Research "Firms, Workers, and Human Capital in Ghanaian Manufacturing " pdf

A new process for accurately measuring the returns to education is developed, one that controls for the major problem in this estimation, namely ability bias, using information about t

Firms, Workers, and Human Capital in Ghanaian Manufacturing

A Dissertation Presented to the Faculty of the Graduate School

of Yale University

in Candidacy for the Degree of Doctor of Philosophy

by Garth Douglas Frazer

Dissertation Director: Christopher R Udry

May, 2003

UMI Number: 3084290

Copyright 2003 by Frazer, Garth Douglas

All rights reserved

®

UMI

UMI Microform 3084290 Copyright 2003 by ProQuest Information and Learning Company All rights reserved This microform edition is protected against unauthorized copying under Title 17, United States Code

ProQuest Information and Learning Company

300 North Zeeb Road P.O Box 1346 Ann Arbor, MI 48106-1346

© 2003 by Garth Douglas Frazer All rights reserved.

Abstract

Firms, Workers, and Human Capital in Ghanaian Manufacturing

Garth Douglas Frazer

2003

Understanding the nature of human capital in manufacturing in Africa will play an

important role in facilitating the development of this scctor This dissertation focuses on

three aspects of the human capital of the firm, and the relationship between the firm and

its employees The primary datasct used is a survey of manufacturing firms and their

cmployces in Ghana First, the productive and remuncrative returns to education are

calculated and compared A new process for accurately measuring the returns to

education is developed, one that controls for the major problem in this estimation,

namely ability bias, using information about the contribution of the firm's workers to

production This methodology is applicable to other contexts where linked employer-

employee data are available Second, given the importance of the extended family

network in Africa, and in particular the importance of hiring relatives to work in the

firm, particularly in smaller firms, the impact of these related employees on firm

profitability is measured Specifically, the productive contribution of relatives is

compared to their remuncration, in order to determine their overall impact on firm

profitability Third, the institution of apprenticeship, which is a period of a few years

during which an apprentice Icarns a manufacturing trade, is explored A model is

presented where apprenticeship is an instance of training in firm-specific human capital,

training which increases an individual's productivity in the current firm, but not in any

other firm Individuals invest in this firm-specific human capital if they have a sufficient

probability of obtaining the capital to start their own firm, and replicate the technology

and business practice of the apprenticeship firm Predictions of the model are tested

To Catherine

Contents

Acknowledgements .cccccccnsesccccssresssseeccecessseessncenseserecseserseeescseeeoneoene i LisE Of FliEUS cc- G1 00 0 1 in th nh 06 ii

II mod ïv ae ii

2 Heterogeneous Labor and Returns to Education .-«- 10 2.1 The Problem of Ability Bias cà se nhehhhehree 14

2.1.1 Instrumental Variable 'echniques - cà ee 16

2.1.2 Developing Country Studies cà se nhheeehhherrrre 19 2.2 Modeling the Production of the Firm se hhee 22 2.2.1 Incorporating the Heterogeneity of Labor in the Production

EUHCĐÏON c0 0n cv nh Tnhh nh HH nh nh es 22 2.2.2 Estimating the Productlon Function «c«ccc« xe 30

2.3 Controlling for Ability Bias in the Wage Equation 34

3.3 Estimating the Wage Equation che nhe 88

SN NỀ:: ti DEORE SENSE EERE EEE EEE C REET EERE ES 91

3.5 Estimating Relative Remuneration and Productivity 93 3.6 COHCÏUSÏON 2Q 00v vn nh nh Tnhh nen TH nh nhà ng 98

4 Apprenticeship and Firm-Speciñic Human Capital 111

4.1 The ModlelÌ c co cv TH nh BE ĐH nh nh nh nh kh, 116 4.1.1 The Basic Model cán nn nh nn nh nành nho He ườ 116

4.1.2 Equilibrlum «ch nà kh nh nh Hà nghe ng kho ờn 119

4.3 Results .ccccccccccceeeeee cee eee eee e etn e een E eed cà ng nh to 133

1.0: 157

Fernandes, Eli Berman as well as seminar participants at Yale University, University of Toronto, Queen's University, University of California-San Dicgo, Boston University, University of Texas-Austin, University of North Carolina-Chapel Hill, University of Colorado Boulder, McGill University, University of Alberta, University of Guelph, and Carleton University, and the Northeast Universities’ Development Consortium

I would also like to thank Francis Teal for the opportunity to participate in the

Ghanaian survey, and to Oxford University and the Ghana Statistical Service for the use

of the data from the Ghanaian Manufacturing Enterprise Survey Permission from the Ghana Statistical Scrvice to use the Ghana Living Standards Survey - Wave 4 data is gratofully acknowledged I would also like to thank the MacArthur Rescarch Network

on Incquality and Economic Performance, as well as the Social Sciences and Humanities Research Council of Canada for support for the doctoral rescarch

Many othcrs have supported me in various ways during the period of this dissertation

My teachers and classmates at Yale University have provided a stimulating environment

to pursue doctoral studies, deeply cnriching my understanding of cconomics My family and fricnds have been patient with me during the challenges that accompany the

dissertation process In particular, without the loving and paticnt support of my wife, Catherine, to whom this dissertation is dedicated, it is difficult to imagine how my doctorate could have been completed I would also like to thank my mother, Elaine, and

my late father, Doug, for the many little efforts and sacrifices that they made for me over the years They taught me how to ask sensible questions long before I pursued doctoral studics.

List of Figures

Figure 2.1 - Matcrials Punction for Manufacturing Industty - ii 62 Figure 3.1 - Nonparametric Regression of Relatives on Size eee: 105 Figure 3.2 - Distribution Across Occupations (Relatives vs Ovorall) 106

List of Tables

"0 11589 nh 63 Table 2.2 - Production Function Estimation in Ghanaian Manufacturing 64 Table 2.3 - Returns to Education in Ghanalan Manufacturing - -. «e 65

Table 2.4 - Returns to Education (TV) ccccehhhHhhhhhhehhhhhueriưe 66

Table 3.1 - Comparing Relatives to Othor Workers' Gencral Charactoristics 107 Table 3.2 - Summmary Statistics for the Data ăn ehhhuhheehHhnreerree 108 Table 3.3 - Firm Production Functions (Including Relatives Control) 109 Table 3.4 - Wage Equation (Including Relatives Control) cee 110 Table 3.5 - Wage Equation (Additional Controls) cccsằheehhiheeree 111 Table 4.1 - Sunmary Statistics - GLSS SUPVCY à ehhuuheehhHhderreeeeree 151 Table 4.2 - Summary Statistics for Manufacturing Workers from the GMES

Table 4.3 - Activitics of Apprentices After Apprenticeship .ccccceseeeeeeeeees 153 Table 4.4 - Selection-Corrected Earnings Regressions - GL55 Data 154 Table 4.5 - Returns to Apprenticeship Within the Manufacturing Sector

(GMES data) ch tt kh nhi HH 155

Table 4.6 - Production Function - Ghanaian Manufacturing Scctor 156

Appendix Tables

Appendix Table 2.A1 - Returns to Education (Selection) cece 67 Appendix Table 2.A2 - Estimates of the Returns to Education Using

Appendix Table 2.A3 - Estimates of the Returns to Education Using Family Background Variablos as ÍnslTHIGHES c.à ii nhhhhHhherrderrrderrdrre 69 Appendix Table 2.Ad - Estimatcs of the Returns to Education Using Studics 8B 777 ae 70

1 Introduction

Over the history of development economics, one of the most understudied areas has

been manufacturing in Africa Part of the reason for this neglect has been the rela- tively minor role of manufacturing in African economies While manufacturing has

played a minor role in African countries, it plays a larger role in middle-income de-

veloping countries, and has played a significant role in the economic development of

virtually all industrialized countries, including the newly industrialized countries

Moreover, over the past two decades, despite the recurring droughts in various re-

gions of Africa, agricultural growth has been stronger than that of industry, and in particular manufacturing Overall Africa’s agricultural sector grew at an average of 2.3% from 1980-90 and at 2.7% from 1990-99 (World Bank, 2001) In comparison, manufacturing grew at a pace of 1.7% from 1980-90 and 1.6% from 1990-99 While these differential growth rates might reflect a variety of different factors, including

comparative advantage considerations, the importance of manufacturing in the de-

velopment of virtually all of the industrialized countries suggests that we should try

to understand why this sector lags in Africa The first step in this understanding involves analyzing the behaviour of manufacturing firms in Africa

In addition to the relatively minor role of manufacturing in Africa, another reason for the neglect of manufacturing in research on Africa has been the absence of micro-

' According to the World Bank’s World Development Indicators (2001), the percentage of GDP accounted for by manufacturing (in terms of value added) in Africa was 16 percent, in comparison

to the overall average of 24 percent for the developing countries.

level data on manufacturing firms Fortunately, this has changed in the 1990s with a

series of surveys of manufacturing firms in various African countries This dissertation

seeks to use the data from the survey of firms in Ghana, which collected data covering,

in the end, a span of 9 years (1991 through 1999)

Naturally, the research questions related to manufacturing in Africa are many,

and any dissertation must narrow its focus to come up with effective conclusions

This dissertation seeks to focus on the human capital of African manufacturing firms, developing a further awareness of the relationship between a firm and its employees This relationship has broad implications for growth and development A continu- ing puzzle in development is the question of why capital does not flow in enormous

quantities from rich countries to poor countries, as the standard neoclassical model

would predict if the technologies in these countries are even remotely similar (Lucas,

1990) Lucas suggests that the explanation may lie in differences in human capital

across rich and poor countries That is, the reason why poor countries do not receive

foreign investment is because of a lower level of labor quality, or human capital per

worker To further explore the potential importance of this hypothesis, empirical

work at the micro level is crucial Therefore, in so far as this dissertation seeks to

further understand the way in which the human capital at a firm is valued, and the

importance of various aspects of this human capital, this dissertation is a small step

*The first three years of the survey were coordinated by the World Bank, and organized by Oxford University in conjunction with the University of Ghana and then the Ghana Statistical Service The later years of the survey were continued by Oxford University in conjunction with the Ghana Statistical Service.

toward contributing to the larger question posed, and suggestively answered by Lucas

To address the relationship between a firm and its workers, or the importance of

human capital in manufacturing, the manufacturing firm dataset used in this study

is particularly rich In particular, this dataset is one of a growing number of linked employer-employee datasets, where we have information not only on the firms where workers work, but also on the employees working there In the Ghanaian manufac- turing sector, the individuals in the employ of a firm include its regular paid workers,

as well as apprentices, who are in the process of learning a trade, over a period of

typically a few years In the Ghanaian survey, in addition to collecting data on all

of the usual details of the manufacturing firms (e.g their value-added, number of

employees, size of their capital stock, etc.), a sub-sample of the firm’s workers and

apprentices was surveyed in order to collect more detailed information on these indi-

viduals Combining the information on workers with the information on firms is at

the heart of each of the papers of this dissertation

In the process of exploring the human capital within African manufacturing, part

of the research agenda parallels a broader one, and part of it is specific to Africa, or to

developing countries The second chapter of the dissertation explores a broad question

which has been explored extensively in developed countries by labor economists That

is, what are the economic returns to schooling? Naturally, in the context of Africa,

the returns to schooling are much broader than the economic ones, as schooling is a significant means of building citizenship and national identity in countries that, have

had their independence for typically less than forty years Still, in countries with very limited resources, and competing demands on budget dollars, the economic returns from schooling are important Schooling is typically seen as a means of addressing distributional as well as growth objectives, the former through the universality of

education, particularly primary education, the latter through the impact of education

on productivity, as alluded to by Lucas Its ability to achieve both of these objectives

should be revealed through schooling’s valuation in the market, namely the returns

to education in the form of increased wages

Carefully measuring the increase in an individual’s wage that results from an

increase in a year of schooling is a difficult econometric problem that has been explored

by development and labour economists for at least thirty years The primary issue that arises is the fact that an individual’s level of schooling is correlated with that

individual’s ability Given that an individual’s schooling is typically observed, while

an individual’s ability is typically unobserved, the challenge is to identify that portion

of a person’s increased wage which results from her increased schooling, rather than the higher ability that might have led to that higher level of schooling A variety

of different methods have been proposed for dealing with this issue, and limitations

have been outlined with all of the approaches

The second chapter of the dissertation attempts to measure the returns to educa-

tion in Ghana, and in particular, within the manufacturing sector in Ghana In the

process, it devises a new method for addressing the issue of ability bias, one which

takes advantage of the linked employer-employee nature of the dataset While the

specific details are explored in the chapter 2, in essence it captures the fact that

workers reveal their ability through their contribution to firm product A measure

of average worker ability (averaged across the workers at a firm) is obtained which

is then used in a wage regression to directly control for worker ability Because the measure of ability obtained is an average at the firm level, the wage regressions need

to be grouped to the firm level in order to use this measure While the returns to

education measure obtained describes the returns to education in Ghana, the proce-

dure developed for consistently measuring the returns can be used whenever a linked

employer-employee dataset, which the appropriate variables (such as schooling) is

available

While the methodology of the first main paper of the dissertation, on returns

to education, is applicable with the right data, to virtually any context, the second

and third papers are more closely tied to the African context The second paper

focuses on an institution whose importance is certainly stronger in developing than developed countries, and perhaps strongest in Africa, namely the extended family Manufacturing firm owners in Africa, and in particular firm owners of smaller firms,

frequently hire their relatives to work for them In the manufacturing dataset used

in this dissertation, 4% of the manufacturing workers were related to the firm owner However, this statistic understates the importance of relatives, particularly for smaller firms, as the average proportion of relatives at a firm was 16%, again reflecting the

greater importance of relatives for smaller firms

The reason why hiring relatives as employees might matter reflects both previous

literature and the stories of firm managers On the one hand, in personal interviews

with the author, firm managers related a hesitancy to hire relatives, but a pressure

to do so from relatives The hesitancy reflected the same source as the pressure

to hire That is, once relatives are hired by the firms, it may be difficult for the

firm to fire them, without experiencing recrimination of some sort from the extended

family This fact may reduce the incentives of relatives to work hard for the firm, and

reduce the productivity of relatives On the other hand, on family farms, relatives are

typically found to be more productive than non-relatives Here, the situation is a bit

different in that the relatives ‘employed’ by the farm manager are typically members

of the same household as the farm manager, and therefore their welfare is more closely intertwined Still, while the same argument may not apply as strongly to relatives

working within the firm, as they typically are not members of the same household,*

in so far as relatives belong to the same extended family network, and transfer of

resources occurs within that network, the ‘family farm’ argument applies, albeit to a

lesser degree Therefore, whether relatives are ultimately a burden or a boost for the

firm is an empirical question In fact, to determine whether relatives benefit or harm

firm profitability (or neither) requires data not only on the productivity of relatives,

but also on their remuneration If relatives are less productive than non-relatives, but

3This reflects my personal observations of the relatives working in the firms where I was inter- viewing While separate data on whether the relatives were living in the same household as the firm manager was not collected, I am quite confident that these would be a minority of cases.

are also paid less, in accordance with their productivity, the firm is not harmed, even

if it is forced by the extended family network to hire them For example, this would

reflect the case where firm owners are forced to hire relatives, but the extended family

does not dictate what their wage need be Fortunately, data on the productivity of

relatives can be revealed by examination of firm production functions, and data on

their remuneration is captured from the individual level data While the results

from this paper are not as applicable to manufacturing firms in developed countries

(although they might be relevant for small retail and other small firms), the influence

of the extended family network has long been of interest to development researchiers

The third paper, while in terms of institutional details reflects the African context,

also addresses a wider issue of economic interest The institution examined in the

third paper is apprenticeship Apprentices attach themselves to a master (who is

typically also a firm’s owner-manager) in order to learn the trade practised by the

master The apprenticeship usually lasts approximately three years, typically involves

payment of fees to the master at the beginning and at the end of the apprenticeship,

with little or no wage for the apprentice during the period of apprenticeship In

this paper, apprenticeship is modelled as a form of firm-specific training Because

the technological and business practice training of apprenticeship is applicable only

within the apprenticeship firm, the apprenticed worker is more productive within

that firm However, the apprenticed worker need not be paid more than his or her outside option, and therefore need not be paid more than workers at other firms

The way in which apprenticed workers can reap the returns to apprenticeship is if

they become self-employed, and replicate the technology and business practice of the

apprenticeship firm These institutional details reflect the data, and the predictions generated by this model are confirmed in the data, as detailed in the paper

References

{1] Lucas, Robert E., Jr “On the Mechanics of Economic Development.” Journal

of Monetary Economics, July 1988, 22(1), pp 3-42

[2] World Bank, World Development Indicators 2001 Washington, D.C.: World

Bank, 2001

2 Heterogeneous Labor and Returns to Education

What is the role of education in growth and development? The common presumption

is that education plays a crucial role, with human capital central to many recent models of growth (e.g Paul M Romer, 1986; Robert E Lucas, Jr., 1988), but the

cross-country empirical evidence is far from unanimous on education’s importance

(Jonathan R Temple, 2001) On the other hand, at the micro level, the positive relationship between the education and wages of workers has been found to hold in

dozens of countries and numerous datasets The micro evidence has led eclucation

to be seen in many circles as “the great equalizer,” with universal education seen as achieving distributional as well as productive goals Understanding the importance

of education is particularly pointed in Africa, which has experienced lower growth and

lower levels of education than all other regions over the past thirty years, and yet the

growth of the stock of education has been more rapid in Africa than elsewhere (Vikram Nehru et al., 1995) Exploring the precise role of education in African economies

involves examination of a number of research questions This paper begins with a

very basic, but important one, accurate measurement of returns to education

Measuring returns to education carefully is particularly important in developing

country contexts where government and individual resources are extremely limited,

and education’s competition for budget dollars is intense.’ While consistent estima-

1Some studies which have examined the returns to schooling in the context of developing countries include: Jere R Behrman and Anil B Deolalikar (1991) - Indonesia ; Behrman, et al., (1999) - India; Tekaligne Godana (1997) - Zimbabwe; Tianyou Li and Junsen Zhang (1998) - China; Germano Mwabu and T Paul Schultz (1996) - South Africa, Helen Skyt Nielsen and Niels Westergard-Nielsen

tion of the returns to education has received considerable attention of labor economists

in industrialized country contexts, as well as development economists in some coun-

tries, comparatively little ”best practice” research has been done for African coun-

trics (Simon Appleton, et al., 1996; Jere R Behrman, 1996) The primary problem

in consistent measurement of the wage returns to schooling is that of ability bias

A worker’s ability is typically part of the residual, and yet is correlated with the

worker’s schooling, resulting in biased parameter estimates This study develops a

new technique for controlling for this ability bias, using information from the firm

where workers work, and uses this technique to obtain consistent estimates of the

returns to education in Ghana

Over the past twenty-five years, the model that, has dominated the estimation of returns to education in developing country contexts is the human capital model of Jacob Mincer (1974), which predicts that a worker’s wage should be a function of schooling and work experience Mincer acknowledged the ability bias problem: “Tt

is widely believed that the omission of ability from the earnings function creates a

specification bias: leaving out a variable which is positively correlated with earnings and investment? biases the coefficient of investment (average rate of return) upward.”

(p 189) The vast majority of studies have attempted to address the issue of ability

bias in the context of the Mincer model using “quasi-experiments” and instrumental

variable techniques, where some trait, such as distance to the nearest school, level

of parental education, different schooling laws, or a twin’s level of education, are

assumed to exogenously affect a person’s level of education While these studies have

considerably advanced our understanding of returns to education, each method has

its limitations, as will be summarized in Section 2.1

This study attempts to examine the question from a different perspective entirely-

in particular, that of the firm It takes a result of the Mincer model, that an

individual’s productivity is a function of her schooling, experience, and ability, and

develops the labor term of the production function that is consistent with this model.®

Typically, in production function studies, the labor term is simply an hours measure,

or at best a number of skilled and unskilled workers, so this new specification more

accurately depicts the heterogeneity of labor at a firm With this Mincer-consistent

production function, a measure of worker ability can be obtained, where ability is

defined as all of the productive characteristics of a worker other than schooling and

experience The thesis of this paper is that this is exactly the definition of ability

that matters Under the Mincer-consistent specification, worker ability is shown to

form one component of firm productivity, where firm productivity is defined as in

the current standard of the industrial organization literature as that portion of the residual which can be seen and acted upon by the firm manager To separate worker ability from that portion of firm productivity which is due to management talent

' The inclusion of schooling and experience into a production function is not new to this study, but was first introduced by the working paper version of Mark Bils and Peter J Klenow (2000), and has been used with firm data in Patricia Jones (2001), Arne Bigsten et al (2000), and Hellerstein, et

al (1999} The inclusion of ability, and the incorporation of the Mincer model into the production function is new to this study

or to technological change, management talent is assumed to be fixed for a given manager, and technological change is assumed to be stationary about a trend The

production function is estimated consistently, taking into account the simultaneity

of labor demand and output decisions, with a technique that follows from James Levinsohn and Amil Petrin (2000) and Steven Olley and Ariel Pakes (1996) The measure of worker ability obtained is an average at the firm level, and can be used in

a wage regression, which is grouped from the individual to the firm level, to control

for ability, and to obtain a consistent estimate of the returns to education

The procedure used in this study can be generalized The assumptions of any

wage equation in which the log(wage) is a linear function (in the broadest definition

of this term) of various factors can be incorporated into a production function The

productive contribution of these factors can be compared to their relative remuner-

ation Assuming within-period profit maximization, the coefficients of the factors

in the production function should equal the corresponding coefficients in the wage

equation (Section 2.2.1 and Appendix 2.1) Clearly, what enables this technique is

the availability of linked employer-employee data This type of data is now available

for a number of countries, and has spawned a growing literature attempting to exploit

it.’

The outline of this paper is as follows Section 2.1 outlines methods which have been used to date to attempt to control for ability bias Section 2.2 outlines

7See John M Abowd and Frances Kramarz (1999) for a survey of the linked employer-employee literature

the incorporation of the Mincer assumptions into the production function, and the

procedure for its estimation Section 2.3 describes the wage equation estimation

Section 2.4 describes the data used for estimation, with the results in Section 2.5,

and a brief conclusion in Section 2.6

2.1 The Problem of Ability Bias

Mincer’s widely-used human capital model (1974) predicts that the following rela-

tionship between wages, schooling and experience should hold:

The work experience variable, X, is designed to capture the effects of on-the-job train-

ing which people receive over their worklife, and in Mincer’s model was a quadratic

Given the typical difficulty in measuring this exactly, a variable measuring potential

experience was proposed by Mincer, and is typically used A person’s potential ex-

perience is the length of his post-schooling life, i.e X=G-S-6, where G is a person’s

age, and the number six removes a person’s pre-school years Now, the chief prob-

lem with this regression is the fact that the residual, x;, includes a person’s ability,

which could both independently affect a person’s wage and be correlated with one’s

schooling The complete equation would read:

log W⁄¿ = Ào + ÀsŠ; + Àx0(X;) + ÀAÁ¿ + 0¿ (2)

The potential of individual ability to confound the returns to schooling has been recognized at least since Gary S Becker (1964) The problem in the above specifi-

cation is that, if ability is omitted from the regression, the schooling coefficient will

be biased To consider the likely direction of the bias (given that ability is not as

correlated with experience as it is with schooling), consider the plim for the OLS

coefficient in a regression without experience:

cov( Aj, S;)

plim As_OLS = As + ÀA

Specifically, Ag will be biased upward provided that: i) ability affects the wage,

independent of its effect on schooling, and ii) schooling and ability are positively

correlated A variety of approaches have been used to address this problem, with the major approaches being summarized in a survey article by David Card (1999) As Zvi

Griliches (1977, p 5) notes, “the simplest way of dealing with this problem is to find

a measure of ‘ability’ and include it in such an equation,” and such an approach was

common in the earlier literature (Griliches and William M Mason, 1972; Griliches,

1976; Griliches, 1977) However, as Griliches (1977) notes in the same article, the value of using IQ-measures as ability controls was controversial: “Two polar views

are possible ‘Ability’ is IQ, or something close to it, and the only problem is that

our measures of it are subject to possibly large (test-retest) errors The alternative view is that ‘ability’, in the sense of being able to earn higher wages, other things

equal, has little to do with IQ.” (p 7) In the subsequent 25 years, the latter view has

been more widely held by the literature, as IQ or other test-related measures have not

often been used to control for ability bias.? The current study uses a direct control

°Some studies that have used test measures include: M Boissiere, et al., 1985; Paul Glewwe, 1999; John B Knight and Richard H Sabot, 1990; Glewwe, 1996; McKinley L Blackburn and David

for ability, but not one which is obtained from an IQ-type test, but rather one which

is revealed through an individual’s productivity Before delineating the procedure

employed, a review of the techniques used to date to handle the problem of ability

bias is warranted

2.1.1 Instrumental Variable Techniques

A variety of studies have attempted to exploit institutional features of the school

system across states or regions, times, or individuals, as ”quasi-experiments” where

people have been subjected to differing levels of treatment in terms of the costs or

benefits of schooling These institutional features of the school system have served

as instruments for the schooling variable, to purge it of its correlation with unob-

served ability.” As evident in the results in Appendix Table A2.2, the instrumental

variable estimates are virtually always higher than the OLS estimates and at times

considerably so However, as discussed previously, the theory predicts that the OLS

estimates should be upward-biased Griliches (1977) provides a partial explanation

to the puzzle in that the attenuation bias of measurement error could be biasing the OLS coefficients downward, offsetting the ability bias However, building on the fact

Neumark, 1995 The ability test used in the developing country context, the Raven Progressive

Matrices test, has virtually always been insignificant (Boissiere, et al., 1985; Glewwe, 1999; Knight

and Sabot, 1990; Glewwe, 1996) This either reflects an absence of ability bias in returns to schooling, or that the test has not yet measured ability accurately This paper is motivated by the latter possibility

‘ Angrist and Krueger (1991) use variations in American compulsory schooling laws, combined with an individual’s quarter of birth as identifying instruments Harmon and Walker (1995) use changes in the minimum school-leaving age in the United Kingdom Kane and Rouse (1993) use public tuition and the distance to the closest 2-year and 4-year colleges as instruments

that the reliability ratio of self-reported schooling in U.S datasets is about 90%!”,

Card (1999, p 1841) notes, ”since measurement error bias by itself can only explain

a 10% gap between OLS and IV, however, it seems unlikely that so many studies

would find large positive gaps between their IV and OLS estimates simply because of

mcasurement error.”

Card (1995) suggests another explanation for these high IV estimates, namely heterogencity in the marginal rate of return to schooling in the population If such heterogeneity exists, then the IV estimates measure the returns to education for the

subgroup of the population affected by the instrument For example, in the studies

using minimum school-leaving age as an instrument, the IV estimate provides the return to education for the subgroup of the population for whom the compulsory school-leaving age matters Card (1995, 2001) argues that the subgroups affected by

the institutional innovations in question typically will have higher marginal rates of

return than other subgroups of the population, thus explaining why the IV estimates

are typically larger than those of OLS.!!

At least as common in the literature as the institutional variation studies is the

use of family background variables as instruments Here, the assumption is that

parental schooling, for example, affects an individual’s schooling level, but does not

OT hese reliability ratios have been calculated using the datasets primarily used by U.S and European investigators, but these are primarily the results being presented Reliability ratios for developing country datasets such as Maluccio’s could be different

‘As Ichino and Winter-Ebmer (1999) note, this conclusion is consistent with the local average treatment effects (LATE) interpretation of IV estimates (Imbens and Angrist, 1994), namely that IV identifies the average treatment of those who comply with the assignment-to-treatment mechanism implied by the instrument

independently affect the wage However, the fact that many studies use these family

background variables to control for individual ability (e.g Ashenfelter and Zimmer- man, 1997; Ashenfelter and Rouse, 1998) suggests that parental variables, if other

controls for ability are not used, may remain correlated with the unobserved ability

of the regression (in the residual), and therefore may not be valid instruments A summary of studies that have used family background variables both as additional

controls, and as instruments is found in Appendix Table A2.3

Another set of studies that can be interpreted as instrumental-variable studies are studies of identical (monozygotic) twins, with the twin’s schooling instrumenting for own-schooling The assumption in the twins studies is that twins have identical abilities, because of their identical genetics, and therefore the difference in schooling levels between twins can be treated as a natural experiment Regressing the differ- ences in wages between twins on the differences in their schoolings should remove the

effect of ability (which is presumed identical between them, and therefore differenced

out) Studies of twins date back to Behrman and Taubman (1976) However, the question of whether identical twins do have identical ability has been raised since

Griliches (1979),!2 and has been strongly echoed in recent papers (Rosenzweig and

Wolpin, 2000; Bound and Solon, 1999; Neumark, 1999).!° Still, whether a twin's

124 further criticism of the twins’ studies raised by Griliches (1979) is that measurement error in schooling may be exacerbated by differencing across twins, resulting in an even larger attenuation bias than the least-squares estimates This concern has been addressed in recent twins’ studies, which follow Ashenfelter and Krueger (1994) in using the other twin’s report of the first twin’s

schooling to instrument for the measurement error

'8For example, in their sample of twins, Behrman, Rosenzweig and Taubman (1994) find that virtually half of the twins had birth weight differences of at least 8 ounces Bound and Solon (1999) list a series of medical studies documenting the association between differences in birth

education is an appropriate instrument is not even relevant for the context of Ghana,

or other developing countries, where datasets on twins do not exist A summary of

twins’ studies is provided in Appendix Table A2.4

2.1.2 Developing Country Studies

Few studies of returns to schooling in developing countries use the best-practice stud-

ics of controlling for ability bias which have been delineated In the case of twins, the historic absence of such datasets for developing countries is the obvious cause;

in other cases, the cause is lack of appropriate questions in the surveys Still, a

number of studies have used the aforementioned techniques, and some of these will

be reviewed here

In the past fifteen years or so, the research activity using direct test-based controls for ability bias has actually been stronger in developing countries than elsewhere The test that has been used to identify ability in the developing country context has been

Raven’s Progressive Matrices test (Boissiere, Knight and Sabot, 1985; Glewwe, 1999;

Knight and Sabot, 1990; Glewwe, 1996) Reading and math tests have also been conducted, but are generally interpreted as cognitive skills resulting from education The Raven test has been insignificant in all wage regressions of which I am aware

(Boissiere, Knight and Sabot, 1985; Glewwe, 1999; Knight and Sabot, 1990; Glewwe,

weight and differences in IQ between twins Even if twins’ birth weights are identical, if one

twin possesses a stronger ”work-ethic” (a component of ability) than another, that twin might obtain a higher schooling level (because of its lower psychic cost), and also be rewarded for his/her determination independently in the labor market As Bound and Solon note (1999), citing work in the psychological literature, another plausible source of variation between twins is their psychological need to differentiate themselves from each other

1996) Therefore, while ability test scores have been used in developing countries, cither they have not measured ability accurately, or ability bias does not exist Regarding institutional innovations, given the general lack of strong enforcement

of compulsory schooling laws, when they do exist in developing countries, these have

not been used as instruments However, Duflo (2000) uses an institutional innovation

in the form of a inassive school construction initiative in Indonesia to identify the re-

turns to schooling, and finds a coefficient between 0.0675(0.0280) and 0.106(0.0222),

in comparison to her precisely estimated OLS coefficient of 0.077 Another instru-

ment used in the industrialized country literature is distance to the nearest college

The comparably relevant variable in developing countries is distance to the near-

est secondary school Maluccio (1998) uses this measure, distance to the nearest,

secondary school, as well as whether a private secondary school is located in the

nearest town in his study of the Philippines His results are comparable to those of the distance-to-college instruments, with the IV estimate being considerably higher

than the OLS estimate (IV: 0.145 (.041) vs OLS: 0.0730(.114)), using a standard

specification, plus gender and rural area controls

Using family background variables as control variables (Heckman and Hotz, 1986;

Armitage and Sabot, 1987) or as instruments (Schultz, 1995) is a bit more common in the returns to education literature in developing countries As in the developed coun- try literature, including family background variables as controls typically reduces the OLS estimates, while using them as instrumental variables increases the estimates.!"

l4For example, Heckman and Hotz in a regression for male heads of household in Panama, find

Nevertheless, the vast majority of developing country studies are simple OLS

16

regressions, without correcting for ability bias.'” A further bias which is considered

in some developing country regressions is the issue of selection (Schultz, 1988) In developed countries, the issue of selection bias resulting from participation in the

labor market has historically been strong in the case of women Frequently, this

issue has been avoided by using samples of men, for whom the bias is not nearly

as strong In the case of developing countries, the bias exists for both genders, as the issue of selection into the labor force is coupled with the issue of selection into wage versus self-employment Some studies have carefully handled this issue as well (Schafgans, 2000; Mwabu and Schultz, 2000; Lanzona, 1998; Vijverberg, 1993)

In summary, no consensus exists among labor economists on the best practice for

handling ability bias Developing country studies, in general, lack the attention to

this issue paid in developed country studies (partly due to the absence of appropriate

data, such as that of twins) In this context, this paper attempts to present a

very different, and very direct way of handling the issue of ability bias when linked

that the coefficient reduces from 0.1187(.0069) to 0.0856{.0074) once the education of both parents

is included a standard Mincerian regression, with an indicator for technical training as an additional control Schultz (1995), in a comparison study of Céte d’Ivoire and Ghana, finds that the use

of parental education and occupation, local health infrastructure and food prices as instruments increases the schooling parameter from 0.124(.007) (OLS) to 0.165(.040) (IV) in Cote d'Ivoire and decreases it from 0.0393(.004) (OLS) to 0.0214(.024) in Ghana, although the estimation in Ghana is imprecise Both these regressions also include controls for migration, body mass index and height, but not experience

15Glewwe (1999) presents the national returns to education for Ghana which follow a Mincer regression, and therefore are most comparable to our results, although the data was collected in 1988-89 (Ghana Living Standards Survey (GLSS) - Round 2) He finds that the OLS measure of the returns to education is 8.5% Using the GLSS data which is closest to the time period of the data under investigation here, the GLSS Round 4 data for 1998-99, we find that the OLS estimate has changed little, and is now 8.8% (author calculation)

employer-employee data is available The next section outlines the methodology of

the current study

2.2 Modeling the Production of the Firm

2.2.1 Incorporating the Heterogeneity of Labor in the Production Func-

tion

A frequent specification for production function estimation is the Cobb-Douglas form, which, written in logarithmic form, using lower cased variables, is:!°

where y is the value-added, / the labor, and k the capital of firm f in period ¢ Esti-

mation of equation (4) by least-squares raises two concerns The first, the problem of

simultaneity bias, has been understood in the literature at least since Jacob Marschak

and William H Andrews, Jr (1944), although truly satisfying solutions to this prob- lem have only arisen recently (Olley and Pakes, 1996; Levinsohn and Petrin, 2000) The second is the assumption of homogeneous labor, in that the variable / is typically

specified as the number of employees or the number of worker-hours at a firm, or

at best split into two types at the firm level Considerable effort, therefore, will be

given in this paper to incorporating the heterogeneity of labor into the production function First, however, let us briefly examine the problem of simultaneity, which will also be addressed in this paper

The simultaneity problem arises because the error term includes firm productivity,

'6Throughout this paper, upper-cased Roman letters will refer to the standard form of variables, and lower cased letters their natural logarithms

which is seen by the firm manager and will very likely be correlated with this period’s

labor input, which is typically considered to be freely variable It may not be

correlated with this period’s capital stock, which is generally considered a quasi-fixed

variable Under standard assumptions, this will result in an upward bias on the labor

coefficient, and possibly a downward bias on the capital coefficient (Levinsohn and

Petrin, 2000) The current state-of-the-art for handling this problem, as echoed in a

survey article on the problem by Griliches and Jacques Mairesse (1995), comes from

the work of Olley and Pakes (1996) In short, they separate the error term into firm

productivity, ws,, which is seen by the firm manager, and 7 ,,, a mean-zero component

which is not The production function then becomes:

ype = Bot Bly t+ Bykp Ð 0y + TỊ cụ (5)

In the Olley and Pakes model, the productivity term is derived, in the context of a

dynamic model, to be a function of investment and the firm’s capital stock, and is

calculated as a nonparametric function of these two variables Then, equation (5) can be estimated for observations where investment is non-zero While restricting

to observations with non-zero investment forces Olley and Pakes to lose 8 percent of

their observations, in other datasets (such as the current one) this can force deletion

of a much larger fraction, and sometimes the majority of observations (56 percent

of observations in the Ghanaian dataset under examination) To overcome this

limitation, Levinsohn and Petrin (2000) modify the Olley and Pakes procedure to

use intermediate inputs, instead of investment, in the estimation of firm productivity

This study modifies the Levinsohn and Petrin procedure by including human capital

variables in the production function, and the details of the procedure will be described

in Section 2.2.2

As mentioned, the other problem of simple estimation of equation (4) is the as-

sumption of homogeneity of labor This simplification reflects the limitations of the

data which have typically been available for estimation Fortunately, with the in-

creasing availability of linked employer-employee datasets, which provide data on at

least a sub-sample of, a firm’s employees, we can now ask the question of what is a

more sensible specification for labor’s contribution to firm product

A sensible place to begin is naturally in the labor literature Given its success at

explaining variation in wages across individuals, the human capital model of Mincer

(1974) has dominated the estimation of earnings equations over the past twenty-five

years In a Mincer human capital model, an individual invests in schooling until the

net present value of that investment is zero, that is the foregone present wage is equal

to the discounted value of the increased future wage resulting from an additional year of schooling As mentioned, the equation of estimation that results from the Mincer model is that of equation (2), where the log(wage) is a function of the years

of schooling, years of experience (or age) and ability of individual 7 If an individual

is rewarded according to her productivity, then the factors in (2) (or whatever factors are included in the wage equation being estimated) should also be included in the firm’s production function Moreover, given the success of (2) in explaining wages,

it is worth paying attention to the implications of this functional form for the firm

Equation (2) tells us that a firm’s wage bill is the following:!”

+

Note that each individual’s remuneration is a convex function of schooling, experience,

and ability To consider the implications of this formulation of the firm’s wage bill

on the firm’s production function, consider an individual factor of the above function,

say the overall level of schooling in the firm In particular, note that schooling is not linearly substitutable between individuals The cost to the firm of substituting

a current member of the workforce with a new worker, identical in every respect,

except with an additional year of schooling depends on which individual worker is

replaced, according to the functional form of (6) The same is true of other worker characteristics In fact, if the firm is optimizing (given that the firm is a price-taker

in wages), the productive contributions of workers’ characteristics should reflect the

relative costs of these characteristics In order for this to be the case, the labor term

in the production function should have the same form as (6) That the labor term

in the production function should have the same form as the wage bill for a profit-

maximizing firm is proven in Appendix A Therefore, the production function which

is consistent with a Mincerian wage equation is F[§) Le*9*Às5;*+ÀxXz*ÀA4; FC, Q, MỊ

Here, managerial ability is M, a further component of ñrm productivity (capturing

'7In this case, a linear experience term is used However, note that the analysis can be extended

to a quadratic in experience, as it is for estimation purposes in this paper In that case, the experience-squared term receives the same treatment as the schooling term and the experience term It is omitted from the exposition for simplicity

technological change, to be discussed later) is Q, and the number of workers of type

j with characteristics (S;,X;,A,;) is L; (so that )> 2; = L is the total number of

workers at the firm, and L; = 0 for types that are unused by the firm) Therefore,

defining the price of output as p, the profit function for the firm is:

l= pFI LjcÀ6tÀS8i *Àx X VÀ, K,Q, M] — = Ljerot rs Sit Ax Xs +ÀA4j+Ê; _ op ig (7)

The return to capital is simply r The choice variables for the firm are the

values for each of the L;’s, as well as K, with the first term of the production function

labelled as the quality of labor, or effective labor.!® While, by Appendix A, the

coefficients on an individual’s schooling in the production function must equal that

on the wage equation in the above profit function, this restriction will be tested, rather

than imposed on the data Therefore, the profit function that will be considered is

of the form:

= pFI" Le? *BsSi+BXX5+B aA; K,Q, M] _ s LjeÀ0*Às5¡ TÀx X; TÀA Ày tế —rK (8)

After estimating the production function, and the wage equation separately, the

equalities 6, = As, and By = Ax will be tested Following the literature standard,

the technology for F’ will be Cobb-Douglas, and therefore the production function to

30 and 90, respectively For consistency of exposition here, think of Aj; as taking on values between

0 and 100 The actual domain definitions are inconsequential for the analysis that follows, but are given for completeness All of the arguments that follow carry through to the continuous case

In equation (9), for the summation to calculate labor quality, the only types that

matter are those chosen in positive quantity (ZL; > 0), so that this term can be re-

indexed by the i = 1, , 2 workers chosen at a firm to get:

L

Y= efo(5~ cổ +BgSit By Xit+ByAs )Pu KP (eM e@)Pu ef (10)

¿=1 The subscripts for firm f and time t remain omitted for this discussion of functional form, but will be introduced subsequently If we use the logarithmic Cobb-Douglas

form, by taking logarithms of both sides of the equation, and representing each vari-

able’s natural logarithm using its lower-cased letter, then:

L

y = Bo t log(> efi + BsSitOxXi+BaABn 4 Bk + B(M+Q) +e (11)

¿=1

Now, if we observed measures for each of the variables in the estimation proce-

dure then estimation of (11) by non-linear least squares (for example) would provide

consistent estimation of the parameters Such estimation would require knowledge

of the distribution of workers at the firm (or at least a sample estimate of the distri-

bution), and their schooling, experience, and ability Unfortunately, MW, Q, and each

of the A;’s is unobserved Before being able to use the most recent techniques from the literature on estimating production functions for estimating equation (11), some

further work is needed First note that the factor e* is common across all workers,

aud can be factored, resulting in:

L

ƒ(%, 5Sr,, Xì, Ấ +, Ấn, Ár) = log()) eÖsS¡i*8xXi+BA4¡), A first-order Taylor

¡=1 approximation to f is the following:

ƒ(0,0, 0) + » Si (Elo + » Xj (Sele ) + » Aj (FF 0 )

+

Note that the first term, f(0, 0) = log(Le°) = log L Therefore, the labor term

which appeared to be missing from the specification of (11) is actually there, at least

when considering the Taylor expansion Now, examine the second term:

The third and fourth terms of (13) are comparable, so that the production function

of (12), using this first-order Taylor expansion becomes:

y = Bo t+ BrBu + By log L + ByBsS + By BxX + By BsAt+ Bek +B(M+Q) +e

(14) Note that in the production function estimation method of Olley and Pakes (1996) and

Levinsohn and Petrin (2000), the simultaneity in the production function is handled

by controlling for that portion of the productivity which is seen by the firm manager

and revealed through the firm’s investment or the firm’s use of intermediate inputs, conditional on the firm’s capital stock By that definition, the ability of workers

at a firm is clearly part of the firm’s productivity Therefore, we can reorganize

equation (14) and define firm productivity as w = Ø„/đaA + B,(M+Q) This

equation makes clear the key assumption which allows us to capture the worker ability

measure, namely that firm productivity consists of only these three components-

worker ability, management talent, and the third component, Q, which is designed to

capture technological change The further assumptions on management talent and

on technological change required for identification of worker ability will be outlined

in a later section Then equation (14) can now be expressed in a form that makes

clear how the production function estimation will proceed (letting 85 = 89 + B48 ,):

Use = BO + By log Ly + ByBsSpr+ BybyXp+ Bek twp ten (15)

Once the estimation of (15) is complete, the estimates of wy, can be decomposed

(details later) in order to achieve an estimate of the average worker ability at the firm, in order to control for ability in the wage equation estimation Note that the complicated term which we have defined as f can also be approximated by a second-

or higher-order Taylor expansion Details of these expansions, and a discussion of

their estimation are provided in Appendix B In general, these higher-order Taylor

expansions can easily be used for production function estimation However, in our

case, we wish to not only estimate the production function, but obtain a measure

of worker ability from this estimation The first-order Taylor expansion is the only order for which worker ability can be separated from the other components of firm

productivity, and therefore equation (15) is used for estimation

2.2.2 Estimating the Production Function

The key component of the production function estimation is the handling of firm pro-

ductivity, w Levinsohn and Petrin (2000) advocate replacing the use of investment

in the estimation procedure with intermediate inputs To consider the estimation

procedure, begin by considering a firm’s intermediate input function, 7, = t(wy, kyr)

It, should be noted that simply writing this function, i(wy:,ks,), assumes that firms

19 While these prices affect the input demand functions,

face the same input prices

given that the prices are common across firms, we can estimate the function using

only the two state variables noted.”” If intermediate input use is monotonically in-

creasing in productivity, conditional on the level of the capital stock, and Levinsohn

and Petrin (2000) provide sufficient conditions under which this is true, then the in-

termediate input function can be inverted to obtain an expression for productivity:

Equation (15) can then be rewritten as the following:

ype = Bylpt BsB Spe + Bx Bu Xp + dips, Kye) + "re (16)

P(t pe, kyr) = Bo + Øgkr + wlise, Kye)

Writing the production function in the form of equation (16) not only separates

those things that depend on the intermediate input and the capital stock from those

19\When Olley and Pakes (1996) define this function, they use the index t, to explicitly reflect the fact that the input prices are the same in a given period, ie é (wy:, ky) The lengths of the time periods that they use are either three or four years Given that the entire length of this study's dataset is four years, we are using a single period, and therefore drop the ¢ subscript at the outset, for clarity

“Yas in Olley and Pakes (1996) and Levinsohn and Petrin (2000)

that do not, it also separates the freely variable inputs from the productivity and

capital stock Note that schooling and experience in this formulation are treated as

freely-variable inputs, just as the labor variable input traditionally is Given that

the schooling and experience levels of the firm change with the hiring and firing of employees, it seems more natural to treat these variables as freely variable rather

than quasi-fixed While this specification does not allow for the tenure of employees

to play a role, this is primarily to keep consistency with the determinants of worker

productivity in the Mincer human capital model

Equation (16) is a partially linear model, which can be estimated semiparamet-

rically using a variety of methods to get consistent estimates for the variable-input

coefficients Following the method of Robinson (1988), taking the expectation of equation (16) conditional on %,, k, yields:

Elurdir, Ree] =

Ellnlir, k]8u + E[Šnlip, kr]8s8u + E[X ni k"]8xÖu + 9( ky) (17)

since i) E[ny lise, kp] = 0, and ii) E[d, (ise, kp lige, Kye] = Oise, ky) In the estimation

procedure, these conditional expectations are calculated using kernel density estima- tion with a normal (Gaussian) kernel Subtracting equation (17) from equation (16)

gives:

yp — Elyplisy kp) = (lp — Ellplin kel) By + (Sp — ElSplin, kp) B58

+(Xpp— EX pilin, kl) Bx Bu + 1p (18)

As a result, running no-intercept OLS with the modified variables of equation (18)

will provide a consistent estimate of 8;,, 858;;, and 8y8, Once these parameters

have been obtained, the contribution of the variable inputs can be subtracted from

equation (16), giving a new dependent variable, y*:

y= yt — Bylp — BsBy Spe — By By Xp = Bo + Bek top + "ht (19)

Therefore, in this first stage of the estimation procedure, the coefficients on the freely-

variable inputs are obtained In order to proceed, some minimal assumption is

required on firm productivity, and following Olley and Pakes (1996) and Levinsohn

and Petrin (2000), I assume that it follows a first-order Markov process,

0p = Blwplwpe—i] + Ep (20)

where € ;, is the mean zero innovation in wy, The expectation term and the intercept,

fy can be collected together into the function:

9(0/i—1) = Bo + Ew pele pra] (21)

so that equation (19) can be rewritten as:

Up = Bakye + g(wpr—r) + (Ep + 141) (22)

Fortunately, using the coefficients obtained from the first stage of estimation (equation

(18)) will provide an estimate of g(wy:1), as shall become evident below, and this can then be used in the estimation of equation (22) The restriction made for