Graduate employment and the returns to higher education in Africa
Mahdi Barounia andStijnBroeckeb
a Institute for Research in the Sociology and Economics of Education, Universityof Bourgogne, France
bResearch Department, African Development Bank, Tunis, Tunisia
Abstract
In this paper, we estimate the return to higher education for 12 African countries using recent data and a variety of methods. Importantly, one of our methods adjusts for the effect of higher education on the rate of joblessness, which is substantial in most African countries, and particularly for women. Our results confirm that Mincerian coefficients cannot be interpreted as a true rate of return, and that thelatter (even after taking into account the employment effect) is considerably lower than what has previously been suggested in the literature (less than half). For Sub-Saharan Africa, we also uncover an interesting relationship between a country’s level of education and the return to higher education: contrary to expectations, we find that in countries where a high proportion of the working age population is educated to tertiary level, the return to higher education is highest.
JEL classification: I21, I23, J31
Key words: graduate unemployment, returns to education, higher education
- Introduction
Estimates of the return to higher education are a potentially useful tool for policy-makers. In England, for example, the increase in fees from September 2012 was largely defended on the grounds that the private returns to higher education are high: “The lifetime earnings of graduates are higher than those of non-graduates. The so-called graduate premium, net of tax, is still worth comfortably over £100,000 in today's money.”[1]
At the same time, a recent survey of the returns to higher education in Africa found that “an increasing amount of work needs to be done with respect to returns to higher education in Sub-Saharan Africa” (Diagne and Diene, 2011). The authors argue that most existing studies focus on a limited number of countries only, and that the data are frequently more than 15 years old. An initial objective of the present paper was therefore to fill this gap by looking at the return to higher education in a range of African countries using recent household and labour force survey data. This is one contribution this paper makes to the literature.
A second contribution of the paper is that it adjusts these private rates of return to higher education by the likelihood of finding employment. As argued by Barceinas et al (2000) “increasing the educational level is profitable not only because this allows the more educated to obtain higher wages, but also because the more educated have a lower probability to become unemployed. As a consequence, the returns to education must be computed taking the unemployment probability into account.” Similarly, Colclough et al (2009) remark that the returns to education in developing countries are “typically not adjusted for unemployment among the educated”. Writing in the context of Europe, where unemployment decreases with education, Barceinas et al (2000) conclude that “unadjusted marginal rates of return to education present a general underestimation because the differentials in employment probability […] across levels of education are not taken into account”.[2]
In many African countriesthe unemployment rate for graduates is higher than for the working age population overall (top panel Figure 1). However, because of high rates of inactivity among people with lower levels of education, this is not true of the level of joblessness (bottom panel Figure 1), which is almost always lower among graduates. So, unlike in Europe, a complex relationship exists in Africa between the level of education and labour market outcomes. In nearly all countries analysed in this paper we find that the returns to higher education increase once the risk of joblessness is factored in. This effect is particularly strong for women.
The analysis uncovered some other interesting findings. Most importantly, we show that the return to higher education varies widely with the method used for estimating it. In particular, we find thatestimates of Mincerian coefficients are frequently higher than the true internal rate of return, and should therefore not be confounded with them. This calls for more caution and candidness by researchers in describing the methods that they use as well as in explaining the meaning of their estimates. We follow Heckman, Lochner and Todd (2008) in arguing that Mincerian coefficients should no longer be blindly (and often wrongly) presented as estimates of the “return to education”[3].
Finally, we find that the rate of return to higher education varies considerably across countries. Interestingly, we find that in countries where a high proportion of the working age population has a university qualification, the rate of return to higher education is the highest.
Figure 1: Graduate unemployment and joblessness in Africa
Source: Authors’ calculations based on Minnesota Population Center (2011). Integrated Public Use Microdata Series, International: Version 6.1
The remainder of this paper is structured as follows. Section 2 provides a brief overview of the literature on the returns to higher education in Africa. Section 3 offers a discussion of the methodologies used for estimating the returns to education, their drawbacks and advantages. Section 4 describes the data that we use in our paper, and Section 5 presents the results from the analysis. Section 6 concludes.
- Literature review
Popular wisdom, influenced by the work of Psacharopoulos (1973, 1981, 1985, 1994) and Psacharopoulos and Patrinos (2004a, 2007), has it that the returns to higher education in Africa are lower than those at lower levels of education. These findings have very much influenced donor investments in education across the continent, but have been criticised by a number of authors, including Bennell (1996)who questioned the quality of the data and analysis used in these studies, as well as Schultz (2003) who analysed data from six African countries and found that private returns were actually higher at secondary and post-secondary levels that at primary level.
Since then, a large number of individual country studies have been published, and these have recentlybeenreviewed by Diagne and Diene (2011). The studies they review for the period 2000-10 (ten studies for six countries) reveal an average rate of return to higher education of 19.0%. The estimates obtained by Colclough et al for four African countries average 26.0%, and their literature review suggests an average return of 22.7%. All these estimates are for one year of university education. Furthermore, Diagne and Diene (2011) conclude that the return to higher education is higher than for other levels of education (a finding confirmed by Kingdon et al, 2008; and Colclough et al, 2009), although it has been falling over time.
As Diagne and Diene (2011) and others allude to, one obvious problem with the various estimates obtained in the literature is that there is very little consistency in the methods used to obtain them. There are issues about: how comparison groups are constructed; how higher education is defined; how wages, earnings or income are measured[4]; what control variables are included (if any); what type of data is used; etc…In addition to these relatively minor issues, there are more fundamental methodological concerns with the estimates obtainedin the literature, with the majority of results reported not truly estimates of the “return to education”. This is the subject of the next section.
- Methodology
Although the return to education is a standard concept in the economics of education, there is little consistency in the approach used for estimating it. A variety of methods/measures exist, and researchers frequently use them interchangeably without necessarily being clear about what it is they are estimating, or about the limitations of the method they are applying. This section starts by setting out some of the methods used in the literature and ends by outlining the approach used in this paper.
Method 1: “Elaborate” internal rate of return
As with any other investment, the private rate of return to an investment in a given level of education is estimated by finding the rate of discount that equalises the stream of discounted benefits to the stream of costs at a given point in time. In the case of a university education lasting four years (and assuming a working life of 42 years), the formula is:
/ (i)Where (Eu-Es) is the earnings differential between a university graduate (subscript u) and a secondary school graduate (subscript s). The right-hand side represents the opportunity cost of higher education (i.e. 4 years of foregone earnings at the level of what someone with secondary education would have earned). The right-hand side could also be augmented with other costs associated with education such as tuition fees. However, for the sake of simplicity and because most university systems in Africa do not charge fees, these will be left out in the analysis presented in this paper. Although this may not be realistic, it should be remembered that foregone earnings are the largest cost associated with a university education[5]. It is also worth pointing out that, although fees are left out of the calculation, so are grants, scholarships and bursaries which are common in many African countries, as well as any income that students may earn while studying.
The internal rate of return (IRR) as defined above is calculated by building age-earnings profiles for both secondary school and university graduates (assuming no earnings for the first four years in the case of university graduates), differencing the two, and finding the discount rate that results in a 0 (zero) net present value of this netage-earnings profile. In practice, this “elaborate method” is very data intensive as sufficient observations are need to populate each age/qualification cell and construct well-behaved age-earnings profiles.This is rarely the case in African labour force or household surveys[6] and so this method is generally impractical. One way around this problem is to estimate two simple earnings equations (onefor secondary school and one for university graduates) with earnings as the dependent variable, and age and age squared as explanatory variables:
/ (ii)Using the coefficients of these regressions, smooth age-earnings profiles can be built for both secondary school and university graduates[7]. As before, no earnings are assumed for the first four years of those investing in a university education. Using these age-earnings profiles we can then calculate the internal rate of return as before. As in the case of the full/elaborate method, this method is rarely used in the literature.
Method 2: Simple earnings function/Mincerianmethod
In practice, many researchers have made use of the earnings function/Mincerian method (Mincer 1958; 1974) to estimate the returns to education. This is primarily due to its simplicity:
/ (iii)In this equation, experience is frequently replaced by age and other explanatory variables (like sex)are added to the right hand side – although the inclusion of such variables is not innocuous as many (e.g. sector of employment, marital status, number of children, region) may be considered endogenous and so should not be included in the regression[8]. The equation above would be run on a sub-sample of individuals with secondary and tertiary education only, and so the coefficientβ1would identify the effect on earnings of having a higher education qualification[9]. Frequently, the estimate is then converted into an annual return by dividing by the number of years of higher education.
Most research using the Mincerian method interprets the coefficient on HE as the return to higher education, and uses it as a substitute for the IRR calculated through the method described previously. Heckman, Lochner and Todd (2008) criticise this interpretation and argue that many strong assumptions are required to claim that estimates ofβ1 accurately measure the internal rate of return. One obvious difference with the two methods described previously is that the Mincerian equation assumes no loss of working life with additional years of schooling, so that the coefficient β1 is more accurately interpreted as the earnings premium or mark-up associated with a higher education qualification than as a rate of return. Psacharopoulos (1994) uses numerous empirical studies to show that earnings’ functions and the elaborate method yield very similar results – however Heckman, Lochner and Todd (2008) reach a different conclusion, and so do we in this paper.
Method 3: The “short-cut” method
In addition to the methods described above, Psacharopoulos and Patrinos (2004b) describe a “short-cut” method for calculating the rate of return which does not rely on the availability of individual data. The formula used is simply:
/ (iv)Where are the average earnings of those with a university qualification, and the average earnings of someone with just secondary schooling. The denominator represents the opportunity cost of an investment in higher education (i.e. four years of earnings at the level of a secondary school graduate[10]). Although simple, the method has clear drawbacks in that it assumes flat age-earnings profiles and no discounting of earnings that occur later in life.
Accounting for the risk of joblessness
Given that education has both an impact on earnings as well as the likelihood of employment, it is surprising that very few estimates in the literature have factored in the risk of joblessness in estimating the returns to higher education. In countries with low rates of joblessness, this is perhaps understandable, but in Africa, where the level of joblessness frequently ranges between 20% and 60% of the working age population (Figure 1), this would seem a major omission. One contribution of this paper is that we adjust the estimates of the return to higher education by the risk of joblessness. In practice, we calculate the adjusted internal rate of return by generating age-expected earnings profiles. We predict age-earnings profiles using equation (ii) above. In addition, however, we simulate age-employment profiles using a logit regression of the following form:
/ (v)At each age, we then weight the predicted earnings by the predicted likelihood of being in employment at that age in order to derive an age-expected earnings profile. We do this separately for those with secondary and those with tertiary qualifications.
Ability bias
Much of the recent literature on returns to education has been dedicated to tackling ability bias in estimating the return to schooling. As early as 1964, Denison (1964) expressed scepticism about the causal effect of schooling on earnings, arguing that observed differences in earnings between education groups are more likely to reflect inherent ability differences rather than true productivity differentials. Generally, it is believed that the correlation between schooling and earnings obtained through OLSoverstates the true causal effect of education. A standard solution to this problem is instrumental variablesestimation and a large literature has developed using instruments such as the minimum school leaving age, the geographic proximity of schools, etc… Generally, this literature has found that OLS under- rather than overestimates the return to schooling. One potential explanation for this finding advanced by Card (2001) is that these estimates are frequently obtained using structural changes in schooling systems (for example a raising of the school-leaving age) which affect more marginal students for whom the return to schooling might be higher.
In this paper, we do not attempt to tackle the issue of ability bias. Although instrumental variables methods have proved useful in trying to establish the causal effect of schooling on earnings, they suffer from the same drawback as outlined above for the Mincer equation: except under very restrictive assumptions, they do not estimate rates of return to schooling, nor are they designed to (Heckman, Lochner and Todd, 2005). Given that instrumental variable estimates have generally been found to be larger than standard OLS estimates, and that the OLS estimates we obtain are already considerably larger than the ones obtained through our non-parametric IRR method, this omission does not detract from the main arguments made in this paper.
Summary
Before turning to a description of the data, we briefly summarise the methodologies employed in this paper. We estimate the return to higher education using four different methods: (1)simple Mincerian earnings functions as in equation (iii) – with experience replaced by age; (2) the short-cut method (equation (iv)); (3) the IRR using simple earnings functions as in equation (ii); and the IRR adjusted for employment probabilities.
Across all countries and estimation methods we attempt to be consistent in the definition of variables and assumptions used. We define the working age population as those aged 15-64. We restrict our sample to those with tertiary and, where possible, academic secondary education only. Only the latter are included as we aim to generate a control group of individuals who had the potential to go on to university (but this is not always feasible). Tertiary education includes everything from diplomas and degree through to postgraduate qualifications (masters and doctorates)[11]. This was a pragmatic decision in view of the small sample sizes of higher educated individuals encountered in most African household and labour force surveys.
We ignore selection issues, and do not include any additional explanatory variables in our regressions (although we do estimate the returns separately for men and for women). We also ignore taxes, benefits, tuition fees and scholarships. On the other hand, in our IRR calculations, we assume foregone earnings for a period of four years, equivalent to the average earnings for those whose highest qualification is upper secondary education. Finally, our measure of earnings will depend on the survey employed, but we have tried to focus on earningsfrom the primary occupation where possible (i.e. excluding benefits). Also note that, in order to avoid estimation problems induced by outlier values, we trim our dataset by removing the top and bottom percentile of earnings.[12]