88483 A WORLD BANK STUDY Working toward Better Pay EARNING DYNAMICS IN GHANA A N D TA N Z A N I A Paolo Falco, Andrew Kerr, Pierella Paci, and Bob Rijkers Working toward Better Pay A WO R L D BA N K S T U DY Working toward Better Pay Earning Dynamics in Ghana and Tanzania Paolo Falco, Andrew Kerr, Pierella Paci, and Bob Rijkers Washington, D.C. © 2014 International Bank for Reconstruction and Development / The World Bank 1818 H Street NW, Washington DC 20433 Telephone: 202-473-1000; Internet: www.worldbank.org Some rights reserved 1 2 3 4 17 16 15 14 This work is a product of the staff of The World Bank with external contributions. The findings, interpreta- tions, and conclusions expressed in this work do not necessarily reflect the views of The World Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Nothing herein shall constitute or be considered to be a limitation upon or waiver of the privileges and immunities of The World Bank, all of which are specifically reserved. Rights and Permissions This work is available under the Creative Commons Attribution 3.0 IGO license (CC BY 3.0 IGO) http:// creativecommons.org/licenses/by/3.0/igo. Under the Creative Commons Attribution license, you are free to copy, distribute, transmit, and adapt this work, including for commercial purposes, under the following conditions: Attribution—Please cite the work as follows: Falco, Paolo, Andrew Kerr, Pierella Paci, and Bob Rijkers. 2014. Working toward Better Pay: Earning Dynamics in Ghana and Tanzania. World Bank Studies. Washington, DC: World Bank. doi:10.1596/978-1-4648-0207-2. License: Creative Commons Attribution CC BY 3.0 IGO Translations— If you create a translation of this work, please add the following disclaimer along with the attribution: This translation was not created by The World Bank and should not be considered an official World Bank translation. The World Bank shall not be liable for any content or error in this translation. Adaptations—If you create an adaptation of this work, please add the following disclaimer along with the attribution: This is an adaptation of an original work by The World Bank. Responsibility for the views and opinions expressed in the adaptation rests solely with the author or authors of the adaptation and are not endorsed by The World Bank. Third-party content—The World Bank does not necessarily own each component of the content contained within the work. The World Bank therefore does not warrant that the use of any third-party-owned individual component or part contained in the work will not infringe on the rights of those third parties. The risk of claims resulting from such infringement rests solely with you. If you wish to re-use a com- ponent of the work, it is your responsibility to determine whether permission is needed for that re-use and to obtain permission from the copyright owner. Examples of components can include, but are not limited to, tables, figures, or images. All queries on rights and licenses should be addressed to the Publishing and Knowledge Division, The World Bank, 1818 H Street NW, Washington, DC 20433, USA; fax: 202-522-2625; e-mail: pubrights@worldbank.org. ISBN (paper): 978-1-4648-0207-2 ISBN (electronic): 978-1-4648-0209-6 DOI: 10.1596/978-1-4648-0207-2 Cover design: Debra Naylor, Naylor Design Library of Congress Cataloging-in-Publication Data has been requested. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Contents Acknowledgments ix About the Authors xi Abbreviations xiii Executive Summary 1 Chapter 1 Introduction 3 Chapter 2 What Did We Know about the Determinants of Earnings and Earnings Growth in Ghana and Tanzania? 7 The Determinants of Earnings Levels 7 The Determinants of Earnings Growth 8 Evidence on Low-Pay Persistence and Scarring 9 Notes 9 Chapter 3 Data and Descriptive Statistics 11 Ghana and Tanzania Urban Panel Surveys 11 Construction of Key Explanatory Variables 12 Descriptive Statistics 14 Notes 21 Chapter 4 The Determinants of Earnings Levels 23 Framework and Baseline Specification 23 Education Pays, and Pays More and More 24 Special Challenges for Youth and Women 27 Differences across Sectors 29 Notes 31 Chapter 5 The Determinants of Earnings Growth 33 Framework 34 A Bird’s Eye View of Earnings Growth in Ghana and Tanzania 35 Differences across Sectors 36 Notes 39 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2  v vi Contents Chapter 6 Low-Pay/High-Pay Transitions 41 Descriptive Statistics 42 Econometric Framework 43 Results 44 Notes 47 Chapter 7 Main Findings and Key Policy Implications 49 Message 1: Job Characteristics Are an Important Determinant of Both Earnings Levels and Earnings Growth 49 Message 2: Women and Youth Face Special Challenges 50 Message 3: Skills Acquisition Is a Stepping Stone Toward Better Paying Jobs, at Least in Wage Employment, Especially for Women 50 Message 4: Self-Employment Can Be Desirable 50 Message 5: The Public Sector Wage Premium Is a Potential Barrier to the Efficient Working of the Labor Market 50 Note 51 Appendix A Summary Statistics and Variable Definitions 53 Summary Statistics 53 Variable Definitions 56 Appendix B A Framework for Analyzing Earnings Panel Data 57 Tackling the Endogeneity of Schooling 57 Controlling for Unobserved Fixed Effects 58 Sorting Matters, But Is Not the Entire Story 60 What Do These Regressions Tell Us About Growth?   Asymmetric Sectoral Switching Premia 62 Appendix C A Framework for Analyzing Earnings Growth 65 Econometric Framework 65 Tackling Measurement Error: The Determinants   of Earnings Growth over a Two-Year Period 66 Controlling for Fixed Effects 68 Appendix D A Framework for Analyzing Transitions between Low- and High-Paid Employment 71 Econometric Framework 71 Results 75 Notes 78 Bibliography 79 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Contents vii Boxes 4.1 Key Hypotheses and Main Findings 24 5.1 Key Hypotheses and Main Findings 33 6.1 Key Hypotheses and Main Findings 41 Figures 3.1 Occupational Categories 14 3.2 Average Education by Occupation 15 3.3a Mean Earnings by Occupation (Ghana) 15 3.3b Average Education by Occupation (Tanzania) 15 3.4a Mean Earnings by Occupation and Gender (Ghana) 16 3.4b Mean Earnings by Occupation and Gender (Tanzania) 16 3.5a Mean Earnings by Occupation and Age (Ghana) 17 3.5b Mean Earnings by Occupation and Age (Tanzania) 17 3.6a Average Earning Changes (%) by Type of Transition (Ghana) 20 3.6b Average Earning Changes (%) by Type of Transition (Tanzania) 20 3.7 One-Year Earnings Growth 20 Tables 3.1 Panel Retention Rates 12 3.2 Correlation Coefficients between Skills Proxies 13 3.3 One-Year Transitions between Occupations 18 3.4 Two-Year Transitions between Occupations 18 3.5 Four-Year Transitions between Occupations (Ghana) 19 3.6 One-Year Transitions from Low to High Pay 21 4.1 Earnings Functions (OLS) 25 4.2 Earnings Functions—Controlling for Ability (OLS) 27 4.3 Earnings Functions by Age and Gender 28 4.4 Earnings Functions by Occupation 29 5.1 Determinants of One-Year Growth in Log Earnings 34 5.2 Determinants of One-Year Growth in Log Earnings   by Occupation 37 6.1a Raw Persistence in Ghana 42 6.1b Raw Persistence in Tanzania 42 6.2a Predicted Entry and Persistence Rates, and Mean and Median   Predicted Time in High and Low Pay for Ghana, Bivariate 45 6.2b Predicted Entry and Persistence Rates, and Mean and Median   Predicted Time in High and Low Pay for Tanzania, Bivariate 45 A.1 Summary Statistics (Ghana) 53 A.2 Summary Statistics (Tanzania) 53 A.3 Mean (1-Year) Changes in Log Earnings by Transition   Type (Ghana) 54 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 viii Contents A.4 Mean (1-Year) Changes in Log Earnings by Transition   Type (Tanzania) 55 B.1 Control Function Approach to Instrument Education and   Apprenticeships (First Stage) 59 B.2 Earnings Functions (FE and FD) 61 C.1 Determinants of Two-Year Growth in Log Earnings 67 C.2 FE and FD Estimates of Annual Earnings Growth 68 D.1 Bivariate Specification Tests in Tanzania and Ghana 73 D.2 Maximum Likelihood Coefficient Estimates for Bivariate Probit 76 D.3 Maximum Likelihood Coefficients for Trivariate Model, Ghana 77 D.4 Trivariate Specification Tests for Ghana 77 D.5 Predicted Entry and Persistence Rates and Mean and   Median Predicted Time in Low Pay for Ghana,   Trivariate Normal Model 78 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Acknowledgments We would like to thank the Centre for the Study of African Economies at Oxford University and in particular Francis Teal, Justin Sandefur, Andrew Zeitlin, and Simon Quinn for generously providing access to the data. We are grateful to Louise Cord, Louise Fox, John Giles, Mary Hallward-Driemeier, David McKenzie, David Newhouse, Emanuel Skoufias, Ana Revenga, and especially Francis Teal for useful comments and to the World Bank Gender Action Plan (GAP) Trust Fund for financial support. All errors are our own. The views expressed here are those of the authors and do not necessarily represent the views of the World Bank, its Executive Board, or member countries. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   ix About the Authors Paolo Falco is an ESRC postdoctoral fellow at the University of Oxford (Centre for the Study of African Economies). His research lies at the intersection of labour, behavioral, and experimental economics, with a focus on developing countries. He holds a PhD in economics from Oxford University and a BSc Economics degree from the University College London. He is a Junior Prize Fellow of the Royal Economics Society, and, in recent years, he has been a visit- ing scholar at the École Normale Supérieure (Paris School of Economics) and at the University of Copenhagen. His previous work experiences include collabora- tions with the World Bank, the International Monetary Fund, the Institute for Fiscal Studies, and the UK Independent Commission on Banking. Andrew Kerr is a senior research officer at DataFirst, University of Cape Town. He studied economics and applied mathematics at the University of Kwa Zulu-Natal, South Africa, before completing the MPhil and DPhil in economics at Oxford between 2006 and 2011. His research interests include labor and transport eco- nomics. He previously worked as a consultant for the South African Competition Commission and the World Bank, and he helped manage household and firm surveys in Tanzania and Ghana for the Centre for the Study of Africa Economies. Pierella Paci is lead economist in the Office of the Vice President of the Poverty Reduction and Economic Management Network (PREM). In her 15-year career at the World Bank, Ms. Paci has held technical, operational, and management posi- tions in the Europe and Central Asia region, in the Gender and Development Group, and in the Poverty Reduction and Equity Group where she led the Employment and Migration Team and the Shock and Crisis thematic area. Prior to joining the World Bank Group, she was assistant professor of economics at the University of Sussex (UK) and associate professor of economics at City University, London. Ms. Paci holds a degree in economics from the University of Rome (Italy) and a PhD in economics from the Victoria University of Manchester (UK). She has written extensively in the areas of labor economics, gender economics, inequality and poverty, jobs, and human development. Bob Rijkers is an economist in the Trade and International Integration Unit of the Development Economics Research Group in the World Bank Group. He is Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   xi xii About the Authors ­nterested in political economy, trade, and labor market issues. Since joining the i World Bank full time in 2008, he has worked in the Poverty Reduction Anchor of the PREM network, the Macroeconomics and Growth Unit of the Development Economics Research Group, and the Office of the Chief Economist of the Middle East and Northern Africa region. He holds a BA in science and social sci- ences from University College Utrecht, Utrecht University, and an MPhil and DPhil in economics from the University of Oxford. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Abbreviations ASD aggregate state dependence CF control function FD first difference FE fixed effect GDP gross domestic product GSD genuine state dependence HBS household budget survey LIC low-income country MIC middle-income country ML maximum likelihood NGO nongovernmental organization OLS ordinary least squares RDG reading UPS urban panel surveys WG within group Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   xiii Executive Summary This study uses the Ghana and Tanzania Urban Panel Surveys to examine the determinants of earnings, earnings growth, and low-pay/high-pay transitions in the high growth period 2004–08 and to identify communalities/differences across the two countries. The analysis highlights the importance of job character- istics in determining earnings and earnings growth. On average, even after controlling for ability bias, returns to cognitive skills and education are higher in wage employment than in self-employment, with the civil service and large firms paying the highest wages. However, the high within-sector heterogeneity in earnings means that self-employment is not always inferior to wage employment. Earnings growth is difficult to predict, and its drivers vary across sectors. Over the short horizons of this study, the sharpest changes in earnings were associated with job switches. These findings point toward path dependence in pay trajec- tories, and this conclusion is reinforced by the finding that being in low-paid employment has a scarring effect: it undermines future earnings prospects. Moreover, the determinants of low-pay incidence (that is, the risk of becoming low paid) differ from those of low-pay persistence (that is, the risk of remaining low paid). Women and young workers are especially likely to fall into, and remain trapped in, low-pay activities. Although the data reviewed for this study cover 2004–08, recent develop- ments in the economic situation and structure of the countries may have rein- forced the messages that emerge from this analysis. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2  1 CHAPTER 1 Introduction Improving access to productive employment is a key policy challenge, especially in low-income countries (LICs), where the only asset in abundance is labor. Since returns from working are the sole sources of income for these countries’ citizens, changes in labor earnings are more important in explaining changes in per capita household income than changes in any other source of income (Fields et al. 2003a, 2003b). Thus, policies that increase labor earnings help accelerate poverty reduction and growth. Individual workers can increase earnings by (i) working more hours; (ii) increasing labor productivity in a given job (if self-employed and/or if wages adjust to productivity); and (iii) moving to a job that offers higher returns per hours worked. Yet, the understanding of individual earnings dynamics remains limited (Fields 2008). A good deal of empirical literature has focused on identify- ing engines for, and barriers to, employment generation and, in particular, on the links between gross domestic product (GDP) growth and job creation. However, less attention has been paid to understanding the factors that lead to larger and faster pay increases, and when these factors are studied, the implicit assumption is often that they are the same as those that bring individuals out of low-pay employment. The small but growing body of literature points to strong persis- tence in earnings over time, but it remains unclear to what extent this persis- tence is due to individual heterogeneity rather than to the fact that being in a low-paying job itself undermines future earnings prospects. Shortages of longitudinal data are often quoted as the main reason for the limited understanding of earnings dynamics and their determinants in the devel- oping world. However, evidence from developed countries suggests that being in a low-paying job might have severe scarring effects (Cappellari and Jenkins 2004). Another relatively unexplored issue is the extent to which determinants of earnings vary across types of activities and sectors. For example, it is still unclear how the returns to skills vary across sectors and whether specific skills are valued differently in different sectors. Neither is it known whether the dif- ferences in returns to education between these types of activities reflect ability bias. Empirical evidence on the characteristics of successful entrepreneurs and Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2  3 4 Introduction how these might differ from those of wage workers also remains scarce, even though self-employment in microenterprises is an increasingly important source of income and employment in Africa (Fox and Gaal 2008) and other regions (International Labour Organization [ILO] 2002). These knowledge gaps are of particular concern in LICs, where poverty is less likely to be the consequence of a lack of employment, but rather of limited access to high productivity, well-paid jobs (see, for example, Johansson de Silva and Paci [2012]). In this context, the main policy challenge is to identify barriers to pro- ductivity and wage growth, rather than only employment creation. Building on ongoing research on earnings mobility, this study uses unusually rich longitudinal data from Ghana and Tanzania to identify engines of, and bar- riers to, earnings and earnings mobility. It examines the role of individual charac- teristics—such as gender, age, and skills—and characteristics of the job, but it also focuses on the role of job switches—for example, moves into and out of self- employment. It zooms in particularly on the drivers of transitions between low- paying and high-paying jobs, and addresses questions such as whether being low paid is a transitory or permanent phenomenon, and whether it has a scarring effect on an individual’s employment prospects. The extent to which earnings dynamics differ for women and young adults is also discussed in detail. Tanzania and Ghana provide a very relevant context in which to examine these issues. Tanzania’s National Strategy for Growth and the Reduction of Poverty emphasizes the creation of productive employment opportunities to support poverty reduction, highlighting the potential of self-employment to provide viable earning opportunities. Promoting entrepreneurship is also an important pillar of Ghana’s Poverty Reduction and Growth Strategy II. Job creation and enhancing returns to self-employment are progressively becoming more pressing policy issues, because both countries have experienced rapid growth in the pro- portion of the workforce that is self-employed outside agriculture. In particular, over recent years, as Ghana has become a middle-income country (MIC) and joined the club of oil producers, increasing access to productive employment for a growing part of the population has become an important pillar of the country’s inclusive growth strategy. Moreover, the cross-country comparison of earnings dynamics and labor mar- ket transitions helps shed light on the institutional factors that promote labor market mobility and entrepreneurship. The relevance of these results for policy making extends beyond these two countries. The structure of the Tanzanian and Ghanaian labor markets is typical of LICs, in which self-employment in small- scale activities accounts for a very large proportion of all employment (Kingdon, Sandefur, and Teal 2005). From a pragmatic point of view, the availability of unique, novel data sets (see chapter 3), which allow analysis of previously unex- plored policy issues, makes these countries very appealing case studies. This study next presents a brief review of related literature (chapter 2), fol- lowed by a descriptive overview of the labor markets in the two countries (chap- ter 3). The determinants of earnings levels are examined in chapter 4, and those Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Introduction 5 of earnings growth in chapter 5. Chapter 6 focuses on low-pay/high-pay transi- tions and analyzes whether the experience of being in a low-paying job under- mines an individual’s future earnings prospects. Finally, chapter 7 discusses key policy implications. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 2 What Did We Know about the Determinants of Earnings and Earnings Growth in Ghana and Tanzania? Empirical evidence suggests that the determinants of earnings and their growth in Ghana and Tanzania, as in other countries, depend not only on work­ ers’ human capital and gender but also on the location, sector, scale, and producti­ vity of the workplace. The Determinants of Earnings Levels Several studies show that human capital is an important determinant of earnings in both Ghana and Tanzania. Söderbom et al. (2006) found returns to education in Tanzania of between 6 and 13 percent, based on data from manufacturing sector surveys. Pissarides (2002) found that the return to education in Tanzania was 10 percent when estimated using household survey data from 1991, and 4 percent based on enterprise survey data from the same year. These estimates are roughly in line with those for developed countries (Card 2001). Quinn and Teal (2008) also found the returns to education in Tanzania to be convex, that is, the marginal returns rise with educational attainment. These findings are in line with previous results for Ghana (Rankin, Sandefur, and Teal 2007) and with those from developed economies (Belzil and Hansen 2002). These higher returns to education at higher levels of educational attainment may be driven by differences in activities and sectors rather than growth in earn­ ings within given jobs, since better education is strongly correlated with higher probability of finding better-paid employment. Fafchamps, Söderbom, and Benhassine (2009), for example, found that more than 50 percent of the educa­ tion-wage premia in manufacturing is accounted for by sectoral sorting. The estimated returns to education might also be driven by innate ability bias, due to more able individuals both acquiring more education and having better employ­ ment opportunities. However, separating the impact of innate ability from that of acquired education is rarely possible due to data limitations. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2  7 8 What Did We Know about the Determinants of Earnings and Earnings Growth in Ghana and Tanzania? Gender also plays an important role. Women typically earn less than men, both in wage employment and self-employment. The majority of microenter­ prises in African countries are operated by women, despite female-headed firms being typically less profitable than others. This suggests that women lack alterna­ tive income-earning opportunities (Mead and Liedholm 1998).1 Earnings gener­ ally also rise with age, and youth face difficulty finding high-paid employment, possibly due to lack of labor market experience and lack of financial capital to set up profitable enterprises. Wages and their determinants—especially returns to skills—vary substantially both across and within sectors and across locations. Indeed, a striking feature of developing country labor markets is the heterogeneity in earnings for workers with similar observable characteristics. For example, public and private formal sector employees are typically better paid, enjoy more stable employment, and better benefits than self-employed and informal sector workers. Earnings also rise with firm size: Söderbom et al. (2006), for example, found a strong positive association between firm size and wages in both Ghana and Tanzania, even after controlling for worker characteristics. This result survives the application of fixed-effect estimation techniques to control for unobserved determinants of earnings that might also be driving sorting of workers into larger firms. Using data from manufacturing firms in 10 countries, Fafchamps and Söderbom (2006) show that the positive correlation between firm size and wages is consistent with efficiency wage models based on moral hazard with costly supervision. The existence of pay differentials and differential rates of earnings growth for workers with similar observable characteristics across and within sectors in firms of different size may be indicative of labor market segmentation, or could reflect differences in unobserved skills or compensating differentials. Rankin, Sandefur, and Teal (2007) argue that the large wage differentials observed in Ghana for comparable workers across different sectors represent compelling evidence of existing labor market rigidities that keep wages above market-clearing level and lead to segmentation.2 Evidence from Latin America suggests that self-employ­ ment in small enterprises is largely a voluntary phenomenon (Maloney 1999). However, the extent to which these findings can be generalized to Africa, where countries are poorer and inequality lower remains questionable. The Determinants of Earnings Growth While the determinants of earnings levels have been the subject of a voluminous body of research, the determinants of earnings growth have received far less attention. Fields et al. (2003a, 2003b), using household data from Indonesia, South Africa, Spain, and Venezuela, found that job changes are the most impor­ tant factor behind earnings growth and that the roles of age and education are surprisingly weak. Similarly, Quinn and Teal (2008) found that education and age are not significantly correlated with earnings growth in Tanzania, and earnings rise more quickly for those with low levels of education. These findings raise the Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 What Did We Know about the Determinants of Earnings and Earnings Growth in Ghana and Tanzania? 9 question of whether acquisition of any skills leads to faster earnings growth, or whether specific skills are rewarded differently across sectors. Evidence on Low-Pay Persistence and Scarring Evidence from developed countries suggests also that being in a low-paying job has a negative effect on earnings prospects, a phenomenon referred to as labor market scarring.3 However, it is not clear to what extent these findings can be generalized to the less rigid labor markets of the developing world. Notes 1. Tanzania seems to deviate from this general pattern since most microenterprises in that country are operated by men. 2. For a more extensive overview of potential causes of sector differences, see Kingdon, Sandefur, and Teal (2005). 3. See for example, Cappellari and Jenkins (2004), Cappellari (2002), and Stewart and Swaffield (1999). Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 3 Data and Descriptive Statistics Ghana and Tanzania Urban Panel Surveys The Ghana and Tanzania Urban Panel Surveys (UPS) were designed by the Centre for the Study of African Economies at the University of Oxford to track the labor market experience of a representative sample of urban working-age individuals (ages 15–65) over several years. In both countries, surveys began in 2004 and respondents were subsequently visited at yearly intervals for three years in Tanzania (2004–06) and five years in Ghana (2004–08). The 2007 wave in Ghana, however, was obtained from recall questions administered during the 2008 survey, and this may have undermined its comparability with the other waves. The results presented in this study are robust to excluding 2007. The Ghana survey covers a stratified random sample of urban households from the 2000 census. In Tanzania, the sample was drawn from the households visited by the 2000–01 Household Budget Survey (HBS), conducted by the Tanzania Bureau of National Statistics, with additional randomly selected house- holds added in 2006.1 The surveys have a number of strengths. In addition to the longitudinal dimension, the fact that the UPS were designed to record the net earnings of both wage earners and the self-employed is an obvious advantage, because it enables comparisons between these two categories. The strong comparability of the two surveys also facilitates comparisons across countries. In addition, in 2005 and 2006, respondents undertook several tests specifically designed to measure their mathematical and verbal skills and their noncognitive abilities. This detailed and relatively uncommon information makes it possible to separate the impact of innate intelligence, acquired skills, and education on earnings levels and dynamics. However, a drawback of the data is the relatively high attrition rate, especially in Tanzania. Table 3.1 presents the percentages of respondents interviewed in the first year (2004) who were interviewed again in any of the following years. However, the model predicting attrition suggests that attrition is largely random, and therefore not a strong concern from an econometric point of view.2 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   11 12 Data and Descriptive Statistics Table 3.1  Panel Retention Rates Country 2005 2006 2008 Ghana 0.79 0.63 0.40 Tanzania 0.75 0.45 — Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: — = not available. Another potential limitation is the fact that the data cover only the period up to 2008. Since then, both countries have continued to enjoy rapid growth, and Ghana recently joined the middle-income category and the oil produc- ers’ club. Construction of Key Explanatory Variables This section discusses only the most important variables. For an overview of other variables used in the analysis, see appendix A. Earnings. Earnings are defined as pretax monthly earnings. For wage workers, they include average bonuses and allowances received in any given month. For the self-employed, they are proxied by a measure of monthly profits obtained after guiding them through the concepts of business revenues and costs. Thus, while the earnings of wage workers purely capture the returns to labor, the earn- ings of the self-employed may also reflect returns to capital as well as the contri- butions of unpaid workers (who may be members of the same household). Since only 19 percent of the self-employed report hiring any paid or unpaid workers in both countries, the latter issue is likely to be of second-order magnitude and the regression analysis partially corrects for this bias by controlling for the total number of employees. Occupational Categories. Paid workers are divided into three main categories: (i) self-employed entrepreneurs (with or without employees), (ii) wage earners in private firms, and (iii) civil servants.3 All respondents not employed for pay fall into a residual category of unpaid workers. This includes students, unpaid family workers, unpaid apprentices, working-age individuals who are temporarily or permanently out of the labor force, and unemployed job seekers. Skills Variables. The survey includes four ability tests: a mathematics test, a language comprehension test, a reading ability test, and the Raven’s Matrices, which are designed to measure the cognitive and noncognitive abilities of the respondent. In this study, the term “cognitive skills” refers to those skills that are developed or improved through schooling and education—that is, literacy and ability to perform mathematical calculations. Noncognitive skills, on the other hand, are considered either innate—that is, genetically inherited—or the product of early development. The mathematics test was a combination of practical problems (such as computing the duration of a trip, given distance and speed) and general arith- metic. The language comprehension test required respondents to read a short Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Data and Descriptive Statistics 13 text and answer questions related to its contents, and the reading test required respondents to read aloud a series of words and an entire sentence, and translate a series of English words into their native language. The non- cognitive skills Raven’s Matrices required the respondent to understand the pattern linking several objects within a matrix by simple intuition and logic, and complete the matrix accordingly. Points were awarded for fluency of reading and correct translations. The math and language comprehension tests were administered both in 2005 and 2006, but, despite maintaining compa- rable structure and contents, they were changed between 2005 and 2006. The reading ability and Raven’s Matrices were added in 2006. For compara- bility’s sake and to retain the maximum sample size, the scores of respon- dents who took the test in both waves were averaged so that the skills cap- tured by such tests are time invariant over the relatively short time frame covered by the surveys. Table 3.2 shows pairwise correlations between skills proxies and years in formal education. With the exception of two cases in Tanzania, all correla- tions between scores are positive, but those between the math and the lan- guage comprehension scores are much higher in both Ghana and Tanzania than the correlation between the Raven’s score and linguistic abilities. This provides empirical support for the assumption that both math and literacy skills are functions of schooling, while noncognitive skills (proxied by the Table 3.2  Correlation Coefficients between Skills Proxies Ghana Language Raven’s comprehension Math Reading Matrices Education Language comprehension   1 Math 0.66   1 Reading 0.33 0.51    1 Raven’s Matrices 0.37    0.478 0.29    1 Education 0.37 0.44 0.53 0.19 1 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Tanzania Language Raven’s comprehension Math Reading Matrices Education Language comprehension     1 Math 0.65 1 Reading 0.36 0.27   1 Raven’s Matrices 0.19 0.25 0.11   1 Education –0.01 –0.03 0.52 0.33 1 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 14 Data and Descriptive Statistics Raven’s score) are innate. The hypothesis is also supported by the fact that formal education in Ghana is highly correlated with literacy (particularly reading abilities) and math skills, but only weakly correlated with innate abil- ity (Raven’s). In Tanzania, the correlations between scores in language com- prehension and math and the years in school are negative, but very close to zero and insignificant, while the correlation between education and reading abilities is stronger. What accounts for these differences across countries is not clear. Descriptive Statistics Occupational Categories Figure 3.1 presents the pooled data on the percentage of workers in different occupations from all survey waves. The self-employed are by far the largest group, and wage employment covers approximately 25 percent of the sample in both countries. However, at nearly one-third of wage employment, the pub- lic sector accounts for a considerably larger share of employment in Tanzania than in Ghana. In both countries, close to one quarter of the working-age popu- lation is either unemployed or out of the labor force and public sector employ- ees have the highest average educational attainment followed by private sector wage workers (figure 3.2). The self-employed have the lowest average levels of formal education, even lower than the average education levels of those who are not working. Mean Earnings by Occupation As shown in figures 3.3a and 3.3b, in both countries, mean earnings are highest in the public sector and in large private enterprises and lowest for entrepreneurs without employees and wage earners in small firms.4 Figure 3.1  Occupational Categories Ghana Tanzania Self Self Other unpaid/ employment Other unpaid/ employment unemp/inact 35.72% unemp/inact 44.83% 26.46% 23.84% Unpaid Unpaid family work family work 1.927% 2.434% Student Student 3.645% 10.46% Public wage Public wage 8.714% 4.433% Private wage Private wage 20.49% 17.05% Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Data and Descriptive Statistics 15 Figure 3.2  Average Education by Occupation Self-employment Private wage Public wage Student Unpaid family work Other unpaid/unemp/inact 0 5 10 15 Years Ghana Tanzania Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Figure 3.3a  Mean Earnings by Occupation (Ghana) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50 100 150 GHS Figure 3.3b  Mean Earnings by Occupation (Tanzania) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50,000 100,000 150,000 200,000 TZS Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: GHS = Ghana cedis; TZS = Tanzanian shillings. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 16 Data and Descriptive Statistics Figures 3.4a, 3.4b, 3.5a, and 3.5b show the existence of a significant wage disadvantage for women and workers under age 30 in every sector. In Ghana, the gender gap is largest among the self-employed, while in Tanzania, it is most pronounced in small private firms. The youth disadvantage in Ghana is most striking among the self-employed with employees and small firm workers, while in Tanzania it is most striking in the public sector. Sector Transitions Table 3.3 shows that the most persistent occupational category over a one-year period is self-employment, which has a retention rate of more than 80 percent in both countries. Overall, the percentage of workers who move between occu- pations is surprisingly high. The self-employed and the private wage workers display similar rates of transition out of paid employment (11 to 14 percent), and the rate is only marginally lower for public sector employees (10 percent for Figure 3.4a  Mean Earnings by Occupation and Gender (Ghana) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50 100 150 200 GHS Male Female Figure 3.4b  Mean Earnings by Occupation and Gender (Tanzania) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50,000 100,000 150,000 200,000 TZS Male Female Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Data and Descriptive Statistics 17 Figure 3.5a  Mean Earnings by Occupation and Age (Ghana) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50 100 150 200 GHS <30 >30 Figure 3.5b  Mean Earnings by Occupation and Age (Tanzania) Self-employed without employees Self-employed with employees Private wage in small firm Private wage in large firm Civil service Public enterprise 0 50,000 100,000 150,000 200,000 TZS <30 >30 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Ghana and 11 percent for Tanzania). Transitions into paid employment, on the other hand, lead predominantly to self-employment in Tanzania (27 percent), but are equally distributed between self-employment and private wage employ- ment in Ghana (12 percent for both).5 As expected, mobility is considerably higher over a two-year horizon. As table 3.4 shows, only 69 percent of the self-employed and 51 percent of pub- lic employees in Ghana—and 61 and 45 percent, respectively, in Tanzania— retained their jobs. Perhaps even more surprisingly, more than 25 percent of public employees (29 and 27 percent for Ghana and Tanzania, respectively) moved to the private sector within the two-year horizon, and 65 percent of the unpaid moved into paid employment in Tanzania, although only 43 per- cent did so in Ghana. Finally, transitions over four years, presented in table 3.5 for Ghana, show a further reduction in the retention rates with the strongest flows being those Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 18 Data and Descriptive Statistics Table 3.3  One-Year Transitions between Occupations Ghana Tanzania Priv Pub Priv Pub Self wage wage Unpaid Total Self wage wage Unpaid Total Self 1,193 103 4 163 1,463 Self 414 23 10 69 516 81.54 7.04 0.27 11.14 100 80.23 4.46 1.94 13.37 100 Priv wage 87 568 25 109 789 Priv wage 8 89 48 24 169 11.03 71.99 3.17 13.81 100 4.73 52.66 28.4 14.2 100 Pub wage 7 26 117 17 167 Pub wage 3 17 45 8 73 4.19 15.57 70.06 10.18 100 4.11 23.29 61.64 10.96 100 Unpaid 184 180 22 1,138 1,524 Unpaid 26 10 3 59 98 12.07 11.81 1.44 74.67 100 26.53 10.2 3.06 60.2 100 Total 1,471 877 168 1,427 3,943 Total 451 139 106 160 856 37.31 22.24 4.26 36.19 100 52.69 16.24 12.38 18.69 100 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: Self is self-employment, Priv wage is private wage employment, Pub wage is public wage employment, Unpaid is an unpaid family worker. Table 3.4  Two-Year Transitions between Occupations Ghana Tanzania Priv Pub Priv Pub Self wage wage Unpaid Total Self wage wage Unpaid Total Self 642 120 7 157 926 Self 127 22 9 51 209 69.33 12.96 0.76 16.95 100 60.77 10.53 4.31 24.4 100 Priv wage 79 265 19 109 472 Priv wage 7 33 34 18 92 16.74 56.14 4.03 23.09 100 7.61 35.87 36.96 19.57 100 Pub wage 7 28 49 12 96 Pub wage 2 3 5 1 11 7.29 29.17 51.04 12.5 100 18.18 27.27 45.45 9.09 100 Unpaid 184 177 27 510 898 Unpaid 6 4 3 7 20 20.49 19.71 3.01 56.79 100 30 20 15 35 100 Total 912 590 102 788 2,392 Total 142 62 51 77 332 38.13 24.67 4.26 32.94 100 42.77 18.67 15.36 23.19 100 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: Self is self-employment, Priv wage is private wage employment, Pub wage is public wage employment, Unpaid is an unpaid family worker. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Data and Descriptive Statistics 19 Table 3.5  Four-Year Transitions between Occupations (Ghana) Ghana Self Priv wage Pub wage Unpaid Total Self 144 39 4 59 246 58.54 15.85 1.63 23.98 100 Priv wage 17 37 7 18 79 21.52 46.84 8.86 22.78 100 Pub wage 3 7 7 5 22 13.64 31.82 31.82 22.73 100 Unpaid 50 50 7 83 190 26.32 26.32 3.68 43.68 100 Total 214 133 25 165 537 39.85 24.77 4.66 30.73 100 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: Self is self-employment, Priv wage is private wage employment, Pub wage is public wage employment, Unpaid is an unpaid family worker. in/out of paid employment. At 32 percent, the flows into private wage employment from public sector employment are double those from self-employment. Figure 3.6a and 3.6b report the changes in average earnings associated with different transitions in the two countries. Though it is impossible to draw conclu- sions on the relative financial gains associated with different transitions by simply looking at these figures, it is noticeable that transitions are generally associated with real earnings gains, both in absolute terms and relative to those who stay in a given occupation. For example, on average, the few workers who move from self-employment to a private sector wage job experience higher earnings growth than those who remain self-employed. Interestingly, however, transitions in the opposite direction also lead to higher earnings. Earnings Growth Figure 3.7 shows average one-year earnings growth in different occupations over 2004–08 for Ghana, and 2004–06 for Tanzania. On average, the earnings of the self-employed have grown faster than wages. The caveat with this finding, how- ever, is that measurement error and transitory earnings shocks are both more prominent among the self-employed and therefore might have had an impact. Moreover, in Ghana, wages have grown relatively faster in the private sector than in the public sector, while this difference is negligible in Tanzania. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 20 Data and Descriptive Statistics Figure 3.6a  Average Earning Changes (%) by Type of Transition (Ghana) Self -> self Self -> priv wage Self -> public wage Priv wage -> self Priv wage -> priv wage Priv wage -> public wage Public wage -> self Public wage -> priv wage Public wage -> public wage –0.5 0 0.5 1.0 1.5 Figure 3.6b  Average Earning Changes (%) by Type of Transition (Tanzania) Self -> self Self -> priv wage Self -> public wage Priv wage -> self Priv wage -> priv wage Priv wage -> public wage Public wage -> self Public wage -> priv wage Public wage -> public wage 0 0.2 0.4 0.6 0.8 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: Self is self-employment, Priv wage is private wage employment, Pub wage is public wage employment. Figure 3.7  One-Year Earnings Growth Ghana Self-employed Private wage Public wage Tanzania Self-employed Private wage Public wage 0 .05 .10 .15 .20 TZS Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Data and Descriptive Statistics 21 Table 3.6  One-Year Transitions from Low to High Pay Pay status at time t Pay status at time t-1 Low High Low High Country Tanzania Ghana Low 55.22 44.78 54.29 45.71 High 12.86 87.14 15.22 84.78 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Low-pay/High-pay Transitions In both countries, over 25 percent of paid workers are low paid (29 percent in Tanzania and 27 percent in Ghana).6 However, table 3.6 shows this to be mostly a transitory state in both countries as nearly half of the low-paid move to high- pay occupations within the two years covered by the analysis. In addition, upward mobility is considerably higher than downward mobility: only 15 per- cent (or less) of the high-paid workers falls into low pay during this period. Notes 1. For more information on the sampling strategy, visit http://www.csae.ox.ac.uk/datas- ets/Ghana-Tanz-UHPS/default.html. 2. The analysis also attempted to correct for attrition bias using the inverse probability weighting method proposed by Moffit, Fitzgerald, and Gottschalk (1999). The results (available from the authors), suggest that the qualitative pattern of results is robust to correcting attrition bias, though individual coefficient estimates occasionally changed. 3. Civil servants are defined as those working for the government, international organiza- tions, or nongovernmental organizations (NGOs) and employees in state-owned enter- prises. NGOs workers are categorized as civil sector employees since they are likely to benefit from higher wages and to face a wage-setting structure that is more akin to that prevailing in the civil service sector than that operating in the private sector. 4. Small firms are defined as firms with less than 10 employees. 5. High mobility might reflect measurement error; however, over longer time horizons, measurement error ought to be less of a concern as the signal-to-noise ratio in the data should increase. In other words, over longer horizons, a smaller share of the transitions will be due to measurement error instead of genuine switching. 6. A worker/job is defined as low paid if s/he/it earns/pays less than US$1.25 a day. This is in line with the US$1.25 a day poverty line defined by Chen and Ravallion (2008). However, it is important to note that, contrary to typical poverty analyses, the focus here is on low pay rather that poverty. Thus, the low-pay line is defined in terms of individual earnings rather than equivalized household consumption; thus, the low- earnings threshold is conservative. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 4 The Determinants of Earnings Levels In this chapter we explore the determinants of earnings levels using the Tanzanian and Ghanaian Urban Panel Surveys. We explore the extent to which human capital determines earnings, as well as the challenges faced by women and young people in employment. We also discuss other determinants of earn- ings and whether these determinants differ by sector. See box 4.1 for the key hypotheses we test in this chapter and a summary of its main results. Framework and Baseline Specification The starting point of the analysis is a semilogarithmic Mincerian earnings equa- tion, where the log of monthly earnings is modeled as function of the respon- dent’s gender; height; age and age squared (to allow for nonlinear effects of age on earnings); a self-reported measure of tenure; the log of hours worked; years of schooling and its square (to allow for nonlinear effect of education); firm size; a dummy indicating whether the respondent has ever completed an apprentice- ship; and sector, city, and year dummies. The results of a simple ordinary least squares (OLS) regression are presented in columns 1 (Ghana) and 5 (Tanzania) of table 4.1. A number of patterns are qualitatively similar across countries: • Earnings rise with age and experience. Age-earnings profiles are concave: earn- ings rise quickly with age, but the increase declines as workers grow older, and eventually becomes negative. At age 42 in Tanzania and 43 in Ghana, the turning point is remarkably similar in the two countries. Even after control- ling for age, self-reported tenure is associated with significantly higher earn- ings levels in Ghana, but the estimated association is insignificant in Tanza- nia. This positive correlation between tenure and earnings may reflect the accumulation of job-specific human capital or the fact that better matches between workers and firms are more likely to survive. • They also raise with firm size. The association between firm size and earnings is especially pronounced for entrepreneurs, potentially reflecting returns to Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   23 24 The Determinants of Earnings Levels Box 4.1 Key Hypotheses and Main Findings Hypothesis 1: Education and skills play a crucial role in determining earnings. Main finding: The marginal returns to education are high and increase with the level of educa- tional attainment. • The estimated returns to education are robust to controlling for ability bias. • Cognitive skills are significant predictors of earnings. • Apprentices are either unpaid or paid significantly less than other workers. • Apprenticeships do not lead to higher earnings. Hypothesis 2: Women and youth face special challenges. Main finding: Women/youth earn systematically less than men/older workers and have different returns to education. • The gender differential is highest among the self-employed and lowest in the public sector. • Women’s earnings grow slower with age and reach a peak later in life than those of men. • Women experience positive returns to primary education, while men only benefit from postprimary education. • Returns to education are higher and more highly convex among youth. Hypothesis 3: The determinants of earnings vary by sector. Main finding: The key factors determining earnings vary substantially across sectors and across firms. • The returns to education and cognitive skills are highest in wage employment. • Large firms pay workers with same observable characteristics more than small firms do . capital. However, this could also be interpreted as evidence of the existence of a wedge between productivity and wages paid by large enterprises. • Men tend to earn more than women, even after controlling for educational at- tainment, other observable characteristics, and job type. The gender gap is higher in Tanzania than in Ghana. Education Pays, and Pays More and More The coefficients on education presented in columns 1 and 5 in table 4.1 suggest that education and skills are associated with increased earnings and that the impact rises with the level of education, that is, the returns to education are convex. Furthermore, acquired skills appear more important in determining earnings than innate ability. A common concern when estimating the returns to education is that more able individuals have higher potential wages, but are also likely to obtain more education and this leads to spurious high estimates of the returns to schooling. To control for this potential bias, the analysis initially uses a conven- tional control function (CF) approach, which includes the residuals of a model of educational attainment that uses distance to school as an exclusion restriction. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Levels 25 Table 4.1  Earnings Functions (OLS) Ghana Tanzania Dependent ­variable: log of CF— CF— monthly earnings CF— education + OLS with CF— education + OLS with OLS education apprentice ability OLS education apprentice ability Male 0.240*** 0.163 0.161 0.203*** 0.334*** 0.318*** 0.309*** 0.329*** (0.049) (0.108) (0.108) (0.050) (0.054) (0.056) (0.061) (0.054) Age 0.075*** 0.061*** 0.061*** 0.078*** 0.044*** 0.037** 0.048 0.042*** (0.014) (0.023) (0.023) (0.014) (0.016) (0.018) (0.034) (0.016) (age^2)/100 –0.089*** –0.069** –0.069** –0.093*** –0.045** –0.032 –0.042 –0.041** (0.019) (0.032) (0.032) (0.018) (0.020) (0.024) (0.038) (0.020) Height (cm) 0.005 0.005 0.005 0.006 0.004 0.003 0.003 0.003 (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Years in formal education –0.040** 0.008 0.006 –0.037** –0.056*** 0.007 0.023 –0.053*** (0.019) (0.065) (0.065) (0.019) (0.020) (0.069) (0.074) (0.020) (educ^2)/100 0.004*** 0.004*** 0.004*** 0.004*** 0.007*** 0.007*** 0.007*** 0.007*** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) Apprentice (currently) –0.786*** –0.785*** –0.765*** –0.749*** (0.088) (0.088) (0.134) (0.087) Apprenticeship completed –0.039 –0.041 –0.010 –0.031 0.023 0.018 –0.072 0.021 (0.049) (0.050) (0.132) (0.049) (0.094) (0.095) (0.261) (0.095) Ln (hours) 0.223*** 0.225*** 0.224*** 0.217*** 0.073 0.079 0.078 0.077 (0.061) (0.061) (0.061) (0.061) (0.085) (0.085) (0.085) (0.084) Tenure, self–reported 0.013*** 0.013*** 0.013*** 0.013*** 0.006 0.006 0.006 0.006 (0.003) (0.003) (0.003) (0.003) (0.004) (0.004) (0.004) (0.004) Ln (employees) 0.296*** 0.296*** 0.296*** 0.289*** 0.484*** 0.487*** 0.486*** 0.471*** (0.060) (0.061) (0.061) (0.061) (0.078) (0.078) (0.078) (0.079) Ln (firm size) 0.171*** 0.171*** 0.172*** 0.160*** 0.137*** 0.136*** 0.136*** 0.132*** (0.019) (0.019) (0.019) (0.019) (0.022) (0.022) (0.022) (0.022) Private wage –0.256*** –0.257*** –0.258*** –0.252*** –0.056 –0.055 –0.053 –0.049 (0.065) (0.065) (0.065) (0.065) (0.084) (0.084) (0.084) (0.083) Civil servant 0.566*** 0.564*** 0.564*** 0.503*** 0.474*** 0.474*** 0.476*** 0.450*** (0.087) (0.087) (0.087) (0.086) (0.087) (0.087) (0.088) (0.088) Public enterprise –0.161 –0.158 –0.159 –0.175 0.269** 0.270** 0.273** 0.247** (0.119) (0.120) (0.120) (0.119) (0.128) (0.128) (0.128) (0.125) Residual education –0.047 –0.046 –0.064 –0.080 (0.061) (0.061) (0.068) (0.074) Residual appren- ticeship –0.032 0.092 (0.129)   (0.252) Math score 0.004*** 0.003** (0.001) (0.002) City and year dummies Yes Yes Yes Yes Yes Yes Yes Yes table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 26 The Determinants of Earnings Levels Table 4.1  Earnings Functions (OLS) (continued) Ghana Tanzania Dependent ­variable: log of CF— CF— monthly earnings CF— education + OLS with CF— education + OLS with OLS education apprentice ability OLS education apprentice ability Constant –1.158* –1.296** –1.270* –1.308** 8.586*** 8.395*** 7.928*** 8.471*** (0.639) (0.649) (0.658) (0.632) (0.697) (0.717) (1.442) (0.694) Number of   2,610  2,610   2,610  2,610  1,328  1,328  1,328   1,328 observations R2 0.310 0.310 0.310 0.316 0.304 0.305 0.305 0.308 Adjusted R2 0.304 0.304 0.304 0.310 0.293 0.293 0.293 0.297 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Standard errors are in parenthesis. CF= control function; OLS = ordinary least squares. ***p<0.01       **p<0.05         *p<0.1 However, as discussed in appendix B, the control function approach relies on rather restrictive assumptions, as well as on being able to successfully predict educational attainment. Thus, it is better to adopt a more direct methodology by adding to the equation an explanatory variable for individual ability proxied by performance on mathematics, English, reading, and Raven’s tests.1 The impact of ability bias on the education results appears to be minimal. As shown in columns 2 and 5 of table 4.1, the residual of the model that uses the CF approach to predict educational attainment is never significant. The esti- mated returns to education do not change when the residual is included and, when they do, they become more positive, suggesting an upward rather than downward bias. The third specification—presented in columns 3 and 6—also includes a residual of a model that predicts the probability of being an apprentice and again is not significant. However, the estimated coefficient on doing an apprenticeship rises substantially, perhaps indicating that less able individuals sort into apprenticeships, as suggested by Kahyarara and Teal (2008). The estimated returns to education are robust to including proxies for unob- served skills. Columns 4 and 8 show that the inclusion of an indicator of math- ematical ability does not significantly affect returns to education. Other mea- sures of ability are not included because mathematical ability is highly correlated with other skills (see table 3.2) and inclusion of the other proxies excessively reduces the sample size (see chapter 3). For purposes of comparability, the results of using different ability proxies are presented in table 4.2. The only indi- cators of ability that have a statistically significant positive impact on earnings when entered individually are the Raven’s and English score in Ghana and the English and reading scores in Tanzania. When entered jointly, the proxies tend to lose their individual significance and show the “wrong” sign, presumably because they are highly correlated and the samples are small, especially in Tanzania. However, the ability proxies remain jointly significant in Ghana. More importantly, the inclusion of the proxies does not significantly affect the Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Levels 27 Table 4.2  Earnings Functions—Controlling for Ability (OLS) Dependent Ghana Tanzania variable: log of monthly earnings (1) (2) (3) (4) (5) (1) (2) (3) (4) (5) Math score 0.004** 0.004*** 0.001 0.003* (0.002) (0.001) (0.002) (0.001) Raven’s score 0.002 0.003*** –0.002 –0.002 (0.001) (0.001) (0.002) (0.002) English score 0.002 0.001** 0.002* 0.003** (0.001) (0.001) (0.001) (0.001) RDG score –0.003** 0.000 0.001 0.004*** (0.001) (0.001) (0.002) (0.001) Years in formal education –0.053* –0.044** –0.041* –0.046*** –0.068*** –0.059 –0.057*** 0.004 –0.042 –0.055*** (0.028) (0.018) (0.022) (0.017) (0.020) (0.036) (0.020) (0.030) (0.034) (0.020) (educ^2)/100 0.438** 0.412*** 0.440*** 0.465*** 0.570*** 0.735*** 0.754*** 0.441** 0.622*** 0.739*** (0.170) (0.124) (0.145) (0.122) (0.131) (0.222) (0.157) (0.199) (0.209) (0.161) Apprentice (currently) –0.771*** –0.758*** –0.765*** –0.788*** –0.819*** (0.105) (0.088) (0.095) (0.083) (0.095) Have you ever been –0.027 –0.021 0.008 –0.037 –0.063 0.099 0.020 0.103 0.079 0.030 an appren- tice? (0.062) (0.049) (0.056) (0.048) (0.054) (0.090) (0.091) (0.089) (0.089) (0.092) Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observa- tions 1,663 2,736 2,086 2,853 2,236 1,031 1,394 1,082 1,051 1,363 R2 0.338 0.313 0.309 0.294 0.309 0.338 0.312 0.330 0.333 0.318 Adjusted R2 0.328 0.307 0.301 0.288 0.302 0.321 0.300 0.316 0.318 0.306 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Controls included but not presented to conserve space: age, age squared, tenure, hours worked, firm size, sector, year dummies, city dummies. Standard errors are in parenthesis.OLS = ordinary least squares; RDG = Reading. ***p<0.01    **p<0.05    *p<0.1 estimated returns to education or to doing an apprenticeship. The estimated returns to education are also robust to controlling for Raven’s test scores, which, arguably, is a relatively clean proxy for innate ability. Special Challenges for Youth and Women As shown in table 4.3, differences in the determinants of earnings emerge between men and women, and between younger and older workers:2 • Gender gaps in earnings exist in all sectors. The gap is highest among the self- employed and lowest in the public sector. To the extent that this is driven by discrimination, the public sector appears to be a more gender-sensitive employer. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 28 The Determinants of Earnings Levels Table 4.3  Earnings Functions by Age and Gender Dependent variable: Ghana Tanzania log of monthly ­earnings Women Men Young Old Women Men Young Old Male 0.172** 0.222*** 0.182** 0.358*** (0.068) (0.071) (0.092) (0.062) Age 0.064*** 0.098*** 0.037*** –0.000 0.024 0.069*** 0.046*** 0.003 (0.022) (0.017) (0.010) (0.004) (0.022) (0.022) (0.015) (0.004) (age^2)/100 –0.069** –0.121*** –0.025 –0.071** (0.029) (0.023) (0.028) (0.028) Height (cm) 0.003 0.005 0.008 0.003 –0.004 0.010* 0.004 0.002 (0.005) (0.004) (0.005) (0.005) (0.005) (0.005) (0.007) (0.004) Years in formal education –0.031 –0.068*** –0.051* –0.030 –0.049** –0.064** –0.089** –0.043* (0.025) (0.023) (0.030) (0.023) (0.025) (0.031) (0.037) (0.022) (educ^2)/100 0.004** 0.004*** 0.004** 0.003** 0.007*** 0.008*** 0.009*** 0.007*** (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.003) (0.002) Math score 0.004** 0.003*** 0.006*** 0.002 0.004* 0.001 0.004* 0.002 (0.002) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) Apprentice (currently) –0.540*** –0.844*** –0.675*** –0.759*** (0.132) (0.108) (0.094) (0.252) Apprenticeship completed –0.044 0.025 0.054 –0.075 0.072 –0.078 0.134 0.025 (0.072) (0.065) (0.072) (0.064) (0.128) (0.121) (0.109) (0.101) Ln (hours) 0.209*** 0.170** 0.147* 0.249*** 0.137 –0.030 0.318** –0.008 (0.081) (0.083) (0.076) (0.080) (0.116) (0.117) (0.161) (0.091) Tenure 0.015*** 0.010** 0.024*** 0.012*** 0.009* 0.003 –0.008 0.005 (0.004) (0.004) (0.008) (0.003) (0.005) (0.006) (0.013) (0.004) Ln (employees) 0.301*** 0.218*** 0.187** 0.313*** 0.435*** 0.490*** 0.001 0.579*** (0.089) (0.076) (0.093) (0.073) (0.130) (0.095) (0.166) (0.084) Ln (firm size) 0.188*** 0.163*** 0.211*** 0.123*** 0.223*** 0.059* 0.119*** 0.136*** (0.032) (0.024) (0.028) (0.023) (0.030) (0.032) (0.034) (0.027) Private wage –0.267*** –0.319*** –0.372*** –0.119 –0.248** 0.140 –0.179 –0.010 (0.094) (0.089) (0.091) (0.089) (0.108) (0.126) (0.132) (0.104) Civil servant 0.534*** 0.473*** 0.419*** 0.568*** 0.377*** 0.503*** 0.109 0.532*** (0.136) (0.113) (0.138) (0.104) (0.116) (0.128) (0.144) (0.094) Public enterprise –0.266 –0.219 –0.409** –0.029 –0.001 0.446** –0.229 0.329** (0.212) (0.153) (0.182) (0.158) (0.162) (0.191) (0.198) (0.149) City and year dummies Yes Yes Yes Yes Yes Yes Yes Yes Constant –0.780 –0.888 –0.992 0.579 9.795*** 7.842*** 7.202*** 9.846*** (0.894) (0.783) (0.855) (0.855) (0.992) (0.963) (1.358) (0.742) R2 0.256 0.340 0.361 0.236 0.316 0.245 0.252 0.320 Adjusted R2 0.244 0.328 0.348 0.225 0.295 0.221 0.204 0.306 Number of observations 1,428 1,299 1,073 1,654 716 678 347 1,047 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Standard errors are in parenthesis. ***p<0.01    **p<0.05    *p<0.1 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Levels 29 • Women’s earnings also grow slower with age, but peak later in life than those of men. This finding might reflect lower participation rates among women at a young age and delays in the accumulation of labor market experience. • But women experience higher returns to education than men. Women experience positive returns to having finished primary school, while men only benefit from having finished secondary school and beyond. The result is less pro- nounced in Tanzania, but is important, because it shows that female schooling can yield higher returns at a young age. • Young workers earn significantly less than older workers in private employment, perhaps because they lack experience and financial capital. • Young workers also face the highest unemployment rates. • But returns to education are higher and more highly convex among the young, suggesting that schooling is more important at the early stages of one’s career than later on, when capital accumulation (for example, savings) may help compensate for lack of skills. Skills erosion is another potential explanation for this result. Differences across Sectors The specifications presented so far assume that the determinants of earnings are the same across sectors. To investigate whether skills are rewarded differently in different sectors, and to examine how skills vary across sectors, separate earnings functions are presented in table 4.4 for the self-employed, wage employed, and those in public sector employment. Table 4.4  Earnings Functions by Occupation Ghana Tanzania Self Private wage Public wage Self Private wage Public wage Male 0.317*** 0.167*** 0.051 0.352*** 0.258** 0.165 (0.079) (0.062) (0.129) (0.071) (0.102) (0.107) Age 0.090*** 0.085*** 0.019 0.048** 0.005 0.075** (0.021) (0.020) (0.028) (0.020) (0.032) (0.035) (age^2)/100 –0.106*** –0.103*** –0.010 –0.053** 0.008 –0.067* (0.027) (0.029) (0.036) (0.025) (0.042) (0.040) Height (cm) 0.005 0.005 –0.008 0.003 0.002 0.004 (0.005) (0.005) (0.008) (0.004) (0.007) (0.007) Years in formal education 0.012 –0.085*** 0.000 –0.040 –0.079** –0.042 (0.025) (0.029) (0.065) (0.025) (0.037) (0.039) (educ^2)/100 –0.001 0.007*** 0.002 0.005** 0.009*** 0.006** (0.002) (0.002) (0.003) (0.002) (0.003) (0.003) Apprenticeship completed 0.036 –0.057 –0.173 –0.076 0.246 –0.050 (0.067) (0.061) (0.130) (0.123) (0.171) (0.191) table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 30 The Determinants of Earnings Levels Table 4.4  Earnings Functions by Occupation (continued) Ghana Tanzania Self Private wage Public wage Self Private wage Public wage Apprentice (currently) –0.415** –0.802*** –0.313 (0.185) (0.093) (0.332) Math score 0.002 0.004*** 0.005** 0.003 –0.000 0.001 (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) Ln (hours) 0.255*** 0.054 0.501** 0.198** –0.412** 0.171 (0.074) (0.092) (0.212) (0.098) (0.164) (0.248) Ln (firm size) 0.143*** 0.271*** 0.154*** 0.083 (0.020) (0.102) (0.027) (0.065) Tenure 0.012*** 0.012** 0.015* 0.004 0.007 –0.003 (0.004) (0.005) (0.009) (0.006) (0.006) (0.006) Ln (employees) 0.279*** 0.465*** (0.061) (0.077) Self-employed and not a trader –0.071 –0.174*** (0.066) (0.067) Wage em- ployee in manufactur- ing 0.014 –0.308*** (0.055) (0.118) Public enter- prise –1.263*** –0.109 (0.475) (0.176) City and year dummies Yes Yes Yes Yes Yes Yes Constant –1.672** –0.775 1.392 8.140*** 11.457*** 7.891*** (0.832) (0.871) (1.610) (0.826) (1.455) (1.699) Number of observations 1,571 904 225 879 332 175 R2 0.197 0.534 0.407 0.206 0.374 0.323 Adjusted R2 0.187 0.523 0.349 0.188 0.335 0.239 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Standard errors are in parenthesis. ***p<0.01     **p<0.05     *p<0.1 The findings show that earnings regimes vary across sectors in a number of ways. For example, the gender gap is highest for the self-employed and statisti- cally negligible for public sector employees. The returns to education also rise more rapidly in wage employment than in self-employment, suggesting higher returns to secondary education for wage workers. However, some interesting commonalities also emerge. For example, the age-earnings profiles of the wage earners and self-employed in Ghana are very similar, unlike those in Tanzania. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Levels 31 Notes 1. As explained in chapter 3, the Raven’s test is a cleaner proxy for innate ability than other scores because it measures noncognitive skills, which are arguably less linked to schooling than those measured by the other tests. 2. A young worker is defined as someone younger than the median sample age of 30. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 5 The Determinants of Earnings Growth In this chapter we explore the determinant of earnings growth in the Tanzanian and Ghanaian Urban Panel Surveys. We use the panel nature of the data to understand how earnings change over time and how these changes vary across sectors and differ between workers who change sector and those who don’t. Box 5.1 sets out our key hypotheses and main findings. Box 5.1 Key Hypotheses and Main Findings Hypothesis 1: The variables that determine earnings levels are also strong predictors of changes in earnings. Main finding: It is difficult to identify predictors of earnings growth. • The models work poorly in Tanzania and for the Ghanaian self-employed. • The models work better for wage workers in Ghana. Hypothesis 2: The determinants of earnings growth vary by sector. Main finding: The factors governing earnings growth vary substantially across sectors. In Ghana, education is positively associated with earnings growth for wage workers, but •  not for the self-employed. Earnings growth is much faster for apprentices than individuals with comparable charac- •  teristics. Hypothesis 3: Switching sectors can lead to substantial changes in earnings. Main finding: Switching jobs is the main source of earnings growth. • Moving from small to large firms leads to higher earnings. • Switching sectors also leads to substantial changes in earnings. • Moving to the public sector results in gains. • Moving to self-employment also tends to increase earnings. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   33 34 The Determinants of Earnings Growth Framework Earnings growth is modeled as a function of changes in the time-variant explana- tory variables of earnings levels—that is, firm size, tenure, hours worked, and occupation. Time-invariant characteristics—such as education, gender, and height1—are only included as explanatory variables if they are expected to have an additional impact on the rate of earnings growth over and above their impact on levels. In addition, lagged levels of some time-variant variables—for example, firm size—are included to test whether individuals in larger firms/enterprises experience more rapid earnings growth even if their firm does not grow in size. This may be due to large firms raising productivity faster or age-tenure profiles being steeper, as larger firms benefit more from firm-specific skills.2 The preferred specification, presented in table 5.1, includes controls for gen- der, age, height, educational attainment, math test score, and the lag of firm size; a set of variables reflecting changes in hours worked, tenure, firm size; and a set of dummies capturing sector transitions (or the absence thereof), year and city dummies, and dummies indicating whether an individual is entering, graduating, or currently doing an apprenticeship. Table 5.1 Determinants of One-Year Growth in Log Earnings Dependent variable: change log earnings Ghana Tanzania Male –0.005 –0.035 (0.046) (0.069) L.age 0.001 –0.003 (0.002) (0.004) Height (cm) –0.003 0.001 (0.003) (0.005) Years in formal education –0.014 –0.001 (0.017) (0.024) (educ^2)/100 0.001 –0.001 (0.001) (0.002) Math score –0.001 –0.003 (0.001) (0.002) Became apprentice –0.186 (0.260) Apprentice, graduated 0.338 (0.273) L. apprenticeship completed 0.057 –0.002 (0.043) (0.099) L. apprentice (currently) 0.391*** (0.127) ΔLn (hours) 0.029 0.134 (0.065) (0.102) Δtenure 0.002 0.018* (0.005) (0.010) table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Growth 35 Table 5.1 Determinants of One-Year Growth in Log Earnings (continued) Dependent variable: change log earnings Ghana Tanzania L. tenure 0.002 0.003 (0.003) (0.005) ΔLn (employees) 0.095 0.289* (0.085) (0.174) L.Ln (employees) –0.003 0.061 (0.065) (0.149) ΔLn (firm size) 0.068*** 0.042 (0.024) (0.033) L.Ln (firm size) 0.017 0.015 (0.016) (0.029) Self -> priv wage 0.036 0.091 (0.154) (0.228) Self -> pub wage 0.915 0.649** (0.719) (0.253) L.priv wage –0.115* –0.161 (0.061) (0.108) Priv wage -> self 0.447** –0.049 (0.185) (0.254) Priv wage -> public –0.032 –0.043 (0.100) (0.149) L.public –0.073 0.015 (0.074) (0.144) Public -> self –0.450 0.028 (0.327) (0.149) Public -> priv wage –0.224** –0.220 (0.111) (0.304) Constant 0.693 0.220 (0.426) (0.752) City and year dummies Yes Yes Number of observations 1,461 602 R2 0.070 0.107 Adjusted R2 0.050 0.063 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Standard errors are in parenthesis. L. indicates a lagged variable whilst Δ indicates a change in a variable between current and previous time period. ***p<0.01     **p<0.05    *p<0.1 A Bird’s Eye View of Earnings Growth in Ghana and Tanzania Arguably, the most important finding of this analysis is that in both countries it is difficult to predict earnings growth. This is evidenced by the low R2s of the regressions and the lack of significance of most explanatory variables. These poor results may be due to the very short time horizon over which earnings growth was studied. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 36 The Determinants of Earnings Growth Though the R2s in Tanzania are higher than those in Ghana, the lagged sector of employment and the city dummies (not reported to conserve space) are the only statistically significantly time-invariant variables. Sector changes are also the only significant regressors among those intended to capture the impact of changes in observable characteristics. Although the number of individuals who either became paid apprentices or graduated from a paid apprenticeship was relatively small, some interesting results emerge. Apprentices experience accelerated earnings growth, and graduating from an apprenticeship leads to strongly significant increases in earnings. However, this impact seems to be a one-off event; earnings do not continue to grow faster for workers with apprenticeships, and having finished an apprentice does not lead to higher earnings. Rather, graduation enables the low-paid apprentice to catch up with individuals with comparable observable characteristics. Switching jobs provides the best opportunity for fast earnings growth. Changes in firm size and sector are also very strongly correlated with increases in earnings, and average earnings growth rates vary across sectors, though the firm size and sectoral effects need to be assessed jointly. In Ghana, losing a public sec- tor job leads to earnings losses, but moving from a private sector wage job to self-employment, on average, increases earnings. In both countries, a move from self-employment to a wage job in a small firm has little impact on earnings, but wages increase when moving to a large firm. However, it is important to recognize that the ability to find significant predic- tors of earnings growth may be undermined by the fact that the data are likely to be measured with a substantial degree of noise. Moreover, the fact that earn- ings change the most as a result of job changes is perhaps not surprising in view of this short time horizon. To examine how sensitive these results are to the time horizon under study, appendix C contains the examination of the determinants of earnings changes over a two-year period. In addition, the analysis attempts to control for possible individual-specific patterns of earnings growth by controlling for fixed effects. Overall, the results do not change very much, though the implied sectoral premia are sensitive to the time horizon used. Differences across Sectors Thus far, this analysis has imposed a common earnings growth regime across sectors despite the findings of chapter 4, which show that the determinants of earnings levels vary across sectors. To examine whether the determinants of earnings growth also vary across sectors, we estimate the preferred specification separately for those who started as self-employed, those who started off as wage workers, and those who started in the public sector. Except for wage workers, indicators capturing the impact of becoming an apprentice and graduating from an apprenticeship were dropped since we do not have a suf- ficient number of observations to identify the effects. Similarly, the sample sizes for public sector employees are small in Ghana and Tanzania, and should thus be interpreted with caution—for completeness they are reported in appendix C, but not discussed in detail. The results of separate earnings regres- sions for different sectors are presented in table 5.2. Perhaps the most Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Growth 37 Table 5.2  Determinants of One-Year Growth in Log Earnings by Occupation Ghana Tanzania Self Private wage Public wage Self Private wage Public wage Male –0.024 –0.033 0.065 –0.026 0.033 –0.102 (0.068) (0.071) (0.092) (0.095) (0.118) (0.310) L.age 0.000 0.003 –0.002 –0.000 –0.005 –0.008 (0.004) (0.003) (0.005) (0.006) (0.006) (0.010) Height (cm) –0.003 –0.001 –0.007 –0.002 –0.003 0.035* (0.003) (0.004) (0.007) (0.007) (0.006) (0.020) Years in formal education –0.004 –0.056** –0.100 –0.010 0.023 –0.044 (0.024) (0.025) (0.196) (0.038) (0.046) (0.097) (educ^2)/100 0.001 0.003** 0.004 0.001 –0.002 –0.001 (0.002) (0.001) (0.008) (0.004) (0.004) (0.007) Math score –0.002 –0.002 0.007*** –0.001 –0.005** 0.008 (0.002) (0.001) (0.002) (0.003) (0.003) (0.007) L.Apprenticeship completed 0.044 0.053 0.029 0.032 –0.234 0.001 (0.062) (0.060) (0.152) (0.113) (0.168) (0.602) ΔLn (hours) –0.016 0.274*** 0.206 0.112 0.339 0.631 (0.074) (0.080) (0.217) (0.115) (0.286) (0.643) Δtenure –0.002 0.022*** –0.008 0.027* 0.020** 0.014 (0.007) (0.008) (0.010) (0.015) (0.009) (0.024) L.tenure 0.002 0.002 0.009 0.005 0.002 0.007 (0.004) (0.005) (0.006) (0.007) (0.006) (0.016) ΔLn (employees) 0.059 0.479** 0.281 0.082 (0.090) (0.196) (0.178) (0.855) L.Ln (employees) –0.006 0.029 (0.070) (0.150) ΔLn (firm size) 0.157 0.099*** –0.001 0.080 0.065 –0.247 (0.099) (0.030) (0.027) (0.087) (0.042) (0.180) L.Ln (firm size) 0.007 0.019 0.031 –0.042 (0.020) (0.023) (0.030) (0.159) Self -> priv wage –0.181 0.015 (0.250) (0.296) Self -> public 0.748 0.627** (0.694) (0.308) becameapp 0.028 (0.270) Appgrad 0.572** (0.269) L.Apprentice (currently) 0.357*** (0.125) Priv wage -> self 0.340* –0.148 (0.195) (0.285) Priv wage -> public 0.064 0.029 table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 38 The Determinants of Earnings Growth Table 5.2  Determinants of One-Year Growth in Log Earnings by Occupation (continued) Ghana Tanzania Self Private wage Public wage Self Private wage Public wage (0.119) (0.169) Public -> self –0.342 0.706 (0.233) (0.592) Public -> priv wage –0.126 0.272 (0.112) (0.499) City and year dummies Yes Yes Yes Yes Yes Yes Constant 0.715 0.494 1.182 0.454 0.974 –5.634* (0.583) (0.709) (1.610) (1.050) (0.957) (3.371) Number of observations 897 442 122 404 140 58 R2 0.057 0.230 0.243 0.134 0.181 0.423 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Standard errors are in parenthesis. L. indicates a lagged variable whilst Δ indicates a change in a variable between current and previous time period. ***p<0.01    **p<0.05    *p<0.1 interesting finding is that these models are much better predictors of earnings growth for wage workers than for the self-employed, at least in Ghana. This is shown by the higher R2s on the regressions for employees and the fact that more variables are individually significant in Ghana (though not in Tanzania). There are virtually no significant predictors for earnings growth for the self- employed, save for sectoral switches, because it appears that the few individu- als who moved from self-employment to a public sector job saw starkly increased earnings. These poor results might be due to higher measurement error in the earnings of the self-employed and the fact that they may include returns to both capital and labor. The findings for wage workers are more encouraging, at least in Ghana, where education and its square are jointly strongly correlated with educational attain- ment, suggesting a convex relationship between earnings growth and education. This means that at low levels of education, an additional year of education decel- erates earnings growth, but beyond secondary school, earnings grow progressively more with each additional year of education. On the other hand, earnings growth is negatively correlated with cognitive skills, a finding which is difficult to explain. Doing an apprenticeship is again associated with much higher earnings growth during the apprenticeship, and a significant spike in earnings after gradu- ation, but having completed an apprenticeship does not lead to sustained higher earnings growth. Increasing hours worked and moving to a larger firm are also strongly correlated with earnings growth, and individuals who move from wage employment to self-employment tend to experience an increase in earnings that exceeds the benefits of moving to the public sector. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 The Determinants of Earnings Growth 39 Notes 1. The measure of education and height used in this exercise was constructed to be time invariant (see chapter 3). 2. Note that changes in earnings associated with firm expansion, contraction, or moving between firms of a different size are already controlled for by including changes in firm size as an explanatory variable. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 6 Low-Pay/High-Pay Transitions In this chapter we focus on low-paid work1 using the Ghanaian and Tanzanian Urban Panel Surveys. We explore the extent of low-paid work in the survey data and identify whether it is possible for those in low-paid employment to move towards higher paid work. We also analyze the factors associated with moving into and out of low-paying work and whether doing low-paid work has a scarring effect on future earnings. See box 6.1 for the key hypotheses we test in this chapter and our main findings. Box 6.1  Key Hypotheses and Main Findings Hypothesis 1: Workers risk being trapped in low-paying occupations, and some groups of workers are particularly at risk of falling into low-pay traps. Main finding: Low pay is a persistent condition. • Women and youth are more likely to fall into and remain trapped in low-paying jobs. • Young women are doubly disadvantaged. Hypothesis 2: Being in a low-paid job has a scarring effect on prospects of future earnings. Main finding: The experience of being in a low-paying job is scarring. Falling into low pay undermines individuals’ prospects of obtaining high-paying jobs in •  the future. Hypothesis 3: The process and factors that push individuals out of a low-paying job are not the same as those that pull them out. Main finding: There is path dependence in pay trajectories. • The probability of entering/exiting a job category depends on the initial status. The impact of different factors on the transition probability also depends on both initial and final status. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   41 42 Low-Pay/High-Pay Transitions Descriptive Statistics Tables 6.1a and 6.1b show that in both Ghana and Tanzania, individuals who were low paid at t-1 are four times more likely to be low paid at time t than other workers. In addition, the average probability of being low paid in two consecutive periods is higher than 50 percent in both countries. The persistence rate is sub- stantially higher than that reported in developed countries and is particularly high for women and for youth.2 Finally, is it worth noting that a number of t-1 earners have dropped out of the earnings distribution at time t due to either hav- ing exited the labor force or simply because of not having reported earnings. Because these individuals are roughly as likely to be low paid as those who stay in the sample, their exit does not bias the results.3 Table 6.1a  Raw Persistence in Ghana Low pay at t + 1 No Yes Total No 87.14 12.86 100 All Low pay at t Yes 44.78 55.22 100 Total 72.61 27.39 100 No 82.13 17.87 100 Women Low pay at t Yes 41.79 58.21 100 Total 65.04 34.96 100 No 83.43 16.57 100 Youth Low pay at t Yes 38.97 61.03 100 Total 62.9 37.1 100 Table 6.1b  Raw Persistence in Tanzania Low pay at t + 1 No Yes Total All Low pay at t No 84.78 15.22 100 Yes 45.71 54.29 100 Total 70.26 29.74 100 Women Low pay at t No 76.36 23.64 100 Yes 35.22 64.78 100 Total 56.17 43.83 100 Youth Low pay at t No 83.58 16.42 100 Yes 41.54 58.46 100 Total 62.88 37.12 100 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Low-Pay/High-Pay Transitions 43 Econometric Framework The descriptive statistics presented above do not allow quantification of the extent to which low-pay persistence is due to the workers who are low paid at time t-1 having systematically lower endowments than the rest—that is, individ- ual heterogeneity—as opposed to the experience of low pay itself increasing the probability of being low paid in future—genuine state dependence or scarring. The potential scarring effect of being low paid may arise because being in a low- paying job induces human capital depreciation, but also because, in a market with imperfect information, employers might use previous pay as a signal of abil- ity (Cappellari and Jenkins 2004). Being in a low-paying job may also undermine an individual’s aspirations and productivity or reduce reservation wages.4 However, the sample of low earners at time t-1 is not a random sample of the population, but instead is likely to contain a larger proportion of individuals with a high propensity to be low paid in any period. From the econometric prospec- tive this gives rise to the initial conditions problem (Heckman 1981b). An addi- tional econometric challenge is that there may be nonrandom selection into the subsample of individuals for whom two consecutive earnings levels are observed. This is possible because workers with particular characteristics may be relatively more likely to systematically drop out of the sample after the first period. Modeling Strategy To address these econometric problems, low-pay/high-pay transitions are mod- eled using a first-order Markov model that accounts for the initial conditions problem and nonrandom retention. The modeling framework is an adaptation of the framework used by Cappellari and Jenkins (2004), who treat selection into base-year earnings category and sample retention as issues of multiple endoge- nous selection. An appealing feature of this approach is that it is a switching model that allows the determinants of low pay to have a differential impact, depending on whether the individual in question was in a low-paying or high- paying job in the previous period. In other words, the model helps examine whether the effects of individual characteristics—such as age, gender, and educa- tion—on the probability of moving out of a low-paying job differ from their impacts on the probability of falling into low-paying employment. However, a very severe limitation of this modeling strategy is that it imposes very strong structural assumptions. More specifically, the approach used multi- variate probit models to model low-earnings transitions between two consecutive years, pooling observations across observed transitions. The most general model comprises three equations: (i) an initial conditions equation that models base- year low pay, (ii) a retention equation that determines whether wages are observed both in period t-1 and in period t, and (iii) a transition equation that models low-pay status in period t. These equations are assumed to have a com- mon error process. The correlations between the error terms in the different equations are used to account for unobserved heterogeneity—as explained in more detail in appendix D. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 44 Low-Pay/High-Pay Transitions Model Specification and Performance The bivariate specification of earnings transitions uses personal characteristics (education, gender, age, height, and whether the person has ever been an appren- tice) and job characteristics (sector dummies, tenure, firm size, or number of employees) to control for the initial conditions. Year dummies are used to control for transitions due to systematic increases in earnings associated with time trends, such as technical progress. Dummies for current city are also included and paren- tal education is used as an instrument for the initial conditions equation, although it may itself be correlated with the probability of transiting. Tests of alternative specifications are presented in appendix D. The estimated correlation between unobservables in the two equations is negative in both countries, though insignificant in Tanzania, probably because of the relatively small number of transitions. This suggests that there is “regression to the mean” in the sense that individuals who were low paid at t-1 are more likely to be high paid this year, potentially reflecting error in the measure of earnings. Results These models help examine how personal and job characteristics impinge on the probability of moving from low-paying to high-paying jobs and vice versa. The endogenous switching model implies that regressors have different effects, depending on whether the analysis is conditioning on high or low pay. Since we are most interested in low pay, the analysis focused on the low-income persis- tence rate sit (the probability of being low paid at t, conditional on being low paid at t-1) and the low-income entry rate eit (the probability of being low paid at t, conditional on being high paid at t-1). Appendix D explains how to compute these rates. The analysis also documented expected durations of high-paid and low-paid employment spells under the assumption that the economic environ- ment is stationary, that is, it does not change over time. In addition, the models measured state dependence. The Determinants of Transition Probabilities The analysis explored the impact on the predicted probabilities of low-pay per- sistence and entry (sit and eit), allowing personal and job characteristics to vary. The results for Ghana are shown in table 6.3 and those for Tanzania in table 6.4. Two broad messages emerge from the findings. The first important message is the existence of significant evidence of path dependence in low-pay/high-pay transitions in both countries. This is shown by the fact that the impact of the explanatory variables on low-pay persistence probabilities is not statistically equal to that on the propensity to remain in (relatively) high-pay employment. The second important findings is that there is also cross-country and within-country heterogeneity in persistence and entry rates of low pay and in their determinants. In Ghana, if high-paid at t–1, the reference individual has a predicted proba- bility of 13 percent of entering low-paid employment in the next period (first Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Low-Pay/High-Pay Transitions 45 Table 6.2a  Predicted Entry and Persistence Rates, and Mean and Median Predicted Time in High and Low Pay for Ghana, Bivariate Low entry Mean Median Low Mean Median (%) high high persist (%) low low Reference 13 7.6 4.9 50 2.0 1.0 Educ = 12 10 9.9 6.5 51 2.0 1.0 Female 23 4.3 2.6 59 2.4 1.3 Small firm 12 8.6 5.6 47 1.9 0.9 Public employee 7 14.9 10.0 44 1.8 0.8 Age = 40 11 9.3 6.1 43 1.8 0.8 Apprentice 12 8.2 5.3 50 2.0 1.0 Tenure 14 7.3 4.7 46 1.8 0.9 1 employee 9 11.0 7.3 49 2.0 1.0 Age = 20 21 4.8 3.0 63 2.7 1.5 Table 6.2b  Predicted Entry and Persistence Rates, and Mean and Median Predicted Time in High and Low Pay for Tanzania, Bivariate Low entry Mean Median Low Mean Median (%) high high persist (%) low low Reference individual 10 10.2 6.7 24 1.3 0.5 Educ = 12 1 78.4 54.0 2 1.0 0.2 Female 22 4.6 2.8 61 2.5 1.4 Small firm 7 14.3 9.6 46 1.8 0.9 Public employee 3 38.8 26.5 6 1.1 0.3 Age = 40 11 9.3 6.1 16 1.2 0.4 Apprentice 10 10.2 6.7 15 1.2 0.4 Tenure 8 11.8 7.8 26 1.4 0.5 1 employee 6 17.4 11.7 24 1.3 0.5 Age = 20 10 10.0 6.6 40 1.7 0.8 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: The reference person is a 30-year-old, self-employed male with no employees, who has seven years of educa- tion, five years of tenure in his current job, and never been an apprentice. Low entry refers to the probability of being low paid in the next period conditional on being high paid currently. Low persist refers to the probability of continuing to be low paid in the next period. column of table 6.2a). By contrast, the fourth column shows that, if low paid at time t–1, the reference individual has a 50 percent probability of remaining in low-paid employment. As shown in Table 6.2b, the gap is somewhat smaller in Tanzania, but still significant (10 and 24 percent, respectively). Tables 6.2a and 6.2b also show how the probability varies with personal and job characteristics. The low-pay entry rate is higher in Ghana than in Tanzania, but the difference is particularly striking for the persistence rate, which is more than double. Ghana also shows higher within-group heterogeneity in the low- income entry rates, whereas in Tanzania, heterogeneity in persistence rates is higher. In so far as policies concentrate on the low earners, these might have Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 46 Low-Pay/High-Pay Transitions higher payoffs in Tanzania. For Ghana, these results suggest substantial potential payoffs to targeting the earners most at risk of transitioning into the low-paid employment categories, particularly youth and women. This conclusion is strengthened by the work below on measuring state dependence. In line with the findings in the previous sections of this study regarding large returns to education, in Tanzania the largest estimated source of heterogeneity is education. Having 12 years of schooling lowers the probability of entering low pay and of low-pay persistence by more than 90 percent, and this dramatically increases the mean predicted years out of low-pay. By contrast, the strongest effects in Ghana are observed for youth and women, who are nearly twice as likely to fall into low-pay employment and, respectively, 20 and 26 percent more likely to get trapped in it. Young women are doubly disadvantaged. Women face heavy disadvantage in Tanzania too. They are more than twice as likely as men to fall into low pay and 40 percent more likely to remain trapped. Interestingly, the large disadvantage seen for youth in Tanzania is only evident for those already in low pay (60 percent above that of the reference category), while the probability of entry into it (for those who are better paid) is not significantly different. Results from both countries also confirm the advantage of being employed in the public sector: the probability of a public sector employee entering low pay is half that of the reference individual, and the expected persistence in this status is also much lower. The probabilities of entry into low pay are also substantially lower for the self-employed with employees. Testing for Evidence of Labor Market Scarring The descriptive statistics presented above also give us a very crude measure of state dependence that does not control for differences between individuals, either observed or unobserved. To control for these factors, we compute two alternative measures of state dependence: aggregate state dependence (ASD) and genuine state dependence (GSD).5 Using the bivariate specification, the ASD is estimated to be 0.42 in Ghana and 0.45 in Tanzania, while GSD values for both countries are 0.35 and 0.27, respectively. When using the trivariate model to control for retention, both ASD and GSD values are significantly reduced in Ghana, to 0.27 and 0.23, respec- tively. Overall, these results suggest that there are substantial scarring effects of low pay in both Tanzania and Ghana. The values for GSD are similar to those reported for developed countries, but the ASD figures are generally lower. For example, Cappellari (2002), using panel data on Italian workers, finds a GSD value between 0.20 and 0.36, but ASD values between 0.49 and 0.61, depending on specification. The lower ASD values suggest higher mobility across the low- income line in the African economies than in developed countries, in accordance with the existence of a larger group of individuals vulnerable to low earnings.6 However, measurement error may also be a more severe problem in developing country data sets than in developed country data. The fact that ASD is closer to GSD in Ghana and Tanzania than in other countries suggests that scarring plays a larger role than observable characteristics. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Low-Pay/High-Pay Transitions 47 Notes 1. Low-paid jobs are those that pay less than US$1.25 a day. For more details, see chap- ter 3. The alternative definition of 60 percent of median earnings was used to assess the robustness of the results, but since the conclusions were not qualitatively different, these results are not reported to conserve space. 2. The higher persistence rate compared to developed countries may partly reflect the different low-pay line used (Stewart and Swaffield 1999). 3. Results are available from the authors. 4. The reservation wage is the lowest wage rate at which a worker would be willing to accept a particular type of job. A job offer involving the same type of work and the same working conditions, but at a lower wage rate, would be rejected by the worker. 5. Following Cappellari and Jenkins (2004), ASD is defined as the difference between the probability of being low paid at t, for those low paid at t–1, and the probability of being low paid at t, for those not low paid at t–1. GSD is defined as the average dif- ference between the predicted probabilities of being low paid conditional on being low paid and high paid, respectively. Measures of ASD are arguably less prone to measurement error than measures of raw state dependence. However, contrary to the GSD, they do not control for individual heterogeneity. If differences in observed indi- vidual characteristics are the main drivers of initial pay states and transitions between states, the GSD is expected to be lower than ASD, and, in the presence of labor mar- ket scarring, the GSD should be greater than zero. 6. Though it should be remembered very different low-income lines are used in the literature on developed countries. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 CHAPTER 7 Main Findings and Key Policy Implications This study contributes to understanding earnings dynamics in low-income coun- tries (LICs) by examining the determinants of earnings, earnings growth, and low-pay/high-pay transitions in Ghana and Tanzania. It highlights the impor- tance of personal and job characteristics in determining earnings and earnings growth, but also finds large variations across sectors and firm size in the returns to individual characteristics and the prospects for earnings growth. These find- ings point toward strong path dependence in pay trajectories and the existence of a scarring effect. Thus, falling into low-pay employment reduces considerably one’s future labor market prospects. These results go beyond the time period covered by the data (2004–08) and also have a number of important implica- tions for the potential effectiveness of alternative interventions in enhancing the quality of employment opportunities, not only in Ghana and Tanzania, but also in LICs more broadly. Message 1: Job Characteristics Are an Important Determinant of Both Earnings Levels and Earnings Growth In the short run, the most effective way of increasing earnings is to change the type of job, because earnings and potential for growth vary considerably across different types of employment. Returns to education, in terms of both earnings levels and growth, also seem driven to a large extent by differences in the type of jobs available to workers of different ages and education. The scope for wage increases within a given job is much more limited. In addition, the persistence in earnings is high and the initial employment type is an important determinant of where one ends up. Being in a low-paying job has powerful scarring effects that make it difficult to move to better-paid job. Thus, individual earnings trajectories are to a large extent determined by one’s labor allocation across sectors over the life cycle. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   49 50 Main Findings and Key Policy Implications Policy implication: Introducing measures that act as safety ropes and prevent vulnerable individuals from falling into low-pay might yield high returns, since it is more difficult to get people out of low-paying jobs once they have fallen into them. Message 2: Women and Youth Face Special Challenges Women and young workers earn much less than men and older workers, even after controlling for educational disadvantages. They are also much less likely to escape low-paying jobs and have higher risk of falling into low pay. Policy implication: Gender-sensitive and youth-focused social protection poli- cies are a first step to bridging the gender and youth gaps in earnings. Message 3: Skills Acquisition Is a Stepping Stone Toward Better Paying Jobs, at Least in Wage Employment, Especially for Women As in previous studies, this analysis finds a strong correlation between skills acquisition - proxied by education, cognitive skills, and having completed an apprenticeship - and initial earnings. However, the relationship between skills and earnings growth is much weaker, although it varies across employment type. In Ghana, education is significantly positively correlated with earnings growth for the wage employed, but not for the self-employed. In both countries, the returns to education are higher for women. Policy implication: Existing efforts to raise educational attainment may yield high returns and promoting female school enrollment may be especially benefi- cial. In addition, the finding that returns to education are increasing and robust, even when controlling for ability bias, suggests that investing in tertiary education might be an excellent means of capitalizing on the gains of enhanced primary and secondary school completion rates.1 Message 4: Self-Employment Can Be Desirable Although different sectors offer different returns, being self-employed is not always synonymous with having low earnings. Indeed, on average, moves into wage employment are just as likely to lead to better pay as moves to self-employ- ment. This supports the view that self-employment is not always employment of the last resort. Yet, entrepreneurs are on average older than wage workers, hinting at the possibility of capital constraints. Policy implication: Facilitating better access to credit might be beneficial. Message 5: The Public Sector Wage Premium Is a Potential Barrier to the Efficient Working of the Labor Market Civil servants earn more than private sector employees with comparable char- acteristics, and, unsurprisingly, moves into the public sector are associated with increased earnings, while moves out of it tend to be associated with a wage Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Main Findings and Key Policy Implications 51 drop. This suggests that public sector jobs may be rationed. This hypothesis is consistent with the finding that a large percentage of those not working at the time of the surveys are highly educated individuals likely to be queuing for public sector jobs. Policy implication: Reducing the public sector pay premium by stalling earn- ings increases can contribute to a less distorted labor market. Note 1. Of course, these arguments implicitly assume that the returns to education remain constant-which may not be the case if tertiary school enrollment rates increased rap- idly or if girls’ educational attainment caught up with that of boys. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 APPENDIX A Summary Statistics and Variable Definitions Summary Statistics Table A.1  Summary Statistics (Ghana) Variable Mean Observations Male 0.43 6,796 Age 31.32 6,455 Education 8.75 6,795 Height 164.00 6,422 Math score 45.88 5,482 Raven’s score 32.96 4,218 English score 60.28 5,511 Reading score 55.88 4,605 Apprenticeship completed 0.28 6,813 Apprentice (currently) 0.05 6,509 Employees 1.25 3,864 Firm size 13.13 3,819 Hours 46.74 3,994 Tenure 8.43 3,884 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Table A.2  Summary Statistics (Tanzania) Variable Mean Observations Male 0.46 2,412 Age 35.81 2,413 Education 5.05 2,413 Height 163.97 2,197 Math score 61.11 2,233 Raven’s score 19.09 1,848 English score 64.79 1,751 Reading score 59.99 2,163 table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   53 54 Summary Statistics and Variable Definitions Table A.2  Summary Statistics (Tanzania) (continued) Variable Mean Observations Apprenticeship completed 0.07 2,413 Employees 0.14 1,090 Firm size 20.20 1,702 Hours 52.47 1,711 Tenure 9.80 1,899 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Table A.3  Mean (1-Year) Changes in Log Earnings by Transition Type (Ghana) Transition Mean change Number of trans SE w/o -> SE w/o 0.194 750 SE w/o -> SE w 0.486 98 SE w/o -> W <=10 0.387 62 SE w/o -> W >10 0.791 16 SE w/o -> Civil 0.263 2 SE w/o -> Pub Ent 2.661 1 SE w -> SE w/o 0.244 59 SE w -> SE w –0.010 143 SE w -> W <=10 –0.703 4 SE w -> W >10 0.328 7 SE w -> Civil 2.235 1 W <=10 -> SE w/o 0.617 34 W <=10 -> SE w 1.027 13 W <=10 -> W <=10 0.149 162 W <=10 -> W >10 0.263 52 W <=10 -> Civil –0.022 3 W <=10 -> Pub Ent –0.631 1 W >10 -> SE w/o –0.104 11 W >10 -> SE w 1.845 3 W >10 -> W <=10 0.082 21 W >10 -> W >10 0.083 215 W >10 -> Civil 0.035 7 W >10 -> Pub Ent 0.307 6 Civil -> SE w/o –0.502 3 Civil -> W <=10 0.384 1 Civil -> W >10 0.105 7 Civil -> Civil 0.030 66 Civil -> Pub Ent 0.025 4 Pub Ent -> SE w/o –0.234 2 Pub Ent -> SE w 0.345 1 Pub Ent -> W <=10 –0.117 3 Pub Ent -> W >10 0.000 12 table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Summary Statistics and Variable Definitions 55 Table A.3  Mean (1-Year) Changes in Log Earnings by Transition Type (Ghana) (continued) Transition Mean change Number of trans Pub Ent -> Civil 0.171 14 Pub Ent -> Pub Ent 0.136 26 Total 0.195 1,810 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: SE w/o = self-employed without employees; SE w = self-employed with employees; W <=10 = wage employee working in a firm with 10 employees or less; W>10 = wage employee working in a firm with more than 10 employees; Civil = Civil servant; Pub Ent = public enterprise. Table A.4  Mean (1-Year) Changes in Log Earnings by Transition Type (Tanzania) Transition Mean change Number of trans SE w/o -> SE w/o 0.184 285 SE w/o -> SE w 0.455 44 SE w/o -> W <=10 0.568 7 SE w/o -> W >10 0.392 9 SE w/o -> Civil 0.531 4 SE w/o -> Pub En 1.184 2 SE w -> SE w/o 0.111 28 SE w -> SE w 0.049 26 SE w -> W <=10 –0.289 1 SE w -> W >10 –1.511 1 SE w -> Civil . 0 W <=10 -> SE w/o 0.418 3 W <=10 -> SE w 0.019 2 W <=10 -> W <=10 0.134 28 W <=10 -> W >10 –0.020 17 W <=10 -> Civil 0.656 7 W <=10 -> Pub Ent –0.023 2 W >10 -> SE w/o –0.639 2 W >10 -> W <=10 0.038 10 W >10 -> W >10 0.171 34 W >10 -> Civil –0.090 23 W >10 -> Pub Ent 0.036 14 Civil -> W <=10 –1.106 2 Civil -> W >10 0.196 6 Civil -> Civil –0.016 18 Civil -> Pub Ent 1.839 2 Pub Ent -> SE w/o –0.070 2 Pub Ent -> W <=1 –0.016 4 Pub Ent -> W >10 –0.056 5 Pub Ent -> Civil 0.084 9 Pub Ent -> Pub Ent 0.037 11 Total 0.169 608 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: SE w/o = self-employed without employees; SE w = self-employed with employees; W <=10 = wage employee working in a firm with 10 employees or less; W>10 = wage employee working in a firm with more than 10 employees; Civil = Civil servant; Pub Ent = public enterprise. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 56 Summary Statistics and Variable Definitions Variable Definitions Education This is a continuous variable derived from assigning a number of years equiva- lent-based on the structure of the school systems in Ghana and Tanzania-to the highest educational attainment reported. Firm Size This variable captures the number of workers, including the respondent, who work in the same firm. For self-employed entrepreneurs, it is set to 0. Height The variable is constructed to be time invariant by averaging observations on height over time. This helps to smooth measurement error. Tenure This variable is constructed by calculating the number of years between the start of current job-as recalled by the respondent-and the date of the interview. Number of Employees Self-employed entrepreneurs are asked to report the total number of people they employ in their business, including both household members (paid or unpaid) and non-household members. The variable is set to 1 for the self-employed entrepreneurs who don’t hire any employees and for all wage workers (in the public and private sector). Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 APPENDIX B A Framework for Analyzing Earnings Panel Data The starting point is the following semilogarithmic model of earnings: yit = a0 + a1X1it + a2X2i + ui + eit (B.1) where: yit is the natural logarithm of net monthly income. X1it is a vector of time-variant determinants of earnings (including common time trends). X2i is a vector of time-invariant determinants of earnings. ui is a time-invariant, individual-specific determinant of earnings. eit is a time-varying error term. Ordinary least squares (OLS) estimates of equation (B.1) (presented in table 4.2) will be unbiased and consistent if both time-variant and time-invariant determinants of earnings are uncorrelated with the time-invariant (OLS A1) and time-variant (OLS A2) components of the residual: (OLS A1): E[X1iteit] = E[X2ieit] = 0 (OLS A2): E[X1itui] = E[X2iui] = 0 Tackling the Endogeneity of Schooling We ignore the role of the time-invariant fixed effects ui for the time being, and focus on the potential violations of assumption OLS A1 due to the potential endogeneity of education, which might arise if educational attainment is corre- lated with unobserved ability and ability affects earnings. In such a scenario, OLS estimates of equation (B.1) will be biased (since E[X1iteit] ≠ 0). To correct for this potential bias, we use two approaches; the preferred approach is to control for ability directly, by including measures of cognitive skills, which we use as proxies. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   57 58 A Framework for Analyzing Earnings Panel Data In addition, we use a control function (CF) approach, where the residual of a model of educational attainment is used as a regressor in the earnings equation. The idea behind the CF approach is to model the dependence between the unobserved error terms in such a way that the endogeneity bias disappears. The model of educational attainment, Ei, is presented in table B.1 and uses distance to primary school at age 6 and distance to secondary school at age 16 as exclusion restrictions, Zi, that is, they are assumed to be correlated with education and do not have a direct impact on earnings (that is, once education is controlled for, they do not have any additional impact on earnings). In addition, the model controls for the individual’s age and gender. The estimable equation thus becomes: Eit = ϕ 0 + ϕ1 X1it + ϕ 2 X 2i + ϕ 3Zi + ηit (B.2) Predicted residuals from the first-stage regression are then used as controls for unobserved factors affecting both earnings and education in an earnings specifi- cation that is equal to equation (B.1) in all other respects.  + u + ε (B.3) yit = α0 + α1 X1it + α2 X 2i + η it i it where:  = E − (ϕ η it it  +ϕ 0  X +ϕ 1 1it  X +ϕ 2 2i  Z ) 3 i As shown by Wooldridge (2007), under the rather stringent assumptions that: ( ) ( ) ( ) E ε it |X1it , X 2i , Zi , Eit = E ε it |X1it , X 2i , Zi , ηit = E ε it |ηit = ρ ηit  estimates of a0, a1, and a2 will now be unbiased. Note that the first equality holds because Eit and ηit are one-to-one functions of each other. The second equality holds if (εit, ηit) is independent of (X1it, X2i, Zi ) and if we are willing to assume linearity in the conditional expectation E ( ε it |ηit ) . Both these conditions are nontrivial, but generate an estimator that is more efficient than standard IV in nonlinear models. Controlling for Unobserved Fixed Effects In section B.2, we relax the second of the identification assumptions and allow for potential correlation between fixed effects ui and observable determinants of earnings, X1it and X2i: E[ X1it ui ] ≠ 0; E[ X 2i ui ] ≠ 0  Instead, we use two alternative assumptions. First, we assume that the time- variant determinant of earnings is uncorrelated with time-variant unobservables at any other point in time. (WG A1): E[ X1isε it ] = 0 ∀ s, t Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Earnings Panel Data 59 Table B.1  Control Function Approach to Instrument Education and Apprenticeships (First Stage) Country Ghana Tanzania Coef/SE Coef/SE Coef/SE Coef/SE Estimation method OLS Probit OLS Probit Dependent variable: Educational Completed an Educational Completed an attainment apprenticeship? attainment apprenticeship? Male 1.604*** 0.144 0.245 −0.059 (0.257) (0.092) (0.308) (0.137) Age 0.297*** 0.014 0.146 0.153*** (0.076) (0.028) (0.093) (0.049) (age^2)/100 −0.428*** −0.032 −0.243** −0.164*** (0.103) (0.038) (0.118) (0.060) Distance to the nearest primary school at age 6 0.024** −0.001 −0.010 −0.005 (0.012) (0.004) (0.012) (0.006) Distance to the nearest secondary school at age 16 −0.015*** −0.001 −0.430* −0.027 (0.005) (0.002) (0.255) (0.111) Constant 3.405** −0.665 4.242** −4.434*** (1.343) (0.487) (1.825) (1.008) Number of observations 866 866 766 766 R2 0.075 0.028 Pseudo R2 0.01 0.05 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Coef = coefficient; SE = standard error; OLS = ordinary least squares. ***p<0.01   **p<0.05   *p<0.1 This assumption will allow us to employ a within-group estimator, which is effectively equivalent to OLS on the following transformed model:  y  +ε  it = α1 X  it (B.4) 1it T T T 1  =X −1 1  it = yit − where y T ∑ yit , X 1it 1it T ∑X1it and ε  it = ε it − T ∑εit . It should 1 1 1 be noted that in small samples ε  it is negatively correlated with X 1it by con-  struction, leading to the Nickell bias (Nickell 1981), which is typically in the opposite direction of the bias in the OLS estimator. Second, we make the less restrictive assumption that the time-variant deter- minants of earnings are uncorrelated with time-variant unobservables only in the same and in the previous period. (FD A1): E[ X1it ε it −1 ] = E[ X1it −1ε it ] = 0 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 60 A Framework for Analyzing Earnings Panel Data This second variant will justify the estimation of the following model in first differences, using OLS regressions. ∆yit = α1∆X1it + ∆ε it (B.5) Since first differences are not available for the first sample wave, this leads to a reduction in sample size, and, consequently, less precise estimates. Sorting Matters, But Is Not the Entire Story If time-invariant determinants of earnings that are unobserved yet correlated with the explanatory variables exist, the OLS estimates may be biased. For example, the sector differentials may simply reflect differences in unobserved ability. By first differencing the earnings regressions (“first difference [FD] esti- mator”) or using fixed effects estimation (the “within-group estimator” or “fixed effect [FE] estimator”), one can control for such individual-specific unobserv- ables. However, as explained in section B.1, this comes at a cost, because we are then no longer able to assess the impact of time-invariant variables on earnings. In addition, the fixed effects estimator might be downward biased due to the fact that, by construction, the error terms are correlated with the regressors (the so- called Nickell bias). While the FD estimator does not suffer from the same draw- back, it does lead to reductions in sample size and, consequently, less precise estimates. Yet, we still prefer the latter over the FE estimators since it yields unbiased estimates. Since first-differenced estimates examine the determinants of earnings changes as a function of changes in observable characteristics, they also provide a first-pass at the determinants of earnings growth. Table B.2 presents the results of earnings functions estimated by means of fixed effects estimation (columns 1 and 4) and using the within group estimator (columns 2 and 5). The regressions only include time-varying observable charac- teristics. Note that we have dropped age, since, by construction, the change in age from one year to the next will be constant. Overall, the results do not change very much compared to the OLS specifi- cations. To start with, the estimated firm-size effect is relatively robust to controlling for fixed effects, suggesting that the initial observation that workers in large firms earn more than those in small firms is correct. The fact that the effect is lower than in the OLS specifications suggests it partially reflects sort- ing of more able individuals into larger firms (see the discussion in the review of the related literature in chapter 2). Turning to sectoral premia, the results from the within-group estimation show a strong and significant civil service premium in Ghana and a strong public enterprise premium in Tanzania. The effect of working in the private sector, although of the same sign as in previous regressions, does not appear significant. In sum, these results suggest that sec- toral premia reflect both sorting across sectors and genuine differences in remuneration between sectors. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Earnings Panel Data 61 Table B.2  Earnings Functions (FE and FD) Country Ghana Tanzania Estimation method FE FD(1) FD(2) FE FD(1) FD(2) Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE (age^2)/100 −0.008 −0.017 0.001 0.040* 0.029** 0.031** (0.090) (0.077) (0.080) (0.024) (0.014) (0.014) Ln (hours) 0.176** 0.076 0.079 0.138 0.132 0.147 (0.074) (0.065) (0.065) (0.155) (0.103) (0.101) Tenure 0.004 0.000 0.001 0.009 0.015* 0.014* (0.005) (0.005) (0.005) (0.009) (0.008) (0.009) Ln (employees) 0.161* 0.133* 0.123* 0.244 0.255** 0.247* (0.089) (0.073) (0.073) (0.171) (0.126) (0.127) Ln (firm size) 0.125*** 0.082*** 0.084*** 0.065* 0.045** 0.051* (0.037) (0.032) (0.031) (0.036) (0.021) (0.026) Priv wage −0.256** −0.169 0.123 0.184 (0.126) (0.116) (0.196) (0.137) Civil service 0.312* 0.241 0.101 0.146 (0.177) (0.153) (0.257) (0.206) Public enterprise −0.166 −0.056 0.176 0.254 (0.198) (0.163) (0.278) (0.179) Self -> priv wage 0.019 0.044 (0.138) (0.196) Self -> civil 0.538 0.337 (0.821) (0.259) Self -> Pub Ent 2.067*** 0.903*** (0.162) (0.171) Priv wage -> priv wage −0.077** −0.094 (0.031) (0.069) Priv wage -> self 0.533*** −0.131 (0.166) (0.221) Priv wage -> civil −0.006 −0.088 (0.138) (0.141) Priv wage -> Pub Ent −0.202 −0.191* (0.168) (0.099) Civil -> civil −0.133** −0.132 (0.066) (0.088) Civil -> self −0.620*** (0.148) Civil -> priv wage −0.497** −0.230 (0.208) (0.627) Civil -> Pub Ent −0.621*** 1.726 (0.229) (1.337) Pub Ent -> Pub Ent 0.002 −0.122* (0.089) (0.072) table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 62 A Framework for Analyzing Earnings Panel Data Table B.2  Earnings Functions (FE and FD) (continued) Country Ghana Tanzania Estimation method FE FD(1) FD(2) FE FD(1) FD(2) Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Coef/SE Pub Ent -> self −0.016 −0.131 (0.370) (0.332) Pub Ent -> priv wage −0.205** −0.174 (0.103) (0.131) Pub Ent -> civil 0.326 0.043 (0.209) (0.194) Year dummies Yes Yes Yes Yes Yes Yes Constant 1.628 0.184** 0.168** 9.387*** 0.119** 0.142*** (1.140) (0.074) (0.076) (0.719) (0.049) (0.055) Number of observations 3,454 1,753 1,753 1,604 610 610 R2 0.150 0.043 0.060 0.113 0.043 0.063 Adjusted R2 0.147 0.037 0.047 0.108 0.028 0.031 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Coef = coefficient; SE = standard error. FD = first difference; FE = fixed effect. NB First differenced estimates use the first difference of the dependent variable as well as first differenced explanatory variables; for example, Ln (employees) is in fact ΔLn (employees). FE and FD estimates are presented alongside each other to conserve space. ***p<0.01 **p<0.05 *p<0.1 What Do These Regressions Tell Us About Growth? Asymmetric Sec- toral Switching Premia The FD estimator can be interpreted as a growth regression because it exam- ines the determinants of earnings changes. Both the within-group estimator and the FD specification have implicitly assumed symmetry in the effect of sector switching. In other words, the effects of going in and out of a certain sector have implicitly been assumed to be reciprocal. Because this is a strong restric- tion, we set out to test it. Column 3 shows the results of an FD estimation (similar to that used for column 2), where we substituted the first differences of sectoral dummies with switch-specific dummies, capturing the effect of moving from one particular sector into another, thus allowing for asymmetric effects of switching between sectors. Tables A.3 and A.4 presented descriptive data on the amount of sector switching and demonstrated that most people gain from doing so. Here we control for their observable characteristics, thus examining whether changes in earnings reflect changes in observable charac- teristics or pure sector effects. The results demonstrate that the effects documented above are robust when controlling for changes in time-varying observable characteristics. Moving from the private sector to self-employment has a strong positive effect on earnings, yet moving in the opposite direction does not have an effect of the opposite sign. In fact, once we account for the effect of the Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Earnings Panel Data 63 increase in firm size that derives from moving from self-employment to private sector employment, this switch has itself a positive effect. Some evidence of symmetry, on the other hand, seems to exist for movements in and out of the public service; movements into the public sector are associ- ated with substantial pay rises, while movements out of it are associated with significant pay cuts. The number of such switches, however, is too low to heavily rely on these results. Parallel results for Tanzania are included for completeness, but due to the smaller sample size, the number of switches observed is too small to make substantive conclusions. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 APPENDIX C A Framework for Analyzing Earnings Growth Econometric Framework The framework starts from the following general growth model: yit = α0 + λ yi,t −1 + α1 X1it + α2 X 2i + α3 X 2i ⋅ time + ui + ρ i ⋅ time + ε it (C.1) where both time-varying X1it and time-invariant X2i observables have an impact on earnings. The crucial difference with the model in levels presented in chapter 4 is an explicit treatment of income dynamics by including income in the previous period among the explanatory variables. Moreover, by allowing time-invariant observables X2i to have a different impact on earning levels at different points in time (hence the interaction term α3X2i ⋅ time), one can test whether personal and job characteristics impact earnings growth. As pointed out by Deaton (1997, 110), in short panels, it is very difficult to distinguish between persistence in earnings due to unobserved individual hetero- geneity (as captured in ui and ρi) and persistence due to the effect of the lagged dependent variables, as captured by λ. Differencing equation (C.1) helps get rid of the fixed effect ui, yet also induces serial correlation in the error term, which will yield a downward bias in ordinary least squares (OLS) estimates of λ. This observation is a key concern in the large literature on earnings convergence, which typically finds strong evidence for high persistence and regression to the mean. To circumvent the identification problems that introduction of the lagged dependent variable would entail, we make the very strong assumption that changes in earnings are fully persistent, that is, that λ = 1. This assumption is restrictive, thus the model becomes: yit = α0 + yi −1t + α1 X1it + α2 X 2i + α3 X 2i ⋅ time + ui + ρ i ⋅ time + ε it (C.2) Differencing yields the model we estimate: ∆yit = α1∆X1it + α3 X1it −1 + ρ i + ∆ε it (C.3) Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   65 66 A Framework for Analyzing Earnings Growth Equation (C.3) will allow us to identify the effect of time-invariant factors on growth, while controlling for the changes in time-variant determinants of earn- ings levels. Moreover, when modeling growth from period t–1 to t, we can include time-variant factors measured at time t–1 among the time-invariant characteristics in X2i. This approach, explained in Quinn and Teal (2008), is motivated by the observation that time-variant characteristics measured at t–1 that are predetermined are effectively time invariant with respect to growth between t–1 and t, and can therefore be included among the time-invariant regressors. At the risk of belaboring the point, this equation allows discrimination between changes in earnings due to changes in explanatory variables-the “levels” effect of such variables-and changes due to the fact that explanatory variables might have an additional impact on individual earnings growth rates, the “growth” effects of such variables. To understand the difference, an analogy with the growth accounting literature may be of interest. Consider a steady state earnings growth trajectory, where earnings grow at a constant speed g and where indi- vidual and job characteristics are fixed such that X1it = X1it–1 = Xi* then g = α3 X i* + ρ i . In other words, the coefficient a3 measures the effect of the variables Xi* on the long-run growth path. By contrast, the coefficients a1 only affect earnings growth during adjustment to the steady state equilibrium. The above model can be estimated with OLS if we are willing to assume that the error term is uncorrelated with the explanatory variables, for example, if  X 2i ρ i  (OLS A1): E[ X1isε ij ] = 0 ∀ s, t  ε {t, t–1} and E[ X1is ρi ] = E   = 0 ∀ s, t  ε {t, t–1} (OLS A2). As in chapter 4, the fixed effect ρ can be tackled by using fixed effects and first-differences estimators. Tackling Measurement Error: The Determinants of Earnings Growth over a Two-Year Period The lack of strong predictors of earnings growth might partially be due to mea- surement error; if earnings are measured with a great deal of error, this may lead to attenuation bias. To overcome this problem, we estimated regressions where the growth of earnings over a two-year period is used as the dependent variable. The advantage of using a longer time window is that the signal-to-noise ratio in the data ought to be higher, in the sense that the proportion of the observed change in earnings that is due to measurement error should be smaller over a two-year period than over a one-year period. On the other hand, using earnings changes over a two-year period might exacerbate attrition bias. In addition, the differenced sample is much smaller, leading to less precise estimates. The estimates are presented in table C.1—for the purpose of comparability, results using annual changes in monthly income as the dependent variable are also included. As can be seen by comparing the columns, the pattern of results does not change dramatically. However, comparison of these specifications with those presented in table 5.1 does suggest that attrition bias may be a problem. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Earnings Growth 67 Table C.1  Determinants of Two-Year Growth in Log Earnings Ghana Tanzania 2 year 1 year 2 year 1 year Coef/SE Coef/SE Coef/SE Coef/SE Male –0.074 –0.042 –0.627** –0.352** (0.110) (0.072) (0.259) (0.155) L2.age –0.002 –0.001 0.015 0.008 (0.005) (0.004) (0.017) (0.011) Height (cm) 0.005 0.000 0.010 0.000 (0.009) (0.005) (0.024) (0.013) Years in formal education –0.031 –0.012 0.028 0.022 (0.037) (0.025) (0.103) (0.060) (educ∧2)/100 0.003 0.001 –0.008 –0.004 (0.003) (0.002) (0.007) (0.004) Math score 0.001 –0.001 –0.010 –0.005 (0.002) (0.001) (0.010) (0.006) L2.apprenticeship completed 0.032 0.004 0.274 0.182 (0.091) (0.065) (0.307) (0.182) L2.apprentice (currently) –0.375 0.284 (0.757) (0.248) Δ2Ln (hours) –0.101 –0.110 0.211 0.189 (0.092) (0.075) (0.294) (0.250) Δ2tenure –0.002 –0.002 0.024 0.024* (0.007) (0.007) (0.018) (0.014) L2.tenure –0.005 –0.000 –0.028 –0.001 (0.007) (0.005) (0.028) (0.016) Δ2Ln (employees) –0.010 0.018 –0.076 –0.672** (0.113) (0.143) (0.315) (0.324) L2.Ln (employees) 0.202 0.091 0.081 –0.568 (0.167) (0.090) (0.351) (0.455) Δ2Ln (firm size) 0.071** 0.053* 0.131** 0.164*** (0.031) (0.030) (0.066) (0.058) L2.Ln (firm size) 0.003 –0.003 0.168 0.179** (0.030) (0.021) (0.112) (0.074) Self -> priv wage –0.331 –0.148 –0.468 –0.363 (0.281) (0.159) (0.415) (0.251) Self -> public –0.937 0.222 1.411* 0.721** (2.473) (1.156) (0.748) (0.302) Priv wage -> priv wage 0.028 0.002 –0.103 –0.467** (0.133) (0.085) (0.421) (0.232) Priv wage -> self 0.412 0.194 –0.205 0.091 (0.487) (0.304) (0.530) (0.590) Priv wage -> public 0.288 –0.237* 0.448 –0.149 (0.400) (0.137) (0.908) (0.401) Public -> public 0.051 0.018 0.514 –0.305 (0.196) (0.098) (0.616) (0.354) table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 68 A Framework for Analyzing Earnings Growth Table C.1  Determinants of Two-Year Growth in Log Earnings (continued) Ghana Tanzania 2 year 1 year 2 year 1 year Coef/SE Coef/SE Coef/SE Coef/SE Public -> self –0.202 –0.007 (0.466) (0.434) Public -> priv wage –0.123 –0.067 –0.278 –0.615 (0.242) (0.148) (0.735) (0.412) Constant –1.281 –0.031 –0.757 0.652 (1.491) (0.848) (3.595) (2.052) Number of observations 706 706 157 157 R2 0.061 0.067 0.206 0.243 Adjusted R2 0.022 0.028 0.047 0.092 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Coef = coefficient; SE = standard error. ***p<0.01   **p<0.05   *p<0.1 Controlling for Fixed Effects To control for bias arising from unobserved time-invariant, individual-specific effects that impact growth and are also correlated with earnings, we estimate the preferred model by means of fixed effects and first differences estimators. Variables measuring the impact of doing an apprenticeship are removed since there are too few observations to draw reliable conclusions. The results of these regressions are presented in table C.2. Overall, the results do not change substan- tially, save for the estimated sectoral premia. The results of these specifications ought to be interpreted with caution, however, because the number of observa- tions is relatively small. Table C.2  FE and FD Estimates of Annual Earnings Growth Ghana Tanzania FD FE FD FE Method Coef/SE Coef/SE Coef/SE Coef/SE lrearn l.Δage 0.228 0.186 0.102 0.102 (0.362) (0.435) (0.070) (0.137) Δ2Ln (hours) –0.095 –0.053 0.160 0.160 (0.090) (0.107) (0.223) (0.438) Δ2tenure –0.009 –0.003 0.039 0.039 (0.010) (0.011) (0.025) (0.049) Δtenure –0.009 –0.002 0.035 0.035 (0.011) (0.009) (0.029) (0.057) table continues next page Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Earnings Growth 69 Table C.2  FE and FD Estimates of Annual Earnings Growth (continued) Ghana Tanzania FD FE FD FE Method Coef/SE Coef/SE Coef/SE Coef/SE Δ2Ln (employees) –0.232 –0.019 –0.124 –0.124 (0.178) (0.165) (0.517) (1.012) LΔLn (employees) –0.450* –0.146 –0.393 –0.393 (0.265) (0.215) (0.639) (1.253) Δ2Ln (firm size) 0.097*** 0.069 0.069 0.069 (0.038) (0.045) (0.092) (0.181) LΔLn (firm size) 0.077 0.041 –0.076 –0.076 (0.054) (0.057) (0.199) (0.390) Δself->priv wage –0.384 –0.250 –0.445 –0.445 (0.269) (0.319) (0.407) (0.796) Δself->public 0.013 0.403 1.186 1.186 (1.310) (1.509) (0.767) (1.503) Δpriv wage->priv wage –0.989*** –0.540* –1.954*** –1.954 (0.302) (0.300) (0.744) (1.458) Δpriv wage->self 0.963** 0.627 1.131** 1.131 (0.403) (0.447) (0.549) (1.075) Δpriv wage->public –0.085 0.022 1.133* 1.133 (0.378) (0.408) (0.597) (1.170) Δpublic->public –1.117** –0.547 –0.316 –0.316 (0.484) (0.374) (0.701) (1.374) Δpublic->self 0.237 –0.114 (0.544) (0.463) Δpublic->priv wage –0.068 –0.197 0.047 0.047 (0.244) (0.248) (0.589) (1.155) Year dummies Yes Yes Yes Yes Constant –0.252 –5.919 –0.242 –3.484 (0.371) (14.663) (0.156) (5.248) Number of observations 706 1,461 157 602 R2 0.070 0.059 0.107 0.113 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: Coef = coefficient; SE = standard error; FD = first difference; FE = fixed effect. ***p<0.01   **p<0.05   *p<0.1 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 APPENDIX D A Framework for Analyzing Transitions between Low- and High-Paid Employment Econometric Framework The model for low-pay/high-pay transitions is a first-order Markov model that accounts for the initial conditions problem and nonrandom retention by treating them as issues of multiple endogenous selection. The modeling framework is an adaptation of the model proposed by Cappellari and Jenkins (2004). More spe- cifically, we use multivariate probit models to model low-earnings transitions between two consecutive years, pooling observations across observed transitions. There are three parts to the most general model. The first equation models initial low-pay determination at t–1 to control for the initial conditions problem. Lit −1* = β X it −1 + uit −1 where uit −1 = µi + δit −1 and Lit −1 = I ( Lit −1* ≥ τ ) (D.1) where Xit-1 is a vector observable characteristic and uit-1 is an error term that is the sum of an individual-specific effect and white noise δ it −1 , and Lit −1 is an indicator variable indicating whether individual i’s earnings fell below the low- pay threshold in period t–1 or not. The second equation models the probability that individuals whose earnings were observed at time t–1 will also be observed at time t, thus allowing for the possibility that individuals either exit the sample at t or become nonparticipants without earnings.1 Rit * = ψ Wit −1 + ε it where ε it = τ i + ζ it and Rit = I (Rit * ≥ 0) (D.2) where Wit-1 is a vector of observable characteristics that affect the retention propensity and eit is again an error term that is assumed to be composed of an individual-specific effect τi and white noise ζit. If individual i’s retention Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   71 72 A Framework for Analyzing Transitions between Low- and High-Paid Employment probability is lower than a critical threshold, then his/her earnings are not observed in period t. Rit is an indicator variable that indicates whether an indi- vidual was retained. The transition equation models low pay in period t, conditioning on low-pay status at t–1 as a function of observable characteristics Xit. Lit * =  ( )   Lit −1γ 1 + 1 − Lit −1 γ 2  Zit −1 + vit where vit = oi + π it      and Lit = I ( Lit * ≥ τ ) (D.3) This specification is an endogenous switching regression, since the impact of covariates Zit-1 depends on the previous pay state. Again, the error process vit is assumed to be composed of an individual-specific component and a random error process. The error terms from the three equations are assumed normally jointly dis- tributed, which implies a trivariate normal model that can be estimated using simulated maximum likelihood. Allowing correlation between the three equations helps control for possible individual unobserved heterogeneity influencing both initial likelihood of low pay and transition between pay states, and hence to explore whether unobserved differences account for persistence in low earnings. Likelihood Function The truncated trivariate probit model to account for endogeneity due to nonre- tention has the following likelihood function: N  Lit −1γ 1 + (1 − Lit −1 ) γ 2  Zit −1 , ki0 β X it −1 , kiRψ Wit −1 , ρ1 , ρ 2 , ρ3 ) ∑Rit Φ3 ( ki1  lnL =  i =1  (D.4) + (1 − Rit ) Φ2 ( ki0γ Zit −1 , kiRψ wit )   where ki1 = 2 Lit − 1 , ki0 = 2 Lit −1 − 1 , and kiR = 2 Rit − 1 . This model can be computed using maximum simulated likelihood (Cappellari and Jenkins 2003; Train 2003). As observed above, we pool transi- tions across multiple years. To correct for potential violations of the assump- tion that errors are identically and independently distributed, standard errors are clustered. It should be noted that the normality assumption is violated by construction due to the presence of a lagged dependent variable in the transition equation and that same lagged dependent variable being one of the equations estimated. However, the normality assumption is required for tractability. Furthermore, Cappellari and Jenkins (2004) argue that violations of this assumption do not significantly affect the results. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Transitions between Low- and High-Paid Employment 73 Model Specification-Bivariate Models While this framework allows the model to control for the initial conditions and nonrandom retention, below we present bivariate models of earnings transitions that control for the initial conditions problem only. The bivariate models use personal characteristics (education, gender, age, height, whether the person has ever been an apprentice) and job characteristics (sector dummies, tenure, firm size, or number of employees) as well as dummies for current city and survey year as explanatory variables. Parental education is used as an instrument for the conditions equation. Table D.1 shows the model specification tests for the initial ­ bivariate models in both countries. Parental education instruments function well; we cannot reject that they are insignificant in the transition equation, but do strongly reject that they are insignificant in the initial low earnings equation. The estimated correlation between unobservables in the two equations is negative in both countries, though insignificant in Tanzania, which could reflect that, because of a smaller sample size in Tanzania, the true correlation in the popula- tion could not be uncovered. This also suggests that there is “regression to the mean” in the sense that individuals who were low paid last year are more likely to be high paid this year, potentially reflecting an error in the earnings measurement. Transition Probabilities Both the bivariate and trivariate models allow calculation of the low-income persistence rate sit (the probability of being low paid at t, conditional on being low paid at t–1) and the low-income entry rate eit (the probability of being low paid at t, conditional on being high paid at t–1). Φ2 (γ 1 ` Zit −1 , β `xit −1; ρ ) (D.5) sit ≡ Pr (Lit = 1| Lit −1 = 1) = Φ ( β ` xit −1 ) Φ2 (γ 2 ` Zit −1 , −β `xit −1; − ρ ) (D.6) eit ≡ Pr (Lit = 1| Lit −1 = 0) = Φ ( −β ` xit −1 ) Table D.1  Bivariate Specification Tests in Tanzania and Ghana Test Test statistic P value Ghana Exclusion of parental education from transition equation 0.09 0.95 Exclusion of parental education from initial equation 12.75 0.00 Rho = 0 –2.13 0.033 No state dependence 370.46 0.00 Tanzania Exclusion of parental education from transition equation 0.79 0.67 Exclusion of parental education from initial equation 15.99 0.00 Rho = 0 –1.58 0.11 No state dependence 97.82 0.00 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 74 A Framework for Analyzing Transitions between Low- and High-Paid Employment The numerators in (D.5) and (D.6) are the probability that Lit is 1 or 0, respec- tively, at time t, given period t–1 characteristics, with Φ2 being a bivariate normal cumulative density function and ρ measuring the correlation between the transi- tion equation (equation D.3 above) and the initial low-pay equation (equa- tion D.1). These probabilities are conditioned on either being low paid or high paid at t–1, being either 1 or 0, hence these events appear in the denominator in each equation, Φ being a univariate normal cumulative density function.2 The endogenous switching model implies regressors have different effects, depending on whether conditioning on high or low pay. Hence, the main analysis explores the effects of regressors both on sit and eit. In addition, this analysis also examines what changes in covariates imply for predicted mean and median time in low and high pay. If it is additionally assumed to be a stationary environment, we can calculate mean and median duration of low and high pay: the formula for mean duration of low pay is 1 / (1 − sit ) and for the median duration it is log(0.5)/log(sit ). For those in high pay, the mean duration of a spell of high pay is 1/eit and the median duration is log(0.5)/log(eit) (for proofs, see Boskin and Nold [1975]). State Dependence Aggregate state dependence (ASD) is defined here as the difference between the probability of being low paid at t for those low paid at t–1 and the probability of being low paid at t for those not low paid at t–1.    ∑ i ∈{L =1}Pr( Lit = 1 | Lit −1 = 1)   ∑i ∈    { L = 0} Pr ( Lit = 1 | Lit −1 = 0)    ASD =  it − it     ∑i Lit −1       ∑i (1 − L it −1 )    While measures of ASD are arguably less prone to measurement error than measures of raw state dependence, they do not control for individual heterogene- ity. By contrast, genuine state dependence (GSD), defined as the average differ- ence between the predicted probabilities of being low paid, conditional on being low paid and high paid, controls for both unobserved and observed characteristics. N 1 GSD = N ∑ Pr (Lit = 1| Lit −1 = 1) − Pr (Lit = 1| Lit −1 = 0) i =1  GSD is thus the preferred measure of state dependence. If differences in observed individual characteristics are the main drivers of initial pay states and transitions between states, then one would expect GSD to be lower than ASD. If labor market scarring is occurring, then one would expect GSD to be greater than zero. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Transitions between Low- and High-Paid Employment 75 Results Maximum Likelihood (ML) Coefficients from Underlying Model Table D.2 shows coefficient estimates from the underlying maximum likeli- hood estimation of the bivariate models. The coefficient estimates for the other equations are available, but were omitted to conserve space. The first column shows the effects of covariates conditional on being high paid, the second is conditional on being low paid (both in the transition equation), and the third is the p value from a test of whether these effects are significantly different from each other. Coefficient estimates are generally not significant in the transition equation, mainly because we control for initial selection into low pay in a separate equation, where all the coefficients are highly significant. Hence, a more intuitive method of examining the effects of individual and job characteristics is to explore their effects on the predicted probabilities of per- sistence of low pay, sit, and entry into low pay, eit, also allowing these to affect the probability of being low paid in the base period (see chapter 6). Trivariate Model This study only shows results for the trivariate specification for Ghana (table D.3), since the smaller sample size in Tanzania led to estimates that were not robust across specifications and to models that did not always con- verge, given the higher dimension of numerical integration required. Table D.4 shows that in Ghana, the correlation coefficient between the initial earnings equation and the transition equation is negative and significant, as in the bivariate case, and the magnitude is also similar to the bivariate model esti- mated above. Table D.4 also shows that the correlation coefficient between the transition equation and retention equations is positive and significant, suggesting that unobserved characteristics that make an individual low paid in the next period also increase the probability of the survey obtaining an earn- ings measure in the next period. The instruments for selection into initial low earnings work, and results show that the instruments for retention are not significant in the transition equation. Again, our analysis focuses on the effects of changing covariates on both per- sistence and entry probabilities, shown in Table D.5. The results seem to be robust across specifications, with almost all the estimated effects in the trivariate model for Ghana qualitatively the same as in the bivariate model. However, there is now less of a difference in the variance of entry and persistence rates than the bivariate model suggested. Compared to the reference individual, women are nearly twice as likely to enter low pay, and younger workers are again more likely to remain in or enter low pay. Older and more educated workers are less likely to enter low-paid employment in future periods. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 76 A Framework for Analyzing Transitions between Low- and High-Paid Employment Table D.2  Maximum Likelihood Coefficient Estimates for Bivariate Probit Ghana Tanzania P value P value High Low coefficient High Low coefficient earner earner equality earner earner equality Male –0.215** 0.215 0.077 –0.261 –0.138 0.774 (0.106) (0.152) (0.213) (0.294) Age –0.004 0.007 0.351 0.008 –0.020 0.191 (0.005) (0.007) (0.010) (0.014) Years in formal education –0.019 –0.006 0.846 0.201*** –0.118 0.050 (0.034) (0.047) (0.075) (0.101) (educ^2)/100 0.008 0.257 0.619 –1.944*** 0.729 0.113 (0.264) (0.380) (0.719) (1.089) Priv wage 0.032 –0.308 0.358 0.056 0.364 0.745 (0.157) (0.225) (0.427) (0.580) Public –0.187 0.550 0.135 –0.544 0.425 0.496 (0.214) (0.425) (0.707) (0.929) Tenure 0.008 –0.016* 0.107 –0.013 0.030* 0.127 (0.006) (0.010) (0.014) (0.018) Ln (hours) 0.152 0.003 0.479 –0.008 0.224 0.699 (0.118) (0.146) (0.323) (0.390) Ln (employees) –0.170 0.385** 0.084 –0.307 0.638 0.131 (0.128) (0.185) (0.300) (0.428) Ln (firm size) –0.049 0.204** 0.025 –0.117 0.233 0.199 (0.048) (0.082) (0.127) (0.208) Apprenticeship completed –0.089 –0.005 0.696 –0.157 0.818* 0.173 (0.099) (0.130) (0.311) (0.438) Height (cm) 0.002 –0.003 0.718 0.002 –0.022 0.390 (0.007) (0.009) (0.013) (0.018) Number of observations 2,275 1,485 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: ***p<0.01   **p<0.05   *p<0.1 Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 A Framework for Analyzing Transitions between Low- and High-Paid Employment 77 Table D.3  Maximum Likelihood Coefficients for Trivariate Model, Ghana High earner Low earner Male –0.221** 0.105 (0.096) (0.133) Age –0.004 0.003 (0.005) (0.007) Formal education –0.015 0.020 (0.032) (0.047) (educ^2)/100 –0.079 –0.042 (0.256) (0.379) Tenure 0.008 –0.017* (0.006) (0.009) Ln (hours) 0.172 –0.020 (0.126) (0.154) Priv wage –0.009 –0.254 (0.152) (0.216) Public –0.219 0.194 (0.201) (0.439) Ln (employees) –0.133 0.174 (0.124) (0.191) Ln (firm size) –0.028 0.166** (0.045) (0.079) Apprenticeship completed –0.047 0.102 (0.098) (0.128) Number of observations      2,932        2,932 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Note: ***p<0.01    **p<0.05    *p<0.1 Table D.4  Trivariate Specification Tests for Ghana Test statistic P value Exclusion of parental education from transition equation 0.93 0.63 Exclusion of parental education from initial equation 12.75 0.00 ρ1=ρ_transition/initial =0 –1.76 0.08 ρ2=ρ_transition/retention=0 2.13 0.03 ρ3= ρ_initial/retention=0 0.91 0.36 No state dependence 160.42 0 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 78 A Framework for Analyzing Transitions between Low- and High-Paid Employment Table D.5  Predicted Entry and Persistence Rates and Mean and Median Predicted Time in Low Pay for Ghana, Trivariate Normal Model Low entry Mean Median Low Mean low Median (%) high high persist (%) low Reference 8.35 11.98 7.95 29.61 1.42 0.57 Education 5.63 17.77 11.97 23.94 1.31 0.48 Female 14.56 6.87 4.41 38.20 1.62 0.72 Small firm 7.42 13.48 8.99 27.79 1.38 0.54 Public employee 4.21 23.75 16.11 21.21 1.27 0.45 Age 40 7.17 13.94 9.31 27.24 1.37 0.53 Apprentice 8.23 12.15 8.07 33.64 1.51 0.64 Tenure 8.70 11.50 7.62 27.01 1.37 0.53 1 employee 6.38 15.68 10.52 28.18 1.39 0.55 Age 20 11.27 8.87 5.79 35.05 1.54 0.66 Source: World Bank; values arrived at using the Tanzanian and Ghanaian UPSs. Notes: The reference person is a 30-year-old self-employed male with no employees, who has seven years of education, five years of tenure in his current job, and never been an apprentice. Low entry refers to the probability of being low paid in the next period conditional on being high paid currently. Low persist refers to the probability of continuing to be low paid in the next period. Notes 1. Modeling these processes separately (using two probit models instead of one) proved impossible with the data. 2. This result also holds for the trivariate model, by the well-known result that the mar- ginal distribution of an X dimensional normal distribution is an X–1 dimensional normal distribution. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Bibliography Belzil, C., and J. Hansen. 2002. “Unobserved Ability and the Return to Schooling.” Econometrica, Journal of the Econometric Society 70 (5): 2075–91. Boskin, M., and F. Nold. 1975. “A Markov Model of Turnover in Aid to Families with Dependent Children.” Journal of Human Resources 10: 476–81. Cappellari, L. 2002. “Do the ‘Working Poor’ Stay Poor: An Analysis of Low Pay Transitions in Italy.” Oxford Bulletin of Economics and Statistics 64 (2): 87–110. Cappellari, L., and S. P. Jenkins. 2003. “Multivariate Probit Regression Using Simulated Maximum Likelihood.” Stata Journal 3 (3): 278–94. ———. 2004. “Modeling Low Income Transitions.” Journal of Applied Econometrics 19: 593-610. Card, D. 2001. “Estimating the Return to Schooling: Progress on Some Persistent Econometric Problems.” Econometrica 69 (5): 1127–60. Chen, S., and M. Ravallion. 2008. “The Developing World Is Poorer Than We Thought, But No Less Successful in the Fight Against Poverty.” Policy Research Working Paper No. 4703, World Bank, Washington, DC. Deaton, A. 1997. The Analysis of Household Surveys: A Microeconomic Approach to Development Policy. Washington, DC: World Bank. Fafchamps, M., and M. Söderbom. 2006.”Wages and Labor Management in African Manufacturing.” Journal of Human Resources 41 (2): 346–79. Fafchamps, M., M. Söderbom, and N. Benhassine. 2009. “Job Sorting in African Labor Markets.” Journal of African Economies 18: 824–68. Fields, G. 2008. “A Review of the Literature on Earnings Dynamics in Developing Countries.” Mimeo. Fields, G., P. Cichello, S. Freije, M. Menendez, and D. Newhouse. 2003a. “For Richer or for Poorer? Evidence from Indonesia, South Africa, Spain, and Venezuela.” Journal of Economic Inequality 1 (1): 67–99. ———. 2003b. “Household Income Dynamics: A Four-Country Story.” Journal of Development Studies 40 (2): 30–54. Fox, L., and M. Gaal. 2008. Working Out of Poverty: Job Creation and the Quality of Growth in Africa. Washington, DC: World Bank. Heckman, J. 1981a. “Heterogeneity and State Dependence.” In Studies in Labor Markets, edited by S. Rosen, 91-140. Chicago: University of Chicago Press Books. ———. 1981b. “The Incidental Parameters Problem and the Problem: Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Process.” In Structural Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2   79 80 Bibliography Analysis of Discrete Data with Econometric Applications, edited by C. Manski and D. McFadden, 114–17. London: MIT Press. ILO (International Labour Organization). 2002. Women and Men in the Informal Economy: A Statistical Picture. Geneva: ILO. Johansson de Silva, S., and P. Paci. Beyond Job Creation: The Challenges of Developing an Employment Agenda in Developing Countrie. mimeo. Kahyarara, G., and F. Teal. 2008. “The Returns to Vocational Training and Academic Education: Evidence from Tanzania.” World Development 36 (11): 2223–42. Kingdon, G., J. Sandefur, and F. Teal. 2005. “Patterns of Labor Demand in Africa: Africa Region Employment Issues-Regional Stocktaking Review.” Mimeo. Maloney, W. 1999. “Does Informality Imply Segmentation in Urban Labor Markets? Evidence from Sectoral Transitions in Mexico.” World Bank Economic Review 13 (2): 275–302. Mead, D., and C. Liedholm. 1998. “The Dynamics of Micro and Small Enterprises in Developing Countries.” World Development 26 (1): 61–74. Moffitt, R., J. Fitzgerald, and P. Gottschalk. 1999. “Sample Attrition in Panel Data: The Role of Selection on Observables.” Annales d’Economie et de Statistique 55 (56): 129–52. Nickell S. 1981. “Biases in Dynamic Models with Fixed Effects.” Econometrica 49 (6): 1417–26. Pissarides, C. A. 2002. “Human Capital and Growth: A Synthesis Report.” Technical Report 168, OECD Development Centre, Paris. Quinn, S., and F. Teal. 2008. “Private Sector Development and Income Dynamics: A Panel Study of the Tanzanian Labour Market.” Working Paper 2008–09, Centre for the Study of African Economies, University of Oxford, Oxford, UK. Rankin, N., J. Sandefur, and F. Teal. 2007. “Learning and Earning in Africa: Why It Pays to Go to School.” Mimeo, CSAE, Department of Economics, University of Oxford, Oxford, UK, June. Söderbom, M., F. Teal, A. Wambugu, and G. Kahyarara. 2006. “The Dynamics of Returns to Education in Kenyan and Tanzanian Manufacturing.” Oxford Bulletin of Economics and Statistics 68 (3): 261–88. Stewart, M., and J. Swaffield. 1999. “Low Pay Dynamics and Transition Probabilities.” Economica 66: 23–42. Train, K. 2003. Discrete Choice Methods with Simulation. Cambridge, UK: Cambridge University Press. Wooldridge, J. 2007. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 Environmental Benefits Statement The World Bank is committed to reducing its environmental footprint. In support of this commitment, the Publishing and Knowledge Division leverages electronic publishing options and print-on-demand technology, which is located in regional hubs worldwide. Together, these initiatives enable print runs to be lowered and shipping distances decreased, resulting in reduced paper consumption, chemical use, greenhouse gas emissions, and waste. The Publishing and Knowledge Division follows the recommended standards for paper use set by the Green Press Initiative. Whenever possible, books are printed on 50 percent to 100 percent postconsumer recycled paper, and at least 50 percent of the fiber in our book paper is either unbleached or bleached using Totally Chlorine Free (TCF), Processed Chlorine Free (PCF), or Enhanced Elemental Chlorine Free (EECF) processes. More information about the Bank’s environmental philosophy can be found at http://crinfo.worldbank.org/wbcrinfo/node/4. Working toward Better Pay  •  http://dx.doi.org/10.1596/978-1-4648-0207-2 I mproving the returns to labor for low-paid workers is a key policy challenge, especially in low- income countries (LICs) where increases in earnings are the single most important source of poverty reduction and an important engine of shared prosperity. Yet, the understanding of individual earnings dynamics remains limited. The small, but growing, body of empirical literature on the factors leading to larger and faster pay increases points to strong persistence in earnings over time. However, it remains unclear to what extent this is due to differences in individual endowments rather than to the fact that being in low-paying jobs itself undermines future earnings prospects, and to what extent determinants of earnings vary across types of activities and sectors. The knowledge gap is particularly large for LICs due to the limited availability of reliable panel data. Working toward Better Pay uses unusually rich longitudinal data from Ghana and Tanzania to identify engines of, and barriers to, earnings and earnings mobility. It examines the relative role of individual endowments such as gender, age, and skills, as well as characteristics of the job, and also focuses on the role of job switches, for example, moves into and out of self-employment. The analysis also zooms in on the drivers of transitions between low-paying and high-paying jobs and addresses questions such as whether being low paid is a transitory or permanent phenomenon, and whether it has a scarring effect on an individual’s employment prospects. The extent to which earnings dynamics differ for women and young adults is also discussed in detail. Ghana and Tanzania provide a particularly relevant context in which to examine these issues, and the cross- country comparison helps shed light on the institutional factors that promote labor market mobility and entrepreneurship. The audience for Working toward Better Pay is broad: it is an important read for policy makers, academics, and development practitioners interested in reducing poverty and promoting shared prosperity in Ghana and Tanzania. However, its relevance spans well beyond these two countries to include all developing countries where self-employment in small-scale activities accounts for a very large proportion of employment. World Bank Studies are available individually or on standing order. The World Bank Studies series is also available online through the Open Knowledge Repository (https://openknowledge.worldbank. org/) and the World Bank e-Library (www.worldbank.org/elibrary). ISBN 978-1-4648-0207-2 SKU 210207