Ethnic fractionalization in an African labour market

Journal of Development Economics 68 (2002) 355 – 379 www.elsevier.com/locate/econbase

Ethnic fractionalization in an African labour market Abigail Barr a,*, Abena Oduro b a

Economics Department, Centre for the Study of African Economies, University of Oxford, Manor Road, OX1 3UQ Oxford, UK b Centre for Policy Analysis, Ghana Received 1 July 2000; accepted 1 July 2001

Abstract We find evidence of ethnic fractionalization and its consequences in an African labour market. We consider earnings differentials between members of different ethnic groups and between employers’ relatives, unrelated co-ethnics, and other workers in the Ghanaian manufacturing sector. Variations in a standard set of observed workers’ characteristics explain a large proportion of the earnings differentials between ethnic groups. However, the greatest proportion is explained by the ethnic fractionalization of the labour market combined with variations in employers’ characteristics. Workers who are related to their employers earn a premium and there is statistical discrimination in favour of inexperienced co-ethnic workers. D 2002 Elsevier Science B.V. All rights reserved. JEL classification: J71; O15 Keywords: Labour markets; Wage determination; Ethnicity

1. Introduction Discussions about economic policy between international organisations and African governments rarely touch upon issues relating to ethnicity. And yet recent contributions to the literature on cross-country differences in economic performance indicate that ethnic fractionalization, i.e., the segmentation of a population into several groups that are distinct in terms of language and/or culture, is associated with very high economic costs in terms of lower rates of economic growth due to the adoption of dysfunctional macroeconomic policies (Easterly and Levine, 1997), lower levels of trust, weak norms of civic cooperation

*

Corresponding author. Tel.: +44-1865-281-443; fax: +44-1865-281-447. E-mail address: [email protected] (A. Barr).

0304-3878/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved. PII: S 0 3 0 4 - 3 8 7 8 ( 0 2 ) 0 0 0 1 7 - 2

356

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

(Knack and Keefer, 1997), and increased probabilities of civil war (Collier and Hoeffler, 1998).1 Given these findings, surely the time has come to place the economics of ethnicity on the agenda for policy debate. The two most commonly raised arguments against this are that ethnic diversity is pre-determined and cannot be manipulated by economic policy and that ethnic issues are politically sensitive.2 With respect to the first argument, we suggest that, while levels of ethnic diversity cannot be changed, there may be ways of changing their effect on economic outcomes. A necessary prerequisite for identifying policy interventions that might achieve this objective is a deeper understanding of the role and effects of ethnic diversity at the micro-level. We need to know how and why ethnic identity and ethnic boundaries affect the economic decisions that people make during their everyday lives. Through such an investigation, we may be able to identify the conditions under which the negative effects of ethnic diversity on economic outcomes might be minimised. In addition and with respect to the second argument, by increasing our understanding of why ethnicity matters and, wherever possible, linking it to rational choice, we may start to depoliticise the topic. The following analysis contributes to this effort by seeking to establish whether the ethnic fractionalization that we find to be important at the macro-level also affects economic outcomes at the micro-level in an African labour market. If the ethnicity of a worker affects who employs him or how much he earns then, like the population, the labour market can be described as ethnically fractionalized. We focus on the market for labour in the Ghanaian manufacturing sector. Ghana’s index of ethnolinguistic fractionalization is 0.71, placing it close to the average for sub-Saharan Africa of 0.65. The analysis draws from the literature on the economics of discrimination within labour markets. This literature, with its strong empirical component, provides us with a welldeveloped conceptual framework and a set of tools for identifying, categorising, and quantifying the effects of Ghana’s ethnic diversity on manufacturing workers’ earnings. However, this literature focuses almost exclusively on discrimination against black relative to white workers and women relative to men in OECD labour markets. While most of the models proposed do not rule out the possibility that employers may come from ‘disadvantaged’ as well as ‘advantaged’ groups, throughout the literature and especially its empirical dimension, there is an implicit assumption that employers are predominantly

1 Easterly and Levine (1997) and Collier and Hoeffler (1998) use an index of ethnolinguistic fractionalization (ELF), defined as the probability of two randomly drawn individuals for the same country belonging to different ethnic groups; Knack and Keefer (1997) use a measure of ethnic homogeneity, defined as the proportion of the population belonging to the largest ethnic group. 2 Ethnicity became an important political issue in Africa after independence, as pressure grew for the new leaders to create opportunities for indigenous capital, mediate between conflicting ethnic claims on public resources, and enable lagging groups to catch up to those that had secured early economic advantages (Apter, 1965; Cohen, 1969; Bates, 1974; Rothchild and Oluonsola, 1983). Around this time, a number of African countries including Ghana attempted to promote indigenous African business by introducing regulations that pressured Lebanese and Indian entrepreneurs to vacate trading and small scale services (leaving these for African entrepreneurs) and move their capital into manufacturing. More recently, democratisation has often been accompanied by an increase in the politicisation of ethnicity (Glickman, 1995).

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

357

white and male, i.e., from the advantaged group. In the Ghanaian context, it would be entirely inappropriate to assume that employers come predominantly from one ethnic group. Indeed, our data indicates that the ethnic distributions of employers and employees are very similar. Thus, in our investigation we need to take account both of discrimination between different ethnic groups and of discrimination between own and other ethnic groups. Wherever possible we identify and discern between taste-based discrimination, statistical discrimination, and discriminatory outcomes due to networking. However, our primary aims are to establish whether there is labour market segmentation or factionalization on the basis of ethnicity and identify its effects on earnings. The paper has six sections. Following this introduction, Section 2 contains a brief review of the literature on labour market discrimination with particular reference to segmentation. Then in Section 3, we set out our methodology for identifying and testing various hypotheses about ethnic earnings differentials and fractionalization in the Ghanaian manufacturing sector. In Section 4, we describe our data. We present our results in Section 5 and in Section 6 we draw our conclusions.

2. Review of the literature on labour market discrimination Following Altonji and Blank (1999), we define labour market discrimination as ‘a situation in which persons who provide labour market services and who are equally productive in a physical or material sense are treated unequally in a way that is related to . . . ethnicity’ (p. 3168). We endeavour to distinguish between current labour market discrimination, given predetermined worker characteristics, and the effects of prior discrimination on those characteristics due to either discrimination in other markets (O’Neill, 1990; Maxwell, 1994; Neal and Johnson, 1996) or the effects of past labour market discrimination on how workers from different groups prepare for entry into the labour market (Loury, 1977, 1981; Durlauf, 1996; Benabou, 1994, 1996; Lundberg and Startz, 1998). Discrimination can be motivated in several ways. Becker (1971) focused on the effects of a taste for discrimination. In his model discriminating employers behave as if the price associated with hiring a worker from the less favoured group is their wage plus an additional amount which he calls the ‘coefficient of discrimination’. As a result, workers are segregated with those from the less favoured group being hired by the less prejudiced employers and suffering a wage differential that is determined by the preferences of their most prejudiced employer. So here, discrimination leads to fractionalization and earnings differentials. However, discriminating employers earn lower profits. So in the long run, the effects of discrimination on earnings should disappear. A similarly problematic prediction derives from Becker (1971) model in which the employers’ disutility is associated with placing the less favoured group in a certain occupation with occupational segregation and a short run earnings differential as the outcome. Coate and Loury (1993a) present an alternative model in which all employers have the same preferences, thereby removing the tendency for the earnings gap to disappear in the long run. This tendency can also be eliminated by introducing imperfect information in the form of search costs and thereby rendering segregation costly (Borjas and Bronars, 1989; Black, 1995; Bowlus and

358

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Eckstein, 1998). These models do not predict segregation unless it is the employers and not the workers who are conducting the search (Bowlus and Eckstein, 1998). In the Ghanaian context, a taste for discrimination could lead to fractionalization and a premium for workers employed by members of their own ethnic group. However, given that no particular group dominates the role of employer, we do not expect discrimination motivated by taste to lead directly to a wage premium for any particular group.3 This notwithstanding, to the extent that (1) a taste for discrimination in favour of co-ethnics leads to fractionalization and (2) discrimination in credit and other markets leads to variations in the labour demand curves of different types of employer, we may observe earnings differentials between groups. Imperfect information also provides the foundations for models of statistical discrimination. Building on the pioneering work of Phelps (1972) and Arrow (1973) this literature explores the consequences of firms having limited information about the skills and reliability of job applicants, especially young and inexperienced ones, and therefore using correlated and easily observable characteristics such as ethnicity to discriminate between them. One particular finding in this literature is that, due to feedback via, for example, investments in human capital, biased stereotypes might be self confirming (Arrow, 1973; Coate and Loury, 1993b). In the Ghanaian context, statistical discrimination of this form could be leading to both current labour market discrimination and feedback effects and consequent earnings differentials between members of different ethnic groups. Another form of statistical discrimination may contribute to fractionalization and affect the earnings of co-ethnic and related workers. Aigner and Cain (1977) and subsequently Lundberg and Startz (1983) and Lundberg (1991) have explored the effects of group differences in the precision of the information that employers have about individual productivity when that productivity depends on the quality of the match between worker skills and the requirements of the job. Those groups for which more precise information is available will earn a premium. However, if employers learn as workers gain more exposure to the market and if there is no underlying difference in productivity, the premium will be eroded by worker experience. In the literature, this form of discrimination may also lead to ex post differences in productivity across groups. However, this does not apply where the favoured workers are defined by their sameness to their employers rather than by some dimension of individual identity. There is a close conceptual link between this work on statistical discrimination and Montgomery (1991) work on the effect of social networks on labour market outcomes. In both cases, some dimension of social structure is associated with a variation in the amount or accuracy of the information available to employers about prospective employees and vice versa. In the former, the agents have better information about others with whom they share a particular aspect of social identity which may be linked to language, culture, or some other determinant of cognition. In the latter, agents have better information about others with whom they share a social connection. Building on Granovetter (1973) and

3 To the extent that all employers, regardless of their own ethnicity, prefer to employ any particular ethnic group it is more likely to be due to some form of statistical discrimination.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

359

Rees and Schultz (1970), Montgomery (1991) shows that if social networks are important for this reason, ‘workers who are well connected might fare better than poorly connected workers’ (p. 1408). Arrow (1998) makes the connection between this literature and racial discrimination. He also cites Kranton and Minehart (2001), who show that a sufficiently dense network will mimic a perfect market, and argues that evidence of statistical discrimination should be viewed as evidence that networks are both important and imperfect in the sense that they are not sufficiently dense.

3. Methodology We start our investigation in Section 4 with a casual look at who employs who. In particular, we look for evidence of a tendency to employ co-ethnics. However, members of some groups predominating in particular occupations and sectors might also be taken as evidence of market fractionalization. Following this, we move to an analysis of earnings. 3.1. Investigating the variations in earnings between ethnic groups The extent of the variation in earnings between ethnic groups can be established by estimating the following equation lnwi ¼ a0 þ a1 ei þ n1i

ð1Þ

where lnwi is the log of earnings for worker i, a0 is a constant term, ei is a vector of dummies, one corresponding to each ethnic group represented in the sample, a1 is the vector of coefficients associated with those ethnic dummies, and n1i is the error term. The joint significance of a1 tells us whether there is variation across the ethnic groups, the signs and significance of specific elements of a1 tell us whether particular groups earn significantly more or less than the group chosen as a basis for comparison, and the signs and significance of differences between the elements of a1 provide us with similar information about other pairwise comparisons. In an effort to establish how much of the identified earnings differentials are due to variations in predetermined personal characteristics, we then add a vector of worker personal characteristics, xi, to the function lnwi ¼ a0 þ a1 ei þ a2 xi þ n2i :

ð2Þ

Further, to investigate whether the returns associated with the personal characteristics vary across ethnic groups, we introduce a series of interaction terms between ei and xi, lnwi ¼ a0 þ a1 ei þ a2 xi þ a3 ei xi þ n3i :

ð3Þ

Traditionally, the significance of a1 and a3 are interpreted as evidence of current labour market discrimination, while a2xi is assumed to be absorbing the effects of

360

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

variations in personal characteristics, some of which may be due to pre-market discrimination. However, we must be aware of potential omitted variable bias. Omissions of particular concern include controls for innate ability, school quality, worker preferences and comparative advantages.4 In accordance with the literature on labour market discrimination, we use Eqs. (2) and (3) as a basis for our conclusions about whether discrimination is causing ethnic earnings differentials in the Ghanaian manufacturing sector. We then build on Eq. (2) in our efforts to identify the form that this discrimination takes and the role played by fractionalization. Recall that while a taste for discrimination is unlikely to lead directly to earnings differentials between ethnic groups, a taste for employing co-ethnics combined with discrimination in other markets could lead to fractionalization and, as a consequence, to earnings differentials.5 To test whether this is indeed the case, and establish the extent to which ethnic earnings differentials are the result of such a mechanism, we introduce a vector of dummies, gi, corresponding to the ethnicity of the workers’ employers into the earnings function lnwi ¼ a0 þ a1 ei þ a2 xi þ a4 gi þ n4i :

ð4Þ

A significant vector of coefficients, a4, combined with declines in the magnitude and significance of elements of a1 would indicate that ethnic fractionalization due primarily to the employment of co-ethnics is a source earnings differentials. We can take this line of analysis one step further by introducing another vector of employer’s characteristics, hi, which includes variables such as size and capital – labour ratio into the function, lnwi ¼ a0 þ a1 ei þ a2 xi þ a4 gi þ a5 hi þ n5i :

ð5Þ

To the extent that the inclusion of hi reduces the significance of a4, we gain some indication as to why employers from different ethnic groups pay differently. Further, workers may be segregated not only on the basis of co-ethnicity with their employer. Some ethnic groups may be preferentially employed by larger or more capital intensive enterprises, or by public or foreign owned enterprises. If this is the case, the introduction of hi will cause declines in the magnitude and significance of a1.6 Our data set is unusual in that it contains both worker and employer characteristics. However, the range of employer characteristics is limited. Thus, in order to fully account 4 Variations in preferences for particular job characteristics across ethnic groups could provide an alternative explanation for both earnings differentials and segregation. It is, however, encouraging to note that variations in preferences have been less the concern of those interested in black – white differentials than those focusing on male – female differentials. Similarly, the discussion about variations in comparative advantages between groups has primarily been limited to male – female comparisons (e.g. Becker, 1971). 5 Fafchamps (in press) shows that members of different ethnic groups have differential access to suppliers credit, although his analysis focuses on the distinction between African and non-African entrepreneurs rather than finer distinctions between African entrepreneurs from different ethnic groups. 6 Even in the US, where we might expect labour markets to function better, employer characteristics such as sector and size are found to be important determinants of earnings (e.g., Krueger and Summers, 1988; Brown and Medoff, 1989).

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

361

for fractionalization, we also estimate a version of Eq. (5) in which gi and hi are replaced by employer’s fixed effects, di (fixed across workers not time) lnwi ¼ a0 þ a1 ei þ a2 xi þ a6 di þ n6i :

ð6Þ

By comparing a1 in this specification with a1 in Eq. (2), we can gain some insight into the full effect of between-enterprise fractionalization on earnings. Note, however, that the employers’ fixed effects will not pick up the effect of fractionalization within enterprises with respect to occupational allocations. To explore this type of fractionalization we, rather circumspectly, introduce a vector of occupational dummies, oi lnwi ¼ a0 þ a1 ei þ a2 xi þ a6 di þ a7 oi þ n7i

ð7Þ

and once again, monitor the effect on a1. Significant elements in a7 combined with a reduction in the magnitude and significance of elements in a1, could indicate that there is job crowding for some ethnic groups. However, it could also indicate that the observed personal characteristics previously entered into the earnings function are failing to pick up important aspects of human capital, which are correlated with job type. In an attempt to establish whether observed earnings differentials are due to statistical discrimination, we employ a test devised by Altonji and Pierrot (2001). The test assumes that employers learn about workers as the latter’s exposure to the labour market increases and that as the employers learn, workers’ pay becomes more dependent on actual productivity and less dependent on easily observable characteristics such as ethnicity. Thus, in a wage equation that contains interactions between experience and both ethnicity and a hard-to-observe variable that is correlated with productivity, the coefficient on the former will be such that ethnic earnings differentials decline with experience, while the coefficient on the latter will be such that the effect of the hard-to-observe characteristic increases with experience. Adopting this approach, taking mother’s years of education as our hard-to-observe variable, and using Eqs. (2) and (6) as alternative base functions, we arrive at the following two empirical formulations lnwi ¼ a0 þ a1 ei þ a2 xi þ a8a ei f ðki Þ þ a9a mi f ðki Þ þ n8ai ,

ð8aÞ

lnwi ¼ a0 þ a1 ei þ a2 xi þ a6 di þ a8b ei f ðki Þ þ a9b mi f ðki Þ þ n8bi

ð8bÞ

and where ki is years of experience, f(ki) is the experience profile of earnings, and mi is mother’s years of education. 3.2. Investigating the variations in earnings between employers’ kin, co-ethnics, and other workers In order to establish whether employers’ relatives and co-ethnics earn more than other workers, we estimate the following lnwi ¼ b0 þ b1 ri þ b2 ci þ b3 zi þ b6 di þ n9i

ð9Þ

where ri is a dummy that takes the value of 1 if the worker is related to the employer, ci is a dummy that takes the value of 1 if the worker is from the same ethnic group as the

362

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

employer, and zi is a vector of other worker characteristics, i.e., it is the combination of ei and xi. Significant and positive coefficients on ri and ci indicate that relatives and coethnics, respectively, receive positive earnings premiums relative to other workers. Co-ethnicity is defined with respect to shared ethnic identity, while relatedness is defined with reference to a known social linkage. While co-ethnic and even non-co-ethnic workers may have a social linkage with their employer, it is only relatives that definitely have such a linkage. Thus, a significant, positive coefficient on ri should be taken as evidence of a network effect associated either with superior information relating to job matching or a more sustained productivity effect due perhaps to reduced moral hazard or greater esprit de corps (Clague, 1993). In contrast, a significant, positive coefficient on ci should be taken as evidence of either a taste for discrimination in favour of co-ethnics or statistical discrimination based on shared ethnic identity. To establish whether any identified earnings differentials operate through occupational allocations, we again circumspectly introduce a vector of occupational dummies, oi lnwi ¼ b0 þ b1 ri þb2 ci þ b3 zi þ b6 di þ b7 oi þ n10i

ð10Þ

and monitor the effect on b1 and b2. In an endeavour to establish whether any identified earnings premiums are due to statistical discrimination or a similar networking effect, we adapt Altonji and Pierrot (2001) approach once again. In this context, we should not expect to find that hard-toobserve characteristics grow in importance with experience. This is because such characteristics will be more easily observed from the outset by related and co-ethnic employers. For this reason, we introduce only the interaction between experience and the relationship dummies lnwi ¼ b0 þ b1 ri þ b2 ci þ b3 zi þ b6 di þ b8 ri f ðki Þ þ b9 ci f ðki Þ þ n11i :

ð11Þ

If b8 and/or b9 are such that the effects of being related or co-ethnic with one’s employer decline with experience, then we may conclude that the source of any earnings premium afforded to these groups is due to statistical discrimination or a similar networking effect. If either or both premiums do not decline with experience, then we must look for other explanations such as a taste for discrimination or a productivity effect.

4. Data Our data is drawn from the fifth wave of the Ghanaian Manufacturing Enterprise Survey (GMES).7 The sample of enterprises is drawn from four cities in southern Ghana. Each of these cities could be described as a potential melting pot, i.e., as an environment within which Gluckman (1961) would have expected to see ethnicity decline in importance over 7 The first three waves of this survey were conducted as part of the World Bank’s Regional Program for Enterprise Development. The last two were conducted as part of a project on labour markets in sub-Saharan Africa. All five waves were funded by the Department for International Development and conducted by the Centre for the Study of African Economies in collaboration with the University of Ghana and the Ghana Statistical Service.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

363

time. Approximately one third of the enterprises are in Kumasi, a city to the north – west of the capital, Accra. Less than 5% of the sample are in either Cape Coast or Takoradi, on the coast to the west of Accra. All the remaining enterprises in the sample are situated in Accra. For each enterprise, there is a corresponding sample of up to 10 waged workers. In the fifth wave of the GMES, the questionnaires for the entrepreneurs, defined to include owner –

Table 1 Ethnic composition of worker and employer Samples

Akan Asante Fante Akyem Akuapem Kwahu Brong Ahanta Wassa Nzema Assin Denkyira Guan, unspec. Guan, Efutu Guan, Gonja Guan, Buem Guan, Larteh Ga-Adangbe Ga Krobo Ada Ewe Northern Grussie Dagaaba Dagomba Busanga Frafra Kanjaga Kasena Kusasi Mamprusi Wala Builsa Hausa Non-Ghanaian Middle Eastern Asian European Total

Workers

Employers

Freq.

Percent

Freq.

Percent

218 245 50 33 29 15 12 11 10 2 1 13 3 2 1

20.86 23.44 4.78 3.16 2.78 1.44 1.15 1.05 0.96 0.19 0.10 1.24 0.29 0.19 0.10

51 32 6 4 9 5 1 1

26.98 16.93 3.17 2.12 4.76 2.65 0.53 0.53

1

0.53

1

0.53

18 2 1 28

9.52 1.06 0.53 14.81

1 1

0.53 0.53

3

1.59

11 9 4 189

5.82 4.76 2.12 100.00

128 15 14 182

12.25 1.44 1.34 17.42

12 11 10 5 5 2 2 2 1 1 3 7

1.15 1.05 0.96 0.48 0.48 0.19 0.19 0.19 0.10 0.10 0.29 0.67

1045

100.00

364

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Table 2 Ethnic composition of workers and employers tobe used in the analysis Workers

Asante Fante Other Akan Ga-Adangbe Ewe Northern Non-Ghanaian Total

Employers

Freq.

Percent

Freq.

Percent

218 245 182 157 182 61

20.86 23.44 17.42 15.02 17.42 5.84

1045

100.00

51 32 28 21 28 5 24 189

26.98 16.93 14.81 11.11 14.81 2.65 12.70 100.00

managers and general managers or managing directors of corporate enterprises, and workers contained questions about ethnic identity and the incidence of blood relations between workers and their employers. The ethnic structure of the Ghanaian population is complex. There are over 100 distinct ethnic groups some of which combine to make up larger groups. The Akan, for example, is made up of around 20 groups, including the Asante, the Fante, the Akyem, the Akuapem, the Kwahu, and the Brong. Many of the ethnic groups have distinct languages. Others, while sharing their language consider themselves to be distinct for cultural or historical reasons. Our approach during the survey was to ask each entrepreneur and worker which ethnic group they were from. Coding then took place after the fieldwork was complete. Thus, our data captures the ethnic identities that the individuals ascribe to themselves. It is worth noting that none of the respondents had any difficulty deciding on the ethnic group to which they belonged. In the case of corporate enterprises, if the general manager or managing director was not available, although the questions relating to the enterprises accounts and operations were asked of other managers, the ethnicity questions were not asked. As a result, some observations had to be dropped from this analysis. The data required for this analysis was collected from a sample of 1045 workers (see Table 1). A total of 31 Ghanaian ethnic groups are represented in this sample along with the Hausa, a West African group that has been present in Ghana for several generations. We describe the Hausa as Ghanaian throughout the analysis below. For the purposes of analysis, the workers are allocated to six ethnic categorisations (see Table 2): the Asante, a sub-group of the Akan; the Fante, another sub-group of the Akan; Other Akan; the Ga and Adangbe, which are combined throughout and referred to as the Ga-Adangbe as a reminder; the Ewe; and Northern, which includes the Housa as well as eleven groups indigenous to Ghana. Tables 1 and 2 are constructed in such a way that it is easy for the reader to see how the allocations are done.8

8 The Guan, included under Other Akan, was the only sub-group that proved difficult to classify. The Guan is divided into several smaller groups some of which consider themselves to be Akan and some not. Of the groups represented in our sample, the Gonja is the only group to which this applies. They have been included in Other Akan, but might be more appropriately classified as Northern. As only two employees are Gonja, their classification is unlikely to affect the results of the analysis.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

365

The samples of workers is spread across 189 employers. The distribution of these employers with respect to ethnicity is presented in disaggregated form in the final column of Table 1 and in the aggregated form to be used in the analysis in the final column of Table 2. Note that, while we have excluded non-Ghanaians from the sample of workers, 12.7% of the employer sample is Middle Eastern, Asian, or European.9 Excluding these employers from the analysis would greatly reduce the proportion of larger enterprises in our sample. Note that, as we mentioned above, the distributions of workers and employers across the six Ghanaian ethnic categorisations are similar. In both samples, the Asante make up the largest proportion (21% and 26%, respectively), with the Fante as second largest (between 23% and 17%, respectively). The Other Akan, Ga-Adangbe and Ewe groups assume quite similar proportions (between 11% and 17%), while only the Northern group accounts for less than 10% in each sample. In part, this ethnic distribution reflects the geographical focus of the survey. Kumasi is the capital of the Asante region, while Cape Coast and Takoradi are the two largest towns in the Fante region. A relatively small ethnic group, the Ga, are indigenous to the capital, Accra, while the Adangbe traditionally occupy an area to the east of Accra. The Other Akan groups come from the area to the north of Accra and surrounding Kumasi. The Ewe are from the south-eastern part of the country, but have been present in Accra and Kumasi for several generations. Being the largest and most industrialized, these cities have also been a focus for migrants from the North. Our vector of personal characteristics includes years of education and its square, potential years of experience (age-education—6) and its square.10 We also include a gender dummy, a dummy indicating whether the worker is married, two location dummies, and mother’s years of schooling as a control for family background.11 The dummy indicating whether the worker is a relative of his/her employer is constructed with reference to a direct question posed to the workers. The dummy indicating whether the worker is a member of the same ethnic group as his/her employer is based on a coincidence in the detailed ethnic classification. Thus, for example, if a Kwahu worker has an Akyem employer, even though they are both classified as ‘Other Akan’ the co-ethnic dummy will take a value of zero. This dummy has been defined to exclude relatives. So, in effect we have three groups of workers, relatives of the employers, unrelated co-ethnics, and others. Our vector of employer characteristic includes the log of the total number of workers they employ, the log of their capital –labour ratio, eight sub-sector dummies, and two dummies indicating some public ownership and some foreign ownership. Three occupational dummies are introduced into Eq. (7), one for managerial positions, one for clerical 9 Data was collected for 23 non-Ghanaian employees. This is an insufficient number from which to draw conclusions. Thus, they have been excluded from the analysis. 10 When testing for statistical discrimination by interacting experience with ethnicity, the most natural choice of experience variable would be tenure, i.e., the number of years during which the current employer has observed the employee. We rejected this measure for two reasons: first, it is highly likely to be endogenous, and second, we wish to be consistent with Altonji and Pierrot (2001). Potential experience does not seem such an unnatural choice, if one thinks of it as the years for which an employee is expected to account and provide references. 11 Mother’s and father’s education were found to be highly correlated and there were more missing observations in the latter. This may reflect the fact that the majority of the employees in our sample come from matrilineal groups and that in several of the groups, both matrilineal and patrilineal, the father does not traditionally live with the children and their mother.

366

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Table 3 Worker’s characteristics by ethnic group Other Akan Asante

Fante

Ga-Adangbe Ewe

Northern All

Sample size 182 Employee – employer relationship Related to employer (%) 7.69 Same ethnic group (%) 7.69

218

245

157

182

61

1045

14.22 46.33

13.47 20.41

7.64 10.19

12.64 21.98

3.28 8.20

11.00 21.63

Ethnicity of employer Other Akan (%) Asante (%) Fante (%) Ga-Adangbe (%) Ewe (%) Northern (%) Non-Ghanaian (%)

32.97 11.54 17.03 8.79 5.49 0.55 23.63

9.63 59.63 10.09 3.21 1.38 1.83 14.22

19.18 15.10 33.06 10.61 6.94 1.63 13.47

13.38 13.38 12.74 19.75 15.29 2.55 22.93

14.84 11.54 10.99 9.34 34.07 0.55 18.68

14.75 18.03 13.11 3.28 4.92 21.31 24.59

17.7 23.06 17.42 9.47 11.39 2.58 18.37

Worker’s characteristics Monthly earnings Years of education Years of potential experience Female (%) Married (%) Mother’s years of ed.

265,064 11.64 20.89 18.13 68.68 3.56

204,610 11.39 17.86 15.14 65.60 4.41

248,159 10.80 20.67 16.73 65.31 3.73

268,364 11.83 20.04 21.66 75.80 5.01

207,416 11.08 18.45 17.58 65.93 4.23

158,491 6.72 25.20 21.31 72.13 1.61

232,724 11.04 19.90 17.80 68.04 4.00

Occupation Management (%) Services (%) Skilled production (%) Unskilled (%)

15.93 19.23 53.30 11.54

13.30 16.51 57.34 12.84

11.48 20.49 54.10 13.93

12.82 17.31 58.33 11.54

5.49 16.48 64.29 13.74

3.28 13.11 50.82 32.79

11.32 17.84 56.85 14.00

Sector Food (%) Feeds and beverages (%) Furniture (%) Garments (%) Machinery (%) Other metal (%) Plastics and chemicals (%) Textiles (%) Wood processing (%)

16.48 6.59 12.64 11.54 3.30 23.63 4.95 6.59 14.29

14.22 5.96 22.02 8.72 0.46 12.84 9.17 1.83 24.77

26.12 1.63 15.51 6.94 2.45 18.78 5.31 2.45 20.82

29.30 9.55 14.65 8.28 1.27 14.65 10.19 8.28 3.82

12.64 6.04 18.68 12.09 7.69 23.08 6.04 7.14 6.59

24.59 3.28 3.28 6.56 11.48 19.67 4.92 1.64 24.59

20.00 5.45 16.08 9.19 3.44 18.56 6.89 4.69 15.69

Employer characteristics Capital labour ratio No. of employees Foreign ownership (%) State ownership (%)

3.4E + 07 136.66 32.97 9.34

2.3E + 07 89.49 22.48 5.96

2.5E + 07 132.20 30.20 5.31

4.2E + 07 153.97 37.58 12.74

3.7E + 07 99.57 30.22 5.49

2.0E + 07 108.34 29.51 9.84

3.0E + 07 120.26 30.14 7.56

Location Accra (%) Kumasi (%) Coast (%) Non-Ghanaian (%)

64.84 13.19 21.98 23.63

28.44 68.81 2.75 14.22

61.63 11.84 26.53 13.47

88.54 6.37 5.10 22.93

79.67 15.38 4.95 18.68

44.26 44.26 11.48 24.59

61.44 25.65 12.92 18.37

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

367

and sales personnel, and one for unskilled production workers. The basis for comparison is skilled production workers. The classification of employees into occupations was conducted during the interviews with workers and, in many cases, may be quite arbitrary. The size of the enterprise may have affected the number and type of occupational categories used in the process and it is likely that some classifications were made at least partially with reference to the pay the worker received. Thus, the estimated coefficients for Eq. (7) may be subject to endogeneity bias and must be treated with considerable caution. The relationships between workers and employees and the distribution of workers across employers with respect to ethnic origin are explored in Table 3. Eleven percent of workers are employed by a relative and a further 23% are employed by a non-related member of the same ethnic group. Further, workers from every ethnic group are more likely to work for a member of their own ethnic group than for a member of another Ghanaian ethnic group. According to a Chi-squared test, the hypothesis that employees are distributed between Ghanaian employers without regard for ethnicity is rejected at the 0.1% significance level. The pattern is disturbed if we introduce non-Ghanaian employers into the analysis, as Ga-Adangbe and Northerners are more likely to be working for them than for their co-ethnics. Table 3 also contains the mean monthly earnings (in Cedi) and the means of several other variables for each ethnic group. There is considerable variation between groups. On average, the Other Akan (used as a basis for comparison throughout the analysis) earn more than all the other groups with the exception of the Ga-Adangbe with whom they are on a par. They earn 7% more than Fante, 28% more than Ewe, 30% more than Asante, and

Fig. 1. Employees’ years of education by ethnic group.

368

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

67% more than Northerners. Turning to the mean years of education for each group, note that Ga-Adangbe are the most educated with the Other Akan a close second, while on average Northerners have far fewer years of education than any other group. This distinction between Northerners and the other groups is also evident in the frequency distributions for years of education contained in Fig. 1. Here, we can see that the lower average years of education for Northerners is due to the high proportion who have no education.12 This is the only strikingly obvious difference between Northerners’ personal characteristics and those of other ethnic groups.

5. Analysis of earnings 5.1. Investigating the variations in earnings between ethnic groups Specification (1) (first column, Table 4) indicates that Asante earn 0.28 less than Other Akan in terms of log earnings, Fante 0.21 less, Ewe 0.25 less, and Northerners 0.45 less. All of these differences are significant at the 1% level. Further, pairwise comparisons of coefficients indicate that the Ga-Adangbe earn significantly more than all other groups except the Other Akan ( p-values of 0.002, 0.021, 0.006, and 0.001, when compared with the Asante, Fante, Ewe, and Northerners, respectively) and Northerners earn significantly less than all other groups ( p-values of 0.059, 0.012, and 0.036 when compared with Asante, Fante, and Ewe, respectively). In specification (2), we control for a series of worker characteristics. Years of education and experience both enter the earnings function in quadratic form, with the education profile of earnings being convex and the experience profile being concave.13 These relationships are quite stable across all the specifications presented in Table 4. In contrast, the significantly negative premium associated with being female and the positive effect of mothers’ education disappear once we fully control for employer characteristics suggesting that there is sorting between employers on the basis of both gender and family background. Finally, workers in Cape Coast and Takoradi earn significantly less than those in the capital city even after we control for several other employer characteristics, while those in Kumasi earn less primarily because of the smaller size of the enterprises sampled there. Introducing these worker characteristics into the earnings function considerably reduces both the magnitude and significance of the log earnings differentials. The only remaining significant differentials are those between the Other Akan and the Asante, Fante and Ewe. All the rest are not statistically distinguishable from zero at the 10% level. In specification (3) (reported separately in Table 5), we introduce interaction terms between the five ethnic dummies and each of the worker characteristics. We then conduct an F-test for the set of interaction terms corresponding to each characteristic. Only the Fstatistic relating to the interactions between ethnicity and being located in Kumasi was

12

Years in Koranic school are not taken into account when calculating years of education. This is consistent with the findings of Bigsten et al. (1998) that higher levels of education are associated with increasingly higher returns in the manufacturing sectors of several sub-Saharan countries. 13

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

369

significant (5% level). This result is driven by the earnings differential between the Other Akan on the one hand and the Asante and Ewe on the other in that location. However, both Ewe and Other Akan in Kumasi each account for under 3% of our sample. Based on these results, we conclude that there is no significant variation in the returns associated with personal characteristics between ethnic groups and favour specification (2) as a basis for our further investigations. Thus, in Table 6 we use specification (2) to conduct a decomposition of the average ethnic log-earnings differentials.14 We focus on the differentials between the top earning group, Other Akan and the four lowest earning groups, Fante, Ewe, Asante, and Northern. The top half of the table shows the portions of these differentials that are explained by variations in education, experience, other employee characteristics (gender, marital status, and mother’s education) and the portion that remains unexplained by the model. Differences in years of education explain a large portion of the differentials between the Other Akan and the Fante (36%), Ewe (30%) and Northern (78%) groups. In the comparison of the Asante and Other Akan location plays an important role (22%). In general, however, the unexplained portions remain large (between 21% and 67%). Specification (4) (Table 4) contains a set of six dummies corresponding to the ethnicity of the employers. These indicate that in the absence of any other controls relating to the employers’ characteristics, Fante and Non-Ghanaian employers pay significantly more than Other Akan, while Ewe employers pay significantly less. Pairwise comparisons of coefficients indicate that Fante also pay significantly more than Asante, Ga-Adangbe, and Ewe employers ( p-values of 0.072, 0.015, and 0.014, respectively) and Ewe also pay significantly less than Asante, Ga-Adangbe, and Northerners ( p-values of 0.002, 0.014, and 0.010, respectively). Taking account of employer ethnicity eliminates the unexplained earnings differential between Other Akan and Ewe and reduces the size and significance of those between Other Akan and Asante and Fante. Controlling for other employer characteristics as in specification (5) reduces the size and significance of the employers’ ethnic dummies, but not those of the workers. Indeed, the differential between Other Akan and Northerners has grown in size and become significant. This differential remains large and significant when employer-fixed-effects are introduced (specification (6)). Fully controlling for employer characteristics brings out two other significant differentials—Northerners also earn significantly less than the Ewe and the Fante ( p-values of 0.025 and 0.089, respectively). The importance of the employerfixed-effects is clearly demonstrated in the decomposition presented in the bottom half of Table 6. The fixed-effects explain 45%, 56%, and 50% of the differentials between the Other Akan and the Fante, Ewe, and Asante, respectively. However, they explain only 20% of the differential between Other Akan and Northern workers. Finally, note the considerable reduction in the unexplained portion of the differentials once fixed effects are included. Introducing occupational dummies renders the differential between Northerners and all other ethnic groups except the Ewe insignificant. Northerners’ lower earnings could be

14 Having found that any variations in coefficients across ethnic groups were not significant, we decided not to take them into account in the decomposition. Thus, our decomposition is simpler than Oaxaca (1973).

370

Table 4 Earnings and ethnic identity (dependent variable = log of monthly earnings in Cedi, n = 1045) (1) 12.1932  0.2754  0.2088  0.0179  0.2513  0.4528

0.0282

0.0552*** 0.0748*** 0.0768*** 0.0838 0.0777*** 0.0963***

10.8467  0.1507  0.1278  0.0917  0.1675  0.0965 0.0156 0.0043 0.0390  0.0005  0.1332 0.0855 0.0068  0.1575  0.1430

0.3670

(4) 0.1244*** 0.0675** 0.0626** 0.0697 0.0646** 0.0894 0.0138 0.0006*** 0.0064*** 0.0001*** 0.0559** 0.0539 0.0044 0.0490*** 0.0658**

10.8194  0.1338  0.1222  0.0430  0.0598  0.1199 0.0155 0.0041 0.0373  0.0005  0.1415 0.0762 0.0080  0.1538  0.1339 0.0033 0.1362  0.0534  0.2883 0.1246 0.2292

0.4011

(5) 0.1271*** 0.0685 * 0.0627 * 0.0658 0.0665 0.0913 0.0138 0.0006*** 0.0062*** 0.0001*** 0.0552** 0.0529 0.0043 * 0.0559*** 0.0621** 0.0720 0.0620** 0.0768 0.0796*** 0.1498 0.0590***

10.5265  0.1311  0.1042  0.0516  0.0664  0.1980 0.0029 0.0037 0.0235  0.0002  0.0860 0.0345 0.0090  0.0513  0.2819  0.0571 0.0760 0.0289  0.1775  0.0462 0.0050  0.0954 0.1050  0.0030 0.2074

0.4965

(6) 0.2171*** 0.0617** 0.0569 * 0.0608 0.0592 0.0849** 0.0135 0.0006*** 0.0060*** 0.0001** 0.0511 * 0.0496 0.0040** 0.0631 0.0718*** 0.0648 0.0587 0.0795 0.0807** 0.1525 0.0645 0.0501 * 0.0702 0.0131 0.0211***

10.9553  0.0652  0.0596  0.0915  0.0099  0.1920  0.0118 0.0043 0.0371  0.0004  0.0248 0.0213 0.0055

0.7066

(7) 0.1823*** 0.0610 0.0492 0.0610 0.0533 0.0818** 0.0131 0.0006*** 0.0062*** 0.0001*** 0.0492 0.0415 0.0034

11.2652  0.0382  0.0373  0.0418 0.0574  0.1108  0.0087 0.0024 0.0272  0.0003 0.0162 0.0123 0.0043

0.1482*** 0.0529 0.0432 0.0531 0.0479 0.0727 0.0125 0.0006*** 0.0057*** 0.0001*** 0.0471 0.0374 0.0031

0.6454  0.3301  0.0329 0.7673

0.0640*** 0.0427*** 0.0470

Standard errors corrected for heteroskedasticity using White’s (1980) procedure. Specification (5) includes eight sector dummies and specifications (6) and (7) include employer fixed effects. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Constant Asante Fante Ga-Adangbe Ewe Northern Years of education Years of ed. sq. Years of experience Years of experience sq. Female Married Mother’s years of ed. Kumasi Coast E-Asante E-Fante E-Ga-Adangbe E-Ewe E-Northern E-Non-Ghanaian Foreign owned State owned Capital-labour ratio No. of employees Management Unskilled Clerical/Sales R-squared

(2)

Base group Other Akan

Interactions with Asante

Interactions with Fante

Interactions with Ga-Adangbe

Interactions with Ewe

Interactions with Northern

F-test on all interaction terms

Coef.

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

p-Value

0.0963 0.0019 0.0002  0.0212 0.0005 0.1751 0.0926  0.0098  0.3600  0.2487

0.4467 0.0562 0.0022 0.0212 0.0004 0.1744 0.1779 0.0134 0.1694** 0.2365

 0.1221 0.0087 0.0005  0.0248 0.0004 0.0293 0.1298  0.0020 0.0835  0.0245

0.3772 0.0523 0.0021 0.0199 0.0003 0.1921 0.1862 0.0151 0.1733 0.1687

0.2814 0.0044  0.0006  0.0389 0.0007 0.1443 0.2349  0.0154  0.1795  0.0963

0.5010 0.0652 0.0024 0.0262 0.0005 0.2090 0.1948 0.0158 0.2088 0.2561

 0.3416 0.0421  0.0011  0.0069 0.0001 0.2729  0.0160  0.0042  0.3657  0.0819

0.4005 0.0506 0.0020 0.0247 0.0004 0.1898 0.2166 0.0150 0.1775** 0.1764

0.5668  0.0020  0.0003  0.0446 0.0006 0.0677  0.1227  0.0270  0.0155 0.1019

0.5098 0.0566 0.0026 0.0275 0.0004 0.1838 0.2575 0.0208 0.1912 0.3191

0.4580 0.8662 0.9566 0.5133 0.5121 0.7093 0.6822 0.7676 0.0107 0.9063

S.E.

Constant 10.7931 0.3105*** Years of education 0.0009 0.0446 Years of ed. sq. 0.0046 0.0016*** Years of experience 0.0568 0.0158*** Years of experience sq.  0.0008 0.0003*** Female  0.2365 0.1328* Married 0.0146 0.1432 Mother’s years of ed. 0.0134 0.0104 Kumasi 0.0040 0.1365 Coast  0.0846 0.1239

R2 = 0.3877. In the row marked ‘Constant’ and the columns headed ‘Interactions with Asante’, etc., we present the coefficients on the ethnic dummies. In those columns and in later rows are the coefficients on the interactions between the ethnic dummies and each of the personal characteristics listed. Standard errors corrected for heteroskedasticity using White’s (1980) procedure. * Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Table 5 Earnings and ethnic identity, specification (3) (dependent variable = log of monthly earnings in Cedi, n = 1045)

371

372

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Table 6 Decomposition of average ethnic earnings differentials (Basis for comparison = Other Akan)

Sample size Differential in log earnings

Fante

Ewe

Asante

Northern

245  0.21

182  0.25

218  0.28

61  0.45

Analysis of log-earnings differentials based on specification (2) Share of differential due to differences across ethnic groups in. . . education (%) 36 30 experience (%) 1 13 other employees’ characteristics (%) 0 1 location (%) 2 8 Explained (%) 39 33 Unexplained (%) 61 67 Analysis of log-earnings differentials based on specification (6) Share of differential due to differences across ethnic groups in. . . education (%) 25 24 experience (%) 1 17 other employees’ characteristics (%) 0 1 location firm fixed effects (%) 45 56 Explained (%) 71 96 Unexplained (%) 29 4

11 14 3 22 45 55

78  10 3 7 79 21

9 19 2

49  14 2

50 76 24

20 58 42

partially due to occupational segregation. The Ewe are further distinguished once we introduce occupational dummies-Ewes now earn significantly more than Asante ( p-value of 0.080) and Fante ( p-value of 0.058) as well as Northerners ( p-value of 0.019). We interpret this unusual result for Ewe workers as evidence that they are occupationally classified in a different way to other groups. In particular, we suspect that there are some who are performing managerial tasks and being paid accordingly, but who are nevertheless classified as skilled production workers. Note from Table 3 that very few Ewe are classified as management, while a relatively large proportion are classified as skilled production. This interpretation is also born out by the occupational multinomial logit presented in Appendix A, which shows that after controlling for both worker and employer characteristics, Ewe are significantly less likely to be in managerial positions and significantly more likely to be in skilled production worker positions than members of other ethnic groups. Table 7 contains the results of estimating specifications (8a) and (8b) in which we apply the Altonji– Pierrot test for statistical discrimination. Few of the interaction terms between experience and its square and the ethnic dummies and mother’s education are significant. Using a general-to-specific approach, we arrived at two preferred specifications. When employers’ characteristics are not controlled for, the preferred specification contains only the interaction between experience and the Northern dummy and this is significant only when the Northern dummy is excluded (see specification (8a#)). When included, the northern dummy is insignificant. When employers’ characteristics are controlled for, once again it is only the Northern-experience interaction term that survives (see specification

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

373

Table 7 Testing for statistical discrimination on the basis of ethnic identity (dependent variable = log of monthly earnings in Cedi, n = 1045) (8a#)

(8a)

Asante Fante Ga-Adangbe Ewe Northern Asante  experience Fante  experience Ga-Adangbe  experience Ewe  experience Northern  experience Asante  experience2 Fante  experience2 Ga-Adangbe  experience2 Ewe x experience2 Northern  experience2 Mother’s ed.  experience Mother’s ed.  experience2 Worker characteristics Employer characterisitics Employer fixed effects R-squared

(8b#)

(8b)

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

 0.0485 0.0578 0.1159 0.0337 0.3611  0.0152  0.0172  0.0214  0.0156  0.0367 0.0004 0.0003 0.0004 0.0002 0.0005  0.0003 0.0000 yes no no 0.3715

0.1625 0.1547 0.2050 0.1626 0.2255 0.0158 0.0135 0.0198 0.0157 0.0183** 0.0003 0.0003 0.0004 0.0003 0.0003* 0.0012 0.0000

 0.1548  0.1331  0.0961  0.1717

0.0651** 0.0614** 0.0685 0.0630***

0.1142 0.1546 0.1331 0.2025 0.2650  0.0144  0.0168  0.0224  0.0166  0.0275 0.0002 0.0002 0.0004 0.0002 0.0002 0.0002 0.0000 yes no yes 0.7104

0.1396 0.1253 0.1685 0.1393 0.2049 0.0129 0.0108 0.0168 0.0127 0.0164* 0.0003 0.0002 0.0003 0.0003 0.0003 0.0010 0.0000

 0.0651  0.0607  0.0935  0.0085 0.0803

0.0610 0.0493 0.0610 0.0532 0.1216

 0.0050 0.0029*

yes no no 0.3677

 0.0123 0.0043***

yes no yes 0.7088

Standard errors corrected for heteroskedasticity using White’s (1980) procedure. *Significant at the 10% level. **Significant at the 5% level. ***Significant at the 1% level.

(8b#)), although in this case it is not necessary to remove the Northern dummy in order for the interaction term to be significant. These results suggest that the identified earnings differentials are not due to statistical discrimination. However, Northern workers are remunerated less than others for their growing work experiences. This result is consistent with Northerners being less productive perhaps because of language barriers (only 9% of Northerners state that Twi is their most proficient reading language) or lower schooling quality and employers discovering this only with experience. Alternatively, it is consistent with there being discrimination against Northerners by employers who choose not to increase their pay or promote them over time. Further investigations have shown that the lower return on experience for Northern workers is not due to their concentration in unskilled jobs, to them receiving less on-the-job training (31% of Northerners compared to 33% of others have received such training from their current employer), or to a complementarity between years of education and years of experience combined with their lower average years of education.15 15 In this investigation, experience was interacted with years of education, with the occupational dummies, and with a dummy indicating that a worker has at some time received training from their current employer.

374

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

5.2. Investigating the variations in earnings between employers’ kin, co-ethnics, and other workers The results presented in the first column in Table 8 corresponding to specification (9). Being related to the employer is associated with a 23% earnings premium. However, there is no evidence that unrelated co-ethnics earn any more than other unrelated workers. Following the introduction of occupational dummies in accordance with specification (10) (second column of Table 8), the relatives’ premium declines to 18% suggesting either that relatives are allocated better jobs as well as being paid more for whatever job they do or that the relatedness dummy is correlated with unobserved human capital. This earings premium may be due to merely to a preference to employ relatives, i.e., to nepotism. However, a preliminary analysis of enterprise performance suggests that relatives may be more productive. The results relating to specification (11) are presented in the third column of Table 8. Neither of the coefficients on the interaction terms involving the relatedness variable and experience and its square are significant at the 10% level. The earnings premium for relatives Table 8 Earnings, relatedness, and co-ethnicity (dependent variable = log monthly earnings in Cedi, n = 1045) (9)

Constant Asante Fante Ga-Adangbe Ewe Northern Years of education Years of ed. sq. Years of experience Years of experience sq. Female Married Mother’s years of ed. Management Unskilled Clerical/sales Related to employer Same ethnic group Related  experience Related  experience2 Co-ethnic  experience Co-ethnic  experience2 R-squared

(10)

Coef.

S.E.

10.9751  0.0822  0.0591  0.0845  0.0105  0.1854  0.0148 0.0044 0.0380  0.0004  0.0196 0.0142 0.0049

0.1838*** 0.0595 0.0493 0.0608 0.0528 0.0820 ** 0.0132 0.0006*** 0.0063*** 0.0001*** 0.0485 0.0414 0.0034

(11)

Coef.

11.2775  0.0572  0.0384  0.0361 0.0555  0.1071  0.0108 0.0025 0.0278  0.0003 0.0193 0.0081 0.0038 0.6401  0.3212  0.0320 0.2319 0.0704*** 0.1806 0.0136 0.0567 0.0306

0.7111

(11V)

S.E.

Coef.

S.E.

Coef.

S.E.

0.1504*** 0.0512 0.0433 0.0530 0.0475 0.0728 0.0126 0.0006*** 0.0058*** 0.0001*** 0.0467 0.0373 0.0031 0.0637*** 0.0424*** 0.0467 0.0625*** 0.0492

10.9125  0.0795  0.0591  0.0825  0.0065  0.1869  0.0155 0.0045 0.0426  0.0005  0.0234 0.0129 0.0045

0.1848*** 0.0591 0.0493 0.0608 0.0531 0.0830 ** 0.0134 0.0006*** 0.0069*** 0.0001*** 0.0492 0.0413 0.0033

10.9324  0.0784  0.0566  0.0807  0.0072  0.1830  0.0160 0.0045 0.0424  0.0005  0.0228 0.0136 0.0047

0.1824*** 0.0593 0.0493 0.0609 0.0531 0.0827 ** 0.0131 0.0006*** 0.0065*** 0.0001*** 0.0489 0.0416 0.0034

0.7699

0.2561 0.1561 0.2374 0.0705*** 0.3136 0.1246** 0.2951 0.1246** 0.0071 0.0144  0.0003 0.0003  0.0296 0.0107***  0.0289 0.0107*** 0.0005 0.0002*** 0.0005 0.0002*** 0.7143 0.7130

Standard errors corrected for heteroskedasticity using White’s (1980) procedure. All specifications include employer fixed effects. **Significant at the 5% level. ***Significant at the 1% level.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

375

does not vary with worker experience. For this reason, we re-estimated the functions excluding the relatedness interaction terms (fourth column of Table 8). The coefficients on the interaction terms between experience and its square and the co-ethnicity variable reveal that inexperienced co-ethnics receive a positive premium, but that this premium declines as experience increases.

6. Conclusions Our analysis provides evidence of large earnings differentials between ethnic groups within the Ghanaian manufacturing sector. A significant proportion of these differentials can be explained with reference to a fairly standard set of observed workers’ characteristics. In particular, workers’ educational attainments are found to vary across ethnic groups. Northerners, especially, earn considerably less because of their fewer years of education. After controlling for observed personal characteristics, significant differentials remain between the relatively high earning Other Akan and the relatively low earning Asante, Fante, and Ewe. None of these earnings differentials appear to be due to statistical discrimination by employers. Indeed, there is some evidence that inexperienced Northerners are given the benefit of the doubt and are paid less only later in their careers, possibly because language barriers or schooling quality reduce their chances of promotion. Turning to the role of kinship and co-ethnicity in the determination of earnings, there is strong evidence that employers favour their relatives in terms of pay and job allocations. Further, there is some evidence of statistical discrimination in favour of inexperienced, unrelated co-ethnics. There may be a link between these findings and the strong tendency for workers to be employed by members of their own ethnic group. The resulting market segregation, combined with the fact that entrepreneurs from different ethnic groups run very different types of enterprise, explains the greatest part of the earnings differentials. These findings strongly suggest that the Ghanaian labour market is indeed ethnically fractionalized and that this is leading to quite considerable earnings differentials between ethnic groups.

Appendix A. Occupational Attainment See Table 9.

Appendix B. Productivity Table 10 contains OLS regressions, with standard errors corrected using White’s procedure, for a sample of enterprises pooled over 6 years. The dependent variable in each of the estimated equations is the log of real value added per employee. Value added is calculated using the survey data, then deflated to 1991 prices using the CPI and divided by the number of employees before natural logs are taken. All of the equations include the log of the real capital – labour ratio lagged one period, the log of the number of employees

376

Table 9 Multinomial logit model for occupational attainment (n = 1045) (1) Clerical/sales

Unskilled

Coef.

S.E.

Coef.

Coef.

0.949  0.006 0.085 0.000 0.615 0.360 0.005 0.782 0.245  0.124  0.270  1.191  1.535  0.651 0.172  0.718 0.294  0.974 0.206 0.224

0.419** 0.015 0.111 0.001 0.476 0.485 0.031 0.564 0.416 0.487 0.430 0.501** 0.558*** 0.908 0.399 0.571 0.507 0.544* 0.147 0.167

0.478 0.310  0.003 0.012  0.074 0.067 0.001 0.001 1.672 0.261***  0.344 0.266  0.003 0.022  0.142 0.422 0.119 0.284  0.169 0.369 0.284 0.308  0.325 0.358  0.185 0.344 0.126 0.549 0.233 0.275 0.339 0.401 0.366 0.356 0.081 0.385 0.231 0.093**  0.040 0.120 yes yes yes no  860 0.284

 0.154 0.005  0.027 0.000 0.592  0.396  0.035  0.683  0.184 0.478 0.082  0.267 0.203 0.602  0.070  0.488 0.061 0.097  0.060 0.208

*Significant at the 10% level. **Significant at the 5% level. ***Significant at the 1% level.

S.E.

Management

Clerical/sales

Unskilled

S.E.

Coef.

S.E.

Coef.

S.E.

Coef.

S.E.

0.078** 0.005 0.062 0.001 0.287** 0.272 0.024 0.404* 0.308 0.395 0.340 0.390 0.368 0.472 0.316 0.451 0.342 0.372 0.090 0.130

1.203  0.001 0.081 0.001 0.320 0.815  0.031 1.610 0.127  0.389  0.495  1.393  1.921  0.599

0.644* 0.023 0.162 0.002 0.681 0.702 0.042 0.915* 0.578 0.658 0.574 0.738* 0.753** 1.174

0.643  0.007  0.087 0.001 2.240  0.259  0.020 0.185 1.023  0.330 0.620  0.015 0.157  0.195

0.445 0.017 0.089 0.001 0.352*** 0.337 0.027 0.601 0.409** 0.472 0.375* 0.442 0.427 0.675

 0.223 0.003  0.242 0.003 1.249  0.386  0.055  1.447 0.447 0.671 0.069 0.116 0.039 1.099

0.134* 0.007 0.093*** 0.001*** 0.416*** 0.385 0.032* 0.668** 0.476 0.488 0.413 0.476 0.474 0.604*

yes no no yes  611 0.491

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

Years of education Years of ed. sq. Age Age sq. Female Married Mothers years of ed. Related to employer Same ethnic group Asante Fante Ga-Adangbe Ewe Northern Foreign owned State owned Kumasi Coast Capital – labour ratio No. of employees Constant included Employers ethnicity Sector dummies Employer dummies Log likelihood Pseudo R-squared

(2)

Management

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

377

Table 10 Enterprise productivity (dependent variable: log of value added per worker, n = 703) (1)

Constant Capital – labour ratio No. of employees Human capital Kumasi Coast Some foreign ownership Some state ownership E-Asante E-Fante E-Ga-Adangbe E-Ewe E-Northern E-Non-African Proportion of related employees Proportion of co-ethnic employees Proportion of co-ethnic contacts R-squared

(2)

Coef.

S.E.

Coef.

S.E.

10.675 0.137 0.051 0.274 0.072  0.248 0.386  0.020

0.373*** 0.028*** 0.039 0.147* 0.109 0.193 0.152** 0.310

9.786 0.169 0.042 0.503 0.379  0.263 0.436  0.011  0.307 0.045 0.144  0.530 0.971  0.228 0.722  0.182  0.383

0.468*** 0.028*** 0.042 0.144*** 0.124*** 0.203 0.167*** 0.243 0.149** 0.143 0.176 0.176*** 0.278*** 0.257 0.277*** 0.189 0.157**

0.290

0.351

All standard errors reported are corrected for heteroskedasticity using White’s (1980) procedure. All regressions also include six-sectoral dummies and 5-year dummies as explanatory variables. *Significant at the 10% level. **Significant at the 5% level. ***Significant at the 1% level.

lagged one period, the log of the average years of education of the employees lagged one period, two location dummies, dummies indicating some foreign ownership and some state ownership, six-sectoral dummies and 5-year dummies. The value of the capital stock is estimated by using data from the first year that an enterprise was surveyed and then adding reported investment in subsequent years to that stock. Once again, the CPI is used to deflate the value of the capital stock and investments to 1991 prices. Lagging each of the explanatory variables ensures that they are predetermined. However, this is unlikely to fully solve all problems of endogeneity bias as, for each enterprise, all of these variables are correlated over time. The coefficient on the capital – labour ratio is significant at the 1% level and is fairly stable across the two specifications. The magnitude of the estimated coefficients indicates that a 1% increase in the capital stock is associated with an increase in output of between 0.14% and 0.17%. The labour variable is insignificant in both specifications suggesting that returns to scale are constant. The coefficient on the human capital variable is significant in both the specifications, but varies in magnitude depending on whether the ethnicity variables are included. A 1% increase in the average level of education of the employees is associated with between a 0.27% and a 0.50% increase in output. Once we control for ethnicity, we find evidence that Kumasi-based enterprises are more productive than Accra-based ones, while enterprises with some foreign ownership are more productive.

378

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

To the basic model presented in column 1, we add a set of dummies capturing the ethnic identity of the entrepreneurs, the proportion of employees who are related to the entrepreneur, the proportion of employees who, although not related, come from the same ethnic group as the entrepreneur, and the proportion of his/her contacts who come from the same ethnic group. The proportion of employees that are related to the entrepreneur the corresponding coefficient is significant and positive. A 10 percentage point increase in the proportion of relatives in the workforce increases output by 0.07%. This result is particularly striking given that the proportion of relatives is negatively correlated with value added, the number of employees, the capital – labour ratio, and the human capital variable, and uncorrelated with productivity. However, this result must be treated with considerable caution. The data for this variable was collected only in the fifth wave of the GMES. In the regressions, we have assumed that for each enterprise the proportion of relatives in the workforce was constant over time and so applied the same proportion to each of the earlier waves. Acknowledgements We would like to express our gratitude to the Department for International Development for funding the collection of the data and Abigail Barr’s involvement in this study; the World Bank for funding Abena Oduro’s involvement; the Economic and Social Research Council for funding our host institution, the Centre for the Study of African Economies; our field researchers for their perseverance; and the many employers and employees in Ghana for their cooperation and patience. References Aigner, D.J., Cain, G.G., 1977. Statistical theories of discrimination in labor markets. Industrial and Labour Relations Review 30, 175 – 87. Altonji, J.G., Blank, R.M., 1999. Race and gender in labor markets. In: Ashenfelter, O.A., Card, D. (Eds.), Handbook of Labor Economics, vol. 3C. Elsevier, Amsterdam. Altonji, J.G., Pierrot, C.R., 2001. Employer learning and statistical discrimination. Quarterly Journal of Economics 116 (1). Apter, D., 1965. The Politics of Modernization. Chicago Univ. Press, Chicago. Arrow, K., 1973. The theory of discrimination. In: Ashenfelter, O.A., Rees, A. (Eds.), Discrimination in Labor Markets. Princeton Univ. Press, Princeton, 91 – 100. Arrow, K., 1998. What has economics to say about racial discrimination? Journal of Economic Perspectives 12. Bates, R., 1974. Ethnic competition and modernization in contemporary Africa. Comparative Political Studies 6 (4). Becker, G.S., 1971. The Economics of Discrimination, 2nd edn. University of Chicago Press, Chicago. Benabou, R., 1994, ‘Education, income distribution, and growth: the local connection’, Centre for Economic Policy Research, Discussion: 995. Benabou, R., 1996. Heterogeneity, stratification, and growth: macroeconomic implications of community structure and school finance. American Economic Review 86, 584 – 609. Bigsten, A., Collier, P., Dercon, S., Fafchamps, M., Gauthier, B., Gunning, J.W., Isaksson, A., Oduro, A., Oostendorp, R., Pattillo, C., Soderbom, M., Teal, F., Zeufack, A., 1998. Rates of Return on Physical and Human Capital in Africa’s Manufacturing Sector, Working Paper WPS/98.12. Centre for the Study of African Economies, University of Oxford.

A. Barr, A. Oduro / Journal of Development Economics 68 (2002) 355–379

379

Black, D.A., 1995. Discrimination in an equilibrium search model. Journal of Labor Economics 13, 309 – 33. Borjas, G.J., Bronars, S.G., 1989. Consumer discrimination and self-employment. Journal of Political Economy 97, 581 – 605. Bowlus, A.J., Eckstein, Z., 1998. Discrimination and skill differences in a equilibrium search model, mimeo, University of Western Ontario. Brown, C., Medoff, J., 1989. The employer size wage effect. Journal of Political Economy 97, 1027 – 59. Clague, C., 1993. Rule obedience, organizational loyalty, and economic development. Journal of Institutional and Theoretical Economics 149, 393 – 414. Coate, S., Loury, G., 1993. Antidiscrimination enforcement and the problem of patronization. American Economic Review 83, 92 – 98. Coate, S., Loury, G., 1993. Will affirmative-action policies eliminate negative stereotypes? American Economic Review 83, 1220 – 40. Cohen, A., 1969. Custom and Politics in Urban Africa: A Study of Hausa Migrants in Yoruba Towns. University of California Press, Los Angeles. Collier, P., Hoeffler, A., 1998. On the economic causes of civil war. Oxford Economic Papers. Durlauf, S.D., 1996. A theory of persistent income inequality. Journal of Economic Growth 1, 175 – 95. Easterly, W., Levine, R., 1997. Africa’s growth tragedy. Quarterly Journal of Economics 112. Fafchamps, M., 2000. Ethnicity and credit in African manufacturing. Journal of Development Economics, 61(1), 205 – 35. Glickman, H., 1995. Ethnic Conflict and Democratization in Africa. The African Studies Association Press, Atlanta. Gluckman, M., 1961. Custom and Conflict in Africa Blackwell, Oxford. Granovetter, M.S., 1973. The strength of weak ties. American Journal of Sociology 78, 1360 – 80. Knack, S., Keefer, P., 1997. Does social capital have an economic payoff ? A cross-country investigation. Quarterly Journal of Economics 112, 1251 – 88. Kranton, R.E., Minehart, D.F., 2001. A theory of buyer-seller networks. American Economic Review 91(3), 485 – 508. Krueger, A., Summers, L., 1988. Efficiency wages and the inter-industry wage structure. Econometrica 56, 259 – 93. Loury, G.C., 1977. A dynamic theory of racial income differences. In: Wallace, P.A., LaMond, A.M. (Eds.), Women, Minorities and Employment Discrimination D.C. Heath, Lexington. Loury, G.C., 1981. Intergenerational transfers and the distribution of earnings. Econometrica 49, 843 – 67. Lundberg, S.J., 1991. The enforcement of equal opportunity laws under imperfect information: affirmative action and alternatives. Quarterly Journal of Economics 106, 309 – 26. Lundberg, S.J., Startz, R., 1983. Private discrimination and social intervention in competitive labor markets. American Economic Review 73, 340 – 47. Lundberg, S.J., Startz, R., 1998. On the persistence of racial inequality. Journal of Labor Economics 16, 292 – 323. Maxwell, N., 1994. The effect on black – white wage differences of differences in the quantity and quality of education. Industrial and Labor Relations Review 47, 249 – 64. Montgomery, J.D., 1991. Social networks and labor-market outcomes: toward an economic analysis. American Economic Review 81, 1407 – 18. Neal, D.A., Johnson, W.R., 1996. The role of premarket factors in black – white wage differences. Journal of Political Economy 104, 869 – 95. Oaxaca, R., 1973. Male – female wage differentials in urban labor markets. International Economic Review 14, 693 – 709. O’Neill, J., 1990. The role of human capital in earnings differences between black and white men. Journal of Economic Perspectives 4, 25 – 45. Phelps, E.S., 1972. The statistical theory of racism and sexism. American Economic Review 62, 659 – 61. Rees, A., Schultz, G., 1970. Workers in an Urban Labor Market. University of Chicago Press, Chicago. Rothchild, D., Oluonsola, V., 1983. The State Versus Ethnic Claims: African Policy Dilemmas Westview Press, Boulder, CO. White, H., 1980. A heteroscedasticity-consistent covariance matrix estimator and a direct test for heteroscedasticity. Econometrica 48, 817 – 38.