ESSAYS IN LABOUR ECONOMICS - Summit

ESSAYS IN LABOUR ECONOMICS

by

Michele Battisti

B.Sc., University of Trento, Italy, 2005

M.Sc., University of York, UK, 2007

Extended Essays submitted in partial fulfillment

of the requirements for the degree of

Doctor of Philosophy

in the

Department of Economics

Faculty of Arts and Social Sciences

c© Michele Battisti 2012

SIMON FRASER UNIVERSITY

Summer 2012

All rights reserved.

However, in accordance with the Copyright Act of Canada, this work may be

reproduced without authorization under the conditions for “Fair Dealing.”

Therefore, limited reproduction of this work for the purposes of private study,

research, criticism, review and news reporting is likely to be in accordance

with the law, particularly if cited appropriately.

APPROVAL

Name: Michele Battisti

Degree: Doctor of Philosophy

Title of Thesis: Essays in Labour Economics

Examining Committee: Dr. Andrew D. McGee

Chair

Dr. Simon D. Woodcock, Senior Supervisor,

Associate Professor of Economics

Simon Fraser University

Dr. Jane Friesen, Supervisor,



Dr. Brian Krauth, Supervisor,



Dr. Christopher D. Zatzick, Internal Examiner,

Associate Professor of Business Administration


Dr. Nicole M. Fortin, External Examiner,

Professor of Economics

University of British Columbia

Date Approved: May 16, 2012

ii

Partial Copyright Licence

Ethics Statement

The author, whose name appears on the title page of this work, has obtained, for the research described in this work, either:

a. human research ethics approval from the Simon Fraser University Office of Research Ethics,

or

b. advance approval of the animal care protocol from the University Animal Care Committee of Simon Fraser University;

or has conducted the research

c. as a co-investigator, collaborator or research assistant in a research project approved in advance,

or

d. as a member of a course approved in advance for minimal risk human research, by the Office of Research Ethics.

A copy of the approval letter has been filed at the Theses Office of the University Library at the time of submission of this thesis or project.

The original application for approval and letter of approval are filed with the relevant offices. Inquiries may be directed to those authorities.

Simon Fraser University Library Burnaby, British Columbia, Canada

update Spring 2010

Abstract

My thesis focuses on the determinants of wage differentials. I analyse some of the immediate

causes for wage differentials including different types of experience and coworker characteris-

tics. In order to understand poverty and intergenerational wage inequality for marginalized

groups we need to look at academic achievement in early schooling, I also investigate how

reading achievement of non-affluent groups may be improved. The first chapter focuses

on the effect of industry experience on wages. I estimate a simultaneous equation model

using a large panel of Italian workers for the years 1986-2004. I find that wage returns to

industry experience are much higher than wage returns to job seniority, and that returns to

general labour market experience dominate the effects of both industry experience and job

tenure. The second chapter investigates the effect of coworker characteristics on wages. I

measure coworker characteristics by the average labor market value of coworkers’ observed

and unobserved characteristics The effect of interest is identified from within-firm changes

in workforce composition, controlling for person effects, firm effects, and sector-specific time

trends. My estimates are based on a very large linked employer employee dataset of workers

and firms from the Italian region of Veneto for the years 1982-2001. I find that a 10-percent

increase in the average labour market value of coworker skills is associated with a 3.6 percent

wage premium. Between 10 and 15 percent of the immigrant wage gap can be explained by

differences in coworker characteristics. The last chapter investigates the effects of providing

school districts with supplemental funding to support the language development of students

who speak a non-standard English dialect. Exploiting the staggered uptake of this funding

across school districts in British Columbia we find that the policy substantially improved

the reading scores of Aboriginal students between grade 4 and grade 7.

Keywords: Wage Growth; Industry Experience; Coworker effects; Linked Employer-

Employee Dataset; Skill Segregation; Peer Effects; Aboriginal Education

iii

Acknowledgments

First I would like to thank Simon Woodcock, Jane Friesen and Brian Krauth for their

excellent supervision.

For the first chapter of this paper, I would also like to thank Benoit Dostie and Jennifer

Hunt for helpful comments and suggestions on earlier versions of this work. I would also

like to acknowledge the participants at the 2009 CAED Conference (Tokyo, Japan), at the

2009 CEA conference (Toronto, Canada), at the 2009 IZA Summer School (Buch am Am-

mersee, Germany), at the 2011 AIEL conference (Milan, Italy) and 2011 EALE conference

(Paphos, Cyprus) for comments and discussions on earlier versions of my work. Researchers

at the LABORatorio R. Revelli (Collegio Carlo Alberto, Moncalieri, Italy) and in particular

Claudia Villosio and Roberto Quaranta have been of great help during my time there as a

visiting researcher.

Peter Arcidiacono and Joshua Kinsler have been extremely generous and helpful sharing

material from their own work and answering my questions, related to the second chapter

of this thesis. A special thanks to Krishna Pendakur for access to computing resources,

to Giuseppe Tattara for making the dataset available to me and to Elisabetta Trevisan for

precious assistance. I would also like to thank the participants of the 2011 SEAe Conference

(Malaga, Spain) and of the 2012 RES PhD Conference (London, UK).

The data used in the third chapter of this thesis were assembled by Maria Trache at

Edudata from administrative records provided by the B.C. Ministry of Education. Funding

by SFU Community Trust Endowment Fund, by CLSRN, HRSDC, and by the B.C. Ministry

of Labour and Citizens’ Services is gratefully acknowledged.

Working together with my classmates Benjamin Cerf Harris, Mohsen Javdani and Pierre

Ebariste Nguimkeu has been of invaluable support and great pleasure.

iv

Contents

Approval ii

Abstract iii

Acknowledgments iv

Contents v

List of Tables ix

List of Figures xi

1 Individual Wage Growth and Industry Experience 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 Estimating Returns to Experience and Seniority . . . . . . . . . . . . 4

1.2.2 My Empirical Specification . . . . . . . . . . . . . . . . . . . . . . . . 6

1.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.3.1 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.3.2 The WHIP Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.3.3 Sample Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.4 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

v

1.4.1 Wage Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.4.2 Hazard Kernel Estimates of Firm Tenure . . . . . . . . . . . . . . . . 14

1.5 Regression Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.5.1 Males . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.5.2 Females . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.7 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.7.1 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.7.2 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2 High Wage Workers and High Wage Peers 36

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.2 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.3 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.4 Data and Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . 45

2.5 Sample Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2.6 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.7 Regression Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.8 Post-estimation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

2.8.1 Person Fixed Effects and Coworker “Quality” . . . . . . . . . . . . . . 50

2.8.2 Variance Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.8.3 Fixed Effects across Specific Groups . . . . . . . . . . . . . . . . . . . 53

2.8.4 Gender Wage Gap and Peer ‘Quality’ . . . . . . . . . . . . . . . . . . 53

2.8.5 The Immigrant Wage Gap . . . . . . . . . . . . . . . . . . . . . . . . . 54

2.9 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.10 Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3 The ESD Policy for Aboriginal Students in B.C. 64

vi

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.2 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.3 Institutions and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.3.1 Organisation and Funding of the B.C. Education System . . . . . . . 67

3.3.2 English as a Second Dialect policy in B.C. . . . . . . . . . . . . . . . . 67

3.3.3 The Foundation Skills Assessment Exams . . . . . . . . . . . . . . . . 68

3.3.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.3.5 ESL Funding and Student Achievement in B.C. public schools . . . . 69

3.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.4.1 Our Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.4.2 Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3.5.1 Test Participation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3.5.2 Test Score Gains: Main Results . . . . . . . . . . . . . . . . . . . . . . 77

3.5.3 Test Score Gains: Quantile Regression Results . . . . . . . . . . . . . 79

3.5.4 Test Score Gains: non-Aboriginal Students . . . . . . . . . . . . . . . 79

3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.7 Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.7.1 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.7.2 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Bibliography 88

Appendix A Tables 97

A.1 Distribution of Firms by Sector in WHIP . . . . . . . . . . . . . . . . . . . . 97

A.2 Distribution of Firms by Sector in VWH . . . . . . . . . . . . . . . . . . . . . 98

A.3 First Stage regression and Exam Participation of Aboriginal Students . . . . 99

vii

Appendix B Technical Appendices 101

B.1 Iterative procedure for estimating spillover effects . . . . . . . . . . . . . . . . 101

B.2 Structure of the VWH dataset . . . . . . . . . . . . . . . . . . . . . . . . . . 103

B.3 Sample Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B.4 Robustness check: small firms and large firms . . . . . . . . . . . . . . . . . . 105

viii

List of Tables

1.1 Number of Jobs and Conditional Average Duration . . . . . . . . . . . . . . . 23

1.2 Number of Sectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.3 Summary statistics for Job Covariates - Males . . . . . . . . . . . . . . . . . 24

1.4 Summary statistics for Job Covariates - Females . . . . . . . . . . . . . . . . 24

1.5 Summary statistics for Year-level Covariates - Males . . . . . . . . . . . . . . 24

1.6 Summary statistics for Year-level Covariates - Females . . . . . . . . . . . . 24

1.7 Wage Equation for Males . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

1.8 Job Hazard Equation for Males . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.9 Sector Hazard Equation for Males . . . . . . . . . . . . . . . . . . . . . . . . 27

1.10 Variance components and Parameters for Males . . . . . . . . . . . . . . . . . 28

1.11 Wage Equation for Females . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.12 Job Hazard Equation for Females . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.13 Sector Hazard Equation for Females . . . . . . . . . . . . . . . . . . . . . . . 31

1.14 Variance components and Parameters for Females . . . . . . . . . . . . . . . . 32

2.1 Main regression results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

2.2 The contribution of gender and immigration status to the person effect . . . . 61

2.3 Standardised wage, θ and ψ gaps for different groups . . . . . . . . . . . . . . 61

2.4 Gender and quality of peers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2.5 Birth place and quality of peers . . . . . . . . . . . . . . . . . . . . . . . . . . 63

ix

3.1 Per student operating grants to B.C. public school districts. . . . . . . . . . . 81

3.2 Participation in FSA exams, grade 7 students 2002-2004. . . . . . . . . . . . . 81

3.3 Achievement levels and growth, grade 7 students 2002-2004. . . . . . . . . . . 82

3.4 Participation in FSA exams, grade 7 students 2002-2004. . . . . . . . . . . . . 82

3.5 District ESD programs and students in ESD, grade 7 Aboriginal students

2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.6 Effect of ESD programming on grade 7 exam participation, grade 7 Aboriginal

students 2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.7 Effect of ESD programming on exam results, grade 7 Aboriginal students

2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

3.8 Quantile regressions for reduced form effect of ESD programming on exam

results, grade 7 Aboriginal students 2002-2004. . . . . . . . . . . . . . . . . . 84

3.9 Reduced form effect of ESD programming on exam results, grade 7 non-

Aboriginal students 2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . 85

A.1 Economic sector of the firm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

A.2 Firm sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

A.3 First stage regression results. . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

A.4 ESD and probability of taking the FSA exam, grade 7 Aboriginal students

2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

B.1 Regression on different samples . . . . . . . . . . . . . . . . . . . . . . . . . . 106

x

List of Figures

1.1 Experience Profile based on annual data . . . . . . . . . . . . . . . . . . . . . 33

1.2 Industry Experience Profile based on annual data . . . . . . . . . . . . . . . . 34

1.3 Job Tenure Profile based on annual data . . . . . . . . . . . . . . . . . . . . . 34

1.4 Survivorship Function for Job Tenure . . . . . . . . . . . . . . . . . . . . . . 35

1.5 Hazard Function for Job Tenure . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.1 Average monthly wages by gender and proportion of females . . . . . . . . . 56

2.2 Average monthly wages by foreign born status and proportion of foreign born 57

2.3 Standard deviation of log monthly wages over time . . . . . . . . . . . . . . . 57

2.4 Average proportion of peers that are foreign born . . . . . . . . . . . . . . . . 58

2.5 Correlation between person effects and firm effects over time . . . . . . . . . 58

2.6 Decomposition of the gap between the wage of foreign born and Italian born 59

3.1 Percentage of grade 7 Aboriginal students in ESD (SESD stands for “Stan-

dard English as a Second Dialect”, equivalent to ESD) in first year that

district assigns more than 5% to ESD (and more than 10 grade 7 students in

total), 2002-2004. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

3.2 Percentage of grade 7 Aboriginal students in ESD 1999-2004, selected districts. 87

xi

Chapter 1

Individual Wage Growth and

Industry Experience

1.1 Introduction

There is solid evidence of a positive correlation between wages and tenure, both at the

firm and industry level. This finding would seem to suggest that staying in the same firm

for the whole career would maximize wage growth. At the same time there is extensive

evidence showing that workers who earn high wages are relatively mobile. This prima

facie contradicting result suggests that both human capital accumulation and match quality

considerations are important in the wage determination process: workers learn useful skills

as they keep their jobs longer but are also likely to find a better match when they move

more. Estimating returns to firm tenure is complex because treating all job-to-job transitions

equivalently may be misleading, and this is due to the fact that the degree to which human

capital can be transferred from firm to firm is highly heterogeneous. Yet quantifying the

role of firm tenure and industry1 experience for wage growth is critical for researchers and

policy makers alike. For researchers it sheds light on the role of different types of human

capital for mobility and wages. For policy-makers knowing the “rewards” of tenure and

industry experience is valuable for labour market mobility, contracts and compensation

policies. If industry experience is important for individual wage growth then public policy

1“Industry” and “sector” are used interchangeably.

1

CHAPTER 1. INDIVIDUAL WAGE GROWTH AND INDUSTRY EXPERIENCE 2

might consider specific interventions for displaced workers employed in a shrinking sector

as opposed to displaced workers in a booming sector.

I present novel evidence of the wage returns to industry experience by developing and

estimating a model where wages, industry experience and job duration are simultaneously

determined. In particular, I estimate a 3-equation model where I allow worker and job

characteristics that affect wages to also play a role in match duration and sector experience.

To estimate this model I use administrative data for a large sample of young Italian workers

from the Worker Histories Italian Panel for 1986-2004. This paper offers the first estimate

of the returns to industry experience using Italian panel data. To the best of my knowledge it

is also the first simultaneous estimation of wage growth, firm tenure and industry experience.

The main results show that wage returns to industry experience are much higher than wage

returns to job seniority: on average mobility across sectors is associated with a higher short-

term wage penalty than mobility within the same economic sector. In addition, In find that

returns to general work experience dominate the effects of industry experience and of job

tenure. For males, returns to industry tenure are around four times higher than returns to

job seniority for the first years on the job. Returns to labour market experience are four

percent per year for the first five years.

My model also allows to test for the endogeneity of tenure in the wage equation. I find

evidence that wages and employment duration are simultaneously determined. In particular,

individuals with characteristics that are associated with higher wages also stay on the job

longer, and “good” matches (matched with higher wages, conditional on worker and job

characteristics) are less likely to be destroyed. The hypotheses of no simultaneity of seniority

and industry experience across equations are rejected at any conventional significance level.

The concern that firm tenure may be endogenous in the wage equation, i.e. the possibility

of unobservables affecting both wages and employment duration, has been considered since

Abraham and Farber (1987). They find positive returns to seniority to be an artifact of

sample selection in the sense that matches that pay conditionally higher wages from the

start are more likely to survive. As a consequence, a simple Ordinary Least Squares (OLS)

regression of wages on tenure will yield biased estimates. Altonji and Shakotko (1987)

propose an Instrumental Variable (IV) technique to account for this endogeneity.2 Neal

2Among many others, Topel (1991) offers related evidence of the importance of firm-specific human capitalwhile using longitudinal datasets to account for the endogeneity problem discussed above. They find that


(1995) first argues that industry experience might also be an important determinant of

wage growth. Both papers find that industry tenure is a key determinant of wages. Neal

(1995) uses a survey of displaced workers for which industry experience can be taken as

exogenous at the time of displacement.

More recently Dustmann and Meghir (2005) employ a similar strategy to study the wage

impacts of different sources of human capital using data from Germany, using age as an

instrument for experience. Dustmann and Meghir (2005) find that returns to sector tenure

are positive for skilled workers, but are not significantly different from zero for unskilled

workers. On the one hand, studies that use displaced workers allow us to isolate involuntary

match destructions and thus offer reliable estimates. On the other hand, their sample

cannot be considered representative of the working population as a whole, which impacts

the generalisability of their results.

A more direct focus on industry experience emerged after Parent (2000), who uses the

same IV technique as Altonji and Shakotko (1987) to investigate the role of sector tenure

on wages. More recently Cingano (2003) uses a similar set of instruments and an equivalent

control function approach as Dustmann and Meghir (2005), and data from two Italian

provinces to estimate the effects of experience in a certain industrial district on wages. He

finds negative and insignificant effects, concluding that district-specific skills do not seem

to matter for wage growth even among very similar firms.3

Estimating a random-effect simultaneous equation model for wage growth and firm

tenure is possible under a weaker set of assumptions compared to equivalent IV estimates,

and also allows us to estimate the extent of the correlation between wages and job duration,

rather than simply allowing for that correlation to exist. Lillard (1999) then uses U.S. data

to estimate a simultaneous model with a wage equation and an job duration equation. He

finds evidence suggesting that tenure might be endogenous in the wage equation, and that

lower bounds for wage returns of firm seniority are around 2.5 percent a year on average. Topel and Ward(1992) on the other hand stress the importance of job mobility as a source of wage growth for young Americanmales.

3Kambourov and Manovskii (2009) include the role of occupations as well as industries and firms forhuman capital accumulation of workers using the PSID for 1968-1980. Because of data limitations, I am notable to control for occupations and so I focus on sectors only, and so returns to tenure and sector tenure maybe viewed as a combination of the effects of firms, sectors and occupations. While occupations are certainlyrelevant in this context and are typically found to be important, mobility across sectors is more objectivelymeasurable and thus arguably more directly relevant for policies around structural changes in the labourmarket.


this endogeneity might be driven by unobservables. Dostie (2005) using French data finds

that wages and job tenure are simultaneously determined. Both Lillard (1999) and Dostie

(2005) focus on the returns to firm tenure and do not include industry experience in their

analysis. This paper develops a three-equation model that builds upon Lillard (1999) for

estimation strategy and identification, while focusing primarily on wage and mobility effects

of industry experience.

1.2 Empirical Model

1.2.1 Estimating Returns to Experience and Seniority

In order to illustrate the possible issues with estimating returns to experience using obser-

vational data, it may be useful to describe the ideal thought experiment we would run to

identify such effects. In this thought experiment there is a fixed set of workers and a fixed set

of jobs. Workers are randomly assigned to jobs. In order to separately identify the effects of

labour market experience, job tenure and sector tenure, we also need workers to move from

job to job. Therefore, in the thought experiment the matches that we created randomly are

also terminated in a random fashion. Under these conditions, a simple OLS regression of

wage on experience, industry-experience and seniority measures the causal effect of tenure

and experience on wages.

Such experiments cannot be run in reality, and so we need to use observational data,

where randomisations as those described above do not take place. Both sides of the market

are actively seeking their best opportunities (as in Jovanovic 1979): workers do not choose

their job randomly and firms do not hire workers randomly. Over time, the set of jobs that

survives is self-selected. For example, workers might keep searching for new opportunities

while employed as in Pissarides (1994), and quit their current job if they receive a sufficiently

attractive offer. Therefore, higher wages may increase the probability of staying in the

current job. Moreover, both firms and workers can decide to interrupt a match and will

not do so randomly. For example, either side of the market may learn about productivity

over time as in Postel-Vinay and Robin (2002), or characteristics of the worker may be

observable but not contractible as in Peters (2010). In any of these cases there are self-

selection processes at work. My empirical model estimates the effect of firm tenure, industry

experience and labour market experience on wages under a much weaker set of conditions


than those described above for a simple OLS procedure.4

Let i = {1, ..., N} identify a worker, and t = {1, ..., T} a time period. Let J(i, t) be the

firm worker i is employed by, in period t. In the following, J(i, t) ≡ j is used for simplicity.

Although I sometimes refer to the match j as a “job”, it is intended simply as the match of

one firm and one worker. A “job” (as I define it) lasts as long as the time at which firm and

worker separates. Because of limited information on tasks inside the firm, I chose to consider

all occupational changed inside the firm as part of the same “match” in order to account

for the fact that changes in contract arrangement within the firm are a manifestation of the

returns to firm tenure, and are used to reward tenure especially in contexts where there are

frictions on wage flexibility.

Equivalently K(J(i, t)) denotes the sector of worker i in period t, and K(J(i, t)) ≡ k. A

useful starting point for illustrative purposes is a linear wage model such as:

wijt = γ1(seniorityijt) + γ2(sectorseniorityikt) + γ3(experienceit) + εijt (1.1)

where wijt is the real wage of worker i in match j in period t. The variable seniorityijt

denotes the duration of the match j up to period t. The covariate sectorseniorityikt denotes

the experience accumulated by worker i in sector k up to period t, and experienceit is the

total labour market experience of worker i up to time t.5

The error term εijt can be decomposed into a firm-specific time-invariant component

θi, capturing the effect of person-specific time-invariant characteristics, a match-specific

component δij (which captures a firm and sector effect as well) and a component that is

match-, time- and person-specific, denoted by νijt below:

εijt = θi + δij + νijt (1.2)

An OLS procedure yields unbiased estimates of γ1, γ2 and γ3 in equation (1.1) only if

experience, industry experience and job seniority are uncorrelated with θi, with δij and with

νijt. In other words, OLS estimates are biased unless workers were randomly assigned to

4For a comprehensive discussion on how to treat unobserved worker, firm, and match heterogeneity in amodel of wage differentials see Woodcock (2008).

5Because firms do not change sectors in my dataset, I drop the subscript k when redundant in equation(1.1) and in other equations below. For example, I write εijt instead of εijkt.


sectors and firms, and matches were randomly destroyed. My empirical strategy on the other

hand allows individual level unobservables to affect wages, job duration and sector duration

at the same time. It also allows job-level unobservables to affect length and profitability of

the job.

1.2.2 My Empirical Specification

My empirical model is composed of a wage equation, a job tenure hazard equation and an

industry hazard equation. Below, I describe them separately and then discuss how they are

linked together.

Wage Equation

I specify the wage equation as follows

ln(wijt) = α0 + α′1seniorityijt + α′2sectorseniorityikt (1.3)

+(1 + θ1i)α′3experienceit +

T∑t=2

ιwt yeart

+T∑t=2

κwt sectort + θ2i + δij + νijt

where wijt is the real wage of person i at time t.6

The regressors seniorityijt, sectorseniorityikt and experienceit are parameterised

as piecewise-linear splines, where nodes are chosen to provide the best fit to the data.

Piecewise-linear splines allows more flexibility in the effects and better fit to the data com-

pared to simple polynomials in this case.

In equation (1.3) θ1i and θ2i are random person effects with zero conditional mean; yeart

denotes a dummy variable for year t. I include year fixed effects so that the estimates are

not driven by time trends in wages and mobility. The variable sectort is a dummy for each

industry. Sector fixed effects ensure that estimates are not driven by correlation between

industry experience and unobserved sector characteristics; δij captures match-specific unob-

served heterogeneity. The α’s, ι’s and κ’s above are parameters to be estimated. Finally, νijt

6Unless stated otherwise, I use the same notation as above.


is the person-match-time specific error term, which is assumed to have mean zero conditional

on all the other regressors.

Job Duration Hazard Model

Employment duration is estimated using a hazard model based on Kiefer (1988). The

baseline hazard duration dependence is piecewise linear (piecewise Gompertz).7 For person

i employed in job j in year t, the hazard model is

ln(hij(τ)) = β0 + β′1seniorityijt + β′2experienceit (1.4)

+

T∑t=2

ιjtyeart +

T∑t=2

κjtsectort + θ3i + φδij

where ln(hij(τ)) is the conditional log hazard, i.e. the probability to observe a job separation

for a match of length τ at time t, conditional on that match being active.8 I can control

for time-invariant personal unobserved characteristics affecting job mobility through the

person effect θ3i. The random match effect δij from equation (1.3) with the load parameter

φ accounts for potential cross-equation correlation between the job-level wage components

and the job-level turnover hazard. The remaining regressors and parameters are defined as

in equation (1.3).

Sector Duration Hazard Model

For person i employed in sector k in year t, the hazard model is

ln(hsik(τ)) = γ0 + γ′1sectorseniorityikt + γ′2experienceit (1.5)

+

T∑t=2

ιstyeart +

T∑t=2

κstsectort + θ4i

where ln(hsik(τ)) is the conditional log hazard, i.e. the probability of employment in sector

k ending at time t, conditional on that sector spell not having been destroyed earlier.

Equivalently as above, θ4i is a person random effect, which is assumed to be perfectly

7See Pollard and Valkovics (1992) and Lillard and Panis (2003b) for additional information on the Gom-pertz distribution.

8In the data, I cannot distinguish between quits and layoffs and so I need to treat them equivalently.


mobile across sectors: workers are not innately more suitable to a specific sector. The

alternative modelling choice would have been perhaps more general, but it would likely have

been infeasible to estimate. In addition, it would have resulted in the need to estimate one

hazard equation separately for each sector and would have made it impossible to compare the

role of unobserved time-invariant ability across equations. Equation (1.5) does not include

a match random effect: I assume that match quality may affect the probability of changing

jobs, but having taken this effect into account, it has no further effect on the probability

of changing sectors. This is equivalent to assuming that workers and firms are aware that

match effects are transitory in nature.9 The other regressors and parameters are equivalent

to equation (1.3).10

Error Structure and Assumptions on Parameters

I assume a first-order autoregressive error in equation (1.3):

νiJ(i,t)t = ω · νiJ(i,t−1)t−1 + uiJ(i,t)t (1.6)

where uiJ(i,t)t ∼ N(0, σ2u). Errors may thus be correlated within a worker’s career, beyond

the correlation induced by the presence of a person effect. I also assume that the random

match effect is normally distributed:

δij ∼ N(0, σ2δ )

Two sets of elements introduce simultaneity in the three-equation model described above.

First, the individual effects are allowed to covary across equations (1.3), (1.4) and (1.5):

(θ1i, θ2i, θ3i, θ4i)′ ∼ N(0,Σθ,θ)

9Note that this assumption is implicitly needed in all models that include match effects.10The two hazard models described above concern the overall probability of job transitions and sector

transitions. Therefore, the job hazard model of equation (1.4) includes both job transitions within a sectorand transitions of job and sector. The sector hazard model of equation (1.5) includes transitions of job andsector (sector-only transitions are not possible). This introduces a positive correlation between the personeffects in the job hazard model θ3i and in the sector hazard model θ4i, since some of the transitions are thesame transitions in both models. The reader should keep this in mind when interpreting the magnitude ofmy estimate of that correlation. In particular, my estimate is larger than what I would find if I estimatedthe job hazard model using transitions within a sector only.


If industry experience and job tenure are exogenous in the wage equation there would be

no cross-equation correlation between the θ•i’s. I allow for time invariant characteristics that

affect wages to influence match duration and sector experience as well. On the other hand,

IV techniques such as in Kambourov and Manovskii (2009) need to identify instruments that

need to be correlated to all fixed effects, and therefore need to make stronger assumptions

concerning the form that the endogeneity can take, as discussed in Pavan (2011).

Secondly, the load factor φ in equation (1.4) measures the effect of δij (defined in equation

1.3 and estimated in the wage model) on the proportion of match destruction. A significant

estimate for φ suggests that unobserved factors varying at the firm or at the job level

systematically affect wages and employment duration. The hypotheses on the θ•i’s and on

φ are tested separately and jointly using a Likelihood Ratio test.

Random effects estimations do not require restrictions on the correlation of person and

match effects across equations. This in turns makes it possible to evaluate the magnitude of

the endogeneity by looking at the correlations of random effects within and across equations.

All parameters of interest are identified off a mix of within person and match variation (i.e.

variation that comes from observing multiple matches for each worker and multiple wage

observations for each match) and between person and match variation (i.e. variation that

comes from the fact that I observe multiple worker for each year and for each sector).

Therefore, technically all effects would be identified off functional form even if I did not

observe multiple jobs in multiple sectors, and many wage observations for each worker’s

match.

However, in my dataset I observe multiple jobs per person and yearly wage observations

for each job, so that I am not relying on functional form alone for identification. I can use

wage variation within a job as a source of identification of the effects of job seniority on wages,

and wage variation within a person’s career across sectors helps me to separately identify the

wage effects of industry experience and of labour-market experience. The effects of labour-

market experience are in turn identified off in part using workers for whom I observe more

than one employer. A more comprehensive discussion of identification of random effects

models is available in in Greene (2003, pages 295-298) and Lillard (1999).

For equations (1.4) and (1.5) the individual component is identified in part using multiple

spells for each worker before the last spell observed, which may be right-censored because

the most recent match might be alive at the end of the dataset. Since workers may or


may not change sector when they change jobs, I can identify all parameters in the sector

hazard equation separately from those of the job tenure equation. The variance of the

person-specific heterogeneity term can be identified because we observe multiple jobs for

each worker.

1.3 Data

1.3.1 Institutional Background

The empirical investigation in this paper uses a long panel of Italian workers. Italy is often

portrayed as a country where collective bargaining is the main mechanism for wage determi-

nation. The reality is more complex: there are many potential sources of wage differentials.

National regulations concern general issues common to all sectors and all firms, and are

typically silent on specific compensation levels. Contracts signed with trade unions are

usually at the industry level, and specify non-binding minimum wage levels, representing

an industry-specific floor for total compensation. In addition, because minimum wages are

occupation and rank-specific, promotions can affect the relevance of the contractual mini-

mum wages (Cingano, 2003). At the firm level, both firm-level agreements and individual

bargaining are present, and wage premia are found to be highly heterogeneous across firms

(Erickson and Ichino, 1993). An extensive description of the institutional features of the

Italian labour market is beyond the scope of this paper. Addessi and Tilli (2009), Beccarini

(2009) and Schindler (2009) offer a more comprehensive analysis.

After a long period of high unemployment despite positive economic growth in the 1980s,

in the 1990s Italy experienced an increase in labour force participation and and a fall in

the unemployment rate. This can be traced back to the consistent growth of temporary

and part-time employment, especially for young workers. Increased flexibility has been

introduced “at the margin” through a series of reforms that affected primarily new entrants

in the labour force Schindler (2009). The empirical analysis below is based on a younger-

than-average segment of the working population, who face a labour market that is more

flexible in terms of wages and job security, and where short-term contracts are increasingly

common.


1.3.2 The WHIP Dataset

I estimate the simultaneous equation model above using the Work Histories Italian Panel

(WHIP). WHIP is a database of individual work histories for the years 1985-2004, based

on administrative archives from the Istituto Nazionale della Previdenza Sociale11 (INPS),

which is the main institution for social security in Italy.12 By law, all employees in the

private sector, some categories of employees of the public sector and most self employed

need to be enrolled in INPS, with the exception of certain categories of professionals, such

as doctors, lawyers, notaries and journalists, who have alternative social security funds.

The reference population of WHIP consists of all individuals who worked in Italy in any

of the years of the panel. From this population, the WHIP sample is constructed using

four birth dates for each year, so that the sampling ratio is around 1:90. This results in a

dynamic population of about 370,000 people. WHIP includes information about the main

episodes of the working careers of people in the sample, such as dates of match creation

and match destruction for each employment spell, wage received, non-working spells, other

benefits received by the employee such as unemployment and mobility benefits, special

arrangements, occupation13, location. Individual data also include gender, year and region

of birth, amount of unemployment compensations, maternity leave compensation and other

social assistance programs. Being an administrative registry of employment relations, the

WHIP dataset does not include educational attainments of workers. All jobs are identified by

a unique job identifier.14 This paper uses employees of the private sector only, for which the

database also provides some information about employers such as firm size, region, sector15

where each worker is employed. This version of my work does not model unemployment

specifically. This is equivalent to assuming that unemployment has no effect on the set of

11National Institute for Social Security.12WHIP is managed by Laboratorio Revelli Centre for Employment Studies - that has been

constructed thanks to an agreement between the INPS and the University of Torino. Seehttp://www.laboratoriorevelli.it/whip. Detailed descriptions of the WHIP dataset are available from Contini(2002) and Contini and Trivellato (2005).

13However, only five different occupations are possible, and so are of limited usefulness and capture con-tractual pay scale as opposed to occupation in terms of a set of tasks. I am not using this information inthis version of the paper. Including occupational dummies does not substantially affect the results.

14For legal reasons firm identifiers are not included in the dataset, and thus it is not possible to identifyworkers that share the same employer.

15The classification used for this version of the dataset includes 34 sectors and it is based upon the onedigit Ateco91 system.


skills of workers: I do not investigate the possibility that unemployed workers might acquire

labour market skills, and also that their skills deteriorate.

1.3.3 Sample Restrictions

Industry experience, labour market experience and job tenure are left-censored for all indi-

viduals in WHIP because no information is available on employment spells before 1985. I

restrict the sample to workers for whom I can assume that I can observe their whole careers:

I drop all individuals that are employed in the first year of the panel, 1985, and then I

restrict the sample to individuals that are born in 1961 or later. All of the results below

are based on a population of workers that are on average younger than the overall Italian

labour force: the oldest worker in the regression sample is 25 years of age is 1986, and thus

is 43 years of age in 2004, the last year of the panel.

The final sample consists of 82,114 male workers and 56,914 female workers. The total

number of job spells is 207,501 for males, of which 20.5 percent are right-censored,16 and

134,941 for female, of which 21.1 percent are right-censored. It includes 536,277 yearly

wage observations for male workers, and 358,591 for female workers. Wage measures are

converted into year-2004 Euros by using aggregate data of the Consumer Price Index from

Istat (2009). To make jobs of different lengths comparables in terms of wages, I construct

annual Full Time Equivalent wages for all workers.17

As mentioned above, WHIP includes information about start date and end date of each

job but wages are recorded only once a year. I identify a a dominant job for every worker

and every year18 to avoid overweighting observations for short employment spells and to

avoid imputing wage patterns within a year. Therefore, the dataset I use for all regressions

has at most one observation for each worker for each year.

16I.e. 20.5 percent of spells are still active in the last year of my dataset.17I divide total wages by the number of days worked and then multiply the result by 312, the total number

of days of full time workers in one year.18I eliminate all jobs with less than five full-time equivalent working days, then I rank jobs by number of

effective full-time-equivalent days and then by duration and wages.


1.4 Summary Statistics

In my sample, 61 percent of the workers are male and 39 percent are female. Around 90

percent of the workers are in a full-time job. For males, Construction is the largest sector

(18.2 percent of workers), followed by Wholesale and Retail Trade (13.8 percent) and by

Banking and financial intermediaries (10 percent).19 Comparing the distribution of workers

across sectors in my regression sample with that of the 2001 Italian Population Census Istat

(2005) we note that construction, wholesale and retail trade are overrepresented in my

sample, while banking and other services are slightly underrepresented. The discrepancies

are likely to be due to the fact that workers in my sample are much younger than workers in

the population. In my sample, females are most likely to be employed in the Wholesale and

Retail Trade sector (19.9 percent), and in the Banking sector (16.0 percent).20 Comparing

the distribution of workers across sectors with that of the 2001 Italian Census for females, I

find that Hotels and restaurants are overrepresented in my sample, while industry in general

is slightly underrepresented.

Given that the identification of person and job random effects is based upon workers

moving between firms, it is crucial to investigate how much mobility is observed in the data.

Table 1.1 shows that we observe one employment spell for 37.6 percent of male workers in

the sample, two spells for 24.3 percent, three spells for 15.4 percent of the sample. Therefore,

I observe more than 60 percent of workers in my sample working for more than one firm.

Female workers are slightly less mobile than male workers in our data. However, I still find

that around 60 percent of them change employer. Overall, there is a substantial amount of

mobility, which is reassuring for the solidity of my estimation strategy. Figure 1.2 shows

that we observe around 44 percent of males in more than one sector, and around 17 percent

in at least three sectors. The corresponding figure for females are only slightly smaller.

Table 1.3 shows that employment spells of males last on average just over two years.

In the sample used here, about 20 percent of the spells are right-censored. Male workers

enter employment spells with about 20 months of experience on average, and with around 8

months of experience in the same sector. Equivalent figures for female workers are presented

19Table A.1 shows the distribution of workers by sector and gender. Additional information on the codifi-cation of sectors in WHIP is available in Leombruni and Quaranta (2011).

20Table A.1 is based on workers’ first job only. This is done for clarity, and these distributions are notsensitive to this choice.


in table 1.4. Females stay on the job slightly longer than males, and enter an employment

spell with slightly less experience in the labour market and in the sector.

As shown in Tables 1.5 and 1.6, male and female workers in this sample have an average

gross income of 19,700 Euros and 17,900 Euros respectively. These incomes are calculated on

a full-time full-year equivalent using real wages in 2004 Euros. At the start of the year, male

(female) workers have on average 3.66 years (3.56 years) of experience in the labour market,

2.72 years (2.72 years for females) of experience in the sector, and have accumulated tenure

on the job of 2.02 years (2.02 years for females). Looking at the mean alone is misleading:

around 45 percent of workers are employed in a firm that has less than 10 employees, only

15 percent of workers are employed by firms that have more than 300 employees.

1.4.1 Wage Profiles

Figure 1.1 shows that there is a strong positive correlation between experience and log wages.

The difference in wages between males and females is large and increases with the level of

experience for the first ten years. At the beginning of their careers, males and females have

similar wage levels, but at around ten years of experience males earn around 20 percent

more than females. Women with 15 years of experience have average wages that are similar

to those of men with around half as much labour market experience.

Figure 1.2 presents the unconditional correlation between log wages and experience ac-

cumulated in the same industry. The pattern is similar to Figure 1.1, although the gap

between males and females is even larger here and increasing for all levels of industry ex-

perience. Figure 1.3 shows the equivalent log wage profile for match duration. In this case

all of the gap between males and females is accumulated in the first few years of job tenure,

and it is constant afterwards.

1.4.2 Hazard Kernel Estimates of Firm Tenure

Dropping right-censored spells,21 the median duration of a job is around one year for males

and 1.17 years for females. The 75th percentile is 3.25 years for males, 3.59 for females.

Median tenure in a sector is 1.83 years for males and 2 years for females. The 75th percentiles

21I.e. matches that are still active the the of my panel, such that I do not observe their end date. Censoredspells are 20.7 percent of all spells.


is 5.94 years for males, 6.16 for females.

Figure 1.4 shows that the survival probability of jobs falls very rapidly in the first years of

spell duration and declines very gently afterward. At two years of job tenure, more than half

of the matches have already been destroyed. Around one fourth of the matches observed in

the regression sample last more than four years. These patterns are almost indistinguishable

between men and women. A kernel density estimation22 of the hazard rate constructed in

Figure 1.5 shows the probability of match destruction at each level of tenure, conditional

on that match having survived up to that point in time. The hazard rate is high in the first

few years of a match, starting off at over 0.3 and falling to 0.2 at four years of job tenure.

Afterwards it keeps diminishing falling to 0.1 at 12 years of tenure. While patterns are very

similar between males and females, Figure 1.5 does reveal that matches of female workers

are significantly less likely to be destroyed in the first four years of tenure, consistently with

the mobility patterns described above.

1.5 Regression Results

I estimate the three-equation model discussed above using aML (Applied Maximum Likeli-

hood), a software developed by Lillard and Panis (2003a). Because the likelihood of hazard

models does not have a closed form solution, I approximate the integrals in the likelihood

function using a numerical integration algorithm based upon the Gauss-Hermite Quadrature,

which selects a number of support points (I use 4 points for my main simultaneous equation

model) and weights such that the weighted points approximate a normal distribution.23 As

suggested in Lillard and Panis (2003a) and since the distribution of the individual random

effects is of dimension four, I transform the covariance matrix into Cholesky-decomposed

parameters in order to ensure that the covariance matrix remains positive definite.

Estimates of equations (1.3), (1.4) and (1.5) are presented in three separate tables for

male and female workers separately.24 Unless mentioned otherwise, all coefficients described

22All of these kernel estimations use the Epanechnikov kernel.23See Abramowitz and Stegun (1972) for more details.24There are two reasons for having estimated this model for females and males separately. Firstly, it is

of direct interest to look at the estimates for males and females separately because of the overall differencesbetween labour-market performances and dynamics for the two genders. Secondly, possible dynamic selectioneffects might reduce the generality of results for females. This concern may be especially relevant in the Italian


below are statistically significant at the 1 percent level.

1.5.1 Males

Table 1.7 presents estimates from equation (1.3) for male workers. In the column SIM,25

the first two years of industry experience are associated with an average wage increase

of 2.1 percent per year. The years between the second and the fifth are associated with

slightly negative marginal effect on wages: while some industry experience has positive

returns, workers with an intermediate level of industry experience are not paid more than

workers with less industry experience. It is possible that while highly mobile workers might

be driven by choice and high motivation, intermediate levels of industry experience might

signal a previous layoff. The average marginal effect of industry experience on wages is

small, positive and stable after the fifth year at 0.7 percent a year.

Controlling for industry experience, job tenure has a small effect on wages: the first two

years are associated with an average wage increase of 0.5 percent per year; the equivalent

effect falls to 0.3 percent per year in the following three years, and it is not significantly

different from zero for the years 5th-10th. After the tenth year on the job, the effect is

slightly negative, and significant at the 5 percent level, suggesting that staying on the same

job for very long may be detrimental for wages. The effect of labour market experience on

wages are large and stable across our three specifications. The marginal yearly effect for the

SIM specification is 4.3 percent for the first five years and around 1.5 percent afterwards.

The total effects over the first ten years of a worker’s career are 6.8 percent for industry

experience, 2.4 percent for job tenure and 29 percent for general work experience. If wages

reflect marginal productivity of workers, which in turn is a function of human capital accu-

mulation, these results suggest that general human capital and sector-specific human capital

are both more important that firm-specific human capital. It is of interest to note that the

effects of experience are large even after many years in the labour market.

context where females have among the lowest participation rates in the OECD. Workers with skills that arevalued more in the labour market might be more likely to work, and therefore more likely to be observedin my sample, thus making my estimates of the returns to experience of limited applicability for the overallpopulation.

25Column “SIM” refers to the most general specification: the three- equation simultaneous model whereindividual and match effects are allowed to be correlated across equations. Therefore, for column SIM allthree tables refer to a single estimation.


Column W2 includes controls for unobserved heterogeneity at the individual and match

level but does not allow sector tenure and job tenure to be endogenous. Comparing column

W2 and column SIM allow to look at the impact of endogeneity on my estimates. Failing

to control for endogeneity does not lead to a large overestimation of the effect of tenure on

wages, compared to typical estimates using datasets from the US and from other European

countries (see Kambourov and Manovskii 2009 for a good review of the literature). This

difference may be due to the lower level of mobility in the Italian labour market compared

to the US labour market, which might generate a lower correlation between the firm effect

(which is embedded in the match effect in this paper) and the level of job tenure workers

acquire. In other words, in the US low-wage firms might be more likely to lose workers to

to lay them off, and might not be able to attract high-wage workers. In Italy, it seems that

labour market frictions for firing decision might mitigate these tendencies.

Table 1.8 presents the results for the hazard regression for spell duration. The first two

years of job seniority are associated with a lower probability of match destruction. How-

ever, the estimates are much closer to zero once individual heterogeneity and simultaneity

are introduced, falling from around 32 percent in model J1 to around 5 percent in SIM.

The average worker that is in a longer lasting jobs differs systematically from the average

worker that has shorter employment spells. Therefore, in a model that does not control

for unobserved heterogeneity tenure acts largely as a proxy for worker quality and match

quality.

Focusing on the SIM column, seniority has a negative impact on the probability of match

destruction; the marginal effect is strongest for the years second to fifth. It is informative to

interpret these estimates within the context of a framework where where it might take time

for employers and employees learn about match-specific productivity as in Jovanovic (1979)

and Jovanovic (1984). The longer a match survives the more likely it is that it survives

further. The years between the second and the gift have the largest effect, which seems to

suggest that there may be substantial learning in that range of spell duration.

Estimates for the effect of labour market experience on the employment hazard rate

for SIM show that each of the first five years in a worker’s career is associated with a 4.4

percentage-point lower log hazard rate. The following five years on the other hand are

associated with a rise in the exit rate. Workers have the highest probability of leaving their

job very early in their careers or after their first five years. While the former might be


driven by lower-quality matches for inexperienced workers, the latter may be related to the

fact that workers with more than five years of experience are in a better bargaining position

with a new employer. Their better outside option might in turn increase their exit rates.

The estimates for the sector seniority hazard model (equation 1.5) for male workers are

outlined in Table 1.9. In SIM, the effect of industry experience on the conditional probabil-

ity of leaving a sector is negative and large for the first two years and positive and smaller

afterwards. If it takes time for the agents involved to learn the relevant productivity param-

eters, then lower levels of industry experience are associated with a lower exit probability,

while as industry experience gets higher it is associated with a higher exit probability, even

higher than the initial exit rate after around eight years of sector tenure. Similar patterns

can be observed for labour market experience: ceteris paribus, more labour market expe-

rience increases the conditional probability of leaving a certain sector. Workers with the

same experience in one sector but more labour market experience are more mobile. This is

not surprising given that opportunities in other sectors might increase with labour-market

experience.

Tables 1.10 presents the estimates for the variances and covariances of the heterogeneity

components and of the error structure. In the SIM column, the value of σθ1 shows that

there are unobservable characteristics of male workers that affect the wage returns to labour

market experience.26 It is worth stressing that I estimate wage returns to labour market

experience to be highly heterogeneous: workers with a draw of θ1 that is one standard

deviation above the mean earn a marginal return of 8.7 percent for each of the first fiver years

of labour market experience. A worker with a draw that is one standard deviation below

the mean receives a return to labour market experience that is very close to zero.27 The

parameter σθ2 shows that there are individual unobservables that matter for wages above and

beyond heterogeneity in returns to labour market experience. The parameters suggest that

individual unobservables affect match duration and the accumulation of industry experience.

These results are similar to those of Abowd et al. (1999), Lillard (1999) and Dostie (2005).

All correlation coefficients between individual heterogeneity variance components are

significantly different from zero: I can reject the null hypothesis of no simultaneity across the

26One obvious example is educational attainments.27Investigating this type of heterogeneity is of interest to policy makers and researchers, and it represent

one of the advantages of employing a random effect model.


three equations at all conventional significance levels. The correlation coefficient between the

person random effect in the job hazard model and in the wage equation ρθ2θ3 is negative and

significant, which implies that on average high-wage individuals also have a lower conditional

probability of job destruction. These results are consistent with the implication of a search

model a la Mortensen and Wright (2002) where matches that last longer are of higher quality

in terms of productivity. The estimate for ρθ2θ4 shows that the equivalent is true also for

industry experience: workers who have conditionally lower wages are also more likely to

leave a sector.

Looking at the match heterogeneity variance component, the negative and significant

estimate for the parameter φ (capturing the importance of match quality in the job duration

equation) imply that there are ”good” matches28 with higher wages (conditional on all ob-

servables) and lower average conditional probability of destruction, i.e. that last on average

longer. Job duration cannot be taken as exogenous in the wage equation: longer-lasting

matches are not a random sample of all matches, and this is due in part to unobservables.

This result is in support of the results in Dostie (2005), Abowd et al. (1999) and Lillard

(1999). Quantitatively, a job that has a match effect that is one standard deviation higher

than zero in the wage equation, equivalent to a wage gap from average of around 4,600

Euros of year 2004, has a predicted probability of destruction that is around 9.5 percentage

points lower.29

The hypotheses of exogeneity of job and industry experience in the wage equation can

also be tested jointly using a Likelihood Ratio test that compares the likelihood func-

tion of the restricted model based on the implicit assumption of no simultaneity (column

“W2+J2+S2” in the regression tables) against the unrestricted model (column “SIM”).

This test easily rejects the null hypothesis of no simultaneity at any conventional signifi-

cance level.

28As in Lillard (1999) and Dostie (2005) the match effect includes both a firm effect and a pure matcheffect.

29Calculated as (−0.455)× (0.209).


1.5.2 Females

Table 1.11 presents the estimates of equation (1.3) for female workers.30 The estimates of

the SIM model show that the first two years of industry experience are associated with an

average wage effect on 2.5 percent a year. The years between the second and the fifth are

associated with slightly negative marginal effect on wages, and the effect is stable thereafter

at 0.5 percent. This is largely in line with results for males. Wage returns for the first

2 years of job seniority are higher than males’ at 1.2 percent; they fall to a negative 0.3

percent for years 3-5, and levels off at positive 0.4 percent a year after more than ten years.

The wage returns of labour market experience are much lower for females than for males.

Focusing on the results of the simultaneous model in column SIM, the first five years show an

average yearly effect of 1.6 percent and at 0.6 percent afterwards. This is consistent with the

unconditional experience wage profile shown in figure 1.1 where the gap between males and

females is growing in the number of years of labour market experience. Endogenous selection

into the labour force is a more serious concern for females than males, who are typically

found to have a rather inelastic labour supply. Therefore, these estimates suggests that the

reason for large returns to experience is not simply an artifact of endogenous selection into

employment.

Sectors and jobs that are more common for males value previous labour market expe-

rience more than those that are dominated by females’. The total wage returns over ten

years are 1.5 percent for industry experience, 2.5 for firm tenure and 11 percent for labour

market experience for females workers. In other words, sector experience and firm tenure

seem to have a very small effect on the wages of female workers. Further research is required

to investigate the consistency of these findings and to identify the specific mechanisms by

which females’ wages are much less affected by labour-market experience.

Tables 1.12 presents the results for the hazard model of employment duration for females.

Estimates are qualitatively very similar to those for males, albeit with smaller magnitudes.

These differences could be due to events affecting mobility and labour market participation

of females, such as maternity, health problems in the family, elderly care etc. The estimates

for equation (1.5) for females are in Table 1.13. The marginal effects of industry experience

on the probability of changing sectors is qualitatively similar to the estimates for males, and

30The focus here is specifically on the aspects of the estimates in which males and females show differentpatterns.


the smaller coefficients for females are similar to those discussed above for the case of job

tenure. Having between five and ten years of experience increases the hazard rate more for

females than for males. This is not surprising since the effect of labour market experience

on wages is significantly smaller for females.

Estimates in Table 1.14 suggest that individual unobservables are important for wages,

sector and job mobility of females as well. All correlation coefficients between individual

random effects are significantly different from zero. Two coefficients have the opposite sign in

comparison to the estimates for males (ρθ1θ3 and ρθ1θ4 are both positive for females while they

are negative for males): female workers with higher conditional returns to experience also

have higher probability of leaving the job and the industry they are employed in. Overall,

female workers have low returns to experience compared to males. The females that have

higher returns to experience seem to be more similar to males in terms of mobility patterns,

in that they have higher job and sector mobility than other females.

The estimate for the match heterogeneity component φ is negative and significant as it

is the case for males. The estimated coefficient implies that a job that has a match effect

one standard deviation higher than zero in the wage equation, equivalent to a gap from

the average of around 3,600 Euros of year 2004, has a predicted probability of destruction

that is around 7.2 percentage points lower.31 The Likelihood Ratio Test rejects the null

hypothesis of no simultaneity at any conventional significance level.32

1.6 Concluding Remarks

In this paper I use panel data for a sample of Italian workers in years 1986-2004 to estimate

the effect of industry experience on wages taking account of heterogeneity at the individual

and match level. The main results show that industry experience has a much stronger

impact on wages than job tenure. Estimates show that wage returns to job seniority are

very small, and that the returns to industry experience are highly nonlinear, concentrated

in the first years and very small afterwards. On the other hand, wage returns to labour

31(−0.394)× (0.183).32I have run some additional specifications for males and females. Including firm size in the wage regression

shows that, consistent with previous literature (see e.g. Troske 1999 for evidence using matched data), largerfirms pay higher wages ceteris paribus. However, its inclusion does not have any sizeable effect on the otherestimates. As discussed earlier, the inclusion of occupation controls changes the estimates very marginally.


market experience dominate returns to seniority and are very large especially for males.

My empirical model also allow me to test for the presence of endogeneity in the wage

regression. The results provide clear evidence of the endogeneity of job seniority and industry

experience in the wage equation coming from the effects of unobserved individual and match

unobserved heterogeneity: high-wage workers stay on the job longer, good matches last

longer.

The Italian labour market is considered among the most rigid in the OECD countries.33

Nevertheless, job search and job match considerations are found to be important deter-

minants of wages of workers in the early part of their careers. These results imply that

earning losses from a layoff depend on the opportunity that the worker has within the same

sector, because mobility across labour-market sectors is associated with a higher short-term

wage penalty than mobility within the same sector. Labour market policies might consider

differentiated interventions for displaced workers employed in a shrinking sector as opposed

to displaced workers in a booming sector.

This paper has an empirical focus and is largely silent about the possible mechanisms

through which labour-market experience, experience within one industry and job seniority

affect wages. Assuming experience and seniority affect wages through human capital accu-

mulation, my results suggest that industry-specific human capital is more important than

firm-specific human capital. There are a number of competing explanations, and future re-

search is needed to discriminate among some of these explanations. Estimates may at least

in part be driven by the role of labour market networks: industry experience might matter

for wages through its possible effect on a worker’s outside option, which may depend on a

worker’s network. Within the Italian context it would also be important to investigate the

role of trade unions, which are in some cases sector-specific, but for the most part operate

across sectors in Italy.

33See for example Contini and Trivellato (2005).


1.7 Tables and Figures

1.7.1 Tables

Table 1.1: Number of Jobs and Conditional Average Duration

Males Females

Number of jobs Percentage Job Duration Percentage Job Duration

One job 37.6 2.95 40.4 2.99Two jobs 24.3 2.60 24.6 2.64Three jobs 15.4 2.14 15.4 2.18Four jobs 9.5 1.80 8.9 1.83Five jobs 5.8 1.54 5.0 1.56More than five jobs 7.4 1.10 5.8 1.04Total 100.0 2.01 100.0 2.10

Frequencies 82,114 56,914

Unit of observation is the worker

Source: Author’s calculations from WHIP dataset.

Table 1.2: Number of Sectors

Percentages

Number of sectors Males Females

One sector 55.7 58.1Two sectors 27.4 27.6Three sectors 11.3 10.3More than three sectors 5.6 4.0Total 100.0 100.0

Frequencies 82,114 56,914

Unit of observation is the worker.



Table 1.3: Summary statistics for Job Covariates - Males

Variable Mean (Std. Dev.) Min. Max. N

Job duration (years) 2.01 (2.82) 0.08 18.92 208208Dummy for censored job spell 0.20 (0.40) 0 1 208208Sector spell duration 3.60 (3.89) 0.04 18.87 208208Experience at the start of the spell 1.62 (2.62) 0 17.91 208208Experience in the sector 0.71 (1.76) 0 17.91 208208

Unit of observation is the job.


Table 1.4: Summary statistics for Job Covariates - Females


Job duration (years) 2.10 (2.83) 0.08 18.92 135408Dummy for censored job spell 0.21 (0.41) 0 1 135408Sector spell duration 3.60 (3.90) 0.04 18.87 135408Experience at the start of the spell 1.56 (2.60) 0 17.78 135408Experience in the sector 0.69 (1.77) 0 17.34 135408

Unit of observation is the job


Table 1.5: Summary statistics for Year-level Covariates - Males


Real FTE wage 19738.47 (9207.73) 107.49 199393.5 537127Experience 3.66 (3.83) 0 17.91 537127Sector tenure 2.72 (3.39) 0 17.91 537127Job tenure 2.02 (3.00) 0 17.91 537127

Unit of observation is the worker-year. FTE: Full Time Employment


Table 1.6: Summary statistics for Year-level Covariates - Females


Real FTE wage 17859.53 (7454.51) 101.13 198854.67 359186Experience 3.56 (3.74) 0 17.91 359186Sector tenure 2.72 (3.37) 0 17.91 359186Job tenure 2.02 (2.93) 0 17.91 359186

Unit of observation is the worker-year. FTE: Full Time Employment



Table 1.7: Wage Equation for Males

Dependent variable: ln(wiJ(i,t)t)

Variables ModelsW1 W2 SIM

Constant 9.700*** 9.628*** 9.603***(0.016) (0.017) (0.017)

Industry Experience0-2nd year 0.034*** 0.024*** 0.021***

(0.001) (0.001) (0.001)3rd-5th year 0.003*** -0.005*** -0.003***

(0.001) (0.001) (0.001)6th-10th year 0.004*** 0.005*** 0.007***

(0.001) (0.001) (0.001)11th year + 0.004* 0.004*** 0.007***

(0.002) (0.001) (0.001)Job Seniority0-2nd year 0.004*** 0.004*** 0.005***

(0.001) (0.001) (0.001)3rd-5th year 0.007*** 0.002* 0.003***

(0.001) (0.001) (0.001)6th-10th year 0.007*** 0.000 0.001

(0.001) (0.001) (0.001)11th year + 0.000 -0.003** -0.003**

(0.002) (0.001) (0.001)Experience0-5th year 0.034*** 0.042*** 0.043***

(0.001) (0.001) (0.001)6th-10th year 0.017*** 0.015*** 0.015***

(0.001) (0.000) (0.000)11th year + 0.015*** 0.014*** 0.014***

(0.001) (0.000) (0.000)

Number of yearly wage observations: 536,277Time and sector fixed effects in all regressions

W1: Wage model without Unobserved Heterogeneity components

W2: Wage model with Unobserved Heterogeneity components

SIM: 3-Equation Simultaneous model

Asymptotic Standard Errors in Parenthesis

Significance: *=10%; **=5%; ***=1%



Table 1.8: Job Hazard Equation for Males

Dependent variable: ln(hiJ(i,t)t(τ))

Variables ModelsJ1 J2 SIM

Constant 0.06 0.000 0.398***(0.043) (0.052) (0.062)

Job Seniority0-2nd year -0.317*** -0.140*** -0.051***

(0.005) (0.005) (0.006)3rd-5th year -0.114*** -0.078*** -0.140***

(0.005) (0.005) (0.005)6th-10th year -0.069*** -0.054*** -0.072***

(0.006) (0.006) (0.007)11th year + -0.011 -0.004 -0.007

(0.012) (0.012) (0.013)Experience0-5th year -0.141*** -0.170*** -0.044***

(0.002) (0.002) (0.003)6th-10th year -0.018*** -0.003 0.059***

(0.003) (0.003) (0.004)11th year + -0.066*** -0.061*** -0.007

(0.006) (0.006) (0.007)

Number of job observations: 207,501Time and sector fixed effects in all regressions

J1: Job Hazard model without Unobserved Heterogeneity components

J2: Job Hazard model with Unobserved Heterogeneity components



Significance: *=10%; **=5%; ***=1%



Table 1.9: Sector Hazard Equation for Males

Dependent variable: ln(hsik(τ))

Variables ModelsS1 S2 SIM

Constant -0.414*** -0.360*** -0.015(0.042) (0.055) (0.072)

Industry Experience0-2nd year -0.421*** -0.260*** -0.152***

(0.005) (0.005) (0.006)3rd-5th year -0.028*** 0.054*** 0.008*

(0.004) (0.004) (0.004)6th-10th year -0.002 0.067*** 0.061***

(0.003) (0.004) (0.005)11th year + 0.002 0.079*** 0.096***

(0.005) (0.006) (0.007)Experience0-5th year -0.092*** -0.092*** 0.036***

(0.003) (0.003) (0.004)6th-10th year -0.022*** -0.017*** 0.056***

(0.003) (0.004) (0.004)11th year + -0.006 -0.037*** -0.008

(0.005) (0.005) (0.007)


S1: Sector Hazard model without Unobserved Heterogeneity components

S2: Sector Hazard model with Unobserved Heterogeneity components



Significance: *=10%; **=5%; ***=1%



Table 1.10: Variance components and Parameters for Males

ModelsW1+J1+S1 W2+J2+S2 SIM

Individual Heterogeneity Varianceand Covariance Componentsσθ1 1.013*** 1.023***

(0.014) (0.014)σθ2 0.221*** 0.222***

(0.001) (0.001)σθ3 0.500*** 0.898***

(0.005) (0.005)σθ4 0.719*** 1.283***

(0.005) (0.007)ρθ1θ2 -0.402*** -0.408***

(0.005) (0.005)ρθ1θ3 0.158***

(0.006)ρθ2θ3 -0.114***

(0.005)ρθ1θ4 0.205***

(0.005)ρθ2θ4 -0.194***

(0.004)ρθ3θ4 0.947***

(0.001)Match Heterogeneity variance componentsσδ 0.209*** 0.209***

(0.000) (0.000)φ -0.455***

(0.017)Error structureω 0.881*** 0.409*** 0.407***

(0.000) (0.001) (0.001)σν 0.161*** 0.140*** 0.139***

(0.000) (0.000) (0.000)ln-L -1423933.75 -1397664.25 -1368727.7


Significance: *=10%; **=5%; ***=1%



Table 1.11: Wage Equation for Females

Dependent variable: ln(wiJ(i,t)t)

Variables ModelsW1 W2 SIM

Constant 9.618*** 9.579*** 9.570***(0.018) (0.019) (0.019)

Industry Experience0-2nd year 0.032*** 0.027*** 0.025***

(0.002) (0.001) (0.001)2nd-5th year 0.000 -0.003*** -0.005***

(0.001) (0.001) (0.001)5th-10th year 0.006*** 0.002** -0.004***

(0.001) (0.001) (0.001)10th year + 0.000 0.000 -0.005***

(0.003) (0.002) (0.002 )Job Seniority0-2nd year 0.014*** 0.013*** 0.012***

(0.001) (0.001) (0.001)3rd-5th year 0.006*** -0.001 -0.003***

(0.001) (0.001) (0.001)6th-10th year 0.006*** 0.001 0.002

(0.002) (0.001) (0.001)11th year + 0.007*** 0.004* 0.004*

(0.003) (0.002) (0.002)Experience0-5th year 0.020*** 0.025*** 0.016***

(0.001) (0.001) (0.001)6th-10th year 0.004*** 0.008*** 0.006***

(0.001) (0.000) (0.000)11th year + 0.011*** 0.010*** 0.006***

(0.002) (0.001) (0.000)

Number of yearly wage observations: 358,591Time and sector fixed effects in all regressions

W1: Wage model without Unobserved Heterogeneity components

W2: Wage model with Unobserved Heterogeneity components



Significance: *=10%; **=5%; ***=1%



Table 1.12: Job Hazard Equation for Females

Dependent variable: ln(hiJ(i,t)t(τ))

Variables ModelsJ1 J2 SIM

Constant -0.200*** -0.305*** -0.041(0.066) (0.074) (0.135)

Job Seniority0-2nd year -0.311*** -0.143*** -0.059***

(0.006) (0.007) (0.007)3rd-5th year -0.072*** -0.031*** -0.088***

(0.006) (0.006) (0.006)6th-10th year -0.050*** -0.034*** -0.060***

(0.007) (0.007) (0.008)11th year + 0.014 0.026* 0.016

(0.013) (0.014) (0.015)Experience0-5th year -0.136*** -0.165*** -0.034***

(0.003) (0.003) (0.004)6th-10th year -0.010*** 0.004 0.073***

(0.004) (0.004) (0.005)11th year + -0.055*** -0.051*** 0.01

(0.008) (0.008) (0.009)


J1: Job Hazard model without Unobserved Heterogeneity components

J2: Job Hazard model with Unobserved Heterogeneity components



Significance: *=10%; **=5%; ***=1%



Table 1.13: Sector Hazard Equation for Females

Dependent variable: ln(hsik(τ))

Variables ModelsS1 S2 SIM

Constant -0.540*** -0.587*** -0.350**(0.062) (0.076) (0.145)

Industry Experience0-2nd year -0.358*** -0.207*** -0.093***

(0.006) (0.007) (0.007)3rd-5th year -0.019*** 0.061*** 0.017***

(0.004) (0.005) (0.005)6th-10th year -0.024*** 0.026*** 0.020***

(0.004) (0.005) (0.006)11th year + -0.003 0.048*** 0.071***

(0.007) (0.008) (0.009)Experience0-5th year -0.099*** -0.105*** 0.033***

(0.004) (0.004) (0.005)6th-10th year 0.004 0.013*** 0.088***

(0.004) (0.004) (0.005)11th year + 0.000 -0.014* 0.017*

(0.007) (0.007) (0.009)


S1: Sector Hazard model without Unobserved Heterogeneity components

S2: Sector Hazard model with Unobserved Heterogeneity components



Significance: *=10%; **=5%; ***=1%



Table 1.14: Variance components and Parameters for Females

ModelsW1+J1+S1 W2+J2+S2 SIM

Individual Heterogeneity variance componentsσθ1 1.327*** 2.059***

(0.039) (0.086 )σθ2 0.191*** 0.193***

(0.001) (0.001)σθ3 0.496*** 0.927***

(0.006) (0.006)σθ4 0.666*** 1.290***

(0.007) (0.009)ρθ1θ2 -0.373*** -0.372***

(0.009) (0.008)ρθ1θ3 -0.294***

(0.009)ρθ2θ3 -0.058***

(0.007)ρθ1θ4 -0.252***

(0.008)ρθ2θ4 -0.125***

(0.006)ρθ3θ4 0.952***

(0.001)Match Heterogeneity variance componentsσδ 0.183*** 0.183***

(0.000) (0.000)φ -0.394***

(0.029)Error structureω 0.718*** 0.277*** 0.277***

0 (0.002) (0.002)σν 0.237*** 0.207*** 0.207***

(0.000) (0.000) (0.000)ln-L -1010086.9 -994355.35 -975799.48


Significance: *=10%; **=5%; ***=1%



1.7.2 Figures

Figure 1.1: Experience Profile based on annual data

9.6

9.8

1010

.2R

eal w

ages

in 2

004

Euro

s (in

Log

s)

0 5 10 15 20Experience in years

Females MalesSource: Elaborations from WHIP Data

Experience profile using average yearly wages


Figure 1.2: Industry Experience Profile based on annual data

9.6

9.8

1010

.2R

eal w

ages

in 2

004

Euro

s (in

Log

s)

0 5 10 15 20Sector tenure in years


Sector tenure profile using average yearly wages

Figure 1.3: Job Tenure Profile based on annual data

9.6

9.8

1010

.2R

eal w

ages

in 2

004

Euro

s (in

Log

s)

0 5 10 15 20Tenure in years


Tenure profile using average yearly wages


Figure 1.4: Survivorship Function for Job Tenure

0.2

5.5

.75

1Su

rviv

al p

roba

bilit

y

0 5 10 15 20Job Tenure in Years

95% CI 95% CIMALE = Female MALE = Male

Source: Elaborations from WHIP dataset, Kaplan Meier method

Years 1986−2004Survivorship function estimates for Job Tenure

The Kernel estimations above are constructed using the Epanechnikov kernel.

Figure 1.5: Hazard Function for Job Tenure

0.1

.2.3

.4H

azar

d R

ate

0 5 10 15 20Job Tenure in Years

95% CI 95% CIMALE = Female MALE = Male

Source: Elaborations from WHIP dataset, Epanechnikov kernel

Years 1986−2004Kernel hazard estimates for Job Tenure

The Kernel estimations above are constructed using the Epanechnikov kernel.

Chapter 2

High Wage Workers and High

Wage Peers

2.1 Introduction

It has long been hypothesised that spillover effects may play an important role in the work-

place (for example, externalities across coworkers are discussed in Marshall, 1890, p. 12).

Understanding spillovers is important for our general understanding of the labor market. It

is also likely to shed light on findings such as those of Abowd et al. (1999) that firms are

important determinants of wage variation across workers, controlling for individual char-

acteristics and type. This topic is increasingly important as firm segregation by worker

characteristics rises (Kremer and Maskin 1996 and Hellerstein and Neumark 2008 for the

US; Kramarz et al. 1996 for France; Lopes de Melo 2009a for Brazil; Bagger and Lentz 2008

for Denmark) and may play a role for the recent growth in wage inequality (Edin et al.,

2008).

The issue of spillover effects in the workplace has attracted some interest among empirical

economists. However, most of the existing research is based on small datasets and very

narrow economic sectors and tasks, and focuses on the effect of peers operating through

effort or on role of team production in specific firms.1 Most of previous studies find that

1Using panel data from 20 steel minimills, Boning et al. (2007) investigate the effects of the adoptionof problem-solving teams, and find a significant positive effect on productivity. More recently, Chan et al.

36

CHAPTER 2. HIGH WAGE WORKERS AND HIGH WAGE PEERS 37

peer pressure and team-based work matter: observed effort levels are higher when a worker

is paired with higher-productivity individuals. The reason for the scarcity of results on the

labour market as a whole is related to the complexity of statistically identifying spillover

effects, which generates steep data requirements. First, workers in the same firm tend

to have similar wages even in the absence of social interactions because they share similar

characteristics and because they are interact with the same environment (‘correlated effects’,

Manski, 1993). This suggests that spillover effects ought to be estimated from changes in

workforce composition within firms. Secondly, relevant coworker characteristics may be

unobserved to the econometrician. Their exclusion might generate a downward bias of the

estimate of the role of spillovers in the labour market.

Until recently, virtually all observational data on the labour market were individual

or household surveys or censuses, making it impossible to link firm’s characteristics and

characteristics of coworkers to any specific worker. Recently, the advent of linked employer-

employee panel datasets, which include information on many workers inside the same firm

and follow the same workers over time, have made it possible to investigate spillover inter-

actions within firms and to account for the role of unobservables.

In this paper I estimate the effects of coworkers’ skills on wages. To the best of my

knowledge there are only two other studies that estimate wage spillover effects in the work-

place using a representative sample of workers and a comprehensive measure of coworker

skills,2 and both have important methodological limitations relative to this paper. Shvydko

(2007) specifies the peer effect via coworkers’ wages, which raises concerns about endo-

geneity.3 Lengermann (2002) estimates spillover effects operating through observable and

(2012) focus on a different question and investigate the role of compensation schemes on peer effects and onthe level of cooperation inside the firm, using data from a Chinese department store. Hamilton et al. (2003)investigate the effect of group composition on the productivity of teams using data from a garment plant,and find evidence of large and heterogenous spillover effects. Bandiera et al. (2009) focus on the effects ofsocial connections between workers and managers on productivity using data from a soft fruit picking farm.They find that social connections increase the productivity of workers. Ichino and Maggi (2000) look at therole of social interaction for shirking behaviour in a large Italian bank, and find group interactions to bevery important. On the other hand, Guryan et al. (2009) test for the presence of peer effects in productivityusing a dataset of professional golf players, and find no evidence of significant peer effects in that context.

2Battu et al. (2003) measure spillover effects in the UK operating through the level of education of cowork-ers, but cannot control for the role of unobservables at the worker or firm level. In a related contribution,Navon (2010) investigates the effect of knowledge diversity on within-plant human capital spillovers using apanel dataset for Israel.

3All unexplained within-firm wage variation that is common across coworkers is used to identify spillovers.


unobservable coworker characteristics, similarly to this paper. He finds that a one standard

deviation increase in an index of coworker skill is associated with wage increases of 3 to 5

percent. However, Lengermann (2002) uses a different estimator than the one considered in

this paper, and his estimator’s statistical properties are unknown.

I estimate a log-linear wage regression that extends the person and firm effects model of

Abowd et al. (1999). My regression includes fixed individual effects that capture the return

to time-invariant worker characteristics, fixed firm effects that allow for unobserved firm-level

heterogeneity, and year by sector effects to control for sector-specific macroeconomic trends.

I measure peer characteristics by the average of the fixed individual effect among workers

working in the same firm in the same time period. This measures the labour market value

of coworkers’ “portable” skills (i.e., the returns to characteristics that are person-specific

and employer-invariant). I estimate the spillover effect arising from coworkers’ observable

and unobservable time-invariant characteristics simultaneously with the other parameters,

using an estimator based on Arcidiacono et al. (2011). The spillover effect is identified

from changes in the composition of the workforce for the same worker in the same firm,

controlling for sector-specific time trends and firm size.

I estimate the model using data from the longitudinal linked employer employee dataset

Veneto Worker History (VWH), which covers the population of private-sector workers of

the Italian administrative region of Veneto for the years 1982-2001.4 I find that spillover

effects are an important determinant of wage variation: a 10-percent increase in my mea-

sure of coworker ‘quality’ is associated with a 3.6-percent wage premium. This means that

increasing coworker ‘quality’ by one standard deviation is associated with a real wage in-

crease between 4 and 8 percent. Through a simple variance decomposition, I also find that

including spillover effects reduces the overall wage variation explained by firm effects by

about one fourth, suggesting that a substantial component of firms’ contributions to wages

is determined by the composition of a firm’s workforce. I also investigate the role of skill

segregation on wage inequality for specific groups of workers in the presence of spillover

effects. I find that around 12 percent of the gender wage gap and 10 to 15 percent of the

immigrant wage gap is due to the labour market characteristics of peers.

4VWH includes wages and individual characteristics of all workers in each firm. Other datasets, suchas the Longitudinal Employer-Household Dynamics (LEHD) dataset for the US and the LIAB dataset ofthe Institut fur Arbeitsmarkt- und Berufsforschung (IAB) for Germany, could be used for this study withrelatively simple modifications to my strategy. I intend to work on those datasets in the future.


2.2 Theoretical Background

The theoretical literature has identified a number of channels through which the quality of

coworkers could affect a worker’s wage. However, it is hard to predict sign and magnitude of

spillover effects in the labour market theoretically. New empirical evidence can generate fu-

ture theoretical work and inform about the relative practical importance of the mechanisms

I outline below.

First, there may complementarities in the production function, such that a worker’s

marginal productivity may depend on the characteristics of her coworkers. One channel

that has received particular attention is the possible effect of human capital heterogeneity at

the firm level on productivity, as analysed in Kremer (1993), Davis and Haltiwanger (1991),

Kremer and Maskin (1996) and Dunne et al. (2000). Navon (2010) finds that knowledge

heterogeneity within a firm indeed affect spillovers. In a related contribution, Moretti (2004)

tests for the existence of human capital spillover effects across firms within cities and finds

productivity spillovers to be positive and significant for hi-tech plants in the US.5

The characteristics of peers might play a role in wage determination even when workers

are not cooperating, through peer pressure. Two recent papers examine the role of peer

pressure in the workplace using laboratory and field data for isolated tasks. Falk and

Ichino (2006) use a lab experiment to investigate social pressure spillovers, and find that

productivity is higher and less dispersed when subjects work in pairs. Mas and Moretti

(2009) use field data from a large US supermarket chain where worker pairs are varied.

Their estimates show that individual effort is positively correlated with the productivity of

nearby workers.

The average quality of coworkers might also affect individual wages through a worker’s

reservation wage, which may operate through preferences and social norms. Workers may

have a preference for working with a certain type of coworkers, and may be willing to accept

a lower wage for that because of compensating differentials, and this may generate either

positive or negative spillover effects depending on workers’ preferences. Kremer and Maskin

(1996) discuss evidence of social pressure for wage equality within the firm. In addition,

reference points may be important for wage determination (see Dittrich et al. 2011 for an

5In a recent paper, Kurtulus (2011) investigates the role of demographic dissimilarity among coworkersusing data from a large U.S. firm. She finds that age and tenure dissimilarity are associated with lowerworker performance. On the other hand, wage dissimilarities are associated with higher worker performance.


overview of the literature on the effect of minimum wages across the wage distribution).

If the wage structure within the firm provides a reference point for all workers, wages of

lower-ability workers will be affected by the skill composition inside the firm. For instance,

Kronenberg and Kronenberg (2011) find that workers are more likely to leave a firm as wage

inequality in the firm increases. If firms internalised this effect, wages of low-skill workers

would be a positive function of the average level of skills in the firm.

Coworkers’ skills may also affect wages through bargaining dynamics. If high-skill work-

ers are able to extract a higher share of the surplus through bargaining, and bargaining

outcomes are positively correlated within a firm (for example because agents can use pre-

vious bargaining outcomes as reference points), then a worker’s wage will increase with

coworker skills. Conversely, if highly-skilled workers are able to extract a larger share of

the a surplus in a context where wages are a fixed share of total revenues, spillover effects

would be negative due to negative bargaining externalities.

Incentive schemes within the firm can also generate interactions between wages and peer

characteristics. In tournament models6 effort (and thus wages) are a function of the charac-

teristics of all workers in the firm. However, the relationship between quality of coworkers

and individual effort does not need to be positive or monotonic, as discussed in Becker and

Huselid (1992), because of the discouragement effect : low ability workers may choose zero

effort if they perceive their probability of winning to be very low.7 In addition, the expected

level of cooperation among workers (and thus total output and individual wages) may also

depend on the distribution of types. In order to assess the relative importance of these

different channels, the existence and magnitude of the spillover effect should be investigated

empirically first, which is what I turn to in the next section.

6Initiated by the seminal work of Lazear and Rosen (1981).7Harbring and Irlenbusch (2003) offer an excellent review of the literature and also present compelling

experimental evidence showing that in a variety of different treatments agents tend to choose very low levelsof effort in general, and very often zero.


2.3 Empirical Model

My empirical model builds upon the structure of the model of Abowd et al. (1999). In the

following, let i denote a worker, j denote a firm and t a time period.8 A worker i working

at a firm j in period t shares that same employer j with other workers, which I refer to

as i’s set of current coworkers, or current peer group. A worker i at time t has the set of

coworkers Nijt at time t, with cardinality Nijt (in each period there are Nijt + 1 workers at

firm j including worker i). One of worker i’s coworkers is denoted by p. My main regression

model is

wijt = Xitβ + Fjtκ+ θi +

1

Nijt∼i

∑p∈Nijt∼i

θp

η + ψj + τt + εijt (2.1)

where the outcome of interest is worker i’s log monthly wage wijt. I denote time-variant

individual characteristics of worker i by Xit, firm size by Fjt, individual time-invariant

characteristics by θi, whose average among peers9 is 1Nijt∼i

∑p∈Nijt∼i

θp. Firm effects are

denoted by ψj and industry-specific year effects by τt. The b × 1 column vector10 β and

the scalars κ and η are parameters to be estimated. The scalar η captures the effect of

average time-invariant individual characteristics of peers on log wages, which is the my the

parameter of interest. Finally, εijt is a transitory mean-zero error term.

As discussed in Manski (1993) and Bramoulle et al. (2009) there are significant challenges

for identifying peer effects in a linear in means model: it is hard to distinguish social effects

from correlated effects. This paper provides solid grounds on which to rule out much of the

possible role for correlated effects. Individual covariates Xit are included because individual

characteristics that have an effect on wages might also be correlated with the average quality

of a worker’s peer group. I also include firm size, denoted by Fjt, in order to make sure

that my estimates of peer effects are not driven by growth and decline in the number of

8Since the estimation follow workers over time, a more precise notation defines the firm where worker iis employed at the t as J(i, t), but I simply use j since all cases are unambiguous.

9Sometimes I refer to this measure as peer ‘quality’. The reader should be cautions with its interpretationhowever. Since I observe wages and not productivity, θ will capture all of the characteristics that makea worker more productive (more able to produce) and the return to those characteristics as well as thecharacteristics that will make him/her more able to extract rents (holding productivity constant). In otherwords, θ captures the market value of portable skills, and so it does not address the underlying mechanismsthrough which that market value may be different for different workers. If a group of workers receives lowerwages even when I control for their individual characteristics, they will have a lower θ.

10Where b is the number of individual time-variant characteristics included in the model.


employees of a firm, which could be a source of bias for example if firms paid higher wages

but attracted lower-ability workers when they grew in size. There may also be common-

environment effects (‘correlated effects’, Manski 1993): some firms might systematically

better at attracting high-wage workers and might also give out higher wages, conditional on

a worker’s fixed effect. I address this issue by including time-invariant firm effects denoted

by ψj in equation (2.1).

Moreover, I include time effects to control for trends in the average ability of peers and

in the outcome variable, which could affect my estimates of spillover effects. For example,

during a boom firms may pay higher wages but may also see the average ability of their

workforce decrease, which would be the case if marginal workers had lower-than-average

skills. In order to allow for time trends to be different in different sectors of the economy,11

I include industry-specific year trends, denoted by τt in equation (2.1). The individual fixed

effect θi measures the ‘market value of portable skills’ or ‘portable component of individual

wages’. Equivalently, 1Nijt

∑p∈Nijt

θp measures the mean of θ among people working with

worker i at time t. For notational convenience I define θijt ≡ 1Nijt∼i

∑p∈Nijt∼i

θp.

The nonlinear least squares problem derived from equation (2.1) is then

minβ,κ,θ,η,ψ,τ

∑i

∑t

[wijt −Xitβ − Fjtκ− θi − θijtη − ψj − τt

]2(2.2)

Equation (2.2) is written under a ‘proportionality’ assumption (same as in Arcidiacono

et al. 2011 and Altonji et al. 2010) on the characteristics included in θi . It gives a structure

to the relationship between the coefficients on each of the components of θi in the direct

effect on wijt as opposed to its indirect effect through peers. The proportionality assumption

states that the relevant importance of each of these components is the same in the direct

effect on own wages and in the peer effect. For example, if two characteristics that are part

of θi have the same effect on the log wage of worker i, those same two characteristics will

also have the same effect when operating through peers.

Under the proportionality assumption I can directly apply Theorem 1 of Arcidiacono

et al. (2011) for consistency and asymptotic normality of η. The key assumption of Theorem

11In the period of my panel different economic sectors have been exposed to labour market regulations,and to exposure to global markets in a very heterogeneous way, and so an average time trend would notadequately control for the relevant macroeconomic context of each industry.


1 requires residual across any two observations to be uncorrelated: net of person effects,

firm effects and time effects, correlations in outcomes across individuals in the same peer

group are assumed to be captured by the peer effects entirely. In my case, this assumption

implies that workers may be different in their unobserved ability, firms may be systematically

different in the average ability of their workforce, there might be yearly time trends that are

different for different sector. However, all of the remaining intertemporal changes in peer

‘quality’ within a firm, controlling for all of the other covariates, are used to estimate the

peer effect coefficient η, and so need to be assumed orthogonal to the error term εijt. This is

equivalent to assuming that there are no time-varying unobservables driving changes in the

composition of the peer group of worker i while at the same time systematically affecting

worker i’s wage.12

Under the assumptions stated above, the nonlinear least squares solution ηNLS is a

consistent and asymptotically normal estimator of the true parameter η as the number

of individuals goes to infinity for a fixed number of time periods. The key elements that

allow Arcidiacono et al. (2011) to prove this theorem is that the vector of individual fixed

effects can be written as a function of the spillover parameter and of the data, so that the

Least Squares problem above can be formulated as an optimization problem with only one

minimand, η. Arcidiacono et al. (2011) can then use Theorem 12.2 of Wooldridge (2002)

for consistency of M-estimators establishing identification and uniform convergence, and

Theorem 12.3 for asymptotic normality. Even though my problem is complicated by the

presence of additional fixed effects, the main logic of their proofs applied here.

There are reasons why equation (2.2) is still restrictive. First, the model is specified as

a linear-in-means model,13 so that I cannot investigate spillover effects operating through

a different moment of the relevant distribution, and I am also not exploring possible het-

erogeneity in spillover effects. In addition, I assume away endogenous effects: peers’ wages

affect a worker’s wage only through the effect of peers’ ability, not directly via their own

12Thereom 1 of Arcidiacono et al. (2011) also requires either homoskedasticity within each peer group orthat heteroskedasticity is uncorrelated with the number of observations available for each worker. In additionto these assumptions, we also need a few standard assumptions: Corr(θ, ε) = 0, E(θ4i ) < ∞, E(εijt) = 0,E(ε4int) < ∞. Finally we need η to lie in the interior of a compact parameter space Γ where the largestelement of Γ needs to be smaller than 2. See Arcidiacono et al. (2011) page 7 for details on these assumptions

13This is by far the most common choice in the peer effects literature. There are a few exceptions thatare worthy of being mentioned because of their role in the peer effects literature. Brock and Durlauf (2001,2003) use the nonlinearity arising in discrete-choice models to distinguish endogenous effects from exogenouseffects.


wages, for example through effort.14 If peers’ effort choice positively affected a worker’s

effort choice, and effort and ability were correlated, my estimates of η in equation (2.2)

would be upward biased.15

In order to estimate equation (2.2) I find the vector of parameters θ and the parameter η

that minimise equation (2.2) iteratively.16 Intuitively, I start from a model without spillover

effects to get a first set of estimate of all fixed effects. I then use these first estimates to

get a first estimates of my regression parameters β, κ and η. Then update the fixed effects

and then switch between updating the fixed effects and updating the parameters, until

convergence.17

Because of the spillover effect the outcome of person i at time t is a function of the

ability of all of i’s co-workers, which are themselves estimated within the model.18 Each

iteration consists of four steps.19 For a general iteration α the four steps are as follows:

1. Estimate ηαOLS , βαOLS and καOLS from θα−1, ψα−1, τα−1 using Ordinary Least Squares;

2. Estimate θα from θα−1, ψα−1, ηαOLS , βαOLS and καOLS using equation (B.1.2);

3. Estimate ψα from θα, τα−1, ηαOLS , βαOLS and καOLS using equation (B.1.3);

4. Estimate τα from θα, ψα, ηαOLS , βαOLS and καOLS using equation (B.1.4).

14Without this assumption on endogenous effects, my estimates can be viewed as a combination of exoge-nous and endogenous effects, i.e. effects operating through peer characteristics and through behaviour.

15In my context endogenous effects are likely to be a function of time-varying covariates, and so therewould be endogeneity problems including them in my wage regression.

16Estimating equation (2.2) in one step is not computationally feasible with a large dataset.17The specific iterative procedure described below builds upon that of Arcidiacono et al. (2011) adapting

it to the labour market context and in particular to the inclusion of firm effects.18The inclusion of additional covariates compared to Arcidiacono et al. (2011) and in particular of firm

effects and year by sector effects does not affect the main logic of the estimation. When the θs are updated,all of the other fixed effects and covariates are treated as columns of data. For additional details see myAppendix.

19See the Appendix for details and for the updating equations I use.


2.4 Data and Institutional Background

I estimate my model using the longitudinal Veneto Worker History (VWH) dataset.20 The

dataset includes virtually all private-sector workers of the administrative region of Veneto in

the North East of Italy21 for years 1982-2001.22 The VWH dataset includes register-based

information on all establishments and employees that have been hired by those establish-

ments for at least one day during the period of observation. The entire employment history

in the period 1982-2001 has been reconstructed for each employee.23 The VWH dataset is

a very large dataset, with the unselected sample containing around 3.6 million workers and

46 million observations at the worker by year level over 20 years.24

The region of Veneto is the third Italian region by GDP and has a population of around

5 million people, around 8 percent of the country’s total. Its economy is characterised by

small manufacturing businesses which are organised on a regional basis by specialisation

and with local integration.25 The region underwent fundamental economic changes in the

last few decades. Until after World War II, the economy of Veneto was largely based

upon farming and saw large out-migration to Germany, Switzerland, the US, Canada and

Australia. The 1960s and 1970s were characterized by intense economic development, and

Veneto is now one of the richest and most industrialised regions of Italy, and a large net

receiver of international migrants. Immigrants currently represent around 10 percent of the

population of Veneto, which is well above the Italian average.26 According to Istat (2011)

20This panel dataset is built at the department of Economics of the University of Venice using the SocialSecuity administrative data of the Istituto Nazionale per la Previdenza Sociale (INPS), which is the mainpublic institute of social security in Italy.

21State and local government employees, farm workers and some category of professionals, such as doc-tors, lawyers, notaries and journalists, are not included because they have alternative social security funds.Additional information on the dataset available in Card et al. (2010) and in Tattara and Valentini (2010)

22The period covered by the dataset is 1976-2001, but because coding errors concerning wages have beenfound for the period 1976-1981, I will only use the 20-year period between 1982 and 2001. The VWH datasethas not been updated for the years after 2001.

23Considering the occupational spells out of the region of Veneto as well for individual regressors.24Additional details on the structure of the dataset are available in the Appendix.25In particular, the region is has been often studied as a model for industrial districts, characterised by

highly specialised small firms, which tend to be geographically very concentrated. Engineering, textile andclothing and furniture being some of the prominent industries. Some of these small firms have experiencedremarkable growth such as Benetton and Luxottica.

26Data from various Italian censuses available from http://www.istat.it/ show that the proportion offoreign residents (defined as people residing in Italy while not holding an Italian citizenship) increased slowlyfrom 1.2 percent in 1961 to 2.2 percent in 1971 to 3.7 percent in 1981, after which it soared to 6.1 percent

http://www.istat.it/


data, the total number of immigrants in Veneto at the end of 2009 is 489,000, 10 percent

of the total population.27 The equivalent figures for 1991 are 25,000 in absolute number,

around one percent of the total population.

Italy is often viewed as a country where collective bargaining is the main mechanism

for wage determination. The reality is more complex, and especially for small firms as

those that dominate the labour market of Veneto there are many potential sources of wage

differentials across workers. National regulations are typically silent about compensation

levels. Trade union contracts specify non-binding minimum wages at the industry level.

Although these are relevant for bargaining inside the firm, they simply represent an industry-

specific floor for total compensation, and in the region of Veneto actual compensation are

usually higher. Because minimum wages are occupation and rank-specific, promotions can

affect the relevance of the contractual minimum wages. Individual bargaining and firm-

level agreements are also important, and wage premia are highly heterogeneous across firms

(Erickson and Ichino, 1993), and usually higher for small firms, where individual bargaining

plays a larger role (Cingano, 2003).

2.5 Sample Restrictions

In order to estimate my model, it is necessary to identify a specific time dimension for

the panel dataset such that in each time period there is at most one observation for each

worker.28 I choose to construct a dataset where there is at most one observation for each

worker in each year.29 I create a wage variable that measures average monthly wages for full

time employment, so that my estimates of wages are driven by variation in compensation

per unit of time rather than by labour supply variation. The main regressor of interest in

in 1991 and to an estimated 7.5 in 2001 (and on a positive and sharp trend in the last decade).27Employed immigrants in 2009 are around 11 percent of all employed. These figures do not include

undocumented migrants. According to Anastasia et al. (2009) however the proportion of irregulars andtemporary migrants in Veneto is less than 10 percent of the total number of immigrants, which is a muchlower proportion than many other Italian regions. And in the future there will be a very large proportion ofthe whole labour force that will be constituted by second generation immigrants, since in 2009 the percentageof children whose parents are not Italian citizens is 21.2 percent Istat (2010). A synthetic account of thedevelopment of migration to Italy is offered in Colombo and Sciortino (2004).

28Failing to do so would result in higher weights given to more mobile individuals, and it would make itimpossible to include a control for time trends.

29I thus eliminate the case in which there is more than one observation for each worker in the same year.See the Appendix for details.


my dataset is a measure of ability of a worker’s coworkers, which can be constructed only if

the firm has at least two workers.30

Separately identifying fixed effects and person effects requires employment histories to

be sufficiently connected. A connected group of firms and workers contains all the workers

that ever worked for any of the firms in the group and all the firms where any of the

workers were ever employed (Abowd et al., 2002). I use an algorithm to identify connected

groups of observations.31 Abowd et al. (2002) then proceed by estimating person and firm

effects within each group to maintain the representativeness of the sample. I simply drop all

observations that are not part of my main connected sample before estimating my model.32

2.6 Summary Statistics

My regression sample has 28,115,529 yearly observations for 231,195 different firms and

3,180,714 workers. Of these workers, 40.8 percent are female, 8.2 percent are foreign born,

31.2 percent are white collar workers. There is substantial worker mobility in my sample:

I observe around two thirds of the workers in more than one firm.33 Figure 2.1 plots

average monthly wages (in 2003 Euros) for full time employment by gender and the gender

distribution over time. The proportion of females increases from 35.5 percent in 1982 to

around 40 percent from 1997 onwards. Monthly real wages increased both for females and

males, but we observe a break in the trend around 1991, with real wages increasing at 2.41

percent a year on average for females and 2.15 percent for males in the years 1982-1991,

and only 0.37 percent for females and 0.10 percent for males a year in the period 1992-2001.

30This eliminated around three percent of all observations, where firms only had one employee. I alsoconstruct a variable for labour market experience and for firm size, please see the Appendix for details.

31I use the algorithm “a2group” written by Ouazad (2007), who in turn develops it from a Fortran im-plementation written by Robert Creecy. I had to make only minor changes to their code to deal with alarger number of firms. The basic functioning of the algorithm mirrors the definition of connected groups:starting from a single firm, the algorithm finds the set of workers that worked for that firm in any timeperiod, and includes those as part of the connected graph. The algorithm then adds all of the firms that setof workers ever worked for, and add all of the workers that worked for those firms to the connected graph.This procedure continues until no additional worker is added to the connected graph.

32Only around 9,000 observations out of over 28 million are excluded from the main connected graph, i.e.just over 2,000 workers are outside the main connected group out of over 3 million, and only around 1,000firms out of over 230,000.

33For 25 percent of all workers I observe two employers, for around 16 percent I observe three employers,for around 10 percent I observe four employers. Five percent of individuals work for 5 firms within the periodof my data, and a further 6 percent has 6 or more employers.


The gap between monthly wages of males and females decreased slightly from 24.3 percent

in 1982 to 20.4 percent in 2001.

Figure 2.2 compares workers born in Italy with workers born abroad. The proportion

of foreign-born workers increases dramatically, from 2.6 percent in 1982 to 9.8 percent in

2001.34 The chart shows the first large influx of foreign born workers and the first sizeable

arrival of people of different ethnicities around 1990, driven mostly by immigrants from

Morocco and Albania. The unconditional wage gap between foreign born and Italian born

was relatively constant at around 400 Euros in the period 1982-1989. Afterwards, it increases

dramatically, driven primarily by falling average wages of foreign born.35 While in 1982-

1989 average yearly growth rates of gross real wages are 1.70 percent for Italian born and

1.98 percent for foreign born, in the period 1990-2001 the equivalent figures are 0.71 percent

for Italian born and -0.33 percent for foreign born.

Looking at wage dispersion, Figure 2.3 plots the standard deviation of the natural log-

arithm of monthly wages in each year within and across firms, both normalized to 100 in

1982.36 Beyond the year-to-year variations due to business cycle fluctuations, there is a

clear upward trend for both measures. Wage dispersion within firms increases by around

8 percent from 0.379 in 1982 and peaks at 0.414 in 1997. Wage inequality across firms

rose relatively more, (consistent with the finding of Kremer and Maskin 1996 for the US)

from 0.331 in 1982 to a maximum of 0.393 in 1999 (an 18 percent increase from 1982),

slightly dropping afterwards. The overall standard deviation of log monthly wages rises by

11 percent in the same period.

The average number of employees of a firm37 is 21.6 in 1982, falls gradually to 18.1 in

1993 and then levels off, attaining 18.5 in 2001. The median size of firms in my sample

34Up to 1989, foreign born are between 2 and 4 percent of the total. For the most part reflects foreignborn with Italian parents returning from Switzerland, Germany and Latin America.

35In 1990, foreign born would earn on average 2,700 Euros for each month they work full time. Aseverywhere else in the paper, the reader should bear in mind that these wages are gross of taxes and forfull-time months. An earner that earns 2,700 Euros per full-time month probably earns around 1,000 or1,100 Euros net of taxes in a normal month. By 2001, eleven years later, their average wages had fallen to2,600 while those of Italian born are over 3,300 Euros, for a staggering gap of 23 percent.

36I calculate the average of the within-firm standard deviation and the standard deviation of the averagewage of each firm. Figure 2.3 is constructed from a dataset that has one observation for each worker in each,so that I can take account of firm size. If I had one observation per firm the statistics on the chart would beentirely driven by small firms.

37As I mentioned above because I need to exclude firms that only have one employee from my sample, thisaverage is higher than the average for the population of firms in Veneto during the same period.


is 6 throughout the period, with a dip at 5 in 1998 only. In 2001, out of 83,173 firms, 25

percent of all firms have either 2 or 3 employees, 75 percent of the firms have 15 employees or

less, only one percent of all firms have more than 200 employees. Table A.2 shows that the

largest economic sectors in terms total of number of firms are commerce, bars and hotels (28

percent of all firms), construction (20 percent of all firms), construction of metal products

(7 percent) and banking and insurance (6 percent).

2.7 Regression Results

The main estimates of equation (2.1) are presented in Table 2.1. I report heteroskedasticity-

robust standard errors and t-statistics for my coefficients.38 Column 1 estimates a model

with a firm fixed effect, a worker fixed effect and a year by industry effect only. Column 2

adds firm size and a polynomial in labour market experience. Controlling for firm effects, the

effect of firm size and labour market experience on wages is very small.39 In Column 3 I add

a the average person effect of peers. The relative estimated effect is 0.358, which implies

that, using the overall standard deviation of θ, which is 0.218, a one-standard-deviation

increase in the average person effect of a worker’s peers is associated with a wage gain of

7.81 percent.

I also calculate the cross-sectional standard deviations of θ for three representative years,

which are equal to 0.221 for 1982, 0.201 for 1991 and 0.199 for 2001. Estimates associated

wage gain for either of these are between 7.12 percent (using the 2001 standard deviation)

and 7.89 percent (using the 1982 standard deviation). An alternative reference distribution

is the average standard deviation of θ within a person’s career, which is 0.104. This might

be more intuitively appealing since the overall distribution of peer quality in the population

38Arcidiacono et al. (2011) gives no guidance on how to calculate the exact standard errors and so showstandard errors and t statistics from the OLS regression of the last iteration. While these are only approxi-mate standard errors, given their size and the size of my dataset, improved standard errors are very unlikelyto make any difference for inference.

39Both in terms of coefficients and in terms of its effect on R2. This is consistent with the discussion inAbowd et al. (1999) firm size is a proxy for something else in the firm that we typically do not observe, andthis is what is driving large estimates of firm size when we use cross sectional variation in wages only. Largefirms seem to be systematically different from small firms but firms do not pay systematically higher wageswhen they grow. Because my experience measure is in part imputed, this may be lower than what I wouldobtain if I could observe labour market experience from the beginning of their careers for all of the workersin my sample.


may not be the natural reference for considering the changes in coworker composition that

workers in my data might actually experience. Using this alternative reference distribution

a one standard deviation increase in peer characteristics is associated with a wage increment

of 3.7 percent. In this case, the conditional wage effect of having a group of peers that is

one standard deviation higher than average is similar to the effect of two years of labour

market experience. My estimates of 3.7 and 7.9 percent effects may be seen as lower and

upper bounds.40

2.8 Post-estimation Analysis

2.8.1 Person Fixed Effects and Coworker “Quality”

The person effect θ that I estimate in my main regression analysis includes skills as they are

valued in the labour market, offering a summary of the product of relevant characteristics

and their returns. Therefore, there is the potential for the person effects to be affected by

different labour market returns to the same characteristics for specific groups within firms.

As discussed briefly above, this makes the use of the term “coworker quality” or “coworker

skills” for an average of person effects potentially problematic since the measure I use is

derived from wages directly, not from productivity measures.

In order to partially address this concern, I regress the person effect estimated in my

main regression analysis on gender and immigration status. My estimated are shown in

Table 2.2. I find that foreign birth status and gender together explain less than 5 percent

of the variation in θ across workers. This is suggestive evidence that the extent to which

θ is driven by wage discrimination based on gender and immigration status is likely to be

limited.41 Given the limitation of my dataset, only speculation is possible concerning the

determinants of the person effect θ. Having excluded a role for discriminatory behaviour

based on gender, the major factors generating heterogeneity in the person effects are likely

to be educational attainment, unobserved skill/ability, motivation, labour market networks

etc. Many of these characteristics are likely to reflect productivity differentials, and therefore

40In the Appendix, I include robustness checks running the same regression on a subsample of the popu-lation, to investigate how much my estimates vary by firm size.

41This of course does not rule out that specific groups of the population may be discriminated against inother markets and that there could be an indirect effect. For example, women may have had differentialaccess to educational opportunities, and that would affect their outcomes in the labour market as well.


support the claim that my measure of “coworker effects” is indeed a measure that provides

information on characteristics that we would expect to matter in this context.

2.8.2 Variance Decomposition

From equation (2.1) I decompose the variance of log monthly wages wijt as follows:

V ar(wijt) = Cov(wijt, wijt) = Cov(wijt,Xitβ + Fjtκ+ θi + θijtη + ψj + τt + εijt)

= Cov(wijt,Xitβ) + Cov(wijt, Fjtκ) + Cov(wijt, θi) + Cov(wijt, θijtη)

+Cov(wijt, ψj) + Cov(wijt, τt) + Cov(wijt, εijt)

This can be normalised dividing both sides by V ar(wijt):

Cov(wijt,Xitβ)

V ar(wijt)+Cov(wijt, Fjtκ)

V ar(wijt)+Cov(wijt, θi)

V ar(wijt)+Cov(wijt, θijtη)

V ar(wijt)

+Cov(wijt, ψj)

V ar(wijt)+Cov(wijt, τt)

V ar(wijt)+Cov(wijt, εijt)

V ar(wijt)= 1

(2.3)

The bottom section of Table 2.1 lists each element of equation (2.3), and shows that the

contribution of individual time-invariant characteristics to the variance of individual wages

is between 44 and 49 percent. Sector-specific year effects, on the other hand, explain between

5 and 6 percent of wage variation. Experience and firm size are of marginal importance.

Firms’ heterogeneity accounts for around 20 percent of wage variation in column 1 of Table

2.1,42 falling to 18 percent in column 2. Once we control for peer quality, the proportion of

wage variation that is explained by firm effects decreases by about 28 percent. In turn, the

average quality of peers explain around 5 percent of the overall wage variation.43 The R2

of column 3 is very similar to that of column 2: the additional 5 percent of the variance of

log wages explained by peer quality is associated with a similar decrease in the proportion

of the variance explained by the firm effect. An important portion of what our usual firm

effects pick up is driven by the level of skills of that firm’s labour force.

42Gruetter and Lalive (2009) estimate a similar model as column 1 of my model and finds an estimate of27 percent.

43Adding complexity to the functional form used and including a function of peers’ worker effect beyondhe first moment is likely to increase this estimate, which should therefore be seen as a lower bound of theeffect of peers’ ability.


The main regression analysis discussed above includes the estimate of the correlation

between the person effect and the firm effect. Abowd et al. (1999) finds small negative cor-

relation coefficients, which is at odds with much of the theoretical literature which predicts

assortative matching between workers and firms, and generated a large and unsettled de-

bate. When I do not include peer effects in my model, I find positive estimates: high-wage

workers tend to work in high-wage firms.44 Once I include peer effects in the model, the

correlation between θ and ψ is just 0.014. Most of the correlation between θ and ψ found in

column 3 is driven by high-wage workers having high-wage peers: the correlation between

θ and θ is 0.459. 45 I compute the correlation between θ and θ for each year of my dataset

and find it to be between 0.36 and 0.45 depending on the year I use.46 My estimates of

the correlation between θ and ψ and between θ and θ are consistent with the discussion in

Lopes de Melo (2009b): sorting in the labour market operates primarily through workers

with similar levels of productivity sharing the same employer, rather than by the matching

of workers and firms per se.

Figure 2.5 expands on these findings investigating the correlation between worker quality

and firm quality by year for the period 1982-2001. The solid line in Figure 2.5 shows that

the overall correlation between the firm effect and the person effect varies greatly within

the sample period. It is negative between 1982 and 1989 and positive afterwards, peaking

at 0.074 in 1997. The change in this correlation over time is driven both by mobility of

existing workers, and by new firms and new workers entering the dataset. In order to

separate these two sources, I include in Figure 2.5 correlations calculated only for the firms

that are active in 1982. The dashed line follows the same pattern as the solid line, and

exhibits slightly higher correlation coefficients: using a constant pool of firms, the increase

in assortative matching across the sample period is stronger, which is consistent with the

results in Mendes et al. (2010) of stronger assortative matching among long-lived firms using

44Note that the comparison is highly imperfect because I use wages per unit of time while the literaturetypically uses total annual compensation.

45Lopes de Melo (2009b) argues that a better measure of sorting is the correlation between the fixed effect ofa worker and that of her coworkers because of possible non-monotonicities between the firm effect and a firm’sproductivity. Lopes de Melo (2009b) discusses a theoretical model based upon Shimer and Smith (2000),which implies that correlation between person effect and firm effect underestimates the extent of sorting inthe labour market. In a related contribution, Eeckhout and Kircher (2009) also find non-monotonicities ofwages around the equilibrium point, reflecting the structure of the firm’s opportunity cost.

46Prior to this study the two available studies that calculate the equivalent correlation find values between0.3 and 0.4 for Brazil (Lopes de Melo, 2009a) and Denmark (Bagger and Lentz, 2008).


data from Portugal. Figure 2.5 also includes the correlation between the person effect and

the firm effect using the sample of workers that are active in 1982 only. The dotted line while

following a very similar trend but at a lower level: among workers that have been around

for long, the movement towards assortative matching is slower. The movement towards

assortive matching is disproportionally driven by “old” firms and by “new” workers.

2.8.3 Fixed Effects across Specific Groups

Table 2.3 presents the average of wages and of the estimated fixed effects across genders

and immigrant status. On average, female workers have 25 percent lower wages, 20 percent

lower ‘market value of portable skills’ measured by the fixed person effect θ, 8 percent lower

coworker ‘quality’ and work in firms that pay conditionally slightly lower wages. On the

other hand, on average a foreign born worker has a wage that is 13 percent below that of

a native worker; her person effect θ is 15 percent lower and coworker ‘quality’ is 9 percent

lower. Her firm effect ψ is on average 2 percent lower.

2.8.4 Gender Wage Gap and Peer ‘Quality’

In order to further investigate the role of peers on the gender wage gap I documented

in my data, I decompose the average wage gap between the wages of male and female

workers. Consider the following simple decomposition of the average gender wage gap based

on estimating equation (2.1):

E(wMijt − wFijt) = E(XMit β −XF

itβ) + E(FMjt κ− FFjtκ) + E(θMi − θFi ) + E(θMijtη − θFijtη)

+ E(ψMj − ψFj ) + E(τMt − τFt ) + E(εMijt − εFijt)(2.4)

where the exponents F and M stand for ‘Female’ and ‘Male’ respectively. This decom-

position shows that around 85 percent of the overall wage gap between female and male

workers is due to differences in θ, i.e. differences in individual characteristics and their re-

turns in the labour market.47 Differences in peer ‘quality’ explain 12 percent of the overall

47Note this component of the gap does not necessarily reflect differences in skills, since it is itself acombination of skills and their wage returns. Foreign born may have lower labour market skills but are alsolikely to have lower returns to those unobserved labour market skills, for many reasons which may include


gap: one eighth of the gender wage gap is due to the fact that females have on average

coworkers with lower person effect θ. On the other hand, all other covariates as well as the

unexplained component are very small.

To assess the role of gender on peer exposure in more detail, I regress average peer

quality on gender and a series in controls:

θijt = Femaleiδ0 + θiδ1 + Xijtδ2 + Pijtδ3 + ψjδ4 + υijt (2.5)

where θ, θ and ψ are those I estimated my main model and Female is a dummy for

gender. The matrix Xijt includes a constant, experience and firm size. In addition, Pijt

denotes the proportion of females among worker i’s coworkers at time t. Finally, υijt is a

transitory mean-zero error term and δ0, δ1, δ2, δ3 and δ4 are parameters to be estimated.

Table 2.4 presents the estimates from equation (2.5). Column 3 shows that once I control

for the proportion of females among peers, female workers have conditionally higher-θ peers

compared to males.48

2.8.5 The Immigrant Wage Gap

Figure 2.2 and Table 2.3 documented a large and growing wage gap between foreign born

and native born workers.49 Figure 2.4 shows that foreign born and native workers are also

segregated across firms: in 2001, while native workers work in firms where around 9 percent

of workers are foreign born (the corresponding median is around 5 percent), foreign born

workers work in firms where 22 percent of workers are foreign born (corresponding median

is 16 percent). This patterns suggest that peer effects may contribute substantially to the

wage gap between them.

Figure 2.6 shows a simplified graphical representation of the decomposition in equation

(2.4) over time. As shown in Figure 2.2, the overall gap in log monthly wages between

labour market discrimination as found in audit studies.48Column 4 introduces controls for experience and firm size, column 5 adds the firm effect as well, and

shows that ceteris paribus higher-paying firms have lower-θ workers on average. The main insights fromcolumn 3 are confirmed in columns 4 and 5.

49I focus on the simplest case and simply divide my sample of workers in foreign born and Italian born.The analysis of the role of peers for different groups of immigrants is left to future work. The reader shouldbe also aware that my foreign -born dummy includes second generation Italians born abroad, and thus it isnot equivalent to an immigrant dummy.


foreign born and Italian born rises during the period covered by my dataset. Across 1982-

2001, the majority of this gap is driven by differences in the person effect θ. Average peer

characteristics explains between 10.4 percent in 1982 and 15.9 percent in 1987 of the overall

wage gap. My decomposition also shows that a large part of the wage gap (19 percent on

average) is explained by the firm effect ψ: foreign born disproportionately work in firms

that pay lower wages.

Next I regress peer characteristics on a dummy for foreign born and on other covariates:

θijt = (Foreign born)iδ0 + θiδ1 + Xijtδ2 + Pijtδ3 + ψjδ4 + υijt (2.6)

where Pijt denotes the proportion of foreign born among worker i’s peer group and all

other covariates and parameters are defined as in equation (2.5). Table 2.5 displays the

estimates for equation (2.6). Even controlling for own unobserved ‘type’ θi, the proportion

of foreign born among the peer group, experience, firm size and firm effects, foreign born

still have peers that have lower average person effects. Column 5 shows that wages of foreign

born workers lower by around 0.5 percent due to the characteristics of their peers.

2.9 Concluding Remarks

In this paper I estimate the effect of coworkers’ characteristics on wages. I address the main

sources of possible bias due to group selection (by which workers with certain characteristics

are non randomly distributed across firms) and to the role of unobservables, by using within-

firm variation in the peer group composition net of time trends and allowing peer effects to

operate through all relevant time-invariant worker characteristics. I use a large panel dataset

of workers of the Italian region of Veneto for years 1982-2001. I find peer characteristics to

be an important factor for wage determination: a 10-percent increase in coworker ‘quality’

is associated with a rise in real monthly wages of 3.6 percent. In addition, I find that after

controlling for peer quality the effect of firms’ unobservables on wages decrease by more

than one fourth. Next I find that differences in time-invariant labour market characteristics

of peers explain around 12 percent of the gender wage gap and 10 to 15 percent of the

immigrant wage gap.


2.10 Figures and Tables

2000

2500

3000

3500

Mon

thly

gro

ss w

ages

in 2

003

Euro

s

.35

.36

.37

.38

.39

.4Pr

opor

tion

of fe

mal

es

1980 1985 1990 1995 2000Year

Proportion of females in the workforce FemalesMales

Source: Elaborations from VWH data

Average wages and proportion of femalesMonthly Wages and Female Participation

Source: Author’s calculations from the Veneto Worker History Dataset.

Figure 2.1: Average monthly wages by gender and proportion of females


2400

2600

2800

3000

3200

3400

Mon

thly

gro

ss w

ages

in 2

003

Euro

s

.02

.04

.06

.08

.1Pr

opor

tion

of fo

reig

n bo

rn

1980 1985 1990 1995 2000Year

Foreign born in the workforce Foreign bornBorn in Italy


Average wages and proportion of foreign bornMonthly Wages and Proportion of Foreign Born


Figure 2.2: Average monthly wages by foreign born status and proportion of foreign born

100

105

110

115

120

Stan

dard

dev

iatio

n of

mon

thly

wag

e, 1

982=

100

1980 1985 1990 1995 2000Year

Across firms Within firmsSource: Elaborations from VWH data.

Standard deviation of wages in 1982−2001Wage segregation within and across firms


Figure 2.3: Standard deviation of log monthly wages over time


0.0

5.1

.15

.2.2

5Pr

opor

tion

of p

eers

who

are

fore

ign

born

1980 1985 1990 1995 2000Year

Mean for Italian born Median for Italian bornMean for foreign born Median for foreign born


Proportion of foreign peers: mean and medianForeign born peers


Figure 2.4: Average proportion of peers that are foreign born

−.1

−.05

0.0

5.1

Corre

latio

n co

effic

ient

1980 1985 1990 1995 2000Year

All firms all workers Only firms active in 1982Only workers active in 1982

Source: Elaborations from VWH data.

Correlation between estimated person effects and firm effectsHigh−wage workers and high−wage firms


Figure 2.5: Correlation between person effects and firm effects over time


0 .05 .1 .15 .2 .25Earning gap in log of 2003 real monthly wages

20012000199919981997199619951994199319921991199019891988198719861985198419831982

Source: Elaborations from VWH data.

The role of skills, firms and peersDecomposition of the immigrant wage gap

Theta PeersPsi Residual


Figure 2.6: Decomposition of the gap between the wage of foreign born and Italian born


Table 2.1: Main regression results.

Dependent variable: ln(wijt)

ModelsVariables (1) (2) (3)

Estimated coefficients of covariatesExperience 0.013 0.018

(0.000) (0.000)[773] [631]

Experience2 -0.001 -0.001(0.000) (0.000)[-960] [-729]

Firm size/1,000 0.013 0.013(0.000) (0.000)[526] [541]

Coworker Quality θ 0.358(0.0000)[11,074]

Fixed effectsStandard deviation of the person effect: σθ 0.383 0.413 0.389Standard deviation of the firm effect: σψ 0.230 0.215 0.205Standard deviation of the time effect: στ 0.170 0.201 0.200

Pseudo R2 0.716 0.720 0.722

Standard deviations of θσθ (overall s.d.) 0.218σθ,1982 (cross sectional s.d. for 1982) 0.221

σθ,1991 (cross sectional s.d. for 1991) 0.201

σθ,2001 (cross sectional s.d. for 2001) 0.1991N

∑Ni=1 σθ,i (average of within-person s.d.) 0.104

Corr(θ, ψ) 0.154 0.164 0.014Corr(θ, θ) 0.459

Variance decompositionPerson effect θ 0.462 0.491 0.469Firm effect ψ 0.201 0.181 0.134Time effect τ 0.054 0.058 0.058Experience 0.056 0.082Experience2 -0.077 -0.080Firm size 0.010 0.010Spillover effect η 0.049Unexplained εijt 0.284 0.280 0.278

Nobs = 28,115,529, Nworkers = 3,180,714, Nfirms = 231,195

Approximate robust standard errors in brackets, t-stats in squared brackets

Source: Veneto Worker History Dataset.


Table 2.2: The contribution of gender and immigration status to the person effect

Dependent variable: θi

(1) (2) (3)

Dummy for Female -0.180∗∗∗ -0.193∗∗∗

(0.001) (0.001)

Dummy for Foreign born -0.192∗∗∗ -0.244∗∗∗

(0.001) (0.001)

Interaction: Female * Foreign born 0.108∗∗∗

(0.002)

Constant 4.412∗∗∗ 4.354∗∗∗ 4.434∗∗∗

(0.000) (0.000) (0.000)

Observations 3180714 3180714 3180714R2 0.032 0.011 0.046

Standard errors in parentheses∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001


Table 2.3: Standardised wage, θ and ψ gaps for different groups

Populations log(wage) Person effect θ Spillover effect θ Firm effect ψ

Full sample mean 7.88 4.46 4.46 1.78Full sample standard deviation 0.57 0.39 0.22 0.21

Gender gap 0.25 0.21 0.08 0.01Foreign-born gap 0.13 0.15 0.09 0.02



Table 2.4: Gender and quality of peers

Dependent variable: 1Nijt

∑p∈Nijt

θp

(1) (2) (3) (4) (5)

Dummy for female -0.082∗∗∗ -0.030∗∗∗ 0.037∗∗∗ 0.032∗∗∗ 0.032∗∗∗

(0.000) (0.000) (0.000) (0.000) (0.000)

Individual unobserved 0.247∗∗∗ 0.238∗∗∗ 0.222∗∗∗ 0.221∗∗∗

heterogeneity θi (0.000) (0.000) (0.000) (0.000)

Proportion of females -0.240∗∗∗ -0.234∗∗∗ -0.234∗∗∗

in peer group (0.000) (0.000) (0.000)

Experience 0.003∗∗∗ 0.003∗∗∗

(0.000) (0.000)

Experience2 -0.000∗∗∗ -0.000∗∗∗

(0.000) (0.000)

Firm size/1,000 0.026∗∗∗ 0.026∗∗∗

(0.000) (0.000)

Firm heterogeneity ψ -0.017∗∗∗

(0.001)

Observations 28115529 28115529 28115529 28115529 28115529R2 0.033 0.214 0.285 0.339 0.339

Heteroskedasticity-robust standard errors in parentheses∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001



Table 2.5: Birth place and quality of peers

Dependent variable: 1Nijt

∑p∈Nijt

θp

(1) (2) (3) (4) (5)

Dummy for foreign born -0.094∗∗∗ -0.056∗∗∗ -0.014∗∗∗ -0.014∗∗∗ -0.014∗∗∗

(0.000) (0.000) (0.000) (0.000) (0.000)

Individual unobserved 0.254∗∗∗ 0.244∗∗∗ 0.226∗∗∗ 0.226∗∗∗

heterogeneity θ (0.000) (0.000) (0.000) (0.000)

Proportion of foreign born -0.410∗∗∗ -0.374∗∗∗ -0.377∗∗∗

in peer group (0.001) (0.001) (0.001)

Experience 0.003∗∗∗ 0.004∗∗∗

(0.000) (0.000)

Experience2 -0.000∗∗∗ -0.000∗∗∗

(0.000) (0.000)

Firm size/1,000 0.025∗∗∗ 0.025∗∗∗

(0.000) (0.000)

Firm heterogeneity ψ -0.023∗∗∗

(0.001)

Observations 28115529 28115529 28115529 28115529 28115529R2 0.009 0.213 0.240 0.291 0.292

Heteroskedasticity-robust standard errors in parentheses∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001


Chapter 3

The ESD Policy for Aboriginal

Students in B.C.

Note: A summary of the main results is published as “Non-Standard English Dialects and

the Effect of Supplementary Funding on Educational Achievement” with Mark Campbell,

Jane Friesen and Brian Krauth in the Canadian Journal of Speech-Language Pathology and

Audiology, 2011, Vol. 35(2), p.190-197.

3.1 Introduction

Compared to non-Aboriginal people in Canada, Aboriginal Canadians face a substantially

higher risk of poverty (Mendelson, 2006), poor health outcomes, drug and alcohol addiction,

and suicide (Health Canada, First Nations and Inuit Branch, 2003). While features of the

education system, such as residential schooling, may have contributed to these problems

in the past (Royal Commission on Aboriginal Peoples, 1996), many analysts argue that

the key to breaking the cycle of poverty among Aboriginal Canadians lies in improving

educational attainment among Aboriginal children and youth (e.g. Richards and Vining

2004). This view is supported by evidence from other populations that education is as-

sociated with better health behaviours and outcomes (Kenkel, 1991), substantially lower

rates of incarceration (Lochner and Moretti, 2004), higher earnings (Card, 1999), reduced

teen childbearing, criminal propensity, child abuse and neglect, and improved educational

64

CHAPTER 3. THE ESD POLICY FOR ABORIGINAL STUDENTS IN B.C. 65

attainment and health outcomes of children (Greenwood, 1997), increased voter and civic

participation (Dee, 2004), and reduced reliance on public transfers (Wolfe and Haveman,

2001).

Identifying effective strategies for improving educational outcomes is therefore of great

policy significance. British Columbia (B.C.) began in the 1980s to allocate funds under the

province’s English as a Second Language (ESL) policy framework to support students who

“speak variations of English that differ significantly from the English used in the broader

Canadian society and in school” (British Columbia Ministry of Education, 2009). In prac-

tice, the non-standard dialect speakers who receive “English as a Second Dialect” (ESD)

funding under this policy are almost exclusively students who also self-report Aboriginal

identity. B.C. is not unique in taking this approach: Australia has provided ESD funding

to Aboriginal speakers of non-standard dialects since the 1990s (Eades and Siegel, 1999).

Several U.S. school districts have also developed ESD programs aimed primarily at African-

American children, the leading example being the Los Angeles Unified School District’s En-

glish Mastery program (Adger, 2008), which dedicates regularly scheduled teaching blocks

and frequent “language breaks” to “mainstream English language development” (Los Ange-

les Unified School District, 2009).1 Although socio-linguists and educators have argued that

appropriate educational programs may be effective in supporting the language development

of speakers of non-standard dialects (e.g. Rickford 1999; Ball et al. 2006), we are aware of

no systematic evaluations of such policies.

We implement a strategy that exploits the staggered uptake of ESD funding across

B.C.’s public school districts between 1999 and 2004 to measure its relationship to test

score gains. Our results indicate that the provision of ESD funding has been associated

with improved reading test score gains of Aboriginal students between grades 4 and 7.

Quantile regressions suggest that this benefit is concentrated among students in the lower

quantiles of the distribution of reading test scores gains. In contrast, we find no association

between ESD funding and the numeracy test score gains of Aboriginal students. Since

these funds are intended to support services designed to help students become proficient in

English, the absence of any relationship to numeracy test score gains may not be surprising.

Similarly, we find no relationship between ESD funding and the reading or numeracy test

1An attempt by the Oakland Unified School District to access federal funds by recognizing African-American Vernacular English (AAVE) as a distinct language (Oakland Unified School District, 1997) failedspectacularly while sparking the acrimonious “Ebonics debate” in the late 1990s (Ramirez et al., 2005).


score gains of non-Aboriginal students. These negative results increase our confidence that

we have identified a true policy effect on the reading scores of Aboriginal students.

3.2 Related Literature

The supplemental funding that districts receive under the ESD policy that we evaluate differs

from general funding because it is allocated in recognition of the needs of specific students.

At the same time, the province does not dictate which services are offered to these students,

and some of the funds may be directed to other uses so long as an adequate program is

in place to address the identified needs. Our paper therefore is related to, but distinct

from, two literatures: one interested in measuring the effects of overall education funding

levels on academic achievement (e.g. Barrow and Rouse 2004; Guryan 2001), and one that

evaluates the effects of specific programs or pedagogical practices (e.g. Angrist and Lavy

2001; Lavy and Schlosser 2005). Within this second literature, our work is perhaps most

closely related to Machin and McNally (2008)’s study of the effect of a structured literacy

program on reading and English skills. Machin and McNally find that an inexpensive

pedagogical intervention was highly effective at improving skills among English primary

school children. While the intervention that we evaluate is also aimed at improving literacy

skills, the program studied by Machin and McNally involves few resources and specific

pedagogy. In contrast, the program we examine involves substantial supplementary funding

aimed at a particular group of students to address a particular learning issue, but without

prescribing pedagogy.

The nature of the funding supplement that we evaluate is perhaps most similar to special

education funding, which is also allocated in recognition of the needs of specific students.

Several studies examine the effects of special education programs on academic achievement.

Hanushek et al. (2002) use an individual fixed effects framework to estimate the effects of

moving into or out of special education programs on elementary school test score gains in

mathematics. They find that being placed in special education has a substantial positive

effect on math test score gains in the first year, but that much of this benefit is lost in the

following year. Cohen (2007) uses the variation in rates of learning disabilities induced by the

Chicago accountability framework to measure the effect of special education placement on

student outcomes. Her results indicate that being placed in special education in elementary


school reduces high school absenteeism and drop-out rates; her results for high school GPA

are inconclusive.

3.3 Institutions and Data

3.3.1 Organisation and Funding of the B.C. Education System

In B.C., the Ministry of Education establishes curricula and provides operating and capital

grants to district school boards, who then allocate funds to individual schools. The per-

student funding formula over the period of our study is summarised in Table 3.1. In addition

to a basic grant for each regular student enrolled, districts receive supplementary grants for

self-identified Aboriginal students, for students with assessed disabilities and for students

who are deemed to be eligible under B.C.’s ESL policy. Our primary interest is in this last

policy, which provides a 21% supplement to the base grant for a maximum of five years for

each eligible student.

3.3.2 English as a Second Dialect policy in B.C.

Districts that receive ESD funding for Aboriginal students under B.C.’s ESL Policy Frame-

work have substantial discretion in terms of the services provided, subject to meeting several

broad criteria (British Columbia Ministry of Education, 1999). For each designated student,

districts are required to conduct an annual assessment of proficiency in Standard English,

and to design an annual instruction plan that lists specific services the student will receive

in order to improve that proficiency. An ESL specialist must be involved in service planning

and delivery, and districts are encouraged to use culturally relevant resources to provide

services (British Columbia Ministry of Education, 1999). The Ministry does not specify

which services must be provided, and there is no requirement that the funding be dedicated

exclusively to services for the designated student.

This flexibility allows districts to use these funds in a variety of ways. Examples include

supporting the use of specific pedagogical strategies for vocabulary development (Nechako

Lakes School District No. 91, 2006), hiring specialist teachers that provide support to class-

room teachers and develop program materials (Cariboo-Chilcotin School District No. 27,


2009), offering specialised oral language instruction on a weekly pull-out basis and acquir-

ing reading materials with Aboriginal content (Vancouver Island North School District No.

85, 2008), and integrating strategies for oral language development into regular literacy

programs (Haida Gwaii/Queen Charlotte School District No. 50, 2008).

3.3.3 The Foundation Skills Assessment Exams

We are interested in investigating the effect of ESD funding on student achievement, as

measured by B.C.’s Foundation Skills Assessment (FSA) exams. The FSA exams have been

administered to students in grades 4 and 7 in all public and provincially-funded private

schools since 1999. They are based on a variety of questions, both multiple-choice and

open-ended, and are graded by accredited B.C. teachers. All students except those whose

English skills are not sufficiently developed that they can respond to the test, and some

disabled students, are expected to participate. Assignment to ESD does not in itself affect

the expectation that students will participate in the FSA exams. The exams are relatively

low-stakes for all parties: students’ scores do not contribute to their school grade and play

no role in grade completion, and the results do not affect school or district funding or teacher

pay. However, school- and district-level results are made public and are widely discussed

within both the educational system and the news media (e.g. Cowley and Easton 2003).

3.3.4 Data

Our data are drawn from the B.C. Ministry of Education’s enrolment database,2 and its

FSA exam database. These cross-sectional data sets cover all public school students in

grades 4 through 7 between 1999 and 2004. We use the unique identification code assigned

to each B.C. student to link records across the enrolment and FSA exam databases, and

to construct a longitudinal record for every public school student who is in grade 7 in B.C.

from 2002 through 2004, and who was in grade 4 in B.C. three years earlier.3

2Since responsibility for on-reserve Aboriginal education falls under federal jurisdiction, our data do notinclude the approximately 7.5% of B.C. Aboriginal students who attend on-reserve schools (Friesen andKrauth, 2010).

3A minority of students who are observed in both Grades 4 and 7 repeat grades, skip grades, or are outof province or in a private school for one or more of the intervening years. We keep these students in ouranalysis whenever possible. If the student repeats either Grade 4 or Grade 7, we use the last year in Grade4 and the first year in Grade 7.


The enrolment record includes the student’s current grade, school and district identifiers,

year, gender, self-reported Aboriginal identity, enrolment in a language program (e.g. ESL,

French Immersion, Francophone education), enrolment in a special needs program, and self-

reported language spoken at home. Records in the FSA exam database include the student’s

score on each exam subject, along with a flag indicating whether the student was excused

from writing a given exam.

3.3.5 ESL Funding and Student Achievement in B.C. public schools

Characteristics and Achievement of Aboriginal and non-Aboriginal Students

As Table 3.2 shows, 9.6% of students in the grade 7 student population in our sample period

self-report Aboriginal identity. Aboriginal students tend to be sorted into particular schools.

Although they make up less than 10% of the grade 7 student body, the average Aboriginal

student attends a school in which almost 24% of their grade 7 peers are Aboriginal; the

average Aboriginal student in ESD attends a school in which almost half of their grade 7

peers are Aboriginal.4 Although only 1% of Aboriginal students report speaking a language

other than English at home, over 9% are funded under B.C.’s ESL policy. In contrast, only

5.8% of non-Aboriginal students receive ESL funding in Grade 7, although over 21% report

speaking a language other than English at home.

The administrative data provide a number of indications that Aboriginal students, and

particularly those in ESD programs, face significant challenges. The lower panel of Table

3.2 shows that the incidence of assessed disabilities among Aboriginal students is two and

half times as high as the non-Aboriginal population, and is almost four times as high among

Aboriginal students in ESD programs. The incidence of assessed giftedness shows the reverse

pattern.

Table 3.3 presents our measures of academic achievement for the sub-sample of students

with FSA exam results in both grades. Aboriginal students in grade 7 score close to 0.6

standard deviations on average below non-Aboriginal students on both exams. Aboriginal

students in ESD programs score a full standard deviation below the mean in numeracy and

1.3 standard deviations below the mean in reading. In comparison, the mean test score gap

4Friesen and Krauth (2010) provide a detailed assessment of school-level sorting of Aboriginal and non-Aboriginal students and examine the effects of peer composition on Aboriginal students’ test score gains.


between blacks and whites on standardized tests in Texas elementary schools is about 0.7

standard deviations (Hanushek et al., 2002). The next two rows of Table 3.3 show that

Aboriginal students on average continue to fall behind in numeracy between grades 4 and

7, and to an even greater extent in reading. Aboriginal students in ESD programs actually

catch up in numeracy between grades 4 and 7, but their reading test scores decline by a

further 0.13 standard deviations relative to non-Aboriginal students.

The use of test scores to measure achievement levels or gains has the drawback that it

restricts attention to those students who participated in the exam in one or both grades,

and is thus subject to bias from endogenous participation. This problem will be particularly

acute when studying a population that has relatively low academic achievement. The exam

participation statistics in Table 3.4 show that, although non-Aboriginal exam participation

rates in B.C. are high, they are considerably lower among Aboriginal students, and lower

still among Aboriginal students in ESD. About half of those who do not participate have

been excused from the exam. The other half simply do not take the exam, either because

they are absent from school on exam day or because they do not respond to the exam. This

high nonparticipation rate results in a high proportion of Aboriginal students with missing

gain score data: about 25% of Aboriginal students overall and 44% of ESD participants do

not have valid reading gain score data. The number of missing gain scores is slightly higher

for the numeracy exams.

Patterns of ESD Assignment

Although ESD funds were available under B.C.’s ESL policy as early as the 1980s (British

Columbia Ministry of Education, 1985), few districts took advantage of this source of funding

before the late 1990s. Table 3.5 illustrates the rapid expansion in the uptake of ESD funding

between 1999 and 2004. In 1999, only 2.7% of grade 7 Aboriginal students in the province

were identified as eligible for ESL funding; by 2004, this proportion had risen to 10.7%. The

number of districts assigning at least 5% of grade 7 Aboriginal students to ESD (and at

least 10 grade 7 students in total) grew from 4 in 1999 to 16 in 2004.5

5The incidence of Aboriginal students in ESL programs must be computed from micro data, becausethe Ministry of Education does not publish the numbers. The micro data to which we have access for thisstudy does not extend beyond 2004, so we are unable to determine whether ESD programs have continuedto expand since then. However, our discussion with various Ministry staff have provided no indications ofany significant changes to the pattern of uptake of this policy.


Figure 3.1 shows the proportion of grade 7 Aboriginal students in ESD in the first year

that services were offered to more than 5% of Aboriginal students, in each of the twelve

districts where we observe this (arbitrary) threshold being crossed. In nine of these twelve

districts, the proportion of Aboriginal students assigned to ESD jumped from fewer than

5% to more than 20% in a single year. In Nisga’a and Stikine, the proportion leapt from less

than 5% to over 60% from one year to the next. The scale of this within-district variation

is clearly greater than any plausible variation in underlying student characteristics, and is

therefore likely to reflect changes in district ESD policy.

Even in districts with well-established ESD programs, year-to-year variation in the pro-

portion of Aboriginal students assigned to ESD occurs on a greater scale and at a higher

frequency than the likely variation in student characteristics. Figure 3.2 shows the variation

over time in the proportion of grade 7 Aboriginal students in ESD for four illustrative dis-

tricts. Vancouver is one of four districts that assigned more than 5% of Aboriginal students

to ESD in 1999; in this case we observe considerable year-to-year variation in the proportion

of students in ESD, but no clear trend. The other three districts show large discrete jumps,

along with considerable year-to-year variation in subsequent years.

3.4 Methodology

In the section above we have shown that the proportion of Aboriginal students assigned to

ESD varied substantially both within and between districts during our sample period. This

section describes how we exploit this variation to measure the relationship between ESD

funding and student achievement, and attempt to identify the causal effects of ESD funding.

Next we present our empirical model, discuss the substance of our identifying assumptions

and present the estimating equations that we derive from this model.

The population under analysis is Aboriginal students attending B.C. public schools for

grades 4 through 7 during our sample period. Because our models include district fixed

effects, we also restrict attention to students who are in the same district from grade 5 to

grade 7.


3.4.1 Our Empirical Model

This section describes an empirical model that incorporates the three regression specifica-

tions that we estimate below. We do not estimate this model directly. Instead, we use

the model to define the effects we are aiming to measure and to discuss the substantive

conditions under which the fairly simple regressions we estimate will consistently measure

these effects.

Index students by i, districts by d, and time by t. Let t(i) be the school year in which

student i takes grade 7, and let d(i) be the student’s district between grades 5 and 7. Our

model features an outcome equation and an ESD assignment equation:

yi = bd(i) + γt(i) + βxXi + βDESDi + βIpi + ui (3.1)

ESDi,t = ad(i) + δt + πXXi + πppd(i),t + vi,t (3.2)

In the assignment equation (3.2), ESDi,t is an indicator of student i’s ESD assignment in

year t, ad(i) and δt are district and time fixed effects, Xiis a vector of background characteris-

tics, and pd,t is the proportion of Aboriginal students assigned to ESD in district d(i) in year

t. In the outcome equation (3.1), yi is a grade 7 outcome of interest, bd and γt are district

and time fixed effects, ESDi ≡(ESDi,t(i)+ESDi,t(i)−1+ESDi,t(i)−2 )

3 is the proportion of years

that student i was in ESD in grades 5 through 7, and pi ≡(pd(i),t(i)+pd(i),t(i)−1+pd(i),t(i)−2 )

3 is

the average proportion of Aboriginal students assigned to ESD in district d(i) over the years

that student i was in grades 5 through 7. We are primarily interested in two parameters: βD

captures the direct effect of ESD on students who are themselves assigned to ESD, and βI

captures the indirect effect of the district-level proportion of Aboriginal students assigned

to ESD on the average Aboriginal student. The indirect effect could take the form of gen-

eral fiscal/resource spillovers (since districts receive additional funds for each ESD student),

program spillovers (since district ESD programming may include development of new learn-

ing materials that benefit all students), or peer effects (since academic improvements by

high-risk ESD students may improve the classroom learning environment).

In order to identify any parameters of interest, we require that our measure of the district-

level proportion of Aboriginal students assigned to ESD be strictly exogenous conditional

on the fixed effects:


E(ui|{Xj , pj}d(j)=d(i)

ad(i), bd(i)

)= 0 (3.3)

E(vi,t|{Xj , pj}d(j)=d(i)

ad(i),bd(i)

)= 0,

t ∈ {t (i) , t (i)− 1, t (i)− 2} (3.4)

Equations (3.3) and (3.4) mean roughly that, given the district, variation in the district

proportion of Aboriginal students assigned to ESD while a student was in grades 5-7 is un-

related to variation in unobserved factors affecting either student outcomes (3.3) or student

ESD assignment (3.4).

Equation (3.4) would be violated if we were to measure the district ESD assignment rate

using students in our estimation sample, as pj would be a function of vi,t. Instead, we use

available data on assignment rates in grades outside of our estimation sample to construct

these measures.6 Specifically, we measure the district ESD assignment rate using grade 7

Aboriginal students in 1999 to 2001 and grade 4 Aboriginal students in 2002 to 2004.7

Although using students from grades outside of our estimation sample to construct our

measure of district ESD assignment rates eliminates any mechanical relationship between

this measure and the unobserved characteristics of students in our sample, equation (3.4)

would still be violated if within-district changes in ESD assignment policies were driven by

trends in student needs for such programming that are common across grades. As discussed

above, there are a number of discrete jumps in ESD designation rates, and that changes

in student needs are unlikely to match these jumps in timing or magnitude. Even within

districts that have assigned substantial proportions of Aboriginal students to ESD for many

years, we do not observe any clear trends in ESD assignment that might suggest changes

in the underlying composition of student characteristics within a district that are common

across grades.

6While our estimation sample covers the three cohorts of students who complete grade 7 between 2002and 2004, our complete data set includes grade 4 and grade 7 data for all years from 1999 to 2004.

7More specifically, we define program size as: pd,t ≡

{1.66

∑i ESDi,tI(t(i)=t,d(i)=d)∑

i I(t(i)=t,d(i)=d), t < 2002∑

i ESDi,tI(t(i)=t+3,d(i)=d)∑i I(t(i)=t+3,d(i)=d)

, t ≥ 2002

where I(·) is the indicator function and 1.66 is a scaling constant to account for the fact that grade 4 ESDrates are about 66% higher (province-wide over the period 1999-2004) than grade 7 ESD rates.


The zero-conditional-mean assumption of equation (3.3) would be violated if within-

district changes in district ESD assignment policy were systematically implemented in com-

bination with other outcome-relevant changes. Changes in district ESD practices may have

been precipitated in some cases by idiosyncratic factors that raised district awareness of

the provincial policy or of other districts’ implementation of ESD. However, if changes in

district ESD assignment rates were precipitated by changes in school district leadership or

shifting district priorities, or if other relevant policies were introduced at the same time, the

effects of these changes will be confounded with the effects of interest. While we cannot rule

out these possibilities, we are aware of no specific policy changes that coincided with ESD

assignment practices. In addition, we present some evidence on the question by looking at

multiple outcomes. One sign of unobserved policy variation would be if we were to find

substantial “effects” of district ESD policy on outcomes unrelated to ESD services.

3.4.2 Regressions

Reduced Form Model

Given equations (3.1), (3.2), (3.3) and (3.4) it is simple to derive the reduced form regression

model by substituting equation (3.2) into equation (3.1) to get:

yi =[bd(i) + βDad(i)

]+[γt(i) + βDδt(i)

]+ [βX + πXβD]Xi [βI + πpβD] pi + [ui + βDvi]

= cd(i) + λt(i) + αXXi + αRF pi + εi

(RF1)

Apply (3.3) and (3.4) to get:

E(εi|{Xj , pj}d(j)=d(i)

cd(i)

)= E

(vi + ui|{Xj , pj}d(j)=d(i)

cd(i)

)= 0 (RF2)

Equations (3.1), (3.2), (3.3) and (3.4) imply that the regression model (RF1) consistently

estimates the reduced form effect, αRF = βI+ πpβD. Although the reduced form regressions

provide no distinction between direct and indirect effects, they represent our primary results

because they derive from the most plausible identifying assumptions.


Instrumental Variables Model

We also report estimates that use the district-level ESD designation rate as an instrumental

variable for a student’s own ESD assignment. The IV regression model is:

yi = bd(i) + γt(i) + βXXi + βDESDi + ui (IV1)

E(ui|{Xj , pj}d(j)=d(i)

bd(i)

)= 0 (IV2)

The parameter βD can be interpreted as the direct effect of ESD assignment on student

outcomes. Given equations (3.1)-(3.4) and the instrument relevance condition, πp 6= 0,

consistency of the IV estimator also requires that there are no indirect effects:

βI = 0 (3.5)

Equation (IV1) follows from substitution of (3.5) into (3.1), while equation (IV2) follows

directly from (3.3). The examples of ESD implementation include several cases where

districts used ESD funds to pay for learning materials or other resources that would benefit

all students. The IV results therefore should be interpreted with caution.

OLS Model

Finally, we also report OLS estimates of the model:

yi = bd(i) + γt(i) + βxXi + βDESDi + βIpi + ui (OLS1)

E(ui|{Xj , pj , ESDj}d(j)=d(i)

bd(i)

)= 0 (OLS2)

The parameter βD can be interpreted as the direct effect of own ESD assignment (as

in the IV regression), and the parameter βI can be interpreted as the indirect effect of

the district ESD designation rate on the average Aboriginal student. In order for the OLS

estimator to be consistent, we need equation (OLS2) to hold. Given the ESD assignment

equation (3.2), this implies that:

E(ui|{Xj , pj , ESDj}d(j)=d(i)

bd(i)

)= E

(ui|{Xj , pj , vj}d(j)=d(i)

bd(i)

)= 0 (3.6)


Equation (OLS1) is identical to (3.1), while substitution of (3.2) into (OLS2) yields

(3.6). Equation (3.6) is violated if there are unobserved factors that affect a student’s ESD

assignment and are related to unobserved factors that affect his or her exam results. This is

highly likely; for example, skills in Standard English are an important determinant of both

exam outcomes and ESD assignment. The results from the OLS model therefore should be

interpreted as descriptive only.

3.5 Results

3.5.1 Test Participation

If the introduction and/or expansion of ESD within districts coincided, by accident or design,

with changes in exam participation patterns, ESD may be associated with improvements in

measured test scores even in the absence of improvements in academic achievement. Nothing

about ESD assignment itself has a direct effect on the ability of schools to keep students

out of the exams, and B.C. does not have a formal test-based accountability system that

would create incentives for schools to exclude low-achieving students from tests. However,

the publication of “school report cards” by an independent think tank (e.g. Cowley and

Easton 2003) provides some incentives for schools to raise test scores.8 In order to address

this concern, we explore the relationship between test participation and ESD assignment

rates before turning to our models of test score gains. We do so by implementing the three

econometric models described in above, but with participation in the grade 7 exam as the

outcome variable. We control for grade 4 participation and introduce an interaction term

to allow for the effect of ESD to vary by grade 4 participation. These controls have been

added to the model because participation in grade 7 only affects the availability of data on

test score gains for students who participated in grade 4. Results for a simplified model

without these variables (see Table A.4) are similar to those reported here.

The results are presented in Table 3.6. The OLS regression (equation OLS1) treats both

the number of years the student spent in ESD between grades 5 and 7 and the average district

ESD assignment rate as exogenous explanatory variables. This regression provides a biased

8Jacob (2005) and Figlio and Getzler (2006) have both found evidence that schools in jurisdictions withtest-based accountability systems use disability designations to strategically reduce exam participation bylow-achieving students.


estimate of the policy effect due to endogenous assignment, and should be interpreted as

descriptive only. The OLS estimates show that Aboriginal students with ESD designations

are somewhat less likely than other Aboriginal students to write both the reading and

numeracy exams.

The IV regression (equation IV1) uses the district ESD rate as an instrument for the

number of years the student spent in ESD, and aims to measure the effect of one year of ESD

funding under the assumption there are no indirect effects. The reduced form regression

(equation RF1) uses the district ESD assignment rate as an explanatory variable rather than

as an instrument. The coefficient on the assignment rate can be interpreted as capturing

both direct and indirect effects of ESD policy on participation among Aboriginal students.

The signs of the OLS point estimates are reversed in the IV and reduced form estimates,

implying that the introduction or expansion of ESD is associated with increased grade 7

exam participation in both subjects. This association is weak, and is only statistically

significant in one case: numeracy exam participation for students who did not take the

grade 4 numeracy exams.

While these results indicate that ESD assignment rates are not associated with overall

reductions in exam participation by Aboriginal students, we cannot rule out an association

between district ESD assignment policy and a change in the composition of the test-taking

group. However, any such compositional change would have to leave the size of the test-

taking group either unchanged or larger.

3.5.2 Test Score Gains: Main Results

Table 3.7 presents our results for the value-added model of test score gains. For each exam

subject we estimate the same three regressions. All regressions control for the student’s

gender and for district and year fixed effects. The reported standard errors are clustered by

district.

The OLS results show that ESD and non-ESD students have similar value-added reading

exam outcomes, while ESD students have better value-added numeracy exam outcomes than

non-ESD students. Again, these results should be interpreted as descriptive only; they are

consistent with districts targeting students who are facing particular difficulties in language-

related subjects for assignment to ESD.


The IV results, which measure the effect of one year of ESD funding under the assump-

tion there are no indirect effects, suggest that the policy is quite effective in achieving its

goals. The estimated effect of ESD assignment on value-added reading exam outcomes is

positive, statistically significant, and quite large: a 0.33 standard deviation improvement

in test scores from a single year in ESD. The corresponding effect on value-added numer-

acy exam outcomes is nearly zero, as one might expect of a policy that primarily targets

language development.9

The RF results for reading show that the reading test score gain of the average Aboriginal

student is greater when the district receives funding for a greater proportion of Aboriginal

students. As with the IV estimates, the implied effect is statistically significant and fairly

large: a 10 percentage point increase in the proportion of ESD students in the district is

associated with an average increase in test score gains of 0.048 standard deviations. Our

estimates imply that being in a district with an average ESD assignment rate while in

grades 5, 6 and 7 is associated with an average increase in the reading test score gain of

Aboriginal students of around 0.1 standard deviations.10 This is a sizeable effect, and it

corresponds to a decrease of around 20% in the gap between Aboriginal and non-Aboriginal

students in British Columbia. The corresponding estimates for the numeracy exam indicate

no statistically significant relationship between the proportion of ESD students in a district

and value-added numeracy exam outcomes.

These results in Table 3.7 suggest that ESD funding has been used in B.C. to support

services that are effective. Districts that increased the proportion of Aboriginal students that

they deemed eligible for ESD funding saw improved reading exam outcomes for Aboriginal

students. These improved outcomes can be attributed to a large direct effect on program

participants, or to a combination of direct and indirect effects. Given the magnitude of the

IV estimate, the latter interpretation is more plausible.

9Estimates from the first-stage regressions are presented in Table A.3.10On average, districts that offer ESD assign 22 percent of their grade 4 Aboriginal students to ESD; the

estimated effect of such a program is (0.48)(0.22)=0.11.


3.5.3 Test Score Gains: Quantile Regression Results

A natural additional step is to investigate the effects of ESD funding across the distribution

of reading and numeracy test score gains. In particular, ESD services are meant to tar-

get students with weak Standard English skills, so we might expect the improved reading

outcomes documented in the previous section to appear primarily in the lower end of the

outcome distribution. We use quantile regression to develop some evidence on this question.

Table 3.8 presents regression results for three quantile regressions, for the 25th percentile,

median and 75th percentile. These regressions have the same specifications as the reduced

form regressions in Table 3.7, and include district and year dummy variables. Standard er-

rors are estimated by the simple bootstrap, i.e. without adjustment for clustering, and will

be biased downwards. Caution should also apply to interpreting the coefficient estimates.

The econometric model described above does not imply that the conditional quantiles iden-

tify causal effects. The results in Table 3.8 provide additional information on the statistical

relationships underlying our main effect estimates, and are thus suggestive but not definitive

about the effects of ESD programs.

The results in Table 3.8 are consistent with the large average effects on reading test score

gains discussed in the previous section, and suggest that the effects of ESD are stronger at the

bottom of the reading test score gain distribution. For the bottom quartile, the coefficient

on district % ESD is around 0.6 standard deviations. The corresponding coefficient for

the median quartile is also larger than the average effect presented in Table 3.7, while the

coefficient for the top quartile is smaller. The results in Table 3.8 for numeracy are also

consistent with the results in Table 3.7: ESD funding has little relationship with numeracy

test score gains.

3.5.4 Test Score Gains: non-Aboriginal Students

Our analysis above essentially interprets any association between the introduction or expan-

sion of ESD and movements in the test scores of Aboriginal students as evidence of a causal

effect of ESD funding. This interpretation is problematic if ESD is correlated with unob-

served district policy changes that affect test scores. While we cannot rule out all alternative

explanations of this form, we can evaluate whether some consequences of our interpretation

hold in the data. As shown in Table 3.7, we find that ESD improves Aboriginal students’


reading scores but has no effect on their numeracy scores. Similarly we would expect to find

little effect, if any, of ESD on the reading or numeracy scores of non-Aboriginal students.

Results from our reduced form model for non-Aboriginal students in Table 3.9 show that

the proportion of Aboriginal students assigned to ESD within a district has no substantial

or statistically significant effect on the reading or numeracy test scores of non-Aboriginal

students, in either numeracy or reading.

3.6 Conclusion

While there is widespread agreement among researchers and policymakers about the impor-

tance of improving educational outcomes for low-achieving students, there is little consensus

about which policies are most likely to bring about the desired improvements. Our results in-

dicate that supplemental funding targeted towards Aboriginal students in British Columbia

who speak non-standard forms of English provides highly effective support for their literacy

skill development. Since non-standard dialect speakers in other jurisdictions arguably face

similar challenges at school, these results may be of considerable interest to policy-makers

elsewhere.

However, while our approach produces a clear result with respect to the success of this

policy, it does not reveal the mechanism through which that success is achieved. Our results

can therefore provide little guidance to educators who are developing specific programs and

services for non-standard dialect speakers. Finally, we cannot rule out the possibility that it

is not the funding per se, but rather changes to leadership or priority-setting that coincided

with changes in ESD designation rates The clear lesson that we can draw from our results is

that targeting additional attention and resources to Aboriginal students within the education

system can have a substantial effect on their learning outcomes.


3.7 Tables and Figures

3.7.1 Tables

Table 3.1: Per student operating grants to B.C. public school districts.

Category before March 2002 after March 2002

Base amount 3,042 5,308

Aboriginal supplement 755 – 1,030* 950

ESL/ESD supplement 1,230 (Year 1) 1,1001,060 (Years 2-5)

Special needs supplements

Dependent 31,910 30,000

Low incidence/high cost 12,460 15,000

Severe behaviour 6,014 6,000

High incidence/low cost 3,132 0

Gifted 341 0

*Amount per student depends on total number of Aboriginal students in the District.

Source: British Columbia Ministry of Education (2002), page 4.

Table 3.2: Participation in FSA exams, grade 7 students 2002-2004.

Variable Non-Aboriginal Aboriginal Aboriginal in ESD

# of observations 125,956 13,414 1,206

% of total 90.4 9.6 0.87

% Aboriginal peers 8.1 23.9 46.0

% speaking non-English 21.2 1.0 5.1

language at home

% currently ESL/ESD 5.8 9.0 100.0

% disabled 7.9 19.0 30.8

% gifted 2.6 0.8 0.3

Source: B.C. Ministry of Educations enrolment and FSA exam database.


Table 3.3: Achievement levels and growth, grade 7 students 2002-2004.

Variable Non-Aboriginal Aboriginal Aboriginal in ESD

Grade 7 numeracy score 0.01 -0.57 -0.99

Grade 7 reading score 0.01 -0.58 -1.26

Gain in numeracy score -0.05 -0.09 0.02

Gain in reading score 0.00 -0.08 -0.13


Table 3.4: Participation in FSA exams, grade 7 students 2002-2004.

Variable Non-Aboriginal Aboriginal Aboriginalin ESD

% taking grade 7 numeracy 90.2 77.2 61.6

% taking grade 7 reading 91.1 80.4 66.3

% excused grade 7 numeracy 4.5 11.2 21.8

% excused grade 7 reading 4.5 10.4 21.3

% without numeracy gain data 12.6 28.4 48.1

% without reading gain data 11.1 25.5 43.6


Table 3.5: District ESD programs and students in ESD, grade 7 Aboriginal students 2002-2004.

Aboriginal students in ESD Districts with > 5% ESD*

Percent Number Percent Number

1999 2.7 121 6.8 4

2000 3.4 152 6.8 4

2001 3.8 165 6.8 4

2002 6.5 292 11.9 7

2003 9.5 400 18.6 11

2004 10.7 529 27.1 16∗Districts where at least 5% of (and at least 10)

grade 7 Aboriginal students are in ESD.

Source: B.C. Ministry of Educations enrolment

and FSA exam database.


Table 3.6: Effect of ESD programming on grade 7 exam participation, grade 7 Aboriginalstudents 2002-2004.

Exam Participation

Numeracy Reading

Variable OLS IV RF OLS IV RF

Grade 4 0.42*** 0.42*** 0.42*** 0.42*** 0.42*** 0.42***exam participant (0.02) (0.02) (0.02) (0.02) (0.02) (0.02)

Years in ESD -0.08*** 0.05 -0.07*** 0.03* Grade 4 Partic. (0.02) (0.06) (0.01) (0.07)

District % ESD 0.27* 0.10 0.20 0.07*Grade 4 Partic. (0.14) (0.12) (0.14) (0.13)

Yrs in ESD -0.05*** 0.07* -0.05** 0.05*Grade 4 non-Partic (0.02) (0.04) (0.01) (0.05)

District % ESD 0.46** 0.28* 0.41* 0.21*Grade 4 non-Partic (0.19) (0.16) (0.21) (0.17)

Male -0.03*** -0.03*** -0.03*** -0.04*** -0.04*** -0.04***

(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)

# Students 10291 10291 10291 10291 10291 10291

# Districts 59 59 59 59 59 59

Standard errors (clustered by district) in parentheses.

*** p<0.01, ** p<0.05, * p<0.1

All regressions include district and year fixed effects.

IV regressions use district % ESD as an instrument for years in ESD.



Table 3.7: Effect of ESD programming on exam results, grade 7 Aboriginal students 2002-2004.

Test Score Gains

Numeracy Reading


Years in ESD 0.06*** -0.09 -0.02 0.33*

(0.02) (0.27) (0.02) (0.18)

District % ESD -0.22 -0.13 0.51* 0.48*

(0.39) (0.38) (0.27) 0.27)

Male -0.02 -0.01 -0.02 -0.10*** -0.10*** -0.10***

(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)

First-stage F-statistic 48.00 31.70

(p-value) (0.00) (0.00)

# Students 7507 7507 7507 7803 7803 7803

# Districts 59 59 59 59 59 59


*** p<0.01, ** p<0.05, * p<0.1




Table 3.8: Quantile regressions for reduced form effect of ESD programming on exam results,grade 7 Aboriginal students 2002-2004.

Test Score Gains

Numeracy Reading

Quantiles Quantiles

Variable .25 .50 .75 .25 .50 .75

District % ESD -0.22 0.17 0.17 0.59** 0.54* 0.34

(0.52) (0.29) (0.38) (0.29) (0.29) (0.29)

Male -0.03 -0.00 -0.00 -0.09*** -0.07*** -0.10***

(0.03) (0.03) (0.02) (0.02) (0.02) (0.03)

# Students 7510 7510 7510 7806 7806 7806

# Districts 59 59 59 59 59 59

Bootstrap Standard Errors (not clustered) in parentheses.

*** p<0.01, ** p<0.05, * p<0.1




Table 3.9: Reduced form effect of ESD programming on exam results, grade 7 non-Aboriginalstudents 2002-2004.

Test Score Gains

Variable Numeracy Reading

District % ESD 0.05 0.03

(0.20) (0.22)

Male 0.02*** -0.09***

(0.01) (0.01)

# Students 89697 91219

# Districts 59 59


*** p<0.01, ** p<0.05, * p<0.1


Source: B.C. Ministry of Educations enrolment



3.7.2 Figures

0 20 40 60 80Percentage of Aboriginal students in SESD

2004

2003

2002

StikineCowichan Valley

Gold TrailSunshine Coast

Surrey

AlberniHaida Gwaii

QuesnelCariboo-Chilcotin

Nisga'aCampbell River

Prince Rupert


Figure 3.1: Percentage of grade 7 Aboriginal students in ESD (SESD stands for “StandardEnglish as a Second Dialect”, equivalent to ESD) in first year that district assigns morethan 5% to ESD (and more than 10 grade 7 students in total), 2002-2004.


020

4060

Per

cent

age

of A

borig

inal

stu

dent

s in

SE

SD

Vancouver Campbell River Gold Trail Nisga'a

1999

2000

2001

2002

2003

2004

1999

2000

2001

2002

2003

2004

1999

2000

2001

2002

2003

2004

1999

2000

2001

2002

2003

2004


Figure 3.2: Percentage of grade 7 Aboriginal students in ESD 1999-2004, selected districts.

Bibliography

Abowd, J. M., R. H. Creecy, and F. Kramarz (2002, April). Computing person and firm effects using

linked longitudinal employer-employee data. Center for Economic Studies, U.S. Census Bureau,

Technical paper No. TP-2002-06.

Abowd, J. M., F. Kramarz, and D. N. Margolis (1999). High wage workers and high wage firms.

Econometrica 67 (2), 251 – 333.

Abraham, K. G. and H. S. Farber (1987). Job duration, seniority, and earnings. American Economic

Review 77 (3), 278 – 297.

Abramowitz, M. and I. Stegun (1972). Handbook of mathematical functions. Dover Publications,

Inc., New York.

Addessi, W. and R. Tilli (2009). Labor market reforms as a source of the recent italian puzzle.

Journal of European Economy 8 (1), 5 – 20.

Adger, C. (2008). Cited by Naush Boghossian in “LAUSD facing challenge of English at home,

class”. Los Angeles Daily News of February 17, 2008 .

Altonji, J. G., C.-I. Huang, and C. R. Taber (2010). Estimating the cream skimming effect of school

choice. NBER Working Papers: 16579.

Altonji, J. G. and R. A. Shakotko (1987). Do wages rise with job seniority? Review of Economic

Studies 54 (3), p437 – 459.

Anastasia, B., L. Bertazzon, V. Fincati, M. Gambuzza, M. Rasera, and G. Savini (2009). Immi-

grazione straniera in Veneto - Rapporto 2009. Technical report, Regione del Veneto e Veneto

Lavoro.

Angrist, J. D. and V. Lavy (2001). Does Teacher Training Affect Pupil Learning? Evidence from

Matched Comparisons in Jerusalem Public Schools. Journal of Labor Economics 19 (2), 343 – 369.

88

BIBLIOGRAPHY 89

Arcidiacono, P., G. Foster, N. Goodpaster, and J. Kinsler (2011, March). Estimating spillovers

using panel data, with an application to the classroom. Mimeo, Department of Economics, Duke

University.

Bagger, J. and R. Lentz (2008). An empirical model of wage dispersion

with sorting. University of Wisconsin-Madison Working Paper, available from

http://www.ssc.wisc.edu/ rlentz/ressources/wagedisp.pdf.

Ball, J., B. Bernhart, and J. Deby (2006). “First Nations English Di-

alects: Exploratory project proceedings”. Retrieved July 4 2009 from

http://www.earlylearning.ubc.ca/documents/2006/Dialects-Proceedings-Ap-06bb.pdf.

Bandiera, O., I. Barankay, and I. Rasul (2009). Social connections and incentives in the workplace:

Evidence from personnel data. Econometrica 77 (4), 1047 – 1094.

Barrow, L. and C. E. Rouse (2004). Using market valuation to assess public school spending. Journal

of Public Economics 88 (9-10), 1747 – 1769.

Battu, H., C. R. Belfield, and P. J. Sloane (2003). Human capital spillovers within the workplace:

Evidence for great britain. Oxford Bulletin of Economics and Statistics 65 (5), 575 – 594.

Beccarini, A. (2009). The impact of labour market partial reforms on workers’ productivity: The

italian case. International Journal of Applied Economics 6 (2), 1 – 9.

Becker, B. E. and M. A. Huselid (1992). The incentive effects of tournament compensation systems.

Administrative Science Quarterly 37 (2), 336–350.

Boning, B., C. Ichniowski, and K. Shaw (2007). Opportunity counts: Teams and the effectiveness of

production incentives. Journal of Labor Economics 25 (4), 613 – 650.

Bramoulle, Y., H. Djebbari, and B. Fortin (2009). Identification of peer effects through social

networks. Journal of Econometrics 150 (1), 41 – 55.

British Columbia Ministry of Education (1985). A Manual of Policies, Procedures and Guidelines.

Victoria, B.C.

British Columbia Ministry of Education (1999). ESL: A Guide for ESL Specialists 1999. Retrieved

June 7 2009 from www.bced.gov.bc.ca/esl/policy/special.pdf.

British Columbia Ministry of Education (2002). 2002/03 Operating Grants Manual to British

Columbia School Boards. www.bced.gov.bc.ca/k12funding/funding/02-03/estimates/operating-

grants-manual.pdf.

BIBLIOGRAPHY 90

British Columbia Ministry of Education (2009). English as a Second Language Policy and Guidelines.

Retrieved on July 2 2009 from www.bced.gov.bc.ca/esl/policy/guidelines.pdf.

Brock, W. A. and S. N. Durlauf (2001). Discrete choice with social interactions. Review of Economic

Studies 68 (2), 235 – 260.

Brock, W. A. and S. N. Durlauf (2003). Multinomial choice with social interactions. National Bureau

of Economic Research, Inc, NBER Technical Working Paper: 288.

Card, D. (1999). The Causal Effect of Education on Earnings, pp. 1801 – 1863. U CA, Berkeley:

Handbook of Labor Economics Vol. 3A.

Card, D., F. Devicienti, and A. Maida (2010). Rent-sharing, holdup, and wages: Evidence from

matched panel data. National Bureau of Economic Research, Inc, NBER Working Papers: 16192.

Cariboo-Chilcotin School District No. 27 (2009). ESL/ESD in School District 27. Retrieved on June

6, 2009 from http://www.sd27.bc.ca/.

Chan, T. Y., J. Li, and L. Pierce (2012). Compensation and peer effects in competing sales teams.

SSRN Working Paper Series.

Cingano, F. (2003). Returns to specific skills in industrial districts. Labour Economics 10 (2), 149 –

164.

Cohen, J. (2007). Causes and consequences of special education placement: Evidence from Chicago

public schools. Brookings Institution Working Paper, Washington D.C.

Colombo, A. and G. Sciortino (2004). Italian immigration: the origins, nature and evolution of

italy’s migratory systems. Journal of Modern Italian Studies 9 (1), 49 – 70.

Contini, B. (2002). Osservatorio sulla Mobilita del Lavoro in Italia. Il Mulino.

Contini, B. and U. Trivellato (2005). Eppur si muove. Dinamiche e Persistenze nelo Mercato del

Lavoro Italiano. Bologna: Il Mulino.

Cowley, P. and S. T. Easton (2003). Report Card on British Columbia’s Elementary Schools 2003

Edition. Vancouver: The Fraser Institute.

Davis, S. J. and J. Haltiwanger (1991). Wage dispersion between and within u.s. manufacturing

plants, 1963-86. Brookings Papers on Economic Activity , 115 – 180.

Dee, T. S. (2004). Are There Civic Returns to Education? Journal of Public Economics 88 (9-10),

1697 – 1720.

BIBLIOGRAPHY 91

Dittrich, M., A. Knabe, and K. Leipold (2011). Spillover effects of minimum wages: Theory and

experimental evidence. CESifo Group Munich, CESifo Working Paper Series: 3576.

Dostie, B. (2005). Job turnover and the returns to seniority. Journal of Business and Economic

Statistics 23 (2), p192 – 199.

Dunne, T., J. Haltiwanger, and L. Foster (2000). Wage and productivity dispersion in u.s. manu-

facturing: The role of computer investment. National Bureau of Economic Research, Inc, NBER

Working Papers: 7465.

Dustmann, C. and C. Meghir (2005). Wages, experience and seniority. Review of Economic Stud-

ies 72 (1), p77 – 108.

Eades, D. and J. Siegel (1999). Changing attitudes towards Australian Creoles and Aboriginal English,

pp. 265–278. J.R. Rickford and Romaine (Eds.), Creole genesis, attitudes and discourse: Studies

celebrating Charlene J. Sato. Amsterdam: John Benjamins.

Edin, P.-A., B. Holmlund, and O. N. Skans (2008). Wage dispersion between and within plants:

Sweden 1985-2000. NBER Working Paper 13021.

Eeckhout, J. and P. Kircher (2009). Identifying sorting - in theory. Penn Institute for Economic

Research, Department of Economics, University of Pennsylvania, PIER Working Paper Archive.

Erickson, C. and A. Ichino (1993). Wage differentials in Italy: market forces, institutions, and

inflation. IGIER Working Papers no. 34.

Falk, A. and A. Ichino (2006). Clean evidence on peer effects. Journal of Labor Economics 24 (1),

39 – 57.

Figlio, D. N. and L. S. Getzler (2006). Accountability, Ability and Disability: Gaming the System?

Advances in Applied Microeconomics 14, 35–49.

Friesen, J. and B. Krauth (2010). Sorting, peers, and achievement of aboriginal students in british

columbia. Canadian Journal of Economics 43 (4), 1273 – 1301.

Greene, W. H. (2003). Econometric Analysis (5th ed.). Prentice Hall.

Greenwood, D. T. (1997). New developments in the intergenerational impact of education. Interna-

tional Journal of Educational Research 27 (6), 503 – 11.

Gruetter, M. and R. Lalive (2009). The importance of firms in wage determination. Labour Eco-

nomics 16 (2), 149 – 160.

BIBLIOGRAPHY 92

Guryan, J. (2001). Does Money Matter? Regression-Discontinuity Estimates from Education Finance

Reform in Massachusetts. National Bureau of Economic Research, Inc, NBER Working Papers:

8269.

Guryan, J., K. Kroft, and M. J. Notowidigdo (2009). Peer effects in the workplace: Evidence

from random groupings in professional golf tournaments. American Economic Journal: Applied

Economics 1 (4), 34 – 68.

Haida Gwaii/Queen Charlotte School District No. 50 (2008).

Achievement Contract. Retrieved on June 6, 2009 from

http://www.sd50.bc.ca/index.php/publisher/file/action/download/frmAssetID/468/.

Hamilton, B. H., J. A. Nickerson, and H. Owan (2003). Team incentives and worker heterogeneity:

An empirical analysis of the impact of teams on productivity and participation. Journal of Political

Economy 111 (3), 465 – 497.

Hanushek, E. A., J. F. Kain, and S. G. Rivkin (2002). Inferring Program Effects for Special Pop-

ulations: Does Special Education Raise Achievement for Students with Disabilities? Review of

Economics and Statistics 84 (4), 584 – 599.

Harbring, C. and B. Irlenbusch (2003). An experimental study on tournament design. Labour

Economics 10 (4), 443 – 464.

Health Canada, First Nations and Inuit Branch (2003). A Statistical Profile of the Health of First

Nations in Canada. Ottawa.

Hellerstein, J. K. and D. Neumark (2008). Workplace Segregation in the United States: Race,

Ethnicity, and Skill. Review of Economics and Statistics 90 (3), 459 – 477.

Ichino, A. and G. Maggi (2000). Work Environment and Individual Background: Explaining Regional

Shirking Differentials in a Large Italian Firm. Quarterly Journal of Economics 115 (3), 1057 –

1090.

Istat (2005). 14mo censimento della popolazione e delle abitazioni del 2001. Summary Tables

available from http://dawinci.istat.it/, last accessed on September 18, 2011.

Istat (2009). Prezzi al consumo: dati. Istituto Nazionale di Statistica,

http://www.istat.it/prezzi/precon/dati/.

Istat (2010). Natalita e fecondita della popolazione residente: caratteristiche e tendenze recenti.

Available online from http://www.istat.it/it/archivio/38402.

BIBLIOGRAPHY 93

Istat (2011). Tavole regionali. http://www.istat.it/it/archivio/12973, Last accessed on October 21,

2011.

Jacob, B. A. (2005). Accountability, Incentives and Behavior: The Impact of High-Stakes Testing

in the Chicago Public Schools. Journal of Public Economics 89 (5-6), 761 – 796.

Jovanovic, B. (1979). Job matching and the theory of turnover. Journal of Political Economy 87 (5),

p972 – 990.

Jovanovic, B. (1984). Matching, turnover, and unemployment. Journal of Political Economy 92 (1),

108 – 122.

Kambourov, G. and I. Manovskii (2009). Occupational specificity of human capital. International

Economic Review 50 (1), 63 – 115.

Kenkel, D. S. (1991). Health Behavior, Health Knowledge, and Schooling. Journal of Political

Economy 99 (2), 287 – 305.

Kiefer, N. M. (1988). Economic duration data and hazard functions. Journal of Economic Litera-

ture 26 (2), p646 – 679.

Kramarz, F., S. Lollivier, and L.-P. Pele (1996). Wage inequalities and firm-specific compensation

policies in france. Annales d’Economie et de Statistique (41-42), 369 – 386.

Kremer, M. (1993). The o-ring theory of economic development. Quarterly Journal of Eco-

nomics 108 (3), 551 – 575.

Kremer, M. and E. Maskin (1996). Wage inequality and segregation by skill. NBER Working Paper

5718.

Kronenberg, K. and T. Kronenberg (2011, March). Keeping up with the Joneses by finding a better-

paid job - The effect of relative income on job mobility. MPRA Paper No. 29426, posted 08.

March 2011 / 13:22.

Kurtulus, F. A. (2011). What types of diversity benefit workers? empirical evidence on the effects

of co-worker dissimilarity on the performance of employees. Industrial Relations 50 (4), 678 – 712.

Lavy, V. and A. Schlosser (2005). Targeted Remedial Education for Underperforming Teenagers:

Costs and Benefits. Journal of Labor Economics 23 (4), 839 – 874.

Lazear, E. P. and S. Rosen (1981). Rank-order tournaments as optimum labor contracts. Journal of

Political Economy 89 (5), 841 – 864.

BIBLIOGRAPHY 94

Lengermann, P. (2002, June). Is it who you are, where you work, or with whom you work? reassessing

the relationship between skill segregation and wage inequality. U.S. Census Bureau, LEHD Project

Technical paper No. TP-2002-10.

Leombruni, R. and R. Quaranta (2011). La codifica di settore in whip - problemi correnti e studio

di un algoritmo di ricostruzione della codifica ateco 2002. WHIP Technical Reportno.2/2011,

Laboratorio Riccardo Revelli.

Lillard, L. A. (1999). Job turnover heterogeneity and person-job-specific time-series wages. Annales

d’Economie et de Statistique (55-56), p183 – 210.

Lillard, L. A. and C. Panis (2003a). aml multilevel multiprocess statistical software, release 2.0.

Lillard, L. A. and C. Panis (2003b). aML User Guide and Reference Manual. EconWare, Los Angeles,

California.

Lochner, L. and E. Moretti (2004). The Effect of Education on Crime: Evidence from Prison Inmates,

Arrests, and Self-Reports. American Economic Review 94 (1), 155 – 189.

Lopes de Melo, R. (2009a). Estado de uma nacao. chapter Emparelhamento no Mercado de Trabalho

Brasileiro: Evidencia e uma Explicacao. Ipea.

Lopes de Melo, R. (2009b). Sorting in the labor market: Theory and measurement.

Los Angeles Unified School District (2009). Academic English Mastery and Closing the Achievement

Gap. Retrieved on June 6, 2009 from http://www.lausd.net/District 5/aemp.htm.

Machin, S. and S. McNally (2008). The Literacy Hour. Journal of Public Economics 92 (5-6), 1441

– 1462.

Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem. Review of

Economic Studies 60 (3), p531 – 542.

Marshall, A. (1890). Principles of Economics. New York: Macmillan.

Mas, A. and E. Moretti (2009). Peers at work. American Economic Review 99 (1), 112 – 145.

Mendelson, M. (2006). Aboriginal Peoples and Post-Secondary Education in Canada. Ottawa: Cale-

don Institute of Social Policy.

Mendes, R., G. J. van den Berg, and M. Lindeboom (2010). An empirical assessment of assortative

matching in the labor market. Labour Economics 17 (6), 919 – 929.

BIBLIOGRAPHY 95

Moretti, E. (2004). Workers’ education, spillovers, and productivity: Evidence from plant-level

production functions. American Economic Review 94 (3), 656 – 690.

Mortensen, D. T. and R. Wright (2002). Competitive pricing and efficiency in search equilibrium.

International Economic Review 43 (1), p1 – 20.

Navon, G. (2010). Human capital spillovers in the workplace: Labor diversity and productivity.

Israel Economic Review 8 (1), 69 – 90.

Neal, D. (1995). Industry-specific human capital: Evidence from displaced workers. Journal of Labor

Economics 13 (4), p653 – 677.

Nechako Lakes School District No. 91 (2006). English as a Second Dialect 2005/06. Retrieved on

June 6, 2009 from http://netra2.sd91.bc.ca/sd91/student-services/esd2005-06.pdf.

Oakland Unified School District (1997). Resolution No. $597-0063. Retreived on May 7, 2009 from

http://www.linguistlist.org/topics/ebonics/ebonics-res1.html.

Occari, F. and S. Pitingaro (1997). Demografia di impresa e mobilita del lavoro:una stima della

componente spuria sulla base degli archivi Inps. Consiglio Nazionale delle Ricerche, Working

Paper, Occupazione e livelli di attivita in Italia.

Ouazad, A. (2007). Program for the estimation of two-way fixed effects. available at

http://personal.lse.ac.uk/ouazad/.

Parent, D. (2000). Industry-specific capital and the wage profile: Evidence from the national longitu-

dinal survey of youth and the panel study of income dynamics. Journal of Labor Economics 18 (2),

p306 – 323.

Pavan, R. (2011). Career choice and wage growth. Journal of Labor Economics 29 (3), 549 – 587.

Peters, M. (2010). Noncontractible heterogeneity in directed search. Econometrica 78 (4), 1173 –

1200.

Pissarides, C. A. (1994). Search unemployment with on-the-job search. Review of Economic Stud-

ies 61 (3), p457 – 475.

Pollard, J. and E. Valkovics (1992). The gompertz distribution and its applications. Genus.

Postel-Vinay, F. and J.-M. Robin (2002). Equilibrium wage dispersion with worker and employer

heterogeneity. Econometrica 70 (6), 2295 – 2350.

Ramirez, D. J., T. G. Wiley, G. de Klerk, E. Lee, and . Wayne E. Wright, eds. (2005). Ebonics: The

Urban Education Debate. Multilingual Matters, Ltd.

BIBLIOGRAPHY 96

Richards, J. and A. Vining (2004). Aboriginal Off-reserve Education: Time for Action. Toronto:

C.D. Howe Institute.

Rickford, J. R. (1999). Language diversity and academic achievement in the education of African

American students – An overview of the issues Making the Connection: Language and Academic

Achievement among African American Students. . McHenry, ILL: Center for Applied Linguistics

and Delta Systems Co., Inc.

Royal Commission on Aboriginal Peoples (1996). Report of the Royal Commission on Aboriginal

Peoples. Ottawa: Indian and Northern Affairs Canada.

Schindler, M. (2009). The italian labor market: Recent trends, institutions, and reform options.

International Monetary Fund, IMF Working Papers: 09/47.

Shimer, R. and L. Smith (2000). Assortative matching and search. Econometrica 68 (2), 343 – 369.

Shvydko, T. (2007, November). Interactions at the Workplace: Peer Effects in Earnings. Department

of Economics, University of North Carolina, Mimeo, available from http://www.unc.edu/ tshvy-

dko/.

Tattara, G. and M. Valentini (2010). Turnover and Excess Worker Reallocation: The Veneto Labour

Market between 1982-1996. Labour 24 (4), 474 – 500.

Topel, R. H. (1991). Specific capital, mobility, and wages: Wages rise with job seniority. Journal of

Political Economy 99 (1), p145 – 176.

Topel, R. H. and M. P. Ward (1992). Job mobility and the careers of young men. Quarterly Journal

of Economics 107 (2), 439 – 479.

Troske, K. R. (1999). Evidence on the employer size-wage premium from worker-establishment

matched data. Review of Economics and Statistics 81 (1), 15 – 26.

Vancouver Island North School District No. 85 (2008). District Literacy Plan 2008-2009. Retrieved

on June 6, 2009 from http://www.sd85.bc.ca/.

Wolfe, B. and R. Haveman (2001). Accounting for the social and non-market benefits of education.

The Contribution of Human and Social Capital to Sustained Economic Growth and Well-being.

Vancouver: University of British Columbia Press.

Woodcock, S. D. (2008). Wage differentials in the presence of unobserved worker, firm, and match

heterogeneity. Labour Economics 15 (4), 772–794.

Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and Panel Data. MIT Press.

Appendix A

Tables

A.1 Distribution of Firms by Sector in WHIP

Table A.1: Economic sector of the firm

Distribution of workersEconomic sector Males FemalesFarming, hunting, fishing 0.3 0.1Food, beverages and Tobacco 3.5 4.0Textile and clothing 2.3 8.5Leather and fur 1.3 2.0Wood Industries 1.8 0.5Paper and press 1.6 1.3Chemicals and related industries 1.3 1.0Rubber and plastic industries 1.7 1.2Products for working non-metal minerals 1.9 0.9Production of metals and metal products 9.5 3.2Production of machineries 3.4 1.3Production of electrical and optical machinery 4.6 3.5Production of means of transportation 1.3 0.5Other manufacturing 3.3 2.3Construction 18.2 1.9Wholesale and retail trade, auto reparations 13.8 19.9Hotels and restaurants 7.9 12.1Transport, warehouses and communications 5.2 2.4Banking and financial intermediaries 10.0 16.0Computing and rental services 1.0 1.8Public Administration, Social and personal services 5.4 15.0Missing values 0.4 0.4Source: Author’s calculations from WHIP dataset.

97

APPENDIX A. TABLES 98

A.2 Distribution of Firms by Sector in VWH

Table A.2: Firm sector

Economic sector of the firm Frequency Percentage of all firmsCommerce, bars and hotels 64,825 28.04Transport and communications 8,196 3.55Banking and insurances 13,586 5.88Public administration and other services 23,448 10.14Extraction of solid fuels 650 0.28Oil and gas extraction 43 0.02Oil and coal industries 101 0.05Production and distribution of electricity and natural gas 147 0.06Water industries 76 0.03Other extractive industries 153 0.07Extraction and processing of minerals 680 0.29Production and first transformation of metals 1,094 0.47Non-metal material processing 3,739 1.62Chemical industries 1,808 0.78Production of artificial fibers 41 0.02Other metal manufacturing 346 0.15Construction of metal products 16,569 7.17Construction and installation of machinery 4,877 2.11Construction, installation and repairs of office equipment 1,651 0.71Construction and installation of equipment 4,308 1.86Construction and assembly of vehicles 582 0.25Construction of transportation machinery 730 0.32Construction of clocks and other precision machinery 1,015 0.44Food industry 4,562 1.97Sugar. alcohol and tobacco industries 1,604 0.69Textile industry 3,963 1.71Leather industry 1,458 0.63Shoes and clothing industries 9,573 4.14Wood and wood furniture industries 6,406 2.77Paper and print industries 2,627 1.14Rubber and plastic industries 2,659 1.15Other manufacturing 3,121 1.35Construction 46,557 20.14

Source: VWH data. Sectors coded using the 3 digit Ateco 81 coding system.


A.3 First Stage regression and Exam Participation of Abo-

riginal Students

Table A.3: First stage regression results.

Variable First Stage Regression

Male 0.04*

(0.02)

% ESD in district 2.36***

(0.33)

Observations 10290

R-squared 0.26


*** p<0.01, ** p<0.05, * p<0.1

Includes district and year fixed effects.

Source: B.C. Ministry of Education databases. enrolment



Table A.4: ESD and probability of taking the FSA exam, grade 7 Aboriginal students2002-2004.

Exam Participation

Numeracy Reading


Years in ESD -0.11*** 0.08 -0.09*** 0.05

(0.02) (0.06) (0.02) (0.05)

% ESD in district 0.44*** 0.19 0.34** 0.12

(0.16) (0.13) (0.13) (0.12)

Male -0.05*** -0.05*** -0.05*** -0.06*** -0.07*** -0.07***

(0.01) (0.01) (0.01) (0.01) (0.01) (0.01)

Observations 10291 10291 10291 10291 10291 10291

Standard Errors (clustered by district) in parentheses, *** p<0.01, ** p<0.05, * p<0.1




Appendix B

Technical Appendices

B.1 Iterative procedure for estimating spillover effects

For notational convenience I define the variable yijt, which denotes the dependent variable of my

model net of all fixed effects and covariates that are not a function of the current θ:

yijt ≡ wijt −Xitβ − Fjtκ− ψj − τt (B.1.1)

As shown below, the key for the estimation is to derive the First Order Conditions of (2.2) with

respect to the worker effect θi after having substituted in using equation (B.1.1):

∑t

yijt − θi − η 1

Nijt∼i

∑p∈Nijt

θp

+∑

p∈Nijt

η1

Nijt∼i

yijt − θp −η 1

Nijt∼i

∑k∈Nijt∼p

θk

= 0

In order to make this implicit equation for θi operational, I solve the equation above for θi moving

all of the terms including θi to the left-hand side of the equation and then solving for θi:

θi =

∑t

{yijt − η 1

Nijt

(∑p∈Nijt

θp

)+∑p∈Nijt

η 1Nijt

[yijt − θp −

(η 1Nijt

∑k∈Nijt∼p

θk

)]}∑t

(1 + η2 1

Nijt

) (B.1.2)

As discussed below, the person fixed effects that are on the right-hand side of the equation above

are those of the previous iteration, and get updated after each θi is updated using equation (B.1.2).

As a consequence, even though my model includes different and additional fixed effects, Theorem 2

101

APPENDIX B. TECHNICAL APPENDICES 102

in Arcidiacono et al. (2011) applies here, since the additional estimated coefficients do not depend

on theta and thus can be viewed as part of the dependent variable at each iteration. Theorem 2

shows that equation (B.1.2) is a contraction mapping, guaranteeing convergence of the estimated

parameters to their NLS counterparts, for any initial vector θ0 if η < 0.41. In particular, unlike

similar two-step procedures such as that developed in Mas and Moretti (2009), in the procedure of

Arcidiacono et al. (2011) measurement error in the covariates does not lead to an attenuation bias

of the regression coefficients. This is due to the fact that the indirect effect of ability on outcomes

through the peer effects is directly accounted for in the estimation procedure.

Arcidiacono et al. (2011) derive this result by stacking the First Order Condition from the

optimization problems for each θ and checking the conditions for the function from one guess at

the vector of individual effects of θ to the next f : θ → θ′ to be a contraction mapping, which is

equivalent to checking the conditions for ρ(f(θ), f(θ′) < βρ(θ, θ′) for some β < 1 and where ρ is a

valid distance function. In each step of the iterative procedure, after having updated each member

of the vector θ using (B.1.2) the procedure updates the firm fixed effect and the year by sector fixed

effect averaging the residuals for each observation over the relevant set of observations, excluding

the fixed effect of interest.

After having updated the vector f individual fixed effects, I can now update the vector of firm

effects

ψj =

∑i∈Nj

[wijt −Xitβ − Fjtκ− θi − η 1

Nijt

(∑p∈Nijt

θp

)− τt

]∑i∈Nj

1(B.1.3)

and that of time effects:

τt =

∑t∈Nt

[wijt −Xitβ − Fjtκ− θi − η 1

Nijt

(∑p∈Nijt

θp

)− ψj

]∑i∈Nt

1(B.1.4)

For updating θi in iteration α I use a modified version of equation (B.1.2) for computational

convenience, using the result in Lemma 2 of Theorem 1 of Arcidiacono et al. (2011):

θαi =

∑t

{η 1Nijt

(∑j∈Nijt

eα−1jt − eα−1it

)+ eα−1it +

(1 + η2 1

Nijt

)θα−1i

}∑t

(1 + η2 1

Nijt

) (B.1.5)

where eit denotes the regression residual from the OLS regression estimates of step 1. Equation

(B.1.5) is obtained from equation (B.1.2) by rearranging terms so as to identify regression residuals

and then substituting them in isolating the terms that include θα−1i . The procedure described

1The result in Arcidiacono et al. (2011) is not a bivariate relationship, so that the result may hold forvalues larger than 0.4 as well, depending on the size of peer groups.


above lower the sum of squared residuals in each iteration, and can therefore be performed until a

predetermined criterion for convergence is reached2.

B.2 Structure of the VWH dataset

The VWH dataset is composed of a worker archive, a firm archive and a job archive. I link the job

archive to the worker archive using the worker identifier they share, and the firm archive to the dataset

using the firm identifier. The worker archive includes a person identifier, and very limited individual

information:3 gender, birth date, birth place,4 and residential address.5 Not uncommonly for Italian

administrative datasets, the VWH dataset includes no information on the workers’ education. This

is not crucial for my estimation however, because all of the time-invariant individual characteristics

are captured by the person effect, and could not be separately included even if available. The firm

archive is richer and includes a firm identifier, firm’s name, activity, address, sector,6 establishment

date, cessation date, number of initial employees, area code and postal code of the headquarter.

The job archive includes a worker identifier, a firm identifier, time of work (year, month, week, day),

duration of the employment relationship, place of work, total yearly real wages in 2003 Euros for each

job in each time period, qualification, contract, level including information on whether the contract

part-time and/or fixed term.

For the analysis in this paper it is crucial to have a correct identification of firms, in a cross

sectional as well as dynamic sense. The VWH dataset has been the product of a careful identification

of firms as economic entities and not simply as legal entities. The variable has been constructed using

the same technique as in Occari and Pitingaro (1997). When a majority share of workers of a large

firm moves to another firm the mobility is considered spurious. For small firms in order to be even

more conservative Occari and Pitingaro (1997) also require that location does not change. When

mobility is considered spurious, the two firms are recognised as the same firm. Additional information

on the dataset and in particular on the construction of the two different firm identifiers are available

from Tattara and Valentini (2010).

2In the case of my estimation, that criterion is that the sum of squared residuals differ by less than 10−7

between two consecutive iterations.3Unfortunately this is common for administrative datasets of this kind.4From which I have manually constructed a country of birth variable for foreign born from the place of

birth.5This is often different from the current address since there are virtually no incentives for people to change

it and so the change may be delayed by many years.6Employers are classified according to the three-digit Ateco 1981 standard classification. The author

would like to thank Prof. Giuseppe Tattara for sending all of the information necessary for translating theAteco 1981 coding into meaningful industry codes.


B.3 Sample Restrictions

As briefly discussed in the body of the paper, I force the structure of the dataset to have at most one

observation for each worker in each year, and therefore I eliminate additional observations of each

worker/year. Apart from cases with missing values in the variables used in the regression, the vast

majority of these case are cases in which there are two different records for the same worker in the

same firm, which is the result of the fact that the data is based upon a firm identifier that does not

take mergers and acquisitions into account. For all cases in which a worker is observed more than

once in the same firm in the same year I construct a new relationship that incorporates these different

relationships and drop duplicates. For the cases in which there are still multiple observations per

worker/year I identify a dominant job keeping the employment relationship with the higher number

of days paid. I need to use the naive firm identifier for merging the worker, job and firm archives

because that is the variable that links them together. After merging and for the remainder of the

paper I use the firm identifier that does not treat changes in the firm’s ownership as deaths and

births of a new firm.

My main regression model includes a measure of firm size. The VWH does not include firm size,

so I construct it from the data, counting all employees employes in a certain firm for each year. This

measure may underestimate actual firm size since a firm’s workforce may include undocumented

workers, or may hire professionals that I cannot observe because it is not part of my dataset.

The VWH does not include an experience variable per se, and so I need to construct it: within

the period of my data, I can see the employment history of all workers and so I can use the total

number of months worked to construct a measure of actual labour market experience. However, for

a portion of workers in my sample I cannot observe the start of their labour market careers. For

this purpose, I divide workers into two categories, depending on whether I can assume that I observe

them from the beginning of their careers. I assume that I see their whole careers if they have no

job in the first three years of my dataset and if they are at most 18 years old in 1985, the first year

I have since I ignore the first three. For the workers for whom I assume that I am observing their

whole labour market career, experience will be equal to observed experience, given by the sum of

months in full time employment up to (not including) year t. For workers that I do not see from

the start of their careers, experience is given by observed experience up to year t plus the average

months of experience accumulated by workers of the same category and gender from their average

minimum age of employment up to the first time I see them in my dataset. I divide workers into

white collar and blue collar workers based on their occupation, in order to control for the different

age of entry in the labour force of white collar workers. Each year, male workers work on average

around 10 full-time months if they are white collar workers, around 9.5 months if they are blue collar

workers. Female workers work around 9 full-time months if they are white collar and around 8.5

months if they are blue collar workers. Average age of entry in the labour force is very similar for


male workers and females workers, at around 22 for white collars, 19 for blue collars. Finally, in the

construction of my experience measure I ignore the possible effect unemployment may have on the

depreciation of labour market skills of workers.

B.4 Robustness check: small firms and large firms

In Table B.1 I reports estimates of equation (2.1) on two different subsamples of my population,

that of workers of very small firms and of very large firms. In these estimates I find smaller peer

effects and lower proportion of the overall wage variance that is explained by spillover effects. This

suggests that my main estimates are not driven by very small firms of very large firms alone. The

second column of Table B.1 is estimated using the sample of firms that have less than ten employees

at a given point in time. For this subpopulation of firms, person and firm effects are important while

spillover effects explain around 2.2 percent of all wage variation. A one standard deviation increase

in the average labour market skills of peers is associated with a wage gain of 6.8 percent. Using the

average within-firm standard deviation, the equivalent figure is 2.9 percent.

The third column of Table B.1 shows estimates for the same model for a sample of the largest

firms only. Compared to the full sample, peer effects are smaller in terms of average wage effects:

while a unitary change in the overall standard deviation is associated with a wage increase of 6.2

percent, the estimate using average firm-level standard deviation in “peer quality” is of 1.9 percent.

They are also relatively unimportant in terms of proportion of the overall variation that is explained

by them, i.e. 2.4 percent.


Table B.1: Regression on different samples

Dependent variable: ln(wijt)

ModelsFull Small firms Large firms

Number of employees < 10 > 1, 000

Estimated coefficients of covariatesExperience 0.018 0.023 0.018Experience2 -0.001 -0.001 -0.001Firm size 0.000 0.004 -0.000Coworker quality θ 0.358 0.184 0.340

Fixed effectsσθ 0.389 0.467 0.443σψ 0.205 0.372 0.280στ 0.200 0.171 0.237

Corr(θ, ψ) 0.014 -0.373 -0.009

Variance decompositionPerson effect θ 0.469 0.506 0.551Firm effect ψ 0.134 0.191 0.188Time effect τ 0.058 0.072 0.044Polynomial in experience 0.002 0.023 0.005Firm size 0.010 -0.000 -0.002Spillover effect η 0.049 0.022 0.024Unexplained εijt 0.278 0.187 0.190

Pseudo R2 0.722 0.813 0.810

Standard deviations of θσθ (overall s.d.) 0.218 0.372 0.181σθ,1982 (cross sectional s.d. for 1982) 0.221 0.405 0.163

σθ,1991 (cross sectional s.d. for 1991) 0.201 0.360 0.147

σθ,2001 (cross sectional s.d. for 2001) 0.199 0.382 0.1901

NtJt

∑Jj=1Njtσθ,j 0.089 0.158 0.056

(weighted average of within-firm s.d.)

Nobs 28,115,529 3,933,459 4,224,592Nworkers 3,180,714 1,026,651 683,624Nfirms 231,195 203,543 178

Note 1: for small firms and large firms, my converge criterion is 10−4

Note 2: Samples are restricted to observations in the main connected group


Documents

ESSAYS IN LABOUR ECONOMICS - Summit