the association between wage and value added per worker arises from the accounting identity that higher wage results in higher value added dollar for dollar. This question can be resolved only by simultan- eous equations estimation on the full (and as yet unknown) structural system. 2For example, a large automobile assembly line with a given amount of capital per worker has a more interdependent technology than a small garment factory with the same capital per worker in which each sewer has his own machine. On the other hand, large firms may also have more routinized or more specialized production processes which imply lower skill needs. -16- Second, large firms may be less desirable places to work (sociolo- gists call this 'anomie') so large firms might have to pay a wage premium to attract labor. Third, there is a connection between firm size and labor union activity. Colombian labor law prohibits the formation of unions in firms with fewer than 25 workers. Therefore, the more large firms there are in an economic sector, the more likely there are to be unions able to exert power to secure higher wages for their members. Data limitations prohibit the direct testing of unions' influence in large firms. Finally, there are other variables affecting wages for which size of firm is probably a proxy. Consider, for example, the effect of monopoly power in the product market.1 In general, we would expect that firms which operate within monopolistic markets would be earning greater pro- fits. Not only would these firms have greater ability to pay higher wages, but because of their greater profits, they would have more incen- 2 tive to grant wage increases in order to avoid a strike. The monopolistic sectors would tend to be composed of relatively more large firms. Hence, we would find that economic sectors with more large firms would pay high- er wages. The monopoly power argument may be tested directly. This leads to: Hypothesis 9. Sectors which are relatively more concentrated and the workers in those sectors have higher wages ceteris paribus. Some empirical evidence on this question is given in a study by Misas (1976) for Colombia.. Theoretical arguments and empirical evidence for the 1For an analysis of the role of product market considerations on wages in less developed countries with particular reference to Colombia, see Heady (1976). 2But on the other hand, their greater profit position also increases their ability to resist striking workers' demands. -17- United States are given in Levinson (1967). Essentially, the rationale is that more concentrated industries tend to have more profitable firms which are better able to pay higher wages, and unions take advantage of this. Hypothesis 10. Sectors which have proportionately more foreign invest- ment or foreign capital and the workers in those sectors have higher wages ceteris paribus. There is no basis in supply and demand analysis for expecting that higher wages would be paid in sectors with large concentrations of foreign investment or capital apart from the possibility that these sectors may be more capital-intensive, more concentrated or have more large firms, which would presumably be reflected in the tests of other hypotheses. However, various politi- cal or institutional explanations can be offered for an independent effect of foreignness. For example, one such reason is the desire of multinational firms in these sectors to maintain good public relations. Another reason is to avoid large wage differentials between foreign executives and nationals of the host country, and in turn, between nationals in executive positions and other personnel, also nationals. Also, it should be pointed out that foreign firms appear to be more profitable on average than locals, and therefore possess greater capacity to pay higher wages in response to pressures from unions or other sources. For all these reasons, we might expect to find foreign firms paying higher wages. Once again, the hypothesis is multivariate, in that the extent of foreign investment or capital is hypothesized to contribute additional independent explanatory power even in the presence of productivity, capital-intensity, and size. -18- - Hypothesis 11. Certain regions cf the country and the workers in those regions have higher wages ceteris paribus. Colombia is a highly-regionalized country. Many public policies seek to channel resources toward locations where the poor are concentrated. Other policies seek to assist the poor in resettling into areas where jobs are better-paying and more plentiful. For policy as well as academic reasons, then, it is important to understand the determinants of regional inequality.1 Research has shown large income differentials---as great as three or four to one---between one part of Colombia and another. Among the many reasons offered for the persistence of these differentials, two stand out. One is euphemistically referred to as "imperfect labor mobil- ity," i.e., the inability or unwillingness of individuals in poor areas to move to higher income places. The second argument is lack of standardi- zation. It is known that different areas' labor forces differ in important respects (e.g., education, industry). It would be expected, therefore, that wages in some areas would be higher than in others because of these compositional effects. Hypothesis 11 is posed ceteris paribus, i.e., that wage differentials acros0 regions are found even within educational or industrial groups. Evi- dence consistent with this hypothesis would tend to refute the lack of standardization argument and would suggest the existence of geographic labor market segmentation. Hypothesis 12. The structures of wages and their correlates differ as between urban and rural areas. The large literature on dualistic economic development2 suggests that ICf. Kuznets (1963) and Williamson (1965). 2Examples are Lewis (1955) and Fei and Ranis (1964). -19- less developed countries' economies may usefully be considered in two segments: an advanced, relatively high-income "modern sector," and a backward, relative- ly low-income "traditional sector." In some models, such as the well-known one by Harris and Todaro (1970), the "modern" and "traditional" sectors are associated with urban and rural areas respectively. The Harris-Todaro model is not alone in postulating different wage setting mechanisms in the two sectors. Urban wages are assumed to be set institutionally at artifically high levels. Rural wages, in contrast, are thought to be set by supply and demand in the-rural labor market. The urban-rural wage gap leads to migration and urban unemployment in equilibrium. As originally formulated, the Harris-Todaro model assumed homogeneous labor. However, at least one extension of the model---by Fields (1975)--- introduces labor force heterogeneity according to educational level. In the Fields model, education may pay off for the individual in modern but not in traditional sectors. The empirical validity of this assumption is open to testing. Another possibility recently modeled by Rosenzweig (1977),and demon- strated for rural India is that competititive neoclassical forces may work to determine wages and the supply and demand for labor within well- defined geographic areas but not across them due to labor force immobili- ty. That is to say, apparently comparable workers may receive very differ- ent wages depending on where they happen to live. These models, and others vhich also allow for real-world market imperfections, suggest that the structures of wages and their correlates -20- may differ as between urban and rural areas. In particular, the models of the preceding paragraphs lead to the prediction that education may be a relatively more important determinant of urban than of rural wages, while geographic factors may be relatively more important in rural than in urban areas. Empirical evidence on these questions is presented below. -21- III. Evidence on the Determinants of Individual and Family Incomes A. Literature Review By now, a considerable number of studies have used microeconomic survey data to examine the determinants of individual and family incomes in Colombia. The major microeconomic data sets are: (i) The employment and household budget surveys conducted in the mid to late 1960s by the Center for the Study of Economic Development (CEDE), Universidad de Los Andes; (ii) The household surveys conducted by the national statistical office (DANE) since 1970; and (iii) The Census of 1973. Major studies using these respective data sets are: (i) Work by Isaza and Ortega (1969), Prieto (1971), and Musgrove (1974, forthcoming) on the CEDE data; (ii) Published studies by Salazar (1971) and C6rdova (1971) and unpublished reports by DANE; (iii) Various government reports published in the Boletin Mensual de Estadistica and elsewhere. The main data sources and some of the tabular evidence are summarized in Fields and Jaramillo (1975). Some studies have been limited to presentation of simple tabular evidence as in Table 2. Others have constructed income (or earnings) functions using multiple regression analysis. The main results of the regression studies are summanized in Table 3. It is clear that the geographical coverages of these surveys, the populations sampled, and the types of variables included vary widely from one study to another. Nevertheless, we observe considerable agreement among these studies in a number of respects. First of all, education is always found to have an important positive effect on -22- TABLE 2. MEAN FAMILY INCOME CLASSIFIED BY VARIOUS FAMILY CHARACTERISTICS, FOUR URBAN AREAS, 1967-68, IN PESOS QUARTERLY Characteristic Mean Family Income Education of Head None 4,022 Primary 5,257 Secondary 11,163 University 27,299 Age of Head Less than 35 7,131 35-49 8,434 50-64 9,848 65+ 11,094 Family Size 1-2 5,369 3-4 7,114 5-7 8,782 8+ 9,563 Occupation Professionals 21,674 Vendors, commercial 8,806 Artesans, craftsmen, and operatives 5,694 Other employees 6,730 Source: Calvo and Fields (1975). TABLE 3. PRINCIPAL RESULTS OF STUDIES USING MICROECONOMIC SURVEY DATA TO CONSTRUCT INCOME FUNCTIONS IN COLOMBIA STATISTICALLY SIGNIFICANT YEAR OF DATA GEOGRAPHICAL SAMPLE DEPENDENT INDEPENDENT 2 AUTHOR AND SOURCE COVERAGE SIZE VARIABLE VARIABLES R Schults (1969) 1965 Bogotl 1,000 Logarithm Educational level, .17 - .24 Survey of - individuals of wage adjusted age, other family Employment and both sexes for a 48 hour income Unemployment work week (women only) (CEDE) Gonadlez (1971) 1967-68 Bogott 918 Income Educational level, .38 * Survey of individuals age, income source . Family Bud- both sexes (capital, independent gets and work, mixed or Expenditures salaried), sex (CEDE) Wsgrove (1974) 1967-68, Bogots, 2,949 Logarithm of Interactive variables .49 Survey of Barranquilla, families imputed "relative involving educational Family Bud- Cali, Madellfa long term income" level and age of gets and - of family family head, head's Expenditures marital and family (CEDE) statis., presence of capital income, nurber of workers in family, city Urrutia (1974) 1967, Bogotf, 331 Income Educational level, Approx- Survey of Bucaramanga, individuals age, sex 45 Occupational Hanizales, both sexes .4* and Geograph- Medellin ical Mobility (CEDE) Kugler (1975) 1970 National 607 Logarithm Educational level .50 National individuals of income and experience Household both sexes level of individual, Survey parents income (DANE) Fields (1976) 1967, Bogots, 331 Logarithm Educational level, .55 Survey of Bucaramanga, individuals, of income experience, sex, Occupational Manizales, both sexes city of residence, and Geograph- Medellfn occupation, parents' ical Mobility education and income (CEDE) Fields and 1973 Census National 860,000 Logarithm Educational level, . Up to .35 Schultz individuals of income age, department, (1977) rural/urban, employer/ employee -24- income. Second, age or experience are also found to be related signi- ficantly positively to income.1 Third, other variables, although statis- tically significant determinants of income, are not very important. Finally, these studies typically explain between forty and fifty percent of the variance in individual incomes.2 For the purpose of testing the hypotheses given in Section II, four research areas are of particular interest: the relationship between individual and industry characteristics, the importance of the individual's own attributes as compared with family background, the relative magnitude of personal and regional effects, and differ- ences in income-determining factors between urban and rural areas. These are treated in turn in the remainder of this section. 1The one exception to this generalization is the study by Urrutia (1974), in which the experience variable has the wrong sign as often as not. This is probably due to the unusual definition of experience which he employed: number of years the individual reports having worked (in all occupations) divided by age. Schultz (1968) reports a notably lower R than the others. This may perhaps be due to the small number of variables included, or to the fact that his study, being the earliest, is based on one of the first surveys in Colombia, with the possibility of correspondingly greater errors in measurement. -25- B. Individual and Industry Characteristics as Determinants of Incomes for Manufacturing Workers ii Urban Colombia The studies cited in TAble 1 revealed that incomes are associated with both individual and industry characteristics. Because those findings were based for the most part on simple tabulations, we lack information on the relative importance of individual characteristics versus industry characteristics as determinants of incomes. Some evidence on this ques- tion follows. Microeconomic data on individual workers in Colombia were drawn from the Survey of Urban Family Income and Expenditure (in the Spanish acronym, PRESFAM) conducted in 1967-68 by the Centro de Estudios sobre Desarrollo Econ6mico (CEDE), Universidad de Los Andes. The survey was carried out in the four major urban areas of Colombia: Bogota, Medell1n, Call, and Barranquilla.1 In all, the survey included some 3,000 families. All manufacturing workers were selected for this analysis,,producing a sample . of 877 individuals. For the resulting sample of currently-employed work- ers, their incomes could be linked to the characteristics of their indus- tries as well-as to their own personal characteristics.2 The dependent variable used is the logarithm of income. The reasons for this choice are: (i) Approximate lognormality in the distribution of income; (ii) Interpretation of the coefficients on each independent variable as percentage effects on income; (iii) Superior fit of the loga- rithmic form in past research on Colombian data. Income is used rather than labor earnings due to non-availability of the latter. The independent variables are of three types. One group (X1) per- tains to the characteristics of the workers themselves. These data are 1These are all large cities. Their respective populations in the most recent preceding census (1964) were: Bogotg, 1,697,000; Medellin, 773,000; Cali, 647,000; and Barranquilla, 498,000. 2It would have been preferable to have used the characteristics of the particular firm in which the worker was employed rather than the two- digit industry but this could not be done with the available data. f -26- derived directly from the PRESFAM survey. The available characteristics 1 are sex, education, and a proxy for experience. These variables are specified as in an earlier paper (Fields (1976)). A second group of characteristics (X2) relates the individual to the specfic two-digit manufacturing industry (e.g., leather goods) in which he or she was employed.2 One way of characterizing the industry is to intro- duce 19 simple dummy variables, each taking on the value one for the sector in which the individual was employed and zero for all others.3 The coeffi- cients on these variables are to be interpreted as the effect of working in that sector rather than in miscellaneous manufacturing. The other way of characterizing the industry is to assign the numerical value of each industry characteristic--for example, capital-labor ratio or size distri- bution of firms---to the individuals working in that industry. These 1Actual labor market experience is not available in this or other Colombian data sets. 2The industries and the number of cases in each were: Industry No. of Cases 1. Foodstuffs 68 2. Beverages 27 3. Tobacco 4 4. Textiles 91 5. Clothing and Footwear 217 6. Wood 30 7. Wooden Furniture 36 8. Paper and its Products 13 9. Printing 26 10. Leather 7 11. Rubber and its Products 15 12. Chemicals 56 13. Petroleum Derivatives 7 14. Non-metallic Minerals 27 15. Basic Metals 12 16. Metallic Products 39 17. Non-electric Machinery 17 18. Electric Machinery 64 19. Trans ortation Material 100 20. Miscelaneous Manufacturing 21 Total 877 3 . The reason there are only 19 dummy variables when there are 20 two-digit industries is to avoid perfect multicollinearity. The omitted category is miscellaneous manufacturing. -27- assignments were made on the basis of data from the 1967 Manufacturing Survey conducted by the National Statistical Department (DANE).1 A third set of factors (X3) pertain to the characteristics of a sector's labor force, specifically, their average education, average experience, and the proportion male. These averages were computed from this same sample of 877 manufacturing workers derived from the PRESFAM data. The general hypothesis is that X,, X2, and X3 all have significant independent effects in a linear model of the determinants of Colombian in- comes, i.e.,Income a + $1 1 2 2 3 X3. The specific hypothesis is that each of the components of X1 12 and X3 contributes significantly to the overall explanatory power of the model in accordance with the hypothe- ases of section II. The microeconomic regression results appear in Table 4. The depen- dent variable in each regression is the logarithm of income. The struc- ture of the table and the text below is first to present results for X1. X2, and X3 separately and then to explore their marginal effects in various combinations. Regression (1) includes only the personal characteristics variables. Education is found to play a significant positive role in explaining income. The estimated coefficient implies approximately a 15% increase in income for each additional year of schooling. This estimate accords with other regression studies and with estimates of the rates of return to educational investments. Experience is included both linearly (EXP) and quadratically (EXPSQ). (The quadratic specification, with expected positive coefficients on EXP and negative on EXPSQ, allows for the possi- bility that income increases at a decreasing rate or even declines after some point.) The data confirm that more experienced workers have higher incomes but the proportional gain to additional experience declines. IThese figures formed the basis of an earlier study by Nohra de Marulanda and myself (1976). The data are reproduced in Table Al of the Fields-Marulanda paper. MULTIPLE REGRESSIONS, INDIVIDUAL AND INDUSTRY CHARACTERISTICS FOR HANUFACTURING WORKERS IN URBAN COLOMBIA, 1967/68 DEPENDENT VARIABLE: LOC OF INCOME (LNINC) (1) (2) (3) (4) (5) (6) (7) (8) EDUC .15566 .14696 .14781 .14803 .14747 (.00752) (.00785) (.00774) (.00788) (.00787) EXP .06869 .06932 .06801 .07003 .06899 (.00785) (.00786) (.00780) (.00782) (.00782) EXPSQ -.00086 -.00088 -.00085 -.00088 -.00087 (.00014) (.00014) (.00014) (.00014) (.00014) SEXD .48753 .41765 .44860 .39811 .40250 (.06651) (.07263) (.06669) (.07250) (.07248) PROD 5.17 2.91 2.14 1.95 (1.59) (1.80) (1.30) (1.48) CAPINT -.01280 -.00768 .00074 -.00198 (.00998) (.01147) (.00817) (.00941) SIZE .00535 .00140 .00231 .00285 (.00479) (.00630) (.00392) (.00517) FOREIGN .02819 .01521 .01005 .00936 (.00504) (.00667) (.00420) (.00547) AVGEDUC .01667 .01030 .00486 .00051 (.00331) (.00418) (.00279) (.00348) AVGE .00234 .00214 .00055 .00060 (.00168) (.00201) (.00140) (.00166) PROPHALE .46221 .40006 .23255 .16448 (.11160) (.11931) (.09991) (.10618) DSEC1 .12924 (.19624) DSEC2 .16556 (.22978) DSEC3 .54589 (.42572) DSEC4 .25332 (.18945) DSEC5 -.11359 (.18149) DSEC6 -.00536 (.22396) DSEC7 -.10602 (.21500) DSEC8 -.27941 (.27482) DSEC9 .07091 (.23128) DSEClo -.07107 (.34028) DSEC11 .43481 (.26343) DSEC12 .39059 (.20128) DSEC13 .40076 (.34059) DSEC14 .45126 (.22720) DSEC15 .12733 (.28292) DSEC16 .04917 (.21167) DSEC17 .13623 (.25531) OSECIS .13773 (.19782) DSEC19 .07512 (.18829) CONSTANT 3.51 5.63 3.93 4.29 3.54 3.43 2.98 3.19 R2 .383 .078 .092 .100 .412 .396 .393 .398 -29- Notes to Table 4. Note: The variables DSEC1 through DSEC19 refer to the following industries: Industry 1. Foodstuffs 2. Beverages 3. Tobacco 4. Textiles 5. Clothing and Footwear 6. Wood 7. Wooden Furniture 8. Paper and its Products 9. Printing 10. Leather 11. Rubber and its Products 12. Chemicals 13. Petroleum Derivatives 14. Non-metallic Minerals 15. Basic Metals 16. Metallic Products 17. Non-electric Machinery 18. Electric Machinery 19. Transportation Material The omitted category is miscellaneous manufacturing. -30- Sex is introduced as a dummy variable, taking on the value one for males and zero for females1 It is found to be highly significant, the estimated coefficient implying that a man would earn nearly 50% more than a woman with comparable education and experience. Together, these variables explain 38% of the variance in (the logarithms of) income in urban Colombia, comparing favorably with other micro level studies. Regression (2) relates income to the characteristics of the firms in the two-digit industry in which the individual works. Incomes are.significantly higher for individuals in industries with high productivity (value added per worker) and high proportions of foreign investment. Interestingly, the size effect, although positive, fails conventional significance tests and capital intensity has a negative (but insignificant) relationship with income, counter to hypothesis. 2 Regression (3) relates income to the characteristics of the workers in the two-digit industry in which the individual is employed. Workers in industries where the labor force is better-educated and where a higher proportion are male earn significantly higher incomes. Higher average experience is associated with higher incomes, but this effect is not statis- tically significant. Neither the average characteristics of the firms in an industry, nor the average characteristics of the workers in the industry, nor the two taken together provides as much explanation for individual incomes as do workers' own characteristics. This follows from a comparison of the R2's: 1Similar results were obtained when the same regression was run on a sample stratified by sex. 2See p. 35 for a discussion of measurement problems which affect the inter- pretation of these results. -31- Regression (2), based on average characteristics of firms R = .08; 2 Regression (3), based on average characteristics or workers R 2 .09; Regression (4), based on average characteristics of firms and workers together R2 - .10; 2 Regression (1), based on workers' own characteristics R . .38. A question that immediately arises from these results is: do the industry characteristics contribute significant additional explanatory power beyond that provided by an analysis of individual workers' characteristics alone? The question may be addressed using the dummy variables procedure described earlier. Let us add dummy variables for the particular two-digit sector to the personal characteristics included in regression (1). The results of this exercise are given in regression (5). The estimated coefficients vary substantially, suggesting 55% higher wages for workers in the tobacco industry (DSEC 3) than in miscellaneous manufacturing and 28% lower wages for workers in paper manufacturing than in miscellaneous manufacturing. (Other sectors lie in between.) Adding these dummy variables increases the R2 from .383 to .412. This increase in explanatory power, although modest, is statistically significant by the usual F test. These findings show that the characteristics of the economic subsector partially explain individual earnings. We do not yet know exactly how important the sector character- istics are as compared with other variables or which sector characteristics matter. Evidence on both these points is presented below. -32- TABLE 5. ANALYSIS OF VARIANCE, MANUFACTURING WORKERS IN URBAN COLOMBIA, 1967/68. Dependent Variable: Logarithm of Income F Ratio Proportion of (Significance Degrees of Main Effects Variance Explained Level) Freedom Education .201 84.1 (.001) 3 Experience .089 27.9 (.001) 4 Sector .047 3.1 (.001) 19 Covariance -.013 Total Explained .324 15.7 (.001) 26 Note: Explanatory categories are as follows: Education, four categories: None, some or all primary, some or all secondary, some or all higher Experience, five categories: fewer than 5 years, 5-14 years, 15-24 years 25-34 years, 35 or more years Sector, twenty categories: As in Table 4, plus miscellaneous manufacturing -33- Considering first the relative importance of sector and individual characteristics, it is instructive to examine the results of an analysis 1 of variance. These results are given in Table 5. Of the 32.4 percentage points of log variance explained, education and experience together account for 29.0 percentage points and sector for only 4.7 percentage poir-s. When these proportions explained are adjusted for degrees of freedom used, the difference is even more pronounced: 4.1 percentage points per degree of freedom for education and experience, 0.2 percentage points per degree of freedom for sector. These differences are reflected in the F statistics as well. The analysis of variance results reinforce the finding from regressions (1) - (4): sector characteristics are not very important rela- tively; of the proportion of variance explained, the bulk is provided by education and age. Although sectoral variables are less important than might have been thought, they still have a significant role to play. Let us now see which of the sectoral variables have the strongest effects. We may in- vestigate this question in either of two ways. One way is to relate the shift effect to industry characteristics; the other is to work with the industry variables directly in the microeconomic earnings function. The first approach checks whether the regression coefficients for the 19 dummy variables in regression (5) of Table 4 are systematically related to the average characteristics of the firms or workers in the sector (vectors X2 and X3 respectively), i.e., if the shift effect of being in a particular sector is associated with measured characteristics of the sector. Using ordinary least squares (which produces consistent but inefficient estimators), the regression results are: Although the results in Tables 4 and 5 are essentially similar, there are some small differences. The reason is that the explanatory variables are expressed in continuous form in the regressions whereas in the analysis of variance the computer program required that the variables be put into categorical form. Limitations on machine capacity forced the exclusion of some explanatory variables. For both these reasons, the analysis of variance appears to have somewhat lower overall explanatory power than does the regression analysis. -34- COEF = -.057 + .282 x 10-5 PROD -.0060 CAPINT -.0013 SIZE + .0162 FOREIGN, -5 . (.070 x 10-5 (.0059) (.0037) (.0057) R2 . .548, -2 R - .419, for the characteristics of industries, 2 COEF - -.083 + .0031 AVGEDUC + .048 AVGEXP .-.101 PROPMALE,R = .036, (.0375) (.018) (.358) -2 R < 0, for the characteristics of workers, and COEF - .085 + .284 x 10-5 PROD -.0095 CAPINT + .0015 SIZE + .021 FOREIGN (.077 x 10-5 ) (.0070) (.0047) (,007) -.038 AVGEDUC + .0047 AVGEXP -.033 PROPMALE, R = .629, - .394. (.030) (.0159) (.297) for the two taken together (standard errors in parentheses). These results indicate that the sector's characteristics help. explain the shift intercept. Those factors which contribute to the explanation are value added per worker and extent of foreign investment. The statisti- cal significance of these variables means that high productivity or foreign firms are paying higher wages to comparable workers (to the extent that comparability is achieved by controlling for schooling, 'experience,' and sex). Note also the statistical insignificance of the other variables. This means that we do not have evidence for either of two propositions: that industries with capital-intensive or large firms pay higher wages ceteris paribus, or that relatively low-productivity workers benefit by working alongside higher productivity ones. The other approach for determining which industry characteristics affect individual earnings is to introduce the industry characteristics as additional explanatory variables in a regression which already includes -35- personal characteristics. This is done in Table 4, in regressions (6) and (7) for the average characteristics of the firms and workers respective- ly, and in regression (8) for the two taken together. The results sustain the importance of foreignness but cast doubt on the importance of value added per worker (PROD), the coefficient of which is statistically insig- nificant in the presence of individual characteristics. Note how small the additional contribution is: an extra .013 in explanatory power for the firms' characteristics, an extra .010 for the average workers' characteristics, and an extra .015 for the combina- tion. Observe also the virtual constancy of the regression coefficients on EDUC, EXP, EXPSQ, and SEXD and the continued insignificance of the other characteristics of firms. The regression results presented in Table 4 lead to the following conclusion: workers' own characteristics are foremost in determining incomes at the individual level in Colombia, at least among urban manu- facturing workers, the characteristics of the industry, as best we are able to measure them, enter secondarily. The small additional contribution of sectors' characteristics to explaining individual incomes is open to a variety of interpretations. One possibility is that the individual's education, experience, and sex are overwhelmingly important in determining income and that sector characteristics are really of a second (or even third) order of impor- tance. But before we accept this interpretation, we should recognize another possibility: that the methodology utilized tends to work against findings firms' characteristics important. Remember that the firm data are based on average values for the two-digit industry as a whole. Sup- pose there are substantial differences within each of these sectors, as -36- would appear to be the case. By using as explanatory variables the characteristics of the two-digit industry in which the individual works rather than his or her own particular firm, we may have irroortant errors in measurement in the independent variables. This kiridd of imprecision would reduce the aipparent explanatory contribution of the poorly- measured variables. Perhaps in the future, Colombian data sets will permit matching of the characteristics of individual workers with those of their own firms, and we can then test among these alternative interpretations. Meanwhile, we may seek other clues by analyzing average wages across sectors, a task performed in Section IV. In summary, on the relative importance of iLndividual and industry characteristics in determining wages and incomes in urban Colombia: (1) Workers' characteristics explain A considerable share of the variance in i.comes. A simple model taking into account just education, experieuce, and sex accounts for 38% of income inequality at the individual level. (2) In contrast, ind,istry characteristics enter secondarily. Only 10% of the variance is explained by regressions including the average characteristics of firms (value added per worker, capital intensity, size distribution of firms, and importance of foreign investment) and the average characteristics of the industry's labor force (average education, average experience, and proportion male). (3) In a multiple regression including measures of the individual's characteristics, the characteristics of the firms in the industry in which he is employed, and the average characteristics of the industry's 1See, for example, the work of Baily (1977) for an in-depth analysis of technological variability within the shoe and brick industries in Colombia. -37- labor force, it is found that the set of individual's characteristics dominates the other groups of variables. Just 3% additional variance is explained when the two groups of industry characteristics are introduced in the presence of the individual's characteristics. (4) Those industry characteristics which do contribute significantly to explaining individuals' incomes in a multiple regression are value added peri worker and the degree of foreign investment in the industry. (5) All other things equal, no statistically significant wage advantage is found for workers in industries with larger proportions of large or capital-intensive firms or in industries with highly-educated or -experienced workers. This does not mean that such effects are necessarily absent in Colombia; rather, it is that if they are present they cannot be detected in the currently-available data. -38- C. Individual and Family Background Characteristics as Determinants of Incomes in Urban Colombia In Colombia, as in many other less developed countries, there has been a great deal of debate concerning the role of education in advancing economic development.1 This issue is particularly important in the context of evaluating the potentiality of human resource strategy in promoting economic and social mobility, narrowing income inequality, and helping to alleviate absolute poverty. Two opposing views stand out. Those in one group, principally economists, have argued that Colombia's educational system is already making an important contribu- tion.2 Others, in general sociologists and political scientists, sustain the opposite position: that education as a factor producing 3 social mobility is little more than a myth. Recent research--by Kugler (1975) and Fields (1976)--offers evidence on the merits of these conflicting claims. Kugler's study is both pathbreaking and insightful. As shown in Table 3, Kugler found three variables---the individual's education, his experience, and (the logarithm of) his father's income--to be statis- tically significant determinants of income. The results of Kugler's study suggest that the individual's own characteristics are more impor- tant than his family background. In Kugler's words: "The results ob- tained indicate that contemporaneous variables, especially education, as well as socio-economic antecedents are important direct determinants of labor incomes, perhaps with more weight to the former than to the 1 Among recent works which merit attention in this field are those of Blaug (1973) and Harbison (1973). . The most prominent exponent of this view in Colombia is the Minister of Energy and former Director of National Planning, Miguel Urrutia. See Urrutia (1974) and Berry and Urrutia (1976, Chapter 9). 3This position is advanced and defended by Parra (1973). -39- latter." In addition, he. found that socio-economic background is an important determinant of educational attainment, suggesting that family background may have an important indirect effect on income. There is some room for doubt about the general applicability of Kugler's conclusion owing to the nature of his sample. The data were taken from a household survey, including information on each person living in the household. The respondents were not asked about their socio-economic origins. Consequently, Kugler was limited to those households in which at least two generations of income earners were living together. The probable effect of such a sampling procedure is to include disproportionately large numbers of young workers and to introduce selectivity bias by oversampling those who are poor enough to be forced to remain in an extended family. This might tend to bias the results in favor of factors which determine short-run econo- mic position and away from factors which determine one's economic success in the longer run. Given these possible biases, I sought in an earlier paper [Fields (1976)] to examine whether these same qualitative results are found for a sample of workers at all stages of their working lives. The data for such an exercise were taken from a survey of occupational and geographic mobility conducted by the Centro de Estudios Sobre Desarrollo Econ6mico (CEDE) of the Universidad de Los Andes in four urban areas of Colombia (Bogota, Medellin, Manizales, and Bucaramanga) in 1967, including 331 workers.1 While parents' income was not asked in the CEDE survey, parents' education and occupation were included. These variables may be related to the individual's own characteristics to see if they have 1 1 A general description of the data may be found in Fields and Jara- millo (1975). For additional details and basic results, see Garcia (1968). These data provide the basis for the paper by Urrutia (1974). I wish to express my gratitude to CEDE, to Dr. Urrutia, and to his assistant, Lia Guterman, for kindly making these data available to me. -40- an independent effect and, if so, how important that effect is. The dependent variable is the logarithm of the individual's income (LNY). The independent variables are of two general types: those that pertain to the individual and those that pertain to his parents. (For specific definitions, see Fields (1976)). In the first group, we have: the individual's education (EDUC), measured in terms of number of years completed; the number of years of vocational educa- tion completed (VOCEDUC); the individual's experience, defined as age minus schooling minus seven, entered both linearly (EXP) and quadra- tically (EXPSQ); an interaction between education and experience (ED*EXP); two dummy variables for the individual's occupation, according to whether the person is in a white-collar occupation or not (OCCUP1), or a commercial occupation or not (OCCUP2); a dummy variable for the person's sex (MALE), taking on the value one for men and zero for women; a dummy variable taking on the value one if the individual migrated to the urban area in which he now resides after the age of twelve (MIG), zero otherwise; and three dummy variables identifying the city of residence: MED if Medellin, zero otherwise; kAN if Manizales, zero if otherwise; and BUC if Bucaramanga, zero otherwise.2 The variables pertaining to the individual's economic origin include: father's and mother's education (PAEDUC and MADEDUC), and two dummy variables for father's occupation (PAOCCUP1 and PAOCCUP2), defined in the same way as the individual's occupation. *1 1In other regressions not reported here, the sample was stratified according to sex. 2With these definitions, BogotA is the omitted city. According to evidence presented by Isaza and Ortega (1969) , the average income in Bucaramanga is equal to that in Bogota, while Medell±n has higher income and Manizales lower income. Hence, it is hypothesized that MED has a positive coefficient, MAN a negative coefficient, and BUC a null coefficient. -41- tABLE 6. FACTORS EXPLAINING INCOME, URBAN COLOMBIA, 1967 Dependent Variable - Logarithm of Income Independent Variables (1) (2) (3) EDUC 1.16130 .15062 (.01963) (.02047) VOCEDUC .00012 .01029 (.04117) (.04150) EXP .06312 .06385 (.01435) (.01481) EXPSQ -.00086 -.00087 (.00023) (.00023) ED*EXP -.00162 -.00158 (.00079) (.00082) OCCUP1 .40224 .35790 (.13586) (.13785) OCCUP2 .32884 .32128 (.07998) (.08220) MALE .58641 .59356 (.07035) (.07106) MIG -.05214 -.04094 (.06745) (.06836) MED .03728 .01019 (.08316) (.08400) MAN -.32309 -.35610 (.09937) (.10021) BUC -.03571 -.05885 (.08913) (.08976) PAEDUC .004001 -.02350 (,02187) (.01686) MAEDUC .09486 .04119 (.02405) (.01929) PAOCCUP1 .41042 .14799 (.21227) (.16121) PAOCCUP2 .19265 .05082 (.11276) (.08818) Constant 4.56102 6.17584 4.55692 R2 .569 .169 .576 -42- Regarding the individual's dwn characteristics, the hypotheses are that LNY is positively related to EDUC, EDUCVOC, EXP, OCCUP1, OCCUP2, and MALE and negatively related to EXPSQ, ED*EXP, MIG, MED, MAN, and BUC, while for parents' characteristics, we would expect LNY to be a positive function of PAEDUC, MAEDUC, PAOCCUP1, and PAOCCUP2.1 If both parents' characteristics and the individual's characteristics are important independent determinants of income, we would expect variables of both types to be significant in a regression that includes both sets. Considering first the relationship between LNY and the individual's own characteristics, we see in equation (1) of Table 6 that most of the variables behave as hypothesized. EDUC, EXP, EXPSQ, ED*EXP, OCCUP1, OCCUP2, and MAN all have the expected sign and are highly statistically significant. The other variables (VOCEDUC, MIG, MED and BUC) have the right sign but are not significantly different from zero. Together, these variables are found to explain 56% of the variance, which sur- passes the coefficient of determination found in earlier studies (see Table 3). Turning now to the relationship between LNY and the education and occupation of one's parents, we find that the results generally conform with the hypotheses (see equation (2) of Table 6), but they are much weaker. Mother's education and father's occupation have the right 'signs and are highly statistically significant. However, the explanatory power of the regression involving only parental character- 2 istics (R = .17) is considerably lower than the earlier one based on the individual's characteristics. This suggests that in a multiple regression involving both types of variables, socioeconomic origin 1See Fields (1976) for justifications for these hypotheses. -43- would be relatively less important than the personal attributes of the individual. In fact, this is just what we find (see equation (3)). Only mother's education is found to have a statistically significant effect in the presence of the individual's own characteristics. Furthermore, the coefficients of determination in equations (1) and (3) are identical to two decimal places. These results lead to the following conclusion: Incomes in urban Colombia are determined by the economically-relevant characteristics of workers, not of their parents. Parental background makes no significant additional direct contribution to the explanatory power of the model. The preceding result does not exclude the possibility that parental background may have important indirect effects, for example, in deter- mining the educational characteristics of workers. Indeed, previous studies in Colombia have demonstrated a strong relationship between socio- economic background of parents and the socio-economic status of their children1 and between parental background and children's education.2 it is interesting, therefore, to examine the extent to which parental background, along with other characteristics of an individual, can explain his or her educational attainment. The simplest way of relating an individual's education to socio- economic background is to cross-tabulate the number of years of education completed by the individual (EDUC) and the mean years of education completed by his or her parents (PAREDUC). These data are shown in Table 7. While there is a pronounced relationship between the two, the 1See Garcia (1968), Lemoine and Pereira (1975), and Kugler (1975). 2 See Rama (1969), Urrutia and Sandoval (1971), Parra (1973), Urrutia (1974), and Kugler (1975). -44- TABLE .7. EDUCATION OF PARENTS AND OF THEIR CHILDREN, FOUR COLOMBIAN CITIES, 1967 Education of Parents (mean) Education of Younger 1 or Generation 0 1-3 3-5 5-8 8-11 more Total 0 1.6% 0.9% 0.3% 0.0% 0.0% 0.0% 4 .8% (12) (3) (1) (0) (0) (0) (16) 1-3 7.5% 10.0% 4.5% 0.9% 0.0% 0.0% 23.0% (25) (33) (15) (3) (0" (0) (76) 3-5 3.6% 11.5% 13.0% 1.5% 0.6% 0.0% 30.2% (12) (38) (43) (5) (2) (0) (100) 5-8 0.9% 4.5% 8.2% 3.0% 0.9% 0.1% 17.5% (3) (15) (27) (10) (3) (0) (58) 8-11 0.3% 2.7% 5.4% 3.9% 2.4% 0.6% 15.4% (1) (9) (18) (13) (8) (2) (51) 11 or more 0.0% 0.3% 2.1% 2.7% 2.7% 1.2% 9.1% (0) (1) (7) (9) (9) (4) (30) Total 16.0% 29.9% 33.5% 12.1% 6.6% 1.8% 100.0% (53) (99) (111) (40) (22) (6) (331) -45- correlation is far from perfect. There are large numbers of workers in the younger generation whose education greatly exceeded their parents', and at every educational level, there are non-trivial numbers of younger workers who attained less education than their parents. In other words, parents' and children's educations are associated with one another but one is not automatically determined by the other. Moving on to consider other variables, we may formulate the follow- ing hypotheses: (1) EDUC is a positive function of father's and mother's education (PAEDUC and MAEDUC); (2) EDUC is a positive function of father's occupational position (PAOCCUP1 and PAOCCUP2), with a larger effect for the former than for the latter; (3) EDUC is a negative function of migratory status (MIG); a ! (4) EDUC is greater for males than for females. The empirical results are given in Table 8. We find: (1) Parents' Education. The parental education variables are strongly significant. Furthermore, the education variables account for nearly all of the explained variance. The simple correlation between EDUC and PAREDUC (the mean of PAEDUC and MAEDUC) is +.687. This implies that in a simple regression, PAREDUC would explain 47% of the variance in EDUC. Only an additional 2% is explained using four extra variables. (2) Father's Occupation. The results only partially confirm the hypotheses. While PAOCCUP2 is always positive and statistically significant, PAOCCUP1 is not. The difference between the estimated coefficients is not signi- ficantly different from zero. (3) Sex. Sex is not a significant explana- tory variable when entered linearly. Likewise, in supplementary regres- sions when the sample is stratified by sex, the regression relationships -46- TABLE 8. FACTORS EXPLAINING EDUCATIONAL ATTAINMENTS, URBAN COLOMBIA, 1967 Dependent Variable - Years of Education Completed Independent Variable PAEDUC 0.28709 (.07585) MAEDUC 0.54113 (.08369) PAOCCUP1 0.94450 (.73776) PAOCCUP2 1.37495 (.39636) MALE 0.31671 (.32363) MIG 0.06933 (.31060) CONSTANT 2.26961 R 2 .494 -47- are not found to be statistically significant. Thus, there is no convincing evidence for the proposition that Colombian parents discri- minate in education by faVoring their sons over their daughters. (4) Migratory Status. Being a migrant from a rural area is not found to have a statistically significant effect in any of the regressions. This may be because this is an urban sample and those rural residents with less education did not migrate. The principal result from this examination of parents' and children's education is that there is a statistically significant and economically impor- tant relationship between the education and occupational position of Colom- bian parents and the education of their children. Along with the earlier finding that education is a major determinant of income, we find that the educational system in Colombia is an important indirect means of transmitting economic status from one generation to another. This result is not surprising, since it is predicted both by human capital theory and by models of social stratification. Still, the quantitative importance of family background is less than might have been thought. In summary, on the relative importance of personal and family background characteristics in determining wages and incomes in urban Colombia: (1) A quite considerable fraction of the variance in incomes can be accounted for by variability in education, experience, and other individual characteristics. In contrast, parental background is found .-48- to have only limited direct explanatory power, with no significant addition beyond what is contributed by personal characteristics alone. Thus, the view held by some that wages and incomes in Colombia are determined more by "whom you know" rather than "who you are" is refuted as a general matter, although it may have some merit for particular socio-economic groups, particularly the well-to-do. (2) Access to education is limited and is received disproportion- ately by the relatively well-to-do. High socio-economic status parents are more likely to educate. their children than parents whose economic position is less favorable. However: (3) Despite the unquestioned relation between parents' socio- economic status and that of their children, at least in the urban areas, many spaces in the schools go to disadvantaged children. The Colombian educational system as a whole promotes economic mobility for substantial numbers of citizens, and is not limited exclusively or even predominant- ly to the children of the rich. The correlations are unmistakable but the situation is far from deterministic. It is precisely because of this non-determinism that knowledge of an individual's own characteris- tics is more helpful in predicting income than is knowledge of that person's family background. (4) Taken together, these findings suggest that inattention to family background varibles in most studies of wage structure in Colom- bia and elsewhere is not a major omission. -49- D. Personal and Regional Effects on Income Structure in Colombia Regional inequality is of interest for a variety of reasons: gauging the degree of a country'. labor market integration; understanding patterns of population movement in general and labor force migration in particular; predicting future urbanization; assessing the ability of the economic and political systems to deal with prevailing inequali- ties; and characterizing the poor so as to channel resources their way. In the development literature, the best-known studies of regional inequality are those of Kuznets (1963) and Williamson (1965). Note- worthy in the Colombian context is thp research on income inequality by Nelson, Schultz, and Slighton (1971), Prieto (1971), Berry and Urrutia (1976), and Fields and Schultz (1977). Incomes in Colombia are known to be associated with place of resi- dence. In the four principal cities, Prieto (1971) reported mean family expenditures (in pesos per three months ) that differed as follows: Bogota Col. $8,150 Barranquilla 7,090 Cali 6,640 Medellin 5,980 Average, $7,230 four cities Isaza and Ortega (1971) found similar differences. Because of these differentials, Musgrove (1974, forthcoming) analyzed incomes in each Colombian city separately. Nationwide, states (or as they are known in Colombia "departments") differ widely in per capita income. Berry and Urrutia (1976) citing work by Marabelli for 1964, report the follow- ing differentials by department: -50- TABLE 9. INCOME DIFFERENTIALS BY DEPART1ENT IN COLOMBIA, 1964 GDP GDP per'Capita per Capita (National Implicit in average-100) Indices Antioquia 101 2,949 Atlintico 101 2,949 Bolirvar 88 2,569 Boyaci 71 2,073 Caldas 91 2,657 Cauca .61 1,781 Cordoba 83 2,423 Cuandinamarca 101 2,949 Distrito Especial (Bogota) 159 4,643 Choco 32 934 Huila 79 2,307 Guajira 52 1,518 Magdalena 99 2,891 Meta 119 3,475 Nariflo 53 1,548 Norte de Santander 90 2,628 Santander 104 3,037 Tolima 105 3,066 Valle 128 3,737 Colombia , Average 100 2,919 Source: Berry and Urrutia (1976, Table 5.2) -51- Urban-rural differentials around 1970 were of a similar order of magnitude according to the national statistical office (DANE (1971)): average income was Col. $1,356 in urban Colombia, Col. $586 in rural Colombia. These sorts of data have led many to the conclusion that distortions in regional labor markets constitute a critical problem for the Colombian economy and that region- and locality-specific labor market policies are needed to enable the poor to participate in economic development. Recent research by Fields(f977a) and by Fields and Schultz (1977) has examined the relative importance of personal and regional effects on income variation in Colombia. These studies, based respectively on urban and national data, lead to the same conclusion: that the importance of interregional inequality has been overstated, since the great bulk of income inequality occurs within regions and not between them. The results will now be reviewed. Fields(1977a) used the data from the PRESFAM survey of 1967-68 for the four major cities of Colombia. At the very simplest level, it is interesting to ask how much of the income variability among the nearly 3,000 sample families is associated with differences across the various cities and how much to differences within them. An Analysis of Variance (ANOVA) was performed in which the dependent variable was the logarithm of family income and the independent variable was the city of residence. The ANOVA results are given in ANOVA (1) of Table 10. We find that city is significant statistically but not economically in explaining urban income inequality. The test for statistical significance, the F ratio, surpasses the .01 significance level. Nonethelessi, 1This data set also formed the basis for the comparison between individual and industry characteristics in Section B above. tABLE 10. DECOMPOSITION OF INEQUALITY IN URBAN COLOMBIA BY INCOME DETERMINANTS, 1967-68 Decomposition of Log Variance ANOVA (1) ANOVA (2) Significance Significance Source of Variation Sum of Squares F of F Sum of Squares F (df) of F Main Effect Explained: City 9.8 (0.4%) 3,825 .01 9.2 ( 0.4%) 5.74 (3) .001 Education 876.4 (34.7%) 546.1 (3) .001 Age 106.3 ( 4.2%) 66.2 (3) .001 Covariance - 58.9 (-2.3%) Total, Main Effects 933.0 (36.9%) 193.8 (9) .001 Two-Way Interactions Explained: City-Education 13.3 ( 0.5%) 2.76 .(9) .003 City-Age 4.3 ( 0.2%) .90 (9) .999 Education-Age 21.9 ( 0.9%) 4.54 (9) .001 Covariance 1.4 ( 0.0%) Total, Two-Way 40.9 ( 1.6%) 2.83 (27) .001 Interactions Three-Way Interactions Explained: City-Education-Age 13.0 ( 0.5%) .90 (27) .999 Total Explained 987.0 (39.0%) 29.3 (63) .001 Unexplained 2519.4 (99.6%) 1542.3 (61.0%) Total 2529.2 (100.0%) 2529.2 (100.0%) -53- a negligible share of the variance in log income---only 0,4%---is explained by variation across cities. Nearly all of the inequality in urban Colombia is due to variations within cities. In other words, despite the importance ascribed by some authors to inter- city wage differentials, knowledge of a family's city of residence provides very little information on its income. How much further can we get with other information on the family? The urban data permitted analysis of education and age effects as well. The explanatory variables for this expanded analysis of variance were then: City: Bogotd, Barranquilla, Cali, Medellfn Education of head of household: None, primary (some or all), secondary (some or all); higher (some or all) Age of head of household: Less than 35, 35-49, 50-64, 65 and over. Column (2) in Table 10 presents the results of the inequality decomposi- tion according to these factors. Looking first at the main effects, each explanatory factor helps account for inequality. The significance column shows that each of these effects is statistically significant at the .001 level. However, the contributions of the three sets of factors are by no means equal. Of the 36.9% of the log variance explained by the main effects, education accounts for nearly all of it, 34.7%. By contrast, age accounts for just 4.2% and city 0.4%.1 Education 2 thus overwhelms the other explanatory factors. Immediately below the main effects in Table 10 are the interaction effects. The education-city interactions, for example, allow for the possibility that the effect of education on income might depend on which city one lives in or alternatively that the effect of city on 1The whole is less than the sum of its parts due to negative covariance among the main effects. 2Thus, it may be inferred part of the difference in average incomes among cities is due to differences in educational attainments among the various cities' labor forces. Direct evidence based on national data in Colombia is given in Fields and Schultz (1977). "-54- income might depend on one's level of education. The three sets of two-way interaction effects---city-education, city-age, and education- age--together add significantly to the explanation of inequality, but they account for only 1.6% of the log variance. Thus, the explanatory effects of education, age, and city are not independent of one another, but the degree of interdependence is small. Whether the 1.6% additional explanatory power contributed by the two-way interaction warrants a quadrupling of the number of explanatory categories from 9 to 36 is a matter of some economic judgment. The three-way interactions, how- ever, contribute even less explanatory power, only 0.5%. Even on narrow statistical grounds, their inclusion is not justified. It might be objected that urban data sets are inappropriate for testing the importance of regional inequality, since the greatest varia- tion in labor market conditions is probably not between large cities but rather is between advanced and backward regions or between urban and rural areas. This objection is met in research on national data by Fields and Schultz (1977), the highlights of which will now be reviewed. The Fields-Schultz study is based on an extraordinary data set: a public use sample from the 14th Colombian Census of Population, conduct- ed in October, 1973. The Census enumerated approximately 21 1/2 million persons. A four percent sample of returns was converted to machine readable form for purposes of statistical analysis. These 860,000 cases form the statistical base for the study. Table 11 presents the basic data on the relationship between income and various characteristics of the individual. The gross differentials (without cross-classification) are of the following orders of magnitude: eleven-to-one ratio between persons with higher education and persons -55- TABLE'l. Mean Monthly Incomes in Pesos in Colombia, October, 1973 For Persons with Various Characteristics Department of Residence Rural/Urban Antioquia 1536 Rural 536 Atlantico 1872 Urban 1,676 Bogot& D.E. 2694 Bolivar 1324 Education Boyaci 897 None 610 Caldas 1253 Primary 982 Cauca 819 Secondary 2,337 Cesar 1391 University 6,898 C6rdoba 1039 Cundinamarca 946 Sex Choc6 656 Male 1,540 Huila 1093 Female 1,232 La Guajira 1726 Magdalena 1279 Employment Status Meta 1360 Employer 2,093 Narifto 667 Employee 1,275 Norte de Santander 1073 Age Quindlo 1337 10-19 532 - Risaralda 1372 20-24 1,034 Santander 1151 25-29 1,523 Sucre 1121 30-34 1,830 Tolima 1205 35-44 1,941 Valle 1608 45-54 1,961 55-64 1,710 65 and over 1,279 Source: (Fields and Schultz (1977, Table 1)] with none; four-to-one between prime age workers and the very young; a twenty-five percent differential advantage for men over women; four- to-one ratio between the richest department and the poorest; three-to- one between urban workers and rural workers; and sixty percent more for employers and the self-employed than for wage and salary employees. It is sometimes thought that gross income differentials like those observed in Table 11, arise from failure to standardize for other factors determining income. Regional inequality in particular is attributed by some to differences in the composition of various departments' labor forces by education, sex, or age. Were this view correct, incomes within educational, sex, or age categories would be nearly equal across departments even though the department averages differed. However, when we examine the data and break the sample down into cross tabulated categories across these dimensions, the inter- regional income differentials remain in only somewhat attenuated form. An example from the Fields-Schultz study is given in Table 12. Income increases with education, not only in the country as a whole, but for men and women in every department. Women's incomes are less than men's for each educational group in a given department as well as in the aggregate means. Wide interregional differences are observed in all educational categories for both sexes. Thus, notable inter- regional differences are found in Colombia for more or less homogen- eous groups of workers, as are differences by education and sex. This same qualitative result has also been observed in the case of Venezuela; see Schultz (1975). At issue is the relative importance of these different effects. -57- TABLE 12. Male and Female Monthly Incomes, October 1973, for Colombia By Departnht and Education (in Pesos) HALE FE MA LE BOTH Fucntlon ALI. Education SEXES DEPARTHENT NONE PRIMARY SECONDARY IITCHER MALES NONS PRIMARY SECONDARY AICHER AELES TOTAL Antioquia 703 1092 2732 7997 1580 468 769 1755 3386 1315 1516 (3429) (10542) (3566) (629) (18L66) (223) (1564) (1637) (156) (3580) (23746) Atlantico 820 1358 2710 7104 2034 572 842 1588 2801 (600) (2570) (1434) (295) (4899) (83) (567) (674) (9)1 Bogota 912 1337 2974 8370 2702 559 776 1852 3678 1508 2694 D.E. (815) (8744) (5910) (2081) (17550) (323) (2818) (3205) (583) (6929) (21479) Bolivar 753 1091 2304 7317 - 1389 466 651 1500 3421 1036 1324 (978) (1611) (647) (86) (3322) (127) (320) (278) (28) (753) (4075) Boyaci 343 654 2478 5474 888 252 497 1509 3568 960 897 (910) (2920) (484) (89) (4403) (92) (253) (237) (20) (602) (5005) Caldas 712 961 2447 7754 1282 390 608 1514 3116 1084 1253 (665) (2639) (585) (84) (3973) (42) (305) (301) (22) (670) (4643) Cauca 412 633 1887 5758 809 325 500 1574 3228 876 819 (573) (1564) (236) (50) (2423) (83) (194) . (143) (8) (428) (2851) Cesar 918 1182 2542 8466 1436 547 714 1823 4060 111 1391 (422) (767) (196) (28) (1413) (48) (98) (76) (4) (226) (1639) Cardoba 616 849 2791 6919 1025 377 602 2192 2724 1134 1039 (1268) (1446) (294) (43) (3051) (112) (164) (152) (7) 34 (3486) Cudinaamarca 518 760 2098 5175 926 481 635 1662 2524 1095 946 (1313) (4310) (695) (103) (6421) (100) (384) (370) (20) (874) (7295) Choco 248 714 1795 6621 763 89 253 1550 4180 410 656 (307) (359) (91) (14) (771) (207) (69) (58) (3) (337) (1108) Huila 750 867 2464 6066 1095 477 731 1404 3450 1081 1093 (664) (1681) (281) (46) (2672) (43) (166) (202) (8) (419) (3091) La Guajira 912 1592 2798 7800 1860 486 781 1732 3500 1209 1726 (99) (294) (121) (10) (524) (18) (53) (64) (1) (136) (660) Hagdalena 710 1100 2518 6362 1281 445 841 1651 3896 1268 1279 (816) (1130) (362) (44) (2352) (50) (127) (187) (9) (373) (2725) Meta 750 1070 2730 6863 1380 655 874 1563 4100 1236 1360 (307) (906) (246) (26) (1485) (20) (103) (106) (5) (234) (1719) Narilo 390 513 1937 5606 674 206 370 1548 3434 636 667 (930) (2907) (349) (59) (4245) (173) (561) (184) (20) (938) (5183) Norte de 523 881 2390 6373 1076 316 603 1591 2915 1056 1073 Santander (994) (2371) (481) (72) (3918) (71) (295) (272) (22) (660) (4575) Quindio 802 1066 2276 6604 1402 347 510 1424 2982 983 1337 (284) (1216) (370) (49) (1919) (33) (162) (140) (18) (353) (2272) Risaralda 673 1083 2576 6572 1423 484 828 1376 389 1140 1372 (396) (1840) (505) (66) (2807) (30) (329) (230) (25) (614) (3421) Santander 472 927 2706 6517 1184 304 630 1429 3488 993 1151 (1580) (4206) (891) (159) (6836) (180) (643) (585) (41) (1449) (8285) Sucre 569 1076 3109 5228 1158 481 588 1449 3243 890 1121 (654) (615) (175) (23) (1467) (70) (81) (78) (4) (233) (17W0) Tolima 753 962 2489 7870 1226 401 597 1425 3058 Iryl 12n" (1143) (2894) (579) (86) (4702) (73) (247) (.i60) (16) (696) (5393) Valle 666 1162 2569 8502 1685 479 819 1660 3737 1277 1608 (1755) (8698) (,012) (540) (14005) (216) (1530) (1387) (139) (3272) (17277) TOTAL 634 1027 2670 7806 1546 398 713 1.681 3504 1232 148S (20902) (66230) (21504) (4682) (113318) (2417) (11033) (10926) (1252) (25G28) (13894.6) Regional 649. 997. 2492. 6840. 406. 653 1597 3407 Hean Income Variance 33,339 67,511 126,103 1,389,357 20,987 30,23 3,025 217,184 Coeffi- .288 .253 .133 .364 .244 .116 .133 clout of Variation Re'fonAl 6.43 6.87 7,81 8.82 5.97 6,45 7.37 8.12 Klan Lol,ar.fitn ofIco Varinnac .115 .0708 .0197 .0233 .179 .0850 10118 .0183 of L u'1 ihin of Incumt, [Source: Fields and Schultz (1977, Table 2)] -58- As we have seen, Analysis of Variance (ANOVA) is a convenient way of systematizing the effects of a large number of categorical variables such as those given in Table 12. ANOVAs were performed separately for employees and employers.1 For brevity, only the em- ployee results are reviewed here. The working sample consisted of every fifth individual in the labor force, totaling 16,695 employees.2 The statistical tests reported below were performed on males in the labor force. The dependent variable was the natural logarithm of monthly income in pesos. Of those in the labor force, persons without incomes in the survey month and the unemployed were attributed one peso per month in order to include them in the log variance calculation. The explanatory categories were education, age, and place of residence. Four educational categories were distinguished: none, primary (some or all), secondary (some or all), and higher (some or all). There were seven age categories: 10-19, 20-24, 25-29, 30-34, 35-44, 45-54, 55 and over. Two place of residence variables were analyzed. One is rural/urban. The other is depart- ment of residence (23 in number). 'To determine income, the Census asked: "What was your income in pesos last month?" Thus, labor earnings and non-labor incomes cannot be distinguished. However, as a partial control for receipt of labor income versus non-labor income, three categories of income recipients were distinguished: (i) "employees," consisting of wage laborers, salaried employees, and day workers; (ii) "employers," consisting of employers and the self-employed; and (iii) a residual category of domestic servants and unpaid family workers. Employees' income include mostly labor earnings. For employers, though, the income reported in the Census includes payments for cooperating factors of production (such as land and physical capital) as well as returns to their labors and entrepreneurial talents. 2This was not just to reduce computational costs. The full sample was so large that for some problems the storage requirements for the analysis of variance manipulations exceeded the core capacity of the Yale computer. -59- Some illustrative results are reproduced here in Table 13. The eta statistics give the simple correlation coefficients without holding constant for the effect of other variables. In the simple correlations, education is most closely related to income, followed by rural/urban, and then department and age with virtually equal correlations. While the unstandardized correlations are of some interest, it is also desir- able to look at the marginal effects once the other factors are taken into account. The main effects in the ANOVA are marginal effects holding other variables constant. By conventional statistical standards, they are all highly significant at confidence levels in excess of .999, but this should be interpreted cautiously. Given the very large sample size virtually any basis for grouping the data according to personal, demographic, economic, social, or geographic information would reduce the standard error of estimate sufficiently to satisfy the usual F test for statistical significance. Therefore, the F test has little power in discriminating which of the explanatory variables is relatively most important. Other statistical procedures are available. There are two satisfactory ways of interpreting the relative importance of the various effects when other effects are held constant. First, there is the proportion of the variance in the logarithms of income directly explained by each set of explanatory categories. Second, the marginal F ratio deflates the explained variance by the number of categories considered and formally expresses the resulting reduction in standard error of estimate as a ratio to that anticipated TABLE 13. ANALYSIS OF VARIANCE: MAIN EFFECTS NATIONAL DATA FROM COLOMBIAN CENSUS, 1973 MALE EMPLOYEES Geographic Distinction (1) (2) (3) Zero Departments and Order Rural/Urban Departments Rural/Urban Correla- Proportion Proportion Proportion Main tion of Variance F Ratio of Variance F Ratio of Variance F Ratio, Effects (Eta) Explained Marginal Explained Marginal Explained Marginal Education Level (4) .48 .129 1131 .164 1388 .120 1038 Age Group (7) .31 .064 .282 .071 299 .064 278 Rural/Urban 0 (2) .37 .031 805 .016 404 Departments (23) .32 .043 49 .028 33 Covariance .115 -- .074 -- .140 Total Explained .339 893 .352 288 .367 299 Logarithm of Income Mean 6.52 Variance 1.54 Sample Size 16542 (Number of explanatory categories in parentheses) Note: All effects statistically significant at .001 level. [Source: Fields and Schultz (1977, Table 6A)] -61- from a random set of categories in a normally distributed population. For employees, education provides the most information in predicting personal incomes, in the sense of explaining between 12 and 16 percent of log variance. Its statistical significance is also greatest with F's in excess of 1000. Next in importance according to the F criterion is the one-way rural/urban distinction with an F of 400 to 800, accounting for 1.6 to 3.1 percent of the log variance. The seven age categories receive an F of around 300 and account for six or seven percent of the log variance in incomes. Lowest in explanatory power is region. The full 23 departments explain 4.3 percent of the log variance, with an F of 50. These values are much lower than the proportion of variance explained and F ratios found for education and age. In all, slightly more than one-third of the variance of the logarithm of income among employees is explained by these sets of categories. (For employers, the specific numerical values differ but the same qualitative results hold.) The statistical results from analysis of variance, along with other findings from multiple regressions not reported here, tell a consistent story. Even in a country such as Colombia where inter- regional income differentials are ,as large as three or four to one, the importance of regional effects in explaining income is sur- prisingly small. There is much more income variation within depart- ments or within urban and rural areas than across them. If you wanted to predict an individual's income and could ask only one question, you would get a much more accurate prediction by deter- mining his education than by ascertaining his age, region, or -62- residence in an urban or rural area. Knowledge of the department of residence or rural or urban location contributes relatively little to our understanding of wage structure. Note that this result is based on departments and urban or rural location nationwide and not just on the particular large city of residence,as was the conclusion from Fields (1977a) reported above. In interpreting these results, it should be remembered that national results are average statements. In particular, the con- clusion that education is relatively the most important variable and locational factors the least important may not hold everywhere. For specific groups in the population, some factors may have a greater influence than for others. The Fields-Schultz paper stratified the sample by edu- cation and age and examined the determinants of incomes within and across the various groups. The principal findings are: (1) Across education groups, region's explanatory power is greatest for the lowest educational groups, diminishing at higher levels; age gains in impor- tance as education increases. (ii) Across age groups, education be- comes increasingly important up to middle age; the regional effects are found to be small and exhibit no pronounced trend. One other feature of the Fields-Schultz study using national data deserves mention. Upon careful examination of the pattern of covariances in Table 13 and interactions in other ANOVAs, we find that the largest covariances, and the important interactions involve the rural-urban distinction. This suggests that rural and urban labor markets differ in respect to income structure as well as income level. Further evidence on differences between rural and urban income structures is presented in Section E below. -63- In summary, on .the relative importance of personal and regional effects on wage and income structure in Colombia: (1) Regional inequality is often suspected as a major contributor to overall inequality. Comparing the four major cities in Colombia, average incomes differ by some 30%. Nonetheless, less than 1% of overall inequality is found to be associated with income variation across those cities, while 99+% of inequality is due to variations within large cities. (2) Comparing the departments (i.e., states) in Colombia, the average income is four times greater in the highest income department than in the lowest income one. The average income in urban areas is three times higher than in rural areas. Yet, inter-departmental differen- tials account for just 11% of overall inequality and urban-rural differen- 1 tials for 13%. Together, they account for 19%. This means that:.nationwide, more than 80% of inequality is due to variations within geographic areas and not across them. (3) Marked interdepartmental and urban-rural income differences are found even after controlling for sex, employment status, education, and age. (4) Education and age explain a larger percentage of income in- equality than do geographic variables. (5) Overall, about one-third of income inequality is explained by .a simple linear model based on education, age, department, and rural/ urban. (6) Stratifying by education, regional effects on income inequality are found to be greatest for the least-educated; but even for them, region- al effects do not explain more than about 10% of income inequality within an educational group. -64- (7) Stratifying by age, regional effects are similarly small (less than 10%) and show no pronounced trend as one moves from younger workers to middle age and older workers. (8) Interaction effects contribute only about 4% additional explanatory power to a linear model, probably not worth the addition- al complication of a fivefold increase in number of explanatory varia- bles. Even in the main effects model, substantial covariances are found, particularly in ANOVAs where the rural/urban distinction is present. Let us now investigate the determinants of income structures for urban and rural areas separately. -65- E. Personal and Regional Effects on Income Structure: Comparison of Urban and Rural Areas Colombia's population is about evenly divided between urban and rural locations. In the last section, we saw that urban/rural loca- tion is a significant determinant of incomes in Colombia. In the simple tabulations, the income differential between urban and rural areas is three-to-one. In the linear model without interactions, the urban/rural variable exhibits an important and statistically significant effect. When interactions are allowed for, substantial covariance appears between urban/rural and other explanatory variables. This suggests that the explanatory contribution of the other inde- pendent variables (education, age, and department) may differ as between rural and urban areas. This section explores those differ- ences. The most straightforward way of testing for rural/urban differ- ences is to divide the population into two groups, rural and urban, and to examine the structure of income determinants in each. It is also desirable to distinguish between wage and salary workers on the one hand (so-called "employees") and owners and independent workers on the other (so-called "employers"). These divisions produce four samples: Sample Number Description Sample Size Sample 1 Urban employees 10,591 Sample 2 Rural employees 5,951 Sample 3 Urban employers 3,928 Sample 4 Rural employers 2,162 66- Analysis of variance results are presented in Table 14. We find: (1) The relative explanatory power of education, age, and department differs greatly between the rural and urban samples. (2) In urban areas, for both employees and employers, education and age are the principal explanatory variables; department plays a minor role. More specifically, for urban employees, of the 30.9% of the log variance explained, 17.6% is explained by education, 9.6% by age, and 1.2% by department. Likewise, for urban employers, the respective figures are 25.4% (total), 17.2% (education),4.6% (age), and 1.2% (department). (3) In vural areas, for both employees and employers, department is the principal explanatory variable; - education and age play minor roles. More specifically, for rural employees, of the 17.7% of the log variance explained, 13.8% is explained by department, 1.7% by age, and 1.6% by education. Likewise, for rural employers, the respec- tive figures are 27.3% (total), 23.7% (department), 0.7% (age), and 1.5% (education). (4) Given that education and age are important determinants of income in urban but not in rural areas and that interdepartmental differences are important in rural but not in urban areas, we would expect to find differential labor supply responses. Education apparently raises income more in urban areas. Accordingly, educated persons have little incentive to remain in rural areas, but would instead migrate to the cities in large numbers. Less-educated individuals TABLE 14. Analysis of Variance, Stratified by Urban/Rural and Employee/Employer, Main Effects: Males. National Data from Colombian Census, 1973. SAMPLE 1 SAMPLE 2 SAMPLE 3 SAMPLE 4 Urban Employees Rural Employees Urban Employers Rural Employers Proportion Proportion Proportion Proportion of Variance F Ratio, of Variance F Ratio, of Variance F Ratio, of Variance F Ratio, Main Effects Explained Marginal Explained Marginal Explained Marginal Explained Marginal Education Level (4) .170 864 .016 40 .172 443 .015 27 Age Group (7) .096 245 .017 20 .046 59 .007 6 Department (23) .012 8.0 .138 45 .012 4.2 .237 57 Covariance .032 .005 .025 .014 Total Explained .309 153 .177 41 .254 64 .273 47 Logarithm ot Income Mean 6.88 5.90 7.20 5.93 Variance 1.45 1.03 1.97 2.45 Sample Size 10,591 5,951 3,928 2,162 (Number of Explanatory Categories in Parentheses) Note: All effects statistically significant at .001 level. -68- also have incentive to migrate from low-income to high-income departments. Insofar as the high-income departments are comprised of cities, that migration would be of a rural-to-urban character. Several research studies on the causes and consequences of migration in Colombia are in progress. (5) Comparing employees and employers in rural areas, the income structures are rather different. Although the two groups have similar means (5.90 and 5.93 respectively), the variance is much greater for employers (2.45) than for employees (1.03). This larger variance is accounted for, at least in part, by greater interdepartmental variation.1 This suggests that the labor market for landless rural workers (farm laborers and non-agricultural employees) is more balanced geographically than is the distribution of farming and ranching opportunities. Presumably, these differences are associated with the size distribution of landholdings, ecological zones, and specific cropping patterns, but these speculations remain to be established.2 Compare the relative explanatory power of department for the two groups. 2A review of the literature on rural income distribution in Colombia turned up many tabulations but no suitably disaggregated data on the correlates of rural wage structure. The literature reports that average income increases with the size of the landholding, some regions are richer and experience more rapid growth than others, and returns to education are lower in rural areas than in urban areas. The interested reader is referred to the book by Berry (forthcoming) and the studies cited by him. -69- IV. Evidence on the Determinants of Intersectoral Wage Structure A. Review of Previous Results In the last section, we saw that certain characteristics of the individual and of the industry in which he works contribute to an understanding of the determinants of individual incomes. This section carries out a comparable analysis for the determinants of intersector- al wage structure, where the dependent variable is now the average wage in an industry. Past research studies have reported that certain characteristics of the industry are associated with average wages. Among these are value added per worker,capital intensity, size distribution of firms, and foreign ownership. The best-known studies reporting evidence on these variables are those of Nelson, Schultz, and Slighton (1971) and Berry and Urrutia (1976). Readers interested in familiarizing them- selves with the literature should consult those two books and the references cited therein. The research studies just mentioned were univariate, i.e., they presented average wages in large and small firms or in domestic and foreign firms. The evidence from these studies does not directly test the hypotheses of Section II since the hypotheses are ceteris paribus, not mutatis mutandis. In previous research, Nohra de Marulanda and I (1976) used a multi- variate framework to examine the relationship between a number of industry characteristics and average rEnu.eration in two-digit industries in Colombian manufacturing.1 We found the following variables to be positively and significantly related to average re.mumeration (AVGREM): 1The basic data source is the annual Industrial Survey. The survey covers all types of firms though it tends to undersample small ones. -70- Productivity (PROD), measured by value added per worker; Capital-intensity of production (CAPINT), measured by installed electrical capacity per worker; Foreign investment (FOREIGN), taken as a percentage of total investment; Importance of large firms (SIZE), which is the percentage of firms in the sector with more than 50 workers; and Proportion of white collar workers (WHTCOL), taken as a percentage of the total labor force. The results of a regression using the first four of these variables are '1 reported in Column (1) of Table 15. Each variable exhibits the hypothe- sized positive effect and all are statistically significant at the usual levels. Thus, the hypotheses on the ceteris paribus effects of PROD, CAPINT, FOREIGN, AND SIZE are confirmed by the Fields and Marulanda results. In the following section, some new results are presented. In Fields and Marulanda (1976), we presented regression results with and without WHTCOL. That variable is excluded here because of possible simultaneity problems when workers' job-relevant character- istics are also present in a regression. TABLE 15. MULTIPLE REGRESSIONS IN MANUFACTURING, INTERINDUSTRY RESULTS 1967/68 Dependent Variable: Blue-Collar Dependent Variable: Average Remuneration (AVGREM) Remuneration (REMBLU Method of Estimation OLS OLS OLS OLS OLS OLS 2SLS OLS Independent Variable (1) (2) (3) (4) (5) *(6) (7) (8) PROD .06919 .07220 .06964 .06697 .06405 .06198 .050 (.01462) (.01217) (.01197) (.01163) (.01127) (.01130) (.010) CAPINT 233.03 316.27 291.12 312.94 336.69 345.67 152.46 (93.93). (85.19) (82.08) (78.84) (75.24) (75.40) (68.53) SIZE 171.68 80.54 108.16 117.47 112.88 114.30 106.43 (57.67) (56.29) (49.96) (48.96) (48.77) (48.86) (44.43) FOREIGN 187.07 66.00 74.53 79.72 (90.03) (79.70) (79.60) (79.29) PROPHAILE 8965.31 3142.46 2673.75 (9474.53) (2764.94) (2739.48) AVGEDUC 1312.87 913.56 1059.19 1030.29 1173.52 1170.58 309.75 (1247.72) (357.81) (331.15) (329.27) (296.96) (297.54) (270.53) AVGEXP 531.52 -201.90 (596.77) (192.24) CONSTANT 10,327.13 -9,315.04 7,444.52 1703.78 3859.94 3271.85 3352.13 7580.97 2 R .91 .13 .95 .95 .95 .94 * .90 2 *Not reported because of difficulty in interpreting R in two stage least squares regressions. -72- B. New Evidence on Intersectoral Wage Structure in Colombian Manufacturing This study adds other variables to those included-previously. Earlier studies were limited to the characteristics of the firms in the industry. Three variables measuring characteristics of the workers in the industry are added here: the average education (AVGEDUC) and average experience (AVGEXP) of that sector's labor force and the propor- tion of workers who are male (PROPMALE). These variables---the same as used in the regressions of Table 4--were calculated from the micro- economic PRESFAM data. The hypotheses of Section II and the micro- economic results of Section III lead us to expect that AVGREM is posi- tively related to each of these variables. Thus, the maintained linear model of this section is: AVGREM - a + 81 PROD + 02 CAPINT + 8 SIZE + 04 FOREIGN + 85 PROPMALE + 86 AVGEDUC + 87 AVGEXP + e. To explore the relationship between AVGREM and workers' character- istics, AVGREM was regressed on PROPMALE, AVGEDUC, and AVGEXP. The results are reported in Column (2) of Table 15. Although each of the coefficients has the anticipated positive sign, these effects are not statistically significant. Furthermore, the overall explanatory power of the regression is low, especially as compared with the regression based on firms' characteristics. These findings are interpreted at some length below. Other regressions in Table 15 give the results of combining the characteristics of workers and firms. Column (3) includes all seven in- dependent variables. Comparing these findings with the earlier ones which treated the two sets of characteristics separately, we find: -73= (1) PROD retains a similar effect to that observed in the simpler regression. The estimated coefficient is about .07. If this is inter- preted as a structural estimate, it would mean that as value added per worker increases by one unit, manufacturing firms share about 7% of the increment with their workers. This estimate is lower than previously- observed figures (cf. Heady (1976)), but still significant, both statisti- cally and economically. (2) CAPINT continues to have a strong effect on AVGREM, as hypothe- sized. (3) AVGEDUC has a highly significant effect on AVGREK, as expected. (4) The estimated effects of SIZE and FOREIGN on AVGREM are both substantially reduced in the presence of AVGEDUC. (5) Findings (3) and (4) suggest that models like Regression (1) which include only firms' characteristics and,neglect workers' charac- teristics suffer a classic omitted variables problem in which a variable related to both the dependent variable and to an independent variable is left out of the regression. The effect of such an exclusion is to mis- takenly give part of the "credit" to some other variable, and thus give the appearance of a larger effect than is in fact the case.1 Here, it would appear that AVGEDUC is positively related to AVGREM, SIZE, and FOREIGN, and hence the regression coefficients on SIZE and FOREIGN are biased upward when AVGEDUC is excluded. This finding suggests that although firms that are larger or more reliant on foreign investment pay higher wages, they are also able to attract better workers. It 1 An analogous situation in labor economics arises when individual incomes are regressed on education but not ability. Because more able persons on average receive more education, and more able persons also tend to earn More, the estimated coefficient on the education variable is higher when ability is excluded from the repression than when it is present. -74- is unclear, therefore, whether largeness or foreign ownership make an additional independent contribution to the understanding of inter- sectoral wage structure. We analyze their role further below. (6) As before, the effects of PROPMALE and AVGEXP are not signi- ficantly different-from zero.1,2 (7) In light of the insignificance of certain of the independent variables in Regression (3), variables were eliminated one by one. The results are reported in Columns (4)-(6). Four of the variables--PROD, CAPINT, SIZE, and AVGEDUC--have strong independent effects on manu- facturing wages and together account for 94% of the variance in AVGREM. The independent effects of three other variables---FOREIGN, PROPMALE, and AVGEXP---are not confirmed in the anklysis. 1The existence of labor market discrimination against women in Colombia is not necessarily denied by the absence of a statistically significant relationship between the sex composition of the labor force in a subsector (PROPMALE) and that sector's average wage (AVGREM). It appears from casual observation that the predominant difference be- tween men and women in the Colombian labor force is that women are disproportionately relegated to low-paying occupations (e.g., domestic services). In this case, PROPMALE would show an effect in an inter-sectoral regression to the extent that the occupational distri- bution differs between industries. A more .general test of differential labor market rewards to men and women would compare individual earnings structures by sex. Indeed, the microeconomic evidence of Section III and other research studies leave little doubt that being a woman in Colombia is associated with lower wages, even after controlling for differences between men and women in other characteristics. 2The problem with the experience variable (AVGEXP) may be that the use of average experience is inappropriate, since earnings functions like those in Section III have shown that effect of experience is non-- linear. Ideally, one would like to include more sophisticated refine- ments of the experience variable. It is not possible to do so in this data set without encountering serious degree-of-freedom diffi- culties with only 20 observations. -75- (8) One other variable was introduced into the interindustry analysis. It is sometimes thought that more monopolistic industries pay higher wages. Simple tabular evidence from Colombia supports this view (Misas, 1976). However, multivariate analysis suggests otherwise. In a cross section of 88 three-digit industries in Colombian manufacturing for 1968, taking average remuneration per workers as the dependent variable, the regression results were as follows: REMWRKR - 1937.7 + .139 VAWRKR + 688.6 CONC1 (.013) (633.3) + 667.2 CONC2 + 51.3 CONC3 + 785.5 PROD1 + 1289.7 PROD2 (520.6) (510.2) (596.9) (607.8) 2 --2 + 7.99 PRINDEX - 943.1 PROF, R - .728, R - .701 (2.84) (156.8) where REMWRKR = Average remuneration per worker VAWRKR = Value added per worker CONC1 - 75-100% of output was produced by 3 firms CONC2 - 50-75% of output was produced by 4 firms CONC3 = 25-50% of output was produced was produced by 4 firms PROD1 = Consumer goods industry PROD2 = Intermediate goods industry PRINDEX = Rate of increase of prices between 1958 and 1968 PROF - Gross profit per salary unit Both t and F tests reveal that in the presence of value added per worker, the concentration ratio in the industry contributes no significant additional explanatory power. In summary, these results sustain the general finding from earlier studies that characteristics of firms play an important role in deter- mining average remuneration in Colombian manufacturing industries. As in the paper by Fields and Marulanda(1976), several of the firms' charac- teristics exhibit statistically significant effects even in the presence -76- of other firms' characteristics. However, as compared with the Fields- Marulanda results, the findings here differ in three important respects: (i) The earlier results are strengthened, since firms' characteristics are found to matter even in the presence of workers' characteristics; (ii) Not all of the firm variables which were statistically significant previously retain their statistical significance now, in particular FOREIGN; and (iii) The educational attainment of a sector's work force is found to contribute significantly to the explanatory power of the overall model and also to reduce the apparent importance of some of the other variables. In the appendix to this section, some additional econometric points are considered. Readers not interested in these issues can skip direct- ly to Section V with no loss of continuity. -77- APPENDIX TO SECTION IV: ECONOMETRIC ISSUES IN ANALYZING INTERSECTORAL WAGE STRUCTURE Before accepting conclusions (l)-(4) based on Regressions (3)-(6), we should establish that the results are robust to certain possible econometric difficulties. These are: multicollinearity, simultaneous equations bias, and specification error in the dependent variable. Let us take these up in turn. A. Multicollinearity. A priori, there are grounds for concern that many of the indepen- dent variables are multicollinear, perhaps highly so. For example, large foreign firms may use highly capital-intensive techniques and highly-skilled workers, all of which lead to high value-added per worker. If the multicollinearity were acute, it would be difficult if not impossible to distinguish the effect of one of the collinear variables from that of another. To test the severity of multicollinearity, a number of tests may be performed. First, examining the matrix of simple correlation co- efficients, many of the variables were found to be interrelated but only one correlation coefficient (rCAPINT, PROD= +0.72) was greater than 0.5. Making use of the widely-accepted rule of thumb that collin- earity does not become severe until the correlation coefficient approaches or exceeds 0.9, (unless a linear combination of several variables happens to be collinear with another variable), it seems justified to conclude that multicollinearity is not a serious problem in these regressions. To confirm this result, the multicollinearity tests suggested by Farrar and Glauber (1966) and Johnston (1972) were also carried out. This entails (a) computing partial correlation coefficients between =48- the dependent and the independent variables, and (b) regressing each of the seven independent variables on the remaining six. If multi- collinearity were strong, some of the partial correlation coefficients would remain high and at least one of the independent variables would exhibit a statistically significant correlation with one or more of the others. The results confirm the earlier finding that PROD and CAPINT are collinear but not so much so as to cause serious difficulty. Finally, we may examine the pattern of regression coefficients and standard errors in regressions (3)-(6). Each successive regression eliminates a yariable which was found to be statistically insignificant in the previous regression. If an excluded variable were highly collineaf with one of the included variables, its exclusion would reduce substantially the standard error of estimate of the remaining regression coefficients. No such diminution is observed. Based on these tests, we may conclude that the maintained linear model AVGREM - a + 81 PROD + 82 CAPINT + 83 SIZE + 8 FOREIGN + 85 PROPMALE + 86 AVGEDUC + 8 AVGEXP + e is not subject to strong multicollinearity. B. Simultaneous Equations Bias. Up to now, in using ordinary least squares estimation, we have neglected any possible structure of which the equation AVGREM - a+ 81 PROD + 82 CAPINT +83 SIZE + 0 AVGEDUC + c (i) is a part. There is, however, an obvious definitional relationship involving wages and value added, namely, -79- VALUE ADDED - WAGES + VALUE OF OTHER INPUTS. Dividing through by L, we have, PROD - AVGREK + OTHER. (ii) In ignoring (ii), we impart inconsistency into estimates of equation (i), i.e., the estimated regression coefficients are asymptotically biased. Let us gauge the ssvrity of this bias. Formally, the model given by (1) and (ii) is comparable to the simple Keynesian.system C a + OY +. (iii) Y *C + I (iv) and so the econometric solution to the definitional bias is also comparable. Following Goldberger (1964, pp. 288-293), the first equation may be consistently estimated by Instrumental Variables methods, or the equivalent in an exactly identified equation, Two Stage Least Squares. The results of the Two Stage Least Squares estimation appear in Column (7) of Table 15. We observe a close similarity between the Two Stage Least Square esti- mates and the Ordinary Least Squares estimates. Hence, we may conclude that the simultaneous equations bias caused by neglect of equation (ii) is not serious enough to warrant re-estimation of the previous equations or to cast doubt on the qualitative results reported above. Equation (ii) is not the only structural relationship to have been ig- nored. Another possible source of simultaneous equations bias arises due to the endogeneity of capital intensity. This is because, all other things equal, higher wages would presumably induce firms to substitute capital for labor. We would therefore observe a positive correlation between wage and capital-labor ratio, with part of the causality going from the former to the latter. This poses no problem, given the limited 1In the Keynesian system (iii) and (iv), the consistent procedure is to use I as an instrument for Y in (iii). In our system (i) and (ii), OTHER is used as an instrument for PROD in (i). -80- objective in this paper of establishing correlations rather than causal structure. In future research i.n which causal structures are sought, the most straightforward way to recognize the endogeneity of capital intensity is to treat it as a jointly-determined variable which is a function of the price of capital relative to the price of labor in that particular Industry.1 Unfortunately, secto4-specific data on capital prices are unavailable, so this problem remains insoluble. In any case, there is ample precedent in the literature for neglecting 2 the endogeneity of factor intensity in studies of wage structure. C. Misspecification of the Dependent Variable. The dependent variable in this and other studies is the average remuneration of all workers in the sector in question. Since occupa- tional distribution in Colombia varies substantially from one sector to the next, it is possible that the intersectoral wage patterns are due to systematic differences in the composition of employment rather than to higher wages for comparable workers. The three worker character- aic variables (PROPMALE, AVGEDUC, and AVGEXP) were included to try to achieve a partial standardization. As a further control, we may also see if the model does equally well in explaining wages among just blue-collar workers (obreros). This test and the results were as follows. Average blue-collar remuneration REMBLUE was run on the various combina- tions of independent variables contained in Table 15. An illustrative result is reported in Column (8). The results of columns (6) and (8) 1Logically, it would follow that labor intensity should also be treated as endogenous, which it never is. 2 See, for example, Rosen (1969) or any of the studies cited by Perlman (1969). are quite similar in th.eir overall explanatory power and in the statis- tical significance of the variables. Note though that each regression coefficient is lower for REMBLUE than for AVGREM, which would be expected from the fact that blue-collar workers (obreros) are paid less than white- collar workers (empleados). One other difference is that AVGEDUC is not statistically significant in the REMBLUE regression. This may be be- cause AVGEDUC is the average education of all workers in the sector and not just the blue-collar workers, to which the dependent variable per- rains, or because education is a critical factor determining whether an individual is engaged in a white or blue-collar occupation. All.in all.,.then, .these results generally sustain the earlier- conclusions. -82- V. Synthesizing Microeconomic and Interindustry Results on Wage Structure in Colombian Manufacturing Sections III.B and IV of this paper analyzed the determinants of wages in Colombian manufacturing., The analysis was conducted at two levels: individual and interindustry. The results generally support a number of hypotheses about the influence of the characteristics of work- ers and firms on wages. The two types of analyses suggest that different explanatory variables play a role in wage determination at the two levels. In the microeconomic analysis, the variables found to influence income were: education, experience, and sex of the individual and the "productivity" and importance of foreign investment in the industry in which the individ- ual is employed. At the industry level, average wages were found to be related to "productivity," capital intensity and size distribution of the firms in the industry and the average education of the sector's labor force. At both levels, education of the individual and the "productivity" of the workers in his industry are found to contribute to higher incomes. However, the effects of the other variables are not consistent across the two levels of analysis. The lack of congruence between the two lists requires further examination. Among the possibilities are the following: (1) The two sets of results are based on different data sets. The microeconomic data are drawn from a household survey while the industry data are from surveys of manufacturers. The microeconomic data were gathered from the four major cities while the industry data were collected -83- nationwide. All of the industry data are based on large numbers of workers; for the microeconomic data, some of the sample sizes in parti- cular industries are quite small. (2) The two sets of results are based on different functional forms. The interindustry results are based on a conventional linear model while the microeconomic results are log linear. Even if everything else were the same, when the dependent variable is changed, the explanatory power of particular independent variables change too. (3) Non-linearities in the effects of the several variables may be distorting the results. If there are important intercorrelations among the explanatory variables, and if these correlations cause the effect of one variable to depend on the level of another, non-interactive models such as those used in this paper are misspecified. It is an open'question, subject to empirical research, whether more sophisticated specifications would resolve the disparities between the two sets of results. (4) Perhaps most importantly, the microeconomic and the interindustry analyses deal with different phenomena. The microeconomic analyses seek to explain the individual's own income. Interindustry comparisons seek to explain averages across groups of individuals. It is not obvious why the dependent variables in the two types of studies should be related in the same way to the same independent variables. Here is some information bearing on these points: (1) The mean industry incomes from the urban household data are quite different from the industry means from the nationwide survey of manufacturers. The correlation coefficient is just +0.28. We cannot tell whether the problem is measurement error, small sample size, or -84- TABLE 1 6. INTERSECTORAL REGRESSIONS WITH VARIOUS SPECIFICATIONS OF DEPENDENT VARIABLES, COLOMBIAN MANUFACTURING, 1967. Independent Variables Dependent Variable REM SUBSECINC LOGREM LOGSUBSECINC PROD .069 .0018 .34xl0- .37xlO0- (.015) (.0012) (.09X10 )5 (llxlO_5 CAPINT 233.0 -6.78 .0029 -.0111 (94.0) (7.62) (.0062) (.0073) FOREIGN 187.0 25.3 .0129 .02E9 (90.0) (7.3) (.0059) (.0070) SIZE 171.7 -1.09 .0111 .56xlO130 (57.7) (4.68) (.0038-): (4.5xl0O ) CONSTANT 10,327 506 9.36 5.76 2 .907 .458 .812 .624 (R )(.883) (.314) (.762) (.524) Mean of depend- ent variable 20,060 670 9.84 6.07 Definitions and sources of dependent variables: REHe: Basic salary expenditure plus fringe benefits divided by number of workers as taken from employers' records, nationwide,annual. SUBSECINC: Wage and self-employment income as reported in household survey, aggregated up to two-digit industry level, foul- urban areas, in tens of pesos per three months. LOGREM: Logarithm of REM. LOGSUBSECINC: Logarithm of SUBSECINC. difference between urban and nationwide patterns. We can, however, explore the effect of the differences. One way of examining the difference between the two data sets is to perform two regressions. In one, the dependent variable is the average sector income as reported in the survey of manufacturers (REH), while the other uses the industry average as calculated from the urban house- hold survey (SGESECINC). The independent variables are the same in the two cases. The results appear in the first two columns of Table 16. The much higher proportion of variance explained in the first regression than in the second suggests that the data from the survey of manufacturers may be better for industry averages. But in the final analysis, most of us are probably interested in the incomes of families and individuals, not industries. For the personal distribution of income, the household data not only are better---they are all that will do. (2) At least two differences between the interindustry and micro- economic regressions may be due to different functional forms. One is the effect of capital intensity. Capital intensity in the industry was associated with average income in the interindustry analysis but not in the individual analysis. The individual analysis used a logarithmic form and the interindustry analysis did not. Compare now the first regression in Table 16 (linear model) with the third (log Unear model). In adopting the log linear form, capital intensity loses its statistically significant association with average income. A second difference appears in the effect of value added per worker ("productivity"). Comparing the second and fourth regressions, its effect appears to be greater in the logarithmic than in the linear specification. Additional experimentation with alternative functional forms is called for. -86- (3) Some of the independent variables are closely-related to one another. Here is the matrix of correlation coefficients: PROD FOREIGN CAPINT SIZE PROD 1.000 FOREIGN -.134 1.000 CAPINT +.726 +.042 1.000 SIZE +.397 +.047 +.418 1.000 With correlation coefficients of this order, interactive effects are distinct possibilities. Once again, the need arises for further explora- tion into alternative functional forms. (4) Research into aggregation issues may be fruitful. Part of the answer is to be found in the nature of the groupings This may be illustrated with reference to Figures 1-4. Figure 1 presents a frequency distribution of incomes for everyone in the urban economy (based on the PRESFAM data). Individuals are scattered over the entire range of the income distribution. In Figure 1, however, the worker's characteristics and the industry in which he is employed do not appear. Figure 2 illustrates a pattern of income differentials based on the grouping of individuals by education. On average, income increases with education. More importantly, there is relatively little overlap between incomes of one educational group and those of the next. Conse- quently, at the individual level, education would be a very important determinant of income. In Figure 3, individuals are grouped by industry. Adjacent industries (for example , transportation and foodstuffs) overlap so much in income distribution that it is hard to distinguish one industry from another. Even for the extr-me industries ~'~1* nrn da ff-. -- -l-- r11 1 4-i- -r 1 t-11 -I CI t 1-g I a -- r- - - - - - -- - 11i 1 --Tl1 1 1 -rri- -L T 4 - m li MI 4r'~1 1i :P 1-3 1 41T iFf-;--tr I 1 1T rl4L...LV -88- (chemicals and clothing), the overlap is substantial. In other words, Figure 3 shows that within-industry income variation is greater than between - industry income variation. From the patterns in Figures.2 and 3, it would follow that personal characteristics of an individual would be more important than industry characteristics in determining the individual's own income. This is what the data in Section III show:. However, when we move to industry averages, we are comparing points like those indicated by asterisks in Figure 4, In so doing, within-industry inequality disappears, and our concern is only with the between-industry variability. The positions of the frequency distributions like those in Figure 3 may depend in important ways on the characteristics of the industry such as value added per worker, capital intensity, and so on. In contrast, the characteristics of the workers in the industry may be relatively unimportant in determining industry averages. The findings presented in Section IV are in accordance with this interpretation. These hypothesized relationships are subject to.further examina- tion and verification. In future empirical research, I hope to test these speculations, using both the data analyzed in this paper for 1967-68 and more recent data from the Population Census and Surveys of Manufacturers of 1973. -89- VI. Conclusion It would be redundant to try to summarize specific results, since I have already done so at the end of each subsection. Rather, let me con- clude with some sweeping observations on the results and on future re- search efforts in this area. A. Some General Conclusions 1. Income distribution information at the household level in Colombia is basically sound. This study used microeconomic data from several differ- ent household surveys and one census conducted-by the central government and a private university. Often, the questions were very simple, e.g., the Census determined income by asking: "What was your income in pesos last month?" Garbage data would result in weak results that cannot be sustained across data sets. But these results are strong. That the Colombian models do so well compared to similar models estimated for other countries is itself evidence of data strength. Equally important is the fact that several central results are encountered every time tests are repeated across alternative data sets. Of course, no data set can stand up to improper processing or comparisons with other, less valid ones. The literature on Colombia is not exempt from these difficulties. 2. Statistical models and methods from developed countries do apply to less developed countries. Two examples from this paper are human capital type earnings functions and models of interindustry wage structure. The results for the human capital type earnings function are remarkably consis- tent: each year of education adds about 15% to income; income increases -90- with experience at a decreasing rate; between one-third and one-half of income inequality (as measured by the log variance) can be explained by a simple model of this sort. Repeated tests were not possible in the study of interindustry wage structure, but the one set of results obtained there also does very well in comparison with similar models for countries like the United States. 3. Some seeming relationships do not hold up or are modified upon rigorous test on Colombian data. Here are five examples: a. Migrants to cities are widely thought to be at a disadvantage economically as compared with long term residents. If this is taken as a ceteris paribus proposition, the evidence from Colombian cities does not support it. b. Income and occupational position in Colombia are often thought to be determined by family background rather than the individual's own characteristics. The evidence suggests otherwise: in four major cities, individuals with given characteristics who come from advantaged backgrounds do not do significantly better in the labor market than those from dis- advantaged backgrounds. c. Much weight has been given to regional income inequality in Colombia, and it is often claimed that the Colombian labor market is segmented geographically. Yet, upon careful examination, it is found that the great bulk of inequality in Colombia is within geographic areas (departments) rather than between one department and another. The same is true of a more stringent dualistic model: some 90% of inequality is within rural and urban areas rather than between them. -91- d. It is sometimes maintained that the forces of competition pro- duce a single unified structure of wages which pertains to the entire nation. This is disproven by the finding that in urban areas, education and age are much more important than department as determinants of in- equality, whereas in rural areas, nearly all the inequality that can be explained is explained by department. e. Some would argue that differences in wages paid by different types of industries can be explained by variations in the characteristics of the industries' labor forces. However, evidence at both the micro- economic and intersectoral level shows that workers in some industries re- ceive higher wages than workers in other industries, and that these wage differentials are associated with the characteristics of the firms in the industry, even after holding constant for differences in workers' educa- tion and experience. 4. The microeconometric approach to wage structure gives a promising start toward understanding the Colombian labor market, which in turn contri- butes importantly to understanding the overall distribution of economic well-being. The reader should be careful not to let his concerns ovcr what is not known cause him to lose sight of what has been learned. B. Agenda for the Future An explicit task of this paper is to chart directions for future investigations. Here are some areas which I think merit high priority: 1. Formal modeling of the determinants of wage structure. As the first sections of the paper made clear, it is quite important to have more rigorous -92- theoretical models of how wage differentials arise and are maintained. We need to understand much better than we now do how labor supply and demand interact to produce wages and employment, taking full account of labor force heterogeneity and the rich complexity of LDCs' labor market structures. 2.. More refined econometric models using variables included in this paper. Much of the difficulty in research of this kind comes from lack of theoretical guidance on which variables to use to explain wage structures and how specifically to include them. I have tried to justify each econo- metric procedure insofar as the economic theory of wage differentials permits. Undoubtedly, new improved specifications will emerge once the theoretical underpinnings are clear. 3. More complete econometric modeling using other variables. Some readers may have detected certain omissions, for example, occupation. This was intentional. To have included occupation properly would have required more time and space thar. this project specified, though it would have been easy to have included it improperly. I believe it is better to do nothing on a problem than to do something wrong, since read- ers will not then be misled into thinking they have learned something when they haven't. 4. Explaining underlying relationships. Finding that a variable has a statistical association with income does not tell us why it has that relationship. In a multiple regression framework, a statisti- cally significant effect of education on income does not prove that the human capital theory is correct, nor does a statistically signi.,icant -93- effect of an industry variable prove labor market segmentation. We must look into fundamental causes and identify what are the underlying economic forces that make the proximate relationships hold. 5.. Better and more recent data on personal and industry effects on incomes in Colombia. The section of this study that examined personal and industry effects was based on data from 1967 and 1968. It would be better in future research to use the 1973 Population Census and the Census of Manufacturing for the same year. Better still would be the use of even more recent data from one of DANE's national household surveys coupled with better measures of the characteristics of the actual firms in which individuals are employed rather than their industries. 6. Research on rural labor markets and rural income determination in Colombia. This paper has shown that rural labor markets differ from urban labor markets in income structure as well as income level. While we now know that regional variables are more important than education and age in determining rural income distribution, we do not yet know why these regional variables have the effects they do. Agricultural econo- mists can play a key role in helping to establish what share of the differ- ences are due to land conditions, cropping patterns, factor endowments, and other things not considered here. 7. Research on labor market segmentation. Much effort has gone into arguing that less developed countries' labor markets are not homogeneous, and many have claimed that labor market segmentation is the key to under- standing the heterogeneity of outcomes. Without detailing my concerns, let me just record my uncertainty over the merits of these arguments and -94- my dissatisfaction with the evidence offered in support of them. Seg- mentation theory is proposed as an alternative to conventional labor market theories; I would very much like to see it subjected to rigor- ous formulation and testing. It should perhaps be obvious that these are not easy tasks with easy answers. Even to examine these issues superficially is a sub- stantial job. To do a thorough analysis of any one of them is a major undertaking. For purposes of understanding the structure of wages and labor markets in less developed countries, I believe the optimal research strategy is to take one topic and study it.well. Policy makers, of course, cannot wait and need to make decisions based on the best infor- mation available at the time on a wide range of concerns. I would suggest only that we not c8nfuse these two roles. Let us speak out forcefully when we have enough confidence in our information to know we can, present policy advice and make policy decisions when we must, and have the courage and humility to be silent and search patiently for answers otherwise. -95- BIBLIOGRAPHY Ashenfelter, 0. and Johnson, G. E., "Unionism, Relative Wages, and Labor Quality in U.S. Manufacturing Industries ," International Economic Review, October, 1972. Ayub, M., Income Inequality in A Growth-Theoretic Context: The Case of Pakistan, Unpublished Doctoral Dissertation, Yale University, 1977. Baily, M.A., Technology Choice in the Brick and Men's Leather Shoe Industries in Colombia, Final Report to Agency for International Development, Contract No. AID/otr C-1326, August 1977. Becker, G. S., Human Capital ,New York, Columbia University Press for N.B.E.R., 1964. , Human Capital and the Personal Distribution of Income, W. S. Woytinski Lecture No. 1, University of Michigan, Ann Arbor, 1967. Berg, E. J., "Wage Structure in Less Developed Countries," in Smith, A.D., ed., Wage Policy Issues in Economic Development, London, MacMillan, 1969. Berry, R. A., Agriculture in Colombian Development, forthcoming. and Urrutia, M., Income Distribution in Colombia, Yale University Press, 1976. Blaug, M., "The Correlation Between Education and Earnings: What Does it Signify?," Higher Education, 1972. , Education and the Employment Problem in Developing Countries, Geneva, International Labour Organization, 1973. , "Human Capital Theory: A Slightly Jaundiced View," Journal of Economic Literature, September, 1976. Bowles, S. and Gintis, H., "The Problem with Human Capital Theory: A Marxian Critique," American Economic Review, May, 1975. Calvo, H. and Fields, G. S., "Distribuci6n de Ingresos Urbanos Para Colombia," Informe Preparado por el XXIV Seminario de ECIEL, Lima, Agosto, 1975. Carnoy, M., Education as Cultural Imperialism, New York, David McKay Company, Inc., 1974. Cordova, Polibio, Andlisis Econom6trico de Distribuci6n de Ingresos, Departamento Administrativo Nacional de Estadistica, 1973. -96- Diaz-Alejandro, C., "Trade and the Import Control System in Colombia: Some Quantifiable Features," Economic Growth Center, Yale University Discussion Paper No. 203, May, 1974. Doeringer, P. B. and Piore, M. J., .Internal Labor Markets and Manpower Analysis, Heath, 1971. Dunlop, J. T., "The Task of Contemporary Wage Theory," in G. W. Taylor and F. C. Pierson, ed., New Concepts in Wage Determination, McGraw Hill, 1957. Farrar, D. E. and Glauber, R. R., "Multicollinearity in Regression Analysis: The Problem Revisited," Review of Economics and Statistics, 1967. Fei, J.C.H., and Ranis, G., Development of the Labor Surplus Economy, Irwin, 1964. , Ranis, G., and Kuo, S.W., "Growth and the Family Distribution of Income by Factor Components," Quarterly Journal of Economics, February, 1978. Fields, G.S., A Theory of Education and Labor Markets in Less Developed Countries, Unpublished Doctoral Dissertation, Department of Economics, University of Michigan, Ann Arbor, Mich., 1972. , "Rural-Urban Migration, Urban Unemployment and Underemployment and Job Search Activity in LDC!s," Journal of Development Economics, June, 1975. . "Education and Economic Mobility in a Less Developed Country," Economic Growth Center, Yale University, Center Discussion Paper No. 237, revised version, June, 1976. , "Decomposing LDC Inequality," Economic Growth Center, Yale University, Discussion Paper No. 263, June, 1977. (1977b). , "Income Inequality in Urban Colombia: A Decomposition Analysis," Economic Growth Center, Yale University, Discussion Paper No. 267, Sept. 1977. (1977a). and Jaramillo, H., "A Guide to the Use of Microeconomic Data Sets in Colombia," Economic Growth Center, Yale University,. Sept., 1975. . and Marulanda, N., "Intersectoral Wage Structure in Colombia," Economic Growth Center, Yale University, Discussion Paper No. 251, August, 1976. and Schultz, T.P., "Sources of Income Variation in Colombia: Personal and Regional Effects," Economic Growth Center, Yale University, Discussion Paper No. 262, June, 1977. -97- Garcia, C., Movilidad Ocupacional: Resultados de Una Encuesta, CEDE, Univer3idad de Los Andes, 1968. Goldberger, A. S., Econometric Theory, New York, John Wiley, 1964. Gonzdlez, H., "Determinantes de loa Patrones de Ingreso y Consumo para la Ciudad de BogotA," Tesis de Grado, Universidad de Los Andes, Junio, 1971. Harris, J. and Todaro, M. P., "Migration, Unemployment, and Development: A Two-Sector Analysis," American Economic Review, March, 1970. Harbison, F. H., Human Resources as the Wealth of Nations, Oxford University Press, 1973. Heady, C., "Alternative Theories of Wages in Less Developed Countries: An Empirical Test," Economic Growth Center, Yale University, Discussion Paper No. 254, October, 1976. Isaza, R. and Ortega, F., "Encuestas Urbanas de Empleo y Desempleo: Analisis y Resultados," CEDE, Monograph No. 29, 1971. Johnson, G. "The Supply and Demand for Labor by Educational Category," Southern Economic Journal, 1970. Johnston, J., Econometric Methods, New York: McGraw Hill, 1972. Kalachek, E. and Raines, F., "The Structure of Wage Differences Among Mature Male Workers," Journal of Human Resources, 1976. Kugler, B., "Influencia de la Educacion en los Ingresos del Trabajo: El Caso Colombiano," Revista de Planeaci6n y Desarrollo, Enero, 1975. Kuznets, S., "Quantitative Aspects of the Economic Growth of Nations: VIII. Distribution of Income by Size," Economic Development and Cultural Change, January, 1963, Part II. Lemoine, C. J. and Pereira, F., "La Reproducci6n Social en Colombia: Una Aplicaci6n de las Cadenas de Markov," Compania Colombiana de Datos, Bogota, 1975. Levinson, H. M., "Unionism, Concentration, and Wage Changes: Toward a Unified Theory," Industrial and Labor Relations Review, January, 1967. Lewis, W. A., "Economic Development with Unlimited Supplies of Labor," The Manchester School., May, 1954. -98- Mincer, J., Schooling, Experience and Earnings, Sew York, Columbia University Press, 1974. Misas, G., "Contribuci6n al Esttdio del Grado de Concentraci6n en la Industria Colombiana," Boletin Mensual de Estadistica, DANE, 1976. Musgrove, P., "Urban Household Income and Consumption Patterns in Latin America: A Comparative Analysis of Colombia, Paraguay, Peru and Venezuela," Washington, Brookings Institution, mimeo, June 1974. Musgrove, P., Income and Spending of;,Urban Families in Latin America, Washington, Program of Joint Studies of Latin American Econom.c Integration, forthcoming. Nelson, R. R., T. P. Schultz, and R. L. Slighton, Structural Change in a Developing Country: Colombia's Problems and Prospects, Princeton, N.J.: Princeton University Press, 1971. Nerlove, M., "Household and Economy: Toward a New Theory of Population and Economic Growth," Journal of Political Ecoaomy, March/April, 1974, Part II. Parra Sandoval, R., AnAlisis de Un Mito: La Educaci6n como Factor de Movilidad Social en Colombia, Departamento de Educaci6n, Universidad de Los Andes, Bogotd, 1973. Perlman, R., Labor Theory, New York, John Wiley and Sons, 1969. Perry, G. L., Unemployment, Money Wage Rates, and Inflation, M.I.T. Press, 1966. Prieto, R., Estructura del Gasto y Distribuci6n del Ingreso Familiar en Cuatro Ciudades Colombianas: 1967-68, Centro de Estudios sobre Desarrollo Econ6mico, Universidad de Los Andes, Mayo, 1971. Psacharopoulos, G., Returns to Education: An International Comparison, San Francisco, Jossey Bass/Elsevier International Series, 1973. Rama, G. D., "Educacidn Universitaria y Movilidad Social: Reclutamiento de Elites en Colombia," C.I.D., Universidad Nacional de Colombia, Noviembre, 1969. Reynolds, L. "Relative Earnings and Manpower Allocation in Developing Economies," Pakistan Development Review, Spring, 1969. Rosen, S., "On the Interindustry Wage and Hours Structure," Journal of Political Economy, March/April, 1969. , "Learning and Experience in the Labor Market," Journal of Human Resources, 1972. -99- Rosenzweig, M., "Neoclassical Theory and the Optimizing Peasant: An Econometric Analysis of Market Family Labor Supply in a Developing Country," Economic Growth Center, Yale University, Discussion Paper No. 271, 1977. Salazar, D., "Estudio de los Ingresos Familiares," Boletin. Mensual de Estadistica, No. 238, DANE, Mayo, 1971. Sanjinds, C., "Tendencias del Salario Real Para Obreros y Empleados en el Sector Manufacturero Colombiano, 1960-1970," Documento No. 024, Centro de Estudios sobre Desarrollo Economico, Universidad de Los Andes, Bogota, 1975. -Schultz, T.P., "Returns to Education in Bogota, Colombia," The Rand Corporation, Santa Monica, Calif., RM-5645-RC/AID, September, 1968. , "Determinants of Internal Migration in Venezuela: An Applica- - - tion of the Polytomous Logistic Nodel," paper presented at the Econometric Society World Congress, Toronto, August, 1975. Selowsky, M. , "El Efecto del Desempleo y el Crecimiento sobre la Rentabilidad de la Inversi6n Educacional: Una Aplicaci6n a Colombia," Revista de Planeaci6n y Desarrollo, Julio, 1969. Spence, M., "Job Market Signaling," quarterly Journal of Economics, August, 1973. Stafford, F., "Concentration and Labor Earnings: Comment," American Economic Review, March, 1968. Stiglitz, J.E., "Alternative Theories of Wage Determination and Unemploy- ment in LDCS: The Labour Turnover Model," Quarterly Journal of Economics, May, 1974. , "The Theory of 'Screening," Education, and the Distribution of Income," American Economic Review, June, 1975. , "The Efficiency Wage Hypothesis, Surplus Labour, and the Distri- bution of Income in LDCs," Oxford Economic Papers, July, 1976. Thurow, L.C., Generating Inequality, Basic Books, 1975. Urrutia, ML, "El Desempleo Disfrazado en Bogota," in Centro de Estudios sobre Desarrollo Econ6mico, Universidad de Los Andes, ed., Empleo y Desempleo en Colombia, Bogota, 1968. , Historia del Sindicalismo en Colombia, Ediciones Universidad de Los Andes, Bogota, published in English as The Development of the Colombian Labor Movement, Yale University Press, New Haven, 1969. -100- ______"La Educac16n Coma Factor de Movilidad Social-;"CEDE, Universidad de Los Andes, Documento de Trabajo No. 12, Julio, 1974. ______and Sandoval, C. E., "Distribucio'n de la Educaci6n y Distribuci6n del Ingreso," Revista del Banco de la Repuiblica, Enero, 1971. Wachtel, H. M. and Betsey, C., "Employmeit at Low Wages," Review of Economics and Statistics, May, 1972. Weiss, L. W., "Concentration and Labor Earnings," The American Economic Review, March, 1966. Welch, F., "Education in Production," Journal of Political Economy, Janu- ary/February, 1970. Williamson, J., "Regional Inequality and the Process of National Development: A Description of the Patterns," Economic Development and Cultural Change, July, 1965, Part II.