public–private sector wage gap in australia: variation

Public–Private Sector Wage Gap inAustralia: Variation along the Distributionbjir_773 1..29

Lixin Cai and Amy Y. C. Liu

Abstract

Previous research on public–private wage differentials in Australia has focusedon the mean of the conditional wage distribution. Using six waves of theHousehold, Income and Labour Dynamics in Australia survey, this studyemploys quantile regressions to examine whether the sectoral wage effect variesalong the wage distribution. For females, public sector wage premiums arerelatively stable for almost the entire distribution. For males, they decreasemonotonically and are negative for the top half of the distribution. The decom-position results show that the observed differences in individuals and job char-acteristics explain a substantial proportion of the sectoral wage gap.

1. Introduction

There are many reasons why public and private sectors workers can be paiddifferently. First, the public sector could set wages in a non-competitive waydue to the monopolistic power of governments in setting prices and taxes forthe provision of public services (Reder 1975). Second, the public sector maybe driven by objectives such as vote and/or budget maximization rather thanprofit maximization. Wages in the public sector may also be used to achieveother considerations, such as equity and fairness. Third, the institutionalenvironment for wage setting may differ between public and private sectors.For example, there could be an imperfect labour market in the public sector.Union density is often higher in the public sector than in the private sector.Consequently, unions may have a stronger bargaining power in securinghigher wages for public sector employees in a collective bargaining industrialframework. Fourth, productivity-related characteristics of employees in thetwo sectors may be different. If public sector employees are relatively skilled,they require higher remunerations.

Study of the public–private pay gap has important policy implications ona wide range of labour market issues. For example, higher wages to public

Lixin Cai is at The University of Melbourne. Amy Y. C. Liu is at The Australian NationalUniversity.

British Journal of Industrial Relations doi: 10.1111/j.1467-8543.2009.00773.x••:•• •• 2010 0007–1080 pp. ••–••

© Blackwell Publishing Ltd/London School of Economics 2010. Published by Blackwell Publishing Ltd,9600 Garsington Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.

sector employees may justify outsourcing of some government functions tothe private sector, and may potentially crowd out recruitment efforts in theprivate sector, forcing it to raise wages in order to compete for employees inthe labour market.

Earlier studies on the public–private wage differentials focus mostly on themean of the wage distribution. International evidence suggests that, relativeto the private sector, on average, there is a wage premium for public sectoremployees (between 3 and 11 per cent), and the premium is often higher forfemales than for males (see Borland and Gregory 1999 for a detailed review).Recently, increasingly more studies use quantile regressions to examinewhether the public–private earnings differentials vary along the earningsdistribution. They include, for instance, Blackaby et al. (1999) on the UK,Lucifora and Meurs (2006) on Italy, France and the UK, Melly (2005) onGermany, Mueller (1998) on Canada, and Poterba and Rueben (1994) on theUSA. Typically, these studies find lower pay dispersion in the public sector.Also, they find that public sector employees at the lower end of the wagedistribution enjoy a wage premium relative to private sector employees, butthe reverse holds for employees at the upper end. Further, female publicservants are often found to enjoy a premium across almost the entire wagedistribution, while male public sector employees suffer wage penalty over alarge part of the wage distribution.

Only a few studies have examined public–private earnings differentials inAustralia. Among them, they give conflicting evidence about whether theobserved earnings differentials are attributable to the sectoral effect. Forinstance, using the 1993 Training and Education Experience Survey collectedby the Australian Bureau of Statistics, Borland et al. (1998) show that, rela-tive to their counterparts, the average weekly earnings of full-time publicsector employees were 10–15 per cent higher for males, and 20–25 per centhigher for females. However, they find that the differentials can all beexplained by observed differences in productivity-related individual and jobcharacteristics, suggesting there is no sectoral effect for Australian workers.Moreover, using the 1985 Australian Longitudinal Survey data, Vella (1993)finds a significant wage premium for young female government employeesaged 15–26 years relative to their private sector counterparts even aftercontrolling for observed heterogeneity.

The historically unique feature of the Australian industrial relations(including wage-setting) system makes the topic interesting to an interna-tional audience. The Commonwealth Conciliation and Arbitration Act of1904 governed Australian industrial relations through a system of compul-sory arbitration and conciliation for much of the twentieth century. Underthe arbitration arrangements, unions typically made demands on employers,which would be resolved through tribunals that delivered decisions known as‘awards’ to resolve disputes and set employment conditions. The rates of payand conditions of employment specified in the awards were widely applicableto workers in relevant occupation or industry groups. Therefore, it is notsurprising to find that the wage gap between union and non-union workers

2 British Journal of Industrial Relations

© Blackwell Publishing Ltd/London School of Economics 2010.

and between males and females was traditionally smaller in Australia than inother OECD countries (e.g. Baron and Cobb-Clark 2008). In the case of thepublic–private sector wage gap, Borland et al. (1998) do not find a sectoraleffect for Australian workers in the arbitration era, while a public sector wagepremium is often found in other OECD countries.

Australia has been undergoing significant changes in its industrial relationssystem since the early 1990s. Through a ruling of the Australian IndustrialRelations Commission, Australian wage setting started to shift fromindustry-based awards towards enterprise-based (or workplace-based) agree-ments (Waddoups 2005). The introduction of the Workplace Relations Act(WRA) in 1996 further legitimized this practice. As a result, the proportion ofworkers covered by the traditional award system has fallen. But in May 2006,close to one-fifth of employees were still covered by the awards (AustralianGovernment Department of Employment and Workplace Relations 2007).In addition, the employment and pay conditions provided in the awards arestill bound to all employees, since the awards specify the minimum legalguarantee of employment conditions.

Reflecting the industrial relations reforms, the public sector has also gonethrough significant changes. First, the Public Service Act 1999 provided themost significant and extensive deregulation of public sector employmentgovernance. For instance, it gives agency heads direct power to manage staffusing merit principle to maximize agency performance, shifting the focus toindividual agency for wage determinations (Australian Public Service Com-mission 2003). Second, the public sector size, measured in terms of employ-ment, has reduced significantly since the mid-1980s. The employment shareof the public sector (including commonwealth, state/territory, and local gov-ernments) has dropped from 25.4 per cent in 1985 to about 16 per cent in2005 (Kryger 2006). It is a result of a combination of privatization, outsourc-ing, reduction of permanent employment, increase in part time, causal andcontract employment, and technological changes. Among the OECD coun-tries, Australia and the UK are the only two that have significantly reducedthe share of public sector employment. The public sector size in Australia ismuch smaller than other OECD countries such as Sweden and Denmark(over 30 per cent in 1998) (Jürges 2002) and the UK (about 20 per cent in2005). Nonetheless, the Australian government remains a major employer,with over 1.3 million employees in 2006, approximately 18 per cent of theworkforce (Industry Skills Councils 2006).

The new industrial relations system may lead to differential sectoral effectsacross the conditional wage distribution. The experience of the UK suggeststhat decentralization in wage setting and higher employer’s autonomy inwage determination have contributed to larger public–private wage differ-ences especially in the lower part of the wage distribution (Bender and Elliot1999; Blackaby et al. 1999; Disney and Gosling 2003). Focusing on the publicsector alone, Bender (2003) finds that the pay distribution has narrowed atthe low end but has widened at the upper end after the first round ofenterprise bargaining in Australia. As a result, pay inequality in the public

Public–Private Sector Wage Gap in Australia 3


sector has grown. This would have implications for the wage differentialsbetween public and private sector workers. Therefore, this article contributesto the literature by providing updated estimates on the sectoral wage differ-entials in a changed industrial relations environment.

The rest of the article is arranged as follows. Section 2 describes quantileregression models and the semi-parametric decomposition method. Section 3discusses the data source and model specification. Section 4 presents estima-tion results. Finally, in section 5, we set out our conclusions.

2. Method

Quantile Regression

To investigate whether the public–private pay gap varies at different points ofthe conditional wage distribution, we employ the quantile regression modelsof Koenker and Bassett (1978). Following Buchinsky (1998), we specify theqth (0 < q < 1) conditional quantile of the distribution of the (log) wage w,conditional on a vector of covariates x, as

Q w x xθ β θ( ) = ( ) (1)

Equation (1) assumes a linear relationship between the population condi-tional quantile of w, Qq(w|x) and the covariates x. For a random sample of(wi, xi) for i = 1, . . . , N, equation (1) implies

w x Q xi i i i= ( ) + ( ) =β θ ε εθ θ θ, ,with 0 (2)

where εθi is the error term of the qth conditional (on xi) quantile. In quantileregressions, the only distributional assumption on εθi is that the qth condi-tional (on xi) quantile of the error term equals zero.

For a given q � (0,1), b(q) can be estimated by

ˆ arg min ,β θ β θ ββ

( ) = −( ) − ≤( )( )=∑1

11N

w x w xi i i ii

N

(3)

where 1(·) is the indicator function. b(q) is estimated separately for eachq � (0,1).

Following convention, we first estimate a single-equation quantile regres-sion model of the form similar to equation (2),

w P x Q P xi i i i ii i= ( ) + ( ) + ( ) =α θ β θ ε εθ θ θ, , ,with 0 (2′)

where Pi is a dummy variable equal to one if individual i works in the publicsector and zero otherwise; xi is a vector of other variables that are expectedto affect wages, such as education and experience. The quantile regressioncoefficients can be interpreted as the rates of return to the respective



characteristics at the specific quantile of the conditional wage distribution(Buchinsky 1998; Koenker 2005). Therefore, a(q) measures the public sectorwage premium (or penalty if it is negative) at the qth conditional quantile ofwages, and b(q) measures the effect of other variables at that point of theconditional wage distribution. If the public sector wage premium is the sameacross the conditional wage distribution, we would expect a(q) not to varyfor different qs. If being a public sector employee has no effect on wages, thena(q) should not be significantly different from zero for any q.

The single equation model in equation (2′) assumes that the wage deter-mination process is identical for both public and private sector workers.However, test results shown later suggest that the wage determinants affectpublic and private sector workers differently. To account for the differencesin the returns to wage determining factors between public and privatesector workers, separate wage equations for each group are required. As inthe ordinary least squares (OLS) framework, after estimating the wageequation separately for public and private sector workers using quantileregressions, the differences at various quantiles of the wage distributionsbetween the two groups of workers can be decomposed into the differencedue to observed characteristics and the difference in returns to thecharacteristics.

Decomposition in Quantile Regression

A decomposition method for quantile regression models was initially devel-oped by Machado and Mata (2005). Here we use a modified procedureproposed by Autor et al. (2005) and Melly (2005). In the modified procedure,instead of randomly drawing q and x, we simply estimate quantile regressionsfor a large number of selected qs, such as q1, q2, . . . , qJ, and use the observedsample x to form the required marginal distributions of wages. The followingsteps are involved in decomposing the wage gap between public and privatesector workers at different points of the wage distributions.

Step 1: Estimate bp(tj) and bn(tj), for tj � (0,1) and j = 1, . . . , J, using thepublic sector workers and private sector workers, respectively, to form

xip p

j j

J

i

N p

β τ( ){ }{ }= =1 1and xi

p nj j

J

i

N p

β τ( ){ }{ }= =1 1, where xi

p refers to the observed

characteristics of public sector worker i, xin refers to the observed character-

istics of private sector worker i, Np and Nn refer to the numbers of public and

private sector workers, respectively. xip p

j j

J

i

N p

β τ( ){ }{ }= =1 1provide the predicted

wage density of public sector employees; xip n

j j

J

i

N p

β τ( ){ }{ }= =1 1provide the

counterfactual wage density of public sector workers that would arise if theyretained their own characteristics but were paid as private sector workers.

Step 2: Estimate the qth quantile of the sample xip p

j j

J

i

N p

β τ( ){ }{ }= =1 1, denoted

as Qq(xp, bp(t)), and of the sample xip n

j j

J

i

N p

β τ( ){ }{ }= =1 1, denoted as Qq(xp, bn(t)).



Step 3: Obtain Qq(xp, bp(t)) - Qq(xp, bn(t)). This difference represents thewage gap attributable to the differences in the returns to observed character-istics at the qth quantile, that is the public sector wage effects.1

To estimate the standard errors and confidence intervals of the sectoralwage effects, the bootstrap method can be used to replicate the above pro-cedure. In this study, 500 replications are carried out to estimate the confi-dence intervals, and repeated observations for the same person in differentwaves (i.e. clustering) are taken into account in re-sampling.

3. Data and model specification

Data Source

The empirical analysis uses the first six waves (2001–2006) of the Household,Income and Labour Dynamics in Australia (HILDA) survey. It is a nationalhousehold panel survey. The first wave was conducted in 2001. Then, 7,683households representing 66 per cent of all in-scope households were inter-viewed, generating a sample of 15,127 persons 15 years or older and eligiblefor interview. Of them, 13,969 were successfully interviewed. Interviews forlater waves were conducted about one year apart.

The HILDA survey contains detailed information on individuals’ currentlabour market activity. For those employed, information on job character-istics, such as the size of the workplace and the industry to which theemployee belongs, is also collected. The wages used in this study refer tohourly wages derived from pre-tax total weekly earnings and hours worked inthe main job.2 To avoid the effect of irregular reporting of weekly earningsand hours worked, we excluded those whose hourly wage rate is less than $5.One comparative advantage of HILDA is that the earnings data are notgrouped, thus avoiding possible measurement error due to grouped data.

By defining wages the way it is in this study, we have not examined thedifferentials of total compensations between the two sectors, let alone thenon-financial features of employment arrangements that may compensatethe earnings differentials. For example, employer contributions to pensionschemes for employees may be different between the two sectors. However,no information on employer pension contribution is available in the survey.In the model, we include job characteristic variables, such as part-time andcasual job, occupation and industry, and workplace size, attempting toaccount for the differentials in employment arrangements between the twosectors.

To increase the sample size and thus the accuracy of the estimated distri-bution, we pool the six waves of the HILDA survey currently available.3

Wages are deflated to the first quarter of 2001 using quarterly wage growthrates for males and females separately. Another reason for pooling the datais that sufficiently large sample sizes are important in bootstrapping thestandard errors of the decomposition results.4 Pooling six waves of HILDAraises two econometric issues. One relates to repeated observations, as most



individuals are surveyed more than once. The other is an increase in realwages over time. We include year dummies and use bootstrap methods thataccount for clustering in the empirical work to address these issues.

Our sample includes those wage earners who worked in non-agriculturalindustries. It includes males aged between 25 and 64 (inclusive) years, andfemales aged 25–61 (inclusive) years. Full-time students are excluded. Thereare 28,500 individuals: 14,233 males and 14,267 females. About 25 per cent ofmales and 35 per cent of females in the samples are public sector employees.The summary statistics of the samples are presented in Table A1.

Distribution of Wages

To have a better grasp of the wage distribution across sectors, we estimate thewage density using the kernel estimator and present the results in Figure 1.For both males and females, wages in the public sector appear to have ahigher mean than in the private sector. For males, the dispersion of wages inthe private sector appears to be larger than in the public sector; the oppositeis true for females.

Figure 2 shows the raw wage gap between public and private sectors atdifferent percentiles and at the mean. On average, (log) wages of maleworkers in the public sector are 15 percentage points higher than maleworkers in the private sector; female workers in the public sector have a wage

FIGURE 1Distribution of Wages in the Public and Private Sectors by Sex.

00

.51

1.5

2 4 6 8 2 4 6 8

Male Female

Public Private

Den

sity

(log-) Wage



19 percentage points higher. These mean wage gaps are in line with thosefound in Borland et al. (1998). Clearly, the gap is not uniform across the wagedistribution. For males, the gap decreases from the bottom up to the 15thquantile; it then increases to the 23th quantile and becomes relatively flat upto the 40th quantile. After that, it falls monotonically. The gap for femalesincreases initially, and then falls until the 12th quantile; from the 13th quan-tile, the gap increases up to about the 60th quantile and thereafter fallsmonotonically. For males, the gap is positive from the bottom up to aboutthe 88th quantile of the wage distribution; for females, the gap is positive foralmost the entire wage distribution. The variation of the wage gap along thewage distribution provides a case for using quantile regressions to analyse thepublic–private wage differentials.

Model Specification

We include in the wage equation four education dummies (degree, otherpost-school qualification, year 12, and year 11 and below), work experience(lifetime employment and its square) and a dummy on whether one haslong-term health conditions (representing health capital). In addition tohuman capital, variables on the following characteristics are also included inthe model: demographic characteristics (three dummies for whether one isborn in Australia, an immigrant from an English-speaking or an immigrant

FIGURE 2Raw Wage Gap at Different Quantiles.

-0.2

-0.1

00

.10

.20

.30

.4

(Log

-) w

age

gap

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Quantile

Male (quantile) Female (quantile)

Male (mean) Female (mean)



from a non-English speaking country, a race dummy to identify whether anindividual is an Aborigine or Torres Strait Islander, and a marital statusdummy) and employment characteristics (three dummies to identify casual,part-time or full-time employment), and three occupation dummies forwhite-collar workers (managers and professionals), other white-collarworkers and blue-collar workers. To control for heterogeneity of local labourmarkets and the differential effects of regional living costs on wages, we alsoinclude six state dummies and a dummy indicating capital city residence.There are six dummies to identify workplace size ranging from less than 20 toover 500 employees. The positive relationship between workplace size andwages is well documented (Idson and Feaster 1990; Miller and Mulvey 1996;Morissette 1993). Increasing monitoring costs (which result in higher wagesaccording to efficiency wage theories), greater importance of workplace-specific human capital and teamwork are some explanations discussed in theliterature. A union membership dummy variable is used to capture the unionwage effect. A positive relationship between union membership and wages isoften found in the literature. Finally, year dummies are included to controlfor the trend of increasing real wages over the six waves of the HILDA data.

Summary statistics for the variables used are presented in Table A1. Thesample means reveal very little that is not already well known. For instance,public sector employees enjoy higher wages, larger workplaces have a higherincidence of public sector employees, public sector employees tend to par-ticipate in the workforce longer and be more educated, have white-collartypes of occupation, tend to be union members, are less likely to be migrantsfrom non-English speaking countries, are more likely to be from New SouthWales or the Australian Capital Territory and Victoria, but are less likely tohold casual and part-time jobs. There is some evidence of gender differences.As expected, more females have casual or part-time jobs. This is especiallyapparent among private sector workers. Also, more females are degreeholders and have white-collar jobs than their male counterparts irrespectiveof in which sector they are employed. More female public sector employeesare immigrants from non-English speaking countries.

Econometric Issues

The estimation of a sectoral wage gap typically involves two complicationsresulting from two selection processes. One is the problem of sample selectionarising from the work choice decision; the other is the selection into differentsectors. If these two selection processes are determined by some unobservedfactors that also affect wages, the public–private wage differentials estimatedfrom models that do not account for these possibilities are likely to be biased.

Our approach to accounting for sample selection in quantile regressionsfollows Buchinsky (1998, 2001). That is, we first estimate a single indexselection equation for an individual’s employment status using the semi-parametric procedures as described in Frölich (2006) and Klein and Spady(1993). A power series of (or a polynomial of certain order in) the predicted



index is then included in the wage equation. In our case, we found a poly-nomial of order two was sufficient to account for sample selection. In essence,this selection correction procedure is very similar to the two-step Heckmanprocedure for estimating wage equations using OLS. There are two differ-ences though. First, in our quantile regression framework, the selection equa-tion is estimated using semi-parametric methods that make no assumption onthe distribution of the error term in the selection equation. In the Heckmanprocedure, the selection equation is often estimated as a Probit model, withthe assumption that the error term in the selection equation follows thestandard normal distribution. Second, a polynomial of certain order in thepredicted employment index from the first step is included in the second stepquantile regressions for selection correction. But in the OLS wage models, itis the inverse Mills ratio calculated from the first step that is included in thesecond step wage equation.

For identification purposes, we include the following variables in the selec-tion (i.e. employment status) equation, but not in the wage equation: whetheran individual has a child under five years, whether an individual has a childbetween 5 and 14 years, the total number of dependent children, and whetheran individual is aged 55 or over. Using the presence and total number ofyoung children as instrumental variables in the selection equation is acommon practice in the literature on wage modelling. The instrumentalvariable aged 55 or over may be justified by the fact that in Australia, peopleare allowed to access their superannuation when they reach age 55. Thisprovision itself may provide an incentive for people aged 55 or over to leavethe labour force. Ages also affect wages as is often found in the literature, butin our wage models, age and age square are included to control for the effectof age. When using the variable aged 55 or over as one of the instruments inthe selection equation, we make the assumption that age and its square fullyaccount for the effect of age on wages in the wage equation, and thereforethe aged 55 or over variable has no additional direct effect on wages(Table A2).

However, we have not attempted to account for the potential endogenoussector selection of workers mainly due to lack of valid identifying instru-ments.5 Potentially by studying the public–private pay gap among individualswho we observe in both sectors, we could, to some extent, deal with theselection issue, but the number of these cases is fairly moderate. We couldalso have used the panel nature of the HILDA data to estimate the fixedeffects quantile regression models as proposed by Koenker (2004) to accountfor the endogeneity issue of sector selection. However, unlike in linear modelswhere the unobserved individual fixed effects can be differenced out and arethus omitted from the estimation, the individual fixed effects in a fixed effectsquantile regression model have to be estimated along with the coefficients. Asa result, it is very difficult to implement when the number of parametersinvolved here is so large. Further, only when the endogeneity of sector choiceresults from time-invariant individual heterogeneity, can a fixed effects model(either linear or quantile regression) solve the endogeneity issue. When sector



choice depends on time-variant unobservables, fixed effects models are nothelpful, and instrumental variables are required to deal with the endogeneity.Due to these difficulties, we leave the endogeneity issue of sector choice tofuture research. Accordingly, we interpret the results in this article as adetailed description of the wage gap between sectors rather than a causaleffect of sector on wages.

4. Results

Single-Equation Estimation

Figure 3 presents the coefficient estimates and their 95 per cent confidenceintervals for the public sector dummy variable from both the OLS modeland quantile regressions. The quantile regressions are estimated at each 0.01percentile point. For ease of reading, Table 1 lists the coefficient estimates forthe public sector dummy at selected percentiles and also the estimates fromOLS for males and females.

The OLS estimates show that male workers in the public sector earn a wagethat is 3 per cent lower than their counterparts in the private sector, while forfemale workers in the public sector their wages are about 4 per cent higherthan female workers in the private sector. The OLS results are comparable

FIGURE 3Coefficient Estimates on Public Dummy by Sex.

QR, quantile regression; CI, confidence interval; OLS, ordinary least squares.

-0.4

-0.3

-0.2

-0.1

00.

10.

20.

3

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Male Female

QR lower 95% CI QR gap due to returnQR upper 95% CI OLS lower 95% CIOLS estimate OLS upper 95% CI

(Log

-) w

age

Quantile



with other studies. Take the UK as an example — public servants enjoy awage premium of 5 per cent on average relative to comparable private sectoremployees (e.g. Rees and Shah 1995). It ranges from 2 to 5 per cent for males;but is much higher for females (15–18 per cent).

The story is quite different from the quantile regression estimates. Formales, the sectoral effect exhibits a monotonic decrease across almost theentire conditional wage distribution, ranging from 5 to 3 per cent. For thelowest third of the conditional wage distribution, a positive effect is found,while for the upper 60 per cent of the conditional wage distribution, the effectis found to be negative. The magnitude of the negative effect is fairly large atthe upper end of the conditional wage distribution. For example, in the top20 per cent of the wage distribution, the negative effect is about 10 per centor more. The OLS estimate is close to the estimate at the 60th quantile, butis far from those at the bottom and top ends of the conditional wage distri-bution. For females, the quantile regression estimates are relatively stableand positive over almost the entire wage distribution and in the range of 3–6per cent. The estimates at the very bottom and top ends of the conditionalwage distribution are insignificant. Again, the OLS estimate for femalesprovide misleading inference as to the effect at other parts of the conditionalwage distribution. Using German data, Jürges (2002) also found that incontrast to males, female wage earners in the public sector enjoy a positivewage premium. A negative wage premium is only observed at very highquantiles. Poterba and Rueben (1994) estimate a single log wage equationwith a public sector dummy using quantile regressions for the USA. Theyalso report negative public sector wage premiums at the upper tail of thewage distribution, while a positive premium is evident at the lower end.

Quantile Regression Decomposition

The single-equation estimation results must be interpreted with caution,because they rely on the assumption that the wage determination process is

TABLE 1Estimates of Public–Private Wage Gap in a Single-Equation Model

Males Females

Coefficient s.e. Coefficient s.e.

OLS -0.0321** 0.0133 0.0387*** 0.0095Quantile regression

0.1 0.0529*** 0.0143 0.0333** 0.01370.2 0.0507*** 0.0124 0.0517*** 0.01070.3 0.0293** 0.0117 0.0632*** 0.00950.4 0.0170 0.0126 0.0589*** 0.00820.5 -0.0052 0.0129 0.0553*** 0.00890.6 -0.0368*** 0.0137 0.0615*** 0.00890.7 -0.0777*** 0.0165 0.0578*** 0.00980.8 -0.0985*** 0.0164 0.0461*** 0.01230.9 -0.1322*** 0.0210 0.0199 0.0153

OLS, ordinary least squares. ** denotes significance at 5% level, *** at 1% level.



identical for both public and private sector workers. This assumption may beviolated if being a public sector employee also affects the returns to factorssuch as education. To see whether the model should be estimated separatelyfor each group of workers, we experimented through making interactions ofeach independent variable with the public sector dummy. The F-statistics onthe joint significance of the interaction variables (not shown here) reject thehypothesis that workers in both sectors are subject to the same wage deter-mination process. Therefore, the sectoral wage effects estimated using thesingle-equation model are likely to be misleading; separate wage determina-tion equations for public and private sectors and decomposition methodsare required to provide a more reliable picture of the public–private sectoralwage differentials.

To generate the samples for decomposition purposes, we estimate modelsfor quantiles at [0.001, 0.003 . . . 0.997, 0.999] and at the median. There are501 regressions for each gender and sector group, and thus it is not possibleto report all the estimation results. In the following, we focus on the decom-position results.

Using the procedure described in section 2, we decompose the differencebetween the quantiles of the distribution into the components explained bythe differences in inter-sectoral distribution of observed characteristics (e.g.personal and job characteristics) and by the different returns to the charac-teristics. It is the latter component that can be interpreted as the sectoraleffects, because otherwise, there should not be any difference in the returns.For this reason, the reported results focus on the gap due to returndifferences.

Figure 4 shows the wage gap attributable to return differences at each 0.01percentile point, together with bootstrapped 95 per cent confidence intervals.In bootstrapping the 95 per cent confidence intervals, 100 replications wereused, and the clustering of the observations resulting from the panel data wasalso taken into account. For comparison, the horizontal line in the figureshows the sectoral wage effect estimated using the OLS decompositionmethod. The OLS estimate is computed as x p p nβ β−( ) , using the Blinder–Oaxaca decomposition, where x p refers to the means of the public sectorworker sample; bp and bn refer to the OLS coefficient estimates from publicand private sector workers, respectively. For ease of reading, Table 2 presentsthe results for selected quantiles.

The OLS decomposition shows that for male workers, the contribution ofreturn differences to the wage gap is -0.05, which is significant at the 5 percent significance level, implying that male workers in the public sector earn 5per cent less on average than a comparable worker in the private sector. Thisestimate is larger in size than that found in a single-equation model. Forfemales, the OLS decomposition shows that the sectoral wage effect is about0.03, lower than that found in the single-equation model. For both males andfemales, our OLS decomposition results are different from Borland et al.(1998). They find that the entire observed (mean) wage gap can be attributedto the differences in observed individual and job characteristics. The



differences in the results might not come as a surprise for at least two reasons.First, Borland et al. (1998) use data collected in 1993, about 10 years earlierthan the data used in this study. In 1993, wage setting was very muchcontrolled by the award system, but the data in this study covers the periodwhen enterprise bargaining is widespread. Second, the macroeconomic con-ditions are very different between the two periods. Year 1993 was the timewhen recession hit the bottom, with an unemployment rate of over 12 percent. The period studied here is characterized by a booming economy with anunemployment rate less than half that of the 1993 level. As a result, therelative wage structure between public and private sectors might havechanged, provided that the two sectors have responded to the economicboom differently in terms of how wages are set to attract skilled workers.

The quantile regression decomposition results show that the sectoraleffects are positive for the quantiles from the bottom up to the 50th quantilefor male workers and up to the 82nd quantile for females. The positive effectsare significant at the 5 per cent significance level for the quantiles from thebottom up to the 42nd for males and for the quantiles from the 4th to the 77thfor females. For males, the significant positive effect decreases monotonicallyfrom 16 per cent at the bottom to about 3 at the 42nd quantile; for females,the significant positive effect initially increases from 3 per cent at the 4thquantile to about 7 per cent at the 40th quantile, and falls thereafter to about2 per cent at the 77th quantile. For males, the negative effect becomes

FIGURE 4Wage Gap due to Difference in Returns by Sex.

QR, quantile regression; CI, confidence interval; OLS, ordinary least squares.-0

.6-0

.4-0

.20

0.2

0.4

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Male Female

QR lower 95% CI QR gap due to returnQR upper 95% CI OLS lower 95% CIOLS estimate OLS upper 95% CI

(Log

-) w

age

Quantile



TA

BL

E2

Dec

ompo

siti

onof

the

Pub

lic–P

riva

teSe

ctor

Wag

eG

ap

Mal

esF

emal

es

Raw

gap

Gap

due

todi

ffin

retu

rns

Raw

gap

Gap

due

todi

ffin

retu

rns

Est

imat

es95

%C

IE

stim

ates

95%

CI

OL

S0.

1483

-0.0

493

(-0.

0641

,-0

.034

4)0.

1917

0.02

63(0

.012

2,0.

0404

)Q

uant

ilere

gres

sion

0.1

0.21

250.

0853

(0.0

626,

0.11

01)

0.16

190.

0455

(0.0

267,

0.06

74)

0.2

0.21

460.

0701

(0.0

492,

0.09

58)

0.17

160.

0617

(0.0

447,

0.07

92)

0.3

0.21

960.

0552

(0.0

324,

0.07

96)

0.20

310.

0679

(0.0

510,

0.08

58)

0.4

0.22

140.

0327

(0.0

070,

0.05

55)

0.22

730.

0698

(0.0

536,

0.08

92)

0.5

0.20

300.

0010

(-0.

0244

,0.

0248

)0.

2469

0.06

59(0

.047

7,0.

0847

)0.

60.

1819

-0.0

397

(-0.

0666

,-0

.013

9)0.

2454

0.05

57(0

.035

6,0.

0753

)0.

70.

1380

-0.0

950

(-0.

1250

,-0

.065

9)0.

2367

0.03

84(0

.016

8,0.

0589

)0.

80.

0721

-0.1

744

(-0.

2075

,-0

.140

0)0.

2115

0.01

25(-

0.01

07,

0.03

40)

0.9

-0.0

235

-0.2

922

(-0.

3346

,-0

.243

1)0.

1364

-0.0

483

(-0.

0821

,-0

.012

7)

CI,

confi

denc

ein

terv

al;O

LS,

ordi

nary

leas

tsq

uare

s.



significant from the 57th quantile onwards, and the significant negativeeffect increases from 3 per cent to 47 per cent. For females the negative effectbecomes significant from the 89th quantile onwards; the significant negativeeffect increases from about 4 per cent to about 18 per cent. Similar effectpatterns are found for German workers (Jürges 2002; Melly 2005). Clearly,the estimates from the OLS model cannot reveal the variation of the sectoraleffect across the wage distribution, as found from the quantile regressionmodels. In particular, the opposite sectoral effects at the lower end andthe upper end of the wage distribution cannot be inferred from the OLSmodels.

Table 2 also shows that the part of the wage gap due to differences inobserved individual and job characteristics is substantial. The proportion ofthe observed wage gap attributable to the sectoral effect is relatively small. InTable 2, the quantile where the largest proportion of the gap (40 per cent formales and 36 per cent for females) can be attributed to the sectoral effectis the 10th and the 20th quantiles for males and females, respectively.This suggests that, in line with Melly (2005), public sector employeeshave individual and job characteristics that are more conducive to higherremuneration.

Comparing Figure 4 with Figure 3 and Table 2 with Table 1, we find thatthe patterns of the estimated effects are similar between the single-equationand separate equation models, but the magnitude of the estimated effectsdiffers. The estimates from the latter are generally larger than those in theformer.

Wage Gap by Occupation

In the above OLS and quantile regressions and their associated decomposi-tion, we have controlled for occupation. Nevertheless, it may be argued thattransferability of skills between public and private sectors would be differentbetween occupations. For some occupations, such as accounting and humanresource management, the knowledge and skills required between public andprivate sectors would be very similar, while for some other occupations, thedemand for the skills may be very different between the two sectors. Forexample, the demand for engineers might be very much exclusively comingfrom the private sector. This implies the degree of competition for skillsbetween the two sectors varies depending on occupations, which in turnimplies that the wage gap may also vary accordingly. Particularly, one mightexpect that for those occupations that are subject to strong competitionbetween sectors, there would be a small or no sectoral wage gap, whilewhether the wage gap among occupations that are not exposed to competi-tion should be small or large cannot be predicted a priori.

In this subsection, we divide the samples into two broad occupationgroups, white-collar workers and others, to examine the difference in thesectoral wage gap between occupations. As noted earlier, whiter-collarworkers refers to managers, administrators and professionals. It can be



argued that workers in white-collar occupations would be subject to greaterskill demand competition between the two sectors than workers in non-white-collar occupations. More detailed occupational division is desirable tocapture differences in skill demand competition between the two sectors, butsmall sample sizes would make the estimates unreliable.

Figure 5 shows the raw wage gap by occupation. The differences in rawwage gap between the two groups of occupations are obvious for both malesand females. The raw gap at the mean is larger among workers in non-white-collar occupations than those in white-collar occupations, and this is largelydriven by the fact that at the upper end of the distribution the gap amongnon-white-collar workers is much larger than the gap among white-collarworkers. For male white-collar workers, the positive wage gap at the low halfof the wage distribution is almost completely offset by the negative gap at theupper half of the distribution, leaving the mean raw gap close to zero.

The decomposition results by occupation are presented in Figure 6 andTable 3. For male white-collar workers, the sectoral gap due to return dif-ferences decreases monotonically with quantiles, just like the one estimatedfrom the data that pool occupations (Figure 4), but the rate of decrease of thegap appears to be faster for male whiter-collar workers than for all maleworkers as a whole. For male non-white-collar workers, the sectoral gap dueto return differences is very much flat for most of the distribution, althoughit does fall at the upper end of the distribution. The difference in the sectoralgap due to return differences between the two occupation groups appears to

FIGURE 5Raw Wage Gap at Different Quantiles by Occupation.

-0.2

-0.1

00.

10.

20

.30

.4

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Male Female

Quantile: white collar Quantile: othersMean: white collar Mean: others

(Log

-) w

age

gap

Quantile



be statistically different at the 5 per cent significance level between the 0.5thand 0.92th quantiles. At the mean, male white-collar workers in the publicsector suffer a wage penalty of 6.4 percentage points (Table 3), larger thanthe penalty estimated from all male workers as a whole (Table 2). Malenon-white-collar workers in the public sector have a wage premium of 0.8percentage points at the mean, but the estimate is insignificant at the 5%significant level.

For female white-collar workers, the sectoral gap due to return differencesalso shows a trend of decline when moving up along the distribution, exceptat the two extreme ends, but for female non-white-collar workers, the sectoralgap due to return differences increases for the lower two-thirds of the distri-bution. The difference in the sectoral gap due to return differences betweenthe two occupation groups of females also appears to be statistically signifi-cant between the 0.5th and 0.9th quantiles. At the mean, female non-white-collar workers in the public sector enjoy a much larger wage premium thanfemale white-collar workers (5.7 vs. 1.7 percentage points), and the wagepremium for female non-white-collar workers is significant at the 5 per centlevel.

Do the results support the hypothesis that skill demand competitionbetween the public and private sectors reduces the sectoral wage gap? Forfemales, this seems to be the case, since the gap at the mean is insignificant forwhite-collar workers, but significant for non-white-collar workers, and the

FIGURE 6Wage Gap Due to Difference in Returns by Occupation.

CI, confidence interval.-0

.6-0

.4-0

.20

0.2

0.4

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Male Female

Lower 95% CI: White Gap due to return: WhiteUpper 95% CI: White Lower 95% CI: OthersGap due to return: Others Upper 95% CI: Others

(Log

-) w

age

Quantile



TA

BL

E3

Dec

ompo

siti

onof

the

Pub

lic–P

riva

teSe

ctor

Wag

eG

apby

Occ

upat

ion

Mal

esF

emal

es

Raw

gap

Gap

due

todi

ffin

retu

rns

Raw

gap

Gap

due

todi

ffin

retu

rns

Est

imat

es95

%C

IE

stim

ates

95%

CI

(a)

Whi

te-c

olla

rw

orke

rsO

LS

0.01

34-0

.064

4(-

0.09

51,

-0.0

337)

0.06

480.

0172

(-0.

0069

,0.

0414

)Q

uant

ilere

gres

sion

0.1

0.15

350.

1275

(0.0

775,

0.17

94)

0.12

740.

0651

(0.0

320,

0.10

57)

0.2

0.12

330.

0812

(0.0

380,

0.12

21)

0.11

880.

0513

(0.0

245,

0.08

44)

0.3

0.10

790.

0430

(0.0

013,

0.08

22)

0.10

800.

0424

(0.0

160,

0.07

40)

0.4

0.06

880.

0022

(-0.

0402

,0.

0399

)0.

1005

0.03

31(0

.009

1,0.

0608

)0.

50.

0335

-0.0

419

(-0.

0856

,-0

.001

7)0.

0894

0.02

48(0

.001

1,0.

0492

)0.

6-0

.012

3-0

.095

3(-

0.13

52,

-0.0

475)

0.06

230.

0140

(-0.

0109

,0.

0377

)0.

7-0

.065

5-0

.150

0(-

0.19

16,

-0.1

006)

0.04

570.

0025

(-0.

0315

,0.

0266

)0.

8-0

.118

8-0

.207

2(-

0.25

11,

-0.1

504)

0.01

92-0

.018

7(-

0.05

90,

0.00

81)

0.9

-0.1

9084

-0.2

867

(-0.

3489

,-0

.211

7)-0

.043

2-0

.053

7(-

0.10

93,

-0.0

165)

(b)

Oth

eroc

cupa

tion

sO

LS

0.13

630.

0083

(-0.

0086

,0.

0252

)0.

1038

0.05

67(0

.040

4,0.

0730

)Q

uant

ilere

gres

sion

0.1

0.16

740.

0656

(0.0

365,

0.09

75)

0.09

570.

0295

(0.0

058,

0.05

37)

0.2

0.16

320.

0641

(0.0

390,

0.09

32)

0.09

430.

0545

(0.0

348,

0.07

24)

0.3

0.17

340.

0658

(0.0

394,

0.09

73)

0.10

750.

0712

(0.0

517,

0.08

65)

0.4

0.18

070.

0650

(0.0

353,

0.09

63)

0.10

720.

0797

(0.0

608,

0.09

85)

0.5

0.18

100.

0561

(0.0

244,

0.08

72)

0.12

310.

0876

(0.0

674,

0.10

74)

0.6

0.17

240.

0360

(0.0

010,

0.06

84)

0.12

990.

0947

(0.0

724,

0.11

53)

0.7

0.14

840.

0046

(-0.

0337

,0.

0399

)0.

1385

0.09

77(0

.070

9,0.

1172

)0.

80.

1161

-0.0

420

(-0.

0809

,-0

.007

3)0.

1266

0.09

04(0

.057

2,0.

1140

)0.

90.

0236

-0.1

174

(-0.

1571

,-0

.071

8)0.

1052

0.05

13(0

.013

5,0.

0852

)

CI,

confi

denc

ein

terv

al;O

LS,

ordi

nary

leas

tsq

uare

s.



gap at different quantiles also seems to be smaller for the former than for thelatter (Table 3). However, for males, the opposite appears to be true. Oneexplanation for the lack of evidence may be that the two broad groups ofoccupation classification may not capture the difference in skill demandcompetition between the two sectors, particularly for male workers. Never-theless, the difference in the sectoral wage gap between occupations is aninteresting issue and warrants further research.

5. Conclusion

Using the first six waves of the HILDA survey, this article employs both OLSand quantile regressions and a semi-parametric decomposition method toexamine the sectoral wage gap at the mean and over the entire conditionalwage distribution. Unlike earlier Australian studies, using OLS models, wefound a significant negative sectoral effect for males and a significant positiveeffect for females after controlling for observed individual and job charac-teristics, although the size of the effect is small.

Using quantile regressions, we found a significant wage premium for thepublic sector at the lower part of the conditional wage distribution, and asignificant wage penalty at the upper part of the distribution, particularly formales, a result similar to a number of international studies. The public wagepremium at the lower end of the conditional wage distribution might bedue to the more effective implementation of equal opportunity and anti-discriminatory policies in the public sector, since the government may usepublic sector pay to achieve objectives such as equity and to be a ‘good’employer (Bender and Elliot 1999). A commonly cited reason for public wagepenalty at the upper end of the conditional wage distribution is public oppo-sition to high pay for public servants (Katz and Krueger 1991; Lucifora andMeurs 2006), while the private sector is not subject to such opposition. Thisallows the private sector to use high pay to attract high-skilled workers.Higher private sector remuneration could also be compensating differentialsthat private sector employers use to reduce the turnover rate of high-skilledemployees and/or for less pleasant work environment. For example, somestudies find that overall satisfaction in the public sector is higher than in theprivate sector (Gardner and Oswald 1999; Jürges 2002). Nationwide skillshortages and the booming economy may be further reasons for the muchhigher wages in the private sector than in the public sector at the upper endof the wage distribution. Bargain and Melly (2008) also find a positive effectof the economic upturn on private sector wages in France. They attribute thesectoral wage differential to the sensitivity of private sector wages (and thelack of sensitivity of the public sector) to macro shocks. The recentlybooming Australian economy is largely driven by rapid increases in export ofraw material and commodities produced by the private sector. The boomingof the mining and related industries not only creates high demand for high-



skilled workers, but also generates large revenue for the industries. Thismeans that these private sector employers can afford to pay high wages toattract the required workforce.

It is not clear why female private sector employees are paid more than theirpublic sector counterparts only at the very top end, whereas male privatesector employees are rewarded higher than their public sector counterpartsfor a larger part of the wage distribution. One possible explanation is thatlabour market discrimination against women is more widespread in theprivate sector than in the public sector, affecting most women except for afew at the very top end. Alternatively, the distributional differences of menand women across industries in the private sector may lead to differentpatterns of the sectoral wage effect between men and women if more men arein high-pay industries. Finally, different unobservables between men andwomen could also be a contributory factor. For instance, relative to men,women generally may not be good at bargaining for themselves (Babcockand Lashever 2003). Only the few female executives who have acquiredbargaining skills gain higher pay as the specific salary levels in the privatesector are more likely to be determined by negotiation.

The first explanation can be tested by applying the same decompositionmethod to the data at hand. If the difference in gender wage gap betweensectors is an explanation for the gender difference in the public–privatesectoral wage gap, we would expect the unexplained gender wage gap to behigher in the private than in the public sector at the lowest deciles than at thetop end. However, our decomposition results for the gender wage gap (notshown here), which are consistent with Baron and Cobb-Clark (2008) andKee (2006), show that the unexplained gender wage gap is lower in the privatethan in the public sector at the lowest deciles than at the top. Therefore, thesectoral difference in gender wage gap does not appear to be an explanation.However, the data does not allow us to test the other two propositions. Theexact reasons for the difference in the sectoral wage effects between males andfemales require further investigation.

The decomposition results indicate that differences in observed character-istic explain a substantial proportion of the overall public–private sectorwage gap. The sectoral effect accounts for only a relatively small proportion,with its impact mostly confined to the lowest end of the conditional wagedistribution, that is public sector employees have characteristics that aremore conducive to higher remuneration.

This study has limitations. First, due to the data constraint, the problem ofselection into different sectors could not be dealt with here. If wages andselection into a particular sector are affected by some correlated unobserv-ables, the estimates reported here might be biased. By not allowing endog-enous sector choices, we may underestimate the mean premium, as shown byMelly (2005). Second, the quantile regression results rely on the assumptionthat the covariates, particular sectoral status, are not related to the mean ofthe unobservables. The estimated sectoral wage effects would be biased if thisassumption does not hold. Third, as our data have no information on work



effort (often lower in the public sector) and non-wage benefits (often higherin the public sector), our results are likely to underestimate the true publicsector premium. Fourth, Chatterji et al. (2007) find that workplace charac-teristics such as presence of performance-related pay, company pensionschemes and family-friendly employment practices (e.g. paternity leave andmaternity leave with pay) are important in explaining the public–privatewage gap. Lack of available data meant we could not include more job-related characteristics than those used in the model. Therefore, the wagedifferentials examined here may not reflect the total compensation differen-tials between the two sectors. Finally, we do not distinguish public sectoremployees employed by the federal government from those employed by thestate or local governments. Large wage differences could exist between dif-ferent levels of government employees (Poterba and Rueben 1994).

Final version accepted on 25 October 2009.

Acknowledgements

Lixin Cai thanks the Economics Program at the Research School of SocialSciences of the Australian National University for hosting his visit whenpart of this study was conducted. The authors thank Markus Frölich forsharing his code for estimating binary dependent variable models usingnon-parametric methods. The article uses the data in the confidentializedunit record file from the Department of Families, Housing, CommunityServices and Indigenous Affairs’ (FaHCSIA) HILDA Survey, which ismanaged by the Melbourne Institute of Applied Economic and SocialResearch. The findings and views reported in the article, however, are thoseof the authors and should not be attributed to either FaHCSIA or theMelbourne Institute.

Notes

1. This decomposition uses private sector wages as the norm. Alternatively, publicsector wages and sector-pooled wages can also be used as the norm. The decom-position results using alternative norms show similar patterns and are presented inFigure A1. Since it is more likely that wages in the private sector are determined bymarket forces and thus are subject to less distortion, private sector wages as thenorm is more appropriate.

2. Weekly earnings are not directly asked for in the survey. Instead, they are derivedfrom two questions: the first asks what the amount of the most recent usual payfrom the main job is, and the second asks about the period the pay covers: whetherit is for a week, a fortnight, a month or a year. Hours usually worked in the mainjob, including overtime, are directly asked in the survey.

3. Pooling data across waves may not be appropriate if the public–private gap hasa time trend. Figures A2(a) and A2(b) plot both the raw wage gap and the



decomposed wage gap by wave. From the figures, a time trend does not appear tobe present. Most industrial relations reforms took place in the 1990s. The onlysignificant change to the industrial relations is the introduction of the WorkplaceRelations Amendment Act 2005, known as Work Choices. Since Work Choicesonly came into effect in March 2006, the full impact of it on the labour market hasnot yet been reflected in the six waves of the HILDA.

4. The bootstrapping method is difficult to carry out if the sample size is too small.This is because sampling draws did not always contain observations that had thecharacteristics used in the model if only one wave data were used. For example,since only a few private workers are indigenous in any one wave, a redrawn privateworker sample may not have an indigenous worker. As a result, the original modelthat includes indigenous status as a covariate cannot be estimated using thisredrawn sample. While STATA goes ahead to estimate bs by automatically drop-ping these variables, the number of variables for public and private samples, xp andxn, respectively, will no longer be the same. One could not calculate the counter-factual wages of public sector employees in bootstrapping, since xpbn becomesunconformable. Pooling the six waves of data helps to avoid the problem.

5. Melly (2005) uses the father’s public sector employment as an instrument. God-deeris (1998) uses political activities in college and self-reported political orienta-tion to control sample selection. Unfortunately, HILDA does not collect suchinformation. Occupation of parents when the person was 14 years is available inthe data. We attempted to use parental occupation as instruments, but foundparental occupation was generally insignificant in explaining males’ sector choice,while for females, mother’s, but not father’s, occupation was sometimes signifi-cant. In addition, we are sceptical about the validity of parental occupation asinstruments. Parental occupation is likely to be affected by unobserved ability,which, in turn, is likely to be highly correlated between parents and their children.

References

Australian Government Department of Employment and Workplace Relations(2007). Agreement Making in Australia under the Workplace Relations Act 2004–2006. Canberra: Department of Employment and Workplace.

Australian Public Service Commission (2003). ‘The Australian Experience of PublicSector Reform’. APS Commission Occasional Paper No. 2. Commonwealth ofAustralia. Canberra: Australian Public Service Commission. Available at http://www.apsc.gov.au/about/exppsreform.htm

Autor, D. H., Katz, L. F. and Kearney, M. (2005). ‘Rising Wage Inequality: The Roleof Composition and Prices’. NBER Working Paper No. 11628, National Bureau ofEconomic Research.

Babcock, L. and Lashever, S. (2003). Women Don’t Ask: Negotiation and the GenderDivide. Princeton, NJ: Princeton University Press.

Bargain, O. and Melly, B. (2008). ‘Public Sector Pay Gap in France: New EvidenceUsing Panel Data’. IZA Discussion Paper No. 3427, Institute for the Study ofLabour (IZA).

Baron, J. and Cobb-Clark D. (2008). ‘Occupational Segregation and the GenderWage Gap in Private- and Public-Sector Employment: A Distributional Analysis’.IZA Discussion Paper No. 3562, Institute for the Study of Labour (IZA).



Bender, K. A. (2003). ‘Examining equality between public- and private-sector wagedistributions’. Economic Inquiry, 41: 62–79.

—— and Elliot, R. F. (1999). ‘Relative earnings in the U.K. public sector: the impactof pay reform on pay structure’. In B. Elliott, C. Lucifora and D. Meurs (eds.),Public Sector Pay Determination in the European Union. London: Macmillan, pp.285–328.

Blackaby, D. H., Murphy, P. D. and O’Leary, N. C. (1999). ‘The payment of publicsector workers in the UK: reconciliation with North American findings’. EconomicsLetters, 65: 239–43.

Borland, J. and Gregory, R. (1999). ‘Recent development in public sector labourmarket’. In O. Ashenfelter and R. Layard (eds.), Handbook of Labour Economics.Amsterdam: North Holland, pp. 3573–630.

——, Hirschberg, J. and Lye, J. (1998). ‘Earnings of public sector and private sectoremployees in Australia: is there a difference?’. Economic Record, 74: 36–53.

Buchinsky, M. (1998). ‘Recent advances in quantile regression models: a practicalguideline for empirical research’. Journal of Human Resources, 33: 88–126.

—— (2001). ‘Quantile regression with sample selection: estimating women’s return toeducation in the U.S.’. Empirical Economics, 26: 87–113.

Chatterji, M., Mumford, K. and Smith, P. (2007). ‘The Public-Private Sector GenderWage Differential: Evidence from Matched Employee-Workplace Data’. Discus-sion Paper No. 3158, Institute for the Study of Labour (IZA).

Disney, R. F. and Gosling, A. (2003). ‘A new method for estimating public sector paypremia: Evidence from Britain in the 1990s’. CEPR Discussion Paper 3787.London: Centre for Economic Policy Research.

Frölich, M. (2006). ‘Non-parametric regression for binary dependent variables’.Econometrics Journal, 9: 511–40.

Gardner, J. A. and Oswald, A. J. (1999). ‘The Determinants of Job-Satisfaction inBritian’. Mimeo. Department of Economics, University of Warwick.

Goddeeris, J. H. (1998). ‘Compensating differentials and self-selection: an applicationto lawyers’. Journal of Political Economy, 96: 411–28.

Idson, T. L. and Feaster, D. J. (1990). ‘A selectivity model with employer size wagedifferentials’. Journal of Labour Economics, 8: 99–122.

Industry Skills Councils (2006). ‘Public sector: overview’, in Government Skills Aus-tralia. Canberra: The Industry Skills Council for Government and CommunitySafety. Available at http://www.governmentskills.com.au/content/view/258/538/

Jürges, H. (2002). ‘The distribution of the German public-private wage gap’. Labour,16: 347–81.

Katz, L. and Krueger, A. (1991). ‘Changes in the Structure of Wages in the Publicand Private Sectors’. Discussion Paper No. 1547, Harvard Institute of EconomicResearch: Harvard University.

Kee, H. J. (2006). ‘Glass ceiling or stick floor? Exploring the Australian gender paygap’. Economic Record, 82: 408–27.

Klein, R. and Spady, R. (1993). ‘An efficient semiparametric estimator for discretechoice models’. Econometrica, 61: 387–421.

Koenker, R. (2004). ‘Quantile regression and longitudinal data’. Journal of Multivari-ate Analysis, 91: 74–89.

—— (2005). Quantile Regression. Cambridge: Cambridge University Press.—— and Bassett, G. (1978). ‘Regression quantiles’. Econometrica, 46: 33–50.Kryger, T. (2006). ‘The Incredible Shrinking Public Sector’. Research Note No. 29

2005–06. Parliament Library, Parliament of Australia. Canberra: Commonwealth



of Australia. Available at http://www.aph.gov.au/Library/Pubs/RN/2005-06/06rn29.htm

Lucifora, C. and Meurs, D. (2006). ‘The public sector pay gap in France, GreatBritain and Italy’. Review of Income and Wealth, 52: 43–59.

Machado, J. A. F. and Mata, J. (2005). ‘Counterfactual decomposition of changes inwage distributions using quantile regression’. Journal of Applied Econometrics, 20:445–65.

Melly, B. (2005). ‘Public-private sector wage differentials in Germany: evidence fromquantile regression’. Empirical Economics, 30 (2): 505–20.

—— (2005). ‘Decomposition of differences in distribution using quantile regression’.Labour Economics, 12: 577–90.

Miller, P. and Mulvey, C. (1996). ‘Unions, firm size and wages’. Economic Record, 72:138–53.

Morissette, R. (1993). ‘Canadian jobs and firm size: do smaller firms pay less?’Canadian Journal of Economics, 26: 159–74.

Mueller, R. (1998). ‘Public-private sector wage differentials in Canada: evidence fromquantile regressions’. Economics Letters, 60: 229–35.

Poterba, J. and Rueben, K. (1994). ‘The Distribution of Public Sector Wage Premia:New Evidence Using Quantile Regression Methods’. NBER Working Papers 4734,National Bureau of Economic Research.

Reder, M. (1975). ‘The theory of employment and wages in the public sector’. InD. Hamermesh (ed.), Labour in the Public and Nonprofit Sectors. Princeton, NJ:Princeton University Press, pp. 1–48.

Rees, H. and Shah, A. (1995). ‘Public-private sector wage differential in the U.K.’.The Manchester School of Economic and Social Studies, 63: 52–68.

Vella, F. (1993). ‘Non-wage benefits in a simultaneous model of wages and hours:labour supply functions of young females’. Journal of Labour Economics, 11: 704–23.

Waddoups, C. J. (2005). ‘Trade union decline and union wage effects in Australia’.Industrial Relations, 44: 607–24.



Appendix

FIGURE A1Gap Due to Differences in Returns by Using Different Norms.

-0.4

-0.2

00.

2

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Males Females

Private as norm Public as normPooled as norm

(Log

-) w

age

Quantile



FIGURE A2(a) Raw wage gap at different quantiles by wave.

-0.2

-0.1

00

.10

.20.

30

.4

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Males Females

Wave 1 Wave 2Wave 5 Wave 6

(Log

-) w

age

gap

Quantile

(b) Gap due to differences in returns by wave.

-0.6

-0.4

-0.2

00

.2

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Males Females

(Log

-) w

age

gap

Quantile

Wave 1 Wave 2Wave 5 Wave 6



TA

BL

EA

1Su

mm

ary

Stat

isti

csof

the

Sam

ples

Mal

eF

emal

e

Pub

licP

riva

teA

llP

ublic

Pri

vate

All

Mea

ns.

d.M

ean

s.d.

Mea

ns.

d.M

ean

s.d.

Mea

ns.

d.M

ean

s.d.

Hou

rly

wag

es23

.816

10.6

2821

.576

17.2

3422

.153

15.8

3021

.229

10.1

2117

.831

10.6

4219

.035

10.5

85M

arri

ed0.

820

0.38

40.

742

0.43

80.

762

0.42

60.

718

0.45

00.

714

0.45

20.

715

0.45

1D

egre

e0.

431

0.49

50.

221

0.41

50.

275

0.44

70.

495

0.50

00.

237

0.42

50.

328

0.47

0O

ther

post

-sch

oolq

ualifi

cati

on0.

357

0.47

90.

427

0.49

50.

409

0.49

20.

236

0.42

40.

277

0.44

80.

262

0.44

0Y

ear

120.

088

0.28

40.

118

0.32

30.

111

0.31

40.

098

0.29

70.

165

0.37

10.

141

0.34

8L

ife-

tim

eem

ploy

men

t24

.257

10.1

7021

.812

10.8

7422

.441

10.7

5020

.409

8.93

318

.721

9.55

219

.319

9.37

2L

ife-

tim

eem

ploy

men

tsq

uare

691.

826

507.

308

593.

999

531.

429

619.

176

527.

048

496.

310

393.

958

441.

706

409.

431

461.

064

404.

843

Indi

geno

us0.

015

0.12

30.

009

0.09

30.

010

0.10

10.

023

0.14

80.

009

0.09

50.

014

0.11

7Im

mig

rant

sfr

omE

nglis

h-sp

eaki

ngco

untr

y0.

115

0.31

90.

118

0.32

30.

118

0.32

20.

094

0.29

20.

113

0.31

70.

106

0.30

8Im

mig

rant

sfr

omno

n-E

nglis

hsp

eaki

ngco

untr

y0.

087

0.28

20.

128

0.33

50.

118

0.32

20.

094

0.29

20.

132

0.33

90.

119

0.32

3N

ewSo

uth

Wal

es/A

ustr

alia

nC

apit

alT

erri

tory

0.37

10.

483

0.30

40.

460

0.32

10.

467

0.33

90.

473

0.32

00.

466

0.32

70.

469

Vic

tori

a0.

226

0.41

80.

267

0.44

20.

256

0.43

70.

234

0.42

40.

266

0.44

20.

255

0.43

6Q

ueen

slan

d0.

200

0.40

00.

208

0.40

60.

206

0.40

40.

200

0.40

00.

209

0.40

70.

206

0.40

4So

uth

Aus

tral

ia0.

081

0.27

20.

085

0.27

90.

084

0.27

70.

085

0.27

90.

084

0.27

80.

085

0.27

9W

este

rnA

ustr

alia

/Nor

ther

nT

erri

tory

0.10

00.

300

0.11

10.

314

0.10

80.

310

0.09

90.

299

0.09

30.

290

0.09

50.

293

Tas

man

ia0.

022

0.14

70.

026

0.15

90.

025

0.15

60.

042

0.20

10.

027

0.16

30.

033

0.17

8C

apit

alci

ty0.

595

0.49

10.

650

0.47

70.

636

0.48

10.

586

0.49

30.

646

0.47

80.

625

0.48

4P

art-

tim

e0.

081

0.27

30.

081

0.27

30.

081

0.27

30.

376

0.48

40.

448

0.49

70.

422

0.49

4C

asua

l0.

066

0.24

70.

155

0.36

20.

132

0.33

90.

123

0.32

80.

293

0.45

50.

233

0.42

3P

art

tim

ean

dca

sual

0.04

00.

196

0.05

90.

236

0.05

40.

226

0.10

70.

309

0.23

70.

425

0.19

10.

393

Whi

te-c

olla

rw

orke

rs0.

466

0.49

90.

275

0.44

70.

324

0.46

80.

564

0.49

60.

252

0.43

40.

363

0.48

1O

ther

whi

te-c

olla

rw

orke

rs0.

357

0.47

90.

282

0.45

00.

301

0.45

90.

404

0.49

10.

615

0.48

70.

540

0.49

8B

lue-

colla

rw

orke

rs0.

177

0.38

20.

442

0.49

70.

374

0.48

40.

032

0.17

70.

133

0.34

00.

097

0.29

7H

avin

ghe

alth

cond

itio

ns0.

176

0.38

10.

149

0.35

60.

156

0.36

30.

151

0.35

80.

140

0.34

70.

144

0.35

1F

irm

size

<20

0.18

80.

391

0.36

70.

482

0.32

10.

467

0.17

60.

381

0.45

20.

498

0.35

40.

478

firm

size

20–9

90.

306

0.46

10.

311

0.46

30.

310

0.46

20.

372

0.48

40.

279

0.44

80.

312

0.46

3F

irm

size

100–

199

0.13

90.

346

0.10

90.

312

0.11

70.

321

0.10

10.

301

0.09

90.

299

0.10

00.

300

Fir

msi

ze20

0–49

90.

130

0.33

60.

106

0.30

70.

112

0.31

50.

108

0.31

10.

089

0.28

50.

096

0.29

4F

irm

size

500+

0.23

00.

421

0.10

10.

302

0.13

40.

341

0.22

80.

420

0.07

00.

255

0.12

60.

332

Fir

msi

zeun

know

n0.

007

0.08

60.

006

0.07

90.

007

0.08

10.

014

0.11

70.

011

0.10

40.

012

0.10

9U

nion

wor

kers

0.60

10.

490

0.26

40.

441

0.35

10.

477

0.54

00.

498

0.20

50.

404

0.32

40.

468

Wav

e2

0.17

00.

376

0.16

90.

375

0.17

00.

375

0.16

40.

370

0.16

60.

372

0.16

50.

372

Wav

e3

0.16

80.

374

0.16

60.

372

0.16

60.

372

0.16

90.

374

0.16

20.

369

0.16

50.

371

Wav

e4

0.16

40.

370

0.15

90.

365

0.16

00.

367

0.15

90.

365

0.15

80.

365

0.15

80.

365

Wav

e5

0.16

30.

370

0.16

30.

370

0.16

30.

370

0.16

60.

372

0.16

50.

371

0.16

50.

372

Wav

e6

0.16

00.

367

0.16

60.

372

0.16

50.

371

0.17

00.

375

0.17

10.

376

0.17

00.

376

Sele

ctio

nin

dex

0.36

90.

102

0.37

70.

095

0.37

50.

097

0.30

60.

101

0.33

60.

087

0.32

60.

093

Sele

ctio

nin

dex

squa

re0.

146

0.06

40.

151

0.06

20.

150

0.06

30.

104

0.05

50.

121

0.05

40.

115

0.05

5N

o.of

obse

rvat

ions

3,66

310

,570

14,2

335,

058

9,20

914

,267



TA

BL

EA

2P

robi

tan

dSe

mi-

Par

amet

ric

Est

imat

ions

ofth

eSe

lect

ion

Equ

atio

n

Mal

esF

emal

es

Pro

bit

Sem

i-pa

ram

etri

cP

robi

tS

emi-

para

met

ric

Coe

ffici

ent

s.e.

Coe

ffici

ent

s.e.

Coe

ffici

ent

s.e.

Coe

ffici

ent

s.e.

Age

-0.6

432

0.07

72-0

.791

80.

2827

-0.5

758

0.05

33-0

.813

20.

2646

Age

squa

re-0

.246

10.

0211

-0.2

520

0.08

85-0

.057

70.

0162

-0.0

666

0.03

01M

arri

ed0.

3626

0.03

210.

3632

0.12

040.

0234

0.02

330.

0239

0.03

39D

egre

e1.

0252

0.04

181.

1271

0.38

130.

7464

0.02

820.

9506

0.28

82O

ther

post

-sch

oolq

ualifi

cati

on0.

2911

0.03

130.

2671

0.09

630.

2816

0.02

570.

3727

0.11

88Y

ear

120.

4506

0.04

780.

4383

0.15

650.

2627

0.03

130.

3414

0.11

28L

ife-

tim

eem

ploy

men

t1.

167

0.07

241.

3155

0.45

281.

5187

0.04

252.

0306

0.62

11L

ife-

tim

eem

ploy

men

tsq

uare

0.03

20.

014

0.03

230.

0191

-0.1

808

0.00

99-0

.239

10.

0739

Imm

igra

nts

from

Eng

lish-

spea

king

coun

try

0.00

660.

0412

0.04

320.

0478

-0.1

138

0.03

34-0

.147

70.

0653

Imm

igra

nts

from

non-

Eng

lish

spea

king

coun

try

-0.1

688

0.03

96-0

.136

70.

0623

-0.3

042

0.02

98-0

.376

20.

1185

Vic

tori

a-0

.010

10.

0357

-0.0

496

0.04

290.

0094

0.02

690.

0144

0.03

86Q

ueen

slan

d-0

.081

90.

0374

-0.0

634

0.04

68-0

.091

80.

0284

-0.1

167

0.05

25So

uth

Aus

tral

ia-0

.136

90.

0483

-0.1

836

0.08

12-0

.101

30.

0374

-0.1

383

0.06

69W

este

rnA

ustr

alia

/Nor

ther

nT

erri

tory

0.04

880.

0482

0.05

190.

0577

-0.1

729

0.03

58-0

.232

40.

0868

Tas

man

ia-0

.156

90.

0748

-0.1

526

0.09

400.

1063

0.06

010.

1551

0.09

93C

apit

alci

ty0.

2273

0.02

880.

2668

0.09

460.

0287

0.02

180.

0363

0.03

22H

avin

ghe

alth

cond

itio

ns-0

.884

30.

0277

-0.8

258

0.28

10-0

.676

10.

0243

-0.8

818

0.26

93H

ave

child

ren

unde

rfiv

eye

ars

–0.0

408

0.05

28–0

.043

30.

0642

–0.9

941

0.03

24–1

.296

80.

3991

Hav

ech

ildre

n5–

14ye

ars

0.09

60.

0536

0.10

190.

0712

–0.0

103

0.03

3–0

.028

00.

0484

Num

ber

ofch

ildre

n-0

.086

80.

0246

-0.1

024

0.04

38-0

.135

20.

0159

-0.1

712

0.05

71A

ged

55or

over

-0.3

283

0.06

58-0

.294

80.

1232

-0.2

795

0.05

39-0

.369

70.

1378

Wav

e2

0.06

620.

0422

0.06

430.

0507

-0.0

105

0.03

320.

0093

0.04

77W

ave

30.

2091

0.04

310.

2247

0.08

870.

0582

0.03

350.

0865

0.05

41W

ave

40.

2644

0.04

430.

2954

0.11

030.

065

0.03

40.

0894

0.05

49W

ave

50.

3516

0.04

490.

3710

0.13

320.

1563

0.03

420.

2124

0.07

87W

ave

60.

3657

0.04

510.

3984

0.14

120.

1976

0.03

430.

2616

0.09

07C

onst

ant

-0.1

990.

0688

-0.0

010.

0522

Non

-lab

our

inco

me

-1-1

-1-1

Not

e:F

orid

enti

ficat

ion,

the

cons

tant

term

cann

otbe

incl

uded

inth

ese

mi-

para

met

ric

esti

mat

ion,

and

the

coef

ficie

nton

one

ofth

eco

ntin

uous

vari

able

sha

sto

beno

rmal

ized

tobe

one.

Her

e,w

eno

rmal

ized

the

coef

ficie

nton

non-

labo

urin

com

eto

be-1

.F

orco

mpa

riso

n,w

eal

sono

rmal

ized

the

coef

ficie

ntof

non-

labo

urin

com

eto

be-1

inth

epr

obit

mod

el.



public–private sector wage gap in australia: variation

Documents