Methodological Issues Related to the Analysis ofGender Gaps in Employment, Earnings and

Career Progression∗

Project carried out for the European Commission

Employment and Social Affairs DG

Final Report

Miriam Beblo†, Denis Beninger‡, Anja Heinze§ and François Laisney¶

October 29, 2003

∗Acknowledgements:This project has been financed by the European Commission, Employment and Social AffairsDG. We thank Eurostat for providing us with data from the European Community HouseholdPanel (ECHP). We thank Frank Siebern-Thomas and participants of several meetings at theEuropean Commission for helpful comments. We also thank Mette Deding, Anna d’Addio,Björn Döhring and Monika Oels for their discussion of this report at the workshop on genderpay gaps, September 12, 2003. Their comments are reflected in this final revision. We aregrateful to Doris Weichselbaumer who advised us on the literature and to Franziska Mientusand Susanne Steffes for research assistance. Views expressed represent exclusively the positionsof the authors and do not necessarily correspond to those of the European Commission. Allerrors remain our own.

†ZEW, Mannheim. Beblo@zew.de‡ZEW, Mannheim. Beninger@zew.de§ZEW, Mannheim. Heinze@zew.de¶Université Louis Pasteur, Strasbourg, and ZEW, Mannheim. Laisney@zew.de

1

Contents

1 Introduction 5

2 Analysing gender pay gaps in Europe 62.1 The unadjusted wage gap . . . . . . . . . . . . . . . . . . . . . . 62.2 Approaches and their relative merits . . . . . . . . . . . . . . . . 8

3 Methodological review 113.1 Estimation of wage regression models . . . . . . . . . . . . . . . . 11

3.1.1 OLS-Regression and its limitations . . . . . . . . . . . . . 123.1.2 The sample selection problem . . . . . . . . . . . . . . . . 133.1.3 Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . 153.1.4 Endogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . 173.1.5 The sample selection model by Lewbel . . . . . . . . . . . 213.1.6 Panel wage estimation with selectivity . . . . . . . . . . . 213.1.7 Panel estimation in levels: Wooldridge (1995) . . . . . . . 223.1.8 Panel estimation in differences: Kyriazidou (1997) . . . . 233.1.9 The panel estimator of Lewbel (2002) . . . . . . . . . . . 243.1.10 Summary on selectivity, endogeneity and heterogeneity . . 243.1.11 Further developments . . . . . . . . . . . . . . . . . . . . 25

3.2 Decomposition of the gender wage gap . . . . . . . . . . . . . . . 263.2.1 Decomposition by Oaxaca and Blinder . . . . . . . . . . . 263.2.2 Choice of a non-discriminatory wage structure . . . . . . 303.2.3 Integration of the sample selection correction . . . . . . . 313.2.4 Decomposition by Juhn, Murphy and Pierce . . . . . . . . 343.2.5 Extended decomposition by Juhn, Murphy and Pierce . . 363.2.6 Decomposition by Brown, Moon and Zoloth . . . . . . . . 36

4 Choosing methodologies ... 394.1 ... with regard to the wage regressions . . . . . . . . . . . . . . . 394.2 ... and with regard to the wage decomposition . . . . . . . . . . . 41

5 Data 415.1 Data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415.2 Data constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 425.3 Sample definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

6 Results 446.1 Wage equation estimation . . . . . . . . . . . . . . . . . . . . . . 456.2 Wage decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 46

6.2.1 Oaxaca-Blinder . . . . . . . . . . . . . . . . . . . . . . . . 466.2.2 Decomposition by factors . . . . . . . . . . . . . . . . . . 496.2.3 Juhn-Murphy-Pierce: within-country decomposition . . . 536.2.4 Juhn-Murphy-Pierce: between-country decomposition . . 556.2.5 Brown-Moon-Zoloth . . . . . . . . . . . . . . . . . . . . . 56

2

3

7 Conclusion 56

References 59

App e ndix 1: J uhn- Murp hy- Pi e rce figure s 66

App e ndix 2: D es cr ipti ve s t ati s tic s for t he EU and Ge rmany,a nd e s t imat io n re sul ts f or G er ma ny 75

App e ndix 3 a: Es ti mat ion r esul ts for the EU 91

Appendix 3b: Estimation results for the EU using sectorinformation 109

App e ndix 4a : Es ti ma tion resul ts f or select ed countries 1 23

1

App e ndix 4 b: D ec o mp os i tion by f ac tors for s el e cte d c ountri e s 133

31

Executive summaryIn the European Union women earn on average 16% less than men. To learnabout the factors related to this raw wage gap, it is useful to take a detailed lookat both male and female characteristics. This study contributes to explainingthe difference in earnings by gender and improves the analysis of the factorsrelated to the gender pay gap in the European Union. The typical procedurewhen estimating pay gaps is to first investigate the determinants of women’s andmen’s wages and then use the empirical results in order to draw conclusions ondifferences in the wage formation processes of both sexes. Two methodologicalissues have to be dealt with. First, male and female wage equations have to beestimated consistently. This requires proper treatment of various methodologi-cal problems such as self selection, heterogeneity and endogeneity. Self selectionexists if the working individuals do not form a random subgroup of the sampledpopulation but differ systematically from those not employed. Heterogeneity-biased estimates of the determinants of the wage rate may occur if the wage rateis related to unobserved individual characteristics such as motivation or ability,that are in turn correlated with observed explanatory variables. Endogeneityis a problem if explanatory variables, such as working in the public sector, arenot given exogenously but subject to an individual’s decision and, hence, alsoat risk of being correlated with unobserved factors. Their impact on the wagerate is therefore likely to be estimated with bias when using simple estimators.The second methodological issue concerns the appropriate decomposition of thegender pay gap that allows meaningful interpretation of its components.In this study we propose different techniques to assess the gender pay gap in

the European Union while exploring different estimation methods of the wageequation and different decomposition approaches of the wage gap. We presentwage estimation methods that account for selectivity, endogeneity and hetero-geneity. We decompose the pay gap both at the mean, following Oaxaca (1973)and Blinder (1973), and across the wage distribution, as proposed by Juhn, Mur-phy and Pierce (1993). As the literature on wage equation estimation is veryrich, we concentrate on methods most used in the gender gap literature (i.e.OLS and Heckman) on the one hand, and on the other hand on very recentlydeveloped methods (Lewbel 2002 e.g.). The latter allow, at least theoretically,to take account of the three methodological issues mentioned, in original andinnovative ways.The empirical application, based on the European Community Household

Panel (ECHP) for the five largest European countries (France, Germany, Italy,Spain and the United Kingdom), shows that at most half of the difference inearnings between the sexes can be attributed to differences in characteristics.This confirms the findings of other studies, such as the report Employment inEurope 2002 (European Commission 2002a). However, the size of this endow-ment effect differs considerably between countries and depends on the choice ofestimator. Our results suggest that correcting for self-selection has a significantimpact on both the wage estimates and the pay gap decomposition. Further-more, the results are sensitive to the choice of estimator, that is to the treatmentof self-selection in estimation. We recommend the Lewbel approach because itimposes fewer arbitrary restrictions on the model than the Heckman approach,and performs better in terms of predictions with our data.Another main result of the study is derived from the pay gap decomposition

4

over quantiles of the wage distribution. Remarkable differences are revealedwithin as well as between countries. A further recommendation derived fromour analysis would therefore be to pay careful attention to differences over thewage distribution when drawing policy conclusions. Focusing only on the meanpay gap may conceal politically relevant aspects of the problem.

1 Introduction

Gender gaps in employment, earnings and career progression have been on theEuropean Union’s policy agenda for several decades. Despite engaged com-mitments to promote gender equality, e.g. through the European EmploymentStrategy and national governmental measures, women are still less favoured onthe labour market than men. In 1998 women’s average gross hourly earningsamounted to 76-94 percent of men’s earnings throughout the European Union(European Commission 2002a). Furthermore, women encounter more difficultiesin their career progression and are less likely to be promoted (OECD Employ-ment Outlook 2002). They are under-represented in managerial occupationsand in jobs with a supervisory role.Although the observed or raw difference in wages between men and women

provides an overall picture of the actual gender pay gap, encompassing bothdifferences in endowments and differences in remuneration, it is worthwhile toanalyse the factors related to the raw wage gap in more detail. On average, menand women have different human capital endowments. To start with, men andwomen differ in their education levels and tracks. Second, women and men havequite distinct work histories. Traditionally, women are more likely to interruptgainful employment for family reasons, leading to less work experience com-pared with men. The quasi totality of part-time jobs are occupied by womenand their overall participation rate remains steadily below that of men. Finallythe segregation of men and women with respect to occupation or industry mayexplain a part of the wage gap. Empirical studies consider these differencesby estimating an adjusted pay gap that controls for individual and job charac-teristics.1 However, many published studies fail to take account of importantmethodological issues.The aim of our study is to explain the difference in earnings by gender and

improve the analysis of the factors related to the gender pay gap in the EuropeanUnion.2 Our focus is a methodological one. The typical procedure when estimat-ing pay gaps is to first investigate the determinants of women’s and men’s wagesand then use the empirical results in order to draw conclusions on differences inthe wage formation processes of women and men. Thereby two methodologicalissues have to be dealt with. First, male and female wage equations have to beestimated consistently. This requires proper treatment of various methodologi-cal problems such as self selection, heterogeneity and endogeneity. Self selectionexists if the working individuals do not form a random subgroup of the sample

1The 2002 edition of “Employment in Europe” (European Commission 2002a), for instance,finds that gender differences in these characteristics explain only little of the unadjusted wagegap in most of the EU member states.

2That is, we focus on gender gaps in earnings. Gender gaps in employment are not explicitlylooked at, but the issue is taken account of by considering the selection into employment.Differences in career progression are only considered as they are reflected in the pay gap sincethe ECHP does not allow to study career progression in detail.

5

population but differ systematically from those not employed. Heterogeneity-biased estimates of the determinants of the wage rate may occur if the wage rateis related to unobserved individual characteristics such as motivation or ability,that are in turn correlated with observed explanatory variables. Endogeneityis a problem if explanatory variables, such as working in the public sector, arenot given exogenously but subject to an individual’s decision and, hence, alsoat risk of being correlated with unobserved factors. Their impact on the wagerate is therefore likely to be estimated with bias when using simple estimators.The second methodological issue concerns the appropriate decomposition of thegender pay gap that allows meaningful interpretation of its components.We use a new microeconometric approach to estimate wages for women and

men that takes account of the selection process into the labour market. By ap-plying the Lewbel estimator (Lewbel 2002), we derive consistent wage estimatesfor women and men. Moreover, we are able to provide new empirical evidenceon the gender pay gap in Europe by using different decomposition methodswhile drawing on comparable data for all EU countries. The study thereforecomplements previous work published in Employment in Europe 2002, Employ-ment Outlook 2002 and by the Group of Experts on Gender and Employment(European Commission 2002b).This report starts off with an introductory chapter on the gender pay gaps in

the EU in general and the approaches made to disentangle the various effects re-lated to these gaps. Chapter 3 provides a survey and assessment of the methodsfor estimating wage equations and decomposing gender wage gaps found in theliterature. Within this methodological review we present our preferred Lewbelestimation model and extensions for panel data analysis. Chapter 4 sums upthe methodologies we are applying. Data description is given in chapter 5. Inchapter 6, we provide empirical evidence using the Lewbel approach, comparedwith the results from common econometric procedures. Different decomposi-tions of the wage gap are presented to foster understanding of the reasons forthe observed deviations in earnings. The variety of results obtained by differ-ent econometric procedures, including cross-section and panel data analysis, isillustrated with German wages as an example. In addition, we provide esti-mation and decomposition results based on cross-section Lewbel regressions forthe other four largest EU member states (France, Italy, Spain, and the UnitedKingdom) as well as at the overall EU level. We present the results of theOaxaca-Blinder and Juhn-Murphy-Pierce decompositions for all cross-sectionestimation techniques applied to German data, and based on the Lewbel esti-mation only for the other country samples. (For the wage regression results thereader is referred to Appendix 2 (Germany), Appendix 3a and b (EU level) and4a (France, Italy, Spain, and the United Kingdom).) Chapter 7 concludes.

2 Analysing gender pay gaps in Europe

2.1 The unadjusted wage gap

We start off by taking a look at the raw or unadjusted gender gap in grossearnings across the EU member states (for details on the data source see sec-tion 5). According to the ECHP and as listed in Table 1, the average hourlywage gap between women and men in the EU varies between 26 percent in

6

Table 1: Average raw wage gap between men and women in the EU countries

Country Abs. wage gap Rel. wage gap(in euro) (in %)

Austria 2.13 20.1Belgium 0.98 7.7Denmark 1.72 11.7France 1.44 13.0Finland 1.81 17.7Germany 2.68 21.0Greece 0.72 9.5Ireland 2.48 20.2Italy 0.67 6.8The Netherlands 3.70 22.1Portugal 0.38 6.5Spain 1.37 14.9United Kingdom 3.38 26.5Overall EU 1.82 16.3

Data source: ECHP, country files 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Absolute wage gap = Male minus female gross hourly wages in euro. Relativewage gap = Absolute wage gap / male wage rate.We use the individual and country weights (ratio of sample size and size of the popula-tion 16 years and above) provided in the ECHP for all graphs, statistics and estimationsin the paper.

the United Kingdom and hardly 7 percent in Portugal. Our relative measuresmatch more or less the ranking of countries in the Employment in Europe 2002report (European Commission 2002). The small deviations may be due to thediffering sample definitions of both studies. Whereas our country samples con-tain 25 to 55 year old who report positive earnings and are employed at least8 hours per week, the report of the European Commission includes those aged15 to 64 working 15 hours or more. Interestingly we observe larger pay gapsin the Northern and Central European countries and smaller ones in SouthernEurope. A possible explanation for this pattern is the selection of women intoemployment that may be greater in the South than elsewhere.While these aggregate numbers give us a first idea of the gender pay gap in

the EU countries, they conceal the prevalence of wage differences for differentgroups within each state. Therefore, the country distributions of the gender gapare given in Figure 1. The raw wage gap is displayed by percentiles accordingto the wage rankings of all women and men in the respective country sample.Women and men are ordered separately by their wage rate and then comparedpairwise in each percentile of the distribution. That is, the mean wage of the1 percent lowest paid women is compared with the mean wage of the 1 percentlowest paid men. This gap is displayed as the first observation to the left ofthe panel. Please note that the wage distributions of women and men are nottypically congruent, that is, male wages are usually spread over a wider range.

7

The extent to which male and female wages are spread differently is reflected inthe decile gaps.We can distinguish three main groups of patterns: a decreasing, an increasing

and a U-shaped distribution of the raw wage gap. While the Scandinavian coun-tries Denmark and Finland have increasing wage differences, Southern Europewith Italy and Portugal faces a decreasing pattern. The Netherlands provide anextreme example of a U-shape. As with the average pay differentials, Portugal isthe only country where the wage gap curve drops below 0 (for higher earnings).Figure 2 shows German and EU males’ and females’ wage densities. For

Germany we see that the males’ is shifted to the right, denoting a negativegender pay gap for women. Females’ wage density is more concentrated andless squewed to the right. That explains the U-shaped distribution of the rawgap for Germany.The unadjusted wage gap provides a comprehensive measure of the earnings

inequality between men and women. It summarises both differences in charac-teristics and differences in the remuneration of these characteristics (as well aspotential direct and indirect discrimination). Nevertheless, it may be helpful toinvestigate the components of the gender pay gap separately, particularly if theaim is to better target policy measures at reducing the earnings gap.

2.2 Approaches and their relative merits

The ways in which the factors related to a pay gap are analysed in the literatureare diverse. The crudest approach consists in including a gender dummy in asingle wage regression for women and men. The underlying assumption here isthat female and male wages differ by a fixed amount (shift parameter), but thathuman capital characteristics and other explanatory variables have the sameimpact on women’s and men’s wages.A more flexible approach to investigate the earnings gap has been derived

from human capital theory (Becker 1964), where an individual’s wage rate re-flects her productivity potential based on her human capital characteristics.According to Oaxaca (1973) and Blinder (1973), any wage differential betweentwo groups of people (defined by gender, race, ethnicity etc.) can therefore bedecomposed into two parts. The first is explained by differences in human cap-ital endowments of both groups, the second reflects differences in prices, that isthe remuneration of these endowments. This latter part of the wage differentialis often interpreted as an estimate of wage discrimination.3 Hence, potentialdifferences in the wages of women and men may stem both from differences inhuman capital endowments and other job-related variables (endowment effect)and from a difference in the values that are assigned to women’s and men’s char-acteristics (remuneration effect). The usual endowment factors mentioned in theliterature on gender pay gaps include age, educational attainment, work expe-rience, firm tenure, employment interruptions, occupation, occupational status

3There is an ongoing debate on the interpretation of the gender wage gap or parts of itas discrimination. It is argued that the different endowment of women and men itself mayalready be the result of discrimination because of feedback effects (European Commission2002c). On the other hand, the residual (“unexplained”) part of the gap may still consist ofunobserved differences in human capital characteristics. Due to these ambiguities we prefernot to speak of discrimination, and neither of “explained” or “unexplained” part of the gap,but use the terms endowment and remuneration effect instead (as well as other effects to beintroduced later on).

8

Austria

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Belgium

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Denmark

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Finland

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

France

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Germany

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Greece

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Ireland

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Italy

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

The Netherlands

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Portugal

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Spain

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

UK

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

EU

-0.3

0.1

0.5

0.9

0 10 20 30 40 50 60 70 80 90

percentile

Figure 1: Distribution of the raw wage gap in the EU countriesData source: ECHP, country files 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Male minus female gross hourly wage rate per percentile of the wage distribution.

9

0.0

5.1

.15

dens

ity

0 10 20 30 40hourly wage rate

female wage rate male wage rate

Germany0

.05

.1.1

5de

nsity

0 10 20 30 40hourly wage rate

female wage rate male wage rate

EU

Figure 2: Densities of women’s and men’s wages in Germany and the EU (1998)Data source: ECHP, country data files 1998. Sample of 25-55 year old women andmen, who are employed at least 8 hours per week.Note: Kernel density wage estimation for German men and women using an Epanech-nikov kernel function.

10

and firm variables such as firm size and industry sector. Most empirical studiesestimate an adjusted pay gap that controls for these variables and thereby takeaccount of the observed differences in personal and job characteristics of womenand men. The better the wage determination process can be identified, the moreknowledge about the factors related to the gender pay gap can be gained, andthe better policy measures can be targeted.Since the introduction of the wage decomposition methodology suggested

by Blinder (1973) and Oacaxa (1973), more complex decomposition methodshave been developed that try to shed light on the second, remuneration, part ofthe wage gap by taking also the residual wage distribution into account (Juhn,Murphy and Pierce 1993) or by treating e.g. occupational or sector segregationas endogenous (Brown, Moon and Zoloth 1980). When applying unbiased wageequation models, the basic Oaxaca-Blinder decomposition only requires a singleestimation for the male sample from which the residual remuneration effect canbe derived. Other decomposition approaches generally require separate wageestimations for men and women.However, since each of these gender pay gap decompositions is based on the

estimation of a wage equation, it is of first importance to estimate the wagerate using the most appropriate estimation technique. The following sectiontherefore introduces different methods for both wage estimation and wage de-composition.

3 Methodological review

In this section we review the most important contributions to the gender gapliterature with the aim of providing a methodological overview of both the de-composition approaches and the underlying wage equation estimations. Theexposition of the wage models is restricted to applications found in the litera-ture on wage gaps. Notwithstanding, there exist other and also more advancedwage regression techniques which have not been applied in the literature (onwage gaps). We build on the meta-analysis by Weichselbaumer and Winter-Ebmer (2002) and on the survey by Kunze (2000). The synthesis report by theEU “Group of Experts on Gender and Employment” (European Commission2002c) also provides an extensive review of estimated gender pay gaps through-out Europe. In contrast to the latter, we focus only on the methodologicalaspects of estimating gender wage gaps and do not review the respective esti-mation results.This section is organised in two parts. In the first subsection, different

wage regression models, including the one we propose, are introduced. In thesecond subsection, different decomposition techniques of the gender wage gapare presented with a summary of the main applications found in the literature.

3.1 Estimation of wage regression models

We start with an ordinary least squares regression model, the underlying as-sumptions and limitations of which are discussed in detail. The problems ofsample selection, heterogeneity and endogeneity may be circumvented by alter-native specifications of the wage equation. Finally, we propose the endogenous

11

sample selection model by Lewbel (2002) for cross-section analysis and discusssome extensions to panel data estimation.

3.1.1 OLS-Regression and its limitations

The starting point of our review is a simple wage regression model based on thehuman capital theory (Becker 1964, Mincer 1974). In the original set-up theindividual wage rate is explained by human capital variables such as educationand work experience. Most empirical studies today include also job attributes,labour market features and demographic characteristics. Heckman et al. (2003)provide a profound critique of this approach, underlining the important differ-ences that arise between cohort-based and cross-sectional estimates of returnsto schooling, as well as the crucial roles of expectation formation and sequentialresolution of uncertainty.The wage equations for men and women are specified as follows:

lnWMi = XM

i βM + εMilnWF

i = XFi β

F + εFi (1)

where i indexes individuals within the male and female samples. For simplic-ity the distinction between men and women in the wage equation is neglectedhereafter and a general wage equation is considered. The endogenous variableis the logarithmic wage, lnWi. Vector Xi contains all explanatory variables.The error term represents an iid idiosyncratic error term with mean zero andconstant variance σ2ε.Estimation techniques are generally aimed at providing consistent estima-

tors. The simple wage model specified in the equation system (1) is often es-timated by ordinary least squares. Yet this method only provides consistentcoefficient estimates if the following orthogonality conditions are fulfilled:

E [εi|Xi, I∗i > 0] = 0, (2)

where I∗i denotes a latent index variable which is positive if an individual i isemployed and non-positive otherwise.4

For the orthogonality condition to be satisfied, restrictive assumptions arenecessary. In particular, no misspecification may arise from omitted variables,endogeneity and sample selection. Sample selection is a source of violation of theorthogonality condition. The sample of working people excludes, by definition,those who do not participate in the labour market and therefore may not bea random selection of the overall population. If the participation decision iscorrelated with the earnings function, the expected value of the error termof the latter may not be zero. If, for instance, work experience is positivelyrelated to participation as well as to the wage rate, the coefficients of the wage

4Consistency can be obtained under less restrictive conditions, but condition (2) simplifiesexposition.Consistency is one of the most important asymptotic properties of an estimator. The

property of consistency ensures that the estimation rule will provide an estimated coefficientclose to the true parameter value with a high probability if the sample size is large enough(see Judge et al. (1988, p.83) e.g. for more details on the concept of consistency)..

12

regression are likely to overestimate the return to experience. To correct thisselectivity bias, a sample selection model of earnings is applied which takes theparticipation decision into account.Furthermore, earnings may be determined by several unobservable factors.

Intelligence and motivation are likely to have an effect on wages but they aredifficult to capture. As a consequence, the coefficient estimates of the observedvariables may be biased. This unobserved individual heterogeneity may betaken into account by panel data techniques or by random parameter estima-tion assuming either a continuous distribution (MacFadden 1989) or a discretedistribution (Hoynes 1996). Panel models and their application are discussed inthis chapter.Finally, the endogeneity of explanatory variables is also a frequent source

of specification errors. For instance, work experience may be a function ofprevious earnings. If experience is also correlated with the current wage ratethis implies a simultaneity bias in the wage equation. Besides experience, otherendogenous variables like education may also be affected. To remove that bias,the endogenous variables may be instrumented. An alternative solution is toperform simultaneous estimation of the wage rate and the participation decision.Instrumental variable estimation is picked up in Subsection 3.1.4. Especially,the different types of instrumentation found in empirical analyses of wage gapsare presented.Subsections 3.1.2 to 3.1.4 are voluntarily kept at a simple level of exposition.

The informed reader may move directly to subsection 3.1.5.

3.1.2 The sample selection problem

As stated above, a wage estimation may provide inconsistent coefficient esti-mates of the wage determinants if the sample does not result from exogenousselection. We start with the selection process into the labour market. Let I∗idenote a latent index variable representing individual i’s propensity of partic-ipation. Thus the index variable can also be interpreted as a measure for thepropensity of inclusion in the wage sample. It is assumed that this index vari-able is a function of personal characteristics. This function is assumed linearfor the Heckman and the Propensity score estimators:

I∗i = Viγ + ui, (3)

where Vi represents a vector of human capital and demographic variables whichshould differ from those in the wage equation and ui is an iid idiosyncraticerror term with mean zero and constant variance σ2u. The latent variable is notobservable. Depending on a critical value, mostly set to zero, the participationdecision is:

if I∗i > 0, Viγ + ui > 0 i will participate

otherwise I∗i ≤ 0, Viγ + ui ≤ 0 i will not participate. (4)

The equations (4) illustrate that the sample of individuals whose wages areobserved is not a random sample. It follows that the conditional expectation ofwages is:

13

E [lnWi|Xi, I∗i > 0] = Xiβ +E [εi|Xi, I∗i > 0] . (5)

In most cases the term E [εi|Xi, I∗i > 0] does not equal zero, which is anecessary condition for consistent OLS-estimation.5

If the estimation is based on a non-random sample, other methods thanOLS-estimation have to be applied. In the most commonly used procedure,suggested by Heckman (1979), an artificial regressor is added to the initial wagespecification (1). Indeed, under joint normality of ε and u

E [εi|Xi, I∗i > 0] =ρσεσuE [ui|Xi, I∗i > 0] = ρσε

ϕ (Viγ)

Φ(Viγ), (6)

where ϕ and Φ denote the standard normal density and distribution functions,respectively. ρ represents the correlation coefficient of the wage and participa-tion equation. Adding an error term ε∗i , which is equal to

ε∗i = εi −E [εi|Xi, I∗i > 0] , (7)

the market wage rate function to be estimated is:

lnWi = Xiβ + ρσελi + ε∗i , with λi =ϕ (Viγ)

φ(Viγ). (8)

While the value of the normal hazard λi is generally not known, a consistentestimate bλi can be obtained by probit estimation of the probability that an indi-vidual is working. Subsequently the variable bλi is calculated for each individualseparately and added to the list of regressors for lnWi , as denoted in equa-tion (8). Finally, OLS estimation of this equation provides consistent results.It should be noted, however, that the additional term in the wage equation isdependent on λi and thus on Vi. Table 2 lists studies on the gender wage gapthat correct for sample selection. We indicate whether the selection correctionterm proved statistically significant.The sensitivity of the coefficient of the selection factor depends on the spec-

ification of the selection equation, which highlights a potential weakness of theHeckman correction technique. The procedure requires the availability of validinstruments, i.e. variables which contribute to determine the propensity to workbut are not related to wages. In practice, such exclusion restrictions are difficultto find and collinearity problems are likely to prevail. Lauer and Steiner (2000)test different combinations of instruments. The normal hazard proves mostlysignificant in their wage equation but it does not alter the other coefficientsin a significant way. Furthermore, the consistency of the two-step estimatorhinges on the assumption of multi-normality of the error terms.6 In order tocircumvent this latter restriction, other correction terms have been suggestedsuch as the propensity of participating (Olsen 1980). In the latter approach the

5For more details see Heckman (1979).6An extensive critique of the Heckman procedure is provided by Puhani (2000).

14

propensity score instead of the normal hazard is included in the wage equation.An empirical survey about various correction terms is provided in Vella (1998).Another way to tackle the selection problem is to estimate the participation

and wage equations simultaneously. The advantages and drawbacks are twofold:if none of the equations is misspecified, simultaneous estimation yields efficiencygains. On the other hand, misspecification of one equation may contaminatethe other(s), resulting in inconsistency.

3.1.3 Heterogeneity

Another bias can arise because of heterogeneity. Such a bias appears if unob-served individual characteristics which affect the wage, such as motivation orability, are correlated with the explanatory variables, such as work continuity.For example a more intermittent worker with less motivation could earn less ifintermittence reduces remuneration based on work effort.One way to deal with unobserved individual heterogeneity is to exploit panel

data. There are two main possible assumptions about the unobserved individualeffects in panel data estimation. In the random effects model individual effectsare considered as a component of the error term for which a particular distri-bution is assumed. In fixed effects models the regression is supplemented by anindividual specific characteristic which varies over individuals but is constantover time. In general the wage equation can be specified as

lnWit = Xitβ + Ziδ + αi + εit (9)

whereXit and Zi are vectors which include time-variant and time-invariant char-acteristics of individuals. The variable αi indicates the unobservable individual-specific effects, whereas εit reflects unobservable effects varying both across in-dividuals and over time.In the random effects model the information about the distribution of the

individual effect is used to create an individual-specific term. By adding the un-observable person-specific effect to the general error term the individual-specificterm is generated.The assumptions about the composite error term are:

E (εitαi) = 0 for all i, t (10)

E (εitεjs) = 0 unless i = j, t = s

E (εit) = 0 for all i, t.

The wage equation (9) can be estimated by a GLS estimator, which takesaccount of the variance-covariance matrix of the composite error term in anoptimal way.7 If αi is not correlated with the regressors and (εi1,...,εiT ,αi) ⊥(Xi1,...,XiT , Zi), this GLS estimator is the best linear unbiased estimator.Contrary to the random effects model, no assumption on the distribution of

the individual effects, given the regressors, is made in the fixed effects model.The model contains a person-specific dummy variable to control for unobserv-able factors that may affect the individual wage rate. In the wage equation (9)

7For more details see Baltagi (2001, p.13).

15

Table 2: Examples with selectivity bias correction

Authors Data source Heckman- Heckman- SignificantProcedure Procedure selectionfor men for women coefficient

Dolton, National postal x xMakepeace (1986) survey (UK)Dolton, General household x xMakepeace (1987) survey (UK)Miller (1987) Canadian census x x for femalesCallan (1991) ESRI Survey of x no

Income Distribution,Proverty and Usage ofState Services

Kidd, Australien Bureau x xViney (1991) Statistics (ABS)Miller, Australien Census x x xRummery of Population(1991) and HousingWright, General household x xErmisch (1991) survey (UK)Gyimah-Brempong, CPS (U.S.) x x noFichtenbaum,Willis (1992)Palme, Swedish Level of x x xWright (1992) Living SurveyAshraf, PIDE Socio-Economic x x for malesAshraf (1993) Household Survey of

Rawalpinidi CityChristofides, Canadian Labour x x xSwidinsky (1994) Market Activity

SurveyHersch, PSID x noStratton (1997)Arabsheibani, RLMS (Russia) x xLau (1999)Black, Data from International x noTrainor, Social SurveySpencer (1999) Programme ISSPOglobin (1999) RLMS (Russia) x x noJohansson, HINK x x noKatz,Nyman (2000)Neuman, Isreali Census of x x xOaxaca (2001) Population and Housing

16

αi is now a fixed variable and εit is the traditional error term. The influenceof the explaining variables can be consistently estimated by a fixed effects (FE)estimator. At first it is necessary to eliminate the fixed effects by transformingeach observation either by mean deviation or a first-difference operator. How-ever, these operators also sweep out the time-variant explaining variables. Inthe mean deviation procedure, the individual’s variable means are subtractedfrom each observation. Alternatively, αi can be eliminated by the first-differencemethod where a lagged variable value is subtracted from each observation. Thenan OLS-estimation is carried out on the transformed observations to determinea consistent estimator of β.8

The choice of a random or fixed effects estimator depends on the possible cor-relation between the explanatory variables and the individual effects. As alreadymentioned, the GLS estimator is biased in the random effects model if there isa correlation between the individual-specific effects αi and the variable vectorsXit and Zi, because αi is a component of the error term. In this case, the fixedeffects estimator is more suitable because it is consistent. The Hausman-test isoften used to decide about the appropriate estimation technique. In this test theGLS-estimator is efficient and consistent while the FE-estimator is consistentbut inefficient under the H0-hypothesis (no correlation). Under the alternativehypothesis the FE-estimator is also consistent whereas the GLS-estimator is bi-ased so that the latter cannot be employed. Intermediate situations, where αiis assumed uncorrelated with only parts of X and Z, are studied by Hausmanand Taylor (1981), and discussed in the next subsection.Table 3 includes examples of the empirical literature including FE- or GLS-

estimations.

3.1.4 Endogeneity

An endogeneity bias arises if the error term is correlated with at least one ofthe explanatory variables. In the case of cross section estimation this means

E [εi|Xi] 6= 0. (11)

One way of dealing with endogeneity is to include instrumental variables inthe estimation. Thereby the variable xk, where xk ∈ x is correlated with theerror term (E [ε|xk] 6= 0), is replaced by an instrument or a vector of instrumentsZ. A consistent estimate is ensured only under the following conditions:

E [εi|Zi] = 0 (12)

E [Zixik] 6= 0 (13)

That is, the instrumented variable may not be correlated with the errorterm, while there is sufficiently high correlation with the endogenous variable.

8The mean-deviation transformation only leads to a consistent estimator if the regressorsare strictly exogenous, E (εitxis) = 0 ∀ t, s, while the first-difference estimator only requirespredetermined regressors E (εitxis) = 0 ∀ t ≥ s.

17

Table 3: Examples with longitudinal/panel data, correcting for unobserved het-erogeneity

Authors Data source FE REChoudhury (1993) PSID xKim, PSID x xPolachek(1994)Polachek, PSID x xKim (1994)Naur, Sub-sample of a Danish xSmith (1996) longitudinal data baseHersch, PSID xStratton (1997)Brookes, Panel Comparability Project xHinks, (employed persons fromWatson (1999) Germany and UK)Hansen, Swedish Survey “Household xWahlberg (1999) Market and Nonmarket

Activities (HUS)”

More exactly, to identify variables as instrument variables, the following condi-tions must be fulfilled. The exclusion restriction predicts that the instrumentsincluded in Zi do not determine the wages. Furthermore the order conditionsmust be fullfilled which means if there are k endogenous variables in the model,there should be at least k instruments or k exclusion restrictions. According tothe rank condition, the instruments have to be correlated with or determine theendogenous variables.The instrumental estimator (IV-estimator) has the following form:

biv = (Z0X)−1 Z0 lnW. (14)

In the estimation of wage equations, work experience, housework time, work-ing time intermittence and education are often considered as potentially endoge-nous variables. In general two groups of instruments are applied: exogenousvariables and transformed endogenous variables which comply with the orthog-onality assumption by construction.The most popular exogenous variables which are used in empirical studies

are the number of children, the age, the region, the education of parents andthe occupation. While the correlation between those variables and the endoge-nous ones may be plausible, the correlation between the error term and theinstrumental variables cannot be excluded. For example age is often used as aninstrument for work experience. In this case, the age should not appear in thewage regression (“exclusion restriction”). However, a person’s age may have anadditional effect on the wage if it reflects strength and mental agility indepen-dently of the work experience. In this way the exogeneity assumption can bediscussed for each of the suggested instrumental variables.Within the framework of panel data, endogeneity can arise in different ways,

18

calling for different estimation procedures. First, the idiosyncratic error termand the time-varying and time-invariant regressors can be correlated.

E [εitXit, εitZi] 6= 0. (15)

In this case the application of instrumental estimation is necessary. The mostdetailed discussion of the application of different instruments in connection withgender wage gap estimation is found in Kim and Polachek (1994). They suggestlagged differences of endogenous variables as instruments for work experienceand time out of work.In the case of panel data it is also possible that the individual-specific ef-

fects and some time-varying variables are correlated. One way to eliminate theaccruing bias is therefore to eliminate the correlation. As described in the lastsection, the application of a fixed effects model is a possibility to avoid a correla-tion between the individual effect and the regressors. The fixed-effect estimatoris consistent for β because αi is eliminated by first difference or mean deviationtransformation. Kim and Polachek (1994) provide an extensive discussion ofthe endogeneity issue. Particulary, the authors expose that both endogeneityand heterogeneity are related to one another because in both cases there is acorrelation between the error term in the wage equation and the regressors.Furthermore different tests are described to discriminate between the sources ofthe bias.Another possibility for endogeneity to arise is by correlation between αi and

the time-invariant variables. Hausman and Taylor (1981) developed a method todeal with the correlation between explanatory variables and individual-specificeffects within the framework of panel data analysis. The method uses againthe fact that there are time-varying and time-invariant regressors Xit and Zi.Furthermore it is possible to distinguish the columns of X and Z which areasymptotically uncorrelated with αi from those which are not. The coefficientsof time-varying variables are determined by the fixed effect estimator. The re-sults are consistent because the individual-specific effects are eliminated by thefirst-differences or mean derivation transformation, just like the time-invariantregressors. To obtain consistent estimates of the parameters of endogenoustime-invariant variables, the time-varying exogenous variables are used as in-struments. A necessary condition for this is clearly that there are at least asmany exogenous time-varying variables as there are endogenous time-invariantvariables. The estimation procedure should perform better than traditional in-strumental variable methods, because it does not rely on excluded exogenousvariables for instruments.Table 4 presents some studies dealing with the endogeneity problem. The

table lists possible endogenous variables and the methods employed in order toobtain consistent estimates.Instrumental variable estimation often involves problems concerning the ro-

bustness of the estimation results. Sensitivity tests may be applied for instanceby estimating specifications which employ several alternative instrument sets.A defect of the IV-estimates is that their standard errors are often so large

that the IV-estimates do not differ significantly from the biased OLS-estimates.This problem occurs particularly if the endogenous variables and their instru-ments are only weakly correlated.

19

Table 4: Studies with IV-estimation

Authors Data source Endogenous Instrumental variablesvariable

Callan ESRI Survey of experience no details(1991) Income Distribution,

Poverty and Usage ofState Services

Kim, PSID experience, RE-model (GLS/IV):Polachek hometime, parental education, race,(1994) age and occupation

First-difference approachin FE-model: laggedexplanatory variables

Naur, Sub-sample of a experience apply Hausman andSmith Danish longitudinal Taylor (1981)(1996) data baseHersch, PSID housework time education, experience, tenure,Stratton region, children, nonlabor(1997) income, spouse’s education,

age, wage, residence variablesBlack, Data from experience Probit estimation of partici-Trainor, International Social pation, predicted probabilitySpencer Survey Programme of working in t-1, t-2, ...(1999) ISSPHansen, Swedish survey schooling apply Hausman andWahlberg “Household Market Taylor (1981)(1999) and Nonmarket

Activities (HUS)”

20

3.1.5 The sample selection model by Lewbel

In order to obviate the self selection problem we propose to use the endogenoussample selection model introduced by Lewbel (2002) which, to our knowledge,has not yet been applied to the estimation of gender wage gaps. It providesa two stage least squares estimator of the coefficients, where regressors are al-lowed to be endogenous, mismeasured, or otherwise correlated with the modelerrors. Unlike the Heckman two-step estimator, no structure is imposed on thedistribution of the error terms, permitting a more general form of unknownheteroscedasticity. The estimator does not even require an estimation of the se-lection equation, in other words the participation decision. Yet, the latter maybe estimated separately in order to gain information on the factors influencingthe labour market activity of women and men in Europe.In the Lewbel procedure identification is obtained by observing a “special”

variable S. S is supposed to be continuously distributed with large support.The selection process, that is the participation decision, needs to be monotonicin S. The “special” variable affects treatment in certain ways but does notaffect the dependent variable. Hence S can be interpreted as an instrument fortreatment. As for our application of a wage regression, an appropriate choicefor S may be the unearned income of the individual. Define

W =I

f(S|U) , (16)

where I is the participation indicator, f(S|U) is the conditional probabilitydensity function of S given U , where U ∈ X. (In the case when X containsendogenous variables, one can use U ∈ X,Z instead, where the Z are potentialinstruments for X.).In a first stage, f(S|U) is estimated either parametrically or non-parametrical-

ly. Lewbel shows that, by using f(S|U) as a weighting function in the wageestimation of the second stage, one can obtain a consistent estimator for theβ-parameters of the wage equation, either by least squares or by instrumentalvariable estimation if some of the regressors are endogenous.Lewbel proposes several extensions of this two-stage least squares estimator.

The extension to panel data is presented in the sequel.

3.1.6 Panel wage estimation with selectivity

As discussed in subsection 3.1.3, panel data techniques can be used to controlfor unobserved time-invariant heterogeneity. In many applications, in particularwhen estimating wage equations, a selectivity bias may occur simultaneously.Different estimators have been developed recently to deal with both sources ofestimation bias.The model can be written as

lnWit = Xitβ + αi + εit ∀ 1, ..., N ; t = 1, ..., T (17)

I∗it = Vitγ − ηi − uit ∀ 1, ..., N ; t = 1, ..., T , (18)

where αi and ηi are unobservable and time invariant individual effects whichmay be correlated with the observed variables Xit and Vit. We consider a modelwith fixed individual effects.

21

If we suppose that non-participating people are randomly distributed overthe wage distribution, a pooled OLS estimator may be used to estimate thewage equation (17). If the selection process is time constant, a simple estimatorin differences leads to consistent estimates. In that case, the selection processneed not be specified. Those assumptions are however restrictive.All three estimators we present in this section (Wooldridge, 1995, Kyriazi-

dou, 1997 and Lewbel, 2002) allow for selectivity in the wage estimation, andallow for additive individual effects in both the participation and the wage equa-tion.The first estimator has been proposed by Wooldridge (1995). It relies on a

full parameterisation of the sample selection mechanism, and requires a speci-fication of the functional form of the conditional mean of the individual effectsin the equation of interest. It does not impose distributional assumptions onthe error terms and the fixed effects. The second estimator has been proposedby Kyriazidou (1997). The basic idea of this estimator is, for each individual,to match observations which have the same selection effect in two time periods.Differencing will wipe out both the individual heterogeneity and the selectionterms. The third estimator has been developed by Lewbel (2002). It consists inan extension to panel data of the estimator presented in the preceding section.

3.1.7 Panel estimation in levels: Wooldridge (1995)

The estimator proposed by Wooldridge (1995) can be considered as a panel-extended Heckman two-step estimator. The assumptions are the following:

• the regression function of ηi on Vi is linear:

ηi = Viδ + ci,

where Vi = Vi1, ..., ViT , δ = δ1, ..., δT and ci is a random component;

• the errors of the selection equation, vit = uit + ci, are independent of eViand normally distributed

¡0,σ2t

¢, where eVi = (Xi, Vi), Xi = Xi1, ...,XiT ;

• the regression function of αi on Xi and vit is linear:

E (αi|Xi, vit) = Xiψ + φvit,

where ψ = ψ1, ...,ψT ;

• εit is mean independent of eVi conditional on vit and its conditional meanis linear in vit:

E³εit|eVi, vit´ = E (εit|vit) = ρvit.

Since vit is unknown, the expectation value of vit given participation and eVihas to be computed in a first step. Denoting Iit as the participation indicator,we have:

E³vit|eVi, Iit = 1´ = λ (Viδ) ,

22

where λ (.) represents the normal hazard. Under the four above assumptions,Wooldridge derives an explicit expression for the wage expectation given ob-servables and participation:

E³lnWit|eVi, Iit = 1´ = Xiψ +Xitβ + γtλ (Viδ) .

bθ = ³bψ, bβ, bγ´ is obtained by pooled OLS:bθ = Ã NX

i=1

TXt=1

IitbΨ0itbΨit!−1Ã NX

i=1

TXt=1

IitbΨ0it lnWit

!,

where bΨit = (1,Xi,Xit, 0, ...0,bvit, 0, ..., 0).The variables Xit have to be time-variant in order to separate ψ and β. The

identification of β is obtained here through the assumption of strict exogeneityof the regressors conditional on the unobserved effect. However Wooldridgementions that the latter condition can be relaxed.

3.1.8 Panel estimation in differences: Kyriazidou (1997)

Kyriazidou (1997) introduces a two-step estimator which “differences out” boththe sample selection effect and the unobservable individual effect in the equationof interest, i.e. the wage equation in our case. Rewrite the main equation (17)as a “partially linear regression”:

lnWit = Xitβ + αi + λit + vit,

where λit represents the sample selection and vit ≡ εit − λit.The key idea of Kyriazidou’s method is to assume, for an individual i par-

ticipating in both periods t and s (t 6= s) (Iit = Iis = 1), and who has Vitγ =Visγ , that the sample selection is the same in both periods (λit = λis). A suf-ficient condition for this is the conditional exchangeability assumption, whichassumes that the vectors of error terms (εit, εis, uit, uis) and (εis, εit, uis, uit)are identically distributed conditional on eVit, eVis,αi, ηi. Thus, the conditionalexchangeability assumption allows to “difference out” both nuisance terms αiand λit from the wage equation.Therefore β can be estimated by OLS using differentiated variables from a

subsample of those observations that have Φits = 1 (Iit = Iis = 1) and Ψits =1 (Vitγ = Visγ). However Ψits is not observable and has to be estimated in afirst step. The first step consists in estimating γ consistently. Since the set ofobservations for which Vitbγn = Visbγn will be empty as soon as Vit contains acontinuous regressor, bΨits = 1 (Vitbγ = Visbγ) will be replaced by the followingapproximation:

bΨitsn = 1

hnK

µ(4Vts) bγn

hn

¶where hn is a function decreasing in n, K (.) is a symmetric “kernel den-sity” function and 4Vits = Vit − Vis. Thus bΨitsn is a weight that is maxi-mum if 4Vitsbγn = 0 and declines to zero as the magnitude of the difference|Vitbγn − Visbγn| increases.

23

Table 5: Comparison of panel estimators

Estimator Estimation in Sample selection effects Sample requirementsWooldridge levels parameterised Iit = 1Kyriazidou time difference unspecified and estimated Iit = Iis = 1; Vitγ ∼= VisγLewbel time difference unspecified and not estimated Iit = Iis = 1

The second step consists in estimating β by weighted OLS

bβn =

ÃnXi=1

1

T − 1Xs<t

bΨitsn (4Xits)0 (4Xits)Φits!−1 ·nXi=1

1

T − 1Xs<t

bΨitsn (4Xits)0 (4 lnWits)Φits

where 4Xits = Xit −Xis and 4 lnWits = lnWit − lnWis.In our application we estimate bγn by conditional logit, we pose hn = n−1/5

and choose the normal density as the kernel function (K (υ) = φ (υ)).

3.1.9 The panel estimator of Lewbel (2002)

The Lewbel panel estimator is basically close to Kyriazidou’s. It differs in howthe weights are computed. Similarly to the cross section case, and contrary toKyriazidou’s estimator, the weights are not computed from the estimation ofthe participation equation but are provided by the unspecified estimation of theconditional density function fvt of the special variable S. Therefore

bΨit = Iitbfvt (Sit|Xit) .Define by bΨt the mean of bΨt. Lewbel shows that β can be estimated

by a linear two-stage least squares regression of bΨitWit/bΨt − bΨisWis/bΨs onbΨit eXit/bΨt− bΨis eXis/bΨ, potentially using instruments Zi. In our application weestimate density function fvt of S parametrically as well as non-parametrically.The main features of the three panel estimators we consider are summarised

in Table 5.Dustmann and Rochina-Barrachina (2000) compare the Wooldridge and

Kyriazidou estimators with the one developed in Rochina-Barrachina (1999).The latter is a time difference estimator and adds a distributional assumptionfor the error term but does not require the conditional exchangeability assump-tion. Jensen et al. (2002) propose a Monte Carlo comparison (limited to T = 2)of the small sample properties of these estimators.

3.1.10 Summary on selectivity, endogeneity and heterogeneity

Summarising the methodological discussion above, Table 6 lists the main prob-lems encountered when estimating wage equations, the consequences if they arenot corrected for and the econometric methods that may be applied.

24

Table 6: Overview of methodological problems when estimating wage equations

Problem Definition Consequences Appropriate methodSelectivity non participants in case of positive all except OLS

are not randomly selection: over-estimateddistributed over wages, hence under-the population estimated pay gap

Endogeneity error term is biased estimated Instrumental Variablecorrelated with at coefficients estimation (IV)least one of the all methods could beexplanatory variables extended to IV

Heterogeneity correlation between impact of the correlated FE panel estimators:unobserved individual explanatory variable - Wooldridge (1995)characteristics and is under-estimated if - Kyriazidou (1997)the wage or correlation is negative; - Lewbel (2002)explanatory variables vice versa

Note: Other econometric problems that may occur include e.g. measurement errors,heteroskedasticity.In our application, we do not correct for endogeneity since appropriate instrumentsare not available in the data set.

3.1.11 Further developments

Most recent publications in the gender pay gap literature aim at highlightingthe structure of the wage gap throughout the whole wage distribution. Whileone possibility to take into account distribution effetcs is to use one of the esti-mation methods presented and to apply a Juhn-Murphy-Pierce decomposition,an alternative way is to examine the effects of differences in characteristics andremuneration at different points in the distribution. That is, to apply quantileregression techniques. For example, Albrecht et al. (2003) estimate the effect ofdifferent variables on the wage rates of women and men at all percentiles of thedistribution. The advantage of a quantile regression over, say, ordinary leastsquares, is that it allows to estimate the marginal effect of a covariate on thewage at various points of the distribution, instead of only at the mean. The re-sults of this estimation are used to decompose the difference between the maleand female wage distributions into the endowment and remuneration effects,and, possibly, selectivity and unobservables effects. The applied decompositionis thus in the spirit of the Oaxaca-Blinder approach. But rather than identify-ing differences at the mean of the distribution, they are explained, quantile byquantile. Several methods are discussed in the literature for decomposing differ-ences in distributions. Two examples are the approaches of Machado and Mata(2003) and Bourguignon et al. (2002) which are both based on the quantileregression technique, not used here.

25

3.2 Decomposition of the gender wage gap

Let us define the raw, unadjusted or gross male-female wage differential as thedifference in logarithmic mean wages:

∆ lnW = lnWM − lnWF . (19)

The general procedure of all decomposition methods is that, first of all, wageequations are estimated given individual and other (e.g. firm) characteristics.The estimated price vector bβ and the average human capital and job character-istics for males and females are used to compute weighted differences in meancharacteristics. This “explained” part of the wage gap is assumed to reflect pro-ductivity differences. We call it the endowment effect. The adjusted pay gap isthen measured as the difference between the total wage differential observed andthe fraction explained by differences in human capital endowments of womenand men. This remaining part measures differences in the remuneration of thecharacteristics and is therefore often referred to as a measure of discrimination.We call this second term the remuneration effect.In the literature, different variations of this general procedure can be found.

They can be summarised in three main methods of decomposition. First, there isthe method which Oaxaca (1973) and Blinder (1973) developed independentlyin the earliest studies of wage discrimination. This decomposition techniqueis relatively simple because the analysis of the gap is realised at the mean.There are several extensions to this approach regarding the non-discriminatorywage structure, whereby the core of the method is in use until today. Anotherdecomposition method has been developed by Juhn, Murphy and Pierce (1993)which also takes the residual wage distribution into account. Controlling for theoccupation of individuals and treating the distribution of men and women acrossoccupations as exogenous is a common aspect of both methods. However, asoccupational segregation may be endogenous and reflect discrimination in itself,critique has been raised on this issue and it has been the starting point for thedecomposition technique of Brown, Moon and Zoloth (1980). These authorsinclude an occupation selection equation besides the wage regression equation,which is meant to take up a possibly endogenous distribution of occupations.The raw wage gap can then be decomposed in a part due to within-occupationwage differences and another part due to between-occupation wage differences.The remainder of this section pursues, for the most part, the exposition of

Kunze (2000).

3.2.1 Decomposition by Oaxaca and Blinder

In this section the decomposition techniques based on the studies of Oaxaca(1973) and Blinder (1973) are introduced. First, we present the basic method.Subsequently we discuss some extensions.The basic method applies to the determination of wage differentials at the

mean, and it was developed for cross-sectional data. Most authors estimateseparate wage regressions for women and men, although a single regressionshould be sufficient as we will see later on:

lnWMi = XM

i βM + εMi , (20)

26

lnWFi = XF

i βF + εFi . (21)

Superscripts indicate the sex,M for males and F for females. The estimatesare used to compute the raw wage gap at the mean. Those are given for eachgroup by

lnWM= X

MbβM , (22)

lnWF= X

F bβF , (23)

where lnWM=

NMPi=1 lnWi/NM and NM stands for the number of males in the

sample.9 The vector XMrepresents the average human capital characteristics

of the males. For the females, the terms are defined accordingly.To decompose the raw wage gap it is further necessary to make assumptions

on a competitive price vector which operates as standard in valuing the differentcharacteristics. This price vector should reflect the remuneration of human cap-ital characteristics in absence of discrimination. In the basic approach either the

wage structure of males bβMor females bβF is employed as the non-discriminatingwage structure. In his paper from 1973, Oaxaca argued that the reality is mostlikely to correspond to a situation that lies between those two extreme cases.Later studies try to picture the real non-discriminating wage structure in a bet-ter way. These variations will be discussed in the next section. In the followingthe price vector of males is used because most studies which applied the Oaxaca-Blinder method proceed in this way. The predicted mean wage for women iscomputed with coefficient estimates from the male wage regression and averagecharacteristics of females:

lnW 1F= X

F bβM . (24)

The unadjusted wage gap can be decomposed by calculating two differences.

The first differencelnWM − lnW 1

Findicates by how much the mean wage

for men exceeds the mean hypothetical wage for women in the absence of dis-

crimination. The second term lnW 1F − lnWF

shows the distance between thehypothetical wage for women and their actual mean wage. The following de-composition is reflected if both differences are added:

9 If sampling weights are available for different individuals or groups of individuals in orderto ensure sample representativity, they should be used in computing weighted instead of simplemeans.

27

nlnW

M − lnWFo

| z raw wage gap

=nlnWM − lnW 1

Fo+nlnW 1

F − lnWFo

(25)

=nbβMXM − bβMXF

o+nbβMXF − bβFXF

o= bβM ³XM −XF

´| z endowment effect

+XF³bβM − bβF´| z

remuneration effect

The first term on the right side of equation (25) presents the endowmenteffect of the wage differential between males and females. It arises from differ-ences in the average characteristics. Because of this, the term is often consideredas the “justified” or “explained” part of the gap. The second term, the remu-neration effect due to differences in estimated coefficients is interpreted as ameasure of discrimination. If men and women had the same characteristics atthe mean, the existing raw wage gap would only be caused by the difference inremuneration of these characteristics.It is important to stress that the decomposition into endowment and remu-

neration effects is conditioned by the list of explanatory variables included inthe wage regression. Beyond the endogeneity problems already touched upon, aproblem of variable selection thus arises. An extreme view is that “all relevantvariables measuring individuals’ productive endowments” should be included(Berndt 1991, p. 184). A more pragmatic strategy when elaborating policyrecommendations could be to involve the policy makers in the definition of theexplanotory variables to be used. A prevailing opinion, already expressed byOaxaca (1973), seems to be that the greater the number of control variablesis, the smaller the endowment effect will be. While the accumulated empiricalevidence tends to support that opinion we would like to stress that it is by nomeans guaranteed that including more regressors has that effect. Suppose in-deed that we carry out the decomposition (25) on the basis of a set of variablesX1 and that new set of relevant variables X2, orthogonal to X1, becomes avail-able. The orthogonality ensures that the coefficient of XM

1 in the regressionof lnWM on XM

1 and XM2 is the same as in the original regression of lnWM

on XM1 alone. Hence the difference in the endowment effects measured with

both sets of variables is simply bβM2 ³XM2 −X

F2

´, and its sign is unpredictable

a priori. Thus, while including X2 on top of X1 mechanically improves theexplanation of wages, nothing can be said about its impact on the endowmenteffect. An extreme example is given by the Italian case: the endowment effectis negative (see section 6.2.1).By now, the method just described has become the most common proce-

dure to estimate wage discrimination (see Table 7 for examples). As alreadymentioned, in most studies the wage structure of males is assumed as non-discriminating. The authors mostly argue that in the economy men form thelargest group of workers and therefore face virtually no discrimination. In thecourse of time, the assumption of the non-discriminating wage structure hasbeen discussed intensively and new approaches have been introduced.

28

Table 7: Examples with Oaxaca-Blinder decomposition

Authors Data source Non-discr.wages Type of wagemale female regression

Blinder (1973) PSID x OLSOaxaca (1973) Survey of Economic x OLS

Opportunity 1967Dolton, National postal x OLS andMakepeace (1986) survey (UK) HeckmanDolton, General household x OLS andMakepeace (1987) survey (UK) HeckmanMiller.(1987) Canadian census x OLS and

HeckmanBaker, Canadian censuses, x OLS andBenjamin, Survey of Consumer HeckmanDesaulniers, FinancesGrant (1995)Callan(1991) ESRI Survey of Income x OLS and

Distribution, Poverty and Heckman, IVUsage of State Services

Miller, Australian Census of x x OLS andRummery (1991) Population and Housing HeckmanWright, General household survey x OLS andErmisch (1991) (UK) HeckmanPalme, Swedish Level of Living x HeckmanWright (1992) SurveyChoudhury (1993) PSID x OLS and

FEKim, PSID x OLS, FEPolachek (1994) and RE, IVHarkness (1996) General household survey x OLS

(UK) and British householdpanel survey

Naur, Sub-sample of a Danish x RESmith (1996) longitudinal data baseHersch, PSID x x OLS, IV,Stratton (1997) FEBlack, Data from International x OLS andTrainor, Social Survey Programme HeckmanSpencer (1999) ISSPJohansson, Swedish Household x x OLSKatz, Income Survey (HINK)Nyman (2000)Neuman, Israeli Census of x HeckmanOaxaca (2001) Population and HousingEuropean Commission ECHP x OLS(2002)

29

3.2.2 Choice of a non-discriminatory wage structure

The idea that discrimination not only lowers the wages of the “minority” group(females), but also leads to higher wages for the other group (males), forms thestarting point of these methods. The studies attempt to identify the separateeffects of discrimination of different groups.In general the methods differ with respect to the implicitly assumed non-

discriminatory competitive wage structure. Nevertheless it is possible to expressthem in a general form.First, it is necessary to assume that there is a vector β∗ which reflects the

rates of return on human capital characteristics in the absence of discrimination.Then equation (25) can be reformulated as:

nlnW

M − lnWFo= β∗

³XM −XF

´+X

F³bβM − β∗

´+X

M³β∗ − bβF´ .

(26)

While the well-known first term estimates the productivity difference betweenthe two groups, the second term measures the amount of “discrimination” infavour of males, the third element the amount of “discrimination” against fe-males. Various proposals have been published in the literature of how to esti-mate β∗. Oaxaca and Ransom (1994) suggest the following versatile represen-tation of an estimated non-discriminatory wage structure:

β∗ = ΩbβM + (I −Ω)bβF , (27)

where I describes a diagonal unit matrix and Ω is a weighting matrix. Thetwo extreme cases in the study of Oaxaca (1973) correspond to Ω = I andΩ = 0. The authors point out that the non-discriminatory structure should liesomewhere between those two possibilities.Reimers (1983), for instance, chooses simply Ω = I/2 as weighting matrix ,

or

β∗ = (bβM + bβF )/2. (28)

Another possibility, suggested by Cotton (1988), is to assume that ΩC =fMI. The scalar fM denotes the share of the majority group (here: males)in the total (working) population. Cotton argued that the non-discriminatorystructure should be more similar to the structure that holds for the larger group.This procedure of using labour force weights to construct a non-discriminatorywage structure as convex linear combination of two separate wage structuresis less extreme. The variant of Neumark (1987) and of Oaxaca and Ransom(1994) appears more natural. They define the β∗ as vector of rates of returnwhich would be obtained by estimating equation (20) or (21) over the entirepopulation (males and females combined). The weighting matrix in equation(27) then has the form:

Ω0 = (X0X)−1(XM0

XM), (29)

where X represents the matrix of individual human capital characteristics in thepooled sample. The authors argue that this pooled method does not restrict

30

the non-discriminatory wage structure to be bracketed by the two separatelyestimated wage structures. Furthermore, this approach nests the weightingschemes of Cotton and Reimers.Oaxaca and Ransom (1994) analyse the empirical consequences of using

the female, the male or the Cotton and Neumark wage structure as the non-discriminatory structure. In comparison, the differences in the estimated com-ponents of the unadjusted wage gap are recognizable. The adoption of the fe-male wage structure as the competitive standard leads to a larger discriminatoryand smaller productivity differential than the use of male coefficient estimates.The Cotton and Neumark (pooled) method also yields quite different estimatesof the productivity differential. These differences imply a division of the dis-criminatory wage differential into male overpayment and female underpaymentcomponents.Silber and Weber (1999) examine the differences between five decomposi-

tion procedures. In addition to the approach of Oaxaca and Ransom (1994)they also took Reimers’ method into account. The tests are based on bootstraptechniques. In contrast to Oaxaca and Ransom (1994), both studies concludethat the component measuring the degree of discrimination against the “minor-ity group” (females) was the highest when the “majority group” (males) servedas reference and that the component estimating the degree of discriminationin favour of the “majority group” was the highest when the “minority group”served as reference. Altogether, no general results or robust conclusions can bederived, though. Because the choice of the reference wage structure crucially af-fects the results, other variants of the Oaxaca-Blinder decomposition have beensuggested.10

To summarise, the empirical evidence shows that the gender pay gap decom-position is often sensitive to the choice of the competitive price vector. Table 8lists some studies which apply variations of the basic Oaxaca-Blinder decompo-sition method.

3.2.3 Integration of the sample selection correction

This section discusses how to handle the selectivity bias correction within thedecomposition of the raw wage gap. As already mentioned, in some studiesa correction is applied for both women and men while most analyses only in-clude the correction term in the female wage regression. In the following thedecomposition method is applied for the case where a selectivity bias correc-tion is performed for both sexes. If a random sample for men is assumed, thecorrection term is set to zero.The selection correction terms enter the wage decomposition as follows:

lnWM − lnWF

= bβM(XM −XF) +X

F(bβM − bβF ) + (bθMbλM − bθFbλF ), (30)

where bθ is an estimate of ρσε and bλ is the mean of the estimated hazard.The consequences of a selectivity bias correction are twofold. First, the esti-

mated rates of return, bβM and bβF , differ due to the introduction of the selectionvariable. Second, selection correction term adds to the familiar endowment andremuneration components. However, it is not obvious how the last term in

10 See among others Lauer (2000), Licht and Steiner (1991) or Bonjour (1997).

31

Table 8: Variations on Oaxaca-BlinderAuthors Data source Decomposition methodNeumark National Longitudinal Oaxaca with male wage(1987) Survey of Young Men and structure, Oaxaca with

Young Women (NLS) female-male wage structure,Neumark

Ashraf, PIDE Socio-Economic Oaxaca with male wageAshraf Household Survey of structure, Oaxaca with(1993) Rawalpindi City female-male wage structure,

(Pakistan) CottonOaxaca, U.S Current Population Oaxaca with maleRansom Survey, Data from a wage structure, Oaxaca(1994) specific U.S. firm with female-male

wage structure,Cotton,Neumark

Sharpe, National Longitudinal Reimer, CottonAbdel-Ghany Survey of Young Men and(1996) Young Women (NLS)Silber, Isreali census Oaxaca withWeber male wage structure,(1999) Oaxaca with

female-male wage structure,Reimer, Cotton,Neumark

32

equation (30) should be treated in the overall decomposition scheme, that is,whether it should be attributed to differences in endowments or included in theremuneration effect. Several variants are found in the literature. In most stud-ies, the last term on the right-hand side of the equation (30) is subtracted fromthe observed wage gap on the left-hand side. In this form the left-hand sideprovides a measure of the difference in potential or offered wages, in contrastto observed wages realized only by those participating in the labour market(see among others Oglobin 1999).11 The studies which proceed in this way findthat the existence of a sample selection bias implies that the “offered wage gap”greatly exceeds the observed wage gap. However, this empirical result is ob-

tained with bθM = 0 and bθF positive (no selection for men, positive selectionfor women, that is, positive correlation between unobservables in the wage andparticipation equations) and is therefore by no means general. Note also thatwith other selectivity correction approaches, the relative magnitude of the of-fered and observed wage gap may not relate directly to the sign of a coefficient.With Olsen’s (1980) approach, for instance, the additional regressor designedto correct for selection is the predicted linear index of the participation equa-tion. Given that this is inversely related to the normal hazard included in theHeckman correction, one expects a negative coefficient for this regressor. How-ever, the sign of the mean of the participation index can differ depending on theparticipation rate.The impact, on the remuneration and endowment effects, of taking the cor-

rection of the sample selection into account, is ambiguous. For instance in theinvestigation of Oglobin (1999) both the remuneration and the endowment ef-fects decline in comparison with the results of an OLS regression. But botheffects increase in the study of Miller (1987). Miller und Rummery (1991) showthat the effects may point in opposite direction. In their study the endowmenteffect declines and the remuneration effect increases.Another treatment of the selectivity bias correction term is proposed by

Dolton and Makepeace (1986), who treat the correction term as a regular ex-

planatory variable. The difference in bλ between men and women weighted withbθM enters into the endowment effect. The distance between the estimated co-efficients for the correction term add to the remuneration effect. In contrastto this, Dolton and Kidd (1994) and Choudhury (1993) choose a decomposi-tion similar to equation (30) and interpret the third term as representing theappropriate correction for non-random sampling.Summarising, sample selectivity may be taken into account in various ways.

It is not possible to point out any one of these as the right one, since theappropriate procedure depends on the specific empirical problem and the data athand, as discussed by Neuman and Oaxaca (2001). However, empirical studieslargely support the significance of the selection bias correction, particularly forwomen. Following Dolton and Kidd (1994) and Choudhury (1993), we will treatthe selectivity bias as an additional effect to the endowment, remuneration (andunobservables) effects.Contrary to OLS or Heckman two-step estimators (at least the OLS ver-

sion of the latter), the sample mean of the unobservables (ε) is not zero for theLewbel estimator. Thus, even a wage decomposition at the mean yields an un-11The offered wage coincides with the observed wage for participants, but is not observed

for non-participants.

33

observables effect if we use Lewbel estimates. This effect cannot be interpretedstraightforwardly. However, it may capture part of the selection effect.

3.2.4 Decomposition by Juhn, Murphy and Pierce

While the Oaxaca-Blinder decomposition was developed for cross section wagemodels, the method of Juhn, Murphy and Pierce (1993) is more suitable forextensions either in time (longitudinal wage models) or in space (for instancecross-country comparisons). It extends the approach of Oaxaca and Blinder bydecomposing the pay gap not only at the mean but over the whole wage distri-bution, thereby accounting for the residual (unexplained) wage distribution.In contrast to the basic Oaxaca-Blinder decomposition, wages now have to

be estimated separately for men and women. Predicted wages are then used toderive hypothetical wage distributions that serve to extend the decompositionof the unadjusted wage gap by a wage structure effect. The common predictionfor male and female wages is:

lncWM = XMbβM , (31)

lncWF = XF bβF . (32)

As in the basic Oaxaca-Blinder approach it is assumed that βM = β∗, hencethe estimated coefficients from the male wage regression are taken as the com-petitive price vector. The decomposition of the raw wage gap then includes threecomponents related to differences in endowments, in estimated coefficients andin the residual wage distribution, respectively:

∆ lnW| z raw

wage gap

= bβM(XM −XF )| z difference inobserved char.

=endowment effect

+ XF (bβM − bβF )| z difference in pricesfor observed char.=remuneration effect

+ (bεM −bεF )| z difference inunobservables

=unobservable effect

(33)

The main feature of the Juhn-Murphy-Pierce approach is the decompositionof the raw gap at different points of the wage distribution. Consequently thisdecomposition method allows an analysis at the quantile level:

∆ lnW = bβM(XMq −X

Fq ) +X

Fq (bβM − bβF ) + (εM − εF ), (34)

where XF

q represents the mean characteristics for each quantile q. The Juhn-Murphy-Pierce decomposition has been applied mostly in studies analysing thewage gap between two groups of workers over time or across countries. For thispurpose the basic decomposition from equation (33) must be extended to twoperiods or to two countries.In the following the application of the method shall be introduced for the

comparison of gender earnings differentials between two countries. In this case

34

the decomposition technique is employed to distinguish the effects of gender-specific factors from those associated with the underlying wage structures ofboth economies. First the wage equation for a male worker from country j isdefined as follows:

lnWMj = XM

j βMj + σMj θMj , (35)

where εMj = σMj θMj and σMj is country j’s residual standard deviation of wages.The vector θMj can be interpreted as the standardised unobservable compo-nent of productivity.12 Usually the additional assumption of equal price vectorsfor women and men within each country is made. Hence, non-discriminatoryreturns to observables are assumed for both sexes βMj = βFj = βj . The male-female log wage gap for country j is then given by:

Dj ≡ ∆ lnW = βMj (XMj −XF

j ) + σMj (θMj − θFj ) = βj∆Xj + σj∆θj . (36)

Now it is possible to decompose the gender pay difference between two coun-tries j and k:

Dj −Dk = βj(∆Xj −∆Xk) +∆Xk(βj − βk) + (∆θj −∆θk)σk +∆θj(σj − σk).

(37)

A difference in the gender wage differential D between two countries k andj may arise from four sources. The first component reflects the contribution ofinter-country differences in observed human capital characteristics to the gendergap. The second term measures the impact of different prices across countriesfor observed productivity characteristics. The third component presents theeffect of the international difference in the relative wage position of males andfemales after controlling for observed human capital characteristics. In otherwords, it reflects differences in the level of unobservables. Finally, the lastterm measures differences in the return to unobservable skills. In this way theJuhn, Murphy and Pierce decomposition allows wage structure factors to bedistinguished from gender-specific factors that explain a part of the wage gap.So the first and third term reflect the inter-country difference in the relativebehaviour or treatment of males and females, and as such may be regardedas gender-specific. By contrast, the second and fourth components are notrelated specifically to aspects of gender, but arise from inter-country differencesin the underlying wage structure, that is the relative prices of productivitycharacteristics in the labour markets.The Juhn, Murphy and Pierce decomposition has been adopted by Blau

and Kahn to analyse international gender earning differentials. An innovativefeature of their studies (1992, 1995, 1996a, 1996b, 1997 and 2000) is to focuson the role of the wage structure as an additional factor influencing the genderwage gap. They compare the U.S. labour market with European countriessuch as the U.K., Germany, Denmark. Another study which also adopted thedecomposition method of Juhn, Murphy and Pierce (1993) stems from Rice(1999). This paper examines the factors that shape earning differentials betweenmen and women in European economies. So does the OECD Outlook 2002.Datta Gupta, Oaxaca and Smith (2001) apply the same method to compare the

12 Standardized with mean 0 and variance 1 for each country.

35

development in the gender wage gap in the U.S. and Denmark between 1985 and1995. They use the overall wage distribution (men and women combined) as thereference distribution. In addition to this, they took also a potential selectionbias into account by including the selection correction term.13 See Table 9 fora summary of studies using the Juhn-Murphy-Pierce decomposition.

3.2.5 Extended decomposition by Juhn, Murphy and Pierce

In order to differentiate the gender pay gap within and between the EU coun-tries, we propose the following extension of the Juhn-Murphy-Pierce decompo-sition. In contrast to equation (37) our decomposition also reveals the intra-country differences in all components. Therefore, we have to relax the assump-tion of equal prices for both sexes within each country, that is we now only as-sume non-discriminatory returns to observable characteristics for men βMj = βj .Hence, the difference of the gender pay gaps of two countries (regions, groupsetc.) can be decomposed into eight components: the inter-country differences inendowments, remuneration, unobservables and the prices for unobservables andthe respective intra-country differences in the endowment gap, remunerationgap, unobservables gap and price gap for unobservables:

Dj −Dk = βj (∆Xj −∆Xk) +¡βj − βk

¢∆Xk (38)

+σj (∆θj −∆θk) + (σj − σk)∆θk

+∆βj¡XFj −XF

k

¢+¡∆βj −∆βk

¢XFk

+∆σj

³θFj − θFk

´+ (∆σj −∆σk) θFk .

As in equation (37), the first component reflects the contribution of countrydifferences in observed human capital characteristics and the second measuresinter-country differences in the remuneration of these characteristics. Genderdifferences in wage positions across countries, after controlling for endowmenteffects, are reflected in the third term, while the fourth reveals the country devi-ations in prices for unobservables. The fifth to eighth term, on the contrary, rep-resent differences within countries. The fifth and sixth capture within-countrydifferences in endowments and remuneration. This decomposition enables usto compare the relative contributions of inter-country and intra-country differ-ences.

3.2.6 Decomposition by Brown, Moon and Zoloth

Inspired by the shortcomings of previous studies, Brown, Moon and Zoloth(1980) suggested a new decomposition method. Former studies took only theproblem of unequal pay for equal work into account. Brown, Moon and Zolothargued that differing occupational distributions are also an important sourceof male-female differentials. By including dummy variables for occupations inthe wage regression, most studies take differences between the occupationaldistributions of men and women as given and ignore the inherent potential

13Note, that equation (30) can be straightforwardly extended to the Juhn-Murphy-Piercedecomposition.

36

Table 9: Examples with Juhn-Murphy-Pierce decomposition

Authors Compari- Data source Included Referenceson over Countries/ Country/

Years YearBlau, Countries ISSP: International Social Survey Germany, U.K., U.S.Kahn Programme, CSCC: Class U.S.,Austria,(1992) Structure and Class Switzerland,

Consciousness data, IDS: Sweden Nor-Income Distribution survey way,Australia

Blau, Countries ISSP: International Social Survey Germany, U.K., U.S.Kahn Programme, CSCC: Class U.S., Austria,(1995) Structure and Class Switzerland,

Consciousness data, IDS: Income Sweden, Norway,Distribution survey, Bank of Australia,Italia Survey for Italia Hungary, Italia

Blau, Countries ISSP: International Social Survey Germany, U.K., U.S.Kahn Programme, CSCC: Class U.S., Austria,(1996a) Structure and Class Consciousness Switzerland,

data, IDS: Income Distribution Sweden, Norway,Survey, Bank of Italia Survey for Australia,Italia Hungary, Italia

Blau, Countries ISSP: International Social Survey Sweden, U.S. U.S.Kahn Programme, CSCC: Class(1996b) Structure and Class Consciousness

dataBlau, Time PISD: Panel Study of Income U.S. 1979, 1988 U.S.Kahn Dynamics, CPS: Current 1988(1997) Population SurveyRice Countries ECHP: European Community Denmark, France, DenmarkWorld Bank Household Panel Study, HHP: Germany, Greece,(1999) Hungarian Household Panel Italy, Spain, Portu-

gal, UK., HungaryDatta Gupta, Time CPS: Current Population Survey Denmark 1983, DenmarkOaxaca, and Danish Longitudinal Sample 1995; 1995;Smith data U.S.1983, 1995 U.S.1995(2001)OECD Countries ECHP: European Community Austria, Ireland, PooledOutlook Household Panel Study Netherlands, Re-(2002) Belgium, Finland, gression

France, Germany,Greece, Italy,Portugal, UK,Denmark, Spain

Holmlund Time LINDA: Longitudinal Individual Sweden 1992, Sweden(2003) Data for Sweden 1998 1992

37

discrimination. Brown, Moon and Zoloth included the probability of attaininga certain occupation (from a multinomial logit model) in their analysis of wagedifferentials.To be able to explain the decomposition approach of Brown, Moon and

Zoloth (1980), the unadjusted gender wage gap shall be represented as thedifference in weighted average log-wages taken across K occupations:

lnWM − lnWF =KXj=1

PMj lnWMj −

KXj=1

PFj lnWFj , (39)

where lnWM , lnWF define the log mean wages for men and women, PMj , PFj

are the proportions of males and females in occupation j, j=1,...,K and lnWMj ,

lnWFj denote log mean wages within the occupation j. In a next step equation

(39) is extended by adding and subtracting the term PFj lnWMj :

lnWM − lnWF =KXj=1

(PMj − PFj )lnWMj| z

inter-occupational effect

−KXj=1

PFj

³lnWM

j − lnWFj

´.| z

intra-occupational effect

(40)

The first term on the right hand side of equation (40) indicates the part dueto differences in the occupational distribution between males and females. In thecase of non-segregation, the proportion of males and females in each occupationwould be equal and this term would vanish. The second term measures the partthat is caused by different mean wages within occupations.Brown, Moon and Zoloth (1980) further decompose the terms in equa-

tion (40) into an explained and an remuneration component, according to theOaxaca-Blinder decomposition. In addition an occupational distribution forwomen is used, which exists only in the absence of discrimination. So the ex-plained and remuneration part of the unadjusted wage gap are each composedof an inter-occupational and an intra-occupational effect.

lnWM − lnWF =

inter-occupational effectz | KXj=1

³PMj − bPFj ´XM

jbβMj +

intra-occupational effectz | KXj=1

PFj

³XMj −X

Fj

´ bβMj| z endowment effect

(41)

+

inter-occupational effectz | KXj=1

³ bPFj − PFj ´XMjbβMj +

intra-occupational effectz | KXj=1

PFj

³bβMj − bβFj ´XFj| z

remuneration effect

,

where the vector bPF includes the predicted proportions of females in occupationsj assuming men’s occupational outcomes, Xj are the mean values of character-istics for occupation j, and bβj are the estimated wage equation coefficients for

38

Table 10: Examples with Brown-Moon-Zoloth decompositionAuthors Data source No. of occupationsMiller (1987) Canadian census 6Dolton, sample of U.K. graduates 6Kidd (1994)Kidd, Canadian Labour Market 9; 17; 36Shannon (1996) Activity Survey

occupation j. The first and the fourth parts of equation (41) are obtained byestimating a reduced form multinomial logit model of occupational outcomes formen that attempts to capture the supply- and demand-side factors determiningthe observed occupational distribution. This model specifies the probability ofmale worker i being in occupation j as a function of the worker’s characteristicsVi:

PMij =exp

¡VMi γMj

¢Pnj=1 exp

¡VMi γMj

¢ (42)

The estimates of this model are used to predict bPF , the proportion ofwomen in each occupation if women were distributed according to the maleoccupational allocation mechanism. Thus the male multinomial logit param-eters are combined with the vector of female characteristics to simulate thenon-discriminatory occupational distribution for women.The Brown, Moon and Zoloth decomposition technique has not yet been

applied often. We refer to Table 10 for a few examples.

4 Choosing methodologies ...

4.1 ... with regard to the wage regressions

Given data and time constraints, we have been able to cope with some, butnot all of the methodological problems discussed above. Most important, fromour point of view, is that the estimation procedure chosen allows us to takeaccount of sample selection in the labour market. This requirement is met bytwo-stage estimation of the participation and the wage equation using either thenormal hazard (Heckman) or the predicted linear index (Propensity score) as aselectivity correction term.14 We also apply the Lewbel approach which has theadvantage of providing a non-restrictive correction of the selection bias. As thespecial variable S we use the non-earned income of the household. This variable

14We prefer two-stage estimation to the direct maximum likelihood (ML) estimation of theHeckman model for two distinct reasons. First, ML relies on joint normality of the errors inthe selection and level equations, whereas the two-step estimator only relies on conditionalmoments which may hold for a wider class of distributions, although derived under jointnormality. Second, using OLS in the second stage has the advantage that the average of theresiduals is zero, which does not hold for the ML Heckman estimator.

39

should have an impact on the decision to work but is less likely to be relatedto the wage rate. We tried also other variables, i.e. total household incomewithout individual earnings or non-earned individual income, but the first oneperformed best. As regards the estimation of the weighting function, neededto obtain a consistent estimator for β in the Lewbel model, we try differentmodels, including an approach based on residuals from an OLS regression, anda non-parametric specification. We perform all three estimations (Heckman,Propensity score and Lewbel) on 1998 cross-section data, taking an OLS esti-mation (without selectivity correction) as reference. We expect that, if sampleselection plays a role, the estimates shall vary according to the regression tech-nique applied. This we expect to be the case for the female sample in particular.For males, on the contrary, all estimates are expected to yield similar results.We also try to take account of the possible endogeneity of the explanatory

variables, particularly the choice of the public versus private sector, by applyingone of the Lewbel extensions (Lewbel IV) with an instrumented public sectorindicator. Since the coefficient estimates and the significance of the public sectorindicator do not change with the instrumentation, we refrain from using theseresults for a subsequent decomposition of the pay gap.15 Due to the lack ofappropriate instruments in the ECHP data set, we are not able to model thesector decision any better. (In the upcoming section on the Brown, Moon andZoloth method we list variables that may serve as instruments for modelling thechoice between the public and the private sector.)Finally, we apply panel data analysis and thus take account of individual

heterogeneity in the wage determination process. Three panel data estima-tions are undertaken: we investigate the models and estimators of Wooldridge(1995), Kyriazidou (1997) and Lewbel (2002). The estimations do not yieldconvincing results. The Kyriazidou and Lewbel estimators draw on individ-ual differences over time and thus require substantial variation over time inthe explanatory variables. Such variables are unfortunately rare in the ECHP.Practically the only significant estimated coefficients turn out to be those of theyear dummies. By contrast, the Wooldridge approach can accomodate slightvariation over time, and indeed more coefficients turn out significant. However,the year dummies are not significant. This is a puzzle, because these dummiesshould capture business cycle effects, and these are expected to have rather largeimpacts on wages. Table 11 lists the features of the competing models, as alsodiscussed in detail in the methodological review. In view of all these problems,we disregard panel estimates in the sequel and compare the cross-section esti-mates obtained from the endogenous sample selection model by Lewbel withthose resulting from (1) OLS estimation, (2) Heckman two-step estimation and(3) Propensity score two-step estimation, taking the German sample as a refer-ence sample. For the remaining four country samples and the full EU data setwe use the Lewbel procedure. Concentrating on the cross-section Lewbel model,we are able to disentangle the components of the gender pay gap while consider-ing at least the most prevailing of the methodological problems discussed above,that is, to take account of the selection effect in a non-restrictive way.

15For illustrative purposes, the estimation results are presented in Appendix 2.

40

Table 11: Overview of regression techniques appliedEstimator Selectivity Prerequisite Participation Individual

correction eq. specified heterogeneitythrough accounted for

OLS - - - -Heckman normal hazard norm. of errors p. -Prop. score propensity score norm. of errors p. -Lewbel weighting valid weight. var. - -Lewbel n.p. weighting valid weight. var. - -Lewbel IV weighting valid weight. var. - -Wooldridge normal hazard norm. of errors p. xKyriazidou differentiat. hazard CE s.p. xLewbel panel weighting valid weight. var. - xLewbel panel n.p. weighting valid weight. var. - x

Note: n.p.: non-parametric, p.: parametrically, s.p.: semi-parametrically. CE: Condi-tional Exchangeability assumption.When testing the normality assumption it was always rejected.The special variable S used for the weighting function has to be a valid weightingvariable.

4.2 ... and with regard to the wage decomposition

We calculate the Oaxaca-Blinder and Juhn-Murphy-Pierce decompositions basedon the wage estimations selected. In both applications we use the coefficient es-timates from the male wage regressions as the reference remuneration, that is,non-discriminatory salary structure. As Ginther and Hayes (2003) point out,men are the usual comparison group in legal proceedings of gender discrimina-tion. Hence a pooled approach, obtained from a weighted average of the maleand female wage structures, is not likely to be used in legal cases concerned withequal opportunities for women and men.Due to the use of Heckman and Propensity score two-step regression tech-

niques we are able to distinguish endowment and remuneration effects from aselection effect that gives us an idea of what the wage distribution of womenwould look like in the absence of sample selection. Data and time constraints ledus to refrain from the Brown, Moon and Zoloth decomposition method. But, asalready mentioned, we do find this approach appealing because of its focus onoccupational segregation and its ability to provide accurate within-occupationand across-occupation wage differentials.

5 Data

5.1 Data set

Our empirical analyses are based on data from the user data base of the Eu-ropean Community Household Panel (ECHP UDB, version December 2002).To meet the requirement of more in-depth knowledge and greater compatibilityof data on social and economic conditions in the European Union, the ECHP

41

was launched as a closely coordinated component of a system of householdsurveys. The ECHP is a standardized survey conducted in member states ofthe European Union under the auspices of the Statistical Office of the Euro-pean Communities.16 It involves annual interviews of a representative panel ofhouseholds and individuals in each country, covering a wide range of topics onliving conditions. This includes comparable information across member states,on income, work and employment, poverty and exclusion, housing, health, andmany other social indicators. The key feature of the ECHP is harmonisationof its methodology, specifically through the creation of a centralised question-naire which serves as point of departure for all national surveys. The ECHP isthus a rich data set, providing a wide range of information for investigating thedistribution of income across European countries.

5.2 Data constraints

Despite its merits the ECHP involves restrictions in many ways, especially forthe analysis of employment and wages. Although a common questionnaire en-sures a standardised concept and content of the national surveys, it does notnecessarily imply the provision of identical variables among countries. Becauseof differing legal and institutional frameworks, the same information, e.g. ongross and net income or social transfers, implies different concepts and defini-tions in the different countries. For instance, while net earnings in Germanymean earnings after taxes and social security contributions, in France they arenet of social security contributions but prior to taxation since, in practice, taxesare computed annually and at the end of the year. As regards other incomesources, only gross amounts are reported in the French data files. We drawon information on gross earnings and net unearned income if given in the dataset. Otherwise, we apply the household tax factor provided by Eurostat. Thisadjustment is neccessary for Finnish and French data. Since the country filesof Luxembourg and Sweden do not contain any information on individual earn-ings, both states cannot be included in the study. As a result, our EU levelanalysis comprises only 13 of the 15 EU member states. As the dependent vari-able in our wage equation we use information on individual gross labour incomeper month divided by the reported number of hours worked per week and theaverage number of weeks per month. Hence, the gross hourly wage rate is onlyapproximated from this aggregate information.Another major drawback is the definition and availability of variables we

can use as explanatory variables in the wage regressions. To mention only afew, education is only defined in broad categories mingling e.g. information onschooling and vocational education. Furthermore, the ECHP does not providethe respondents’ work experience - a variable commonly used in wage estima-tions and known as one of the most important determinants of an individual’shuman capital stock. Although respondents are asked to report the age at whichthey started working this information is not sufficient to derive the individualwork experience. First, the question may be interpreted differently in the coun-tries depending on the existence of an apprenticeship system, student jobs andfeatures of the country-specific labour markets. Second, due to unemployment

16For a detailed description of the ECHP methodology and questionnaires see Eurostat(1996).

42

and child-related interruptions, the age at the first job does not tell much aboutthe years an individual, in particular a woman, has been working. Experience ofunemployment, though reported as a 0/1 information, is a problematic variablebecause unemployment is not clearly defined in the questionnaire. Instead ofreferring to registered unemployment, this question relies on self-assessment bythe respondents.We also do not possess detailed information on children living in the house-

hold, not to mention children born by the respondents. According to the liter-ature, particularly small children have been shown to explain women’s labourmarket participation, that is the selection of women into work, to a large extent.Also, in the lack of actual work experience variables, information on children(in specific age groups) is often used as a proxy for employment interruptionsin female wage equations. The ECHP provides only the number of children upto age 16, with an additional question on those born in the preceding year.Sectoral information is poor and not provided for all countries. This infor-

mation is missing for Germany, e.g., although available in the original data set(GSOEP: German Socio-Economic Panel). Also, occupation is not available forall countries, neither is occupational status. As a result, our estimations of wageequations are quite restricted by the information available. On the other hand,the comparable nature of the ECHP data permits a cross-country analysis forthe EU member states on the highest compatibility level available.

5.3 Sample definition

We select our estimation samples according to the following criteria, applied toeach country. We include respondents from all nationalities between 25 and 55who are presently employed or out of the labour force. The quite restrictiveselection on age is made to prevent the results from being excessively affectedby education and early retirement decisions that may influence participationbehaviour. We exclude the self-employed and people working in family busi-nesses because of the difficulty to have credible information on earned incomefor these categories. Unemployed,17 pensioners, students, those in special train-ing programs or national service (military or civil) as well as people with adisability are dropped, too. These restrictions are aimed at forming a sam-ple of either employed or “voluntarily” not employed people. Furthermore, werestrain wage earners to have positive earnings and work for at least 8 hoursper week. Observations with missing information on household net income orindividual wage income do not enter the analysis. The remaining sample sizesfor all ECHP country files are listed in Table 12. Total sample sizes range frommore than 3000 women and 2500 men in the Italian sample to less than 1000women and men for Denmark. The numbers of employed women deviate fromthe total sample sizes to differing extents while the employment share of menapproaches 100 percent in almost every country. The share of working womenis particularly low in the South of Europe (Spain, Italy and Greece), while it17For those unemployed we assume a willingness to participate in the labour market, that is

not realised for some reason (not able to find a job, e.g.). Alternatively, we could form a sampleby introducing a double selection into participation: one for being unemployed, the other forbeing “voluntarily” unemployed (indicated e.g. by the lack of searching for a job). These twosub-samples may indeed have different characteristics. We choose to exclude the self-reportedunemployed because the information on the unemployment status is rather ambiguous in theECHP. This may have an impact on the wage estimation results, but is often done this way.

43

Table 12: Sample sizes of the ECHP country filesCountry # women % working # men % workingAustria 1,343 70 1,270 100Belgium 1,144 79 1,076 99Denmark 963 97 985 100France 2,260 70 1,935 97Finland 1,431 92 1,310 100Germany 2,608 77 2,576 99Greece 1,680 47 1,206 100Ireland 1,407 58 969 99Italy 3,235 53 2,506 99The Netherlands 2,069 82 2,147 99Portugal 1,943 72 1,778 100Spain 2,616 49 2,176 99United Kingdom 2,166 82 1,740 99EU 24,865 69 21,677 99

Data source: ECHP, country files 1998. Samples of 25-55 year old women and men,who are employed or out of the labour force.

is remarkably high in the Scandinavian countries Denmark and Finland. Alsoremarkable are the unbalanced sizes of the selected samples between the sexesin Southern Europe. One explanation is that men are more often self-employedin rural activities or in services and are therefore dropped. On the other hand,women, especially mothers, are more likely not to participate and thus are keptin our sample. As mentioned earlier, Luxembourg and Sweden are not includeddue to missing information on wage income.Based on both sample and population size, we choose to investigate the

wage gaps of the five largest EU countries Germany, France, Italy, Spain andthe United Kingdom for the year 1998 in more detail. For Germany we alsoapply panel estimation using the years 1994-1999. At the EU level, we performtwo sets of estimations. The first uses the pooled data set of all ECHP countryfiles with income information (that is, except Luxembourg and Sweden), thesecond uses all ECHP country files with additional sector information (hence,Germany, Luxembourg and Sweden are excluded). We choose the 1998 crosssection to facilitate comparison with the results published in Employment inEurope 2002 (European Commission 2002a).For descriptive statistics, such as variable means and standard deviations,

of the German and the EU level sample the reader is referred to the respectivetables in Appendix 2.

6 Results

We investigate various estimation techniques and the resulting decompositioneffects for the German and the pooled EU samples of the ECHP data set.We provide results for both the Oaxaca-Blinder and the Juhn-Murphy-Piercewage decomposition based on four different wage estimations: OLS, Heckman,Propensity score and Lewbel (cross section). In addition to the cross-section es-

44

Table 13: Overview of methods appliedEstimator EU Germany France Italy Spain UKOLS D DHeckman D DProp. score D DLewbel D D D D D DLewbel n.p. ELewbel IV EWooldridge EKyriazidou ELewbel panel ELewbel panel n.p. E

Note: D: wage decomposition, E: wage estimation, n.p. non-parametricFor the EU we provide results including the German sample but without sector infor-mation, and vice versa.

timations we try instrumental variable estimation (Lewbel IV) and panel dataestimators (Wooldridge, Kyriazidou and Lewbel panel) for the German sample.As we do not have valid instruments and variation across time is too small toapply panel data techniques in a satisfactory way, these yield unsatisfactoryresults. We therefore do not use these estimates for a wage decomposition. Per-forming a non-parametric estimation for the first stage of the Lewbel estimatoronly has a small impact on the wage estimation. Regarding the other selectedcountries, Oaxaca-Blinder and Juhn-Murphy-Pierce decomposition results arepresented, based on the cross-section Lewbel estimation only (see Table 13).

6.1 Wage equation estimation

For the German sample of women we perform four different wage equation esti-mations - (1) OLS estimation and two-step estimations according to (2) Heck-man,(3) Propensity score and (4) Lewbel. Since men’s employment participa-tion does not prove to be affected by selectivity problems, male wage equationsare estimated by OLS.18 For decompositions based on the Lewbel model, bothwage equations, for males and females, are estimated with the Lewbel proce-dure. As explanatory variables we include individual characteristics (such asage, education, tenure, information on household composition and regional in-formation) and job characteristics (firm size, sector, occupational group). Theparticipation equation for the Propensity score and Heckman procedures is setup with age, family status, children, non-earned income and various interactionterms as independent variables.19 The selection correction variables turn out tobe statistically significant in both procedures. While the propensity of partici-

18Less than 1.5 percent of the male sample does not participate in Germany.19For each participation or wage equation we start off with a full list of explanatory variables

from which only the statistically significant regressors are kept. The full list for the partici-pation equation includes: two age group dummies, a marriage dummy, a marriage/cohabitingdummy, dummies for secondary and tertiary education, dummies for child age<1, 1-11,monthly non-earned income (household total net income minus own wage income) and aninteraction term for age and child 1-11.

45

Table 14: Goodness of fit for female wagesEstimation model No. of obs. Mean wage Std. dev.Actual wage 2,015 9.80 4.57OLS 1,985 9.82 2.73Heckman 1,982 9.82 2.75Propensity score 1,982 9.82 2.75Lewbel 1,884 9.83 2.84

Data source: ECHP, German data file 1998. Sample of 25-55 year old women, whoare employed at least 8 hours per week.

pating increases a woman’s wage rate in the Propensity score specification, thenormal hazard is negatively correlated with the (log) wage rate. Thus, there isa negative correlation between unobservables in the participation equation andin the wage equation. This result underlines the necessity to correct for sampleselectivity, since the parameters would not be estimated consistently otherwise.As the special variable in the Lewbel model, we use annual non-work net

household income. In addition, we experimented with the respective non-workindividual income as well as total household income excluding the individual’slabour earnings (but including e.g. the partner’s wage income). The first mea-sure performed best with regard to the predicted wages. We opt for the non-earned household income for theoretical reasons also, since this income measureis least likely to be correlated with the individual wage rate.Table 14 summarises the features of the competing wage equation models.

The variance of the predicted wages is largest in the Lewbel model. The Lewbelpredictions therefore represent best the spread of the actual wage distribution.The estimation results for all wage equation specifications (German sample)

and the estimation results of the Lewbel regression for France, Italy, Spain, UKand the EU level are listed in Appendices 2, 3a, 3b and 4a.

6.2 Wage decomposition

6.2.1 Oaxaca-Blinder

The magnitude of the decomposition terms of the raw wage gap differ remark-ably depending on the wage regression model used. According to the OLS esti-mation, a little bit more than 50 percent of the German wage differential can beexplained by different endowments of women and men (see Figure 3). The re-maining half is due to differences in the remuneration of these endowments. Theresults of the Lewbel procedure reflect an endowment effect of similar size. Theselection effect from the Propensity score regression corresponds to a positive

The wage equation is set up with a full list of: age, age squared, dummies for secondary andtertiary education, number of children 1-11 and 12-15, seven occupational group dummies, twofirm size dummies, public sector dummy, 12 sector dummies (for all countries but Germany),permanent contract dummy, part-time dummy, dummy for unemployment spells, dummy EastGermany as well as interaction terms for age and child<1, and age and education. The fulllist of variables corresponds to the variables available in the ECHP potentially interesting forour purpose.Several variables (tenure, education, occupation, public sector, etc.) may well be endoge-

nous but we have paid only limited attention to this problem here.

46

-20

0

20

40

60

80

100

120

OLS Heckman Propensity Score Lewbel

Perc

ent

Remuneration effect Endowment effect Selection effect Unobservable effect

Figure 3: Oaxaca-Blinder decomposition of the gender wage gap in Germany.Data source: ECHP, German data file 1998. Sample of 25-55 year old women andmen, who are employed at least 8 hours per week.

selection on observables and the selection effect from the Heckman regressioncorresponds to a negative selection on unobservables (negative correlation be-tween the unobservables in both equations). Note, however, that this meansthat a woman with a relatively low predicted probability of participation is pre-dicted to earn less conditional on participation than a woman with the sameproductivity endowment X but a higher predicted probability of participation.The Heckman findings for Germany mean that the potential wage gap betweenwomen and men would be lower than observed if those currently not workinghad the same observed characteristics as those currently working. Though, thisdoes not necessarily imply a higher average wage rate if all women worked dueto differing endowments between participating and non-participating women.No selection effect is displayed for the Lewbel estimation since selection is notestimated explicitly, but it shows off indirectly in the unobervables whereasthe unobservable effect is restricted to zero on average for OLS, Heckman andPropensity score estimates. The Lewbel estimate tells us that unobserved char-acteristics of women lead to an offered wage gap which exceeds the observedwage gap, contrary to the findings obtained with the Heckman estimates.In comparison to the German sample, the Oaxaca-Blinder decomposition of

the gender pay gap in the EU looks a bit awkward (see Figure 4). This is due tothe relatively bad specification of the wage equation for the EU sample, wherehardly any variable proves statistically significant (see the estimation results forthe EU in the Appendix). The Heckman selection effect is large and negativeimplying that the unobservables of the participation and wages equations arepositively correlated at the EU level. The differences in the endowment effectresulting from the Lewbel and the other procedures thereby hint at the impor-tance of considering self selection on the one hand and the inappropriatenessof the statistical assumptions implied by Heckman or Propensity score on the

47

-60

-40

-20

0

20

40

60

80

100

120

140

160

OLS Heckman Propensity Score Lewbel

Perc

ent

Remuneration effect Selection effect Endowment effect Unobservable effect

Figure 4: Oaxaca-Blinder decomposition of the gender wage gap in the EUData source: ECHP, pooled country files 1998 (excl. Luxembourg and Sweden). Sam-ple of 25-55 year old women and men, who are employed at least 8 hours per week.

other. In our opinion, the decomposition results at the EU level have to be inter-preted with great caution since a pooled wage equation can hardly be estimatedmeaningfully when between country heterogeneity is presumably large.20

Figure 5 displays the Oaxaca-Blinder decomposition for the EU when addinginformation on sectoral occupation to the list of explanatory variables in thewage equation.21 The decomposition looks quite different now. The endowmenteffect becomes close to zero when using Lewbel estimates while it remains above20 percent when using the other estimates. The selection effect disappearsor is significantly smaller. But Figure 6 suggests, for the Lewbel estimates,that dropping Germany has a larger impact on the wage decomposition thanintroducing sector information.In levels, the offered gender wage gaps estimated with the Lewbel approach

in the other country samples differ from the German sample in both directions.However, according to Figure 7, the fraction explained by different endowmentsof women and men is generally smaller than in Germany. Consequently, theremuneration effect, that is, the relative effect of differences in remuneration, islarger. Italy and Spain have a negative endowment effect. This means that ifItalian and Spanish women had the same characteristics as Italian and Spanishmen they would receive even lower wages than observed. The Italian gap inoffered wages is almost 50 percent higher than that in observed wages, hintingat a strong and negative selection of (low-wage) women into the labour market.

20Obtaining a meaningful estimate at the EU level would mean searching for the optimalmidterm between separate country estimates and complete pooling, entailing time-consumingspecification searches.21As sector information is not available for the German country sample, Germany has to

be dropped in this specification.

48

-20

0

20

40

60

80

100

120

OLS Heckman Propensity Score Lewbel

Perc

ent

Selection effect Unobservable effect Remuneration effect Endowment effect

Figure 5: Oaxaca-Blinder decomposition of the gender wage gap in the EU usingsector informationData source: ECHP, pooled country files 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women and men, who are employed at least 8hours per week.

6.2.2 Decomposition by factors

We now investigate the relative impacts of the factors related to the gender paygap in Germany and at the EU level. These impacts are computed based onthe Lewbel cross-section estimation and an Oaxaca-Blinder decomposition byfactors for each sample.As illustrated in Figure 8, differences in the educational background and

firm tenure of women and men contribute to a positive pay gap in Germany.In addition, the different employment shares of men and women in part-timejobs and occupations like service/sales workers and legislators, senior officialsand managers lead to an increase. Other contributing factors are the differentremunerations of women and men for education and elementary occupations.The different remuneration of age, firm tenure, job interruptions and part timeemployment, on the contrary, reduces the wage gap. Age can be used as aproxy for experience if more detailed information for the latter is not available.In that case age may capture, to some extent, labour market experience andcareer progression. The effect of age in the wage decomposition includes then thefact that women have on average longer career interruptions and therefore earn,at a given age, less than men. Living in East Germany is also associated witha smaller pay gap. To summarise, the different returns to higher educationallevels for women and men have by far the largest impact on the gender wagegap.

According to Figure 9, differences in education do not play such an important

49

-20

0

20

40

60

80

100

120

Decomposition EU withoutGermany, no sector information

Decomposition EU withoutGermany, with sector information

Decomposition EU with Germany,without sector information

Perc

ent

Unobservable effect Remuneration effect Endowment effect

Figure 6: Comparison of Oaxaca-Blinder decompositions using differentEU-samplesData source: ECHP, pooled country files 1998 (excl. Luxembourg and Sweden,respectively Germany). Sample of 25-55 year old women and men, who are employedat least 8 hours per week.Note: Decomposition based on the Lewbel wage regression. Sector information is notavailable for Germany.

50

-60

-40

-20

0

20

40

60

80

100

120

140

160

Germany France Italy Spain UK EU

Perc

ent

Remuneration effect

Endowment effect

Unobservable effect

Figure 7: Oaxaca-Blinder decomposition of the gender wage gap in selectedEU countriesData source: ECHP, country files 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Decomposition based on the Lewbel wage regression using sector informationfor all countries but Germany. The pooled EU sample does not contain Germany,Luxembourg and Sweden.

51

-12 -8 -4 0 4 8 12

east Germany

number of children <16

elementary occupations

plant and machine operators and assemblers

craft and related trades workers

skilled agricultural and fishery workers

service and sales workers

clerks

professionals

legislators, senior officials and managers

middle size firm

small size firm

part time employment

job interruption (due to unemployment)

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

Figure 8: Contributions to the gender pay gap in GermanyData source: ECHP, German data file 1998. Sample of 25-55 year old women andmen, who are employed at least 8 hours per week.Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation.The first bar “age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

52

part at the EU level,22 while the impact of the age variable is substantial (seenote to Figure 9). Accordingly, a difference in age between the sexes contributesto a difference in pay. As regards the endowment effect, also deviations in firmtenure, holding a permanent job contract and having a supervisory or non-supervisory job status increase the wage gap whereas equalizing the proportionsof women and men with a tertiary education level would lead to a decrease.Occupational groups and industry sectors can be distinguished with regard totheir impact on the wage differential. In terms of the remuneration of womenand men, particularly education, firm tenure, working in the public sector andjob status are differently rewarded. Interestingly, women and men face differentreturns to being married and having children, which is a further contributingfactor to the earnings gap.

6.2.3 Juhn-Murphy-Pierce: within-country decomposition

Juhn-Murphy-Pierce decompositions of the gender wage gaps within countriesand at the EU level are displayed in Figures 10 to 25 in Appendix 1. The keyadvantage of the decomposition method of Juhn, Murphy and Pierce is that itallows an investigation of the wage gap over the whole wage distribution. It thusrelates to the Oaxaca-Blinder decomposition by disentangling the pay differenceat the overall sample mean into quantile gaps. To illustrate this feature, let ustake a closer look at the German example. In Figures 10 to 13 the raw wagegap is displayed according to the wage rankings of all women and men in thesample. As in the Figures on the raw pay gap women and men are separatelyordered by their wage rate and then compared pairwise in each decile of thedistribution (the raw wage gap was displayed by percentile). That is, the meanwage of the 10 percent lowest paid women is now compared with the mean wageof the 10 percent lowest paid men. This gap is displayed as the first observationto the left of the figure. As the wage distributions of women and men are notcongruent typically, the decile gaps also reflect the extent to which male andfemale wage distributions are spread differently.In the German case lowest-wage earners face a gender wage gap of about 31

percent. The raw gap first decreases and then increases when moving up theincome distribution. For middle-wage earners it is as low as 20 percent whereasfor high-wage women it amounts to 28 percent.23 As in the Oaxaca-Blinderdecomposition, the raw gap can be decomposed into an endowment and a re-muneration effect. Furthermore, since we are no longer confined to the samplemeans, the effect of unobservables also comes into play. The fraction of the wagegap explained by different endowments of women and men increases over thewage distribution. While taking up only about 20 percent at the lower end ofthe wage distribution, it accounts to more than 80 percent of the gap betweenhigh-wage women and their male counterparts. The remuneration effect showsthe reverse pattern. Different remuneration of characteristics seems to affectmostly women at the lower end of the wage scale. Unobservables take up some

22The reason may be that no industry dummies are included for Germany since this infor-mation is not available.23 It would be interesting to have information on the sampling variability of these effects, in

order to know whether they are significantly different over the wage distribution. That couldbe done by applying a non parametric bootstrap method to draw confidence bands.

53

-12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12

number of children <16

married

other sector

health and social work

education

public administration

real state, renting and business activities

financial intermediation

transport, storage and communication

hotels and restaurants

wholesale and retail trade

construction

manufacture

electricity, gas and water supply

elementary occupations

plant and machine operators and assemblers

craft and related trades workers

skilled agricultural and fishery workers

service and sales workers

clerks

professionals

legislators, senior officials and managers

middle size firm

small size firm

non-supervisory job status

supervisory job status

permant job contract

job interruption (unemployment)

public sector

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

Figure 9: Contributions to the gender pay gap in the EUData source: ECHP, EU country files 1998 (excl. Germany, Luxembourg and Sweden).Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation. Countrydummies serve as additional control variables.The first bar “age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

54

of the wage gap for low income receivers but their effect is close to zero for therest. In contrast to Figure 10, Figures 11 and 12 are based on selectivity cor-rected wage estimations for the female sample. As seen in the Oaxaca-Blinderdecomposition above, the selection effect corresponds to a negative selection onunobservables. It takes up part of the observed wage gap. This means that theoffered wage gap is smaller than the observed wage gap. Remarkably enough,this applies to the whole of the wage distribution to more or less the same ex-tent. As was already the case for the Oaxaca-Blinder representation, the Lewbelestimates differ quite a bit from the Heckman estimates when differentiated overdeciles (see Figure 13). While the increasing pattern of the endowment effectwith rising wage is confirmed, the remuneration effect decreases from the sec-ond decile on for both the Heckman and the Propensity score estimates, but itsprofile is more or less flat according to the Lewbel estimates.At the EU level, the fraction of the wage gap explained by different endow-

ments of women and men also increases over the wage distribution while theremuneration effect decreases with rising income. This is a common result fromall estimation procedures applied. The Heckman selection effect contributesnegatively to the pay gap over the whole distribution.Again, a decomposition for the EU is made based on wage equation estima-

tions with sector information and excluding Germany due to the lack of thisinformation in the German sample. The picture looks a bit different for theselection effects, that are much smaller, and the Lewbel-based decomposition.Over the whole wage distribution the remuneration effect now equals more orless the raw wage gap. That is, nothing can be explained by different charac-teristics of women and men.In France the U-shape of the raw gender wage gap is quite pronounced

(see Figure 22). Starting with some 25 percent in the lowest decile, the gapdiminishes to little more than 10 percent for the rest of the wage distribution,except for the highest earners where it amounts to 18 percent. In contrast tothe German case, the share of the endowment effect varies a lot over the incomedistribution. Also the pattern of unobservables shows more volatility due tothe smaller sample size for France. The quasi-totality of the raw wage gap inFrance is attributed to a different remuneration of women and men. The UKgraph reveals a similar pattern, although the endowment effect takes up a largerfraction of the pay gap, and so does the graph for Spain. For Italy only the wagedifference of the highest earners can be explained by different endowments ofwomen and men.

6.2.4 Juhn-Murphy-Pierce: between-country decomposition

Using our proposed extension of the Juhn-Murphy-Pierce decomposition (equa-tion (38)), we now compare for illustrative purposes the German wage gapdistribution with the French one. (Because of the poor fit of the EU-level wageequation we prefer not to use the results for any comparison.) Figures 26 and 27illustrate the decomposition of the gap across both countries. Eight effects canbe distinguished. The country difference in the raw wage gap ranges between2 and 7 percentage points. That is, for every decile of the wage distributionthe wage gap in Germany exceeds the French one.24 Part of this difference can

24The remark of Footnote 20 applies here as well.

55

be attributed to the inter-country endowment effect (Figure 26) or the intra-country difference in the endowment effect (Figure 27) which are both increasingover the wage distribution. The relative advantage of men in terms of humancapital and job characteristics is greater in France than in Germany, particu-larly in the upper part of the wage distribution. The inter-country remunerationeffect in turn takes up a decreasing fraction of the deviating wage gaps. Thecountry difference in remuneration gaps accounts for a large part of the raw dif-ference, particularly in the lower and middle income range, and turns negativethereafter. The implication is that high-wage French women are more unequallyrewarded for their endowments relative to men than their German counterparts.Differences in unobservables and in the remuneration for unobservables are closeto zero.

6.2.5 Brown-Moon-Zoloth

In our opinion, the decomposition technique by Brown, Moon and Zoloth pro-vides a promising route to deal with endogenous segregation with respect tooccupation, sector or any other possibly selective grouping of jobs. With theBrown, Moon and Zoloth decomposition more insights can be gained, for in-stance, from the determination of the part of the pay gap due to within occupa-tion wage differentials and the part due to gender distinct distributions acrossoccupations. A necessary prerequisite for the application of this technique, how-ever, is the availability of variables that may serve as instruments for explainingthe sorting process into occupations, sectors etc.. As mentioned above, we didnot succeed in modelling the choice of public versus private sector due to thelack of appropriate explanatory variables in the ECHP data set. Such a listwould include more detailed information on the respondents’ children (age),parents (e.g. education level, ocupation), regional information (to depict thedemand side of the labour market), their attitude towards work, career, incomeand family. Only for illustrative purposes, we provide the results of the unsuc-cessful Lewbel IV estimation, where public sector employment is instrumented,in Appendix 2.

7 Conclusion

In this study we propose different techniques to assess the gender pay gap inthe European Union. This concerns the estimation of wage equations as wellas the decomposition of the estimated gaps. Tables 15 and 16, summarizingthe main empirical results of our explorative study, show that at most 50% ofthe difference in payment between the sexes can be attributed to differences incharacteristics. This confirms the findings of other studies, e.g. the Employmentin Europe 2002 report (European Commission 2002a). However, the size ofthe endowment effect differs considerably between countries. It depends onthe information used and on the estimation model and decomposition methodapplied.On the methodological level, we have presented wage estimation methods

that account for selectivity, endogeneity and heterogeneity. We decompose thepay gap both at the mean, following Oaxaca and Blinder (1973), and acrossthe wage distribution as proposed by Juhn, Murphy and Pierce (1993). As the

56

Table 15: Overview of empirical results for all countriesCountries Decomp. Raw wage gap Endowment Remuneration Unobserv.EU OB 16 % 4 % 97 % -1 %

JMP de-/increasing flat flat/decreasing flatGermany OB 21 % 51 % 59 % -10 %

JMP de-/increasing variable flat decreasing/flatFrance OB 13 % 39 % 77 % -16 %

JMP de-/increasing variable de-/increasing variableSpain OB 15 % -15 % 120 % -5 %

JMP variable de-/increasing in-/decreasing variableItaly OB 7 % -50 % 141 % 9 %

JMP flat flat decreasing variableUK OB 27 % 20 % 84 % -5 %

JMP de-/increasing in-/decreasing decreasing de-/increasing

Data source: ECHP, country files and pooled EU country files 1998 (excl. Germany,Luxembourg and Sweden).Note: The percentages give the mean of the raw gap/wage decomposition effects(Oaxaca-Blinder decomposition). The raw wage gap is defined as the difference ofthe hourly gross wage between men and women. The wage decomposition effects arein percentages of the raw gap. The second line of each country gives the overall shapeof the raw gap/wage decomposition effects over the wage distribution (Juhn-Murphy-Pierce decomposition).

literature on wage equation estimation is very rich, we concentrate on methodsmost used in the gender gap literature (i.e. OLS and Heckman) on the onehand, and on the other hand on very recently developed methods (Lewbel 2002e.g.). The latter allow, at least theoretically, to take account of the three mainmethodological issues, selectivity, endogeneity and heterogeneity, in original andinnovative ways.The empirical application, based on the European Community Household

Panel (ECHP) for five European countries (France, Germany, Italy, Spain andthe United Kingdom) as well as at the EU level, shows that correction, especiallyfor selectivity, may have a significant impact both on wage estimates and on thepay gap decomposition: our results suggest that, for Germany, the offered wagegap is smaller than the observed wage gap (for given characteristics) on thebasis of the Heckman estimates, but the reverse using the Lewbel estimates.Therefore, careful attention should be paid to the choice of estimation method.We prefer the Lewbel approach because it is less restrictive on the structureof the data. No structure is imposed on the distribution of the error terms,permitting a more general form of unknown heteroscedasticity.Another main result of the study is derived from the pay gap decomposition

over quantiles of the wage distribution (Juhn, Murphy and Pierce method). Re-markable differences are revealed within as well as between countries. Given thelessons from these Juhn-Murphy-Pierce decompositions, a further recommenda-tion derived from our analysis would be to pay careful attention to differencesover the wage distribution when drawing policy conclusions.Overall, wage estimations for Germany seem to perform best in terms of

57

Table 16: Overview of empirical results for Germany and the EUDecomp. Endowment Remuneration Selection Unobservables

EU (no sector)OLS OB 14 % 86 %

JMP increasing decreasing variableProp. Score OB 15 % 47 % 38 %

JMP increasing decreasing increasing variableHeckman OB 15 % 131 % -46 %

JMP increasing decreasing increasing variableLewbel OB 11 % 101 % -12 %

JMP flat/increasing decreasing variableEU (with sector)OLS OB 22 % 78 % 0 0

JMP increasing decreasing variableProp. Score OB 22 % 79 % -1 % 0

JMP increasing decreasing flat variableHeckman OB 22 % 85 % -7 % 0

JMP increasing decreasing flat flatLewbel OB 4 % 97 % 0 -1 %

JMP flat flat variableGermanyOLS OB 54 % 46 % 0 0

JMP increasing decreasing decreasing/flatProp.Score OB 51 % 62 % -13 % 0

JMP increasing decreasing flat decreasing/flatHeckman OB 51 % 38 % 11 % 0

JMP increasing decreasing flat decreasing/flatLewbel OB 51 % 59 % 0 -10 %

JMP flat/increasing flat decreasing/flat

Data source: ECHP, German data file and pooled EU country files 1998 (first sectionwithout sector information: excl. Luxembourg and Sweden, second section with sectorinformation: excl. Germany, Luxembourg and Sweden).Note: The means and shapes of the wage decomposition effects (Oaxaca-Blinder de-composition and Juhn-Murphy-Pierce decomposition) are given based on the respec-tive wage equation model.

58

the variation of the wage taken up by the explanatory variables. While theFrench estimations suffer from the data problem of an approximated householdincome variable, the results for Italy and Spain may be affected by a furtherselection bias, that is, the choice of self employment. The estimation and de-composition results at the EU level are weak as well. They do not reflect at allthose obtained at the country levels. Furthermore, they are very sensitive tothe omission of a country. This may be due to substantial differences betweencountries, beyond the level of wages, and non co-ordinated wage-related policies.We therefore conclude that a single estimation at the EU level (even includingcountry indicators and country specific information) is less appropriate. A bet-ter estimation technique would include, at the EU level, free country coefficientsfor some variables and restricted coefficients for the remaining variables.Time and data constraints led us to restrict our analyses to the present

state. We could not apply further panel data techniques because of insufficienttime variation of the useful variables, although recently developed estimators(Wooldridge 1995, Kyriazidou 1997 and Lewbel 2002 e.g.) allow to correct forheterogeneity in addition to selectivity and endogeneity. There appears to be aninevitable dilemma between using a data set with a high level of comparabilityacross countries, but only few variables, and using data sets tailored to thespecificity of each country, with much more numerous and informative variables.Only the second type of data set appears to warrant the possibility to applysophisticated estimation techniques at the cost of a lower level of comparabilityof results. Indeed, we found that the information provided in the ECHP issometimes poor, too aggregated, or ambiguous. Therefore we did not succeedin modelling behaviour which is endogenous in the wage determination process,such as the choice of the sectoral occupation. This is, however, a necessaryprerequisite when using a decomposition technique dealing with endogenoussegregation (the Brown, Moon and Zoloth method, e.g.). Moreover, our resultsare very sensitive to the variables used and the quality of information provided,as suggested by Figure 6.Future research may include applications of further wage estimation tech-

niques (e.g. quantile regression) and gap decomposition methods (e.g. Machadoand Mata 2003). Furthermore, confidence bands for the adusted pay gap shouldbe investigated, as well as for the different effects in the wage decomposition.

References[1] Albrecht, J., A. Björklund and S. Vroman (2003), Is there a Glass Ceiling

in Sweden?, Journal of Labor Economics, 21 (1), 145-177.

[2] Arabsheibani, G.R. and L. Lau (1999), ‘Mind the Gap’: An Analysis ofGender Differentials in Russia, Labour, 13 (4), 761-774.

[3] Ashraf, J. and B. Ashraf (1993), Estimating the Gender Wage Gap inRawalpindi City, The Journal of Development Studies, 29 (2), 365-376.

[4] Baker, M., D. Benjamin, A. Desaulniers and M. Grant (1995), The Dis-tribution of the Male/Female Earnings Differential, 1970-1990, CanadianJournal of Economics, 28 (3), 479-501.

59

[5] Baltagi, B. (2001), Econometric Analysis of Panel Data, John Wiley andSons, New York.

[6] Becker, G. (1964), Human capital - A Theoretical and Empirical Analysiswith Special Reference to Education, Columbia Univ. Press, Chicago.

[7] Berndt, E.R. (1991), The Practice of Econometrics: Classic and Contem-porary, Addison-Wesley Publishing Company.

[8] Black, B., M. Trainor and J. E. Spencer (1999), Wage Protection Systems,Segregation and Gender Pay Inequalities: West Germany, the Netherlandsand Great Britain, Cambridge Journal of Economics, 23, 449-464.

[9] Blau, F. D. and L. M. Kahn (1992), The Gender Earnings Gap: Learningfrom International Comparisons, AEA Papers and Proceedings, Gender andLabor-Market Outcomes, 82 (2), 533-538.

[10] Blau, F. D. and L. M. Kahn (1995), The Gender Earnings Gap: SomeInternational Evidence, in: Freeman, R.B. and L.F. Katz (ed.), Differencesand Changes in Wage Structure, University of Chicago Press, Chicago.

[11] Blau, F. D. and L. M. Kahn (1996a), Wage Structure and Gender EarningsDifferentials: an International Comparison, Economica, 63, 29-62.

[12] Blau F. D. and L. M. Kahn (1996b), The Gender Earnings Gap: SomeInternational Evidence, NBER Reprints 2062 (also Working Paper 4224).

[13] Blau, F. D and L. M. Kahn (1997), Swimming Upstream: Trends in theGender Wage Differential in 1980s, Journal of Labor Economics, 15 (1),1-42.

[14] Blau F. D. and L. M. Kahn (2000), Gender Differences in Pay, Journal ofEconomic Perspectives, 14(4).

[15] Blinder, A. S. (1973), Wage Discrimination: Reduced Form and StructuralEstimates, The Journal of Human Resources, 8(4), 436-455.

[16] Bonjour, D. (1997), Lohndiskriminierung in der Schweiz, Berner Beiträgezur Nationalökonomie, Bd. 83, Bern.

[17] Bourguignon, F., F. H. G. Ferreira and P. G. Leite (2002), Beyond Oaxaca-Blinder: Accounting for Differences in Household Income DistributionsAcross Countries, DELTA, WP 2002-04.

[18] Brookes, M., T. Hinks, and D. Watson (1999), Comparisons in GenderWage Differentials and Discrimination Between Germany and the UnitedKingdom, IRISS, WP 1999-02.

[19] Brown, R. S., M. Moon and B. S. Zoloth (1980), Incorporating Occupa-tional Attainment in Studies of Male-Female Earnings Differentials, TheJournal of Human Resources, 15 (1), 3-28.

[20] Callan, T. (1991), Male-Female Wage Differentials in Ireland, The Eco-nomic and Social Review 23 (1), 55-72.

60

[21] Choudhury, S. (1993), Reassessing the Male-Female Wage Differential: AFixed Effects Approach, Southern Economic Journal, 60 (2), 327-341.

[22] Christofides, L.N. and R. Swidinsky (1994), Wage Determination by Genderand Visible Minority Status: Evidence From 1989 LMAS, Canadian PublicPolicy, 20 (1), 34-51.

[23] Cotton, J. (1988), On the decomposition of wage differentials, The Reviewof Economics and Statistics, 70 (2), 236-243.

[24] Datta Gupta, N., R. L. Oaxaca and N. Smith (2001), Swimming Upstream,Floating Downstream: Trends in the U.S. and Danish Gender Wage Gaps,Centre for Labour Market and Social Research, WP 01-06.

[25] Dolton, P.J. and G. H. Makepeace (1986), Sample Selection and Male-Female Earnings Differentials in the Graduate Labour Market, Oxford Eco-nomic Papers, 38, 317-341.

[26] Dolton, P.J. and G. H. Makepeace (1987), Marital Status, Child Rearingand Earnings Differentials in the Graduate Labour Market, The EconomicJournal, 97, 897-922.

[27] Dolton, P. J. and M. P. Kidd (1994), Occupational Access and Wage Dis-crimination, Oxford Bulletin of Economics and Statistics, 56 (4), 457-474.

[28] Dustmann, C. and M. E. Rochina-Barrachina (2000), Selection Correctionin Panel Data Models: an Application to Labour Supply and Wages, IZADiscussion Paper 162.

[29] European Commission (2002a), Employment in Europe 2002, Luxembourg.

[30] European Commission (2002b), The Gender Pay Gap and Gender Main-streaming Pay Policy: Synthesis Report of Gender Pay Equality in EUMember States, report prepared by the Group of Experts on Gender andEmployment commissioned by DG Employment and Social Affairs, Novem-ber 2002, Brussels.

[31] European Commission (2002c), The Adjusted Gender Pay Gap: A CriticalAppraisal of Standard Decomposition Techniques, report prepared by D.Grimshaw and J. Rubery as part of the work by the co-ordinating team ofthe Group of Experts on Gender and Employment commissioned by DGEmployment and Social Affairs, March 2002, Brussels.

[32] Eurostat (1996), The European Community Household Panel (ECHP): Vol-ume 1- Survey methodology and Implementation and The European Com-munity Household Panel (ECHP): Volume 1 - Survey questionnaires: Wave1-3 - Theme 3, Series E, Eurostat, OPOCE, Luxembourg.

[33] Ginther, D.K. and K.J. Hayes (2003), Gender Differences in Salary andPromotion for Faculty in the Humanities 1977-95, Journal of Human Re-sources, 38(1), 34-73.

[34] Gyimah-Brempong, K., R. Fichtenbaum and G. Willis (1992), The Effectsof College Education on the Male-Female Wage Differential, Southern Eco-nomic Journal, 58 (3), 790-805.

61

[35] Hansen, J. and R. Wahlberg (1999), Endogenous Schooling and the Distri-bution of the Gender Wage Gap, IZA DP 78.

[36] Harkness, S. (1996), The Gender Earnings Gap: Evidence From the UK,Fiscal Studies, 17 (2), 1-36.

[37] Hausman, J. A. and W. E. Taylor (1981), Panel Data and ObservableIndividual Effects, Econometrica, 49 (6), 1377-1398.

[38] Heckman, J.J. (1979), Sample Selection Bias as a Specification Error,Econometrica, 47, 153-163.

[39] Heckman, J.J., L.J. Lochner and P.E. Hold (2003), Fifty Years of MincerEarnings Regression, IZA Discussion Paper 775, Bonn.

[40] Hersch, J. and L. S. Stratton (1997), Housework, Fixed Effects, and Wagesof Married Workers, The Journal of Human Resources, 32 (2), 285-307.

[41] Holmlund, H. (2003), Falling Behind, The Wage Structure and the Gen-der Wage Gap in Sweden During the 1990s, Working Paper, StockholmUniversity.

[42] Hoynes, H. W. (1996), Welfare Transfers in Two-Parent Families: LaborSupply and Welfare Participation under AFDC-Up, Econometrica, 64 (2),295-332.

[43] Jensen, P., M. Rosholm and M. Verner (2002), A Comparison of DifferentEstimations for Panel Data Sample Selection Models, Discussion Paper2002-01, University of Aarhus.

[44] Johansson, M., K. Katz and H. Nyman (2000), Wage Differentials andGender Discrimination - Changes in Sweden 1981 - 1998, mimeo.

[45] Judge, G. G., R. C. Hill, W. E. Griffiths, H. Lütkepohl and T.-C. Lee(1988), Introduction to the Theory and Practice of Econometrics, JohnWiley and Sons, New York.

[46] Juhn, C., K. M. Murphy and B. Pierce (1993), Wage inequality and therise in returns to skill, Journal of Political Economy, 101 (3), 410-442.

[47] Kidd, M. P. and M. Shannon (1996), Does the Level of Occupational Ag-gregation Affect Estimates of Gender Wage Gap?, Industrial and LaborRelations Review, 49 (2), 317-329.

[48] Kidd, M. P. and R. Viney (1991), Sex Discrimination and Non-RandomSampling in the Australian Labour Market, Australian Economic Papers,30(56), 28-49.

[49] Kim, M.-K. and S. W. Polachek (1994), Panel Estimates of Male-FemaleEarnings Functions, The Journal of Human Resources, 29 (2), 406-428.

[50] Kunze, A. (2000), The Determination of Wages and the Gender Wage Gap:a Survey, IZA Discussion Paper 193, Bonn.

[51] Kyriazidou, E. (1997), Estimation of a panel data sample selection model,Econometrica, 65, 1335-1364.

62

[52] Lauer, C. (2000), Gender Wage Gap in West Germany: How Far Do Gen-der Differences in Human Capital Matter?, ZEW Discussion Paper 00-07,Mannheim.

[53] Lauer, C. and V. Steiner (2000), Returns to Education in West Germany -An Empirical Assessment, ZEW Discussion Paper 00-04, Mannheim.

[54] Lewbel, A. (2002), Selection Model and Conditional Treatment Effects,Including Endogenous Regressors, mimeo, Boston College.

[55] Licht, G. and V. Steiner (1991), Male-Female Wage Differentials, La-bor Force Attachment and Human Capital Accumulation in Germany,Volkswirtschaftliche Diskussionsreihe 65, Institut für Volkswirtschaftslehre,Universität Augsburg.

[56] Machado, J., and J. Mata (2003), Counterfactual Decomposition ofChanges in Wage Distribution Using Quantile Regression, mimeo, Uni-versidade de Lisboa.

[57] McFadden, D. (1989), A Method of Simulated Moments for Estimation ofDiscrete Response Models Without Numerical Integration, Econometrica,57 (5), 995-1026.

[58] Miller, P. W. (1987), Gender Differences in Observed and Offered Wagesin Canada, 1980, The Canadian Journal of Economics, 20 (2), 225-244.

[59] Miller, P. and S. Rummery (1991), Male-Female Wage Differentials in Aus-tralia: A Reassessment, Australian Economic Papers, 30 (56), 50-69.

[60] Mincer, J. (1974), Schooling, Experience and Earnings, Columbia Univer-sity Press, New York.

[61] Naur, M. and N. Smith (1996), Cohort Effects on the Gender Wage Gapin Denmark, CLS WP 96-05.

[62] Neuman, S. and R. L. Oaxaca (2001), Estimating Labor Market Discrim-ination with Selectivity-Corrected Wage Equations: Methodological Con-siderations and an Illustration from Israel, mimeo.

[63] Neumark, D. (1987), Employers’ Discriminatory Behavior and the Estima-tion of Wage Discrimination, The Journal of Human Resources, 13 (3),279-295.

[64] Oaxaca, R. L. (1973), Male-Female Wage Differentials in Urban LaborMarkets, International Economic Review, 14 (3), 693-709.

[65] Oaxaca, R. L. and M. R. Ransom (1994), On Discrimination and the De-composition of Wage Differentials, Journal of Econometrics, 61, 5-21.

[66] OECD (2002), Women at Work: Who are They and How are They Faring?,OECD Employment Outlook, Chapter 2, 61-125.

[67] Oglobin, C. G. (1999), The Gender Earnings Differential in the RussianTransition Economy, Industrial and Labor Relations Review, Vol. 52 (4),602-627.

63

[68] Olsen, R. J. (1980), A Least Squares Correction for Selectivity Bias, Econo-metrica, 48 (7), 1815-1820.

[69] Olson, C. A. and B. E. Becker (1983), Sex Discrimination in the PromotionProcess, Industrial and Labor Relations Review, 36 (4), 624-641.

[70] Palme, M. O. and R. E. Wright (1992), Gender Discrimination and Com-pensating Differentials in Sweden, Applied Economics, 24, 751-759.

[71] Polachek, S. W. and M.-K. Kim (1994), Panel Estimates of the GenderEarnings Gap, Individual-Specific Intercept and Individual-Specific SlopeModels, Journal of Economics Vol. 61, 23-42 .

[72] Puhani, P. (2000), The Heckman Correction for Sample Selection and itsCritique, Journal of Economic Surveys, 14, 53-68.

[73] Reimers, C. (1983), Labor Market Discrimination Against Hispanics andBlack Men, Review of Economics and Statistics, 65 (4), 570-579.

[74] Rice, P. (1999), Gender Earnings Differentials: The European Experience,The World Bank Development Research Group, WP No.8.

[75] Rochina-Barrachina, M. E. (1999), A new estimator for panel data sampleselection models, Annales d’Économie et de Statistiques, 55-56, 153-181.

[76] Sharpe, D. L. and M. Abdel-Ghany (1996), Discrimination due to Race andGender in the Youth Labor Market: Is it a Double Jeopardy?, Journal ofEconomic and Social Measurement, 22, 43-55 .

[77] Silber, J. and M. Weber (1999), Labour Market Discrimination: Are ThereSignificant Differences Between the Various Decomposition Procedures?,Applied Economics, 31, 359-365.

[78] Vella, F. (1998), Estimating Models with Sample Selection Bias: A Survey,The Journal of Human Resources, 33 (1), 127-169.

[79] Weichselbaumer, D. and R. Winter-Ebmer (2002), The Impact of Markets,Politics and Society on the Gender Wage Gap: A Meta Analysis, WorkingPaper, University of Linz.

[80] Wooldridge, J. M. (1995), Selection corrections for panel data models underconditional mean independence assumptions, Journal of Econometrics, 68,115-132.

[81] Wright, R. E. and J. F. Ermisch (1991), Gender Discrimination in theBritish Labour Market: A Reassessment, The Economic Journal, 101 (406),508-522.

64

65

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration Effect

Source: ECHP, German data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 10: Juhn-Murphy-Pierce decomposition of the gender wage gap in Ger-many (OLS wage estimation)

-0,15

-0,05

0,05

0,15

0,25

0,35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration EffectSelection Effect

Source: ECHP, German data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 11: Juhn-Murphy-Pierce decomposition of the gender wage gap in Ger-many (Heckman two-step wage estimation)

66

App e ndix 1: Juhn- Murphy- Pi e rce figures

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration EffectSelection Effect

Source: ECHP, German data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 12: Juhn-Murphy-Pierce decomposition of the gender wage gap in Ger-many (Propensity score two-step wage estimation)

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gap

Endowment Effect

Unobservables Effect

Remuneration Effect

Source: ECHP, German data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 13: Juhn-Murphy-Pierce decomposition of the gender wage gap in Ger-many (Lewbel two-step wage estimation)

67

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg and Sweden). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 14: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU (OLS wage estimation)

-0,25

-0,15

-0,05

0,05

0,15

0,25

0,35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration EffectSelection Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg and Sweden). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 15: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU (Heckman two-step wage estimation)

68

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration EffectSelection Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg and Sweden). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 16: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU (Propensity score two-step wage estimation)

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gap

Endowment Effect

Unobservables Effect

Remuneration Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg and Sweden). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 17: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU (Lewbel two-step wage estimation)

69

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservable EffectRemuneration Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg, Sweden and Germany). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 18: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU using sector information (OLS wage estimation)

-0,25

-0,15

-0,05

0,05

0,15

0,25

0,35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration EffectSelection Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg, Sweden and Germany). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 19: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU using sector information (Heckman two-step wage estimation)

70

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservable EffectSelection EffectRemuneration Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg, Sweden and Germany). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 20: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU using sector information (Propensity score two-step wage estimation)

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservable EffectRemuneration Effect

Source: ECHP, EU data file 1998 (excl. Luxembourg, Sweden and Germany). Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 21: Juhn-Murphy-Pierce decomposition of the gender wage gap in theEU using sector information (Lewbel two-step wage estimation)

71

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration Effect

Source: ECHP, French data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 22: Juhn-Murphy-Pierce decomposition of the gender wage gap in France(Lewbel two-step wage estimation)

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gap

Endowment Effect

Unobservables Effect

Remuneration Effect

Source: ECHP,Italian data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 23: Juhn-Murphy-Pierce decomposition of the gender wage gap in Italy(Lewbel two-step wage estimation)

72

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration Effect

Source: ECHP,Spanish data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 24: Juhn-Murphy-Pierce decomposition of the gender wage gap in Spain(Lewbel two-step wage estimation)

-0.25

-0.15

-0.05

0.05

0.15

0.25

0.35

1 2 3 4 5 6 7 8 9 10

Raw wage gapEndowment EffectUnobservables EffectRemuneration Effect

Source: ECHP,British data file 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 25: Juhn-Murphy-Pierce decomposition of the gender wage gap in theUK (Lewbel two-step wage estimation)

73

-0.1

-0.05

0

0.05

0.1

1

Raw wage gapInter-Country endowment effect... remuneration effect... unobservables effect... effect due to prices for unobservables

Source: ECHP, German and French data files 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 26: Juhn-Murphy-Pierce decomposition of the gender wage gap betweenGermany and France (based on Lewbel wage estimation), Part 1: inter-countryeffects

-0.1

-0.05

0

0.05

0.1

1

Raw wage gapIntra-country difference in endowment gap... in remuneration gap... in unobservables gap... in price gap for unobservables

Source: ECHP, German and French data files 1998. Sample of 25-55 year olds, employed at least 8 hours per week.

Figure 27: Juhn-Murphy-Pierce decomposition of the gender wage gap betweenGermany and France (based on Lewbel wage estimation), Part 2: intra-countryeffects

74

Table 17: Descriptive statistics for Germany (whole sample)

Men WomenVariable name Variable Mean Std. Dev. Mean Std. Dev.a_y Dummy: age<30 .19 .40 .20 .40a_m Dummy: age 30-45 .53 .50 .51 .50age Age 38.73 8.38 38.99 8.42mar Dummy: married .72 .45 .74 .44mar_coh Dummy: married .82 .38 .84 .37

or cohabitingeduc_3 Dummy: higher education .27 .44 .20 .40educ_2 Dummy: secondary education .57 .50 .58 .49sizehh_015 Number of children <16 .86 1.00 .84 1.01sizehh_1415 Number of Children 14-15 .14 .37 .17 .41sizehh houshold size 3.19 1.26 3.16 1.25child_0 Dummy: newborn .06 .23 .03 .17child_011 Dummy: child<12 .39 .49 .35 .48child_111 Dummy: child 1-11 .34 .47 .32 .47income_ne Yearly nonearned income 11807.86 9580.88 20202.43 13194.17east Dummy: living in east Germany .24 .43 .23 .42cf Dummy: not national .16 .36 .16 .37citizen_other Dummy: non EU citizen .09 .29 .11 .31ce Dummy: EU citizen .07 .25 .05 .23ihhu nonearned household income 1401.35 4906.59 1685.38 5349.64# obs. 2562 2577

Data source: ECHP, German data file 1998. Samples of 25-55 year old women andmen, who are employed or out of the labour force.

75

Appendix 2: Descriptive statistics for the EU and Ger-many, and estimation results for Germany

Table 18: Descriptive statistics for Germany (only working individuals)

Men WomenVariable name Variable Mean Std. Dev. Mean Std. Dev.job_magr Dummy: legislators. senior .09 .28 .03 .07

officials and managersjob_prof Dummy: professionals .12 .32 .10 .30job_cler Dummy: clerks .07 .26 .24 .43job_sale Dummy: service and .04 .20 .14 .34

sales workersjob_wagr Dummy: skilled agricultural .01 .10 .01 .10

and fishery workersjob_wser Dummy: craft and related .33 .47 .04 .19

trades workersjob_wqua Dummy: plant and machine .15 .36 .19 .22

operators and assemblersjob_welm Dummy: elementary occupations .07 .25 .11 .31job_sizelow Dummy: small size firm .19 .40 .24 .43job_sizemid Dummy: middle size firm .55 .50 .43 .50job_public Dummy: public sector .20 .40 .36 .48job_tenure Firm tenure (years) 7.39 6.40 5.95 5.90job_pc Dummy: permanent contract .94 .23 .87 .34job_break Dummy: job interruption .13 .34 .17 .38hours_pt Dummy: part time employment .00 .06 .16 .37wage_grossm Hourly gross wage 12.30 6.50 9.79 4.59# obs. 2388 1747

Data source: ECHP, German data file 1998. Samples of 25-55 year old women andmen, who are employed at least 8 hours per week.

76

Table 19: Descriptive statistics for EU (whole sample)

Men WomenVariable name Variable Mean Std. Dev. Mean Std. Dev.a_y Dummy: age<30 .20 .40 .19 .40a_m Dummy: age 30-45 .50 .50 .49 .50age Age 39.02 8.53 39.41 8.61mar Dummy: married .70 .46 .74 .44mar_coh Dummy: married .80 .40 .82 .39

or cohabitingeduc_3 Dummy: higher education .25 .43 .24 .43educ_2 Dummy: secondary education .35 .48 .32 .47sizehh_015 Number of children <16 .90 1.04 .94 1.06sizehh_1415 Number of children 14-15 .16 .40 .18 .43sizehh household size 3.41 1.37 3.48 1.37child_0 Dummy: newborn .06 .23 .05 .22child_011 Dummy: child<12 .41 .49 .41 .49child_111 Dummy: child 1-11 .35 .48 .36 .48income_ne yearly nonearned income 13971.21 13244.92 20711.17 24215.01cf Dummy: not national .03 .18 .04 .18co Dummy: non EU citizen .02 .14 .02 .15ce Dummy: citizen EU .01 .12 .01 .11d_d Dummy: Germany .12 .32 .10 .31d_dk Dummy: Denmark .05 .21 .04 .19d_nl Dummy: Netherlands .10 .30 .08 .28d_bel Dummy: Belgium .05 .22 .05 .21d_f Dummy: France .09 .29 .09 .29d_uk Dummy: UK .08 .27 .09 .28d_irl Dummy: Ireland .04 .21 .06 .23d_i Dummy: Italy .12 .32 .13 .34d_gr Dummy: Greece .06 .23 .07 .25d_esp Dummy: Spain .10 .30 .11 .31d_po Dummy: Portugal .08 .27 .08 .27d_a Dummy: Austria .06 .23 .05 .23d_fin Dummy: Finland .06 .24 .06 .23ihhu nonearned household income 26871.18 227841.10 37944.19 285625.10# obs. 21645 24814

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Samples of 25-55 year old women and men, who are employed or out of the labourforce.

77

Table 20: Descriptive statistics for EU (only working individuals)

Men WomenVariable name Variable Mean Std. Dev. Mean Std. Dev.job_magr Dummy: legislators. senior .09 .28 .04 .19

officials and managersjob_prof Dummy: professionals .13 .33 .17 .37job_cler Dummy: clerks .10 .30 .24 .43job_sale Dummy: service and .08 .27 .16 .37

sales workersjob_wagr Dummy: skilled agricultural .02 .13 .01 .08

and fishery workersjob_wser Dummy: craft and related .23 .42 .04 .20

trades workersjob_wqua Dummy: plant and machine .14 .34 .04 .19

operators and assemblersjob_welm Dummy: elementary occupations .08 .27 .12 .32job_sizelow Dummy: small size firm .35 .48 .37 .48job_sizemid Dummy: middle size firm .31 .46 .30 .46job_public Dummy: public sector .25 .43 .38 .48job_tenure Firm tenure (years) 8.26 6.97 6.95 6.64job_pc Dummy: permanent contract .89 .32 .83 .38job_break Dummy: job interruption .13 .34 .17 .38hours_pt Dummy: part time employment .01 .12 .16 .36wage_grossm Hourly gross wage 11.08 7.81 9.41 5.30# obs. 18585 14003

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden) Sam-ples of 25-55 year old women and men, who are employed at least 8 hours per week.

78

Table 21: Descriptive statistics for EU (only working individuals excl. Germany)

Men WomenVariable name Variable Mean Std. Dev. Mean Std. Dev.job_super Supervisory job status .15 .36 .08 .27job_nonsup Non-supervisory job status .66 .47 .75 .43sec_agri Dummy: Agricultural .03 .16 .01 .12sec_erng Dummy: electricity, gas and .03 .16 .01 .08

water supplysec_manu Dummy: manufacture .25 .43 .15 .35sec_cons Dummy: construction .12 .32 .01 .11sec_rept Dummy: wholesale and retail trade .11 .31 .12 .32sec_hotl Dummy: hotels and restaurants .03 .16 .04 .20sec_trns Dummy: transport, storage and .09 .29 .03 .18

communicationsec_finc Dummy: financial intermediation .04 .21 .05 .21sec_busn Dummy: real state, renting and .06 .24 .07 .25

business activitiessec_publ Dummy: public administration .11 .31 .09 .29sec_educ education .06 .23 .15 .36sec_hlth Dummy: health and social work .04 .20 .19 .39sec_othr Dummy: other sector .04 .19 .08 .27# obs. 15096 11224

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden) Samples of 25-55 year old women and men, who are employed at least 8hours per week.

79

Table 22: OLS wage estimation for German women 1998lwage_grossm Coef. T-stat.job_magr .1836 3.94job_prof .2526 8.38job_wagr -.4525 -6.00job_cler -.0722 -3.41job_sale -.2790 -10.89job_wser -.1897 -4.47job_wqua -.2249 -5.92job_welm -.3590 -12.55job_public .0736 4.21job_sizelow -.2256 -10.50job_sizemid -.0615 -3.31educ_3 .0992 3.32educ_2 .0634 2.86job_tenure .0259 5.61job_tenure2 -.0006 -2.43east -.2417 -13.65constant 2.2417 74.20obs 1985R-squared .4108Adj R-squared .4060

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours per week.

80

Table 23: OLS wage estimation for German men 1998

lwage_grossm Coef. T-stat.lage 3.9321 3.63lage2 -.5282 -3.53lage_ed3 .3369 4.82job_magr .2060 7.21job_prof .1548 5.68job_wagr -.2204 -3.23job_cler -.0235 -0.79job_sale -.1949 -5.54job_wser -.0913 -4.16job_wqua -.1256 -4.95job_welm -.1864 -5.95job_public -.0631 -3.71job_break -.1264 -6.45sizehh_015 .0222 3.27hours_pt -.3935 -4.04job_sizelow -.2976 -14.78job_sizemid -.1280 -8.34educ_3 -1.1194 -4.36educ_2 .0421 2.23job_tenure .0052 4.23east -.3442 -22.23constant -4.6979 -2.41obs. 2388R-squared 0.4513Adj R-squared 0.4464

Data source: ECHP, German data file 1998. Samples of 25-55 year old men, who areemployed at least 8 hours per week.

81

Table 24: Probit participation estimation for German women 1998

job_part Coef. T-stat.a_y .5354 4.35a_m .3982 4.64lage_ch0 3.5495 3.57lage_ch111 1.4292 3.91mar -.7473 -5.24mar_coh .5489 3.10lincome_ne -.3555 -2.65child_0 -14.3697 -4.12child_111 -6.0151 -4.60e3 .8160 6.91e2 .4454 6.08east .7322 7.56constant 4.0094 3.15obs 2608Pseudo R2 .2719Log pseudo-likelihood -1017.9915

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed or out of the labour force.

82

Table 25: Heckman wage estimation for German women 1998

lwage_grossm Coef. T-stat.job_magr .1866 4.01job_prof .2496 8.29job_wagr -.4650 -6.20job_cler -.0702 -3.32job_sale -.2759 -10.78job_wser -.1852 -4.37job_wqua -.2190 -5.77job_welm -.3412 -11.82job_public .0738 4.22job_pc .0676 2.78job_sizelow -.2362 -10.78job_sizemid -.0733 -3.86educ_3 .0782 2.52educ_2 .0474 2.06job_tenure .0138 10.16east -.2550 -13.72lmbd -.0886 -2.75constant 2.2603 58.91obs 1982R-squared .4143Adj R-squared .4092

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours per week.

83

Table 26: Propensity score wage estimation for German women 1998

lwage_grossm Coef. T-stat.job_magr .1853 3.98job_prof .2500 8.31job_wagr -.4621 -6.16job_cler -.0698 -3.30job_sale -.2743 -10.70job_wser -.1852 -4.37job_wqua -.2193 -5.78job_welm -.3405 -11.77job_public .0739 4.23job_pc .0697 2.88job_sizelow -.2361 -10.77job_sizemid -.0726 -3.83educ_3 .0856 2.81educ_2 .0544 2.41job_tenure .0140 10.33east -.2515 -13.74xbhat .0217 2.60constant 2.1944 65.72obs 1982R-squared .4140Adj R-squared .4090

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours per week.

84

Table 27: Lewbel regression for Germany 1998, 1. stage: density function forwomen

Coef. T-stat.lage -15.1919 -2.08lage2 2.2318 2.21educ_3 1.1174 7.45educ_2 .6172 5.15east -.4282 -3.97child_0 .6529 2.57citizen_other -1.221 -7.89cititzen_EU -.6343 -3.24sizehh_015 -.0592 -1.21constant 30.9817 2.36

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours.

Table 28: Lewbel regression for Germany 1998, 2. stage: selectivity-correctedwage estimation for women

lwage_grossm Coef. T-stat.lage 8.6983 2.08lage2 -1.1965 -2.09job_magr .1838 .81job_prof .3042 3.14job_wagr -.0463 -.57job_cler -.3233 -3.75job_sale -.5803 -3.70job_wser .1484 1.14job_wqua -.3357 -2.58job_welm -.5323 -5.63job_sizelow -.2767 -3.57job_sizemid -.1141 -1.61educ_3 -.1357 -1.57educ_2 -.0648 -.82job_tenure .0171 3.32east -.2005 -3.49constant -13.3174 -1.75

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours per week.

85

Table 29: Lewbel regression for Germany 1998 for women with and withoutinstruments

Lewbel Lewbel-IVCoef. T-stat. Coef. T-stat.

lage 8.3008 1.92 4.6822 .87lage2 -1.1398 -1.92 -.6471 -.87job_magr .1879 .80 -.2316 -.60job_prof .3074 3.11 .3121 2.88job_cler -.0460 -.52 -.2455 -1.54job_sale -.3152 -3.43 -.5086 -3.45job_wagr -.5823 -3.63 -.7503 -3.78job_wser .1165 .91 -.1971 -.84job_wqua -.3406 -2.53 -.6541 -2.70job_welm -.5336 -5.60 -.5106 -4.89job_sizelow -.2794 -3.28 -.4447 -3.03job_sizemid -.1215 -1.69 -.0774 -.87job_public*) .0067 .11 -.8008 -1.40educ_3 -.1445 -1.62 .0656 .32educ_2 -.0664 -.83 .0233 .19job_tenure .0163 3.16 .0195 2.60east -.2008 -3.42 -.0980 -1.00constant -12.6135 -1.60 -5.7132 -.57

Data source: ECHP, German data file 1998. Samples of 25-55 year old women, whoare employed at least 8 hours per week.Note: *) Public sector instrumented with child_0, child_111, mar_coh, mar,mar×lage.

86

Table 30: Lewbel regression for Germany 1998, 1. stage: density function formen

Coef. T-stat.mar_coh -1.9635 -2.46mar_coh×lage .6489 3.03sizehh_015 -.1494 -3.42educ_3 1.479 10.59educ_2 .6608 5.27east -.5421 -5.57citizen_other -.8688 -5.69cititzen_EU -.5144 -2.98constant 4.8391 32.50

Data source: ECHP, German data file 1998. Samples of 25-55 year old men, who areemployed or out of the labour force.

Table 31: Lewbel regression for Germany 1998, 2. stage: selectivity-correctedwage estimation for men

lwage_grossm Coef. T-stat.job_magr .4044 4.92job_prof .0683 .90job_cler -.0132 -.16job_sale -.2871 -3.41job_wagr -.1902 -1.34job_twser -.1322 -1.90job_wqua -.1876 -2.27job_welme -.2163 -2.54job_sizelow -.1767 -2.80job_sizemid -.0327 -.69job_break -.0970 -1.81sizehh_015 .0540 2.89educ_3 .2163 3.35educ_2 .1101 2.21job_tenure .0109 3.45east -.3658 -6.52hours_pt -.3220 -2.96constant 2.3886 25.46

Data source: ECHP, German data file 1998. Samples of 25-55 year old men, who areemployed at least 8 hours per week.

87

Table 32: Wooldridge Panel Wage Estimation for Women 1994-1999 for Ger-many

Coef. T-Stat.job_magr .1882 2.54job_prof .2400 8.67job_wagr -.4617 -2.28job_cler -.0747 -2.93job_sale -.2480 -7.01job_wser -.1198 -2.09job_wqua -.1872 -3.49job_welm -.3234 -7.568job_public .0501 3.48sizehh_015 -.0178 -2.91job_pc .0963 1.26job_sizelow -.2274 -11.03job_sizemid -.0685 -4.73educ_3 .1270 2.82educ_2 .0511 1.57job_tenure .0237 4.36job_tenure2 -.0007 -3.17cf .0191 .70east -.2876 -17.66Dummy_95 -.0333 -.34Dummy_96 .0248 .25Dummy_97 .0701 .72Dummy_98 .1027 1.03Dummy_99 .1234 1.24lmbd -.0382 -.75constant 2.1361 25.52

Data source: ECHP, German data files 1994-1999. Samples of 25-55 year old women,who are employed or out of the labour force.

88

Table 33: Kyriazidou Panel Participation Estimation for Women 1994-1999 forGermany

Coef. T-Stat.a_y -1.0994 -7.31a_m -.6547 -6.92lage_ch111 -.8421 -1.77mar -1.7073 -7.21mar_coh 1.4046 4.86ie -1.2932 -12.90child_111 2.0017 1.17educ_3 1.6035 12.22educ_2 .6203 7.59east 2.1239 11.49constant 14.6076 14.63

Data source: ECHP, German data files 1994-1999. Samples of 25-55 year old women,who are employed or out of the labour force.

89

Table 34: Kyriazidou Panel Wage Estimation for Women 1994-1999 for Ger-many

Coef. T-stat.job_magr .0663 3.94job_prof .0153 1.81job_wagr -.0865 -1.55job_cler -.0443 -5.57job_sale -.0462 -2.61job_wser .0173 1.03job_wqua -.0169 -.90job_welm -.0245 -1.32job_public -.0278 -2.31sizehh_015 -.0033 -.68job_pc -.0632 -14.03job_sizelow -.0541 -5.35job_sizemid -.0088 -1.69educ_3 -.0353 -1.07educ_2 -.0166 -.33job_tenure -.0211 -5.86job_tenure2 .0011 5.86cf .0127 .41east -.0053 -.22Dummy_95 -.1019 -24.90Dummy_96 -.1319 -35.96Dummy_97 -.1337 -35.67Dummy_98 -.1198 -31.04Dummy_99 -.1071 -26.51constant .2921 311.34

Data source: ECHP, German data files 1994-1999. Samples of 25-55 year old women,who are employed at lesast 8 hours per week.Note: Variables in differences.

90

Table 35: OLS wage estimation for EU women 1998lwage_grossm Coef. T-stat.lage 3.8151 8.84lage2 -.5094 -8.54job_magr .1332 8.10job_prof .1850 18.88job_wagr -.4135 -12.21job_cler -.1075 -12.43job_sale -.2780 -28.01job_wser -.2912 -18.17job_wqua -.2715 -16.72job_welm -.3482 -29.65job_public .1363 2.83job_break -.0434 -5.56sizehh_1415 -.0211 -3.05mar -.0254 -2.58mar_coh .0334 3.13job_pc .1181 14.45hours_pt .0290 3.52job_super .0767 5.69job_nonsup -.0380 -4.68job_sizelow -.1410 -19.94job_sizemid -.0601 -8.40educ_3 .0942 1.89educ_2 -.0769 -1.47job_tenure .0088 16.37citizen_EU .0536 2.00d_d .2556 6.77d_dk .6064 11.78d_nl .5354 16.28d_bel .3626 8.44d_uk .1216 3.31d_irl .0163 .43d_i .1955 5.60d_gr -.2049 -5.24d_esp -.0761 -2.18d_po -.5842 -17.44d_a .1315 3.36d_fin .1155 2.92

to be continued...

91

Appendix 3a: Estimation results for the EU

Table 36: ... Table 35 continuedlwage_grossm Coef. T-stat.d_e2 .0894 1.59dk_e2 .0936 1.38nl_e2 .0759 .83bel_e2 .1060 1.69uk_e2 .1636 2.67irl_e2 .2586 4.36i_e2 .1219 2.19gr_e2 .1085 1.81esp_e2 .2044 3.54po_e2 .2971 5.18a_e2 .1448 2.48fin_e2 .1204 2.02d_e3 -.0920 -1.64dk_e3 -.0651 -.96nl_e3 -.1845 -2.53bel_e3 -.0131 -.22uk_e3 .0450 .82irl_e3 .3457 5.80i_e3 .1019 1.80gr_e3 .1325 2.23esp_e3 .1092 2.00po_e3 .5586 10.04a_e3 .0395 .62fin_e3 -.0095 -.17d_pub -.0838 -1.64dk_pub -.1639 -2.98nl_pub -.0964 -1.88bel_pub -.1132 -2.12uk_pub -.0567 -1.09irl_pub .0474 .86i_pub -.0821 -1.60gr_pub .0369 .68esp_pub -.0011 -.02po_pub .0792 1.52a_pub -.0945 -1.77fin_pub -.2084 -4.00constant -5.1896 -6.67# obs 13989R-squared 0.6527Adj R-squared 0.6509

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed at least 8 hours per week.

92

Table 37: OLS wage estimation for EU men 1998

lwage_grossm Coef. T-stat.lage 2.0576 5.03lage2 -.2490 -4.42job_magr .1171 10.35job_prof .1431 14.38job_wagr -.3800 -18.41job_cler -.0967 -9.45job_sale -.1932 -17.17job_wser -.1581 -17.80job_wqua -.1915 -19.30job_welm -.2900 -24.94job_public .0384 .74job_break -.0844 -10.76sizehh_1415 .0054 1.95mar .0173 1.85mar_coh .0422 4.14job_pc .0998 11.52hours_pt -.0387 -1.82job_super .1119 11.70job_nonsup -.0429 -5.97job_sizelow -.2007 -31.23job_sizemid -.0945 -15.09educ_3 .0795 1.67educ_2 .0071 .14job_tenure .0064 14.03citizen_EU .0688 3.08d_d .3003 8.69d_dk .4657 10.80d_nl .4853 15.94d_bel .2367 6.34d_uk .1297 3.85d_irl .0963 2.73d_i .1044 3.34d_gr -.2852 -8.45d_esp -.1035 -3.33d_po -.6080 -19.80d_a .1007 2.52d_fin .0604 1.58

to be continued...

93

Table 38: ... Table 37 continuedlwage_grossm Coef. T-stat.d_e2 -.0280 -.52dk_e2 .0422 .69nl_e2 -.0653 -.74bel_e2 .0130 .23uk_e2 .0496 .80irl_e2 .0485 .87i_e2 .0401 .77gr_e2 .0876 1.59esp_e2 .1258 2.36po_e2 .2481 4.46a_e2 .0394 .68fin_e2 .0107 .19d_e3 -.0433 -.83dk_e3 -.0077 -.13nl_e3 -.1034 -1.59bel_e3 .0317 .58uk_e3 .0618 1.21irl_e3 .1806 3.25i_e3 .1445 2.73gr_e3 .0955 1.76esp_e3 .1686 3.32po_e3 .4819 8.70a_e3 .1546 2.42fin_e3 .0302 .54d_pub -.1015 -1.86dk_pub -.1603 -2.57nl_pub -.0371 -.68bel_pub -.0674 -1.20uk_pub -.0235 -.42irl_pub .1419 2.49i_pub -.0301 -.56gr_pub .0835 1.49esp_pub .0231 .42po_pub .1021 1.83a_pub -.1274 -2.27fin_pub -.0875 -1.55constant -2.0465 -2.78# obs. 18558R-squared 0.6409Adj R-squared 0.6394

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old men, who are employed at least 8 hours per week.

94

Table 39: Probit participation estimation for EU women 1998

job_part Coef. T-stat.a_y .6985 18.73a_m .4578 18.67lage_ch0 1.2989 4.52lage_ch111 .9336 8.94mar -.3655 -8.33mar_coh -.1183 -2.29lincome_ne -.1679 -14.14child_0 -5.4986 -5.52child_111 -3.9275 -10.42educ_3 1.1475 39.36educ_2 .6379 27.79d_d .0447 1.05d_dk 1.1678 12.53d_nl .7473 15.87d_bel .0717 1.30d_uk .1044 2.16d_irl -.3532 -7.46d_i -.4936 -12.53d_gr -.7393 -16.33d_esp -.6583 -15.81d_po .1913 4.17d_a -.0992 -2.01d_fin .7164 11.82constant 2.1085 17.26obs 24857Pseudo R2 0.2357Log pseudo-likelihood -11783.176

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed or out of the labour force.

95

Table 40: Heckman wage estimation for EU women 1998

lwage_grossm Coef. T-stat.lage 3.777 8.04lage2 -.5070 -7.78job_magr .1339 8.15job_prof .1843 18.82job_wagr -.4137 -12.23job_cler -.1068 -12.36job_sale -.2766 -27.87job_wser -.2875 -17.95job_wqua -.2701 -16.64job_welm -.3453 -29.39job_public .1336 2.78job_break -.0434 -5.56sizehh_015 -.0070 -1.89job_pc .1162 14.18hours_pt .0307 3.62job_super .0791 5.87job_nonsup -.0362 -4.45job_sizelow -.1409 -19.93job_sizemid -.0605 -8.46educ_3 .1795 2.78educ_2 -.0359 -.65job_tenure .0088 16.51citizen_EU .0536 2.00d_d .4755 6.64d_dk .7283 8.31d_nl .6338 9.69d_bel .4556 5.67d_uk .2435 3.29d_irl .1119 1.42d_i .2617 3.79d_gr -.1953 -2.53d_esp -.1060 -1.48d_po -.5207 -7.64d_a .2713 3.55d_fin .2745 3.39

to be continued...

96

Table 41: ... Table 40 continuedlwage_grossm Coef. T-stat.d_lmd -.3540 -3.58dk_lmd -.2033 -.83nl_lmd -.1173 -1.17bel_lmd -.1508 -1.38uk_lmd -.1907 -1.80irl_lmd -.1683 -1.68i_lmd -.1340 -1.44gr_lmd -.0866 -.89esp_lmd -.0421 -.45po_lmd -.0905 -.91a_lmd -.2195 -2.16fin_lmd -.2934 -1.94d_e2 -.0036 -.06dk_e2 .0549 .72nl_e2 .0554 .59bel_e2 .0796 1.17uk_e2 .1211 1.83irl_e2 .2278 3.51i_e2 .1000 1.68gr_e2 .1052 1.64esp_e2 .2313 3.71po_e2 .2826 4.58a_e2 .0941 1.48fin_e2 .0581 .87d_e3 -.2489 -3.46dk_e3 -.1488 -1.69nl_e3 -.2487 -2.94bel_e3 -.0785 -1.01uk_e3 -.0419 -.58irl_e3 .2740 3.53i_e3 .0489 .68gr_e3 .1097 1.45esp_e3 .1248 1.75po_e3 .5127 7.20a_e3 -.0603 -.76fin_e3 -.1240 -1.59

to be continued...

97

Table 42: ... Table 41 continuedlwage_grossm Coef. T-stat.d_pub -.0835 -1.64dk_pub -.1587 -2.89nl_pub -.0917 -1.79bel_pub -.1094 -2.06uk_pub -.0551 -1.06irl_pub .0529 .96i_pub -.0790 -1.54gr_pub .0383 .71esp_pub -.0079 -.15po_pub .0817 1.57a_pub -.0922 -1.73fin_pub -.2054 -3.95lmbd .1880 2.10constant -5.1705 -6.12obs 13988R-squared 0.6539Adj R-squared 0.6517

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed at least 8 hours per week.

98

Table 43: Propensity score wage estimation for women in the EU 1998

lwage_grossm Coef. T-stat.lage 3.698 7.90lage2 -.4952 -7.63job_magr .1338 8.15job_prof .1842 18.81job_wagr -.4141 -12.24job_cler -.1065 -12.31job_sale -.2763 -27.84job_wser -.2874 -17.94job_wqua -.2701 -16.65job_welm -.3461 -29.47job_public .1379 2.87job_break -.0434 -5.63sizehh_015 -.0053 -1.43job_pc .1164 14.22hours_pt .0311 3.68job_super .0790 5.83job_nonsup -.0369 -4.54job_sizelow -.1414 -20.01job_sizemid -.0607 -8.48educ_3 -.1998 2.96educ_2 -.0379 -.69job_tenure .0088 16.40citizen_EU .0520 1.94d_d .2036 5.04d_dk .5784 9.72d_nl .5204 13.76d_bel .3351 7.37d_uk .8743 2.20d_irl -.0202 -.50d_i .1614 4.30d_gr -.2517 -6.05d_esp -.1210 -3.22d_po -.5963 -16.06d_a .0980 2.36d_fin 0.0647 1.39d_hat .1438 3.61dk_hat .0870 2.08nl_hat .0762 1.94bel_hat .0780 1.90uk_hat .1019 2.52irl_hat .0784 1.86

to be continued...99

Table 44: ... Table 43 continuedlwage_grossm Coef. T-stat.i_hat .0664 1.68gr_hat .0438 1.06esp_hat .0346 .86po_hat .0526 1.23a_hat .1052 2.56fin_hat .1144 2.57d_e2 .0106 .18dk_e2 .0592 .84nl_e2 .0549 .58bel_e2 .0768 1.17uk_e2 .1141 1.78irl_e2 .2295 3.64i_e2 .0974 1.67gr_e2 .0979 1.56esp_e2 .2096 3.44po_e2 .2860 4.71a_e2 .0939 1.52fin_e2 .0642 1.01d_e3 -.2601 -3.54dk_e3 -.1652 -1.96nl_e3 -.2669 -3.03bel_e3 -.1037 -1.34uk_e3 -.0790 -1.08irl_e3 .2551 3.23i_e3 .0225 .31gr_e3 .0789 1.02esp_e3 .0790 1.07po_e3 .4998 6.73a_e3 -.0861 -1.07fin_e3 -.1432 -1.85

to be continued...

100

Table 45: ... Table 44 continuedlwage_grossm Coef. T-stat.d_pub -.0879 -1.73dk_pub -.1634 -2.98nl_pub -.0961 -1.88bel_pub -.1144 -2.15uk_pub -.0597 -1.15irl_pub .0479 .87i_pub -.0837 -1.64gr_pub .0331 .61esp_pub -.0122 -.23po_pub .0757 1.46a_pub -.0970 -1.82fin_pub -.2089 -4.02xbhat -.0877 -2.31constant -5.1113 -6.06obs 13988R-squared 0.6539Adj R-squared 0.6517

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed at least 8 hours per week.

101

Table 46: Lewbel regression for the EU 1998, 1. stage: density function forwomen

Coef. T-stat.lage -6.3248 -1.51lage2 1.0362 1.78educ_3 1.9617 9.58educ_2 .7874 4.12child_0 .4497 3.34child_111 .1635 2.01sizehh_015 -.1559 -4.23citizen_EU -.6153 -2.70citizen_other -1.5399 -8.73mar_coh -3.0473 -3.09mar_coh×lage .8149 3.00d_d -.2732 -1.27d_dk -2.5853 -6.66d_nl -2.6972 -17.18d_bel -2.8033 -10.23d_uk -2.1450 -10.72d_irl -4.0183 -19.04d_i -4.6653 -28.73d_gr -2.6346 -13.79d_esp -.2430 -1.47d_po -4.3211 -25.98d_a .1002 .41d_fin -2.3532 -8.63d_e2 -.3248 -1.19dk_e2 -.2129 -.48nl_e2 -1.3854 -1.59bel_e2 .1571 .42uk_e2 .1158 .34irl_e2 .1769 .59i_e2 .2781 1.16gr_e2 .6700 2.34esp_e2 .1592 .57po_e2 .5598 1.70a_e2 -.2813 -.90fin_e2 -.8281 -2.34

to be continued...

102

Table 47: ... table 46 continuedCoef. T-stat.

d_e3 -1.1578 -3.65dk_e3 -.9854 -2.11nl_e3 -1.4204 -2.13bel_e3 .4460 1.25uk_e3 -.7257 -2.61irl_e3 -.0153 -.04i_e3 -.5237 -1.64gr_e3 .4372 1.33esp_e3 -.1364 -.50po_e3 -.4210 -1.29a_e3 -1.8024 -4.06fin_e3 -.1726 -.50constant 15.0582 2.00

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed or out of the labour force.

103

Table 48: Lewbel regression for the EU 1998, 2. stage: selectivity-correctedwage estimation for women

lwage_grossm Coef. T-stat.job_magr .2395 4.25job_prof .2523 6.70job_cler -.1175 -2.79job_sale -.3524 -7.07job_wagr -.5657 -3.65job_wser -.5019 -7.30job_wqua -.4299 -5.82job_welm -.4716 -8.07job_sizelow -.1837 -5.91job_sizemid -.0774 -2.62educ_3 .1341 2.92educ_2 .0365 .94job_tenure .0174 9.26d_d .2666 1.93d_dk .5416 3.80d_nl .5172 3.47d_bel .3082 2.25d_uk .1602 1.16d_irl .2228 1.50d_i .2520 1.68d_gr -.1147 -.84d_esp -.0436 -.32d_po -.3235 -2.40d_a .1994 1.43d_fin .0220 .16constant 1.9984 12.37

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old women, who are employed at least 8 hours per week.

104

Table 49: Lewbel regression for the EU 1998, 1. stage: density function for men

Coef. T-stat.job_break -.2583 -3.26educ_3 2.0129 9.05educ_2 .6932 3.50mar_coh -4.7427 -9.19mar_coh×lage 1.3055 9.43sizehh_015 -.15334 -5.75citizen_other -1.1656 -6.14citizen_EU -.7116 -3.11d_d -.4458 -1.81d_dk -3.2080 -8.99d_nl -2.6647 -15.88d_bel -3.1749 -11.49d_uk -2.1932 -9.42d_irl -3.7971 -14.96d_i -4.7060 -25.49d_gr -3.5549 -15.16d_esp -.4179 2.26d_po -4.5140 -25.52d_a -.0084 -.03d_fin -2.5494 -9.01d_e2 -.2972 -1.01dk_e2 .3341 .81nl_e2 -.8779 -1.07bel_e2 .4744 1.30uk_e2 -.2229 -.50irl_e2 .2259 .66i_e2 .1293 .51gr_e2 .8410 2.61esp_e2 .2244 .79po_e2 .5989 1.79a_e2 -.2667 -0.71fin_e2 -.2868 -0.82d_e3 -.8497 -2.58dk_e3 .0152 .03nl_e3 -1.8131 -3.31bel_e3 1.0342 2.82uk_e3 -.6898 -2.25irl_e3 -.5093 -1.33i_e3 .1939 .58

to be continued...

105

Table 50: ...table 49 continuedCoef. T-stat.

gr_e3 .5270 1.50esp_e3 .1023 .35po_e3 .4349 1.15a_e3 -.7027 -1.36fin_e3 .1735 .47constant 5.8004 37.82

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old men, who are employed or out of the labour force.

106

Table 51: Lewbel regression for the EU 1998, 2. stage: selectivity-correctedwage estimation for men

lwage_grossm Coef. T-stat.job_magr .2007 6.41job_prof .1693 6.61job_cler -.1120 -4.25job_sale -.2701 -9.23job_wagr -.4709 -8.41job_wser -.2017 -7.76job_wqua -.2386 -8.35job_welm -.3355 -9.65job_sizelow -.2294 -12.79job_sizemid -.1141 -6.28job_break -.1332 -6.80sizehh_015 .0132 2.10educ_3 .1719 7.97educ_2 .0770 4.33job_tenure .0160 15.33d_d .2372 3.61d_dk .4683 6.55d_nl .4910 6.67d_bel .2406 3.67d_uk .1825 2.52d_irl .1773 2.40d_i .1804 2.55d_gr -.1847 -2.96d_esp -.0629 -.99d_po -.4061 -6.32d_a .0837 1.23d_fin .0212 .31constant 2.1877 31.58

Data source: ECHP, pooled EU data file 1998 (excl. Luxembourg and Sweden).Sample of 25-55 year old men, who are employed at least 8 hours per week.

107

108

Table 52: OLS wage estimation for EU women 1998 (using sector information)lwage_grossm Coef. T-stat.lage 3.5655 7.57lage2 -.4640 -7.07job_magr .1023 5.45job_prof .1505 13.17job_cler -.1246 12.15job_sale -.2503 21.26job_wagr -.3365 -7.50job_wser -.3240 16.53job_wqua -.3001 14.97job_welm -.3287 24.01job_public .0672 7.13job_break -.0368 -4.29mar -.0308 -2.71mar_coh .4256 3.97mar_coh×lage -.1055 -3.51job_pc .1297 14.53hours_pt .0387 4.27job_super .0820 5.74job_nonsup -.0507 -5.98job_sizelow -.1264 15.80job_sizemid -.0648 -7.90educ_3 .1160 2.37educ_2 -.0719 -1.40job_tenure .0087 14.73sec_erng .2217 4.41sec_manu .1262 3.95sec_cons .1395 3.40sec_rept .0552 1.74sec_hotl .0522 1.55sec_trns .1505 4.33sec_finc .2650 7.89sec_busn .1154 3.57sec_publ .1442 4.42sec_educ .1883 5.85sec_hlth .0886 2.81sec_othr .0328 1.04d_dk .5358 10.38d_nl .5047 16.24d_bel .3212 7.75d_uk .1041 2.94d_irl .0278 .75d_i .1855 5.53d_gr -.1970 -5.22d_esp -.0740 -2.21

to be continued

109

Appendix 3b: Estimation results for the EU using sectorinformation

Table 53: ... Table 52 continuedlwage_grossm Coef. T-stat.d_po -.5654 17.68d_a .1041 2.77d_fin .0783 1.19dk_e2 .0906 1.31nl_e2 .0708 .78bel_e2 .0978 1.57uk_e2 .1534 2.52irl_e2 .2576 4.4i_e2 .0955 1.75gr_e2 .1167 1.97esp_e2 .1977 3.47po_e2 .3025 5.34a_e2 .1390 2.41fin_e2 .0770 .83dk_e3 -.0876 -1.28nl_e3 -.1896 -2.59bel_e3 -.0508 -.87uk_e3 .0161 .3irl_e3 .3512 6.09i_e3 .0435 .79gr_e3 .1344 2.34esp_e3 .1029 1.94po_e3 .5861 10.88a_e3 -.0130 -.21fin_e3 .0064 .08_con -4.9841 -5.89

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed at least 8 hours perweek.

110

Table 54: OLS wage estimation for EU men 1998 (using sector information)

lwage_grossm Coef. T-stat.lage 2.0917 5.03lage2 -.2492 -4.35job_magr .0865 6.79job_prof .1403 12.27job_cler -.1227 11.10job_sale -.1312 10.35job_wagr -.2505 -9.85job_wser -.1592 15.19job_wqua -.1957 17.24job_welm -.2778 21.74job_public .0408 4.17job_break -.0746 -8.75mar .0520 7.88job_pc .1130 12.44job_super .1268 12.66job_nonsup -.0471 -6.39job_sizelow -.1678 23.50job_sizemid -.0721 10.23educ_3 .1014 2.28educ_2 .0780 9.81job_tenure .0058 11.52ser_erng .3049 11.80ser_manu .2114 10.28ser_cons .2080 9.82ser_rept .1338 6.30ser_hotl .0155 .59ser_trns .2219 10.15ser_finc .4125 17.36ser_busn .2306 10.26ser_publ .2007 8.84ser_educ .1927 7.84ser_hlth .1532 6.27ser_othr .1906 8.04d_dk .4301 15.56d_nl .4928 20.28d_bel .1913 7.10d_uk .1361 5.01d_irl .1318 5.04d_i .0848 3.52d_gr -.2497 -9.72

to be continued

111

Table 55: ... Table 54 continuedlwage_grossm Coef. T-stat.d_esp -.0715 -2.93d_po -.5540 22.72d_a .0470 1.86d_fin .0378 1.04dk_e3 -.0384 -.72nl_e3 -.1228 -1.97bel_e3 .0349 .71uk_e3 .0261 .55irl_e3 .1896 3.75i_e3 .1283 2.61gr_e3 .0775 1.57esp_e3 .1262 2.68po_e3 .4409 8.46a_e3 .1459 2.61fin_e3 .0858 1.36_cons -2.3899 -3.18

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old men, who are employed at least 8 hours per week.

112

Table 56: Probit participation estimation for EU women 1998 (using sectorinformation)

job_part Coef. T-stat.a_y .7041 17.67a_m .4763 18.34lage_ch0 1.5629 5.19lage_ch111 1.0349 9.33mar -.2787 -5.93mar_coh -.2151 -3.90lincome_ne -.1104 11.09sh -.1221 13.54child_0 -6.1955 -5.93child_111 -4.1275 10.32educ_3 1.0858 35.73educ_2 .6011 24.41citizen_other -.3402 -3.87citizen_EU -.3899 -3.36d_dk 1.1631 12.47d_nl .7050 14.88d_bel .0967 1.75d_uk .1144 2.37d_irl -.2656 -5.48d_i -.4630 11.65d_gr -.6814 14.94d_esp -.6024 14.30d_po .2715 5.84d_a -.0575 -1.15d_fin .7091 11.68constant 1.9265 19.02

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed or out of the labourforce.

113

Table 57: Heckman wage estimation for EU women 1998 (using sector informa-tion)

lwage_grossm Coef. T-stat.lage 3.5655 7.57lage2 -.4642 -7.08job_magr .1025 5.46job_prof .1503 13.15job_cler -.1249 12.18job_sale -.2501 21.24job_wagr -.3363 -7.50job_wser -.3239 16.53job_wqua -.3002 14.97job_welm -.3290 24.02job_public .0672 7.13job_break -.0369 -4.30educ_3 .1281 2.58educ_2 -.0645 -1.25job_tenure .0087 14.76mar -.0353 -3.01mar_coh .4387 4.09mar_coh×lage -.1103 -3.66job_pc .1302 14.58hours_pt .0374 4.10job_super .0825 5.77job_nonsup -.0504 -5.93job_sizelow -.1263 15.79job_sizemid -.0646 -7.87ser_erng .2220 4.42ser_manu .1264 3.96ser_cons .1398 3.41ser_rept .0553 1.74ser_hotl .0504 1.50ser_trns .1507 4.34ser_finc .2655 7.91ser_busn .1158 3.58ser_publ .1440 4.41ser_educ .1882 5.85ser_hlth .0884 2.80ser_othr .0329 1.04d_dk .5495 10.49d_nl .5155 16.19d_bel .3234 7.80d_uk .1073 3.02d_irl .0220 .60

to be continued... 114

Table 58: ... Table 57 continuedlwage_grossm Coef. T-stat.d_i .1794 5.31d_gr -.2067 -5.40d_esp -.0832 -2.44d_po -.5610 17.48d_a .1026 2.73d_fin .0882 1.33dk_e2 .0860 1.25nl_e2 .0669 .73bel_e2 .0978 1.57uk_e2 .1530 2.52irl_e2 .2602 4.45i_e2 .0977 1.79gr_e2 .1174 1.98esp_e2 .2020 3.54po_e2 .3020 5.33a_e2 .1408 2.44fin_e2 .0743 .80dk_e3 -.0957 -1.39nl_e3 -.1972 -2.69bel_e3 -.0517 -.88uk_e3 .0149 .28irl_e3 .3549 6.15i_e3 .0468 .85gr_e3 .1383 2.41esp_e3 .1084 2.04po_e3 .5840 10.84a_e3 -.0112 -.18fin_e3 .0010 .01lmbd .0280 1.58_cons -4.9952 -5.91

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed at least 8 hours perweek.

115

Table 59: Propensity score wage estimation for women in the EU 1998 (usingsector information)

lwage_grossm Coef. T-stat.lage 3.5631 7.56lage2 -.4636 -7.07job_magr .1022 5.45job_prof .1503 13.16job_cler -.1248 12.17job_sale -.2500 21.22job_wagr -.3363 -7.50job_wser -.3239 16.53job_wqua -.3001 14.96job_welm -.3285 23.98job_public .0672 7.13job_break -.0371 -4.31mar -.0303 -2.56mar_coh .4279 3.99mar_coh×lage -.1060 -3.52job_pc .1301 14.56hours_pt .0389 4.26job_super .0820 5.74job_nonsup -.0508 -5.99job_sizelow -.1264 15.81job_sizemid -.0647 -7.89educ_3 .1150 2.31educ_2 -.0724 -1.40job_tenure .0086 14.71ser_erng .2222 4.42ser_manu .1263 3.96ser_cons .1398 3.41ser_rept .0554 1.75ser_hotl .0509 1.52ser_trns .1507 4.34ser_finc .2653 7.90ser_busn .1156 3.58ser_publ .1444 4.42ser_educ .1885 5.86ser_hlth .0886 2.81ser_othr .0329 1.04d_dk .5344 10.17d_nl .5037 15.81d_bel .3209 7.73d_uk .1037 2.92d_irl .0280 .76

to be continued...116

Table 60: ... Table 59 continued

lwage_grossm Coef. T-stat.d_i .1857 5.52d_gr -.1964 -5.18d_esp -.0736 -2.18d_po -.5658 17.63d_a .1040 2.77d_fin .0775 1.17dk_e2 .0907 1.31nl_e2 .0709 .78bel_e2 .0978 1.57uk_e2 .1534 2.52irl_e2 .2571 4.39i_e2 .0955 1.75gr_e2 .1145 1.94esp_e2 .1976 3.47po_e2 .3025 5.34a_e2 .1389 2.41fin_e2 .0771 .83dk_e3 -.0874 -1.28nl_e3 -.1896 -2.59bel_e3 -.0506 -.86uk_e3 .0162 .30irl_e3 .3512 6.09i_e3 .0437 0.79gr_e3 .1344 2.34esp_e3 .1030 1.94pro_e3 .5862 10.88a_e3 -.0130 -.21fin_e3 .0066 .08xbhat .0009 .13_cons -4.9828 -5.89

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed at least 8 hours perweek.

117

Table 61: Lewbel regression for the EU 1998 (using sector information) , 1.stage: density function for women

Coef. T-stat.lage 1.1493 4.65educ_3 1.5773 7.60educ_2 .6023 3.11child_0 .4524 3.19child_111 .1720 1.98sizehh_015 -.1729 -4.55citizen_EU -.5677 -1.85citizen_other -1.6977 -7.21mar_coh -3.5392 -3.40mar_coh×lage .9361 3.27d_dk -1.0673 -2.71d_nl -1.1877 -7.46d_bel -1.2998 -4.68d_uk -.6313 -3.11d_irl -2.5016 11.69d_i -3.1547 19.14d_gr -1.1226 -5.79d_esp 1.2674 7.57d_po -2.8096 16.64d_a 1.6186 6.52d_fin -1.9227 -6.95dk_e2 -.0324 -.07nl_e2 -1.2041 -1.37bel_e2 .3408 .91uk_e2 .2994 .88irl_e2 .3590 1.18i_e2 .4561 1.88gr_e2 .8573 2.96esp_e2 .3366 1.19po_e2 .7450 2.24a_e2 -.1083 -.34fin_e2 -.5586 -1.56dk_e3 -.6164 -1.30nl_e3 -1.0127 -1.50bel_e3 .8345 2.30uk_e3 -.3412 -1.21irl_e3 .3646 1.02i_e3 -.1580 -.49gr_e3 .8133 2.44esp_e3 .2388 .86po_e3 -.0432 -.13a_e3 -1.4326 -3.19fin_e3 -.1605 -.46constant .1796 .20

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed or out of the labourforce.

118

Table 62: Lewbel regression for the EU 1998 (using sector information), 2.stage: selectivity-corrected wage estimation for women

lwage_grossm Coef. T-stat.job_magr .2304 3.20job_prof .2300 4.86job_cler -.1205 -2.08job_sale -.3228 -4.45job_wagr -.5098 -2.53job_wser -.4601 -4.34job_wqua -.3979 -3.76job_welm -.4614 -5.38job_public .1042 3.71job_pc .1186 2.15job_sizelow -.1616 -4.15job_sizemid -.0870 -2.44educ_3 .1406 2.18educ_2 .0319 .59job_tenure .0163 8.12d_dk .5734 3.01d_nl .5413 2.74d_bel .3130 1.70d_uk .1700 .92d_irl .2759 1.40d_i .2548 1.28d_gr -.1092 -.59d_esp -.0398 -.21d_po -.2761 -1.53d_a .2165 1.17d_fin .0103 .05constant 1.8367 7.87

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old women, who are employed at least 8 hours perweek.

119

Table 63: Lewbel regression for the EU 1998 (using sector information), 1.stage: density function for men

Coef. T-stat.job_break -.2022 -2.34educ_3 1.6678 7.37educ_2 .5308 2.63sizehh_015 -.1388 -4.75citizen_EU -.8092 -2.41citizen_other -1.4830 -5.58mar_coh -5.1928 -9.17mar_coh×lage 1.4573 9.27mar -.1837 -1.76d_dk -1.6692 -4.60d_nl -1.1224 -6.56d_bel -1.6295 -5.80d_uk -.6540 -2.76d_irl -2.2655 -8.76d_i -3.1591 16.74d_gr -2.0066 -8.39d_esp 1.1275 5.96d_po -2.9614 16.39d_a 1.5549 4.61d_fin -1.9926 -6.92dk_e2 .4820 1.15nl_e2 -.6955 -.83bel_e2 .6411 1.73uk_e2 -.0456 -.10irl_e2 .3979 1.14i_e2 .2941 1.14gr_e2 1.0056 3.07esp_e2 .3948 1.36po_e2 .7699 2.26a_e2 -.1129 -.29fin_e2 -.2153 -.60dk_e3 .3513 .77nl_e3 -1.5019 -2.69bel_e3 1.3795 3.69uk_e3 -.3383 -1.08irl_e3 -.1535 -.39i_e3 .5335 1.58gr_e3 .8730 2.44esp_e3 .4459 1.50po_e3 .7805 2.03a_e3 -.3507 -.67fin_e3 .0 00 8 . 0 0constant 4.2827 27.21

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old men, who are employed or out of the labour force.

Table 64: Lewbel regression for the EU 1998 (using sector information) , 2.stage: selectivity-corrected wage estimation for men

lwage_grossm Coef. T-stat.lage .3562 8.29job_magr .0487 1.33job_prof .1630 5.94job_cler -.1068 -3.94job_sale -.1515 -4.56job_wagr -.3102 -3.03job_wser -.1557 -5.09job_wqua -.1984 -6.23job_welm -.2881 -6.93job_break -.0748 -3.39mar .0328 1.63sizehh_015 .0152 2.09educ_3 .1881 8.04educ_2 .0978 4.81job_tenure .0061 4.26job_pc .1097 4.30job_super .1335 5.23job_nonsup -.0936 -4.94job_sizelow -.1768 -8.94job_sizemid -.0802 -4.10ser_erng .1756 1.69ser_manu .1656 1.69ser_cons .1570 1.58ser_rept .0540 .55ser_hotl -.1368 -1.32ser_trns .1887 1.93ser_finc .3369 3.37ser_busn .1425 1.41ser_publ .1968 2.02ser_educ .1326 1.32ser_hlth .1764 1.72ser_other .1512 1.49d_dk .4185 5.19d_nl .5374 6.77d_bel .2246 3.12d_uk .1574 2.06d_irl .2253 2.78d_i .1446 1.85d_gr -.1638 -2.34d_esp -.0341 -.48d_po -.3526 -4.95d_a .0923 1.26d_fin .0892 .90constant .6536 3.21

Data source: ECHP, pooled EU data file 1998 (excl. Germany, Luxembourg andSweden). Sample of 25-55 year old men, who are employed at least 8 hours per week.

122

Table 65: Lewbel regression for France 1998, 1. stage: density function forwomen

Coef. T-stat.lage 1.1807 4.56educ_3 1.5508 11.68educ_2 .5461 4.40sizehh_015 -.1653 -3.22citizen_other -1.9839 -5.88cititzen_EU -.7111 -2.03mar .3988 3.29constant -.1642 -.17

Data source: ECHP, French data file 1998. Sample of 25-55 year old women, who areemployed or out of the labour force.

Table 66: Lewbel regression for France 1998, 2. stage: selectivity-correctedwage estimation for women

lwage_grossm Coef. T-stat.job_magr -.1754 -1.14job_prof .2827 2.64job_cler -.1794 -1.62job_sale -.3650 -2.57job_wagr -.3828 -.51job_wser -.3628 -1.03job_wqua -.1755 -1.01job_welm -.3564 -2.05job_break -.0185 -.16educ_3 .2307 2.34educ_2 .0649 .67job_tenure .0253 4.56constant 1.9155 12.57

Data source: ECHP, French data file 1998. Sample of 25-55 year old women, who areemployed at least 8 hours per week.

123

App endix 4 a: Estima ti on results f or selected countri es

Table 67: Lewbel regression for France 1998, 1. stage: density function for menCoef. T-stat.

job_break -.6720 -2.80educ_3 1.5983 10.96educ_2 .4627 3.61hours_pt -.9488 -2.21sizehh_015 -.1572 -2.91citizen_other -1.5961 -3.71citizen_EU -1.1459 -3.25mar .3974 3.27constant 4.1621 34.16

Data source: ECHP, French data file 1998. Sample of 25-55 year old men, who areemployed or out of the labour force.

Table 68: Lewbel regression for France 1998, 2. stage: selectivity-correctedwage estimation for men

lwage_grossm Coef. T-stat.lage .2231 1.57job_magr .2980 3.26job_prof .4936 5.56job_cler -.1619 -1.53job_sale -.2803 -1.98job_wagr -.2486 -.75job_wser -.1641 -1.84job_wqua -.0934 -.94job_welm -.2845 -2.00job_break -.1570 -1.87educ_3 .1739 2.34educ_2 .0603 .88job_tenure .0146 2.69idf .1406 1.82constant 1.2369 2.56

Data source: ECHP, French data file 1998. Sample of 25-55 year old men, who areemployed at least 8 hours per week.

124

Table 69: Lewbel regression for Spain 1998, 1. stage: density function forwomen

Coef. T-stat.lage 2.7356 6.06edu_3 1.8269 7.40educ_2 .9390 3.51mar_coh -.9236 -3.65hours_pt -.8006 -1.97constant -3.7096 -2.23

Data source: ECHP, Spanish data file 1998. Sample of 25-55 year old women, whoare employed or out of the labour force.

Table 70: Lewbel regression for Spain 1998, 2. stage: selectivity-corrected wageestimation for women

lwage_grossm Coef. T-stat.job_magr .4119 1.36job_prof .3321 2.46job_clerc -.1982 -1.33job_sale -.3989 -2.55job_wagr -.4766 -2.25job_wser -.4211 -2.79job_wqua -.2372 -1.09job_welm -.4172 -2.13job_public .1347 2.20job_tenure .0103 2.63job_pc .1831 2.21job_sizelow -.1719 -2.26job_sizemid -.0296 -.53_cons 1.8860 7.42

Data source: ECHP, Spanish data file 1998. Sample of 25-55 year old women, whoare employed at least 8 hours per week.

125

Table 71: Lewbel regression for Spain 1998, 1. stage: density function for menCoef. T-stat.

lage 2.2822 4.46job_break .6428 2.53educ_3 1.8017 6.87educ_2 .7405 2.69hours_pt 1.7166 1.91mar_coh -1.3775 -5.03job_nonsup -1.0989 -4.25job_super .6070 1.54citizien_other -4.1339 -2.08_cons -1.3364 -.72

Data source: ECHP, Spanish data file 1998. Sample of 25-55 year old men, who areemployed or out of the labour force.

126

Table 72: Lewbel regression for Spain 1998, 2. stage: selectivity-corrected wageestimation for men

lwage_grossm Coef. T-stat.lage .2963 3.97job_magr .3933 4.47job_prof .2933 4.63job_clerc -.0021 -.03job_sale -.0576 -.77job_wagr -.2234 -1.69job_wser -.1145 -2.01job_wqua -.0961 -1.49job_welm -.2555 -3.96job_break -.1348 -3.94educ_3 .1936 4.48educ_2 .1111 2.59job_tenure .0066 2.16job_pc .1720 4.55job_super .0489 1.16job_nonsup -.1002 -2.68job_sizelow -.3104 -7.98job_sizemid -.1590 -3.79hours_pt .2820 5.60ser_erng .2180 1.52ser_manu .2400 2.45ser_cons .3494 3.37ser_rept .0341 .34ser_hotl -.0359 -.32ser_trns .1160 1.18ser_finc .3073 2.26ser_busn .1654 1.62ser_publ .2174 2.07ser_educ .2200 1.79ser_hlth .1795 1.39ser_other .0935 1.00cons .8319 2.74

Data source: ECHP, Spanish data file 1998. Sample of 25-55 year old men, who areemployed at least 8 hours per week.

127

Table 73: Lewbel regression for the UK 1998, 1. stage: density function forwomen

Coef. T-stat.job_break -.8911 -3.11educ_3 1.1489 8.33educ_2 .8945 4.57job_sizelow -.3743 -2.55job_sizemid -.0033 -.02job_nonsup .0869 .61job_super .3946 2.03sizehh_015 -.2533 -4.45mar_coh×lage 1.9448 6.25mar_coh -6.3474 -5.54citizen_other -.7048 -2.10constant 3.2594 17.02

Data source: ECHP, British data file 1998. Sample of 25-55 year old women, who areemployed or out of the labour force.

Table 74: Lewbel regression for the UK 1998, 2. stage: selectivity-correctedwage estimation for women

lwage_grossm Coef. T-stat.lage 10.4042 4.19lage2 -1.4440 -4.25job_magr .0535 .74job_prof .1503 2.09job_cler -.2265 -3.32job_sale -.5296 -7.94job_wagr .0164 .05job_wser -.6309 -2.95job_wqua -.6267 -6.51job_welm -.6469 -7.81job_sizelow -.1899 -5.14job_sizemid -.0537 -1.26educ_3 .1674 3.91educ_2 .1296 2.31job_tenure .0096 2.54constant -16.4137 -3.64

Data source: ECHP, British data file 1998. Sample of 25-55 year old women, who areemployed at least 8 hours per week.

128

Table 75: Lewbel regression for the UK 1998, 1. stage: density function for menCoef. T-stat.

job_break -1.3806 -5.91educ_3 1.1277 7.42educ_2 .3812 1.34job_sizelow -.4429 -2.92job_sizemid -.2791 -1.80job_nonsup .0140 .08job_super .8095 4.30sizehh_015 -.2192 -3.52mar_coh×lage 1.7097 5.26mar_coh -5.8909 -4.91constant 3.5437 14.88

Data source: ECHP, British data file 1998. Sample of 25-55 year old men, who areemployed or out of the labour force.

Table 76: Lewbel regression for the UK 1998, 2. stage: selectivity-correctedwage estimation for men

lwage_grossm Coef. T-stat.lage 6.7100 1.96lage2 -.9088 -1.92job_magr -.0363 -.46job_prof .0214 .23job_cler -.3825 -3.51job_sale -.4933 -5.79job_wagr -.4962 -2.09job_wser -.3143 -3.48job_wqua -.4158 -4.25job_welm -.5387 -4.83job_break -.2044 -3.94educ_3 .1637 2.87educ_2 .1483 1.45job_sizelow -.1929 -3.39job_sizemid -.1712 -3.05constant -9.7252 -1.56

Data source: ECHP, British data file 1998. Sample of 25-55 year old men, who areemployed at least 8 hours per week.

129

Table 77: Lewbel regression for Italy 1998, 1. stage: density function for women

Coef. T-stat.educ_3 1.3740 6.76educ_2 .9533 7.79mar_coh×lage 1.1738 4.01mar_coh -4.3230 -3.99constant 1.1331 7.25

Data source: ECHP, Italian data file 1998. Sample of 25-55 year old women, who areemployed or out of the labour force.

Table 78: Lewbel regression for Italy 1998, 2. stage: selectivity-corrected wageestimation for women

lwage_grossm Coef. T-stat.lage .5049 2.63job_magr .3660 1.53job_prof .3608 1.48job_cler -.1167 -.50job_sale -.2276 -.93job_wagr -.6543 -.95job_wser -.3993 -1.68job_wqua -.3718 -1.03job_welm -.5598 -1.86constant .4827 .71

Data source: ECHP, Italian data file 1998. Sample of 25-55 year old women, who areemployed at least 8 hours per week.

130

Table 79: Lewbel regression for Italy 1998, 1. stage: density function for menCoef. T-stat.

job_super .6304 3.21job_sizelow -.6066 -4.01job_sizemid -.2824 -1.69educ_3 1.9666 9.33educ_2 .6657 4.89child_111 -.3509 -2.49mar -.8708 -1.77mar_coh×lage .1959 1.45constant 1.5665 8.23

Data source: ECHP, Italian data file 1998. Sample of 25-55 year old men, who areemployed or out of the labour force.

Table 80: Lewbel regression for Italy 1998, 2. stage: selectivity-corrected wageestimation for men

lwage_grossm Coef. T-stat.lage .2661 2.08job_magr .3425 4.59job_prof .1553 2.32job_cler .0616 .99job_sale .0161 .15job_wagr -.2010 -.81job_wser -.0759 -.83job_wqua -.0444 -.41job_welm -.0559 -.58job_super .0798 1.76job_sizelow -.1181 -2.16job_sizemid -.0701 -1.50educ_3 .2641 3.16educ_2 .1180 1.67job_tenure .0108 2.66constant 1.1492 2.58

Data source: ECHP, Italian data file 1998. Sample of 25-55 year old men, who areemployed at least 8 hours per week.

131

132

-12 -8 -4 0 4 8 12

Paris

elementary occupations

plant and machineoperators andassemblers

craft and related tradesworkers

skilled agricultural andfishery workers

service and salesworkers

clerks

professionals

legislators, seniorofficials and managers

job interruption(unemployment)

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

Figure 28: Contributions to the gender pay gap in FranceData source: ECHP, French data file 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation.The first bar ”age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

133

Appendix 4b: Decomposition by factors for selected coun-tries

-12 -8 -4 0 4 8 12

other sector

health and social work

education

public administration

real state, renting and business activities

financial intermediation

transport, storage and communication

hotels and restaurants

wholesale and retail trade

construction

manufacture

electricity, gas and water supply

elementary occupations

plant and machine operators and assemblers

craft and related trades workers

skilled agricultural and fishery workers

service and sales workers

clerks

professionals

legislators, senior officials and managers

middle size firm

small size firm

non-supervisory job status

supervisory job status

part time employment

permant job contract

public sector

job interruption

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

Figure 29: Contributions to the gender pay gap in SpainData source: ECHP, Spanish data file 1998. Sample of 25-55 year old women andmen, who are employed at least 8 hours per week.Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation.The first bar ”age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

134

-12 -8 -4 0 4 8 12

elementary occupations

plant and machine operators and assemblers

craft and related trades workers

skilled agricultural and fishery workers

service and sales workers

clerks

professionals

legislators, senior officials and managers

middle size firm

small size firm

job interruption (unemployment)

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

36.09

Figure 30: Contributions to the gender pay gap in the UKData source: ECHP, British data file 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation.The first bar ”age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

135

-12 -8 -4 0 4 8 12

number of children <16

elementary occupations

plant and machine operators and assemblers

craft and related trades workers

skilled agricultural and fishery workers

service and sales workers

clerks

professionals

legislators, senior officials and managers

non-supervisory job status

supervisory job status

permant job contract

firm tenure (years)

tertiary education

secondary education

age

Percent

endowment effectremuneration effect

-47.89

27.6

Figure 31: Contributions to the gender pay gap in ItalyData source: ECHP, Italian data file 1998. Sample of 25-55 year old women and men,who are employed at least 8 hours per week.Note: Oaxaca-Blinder decomposition based on the Lewbel wage estimation.The first bar ”age” includes both the pure age effect and the constants of the wageregressions. Naturally, the endowment effect is derived from the age only.

136