World Bank Reprint Series: Number 200 J. B. Knight and R. H. Sabot The Returns to Education: Increasing with Experience or Decreasing with Expansion? Reprinted with permission from Oxford Bulletin of Economics and Statistics, vol. 43, no. 1 (February 1981), pp. 51-71. THE RETURNS TO EDUCATION: INCREASING WITH EXPERIENCE OR DECREASING WITH EXPANSION? * J. B. Knight and R. H. Sabot I INTRODUCTION In earnings function analysis, whether of rich countries or poor, two results are uni- versally obtained: a positive coefficient on the education variable (years of schooling) and a positive coefficient, at least initially, on the experience variable (years spent in employment). In each of 53 cases studies in 32 countries surveyed by Psacharopoulos, there was a positive association between the educational attainment of employees and their earnings;1 Blaug reports evidence for some 40 countries that age-earnings profiles (age being a proxy for experieoce) at different educational levels are concave from below.2 There is no general agreement on the explanation for these results: both are predicted by human capital theory but they are also consistent with other explana- tions. The positive coefficient on education could be caused by human capital, by ability, by stochastic screening for ability, or by employer preference for the educated, i.e. 'credentialism'. The initially positive coefficient on employment experience could represent human capital acquired on-the-job or the operation of an internal labour market. In this paper we are directly concerned not with these two phenomena but with a third and closely related phenomenon: the tendency for the coefficient on education to depend on the length of employment experience. Like the other two, such a result is open to more than one interpretation. If education and experience have an inter- active effect on earnings, however, this suggests the need for an explanation which encompasses all three phenomena. The tendency for the coefficient on education to rise as employment experience lengthens is equivalent to earnings-experience profiles that are steeper for the more educated groups of workers. Layard and Psacharopoulos provided evidence of this for the United States,3 and Blaug evidence for Thailand.4 Similarly, Mazumdar found for Malaysia that the earnings profiles of groups at successive educational levels generally diverge with age, at least up to the points at which the profiles peak.5 Psacharopoulos observed no such effect for Morocco, however: earnings-experience profiles were parallel.6 A further test of whether the coefficient on education depends on employ- ment experience is provided by the coefficient on the product of education and *We are grateful to Mark Blaug, Mary Jean Bowman, Chris Gilbert, Benjamin King, Andrew Oswald, Bridget Rosewell and the Editors for comments on an earlier draft. ' Psacharopoulos (1973), chapter 3. 2Blaug (1976), p. 837. 3 Layard and Psachaiopoulos (1974), p. 992. ' Blaug (1974), p. 4. 'Mazumdar (1979), pp. 526-3. 6 Psacharopoulos (1977), pp. 40-1. 51 52. BULLETIN experience in earnings function analysis. Anderson found a significant and positive coefficient on thie product term for El Salvador,7 whereas Mincer obtaineci a significant but negative coefficient in his earnings function, implying a negative relation for the United States.8 In a sophisticated analysis of British data, Psacharopoulos and Layard found that the inclusion of interaction terms produced coefficients that were sensitive to precise specification.9 Using the regression equation underlying their basic model (equation 3.7), however, it can be shown, for plausible ranges of education and employment experience, that the interaction terms raise earnings as either education or employment experience is increased. The impression we get from the literature is that the effect of interaction between education and experience has not been systematically examined. We propose to help remedy this deficiency by means of a case study in a particular developing country, Tanzania. In 1971 one of the autlhors conducted a survey of some 1000 manufacturing employees in Dar es Salaam. A full analysis of this survey is to be found elsewhere;10 here we are concerned witlh only one aspect of the earnings function analysis. The relevant characteristics of the Tanzanian economy are a low average level of education among employees, a rapid expansion of the educational system in the years prior to the survey, and an inflexible labour market in whiclh market forces are weakened as a result of various forms of intervenition. There are two main ways of examining the data for interaction of the sort suggested. One is to introduce an interaction term in the earnings function. Thus, if E represents years of education and L years of employment experience, the term E.L can be included as well as E, L and L2. The interaction term represents the joint effect of E and L over and above the sum of their separate effects. A positive coefficient on E.L can be interpreted in two ways: as showing the coefficient on E rising as L lengthens or the coefficient on L rising as E lengthens. The second approach is to examine the coefficient on E in regression equations stratified by length of employ- ment experience, and the coefficient on L and L2 in regression equlations stratified by years of education. The former method has the advantage of providing significance tests of interaction but runs the risk of collinearity between a product term and its component predictors. We make use of both methods. In examining the coefficient on education we are concerned with the gross, and not the net, returns to education, i.e. with the returns prior to the deduction of costs. We do not therefore attempt in this paper to assess the social value of educa- tional expenditure. Our interest lies in understanding the operation of the labour market and in interpreting the earnings premium associated with education. Evidence is presented in Section Il to show that education and employment experi- ence do indeed have an interactive effect on earnings. Section III sketches out explana- tions of this evidence which stem from established theories. The innovative part of the paper is our development of a new explanation. In Section IV we show that a process of 'filtering down' has occurred in Tanzania: as education has expanded, so 7Anderson (1980), p. 139. This result was confirmed in correspondence with the author. 8Mincer (1974), pp. 92-3. 9 Psacharopoulos and Layard (1979), p. 493 and Table III. '1 Knight and Sabot (1980). THE RETURNS TO EDUCATION 53 educated entrants to the labour market have accepted lesser jobs. Section V examines the role of occupation in the determination of earnings. We show that, because of the role that occupation plays, filtering down can explain the evidence. Section VI con- siders the implications and concludes. II EVIDENCE FOR TANZANIA In order to test for interaction effects the following specifications were made for the sample of 777 African 'regular'll employees in the manufacturing sector of Tanzania in 1971:12 Z=Z(E,L,L2,ZXk) (1) Z = Z(E, L, L ,E.L, WXk) (2) Z = Z(ZEii,EL1, IXk) (3) Z = Z(-7Ei, Y-Lj, ZEi.Li, Y-Xk) (4) where Z= log earnings E = years of education L = years of wage employment experien-.e Ei = education dummy variables, with El = no education (base dummy) E2 = primary standarnis 1-4 E3 = primary standards 5-8 E4 = post-primary education L= employment experience dummy variables, with L l = 1-3.9 years of employment experience (base dummy) L2 = 4-7.9 years of employment experience L3 = 8-11.9 years of employment experience L4 = 12 years or more of employment experience Xk = vector of other explanatory variables.13 A particular advantage of this data set over most others is that it provides information on the actual number of years of experience in wage employment. It is not necessary, therefore, for us to use 'age minus years of education minus 6' as a proxy for length of employment experience, which is the normal procedure in earnings function analysis 14 " Those employed on a monthly basis were defined as 'regular' and those on a daily basis as casual' employees; only a small proportion of respondents were casual employees. "2A random sample was taken of manufacturing firms in Dar es Salaam, stratified by firm size; sampling of employees within firms was random or stratified by department. Our findings of labour market segmentation led us to confine the analysis to African regular employees. '3The variables used are dummies for occupation, formal training, migrant status, sex and age. The occupation dummy variables will be examined separately below. '4The use of this proxy is liable to produce upward bias in the coefficient onE (Blinder, 1976, pp. 13-14). 54 BULLETIN A comparison of equations (1) and (2) and of (3) and (4) enables us to assess the effect of the interaction terms. The results are set out in Table 1. Equation (1) shows the familiar earnings function results: significantly positive coefficients on E and L and a significant negative coefficient on L2. The introduction in equation (2) of the product term E.L reduces the coefficient on E: it remains positive but is no longer statistically significant. The coefficient on L is also reduced in equation (2) but it is still significant. The coefficient on E.L is found to be positive and highly significant. At the mean values of E and L (E = 3.63 and L = 9.02), E raises earnings by 5.4 per cent, L by 24.8 per cent, and E.L by 12.4 per cent. Moreover, the application of the hier- archical F test to equations (1) and (2) refutes the null hypothesis that the interaction term is not significant. TABLE 1 Earnings Functions With and Without Interaction Terms Equation: (1) (2) (3) (4) Coefficients in: E 0.0491** 0.0148 L 0.0556** 0.0338** L2 0.0011** -0.0007** E.L 0,0038** E2 0.0516 0.0412 E3 0.1455** 0.0207 E4 0.7304** 0.4054** L, 0.1462** 0.0804 L3 0.2852** 0.1600 L4 0.4258** 0.2426** E2.L2 -0.0356 E2.L3 -0.0143 E2.L4 0.0720 E3.L2 0.0627 E3.L3 0.1942 E3.L4 0.2344* E4.L2 0.2369* E4.L3 0.4079* E4.L4 0.8437** k2 0.4646 0.4810 0.4818 0.4982 SE 0.3821 0.3762 0.3759 0.3699 F 52.80 52.37 46.09 31.82 Notes: The dependent variable in each of the equations is Z (log earnings); similar results are obtained when W (earnings) replaces Z as the dependent variable. Variables included in the equations but not reported in the table are occupation, formal training, migrant status, sex and age dummies. * Indicates in this, and subsequent, tables that the coefficient is significant at the 5 per cent level. ** Indicates in this, and subsequent, tables that the coefficient is significant at the 1 per cent level. Similar results are obtained when education and employment experience dummy variables (permitting non4inearities) are used in place of E, L and L2 (equations (3) and (4)). Where El (no education) is the base dummy, the coefficients on E2, E3 and THE RETURNS TO EDUCATION 55 E4 are all reduced by the introduction of interaction terms: only the coefficient on E4 remains significantly positive. The coefficients on L2, L3 and L4 (with LX, less than 4 years of employment experience, as the base dummy) are all reduced and only that on L4 is significant. On the other hand, the interaction terms are generally positive and four are significantly so. It is interesting that the interaction-tenn co- efficients rise consistently as employment experience is increased, holding the educa- tion dummy constant, and as education is increased, holding the experience dummy constant. The coefficient on E4.L4 implies that the interaction tenn raises the earn- ings of the group with post-primary education and 12 years or more of employment experience above the earnings of the base dummy group by 133 per cent."5 The introduction of the nine interaction terms in equation (4), however, fails the hier- archical F test of significance. Another method of detecting interaction effects is to stratify the sample in turn by employment experience and education, and to examine how the coefficients on E and on L differ for different strata (Table 2). Consider a division of the sample TABLE 2 Coefficients on Education and Employment Experience in Stratified Earnings Functions Equation: (5) (6) (7) (8) Employment experience (years): 0-3.9 4-7.9 8-11.9 12 or more Coefficient onE 0.031** 0.044** 0.048** 0.078** k2 0.336 0.428 0.393 0.553 W 249 278 343 466 N 180 201 199 198 Equation: (9) (10) (11) (12) Education (years): 0 1-4 5-8 9 or Ymiore Coefficient onL 0.0129** 0.0185** 0.0392** 0.0384* R2 0.346 0.265 0.477 0.562 Equation: (9') (I 0) (11') (12') Education (years): 0 1-4 5-8 9 or more Coefficient on: L 0.0318** 0.0260 0.0685** 0.0626 L 2 -0.0006* -0.0003 -0.0014* -0.0012 K2 0.364 0.262 0.485 0.554 W 260 291 315 884 N 201 226 292 58 Notes: Other independent variables included in the regressions are occupation, formal training, migrant status, sex and age. W is the mean earnings, by employment experience group or by education group, in spm. "SThe proportionate increase (g) exceeds the coefficient (c) on E4.L4: l1Og = 100 (exp(c)-1). See Halvorsen and Palmquist (1980). 56 BULLETIN into four employment experience groups. An earnings function analysis with Z as the dependent variable and E among the independent variables reveals that the coefficient on E rises consistently with employment experience (equations (5)-(8)). Secondly, a stratification of the sample by educational category produces similar results for employment experience. When L2 is excluded from the regression (equations (9)- (12)), the coefficient on L rises with education at each but the final step. With L2 included (equations (9')-(12')), the value of (aL-bL2) at the mean value of L (9.02 years) implies a sharp distinction between the two lower and the two higher educa- tional categories. Not only does a year of education add proportionately more to earnings the higher the employment experience of the worker, but also a year of employment experience adds proportionately more to earnings the higher his educa- tional attainment. III CONVENTIONAL EXPLANATIONS Neither human capital theory nor screening theory in simplest form predicts the positive interaction results. If human capital based on formal education depreciates over time, simplest human capital theory predicts that the influence of education on earnings declines as employment experience lengthens. To explain the increasing returns to education with experience, a human capital theorist would have to argue that the possession of education itself facilitates skill acquisition on-the-job, i.e. prior education is an input into the production function of skill formation. It is also in the spirit of the screening hypothesis that the partial effect of educa- tion on earnings should fall with length of employment experience. If screening by education explains the higher starting wages earned by the educated, the influence of education on earnings deciines as employment experience lengthens: as employers get to know their workers, they can obtain more direct evidence on productivity and need no longer rely on educational quaiificaticns. This tendency may not be strong because it is difficult for an employer to discover the productivity of one individual in a complex production process. Nevertheless, the interaction effect can be expected to be negative."6 The human capital explanation of positive interaction also provides a proponent of screening with a counter-argument, however: personal ability, rather than education, is the input into the production function of skill formation on-the-job. He might thus argue that people are screened for ability on entry to employment and are allocated to jobs on that basis. Some jobs involve more training on-the-job than others; the more educated, because they are more trainable, are allocated to jobs in which they can acquire more human capital through experience. Whether the initial criterion for allocation to jobs is education as a form of human capital or as a screening device, the theory of internal labour markets can provide an explanation of the subsequent increasing importance of education. Job ladders and associated wage scales may be institutionally determined in such a way that an 16 Layard and Psacharopoulos (1974) used this prediction as a test of screening against human capital theory. THE RETURNS TO EDUCATION 57 individual's point of entry determines not only his relative starting wage but also how high and how fast he can climb.'7 To vary the metaphor, people with education are selected for escalators that rise rapidly and others for escalators that rise slowly: it follows that earnings differentials will grow with time. If the interaction effect is indeed to be explained in terms of internal labour markets, that does not necessarily mean that it lacks an economic rationale. It is possible for internal labour markets to be viewed as employers' institutional response to the problem of retaining the benefits of firm-specific training. Incremental scales and seniority rules can have the effects of reducing labour turnover and of securing the co-operation of employees in imparting skills. Internal promotion also provides a screening mechanism for selecting workers with desirable characteristics. Both by reducing labour turnover and improving information on individual employees, internal labour markets reduce the cost of specific traiP 4ng to the employer. To summarize: neither human capital theory nor screening theory in simplest form predicts the interaction effects which we wish to explain. With the addition of further assumptions, however, either theory is capable of explaining the evidence. These further assumptions may, but need not, involve internal labour markets. Even if internal labour markets are involved, the rising coefficient on education as employ- ment experience lengthens is likely to reflect human capital formation, although the coefficient may be a poor measure of the product of this human capital. The Tanzanian survey does not enable us to refute either of these explanations of interaction effects nor to favour one of them rather than the other. What we can do, however, is muster the evidence in support of a new explanation which, if correct, must at least diminish the importance of the conventional accounts. IV FILTERING DOWN The variable L conveys more than just the length of time a worker has been in wage employment. It also classifies workers according to the year in which they entered the wage labour force. For example, workers with 12 years or more of experience first entered the wage labour force in the 1 950s, while those with only a year of experience entered in the 1970s. Thus L is open to interpretation from a time-series as well as a cross-section perspective. When separate earnings functions are estimated for workers in each of the Li categories, variation among these functions in the coefficients on education may reflect differences among entry cohorts in the employment oppor- tunities which they encountered rather than the effects of employment experience per se. These differences could have arisen if labour market conditions had changed over time. In that case what appear to be increasing returns to education with experience could represent decreasing returns to education for successive cohorts of entrants to wage employment. This interpretation of the empirical evidence raises two related questions. Why should the returns to education decrease for successive cohorts of entrants to wage employment? Why should earlier educated entrants be better insulated from the forces 17Thurow (1976), chapter 4. 58 BULLETIN reducing returns than more recent entrants? We hypothesize that the answers lie in two factors. One factor is that the number of school leavers increased faster than the demand for labour in the occupations which earlier cohorts largely entered. This caused the relationship between the educational attainments of workers and their occupations to change. There occurred a process of 'filtering down', i.e. the movement of persons at each level of education into lesser jobs or-the reverse of the coin-the educational upgrading of each occupation as the supply of educated manpower is expanded. This increase in the proportion of educated workers in lower level occupations tended to depress the returns to education. Filtering down occurred unevenly, however. Because those already in post were protected from labour market competition, a rising propor- tion of successive cohorts of entrants with a particular level of education entered lower level occupations. The second factor is the lag in the adjustment of the occupational structure of wages to the increase in the relative supply of educated workers. In a perfectly com- petitive labour market such an increase would result in a general reduction in the premium paid to the educated. Both this reduction and the increased educational attainments of workers in lesser jobs could produce a compression of the occupa- tional wage structure. In the Tanzanian case, during the decade before our survey, such compression was slow to occur because supply and demand had only an in- direct effect on earnings. Downward market pressure on wages was weakened by the security of tenure which incumbents generally enjoyed. Moreover, institutional wage determination was important in the public sector and, as the public sector is the main employer of educated manpower, its infiuence was felt also in the private sector. In the 1960s the educated 6lite were able to counterbalance market forces with political pressures. It is this lag in the adjustment of the occupational wage structure that accounts for the relative insulation of earlier cohorts of educated workers from the downward pressure on returns'to education. We intend in this section to establish that a process of filtering down occurred and that its incidence was primarily among those entering the labour market. The next section deals with the effects of filtering down on the returns to education. We begin by assessing whether in Tanzania the grpwth in the supply of educated manpower outstripped the growth in demand and whether, as a consequence, filtering down took place. During the colonial period per capita expenditure on education and the percentage of school-age children in publicly financed schools was among the lowest in Africa.18 In the 1950s and, as Table 3 shows, particularly after independence in 1961 school enrolments increased dramatically and educational appropriations were a growing proportion of a fast-growing budget. Between the years 1962 and 1970, university enrolment increased at an annual average rate of 18.5 per cent, secondary at 10.4 per cent, and primary at 6.0 per cent. The extent of the adjustment problem for the urban labour market which was posed by the increase in educational outputs can be gauged by comparing numbers of school leavers and of employment opportunities in 1962, in 1970, and over the period 1962-70. The lag between input 18 In 1935 only 2.8 per cent of school-age children attended government or government-aided schools. See Sabot (1979), Table 5.3. THE RETURNS TO EDUCATION 59 TABLE 3 The Balance Between Educational Output and Non-Agricultural Wage Employment Opportunities 1962 and 1I 70 1962 1970 1962-70 University output (East African and Overseas) 195 679 3735 Form 6 output 275 1389 6464 of which: entered university 444 1176 6491 entered labour force -1,59 213 -27 Form4 output 1950 6713 41505 of which: entered form 5 297 1608 9142 entered labour force 1653 5105 32363 Standard 7 output 13730 65624 392799 of which: entered form 1 4972 7740 65048 entered labour force 8758 57884 327751 Total entered labour force 10242 63202 360087 of which: primary leavers 8758 57884 327751 post-primary leavers 1484 5318 32336 Increase in non-agricultural employment -8280 13214 79538 Percentage Enrolment in: increase per annum University (East African and Overseas) 915 3550 18.5 Secondary schools 14175 31217 10.4 Primary schools 518663 827974 6.0 Source: United Republic of Tanzania, The Economic Survey 1970-71, Table 76, The Economic Survey 19 71-72, Table 93, Survey of Employment and Earnings, various issues. to and output from an educational level meant that the proportionate growth of educational outputs was highest at the primary level. The number of students emerging from the top of each educational level (standard 7, form 4, form 6 and university completion) increased from 16 150 in 1962 to 74 400 in 1970. Not all entered the labour force: a proportion of those below university level went on to the next stage of education. This proportion declined sharply over the period, largely owing to increased competition for secondary school places. The rapid increase in numbers- completing primary school resulted in a decline from 36 per cent to 12 per cent in the proportion of standard 7 leavers able to gain access to form 1. Over the 8-year period prior to our survey, non-agricultural wage employment in Tanzania increased by almost 80 000 (Table 3). Over the same period the number of standard 7 leavers entering the labour force was 327 750 and the number of post- primary entrants 32 300. Non-agricultural employment fluctuated from one year to another (e.g. the fall in 1962) but the annual average increase over the period was 10 000. In 1962 the number of primary and post-primary school leavers entering the labour force just exceeded 10 000. In 1970 there were nearly 58 000 primary and 60 BULLETIN more than 5000 post-primary leavers, i.e. there were over six educated labour force entrants for every new wage job created. The great majority of primary school leavers could no longer obtain urban wage employment. In 1970 white-collar jobs comprised only 33 per cent of urban wage jobs, and the annual increment was unlikely to involve a significantly higher proportion. During the 1960s the growth of high-level jobs open to Tanzanian citizens had been augmented by the government policy of localization. For instance,in 1962 39 per cent of senior- and middle-grade posts in the civil service were filled by citizens; 8 years later the proportion had risen to 86 per cent and the number of citizens in these posts had increased by 6200.19 Nevertheless, by 1970 non-menial white-collar jobs for secondary school leavers could no longer be assured. The scope for better-educated workers to oust less-educated incumbents from their jobs is severely circumscribed in Tanzania as in many other countries. The Security of Employment Act of 1964 gave workers considerable job security.20 It prevented an employer from dismissing an employee or reducing his pay except for breach of the disciplinary code. In addition it ensured an effective enforcement mechanism. The act provided for the formation of 'workers' committees' in every establishment in which ten or more NUTA21 members were employed, and laid down procedures by which workers' committees and a conciliation board could become involved in the event of a dismissal. Workers' committees are active in supporting workers' grievances, and in any case the general ethos would be overwhelmingly against the notion that a worker might be dispensed with for no better reason that that a more educated worker was available. The implication is that those in a post generally cannot be replaced, and that filtering down can occur only as new jobs are created or old jobs become vacant. The brunt of filtering down is therefore borne by entrants to the wage labour market, among whom employers are free to choose. The more rapid expansion of educational opportunities than of wage employment opportunities set the stage for filtering down. Scarce wage employment opportunities were rationed among competing job seekers, with the level of education as the principal criterion, and with a rise in the educational attainment of more recent recruits as the predictable result. The educational attainments of different cohorts of entrants to various occupations provide evidence of filtering down in the manu- facturing sector. Table 4 shows the extent to which educational levels in the sample as a whole, and of white collar, manual and different types of manual occupations changed over time. The proportion of workers with different levels of education varies greatly according to the date of entry to wage employment. The value of x2 is statistically highly significant for all groups other than white collar. Uneducated entrants and those with only standards 1-4 were increasingly kept out of manufacturing employment. We compare the earliest cohort, which entered wage employment in the 1950s, with the latest, which entered during the period 1968-71. No fewer than 77 per cent I 9 United Republic of Tanzania, Annual Manpower Report to the President 19 71, p. 79. 20Jackson (1979). 21 National Union of Tanganyika Workers, the single trade union in Tanzania. TABLE 4 Educational Composition of African Regular Employment by Year Entered Wage Employment Category of Year entered No Standards Standards' Forms 1- Less than 5 years Total X2 d.f. Significance employees wage employment education 1-4 5-8 S years or more number All employees 1968-71 14.5 13.0 59.5 13.0 27.5 72.5 131 1964-67 26.2 24.8 44.2 4.9 51.0 49.1 206 1960-63 26.0 35.2 31.1 7.8 61.2 38.9 219 up to 1959 31.8 37.9 24.2 6.1 69.7 30.3 214 770 White collar 1968-71 6.3 0.0 48.8 50.0 6.3 93.8 16 11.75 12 0.4655 1964-67 4.8 4.8 62.9 28.6 9.6 91.5 22 1960-63 3.4 0.0 44.8 51.7 3.4 96.5 29 m up to 1959 3.6 14.3 46.4 35.7 17.9 82.1 29 - k 96 m Manual 1968-71 15.7 14.8 61.7 7.8 30.5 69.5 115 89.15 12 0.0000 1964-67 28.6 27.0 42.2 2.2 55.6 44.4 185 z 1960-63 29.5 40.5 28.9 1.0 70.0 29.9 190 cn up to 1959 36.d 41.4 21.0 1.6 77.4 22.6 186 O 676 t of which: e skilled 1968-71 8.7 17.4 47.8 26.1 26.1 73.9 23 31.64 12 0.0016 1964-67 19.7 24.6 50.8 4.9 44.3 55.7 61 > 1960-63 17.3 44.2 36.5 1.9 61.5 38.4 52 -3 up to 1959 26.4 39.1 31.0 3.4 65.5 34.4 87 0 223 semi-skilled 1968-71 6.8 16.9 72.9 3.4 23.7 76.3 59 33.78 12 0-0007 1964-67 18.8 37.5 42.2 1.6 56.3 43.8 64 1960-63 23.9 44.3 30.7 1.1 68.2 31.8 88 up to 1959 39.7 46.0 14.3 0.0 85.7 14.3 63 274 unskilled 1968-71 36.4 9.1 51.5 3.0 45.5 54.5 33 61.28 12 0.0000 1964-67 48.3 18.3 33.3 0.0 66.6 33.3 60 1960-63 52.0 30.0 18.0 0.0 82.0 18.0 50 up to 1959 52.8 38.9 8.3 0.0 91.7 8.3 36 179 62 BULLETIN of manuai workers in the earliest cohort had less than 5 years of education, to be contrasted with 31 per cent in the most recent cohort. The tendency for this pro- portion to fall is to be found at each manual skill level: even in the case of unskilled workers, it declined from 92 per cent for the earliest cohort to 46 per cent for the most recent. The reason for the fall was a filtering down of uppel-primary and post- primary leavers. The proportion of manual workers with standard 5-8 rose from 21 per cent to 62 per cent. Secondary school leavers increasingly entered the skilled manual occupations. More than a quarter of the 1968-71 entrants to skilled manual employment had some secondary education; among workers who entered these occupations a decade earlier only 3 per cent had been to secondary school. Although it was not as marked as in the manual occupations, a process of filtering down occurred also in white collar occupations. Again comnparing the earliest with the most recent cohort, we find that the proportion of white-collar workers with less than 5 years of schooling fell from 18 per cent to 6 per cent, and the proportion with some secondary schooling rose from 36 per cent to 50 per cent. Filtering down in these jobs could be expected to gain pace during the 1970s as localization came to an end and as more secondary school pupils reached the form 4 exit point. Filtering down is also revealed in the occupational distribution of various cohorts of a single educational group. Consider those whose highest educational attainment was standards 5-8, i.e. upper primary schooling. Twenty years ago a primary school certificate was a virtual guarantee of a white collar job in Tanzania. Thus it is not surprising to find, in Table 5, that for the cohort of workers with upper primary schooling who entered wage employment in the 1950s 73 per cent were in white collar or skilled jobs and only 7 per cent in unskilled employment. Labour market conditions changed dramatically for upper primary leavers during the 1960s. Only 28 per cent of the most recent cohort of entrants were able to obtain white collar or skilled employment. The proportion of upper primary leavers entering manu- facturing who took unskilled jobs increased four-fold over the decade. Note that Table 5 understates the extent of filtering down: it does not include the rising pro- portion of upper primary leavers who were unable to obtain any urban wage employ- ment. V OCCUPATION AND WAGE DETERMINATION Our evidence of filtering down and its uneven incidence is not sufficient in itself to explain the positive interaction results. It is necessary also to understand the role that occupation plays in wage determination. Our first task is to establish that it does indeed play an independent role. We then attempt to explain that role and its implications for the interaction term. Table 1 shows the coefficients on education, employment experience and their interaction obtained from equations (1)-{4): Table 6 shows the coefficients on the occupation dummy variables obtained from the same four equations. The occupation variables are: O, = supervisory; 02 = clerical; 03 = skilled manual; 04 = semi-skilled manual; 05 = unskilled manual (the base dummy). Apart from the semi-skilled cate- gory, the coefficients on the occupation dummy variables are significantly positive and TABLE 5 Filtering Down of Upper Primary School Leavers Year of Employees Mean earnings entry to wage Percentage of total Total White collar Semi-skilled Unskilled Total C employment and skilled x White collar Semi- Unskilled spm Index spm Index spm Index spm Index co and skilled skilled H 0 1968-71 28 42 30 125 278 100 199 100 190 100 218 100 c 1964-67 49 31 20 108 348 125 236 119 207 109 285 130 > 1960-63 50 43 7 54 469 168 346 174 257 135 401 184 0 up to l9S9 73 20 7 55 586 211 457 230 281 148 538 247 Z C' 64 BULLETIN TABLE 6 Coefficients on Occupation in Earnings Functions Equation: (1) (2) (3) (4) Coefficients on: O1 0.6092** 0.5728** 0.6006** 0.5489** 02 0.5223** 0.5196** 0.4607** 0.4424** 03 0.2575** 0.2583** 0.2916** 0.2977** 0.1 0.0335 0.0406 0.0590 0.0683 Notes: With 0 (unskilled manual) as the base dummy, the significance tests relate to the difference 'between each occupation and 0,. For a decription of equations 1-4, see pp. 53-5 and Table 1. large. In addition to the personal econonic characteristics which he brings to a job, the occupation of a worker makes a difference to his earnings. Occupation also affects the coefficients on E and L. Earnings functions stratified by occupation are presented in Table 7 (equations (13)-(16)). With the exception of the skilled manual category, the coefficient on L rises with skill level. The coefficient on E is sigrnificant only for the skilled manual and non-manual categories. The economic value not only of employment experience but also of education depends on the occupation in which they are applied. There is also a tendency for the average years of education and of employment experience to increase with skill level. In human capital theory the job that a person does is irrelevant to his earnings: it is the economic characteristics which he brings to bear on the job that determines his TABLE 7 Coefficients on Education and Employment Experience for Different Occupation Groups Equation: (13) (14) (15) (16) Occupation Unskilled Semi-skilled Skilled Non-manual group: 05 04 03 01+02 Coefficient on: E 0.009 0.001 0.069** 0.122** L 0.033** 0.050** 0.041** L2 -0.001 -0.001 -0.001 -0.002 Mean value of: W 223 259 391 656 E 2.2 3.2 3.7 7.4 L 8.1 8.3 10.4 9.5 N 179 274 223 95 02 0.358 0.182 0.288 0.595 Notes: Other independent variables included in the equations are formal training., migrant status, sex and age. The interaction term E.L was deleted from the regressions after it proved not to be significant. The mean values of W represent average earnings by occupation in shillings per month. THE RETURNS TO EDUCATION 65 productivity. A human capital theorist would therefore have to argue that the signifi- cant occupational dummy coefficients are acting as proxies for human capital not otherwise measured. This explanation, however, begs the question of why occupation and unmeasured human capital happen to be correlated. An alternative approach-falling within the internal labour market framework- suggests that occupation is not simply correlated with human capital but determines it. The characteristics required of a worker if he is to perform a job competently involve much occupation-specific human capital. For some jobs it is the nature of the job which reciuires workers to have much human capital for its effective performarnce, and which also generates unmeasured human capital in the form of training and experience acquired on-tle-job. A third view is that occupational wage structure is influenced by such non-market factors as custom, institutions, and bargaining, and adjusts only with a lag to market forces. The significant occupation coefficients might in that case reflect non-economic factors and previous economic conditions. Yet a fourth explanation is that the occupa- tion coefficients represent payments for ability or other personal characteristics observed by the employer but not by the researcher. This could be the case if employers' recruitment and promo'ion decisions were much influenced by such characteristics. Both the job-specific human capital and the lagged adjustment explanations of significant occupation dummy coefficients have implications for the returns to educa- tion in a process of filtering down. An analytical framework is required for a study of these implications. The appropriate theory has been developed elsewhere:22 it is sufficient here to sketch out the essential elements. The concept underlying the analysis is that of the 'occupational production function', i.e. occupation-specific relationships between education and productivity. Education is likely to be of some value in most occupations but its value is unlikely to be uniform. For any one occupation the relation between years of education (E) and productivity (Y) can be illustrated by means of the occupational production function Y1 in Fig. 1. Below some minimum level of education productivity in the occupation is zero; above it, productivity increases with education. The slope of Y1 may decline continuously or rise to a peak and subsequently decline. Eventually the curve becomes horizontal: further education has no effect on productivity. A second occupation may have similar characteristics but quite different position and slope, e.g. Y2. The effects of educational expansion differ accordirng to whether wages are flexible or rigid. Both cases can be analysed by means of the occupational production function. Neither extreme case is likely to apply in practice; nevertheless, we consider each in turn. Assume that initially all workers with education E1 are employed in occupation 01, that all those with E2 are employed in 02, and that E1, E2 and 01, 02 are the only alternatives. Employees are assumed to receive a wage equal to their marginal pro- ductivity, i.e. E1 workers receive a wage denoted by point a and E2 workers a wage denoted by b. "Knight (1979). 66 BULLETIN Y, W WI ---------------- W2 Ez El E Fig. 1. If there is an increase in the number of workers with E1, a perfectly competitive labour market responds as follows. The increase in supply depresses W1, encouraging employment in 01 and so reducing the marginal product of labour in 01. The process continues until a wage corresponding to point c is reached. Workers with El now filter down into 02: in the perfectly competitive case all those with the same educa- tional attainment receive the same wage. Now consider the alternative assumption of wage rigidity: W1 remains at point a. Filtering down begins immediately the educa- tional expansion occurs. If W2 is fixed at a level corresponding to point b, El workers are preferred to E2 workers in 02 because of their higher productivity (c > b). In the rigid wage case, workers with the same educational attainment receive different wages according to their occupation. Moreover, if there is some degree of job security, the pay of El employees is likely to be positively related to length of employment experi- ence: the incumbents of good jobs hold onto fhem and it is the entry cohorts which filter down. We first examine how in Tanzania earnings vary among occupations within an educational group, and then how earnings vary among educational levels within an occupation. Table 5 sets out not only the occupational distribution of upper primary leavers by cohort of entry to wage employment but also their mean earnings by occupation and by cohort. Within each cohort, mean earnings increase as occupa- tional skill level rises. In other words, at a given educational and experience level, the choice of occupation does indeed make a difference to earnings. On average, upper primary school leavers with 12 or more years of employment experience (the 1959 or earlier cohort) earn 2.5 times as much as those with 1-3 years of experience (the 1968- 71 cohort). Disaggregation by occupation, however, suggests that it would be wrong to attribute this increase wholly to the effect of experience on earnings. Mean earnings within an occupation increase proportionately less with employment experience than do mean earnings for the group as a whole. This is due to the change in the occupa- tional composition of successive entry cohorts. Earlier cohorts are concentrated in THE RETURNS TO EDUCATION 67 high-pay white-collar and skilled jobs; more recent cohorts in semi-skilled and unskilled jobs. The premium apparently associated with experience is therefore an exaggerated measure of the benefit derived from employment experience. The existence in Tanzania of occupational production functions different for each occupation is suggested by data of rn.ean earnings by occupation and by years of education. There is a problem of small numbers and outliers, but occupational pro- duction function curves are fitted freehand to the data by way of illustration in Fig. 2. The occupational earnings differentials within each educational group suggest either that there are differences in the post-school human capital required for the per- formance of different jobs or that there is a disequilibrium wage structure stemming from a failure of, or a lag in, the adjustment of wages to the increased supply of educated manpower. We believe that the latter is at least part of the explanation. w 1400 Whitecollar 1200 - 1000 / 800 / Skilled 400 /X * Whitecollar / ~X Skillted nl0 Semi-skilled 200 L Semi-skilled A Unskilled 0 2 4 6 8 10 12 14 16 E Fig. 2. In Tables 1 and 2 it was found that the coefficient on the education variable increased with length of employment experience, i.e. that there was a positive inter- action effect between education and employment experience. The fact that the co- efficient on the interaction term E.L is statistically significant in the unstratified regression but not in the regressions stratified by occupation23 suggests that the changing occupation composition is responsible for the observed interaction. There is no necessity for a process of filtering down combined with wage rigidity to produce 23 See the notes to Table 7. 68 BULLETIN interaction. For instance, it would have been possible for our inclusion of occupation dummy variables in equations (1)412) to have picked up all the effects of changing occupational composition on the estimated equations. This is most unlikely to have been the case in Tanzania, however, because occupation affects not only the constant term in the regressions (Table 6) but also the values of the coefficients (Table 7). The implications of filtering down can be illustrated as follows. We saw, that entrants with 5 years or more of education increasingly took unskilled and semi- skilled jobs (Table 4). These are the jobs for which the coefficient on the education variable is low (Table 7). The coefficient on education can be viewed as the weighted average of the education coefficients for the different occupations: the greater weight attached to unskilled and semi-skilled employment for the more recent entry cohorts would therefore reduce the coefficient in their case. The effect of filtering down on the education coefficient is evident from the much flatter shapes of the unskilled and semi-skilled than of the skilled and white-collar occupational production functions illustrated in Fig. 2. Filtering down in conditions illustrated by Fig. 2 can also explain the tendency for the coefficient on employment experience to rise with the level of education (equations (9)-(12) and (9')(12') in Table 2). This is because earnings by occupation, and therefore earnings by cohort, differ more for the better educated workers. A basic charge has been levelled against regression analyses which stress the role of occupation in wage determination. It is that endogenous variables such as occupation should not be included as independent variables in earnings functions.24 Education can have both a direct effect on earnings and an indirect effect through the choice of occupation: it can influence a worker's income within an occupation and the occupa- tion in which he is employed. The effect of including occupation as an independent variable is thus to overstate the role of occupation and to understate the role of education in earnings determination. Such bias can be avoided by excluding occupa- tion as an independent variable in the earnings function. Unfortunately, however, a new source of error then arises. If occupation does in fact have an independent effect on earnings as we have argued, then that effect will be missed in the regression analysis: In similar vein the validity of earnings function estimates for stratified sub-samples has been questioned for cases in which the stratification variable is endogenous.25 Only if the labour market is segmented is stratification by occupation valid; there must not be mobility between occupations.26 If a worker is mobile, the earnings determinants within his current occupation do not represent the determinants over his working life. Thus the difference in the occupational composition of the various "4Fields (1980), pp. 242-5 1. "Fields (1980), pp. 251-5. 26 It is also important, when stratifying into sub-samples, tl at the labour force be homogeneous with respect to omitted variables. Fields (1980), p. 254) has :.gued that education or experience may substitute for ability in determining a worker's occupation. If high-education or long- experience workers in low occupations tend to have low ability, the effect is to reduce the co- efficient on education or experience in earnings functions for those occupations. But this argument does not provide a convincing explanation of the tendency for the coefficients on education and experience to rise with occupational level. By the same token, low-education or short-experience workers in high occupations should have high ability, and this should reduce the coefficients for these occupations also. THE RETURNS TO EDUCATION 69 cohorts might be mainly the result of experience acquired or promotions achieved during employment and not, as we have argued, mainly the result of different employ- ment opportunities on entry to the wage labour market. There are two replies to this criticism. First, even if mobility between occupations is substantial, the determinants of earnings within occupations are relevant to under- standing the operation of the labour market; the discovery that significant occupational differences do exist is instructive. Secondly, there is the difficult empirical question of the extent of mobility between occupations, which only longitudinal surveys can answer conclusively. The survey itself provides little evidence on the extent of occupational mobility. Information is available on occupation with the previous employer and current occupation. No fewer than 47 per cent of those previously unskilled remained un- skilled and 35 per cent had become semi-skilled; only 15 per cent had risen above these ranks to skilled manual jobs and 3 per cent to white-collar jobs. Of those who had previously been semi-skilled or skilled, 8 per cent had risen and 8 per cent had fallen in occupational level. The sample is characterized by very low rates of labour turnover. The evidence for upward mobility within the firm, however, is no stronger than that between firms. Some 79 per cent of the sample had been doing the same work since joining the firm. Asked to compare their initial and curient skill levels within the firm, 85 per cent claimed to be at the same skill level and 14 per cent to have risen in skill. The majority of unskilled workers saw no prospect of internal promotion; of the 19 per cent who saw a prospect, two-thirds expected a semi-skilled job. The evidence of the survey is consistent with our argument. Our case rests primarily on the reasons presented in Section IV, however, for believing that, because of the loosening in the labour market for entrants which has occurred with educational expansion, the job prospects of labour market entrants at most educational levels liave deteriorated over time in Tanzania. The most recent cohort in 1971 could not expect to achieve by 1981 the 1971 occupational distribution of the cohort 10 years their senior.27 VI CONCLUSION We have found evidence of positive interaction between education and experience in our earnings function analysis of Tanzanian manufacturing. This result is not unique to Tanzania: it appears to be a more general, but not universal, phenomenon. An explanation in terms of human capital theory can be found; this requires only the further assumption that formal education is itself an argument in the production function for human capital, i.e. formal schooling and post-school human capital formation are complementary. A second explanation, however, related to the first, is that internal labour markets are so structured that a worker's education influences not only his point of entry and initial earnings but also his subsequent progress and 2"This hypothesis will be formally tested using the comparable survey conducted in 1980 by a team including the authors. 70 BULLETIN earnings throughout his working life. Our own interpretation is that the rapid expansion in the stock of educated manpower, combined with wage rigidity and job protection in the labour market, forced the more educated entrants to accept lesser jobs than their predecessors had done: the earnings premium on education for more recent cohorts was therefore lower than that for earlier cohorts. In putting forward this explanation, we have not rejected the conventional explanations. Our account of filtering down and its implications, however, has the virtue of simultaneously solving more than one puzzle. Whether it can be generalized to other countries remains to be researched. The implications of our explanation for the estimation of returns to education are far-reaching. If the apparent rise in the returns to education with employment experi- ence is partly or mainly due to a fall in the returns to the more recent cohorts, then the coefficient on the education variable in our regression equation, by averaging over the sample as a whole, overestimates the marginal retutns to education. It is of course the marginal, not the average, returns that are relevant to both private and public decision-making on educational investment. The coefficient on education among workers in the most recent (1968-71) entry cohort is perhaps a better measure of the marginal returns to education than the coefficient in the unstratified regression. If so, the gross private returns per year of schooling is a third less than previously indicated, i.e. earnings increased by 3.1 per cent for each additional year of schooling (equation (5)) instead of by 4.9 per cent (equation (1)). We recognize that the use of this coefficient carries with it the danger of downard bias if indeed, within a human capital framework, school and post-school investments are positively correlated: the earnings of the more educated are depressed more than those of the less educated while they pay for their training. Our contribution, however, is to explain why the use of the coefficient on education derived in the conventional way would produce up- ward bias in the Tanzanian case. It is no more than that: we do not argue in this paper that the filtering down revealed by the survey was a bad thing, nor that the true net rate of return to education precluded a policy of continued educational expansion. This study, like virtually all econometric research into wage structures in develop- ing countries, is based on data for a cross-section of the labour force. Given that the analysis must be essentially static, there is a tendency to make a virtue out of a necessity by assuming that the snapshot is of the labour market in equilibrium. When it is at odds with the facts, the assumption of static equilibrium can be misleading. Cross-section evidence that makes sense only when viewed in the context of a dynamic labour market in disequilibrium cannot be forced into an equilibrium mould. The conventional interpretations of the rise in the coefficient on education with employ- ment experience may well be a case in point. Institute of Economics and Statistics, University of Oxford. World Bank, Washington DC. REFERENCES Anderson, Lascelles (1980). 'Rates of Return to Human Capital: A Test Using El THE RETURNS TO EDUCATION 71 Salvador Data', American Economic Review, Papers and Proceedings, Vol. 70, No. 2. May. Blaug, Mark (1974). 'An Economic Analysis of Personal Earnings in Thailand', Eco- nomic Development and Cultural Change, Vol. 23, No. 1, October. Blaug, Mark (1976). 'Human Capital Theory: A Slightly Jaundiced View', Journal of Economic Literature, Vol. 14, No. 3, September. Blinder, Alan S. (1976). 'On Dogmatism in Human Capital Theory', Journal of Human Resources, Vol. 11, No. 1, Winter. Fields, Gary S. (1980). 'Education and Income Distribution in Developing Countries: A Review of the Literature', in Timothy King (ed.), Education and Income, World Bank Staff Working Paper No. 402, July, Washington DC. Halvorsen, Robert and Palmquist, Raymond (1980). 'The Interpretation of Dummy Variables in Semi-logarithmic Equations', American Economic Review, Vol. 70, No. 3, June. Jackson, Dudley (1979). 'The Disappearance of Strikes in Tanzania: Incomes Policy and Industrial Democracy', .Journal of Modern African Studies, Vol. 17, No. 2, June. Knight, J. B. (1979). 'Job Competition, Occupational Production Functions, and Filtering Downi', Oxford Economic Papers, Vol. 3 1, No. 1, July. Knight, J. B. and Sabot, R. H. (1980). Why Wages Differ: A Study of Tanzania's Manufacturing Sector, World Bank, Washington DC (mimeo). Layard, R. and Psacharopoulos, G. (1974). 'The Screening Hypothesis and the Social Returns to Education', Journal of Political Economy, Vol. 82, October. Mazumday, Dipak (1979). Urban Labor Market and Income Distribution in Peninsular Malaysia, World Bank, Washington DC (mimeo). Mincer, Jacob (1974). Schooling, Experience and Earnings, National Bureau of Eco- nomic Research, New York. Psacharopoulos, George (assisted by Keith Hinchliffe) (1973). Returns to Education. An International Comparison, Amsterdam, Elsevier. Psacharopoulos, George (1977). 'Schooling, Experience and Earnings: The Case of an LDC', Journal of Development Economics, Vol. 4, No. 1, March. Psacharopoulos, George and Layard, Richard (1979). 'Human Capital and Earnings: British Evidence and a Critique', Review of Economic Studies, Vol. 46. Sabot, R. H. (1979). Economic Development and Urban Migration. Tanzania 1900- 1971, London, Oxford University Press. Thurow, Lester (1976). Generating Inequality, London, Macmillan. THE WORLD BANK Headquarters: 1818 H Street, N.W. U Washington, D.C, 20433, U.S.A. European Office: 66, avenue d'Iena 75116 Paris, France Tokyo Office: Kokusai Building, 1-1 Marunouchi 3-chome Chiyoda-ku, Tokyo 100, Japan The full range of World Bank publications, both free and for sale, is described in the World Bank Catalog of Publications, and of the continuing research program of the World Bank, in World Bank Research Program: Ab- stracts of Current Studies. The mnost recent edition of each is available with- out charge from: PUBLICATIONS UNIT THE WORLD BANK 1818 H STREET, N.W. WASHINGTON, D.C. 20433 U.S.A.