The Australian Gender Wage Gap BY MARY STEPHAN A THESIS SUBMITTED FOR THE DEGREE OF DOCTOR OF PHILOSOPHY OF THE AUSTRALIAN NATIONAL UNIVERSITY JANUARY 2017 COPYRIGHT BY MARY STEPHAN 2017 ALL RIGHTS RESERVED i
DECLARATION I, Mary Stephan declare that this thesis, submitted in fulfilment of the requirements for the award of Doctor of Philosophy, Crawford School of Public Policy, College of Asia and the Pacific, The Australian National University, is wholly my own work unless otherwise referenced or acknowledged. This thesis has not been submitted for qualifications at any other academic institution. MARY STEPHAN January 2017 ii
ACKNOWLEDGMENT There are a few people that need to be recognised for being with me on my intensive journey of undertaking this thesis. My greatest debt is to my supervisor, Professor Bruce Chapman. I thank him for his foresight and vision prior to the commencement of my thesis. I am grateful for his careful guidance throughout my candidature, especially during the early stages. I appreciate the quality checks of my thesis and research. I thank Bruce for his professionalism, support, encouragement, and macro-management during my entire candidature. When situations became tough, Bruce allowed me to find the right tools that helped me clear my own path. This thesis would not have been started if it wasn t for him. Many thanks go to Professor Alison Preston for reviewing my thesis and providing constructive feedback and comments. I thank the examiners for the time and effort taken to examine this thesis. I appreciate the very helpful and constructive feedback that they provided. I thank my mother Samira Hormis for teaching me the value of education and showing me that the best things in life are worth working hard for. I am grateful to my husband Hanna Stephan for his support throughout the research and writing period. I thank him for his patience while I dedicated all of my spare time to the research and the writing of my thesis. I thank my mother in-law Armenouhi Stephan for her domestic support during the majority of my research and thesis writing. iii
ABSTRACT In the Australian labour market, men earn higher wages than women and this difference is persistent. The purpose of this thesis is to examine the gender wage gap and its determinants over time, to measure the gap using different methods compared to prior Australian research, and to measure the gap by sector of employment. The gender wage gap is estimated using the real 1 hourly wages of men and women in the Australian labour market. The thesis begins with a review of Australian and international studies that have measured the gender wage gap and of possible explanations for its existence. The literature review provides a survey of the empirical methods that have been used to measure the gender wage gap and of empirical issues that arise when measuring the gap. Further, the literature review provides a discussion of labour supply and demand side factors that could influence the gap. The analysis in this thesis begins by measuring and comparing the male and female mean and distributional labour market characteristics and the returns to characteristics in 2001 and 2012. The results show that women s labour market characteristics have improved compared to men s characteristics; however, women s relative returns to characteristics have not increased. Next, the gender wage gap is estimated at the mean and along the wage distribution, decomposed, and compared over time. The results show that the mean and distributional gender wage gap has increased over time and the gap is increasingly unexplained by differences in labour market characteristics. The comparison over time extends prior static estimates of the gender wage gap. The gender wage gap is then estimated using panel data and fixed effects methods to account for time-invariant unobserved individual heterogeneity. This is the first time that the Australian gender wage gap has been measured by taking into account individual fixed effects. The fixed effects results are compared with the results of traditional estimation methods and prior research. 1 In 2012 dollars. iv
The comparison shows that time-invariant unobserved individual heterogeneity constitutes 62 per cent of the gender wage gap and that the Australian gender wage gap has previously been overestimated by approximately 7 percentage points. The inclusion of time-invariant unobserved individual heterogeneity in the estimation provides a clearer measure of gender based earnings differentials. The gender wage gap is measured by sector of employment the Australian private and public sector using panel data and time-invariant individual fixed effects methods. The results are compared with findings from traditional methods and prior research. This is the first time that the Australian sector-specific gender wage gap has been estimated by incorporating time-invariant unobserved individual heterogeneity. The results show that traditional estimation methods overestimate the gender wage gap by 6 percentage points in the private sector and 4 percentage points in the public sector. Time-invariant unobserved individual heterogeneity explains 42 per cent of the gender wage gap in the private sector and 41 per cent of the gender wage gap in the public sector. Further, the distributional sector-specific gender wage gap is larger and increases faster in the private sector compared to the public sector. The sector-specific gender wage gap is increasingly unexplained by differences in labour market characteristics along the wage distribution. To the knowledge of the author, this is the first time that very detailed estimates of the Australian gender wage gap have been compared over time. This is also the first time that the Australian gender wage gap has been measured by taking into account timeinvariant individual fixed effects. The results highlight that the gender wage gap has increased over time and that unobserved individual heterogeneity need to be incorporated in the estimation of the Australian gender wage gap as traditional estimation methods tend to overestimate the gap. v
LIST OF ABBREVIATIONS AND ACRONYMS ABS Australian Bureau of Statistics ANZSCO Australian and New Zealand Standard Classification of Occupations AWOTE Australian Weekly Ordinary Time Earnings FE Fixed Effects HILDA Household Income and Labour Dynamics in Australia OLS Ordinary Least Squares UK United Kingdom USA United State of America vi
Contents DECLARATION... ii ACKNOWLEDGMENT... iii ABSTRACT... iv Chapter 1 Introduction... 1 1.1 Scope... 1 1.2 Structure... 3 Chapter 2 Literature review... 7 2.1 Empirical literature... 9 2.1.1 2.1.2 2.1.3 2.1.4 Distributional gender wage gap... 11 Endogeneity... 11 Sample selection... 13 Fixed effects... 17 2.2 Labour market supply... 23 2.2.1 2.2.2 2.2.3 2.2.4 Human capital model... 24 Labour market interruptions... 25 Occupational segregation... 26 Gender segregation in the field of study (education)... 28 2.3 Labour market demand... 29 2.3.1 2.3.2 2.3.3 Organisational networking and group interactions... 30 Crowding and competition... 32 Institutionalisation... 33 2.4 Conclusion... 35 Chapter 3 Methods... 37 3.1 Pooled regression with gender dummy... 37 3.2 Segregation by gender... 37 vii
3.3 Quantile regression... 39 3.4 Sample selection correction... 41 3.5 Panel estimation... 43 3.6 Gender wage gap... 46 3.7 Counterfactual wage decompositions... 46 3.7.1 3.7.2 3.7.3 Mean decomposition... 46 Distributional decomposition... 49 Wellington decomposition... 51 3.8 Appendix A: Decomposition methods... 53 3.8.1 Oaxaca blinder decomposition... 53 3.9 Biewen decomposition... 53 Chapter 4 Data... 57 4.1 Attrition... 57 4.2 The variables... 58 4.2.1 4.2.2 Industry and occupation... 60 Sample selection variables... 62 4.3 Dataset A... 64 4.4 Dataset B... 71 4.5 Dataset C... 74 Chapter 5 The Australian gender wage gap: what has changed over time?... 77 5.1 Introduction... 77 5.2 Empirical framework... 78 5.3 Results... 80 5.3.1 5.3.2 5.3.3 5.3.4 Raw gender wage gap... 80 Pooled regression with gender dummy... 80 Stratification by gender... 82 Sample selection correction... 89 viii
5.3.5 5.3.6 5.3.7 Returns to labour market characteristics... 90 Counterfactual decomposition... 100 Wellington decomposition... 105 5.4 Conclusion... 110 5.5 Appendix B... 113 Chapter 6 The Australian gender wage gap: how much can be explained with fixed effects? 117 6.1 Introduction... 117 6.2 Empirical framework... 120 6.2.1 6.2.2 Heterogeneity... 120 Benefits of fixed effects models... 121 6.3 Results... 124 6.3.1 6.3.2 6.3.3 6.3.4 6.3.5 6.3.6 Diagnostic... 124 Raw gender wage gap... 125 Sample selection bias... 126 Gender segregated estimates... 127 OLS and FE gender wage gap... 135 Gender wage gap decomposition... 138 6.4 Discussion... 141 6.5 Conclusion... 142 6.6 Appendix C... 145 Chapter 7 The gender wage gap in the private and public sector: a fixed effects estimation 150 7.1 Introduction... 150 7.2 Empirical Literature... 153 7.3 Results... 157 7.3.1 Diagnostics... 157 ix
7.3.2 7.3.3 7.3.4 7.3.5 7.3.6 7.3.7 7.3.8 Raw gender wage gap... 158 Sample selection bias... 160 Gender segregated estimates OLS and FE models... 162 Gender wage gap OLS and FE models... 166 Gender wage gap decomposition OLS and FE... 169 Quantile regression... 173 Quantile gender wage gap decomposition... 175 7.4 Discussion... 178 7.5 Conclusion... 180 7.6 Appendix D... 183 7.6.1 Appendix D, Section 1... 186 Chapter 8 Conclusion... 190 8.1 Main findings... 190 8.2 Future research... 195 References... 197 LIST OF TABLES Table 4.1 Variables and definitions... 61 Table 4.2 Variables in the sample selection equation... 63 Table 4.3 Mean descriptive statistics (Dataset A 2001 and 2012)... 67 Table 4.4 Distributional descriptive statistics (Dataset A 2001 and 2012)... 68 Table 4.5 Mean descriptive statistics (Dataset B 2001 to 2012 panel)... 73 Table 4.6 Mean descriptive statistics (Dataset C 2001 to 2012 panel)... 76 Table 5.1 Raw mean gender wage gaps (2001 and 2012)... 80 Table 5.2 Mean returns to men s and women s labour market characteristics (2001 and 2012)... 82 x
Table 5.3 Distributional returns to men s and women s labour market characteristics (2001 and 2012)... 83 Table 5.4 OLS wage equation with sample selection correction... 90 Table 5.5 Mean gender wage gap decomposition (2001 and 2012)... 102 Table 5.6 Quantile gender wage gap decomposition (2001 and 2012)... 105 Table 5.7 Mean Wellington decomposition (change between 2001 and 2012)... 106 Table 5.8 Dynamic mean decomposition... 108 Table 5.9 Distributional Wellington decomposition (change between 2001 and 2012)... 109 Table 5.10 Dynamic distributional decomposition... 110 Table 6.1 Raw gender wage gap... 125 Table 6.2 OLS wage equation with selectivity correction... 127 Table 6.3 Pooled OLS and FE estimates returns to labour market characteristics... 132 Table 6.4 Pooled OLS and FE estimates to labour market characteristics, industry and occupation... 134 Table 6.5 Gender wage gaps OLS and FE... 138 Table 6.6 Counterfactual decomposition OLS and FE gender wage gaps... 141 Table 7.1 Raw gender wage gap (2001 to 2012 panel)... 158 Table 7.2 OLS wage equation with sample selection correction public and private sector... 161 Table 7.3 OLS and FE estimates returns to labour market characteristics... 165 Table 7.4 Gender wage gaps by sector OLS and FE... 169 Table 7.5 Gender wage gap decomposition by sector OLS and FE... 172 Table 7.6 Quantile gender wage gap by sector... 175 Table 7.7 Quantile gender wage gap decomposition by sector... 178 xi
LIST OF FIGURES Figure 5.1 Mean and distributional gender wage gap (2001 and 2012)... 81 Figure 5.2 Mean earnings functions (2001)... 87 Figure 5.3 Mean earnings functions (2012)... 88 Figure 5.4 Quantile gender wage gaps (2001 and 2012)... 103 Figure 6.1 Average real hourly wages by age and gender... 126 Figure 7.1 Average real hourly wages by age and gender... 160 Figure 7.2 Quantile gender wage gap by sector... 175 xii
Chapter 1 Introduction 1.1 Scope The gender wage gap is evident in the labour market when men and women are not paid equally and refers to the situation where women s earnings are less than men s. Prior to the introduction of the 1969 Equal Pay Case in Australia, tribunals enforced discrimination against women in pay. Following the 1969 Equal Pay Case, Australian women s relative earnings improved, leading to a narrowing of the gender wage gap. Currently, the Australian gender wage gap of 18 per cent is higher than the Organisation for Economic Co-operation and Development s average of 15.5 per cent (OECD 2015). Australian and international economists have spent decades measuring and explaining the gender wage gap (Weichselbaumer & Winter-Ebmer 2007), while policy-makers have attempted to identify methods of closing the gap (Cobb-Clark 2012). The gender wage gap has been reported in a number of countries (Dixon 2000; Siphambe & Thokweng Bakwena 2001; Kidd & Shannon 2002; Rubery, Grimshaw & Figueiredo 2005; Daly 1990; Gregory & Borland 1999) and empirical studies have found that it is persistent (Chang & Miller 1996; Preston 2003). In measuring the gender wage gap, economists have identified that the gap varies along the wage distribution (Miller 2005; Kee 2006; Baron & Cobb-Clark 2010), is larger for the self-employed (Eastough & Miller 2004), and is different between the labour market sectors of employment (Kee 2006; Baron & Cobb-Clark 2010). As there are economically valid reasons for a difference in earnings between workers or groups of workers, the presence of a gender wage gap does not imply the presence of discrimination. However, a wage gap is typically attributed to discrimination when identically endowed employees with the same productive skills are treated differently based on characteristics that are not related to their productivity such as gender, ethnicity, and marital status. After statistically accounting for the productivity related labour market characteristics of men and women at the mean and along the wage distribution, the remaining earnings difference is claimed to be the gender wage gap, and therefore, gender based discrimination of earnings. These statistical residuals are often decomposed at the mean 1
(Blinder 1973; Oaxaca 1973) and along the wage distribution (DiNardo, Fortin & Lemieux 1995; Machado & Mata 2005) into two components; one that can be explained by gender differences in endowments and one component that is attributable to difference in coefficients. The identified explanations for the gender wage gap can be categorised into three main groups. The first is that the gender wage gap is a result of the differences between the human capital characteristic of men and women such as education, tenure, and work experience. Second, the gender wage gap is attributed to gender differences in job characteristics such as the industry, occupation, and sector of employment. Third, the gender wage gap is claimed to be the result of a difference in family responsibilities and commitment between men and women such as, being the primary caregiver for dependent children. Australian studies have measured the mean and distributional gender wage gap at particular points in time and compared the mean gender wage gap over time. However, detailed distributional estimates of the gender wage gap and the explanations for the gap have not been compared over time. A comparison over time will show whether the gender wage gap has narrowed and whether the explanations for the gap have changed. Further, a comparison over time will show whether returns to individuals earnings are in line with the changes in individuals labour market characteristics. While measuring the mean and distributional gender wage gap, Australian studies have only accounted for the observed labour market characteristics of men and women. A small number of studies have also accounted for firm-specific effects. This is limited as time-invariant unobserved individual heterogeneity can influence earnings and the gender wage gap. Estimation of the gender wage gap while accounting for the observed labour market characteristics and the time-invariant unobserved individual heterogeneity will provide a clearer measure of earnings differentials. Lastly, despite the institutional differences between the Australian public and private sector, little attention has been given to differences in the gender wage gap between these sectors of employment. More importantly, the impact of time-invariant unobserved individual heterogeneity on the gender earnings differentials in each sector of the Australian labour market has never been measured. Measurement of the gender 2
wage gap by sector of employment will show whether the gender wage gap is different for the employees of private companies and government organisations. Estimation of these gaps while accounting for time-invariant unobserved individual heterogeneity will provide greater insight into the extent of anti-discriminatory practices in each sector. 1.2 Structure This thesis consists of eight chapters including the introduction and conclusion. In Chapter 2, the Australian and international literature on the gender wage gap and the explanations for the gap will be reviewed. The literature review will present studies that have been undertaken to measure the gender wage gap as well as studies relating to labour supply and demand that can be used to explain the gender wage gap. Prior Australian studies have measured the mean and distributional gender wage gap using cross-sectional and panel data. Despite extensive gender wage gap estimates, little attention has been given to a detailed comparison of the gender wage gap over time. A comparison over time extends prior static estimates by showing whether the gender wage gap and its explanations have changed. Chapter 3 provides a description of the methods that will be used to measure and decompose the gender wage gap within each analytical chapter. Chapter 4 describes the Household Income Labour Dynamics in Australia (HILDA) Survey, which will be used in this thesis to measure and decompose the gender wage gap. Further, Chapter 4 provides the variables chosen and the descriptive statistics of the variables for each analytical chapter. Chapter 5 presents results on the comparison of the gender wage gap and the explanations for the gap between 2001 and 2012. Using the real 2 hourly wages of men and women in 2001 and 2012, the analysis incorporates the measurement of men s and women s labour market characteristics and the estimation of returns to their characteristics at the mean and along the wage distribution. The results in Chapter 5 show that despite improvements in women s labour market characteristics over time, women s relative returns to labour market characteristics have 2 In 2012 dollars. 3
not increased. A comparison of the mean and distributional gender wage gaps in 2001 and 2012 shows that the gap has increased over time both at the mean and along the wage distribution. The 2001 and 2012 mean and distributional gender wage gap is decomposed. The decomposition indicates whether the gap is due to differences in labour market characteristics between men and women (explained component), returns to labour market characteristics (unexplained component), and/or an interaction between these components. The results show that the gender wage gap at the mean and along the wage distribution is mainly attributed to the unexplained component of the decomposition and this effect increases with time and earnings. Further, the change in the gender wage gap over time is decomposed. This determines whether the increase of the mean and distributional gap is influenced by changes in labour market characteristics or returns to labour market characteristics. The results show that an increase in the disparity between the returns that men and women receive for their labour market characteristics (coefficients) over time has driven the increase in the mean and distributional gender wage gap between 2001 and 2012. In Chapter 6, the gender wage gap will be measured using a panel dataset and fixed effects methods to account for time-invariant unobserved individual heterogeneity. The incorporation of time-invariant unobserved individual heterogeneity in the estimation extends previous studies undertaken to estimate the Australian gender wage gap. To show the benefit of accounting for time-invariant unobserved individual heterogeneity, the fixed effects results are compared to estimates of the gender wage gap using the traditional ordinary least squares model and of prior research. The ordinary least squares estimates in Chapter 6 present a gender wage gap of 11 per cent, which is consistent with prior research. The fixed effects estimation presents a gender wage gap estimate of 4 per cent. The results show that time-invariant unobserved individual heterogeneity accounts for 62 per cent of the gender wage gap. These results also show that the ordinary least squares model tends to overestimate the gender wage gap by approximately 7 percentage points. As such, the results highlight the importance of incorporating time-invariant unobserved individual heterogeneity in the estimation of the gender wage gap. 4
The gender wage gap estimates from the traditional model and the fixed effects model are decomposed into three components. That is, components that can be attributed to differences in labour market characteristics (explained component), returns to labour market characteristics (unexplained component), and the simultaneous impact of endowments and coefficients (interaction component). After accounting for the timeinvariant unobserved individual heterogeneity, the explained component of the decomposition increases while the unexplained component declines. These results imply that unobserved heterogeneity contributes to the explanations of the gender wage gap. Further analysis is undertaken in Chapter 6 to incorporate industry and occupational controls in the estimation of the gender wage gap. The analysis shows that these controls do not lead to an increase or a decline of the Australian gender wage gap. Chapter 7 of this thesis presents estimates and comparisons of the gender wage gap in the private and public sector of the Australian labour market. The sector-specific gender wage gap is estimated using real hourly wages (in 2012 dollars) from a panel dataset and fixed effects models. The results are compared with estimates from traditional methods and prior research. The analysis and results presented in Chapter 7 extend prior Australian research by incorporating time-invariant unobserved individual heterogeneity in the estimation of the sector-specific gender wage gap. Similarly to prior Australian estimates of the sector-specific gender wage gap, the results show that the gap is larger in the private sector compared to the public sector. The contributing results from Chapter 7 show that the sector-specific gender wage gap in the Australian labour market has previously been overestimated. The overestimated gender wage gaps are evident following the comparison of the sector-specific ordinary least squares and fixed effects estimates. This comparison shows that time-invariant unobserved individual heterogeneity constitutes 42 per cent of the private sector gender wage gap and 41 per cent of the public sector gender wage gap. As such, the traditional ordinary least squares model leads to an overestimation of the gender wage gap by 6 percentage points in the private sector and 4 percentage points in the public sector. 5
The fixed effects and ordinary least squares sector-specific gender wage gap estimates are decomposed into explained, unexplained, and interaction components. The results show that the inclusion of time-invariant unobserved individual heterogeneity in the estimation of the gender wage gap leads to a substantial reduction of the unexplained component of the decomposition in the private sector but not in the public sector. The estimation of the sector-specific distributional gender wage gap using panel data extends the work of Kee (2006) by using a panel dataset and the work of Baron and Cobb-Clark (2010) by using a longer time series in the panel dataset. In line with prior research, the distributional Australian gender wage gap is larger in the private sector compared to the public sector from the 25 th wage quantile of the earnings distribution. The decomposition of the sector-specific distributional gender wage gap shows that the gap in both sectors cannot be attributed to differences in labour market characteristics, particularly at the top of the wage distribution. Chapter 8 provides a summary of the findings and outlines the labour market and methodology insights gained from the results. Chapter 8 ends by explaining the limitations of this study and providing suggestions for future research that can further contribute to the literature on the labour market and the gender wage gap. 6
Chapter 2 Literature review The 1969 Equal Pay Case introduced equal pay for equal work. This led to an immediate improvement in women s relative earnings (Cobb-Clark 2012). However, the gender wage gap continues to prevail in the Australian labour market. Women s relative wages are shown to be lower amongst the self-employed (Eastough & Miller 2004) and lower at the top of the wage distribution relative to men (Miller 2005; Kee 2006). It has also been shown that wages differ substantially across labour market sectors (Kee 2006; Baron & Cobb-Clark 2010). This persistent gender wage gap is of interest to economic researchers, policy makers, and workers. In particular, policy makers attempt to identify policy initiatives that might successfully close the gap. Australian and international studies have used three types of explanations for the gender wage gap. These explanations are differences between men and women in: human capital characteristics such as education, work history, and work experience; job characteristics such as hours worked, industry, and occupation; and family responsibilities such as being the primary caregiver for dependent children, and being a sole parent. Joshi, Paci and Waldfogel (1999) found a 40.1 per cent gender wage gap in the United Kingdom (UK). The authors revealed that gender differences in human capital, job characteristics, and parenthood accounted for 23.9 percentage points of the gender wage gap. O Neill (2003) reported a 27.9 per cent gender wage gap in the United States of America (USA). The author decomposed the gender wage gap and found that of this gap, 24.7 percentage points were attributed to gender differences in education, work experience, occupation, job characteristics, and parenting related factors. Using 1998 USA data from the Panel Study of Income Dynamics, Blau and Kahn (2007) found a 20.3 per cent gender wage gap. After accounting for education levels, work experience, occupation, industry, union status, and race, the gap reduced to 8.3 per cent. Drolet (2002) used Canadian panel data and found a 19.7 per cent gender wage gap. Around 33 to 50 per cent of this gap was explained by gender difference in education, experience, tenure, family responsibility, industry, occupation, and other variables. Dynamic gender wage gap studies have shown a decrease in the gap over the last 20 years. Blau and Kahn (2006) reported a rapidly declining gap in the 1980s, yet slower 7
declines in the 1990s for the USA. The reduction in the gender wage gap was attributed to an increase in women s labour market participation; changes in the male and female occupational structures; deunionisation; and a decrease in the unexplained component of the gender wage gap. Forecasts of the gender wage gaps for the USA and Australia show that further increases in women s educational achievements will contribute to a reduction of the gender wage gap (Shannon & Kidd 2003). However, a small gender wage gap is still expected to remain in 2040 in the USA and 2031 in Australia. It is argued that labour market participation is an important determinant of earnings (Cassells et al. 2009; Miller 2005). Female labour market participation in Australia has followed an increasing trend over the last few decades. Since 1978, Australian labour market participation has increased by an annual average rate of 2 per cent (ABS 2015). In particular, women s labour market participation increased by an annual average of 3 per cent while men s increased by an annual average of 1 per cent. On average, most women (59 per cent) in the Australian labour market work at a fulltime basis. However, most of the increase in women s labour market participation has been due to the rise in part-time employment, which increased by an annual average of 3 per cent between 1974 and 2014. It can be argued that, relative to men, an increase in women s labour market participation should in part contribute to the narrowing of the gender wage gap (Preston & Jefferson 2007). However, this implies that the gender wage gap was measured using total earnings, rather than hourly earnings. Female labour market participation is driven by a number of factors. Some of these factors are policy based while others are family based. For example, policies such as paid or unpaid maternity leave (Baxter 2008), family-friendly workplaces (Productivity Commission 2009), and formal child care provisions can encourage or discourage women to participate in the labour market. Family based factors are another driver of women s labour market participation and include financial constraints, informal child care provisions, and social gender perceptions towards the roles of men and women in the labour market and the household. In Australia, Gilfillan and Andrews (2010) showed that female labour market participation has increased over the last 30 years. Gordon (2012) revealed an increase in the proportion of women undertaking higher education levels, in particular bachelor 8
degrees or above. Given the increasing trends in female labour market characteristics, an explanation for the persistent Australian gender wage gap is yet to be found. 2.1 Empirical literature Despite the different data sources and estimation methods used to measure the gender wage gap, Australian empirical studies have found evidence of a persistent average gender wage gap (Haig 1982; Jones 1983; Chapman & Mulvey 1986; Langford 1995; Chang & Miller 1996; Preston 2001; Wooden 1999). This gap has been analysed in further detail through the investigation of the differential between part-time and fulltime workers (Preston 2003), gender occupational segregation (Miller 1994; Wooden 1999), and the contribution of firm-specific effects on the gender wage gap (Meng 2004). The literature on the gender wage gap at the mean is well established (Altonji and Blank, 1999, Weichselbaumer and Winter-Ebmer, 2007, Weichselbaumer and Winter- Ebmer, 2005). Meta-analysis studies on the gap have found evidence of a raw gender wage gap of 32 per cent to 35 per cent (Jarrell and Stanley, 2004, Stanley and Jarrell, 1998). Using the 1982 Australian Bureau of Statistics (ABS) Special Supplementary Survey No. 4 and estimating the log wages of men and women separately, Chapman and Mulvey (1986) found a 15 per cent mean gender wage gap in the Australian labour market. The authors decomposed the gender wage gap and found that if women had the same labour market characteristics as men, their hourly earnings would increase by 2 per cent and the gender wage gap would decline to 13 per cent. If women retained their labour market characteristics and were paid like men, their earnings would increase by 14 per cent and the gender wage gap would decline to 4 per cent. Using the 1991 Census Household Sample File, Chang and Miller (1996) estimated a mean gender wage gap of 11 per cent by including a gender dummy variable in a log wage equation. The authors concluded that the gender wage gap may contribute to the explanation of the inter-industry wage structure (Chang & Miller 1996, p.46). Wooden (1999) used the 1993 Survey of Training and Education and found an average gender wage gap of 11.5 per cent. 9
Estimates of the change in the Australian gender wage gap over time present different results depending on the time periods of comparison. Preston (1997) measured and compared the gender wage gap using 1981 and 1991 Census data. The author found that the Australian gender wage gap declined from 29 per cent in 1981 to 19.8 per cent in 1991. An Oaxaca-Blinder decomposition revealed that despite the decline in the gender wage gap over the decade, the gap could be explained less by gender differences in characteristics in 1991 compared to 1981. The change in the gender wage gap of five Australian States was measured by Preston and Crockett (1999) using the 1991 and 1996 Census Household Sample Files and decomposed using the Wellington (1993) method. The authors found that the gender wage gap in NSW, Queensland, and Adelaide declined between the years analysed. This was due to a change in the returns to female labour market characteristics relative to men in those states. By contrast, the gender wage gap in Victoria and Perth increased between 1991 and 1996 despite an improvement in women s labour market characteristics, which was not enough to offset the favourable change towards men s earnings. Preston (2003) measured the change in the Australian gender wage gap for full-time and part-time employees using the 1989/90 Income Distribution Survey and the 1997/98 Survey of Income and Housing Costs. The author found that the gender wage gap declined for both full-time and part-time employees. In contrast to these findings, Preston (2001) used the ABS Average Weekly Earnings Survey to compare the weekly ordinary time earnings of men and women in the Australian labour market between 1992 and 2000. The author found that the wage gap between men and women increased from 7.1 per cent in 1992 to 8.2 per cent in 2000. The mean and distributional change of the Australian gender wage gap in the 2000s is a gap in the literature. The analysis in Chapter 5 of this thesis will contribute to the existing literature by measuring and comparing the mean and distributional gender wage gap between 2001 and 2012. The estimation of an average gender wage gap provides an indication of whether or not a wage gap exists between men and women. However, through the use of the 10
conditional mean, it is assumed that the gender wage gap and its possible causes do not change along the wage distribution. 2.1.1 Distributional gender wage gap In response to this limitation, Australian studies over the last decade have considered the measurement of the gender wage gap along the earnings distribution (Whitehouse 2001; Eastough & Miller 2004; Miller 2005). Distributional gender wage gaps have been measured by implementing the quantile regression technique developed by Koenker and Bassett (1978). The quantile regression decomposition technique has been applied widely in Australian and international studies since the advancement of the technique by Buchinsky (1998a), Buchinsky (1998b), and Buchinsky and Hunt (1999). Whitehouse (2001) used percentile comparisons and highlighted that investigation of wage gaps beyond the mean are required. Eastough and Miller (2004) suggested that a smaller gender pay gap exists among low-wage earners compared to high-wage earners. Miller (2005) also reported a much larger wage gap among high-paid workers than lowpaid workers and found that differences in returns to schooling reduced the earnings gap among the least paid, but increased the gap for all other groups considered. Detailed Australian analysis of the gender wage gap along the earnings distribution include the work of Kee (2006) and Baron and Cobb-Clark (2010). Kee (2006) used Wave 1 (2001) of the Household Income and Labour Dynamics in Australia (HILDA) Survey to measure the gender wage gap along the earnings distribution and found evidence of a strong glass ceiling for women in the private sector. Baron and Cobb-Clark (2010) used a panel dataset by pooling Wave 1 to Wave 6 (2001 to 2006) of the HILDA Survey. The authors measured the gender wage gap along the earnings distributions in the private and public sector and found that irrespective of the sector of employment, difference in labour market characteristics explained the gender wage gap among low-paid workers. However, the difference among the high-paid workers could not be explained by difference in labour market characteristics. 2.1.2 Endogeneity In estimating the wage equation for men and women, it is likely that some of the independent variables are endogenous. Endogeneity occurs when the independent 11
variables in a regression are correlated with the error term. The majority of studies that measure the gender wage gap rely on a restrictive assumption that the independent variables are strictly exogenous and that these time-varying variables are not predetermined (see Kunze 2008 for a review). It can be argued that endogeneity exists within the independent variables of a wage equation including work history/experience, the number of children, and education. This discussion focuses on the endogeneity of work experience as women tend to have more interrupted work histories due to the responsibility of children and family and this has an impact on their earnings. The source of endogeneity for work experience arises from the correlation of work history with unobserved heterogeneity, which is captured in the error term of the regression, and also from non-random sample selection. This endogeneity causes inconsistency in the OLS coefficient estimates of the independent variables. The measurement error associated with the independent variable of interest (experience) can occur for two main data reasons. Firstly, the observed variable can be inaccurate due to reporting, computing or random non-response errors. Secondly, it is possible that the variable does not have an observed counterpart and a proxy must be used. For example, when information on work experience is missing, age or potential work experience are often used as a proxy. However, both of these proxies can cause biased estimates due to their inaccuracy in capturing actual work experience. OLS is consistent if the independent variables and the unobserved heterogeneity (in the error term) are uncorrelated, the independent variables are free from any measurement error, and the sample is a random draw of the population. While these issues are recognised in the literature, dealing with the problems is complicated. In the literature, the main source of endogeneity addressed is the correlation between the independent variables and the time-invariant unobserved individual heterogeneity. This problem can be resolved using fixed effects (as undertaken within this thesis), or using instrumental variable estimators. Fixed effects estimation is discussed in greater detail in Section 2.1.4. The use of instrumental variables to address endogeneity relies on the validity of the exclusion restrictions and the partial correlation of the instrument with the endogenous variable. The difficulty with this method is finding a valid instrument. A common instrument for work experience is the age variable. Studies using 12
longitudinal data derive powerful instruments including Kim and Polachek (1994) and Light and Ureta (1995), which applied the Hausman and Taylor (1981) estimator using de-meaned variables. Other studies use less restrictive instruments such as the lags of the endogenous variables (Arellano & Bond 1991; Arellano & Bover 1995). Kim and Polachek (1994) present one of the most extensive applications of a range of consistent and inconsistent estimators and conclude that variation in results depends on both the estimators and the set of instrumental variables. Sources of endogeneity can be expanded if the error term contains greater complexities than assumed. For example, in addition to person specific time-invariant fixed effects, the error term could contain individual-job effects, individual-firm effects, and timevarying fixed effects. The literature on the firm-specific and job-specific fixed effects is discussed in Section 2.1.4. Kim and Polachek (1994) argued that endogeneity is correlated with heterogeneity. The authors used the correlation between motivation and labour market participation as an example. The authors noted that a motivated worker will increase their labour market participation and decrease their home time. As such, the relationship between endogeneity and heterogeneity is evident as a result of the correlation between the independent variables and the error term in the wage equation. Kim and Polachek (1994) deal with this problem by using fixed effects estimation for heterogeneity and the Chamberlain s generalized specification test for endogeneity. In this thesis, the issues associated with time-invariant unobserved individual fixed effects at the mean are addressed in Chapter 6 and Chapter 7 using fixed effects models and panel data. However, other methods to address endogeneity, such as the instrumental variable method, have not been employed due to the lack of appropriate instruments. Instead, this issue has been left for future research. 2.1.3 Sample selection The decision to participate or not to participate in the labour market raises the issue of sample selection bias (Heckman 1979; Gronau 1973) in estimating the gender wage gap. Receiving a wage is dependent on employment, as such; the inclusion of an individual in the sample requires them to be employed. Depending on the study, additional sample restrictions may apply such as being an employee (rather than self- 13
employed), and working full-time. The sample selection problem is greater for women than for men because the selection bias declines as the selected sample is closer to 100 per cent of the population (Mulligan & Rubinstein 2008) 3. Further, as the decision to work might be related to unobserved factors, which can impact the observed wage, selection might be endogenous and this can potentially result in a sample selection bias in the OLS estimators (Wooldridge 2009, p.324). Sample selection bias arises because the offered wages are influenced by individuals decisions of whether or not to participate in the labour market. The decision to participation in the labour market is dependent on observed and unobserved factors, which can also impact the observed offered wage. An overestimation of the impact of labour market characteristics such as education on earning can arise if one estimates this relationship using a selected sample of labour market participants with observed earnings. The issue arises because some individuals who do not work and therefore, do not earn a wage have also received an education. One of the most common methods used in the literature to correct for sample selection bias is the Heckman (1979, 1976) method that treats the sample selection problem as an omitted variable. This method is referred to as the two-step or limited information maximum likelihood (LIML) method. See Chapter 3, Section 3.4 for the sample selection equations used in this thesis. The early work of Blau and Beller (1988) corrected for sample selection bias using the Heckman (1979) two-step correction method on the trend in the gender wage gap from 1971 to 1981. The authors found that correcting for selection bias reduced the gender wage gap for the sample of white individuals; however, the selectivity variables were insignificant for the sample of black individuals. In the Australian labour market literature, Kidd and Viney (1991) used the ABS Special Supplementary Survey No. 4, 1982 to estimate the gender wage gap while correcting for sample selection bias using the Heckman (1979) two-stage method. The authors found that the female wage equation was subject to sample selection bias and failing to correct for non-random selection would lead to an overestimation of the gender wage gap. 3 Mulligan and Rubinstein (2008) compare the Heckman two-stage correction method with the identification-atinfinity method of correction. 14
Miller and Rummery (1991) measured the gender wage gap of young Australians and controlled for sample selection issues using the Heckman correction. The authors found that the correction for sample selection produced important insights into male-female inequality, although, they noted that the selection correction method lacks efficiency. Specifically, the authors noted that the selection bias correction term (lambda) can be correlated with variables that determine market earnings, such as education. Also, the authors noted that the gender wage gap decomposition is likely to be more sensitive to distributional assumptions than omitted variable bias due to sample selection issues (Lincoln & Miller 1979). More recent work such as Kee (2006) showed no evidence of significant sample selection bias for women at the mean and along the wage distribution in both the private and public sector of the Australian labour market using Wave 1 of the HILDA Survey. Watson (2010) used a panel dataset compiled using Wave 1 to Wave 8 of the HILDA Survey and also did not find a significant self-selection effect for women. To investigate the effect of women s self-selection into full-time employment, Chzhen, Mumford and Nicodemo (2013) used the HILDA Survey to decompose the gender wage gap across the earnings distribution for full-time Australian employees in the private sector. The authors did not find a selection effect in the Australian private sector. That is, women in the sample working full-time in the Australian private sector did not have higher earnings potential than women in general, especially for women at the lower and upper ends of the earnings distribution. The sample selection problem for sectoral choice involves two sets of decisions. The first arises from the decision to work and the second arises from the choice to participate in either sector of the labour market. Some studies that measure the wage gap between the private and public sector and correct for sample selection include Blackaby, Murphy and O Leary (1999), Hyder and Reilly (2005), and Cai and Liu (2011). In these studies, a sample selection correction term for the choice of sector of employment is included within the wage equations for public and private sector workers. The general finding from these studies is that sample selection does not have a significant impact on public sector wage premiums. 15
Blackaby, Murphy and O Leary (1999, p.204) for example, noted that the findings of this paper are not contingent upon the treatment of the endogeneity of sector of employment choice. Similarly, as a result of the insignificant sample selection correction terms in Hyder and Reilly (2005), the authors noted that the differentials based on correcting for selection bias provide few new insights on either the magnitude or evolution of the public sector premium across the conditional wage distribution, (Hyder & Reilly 2005, p.21). Sample selection tests have been undertaken by Rodgers (2004) and Booth and Wood (2008) while estimating the wage gap between full-time and part-time Australian employees. Rodgers (2004) used Wave 1 of the HILDA Survey and a multinomial logit model for sample selection. The author found a statistically significant selection effect for full-time workers but not for part-time workers. In preliminary analysis, Booth and Wood (2008) estimated a sample selection model of participation and wage using Wave 1 to Wave 4 of the HILDA Survey. The authors did not find statistically significant sample selection issues. In general, the correction for sample selection in the Australian labour market literature has presented conflicting results (Miller & Rummery 1991; Kidd & Viney 1991; Lee 1983). This suggests that possible weaknesses of the procedure exist, specifically, the sensitivity of the method to distribution assumptions and identification restrictions (Puhani 2000; Stolzenberg & Relles 1997). It is possible that the application of this method can lead to more bias than it corrects (Le & Miller 1998). Despite the conflicting results of sample selection bias in the literature, tests for nonrandom sample selection will be undertaken in this thesis at the sample mean and along the wage distribution 4. The non-random sample selection bias will be tested in Chapter 5 and Chapter 6 using the Heckman (1979) two-stage method. This method will also be implemented in Chapter 7; however, given the nature of the selection problem, a multinomial logit selection model will be used instead of a binary probit model. The sample selection tests for quantile regressions in Chapter 5 follow Buchinsky (1998a) as described in Chapter 3 Section 3.4. 4 Distributional sample selection tests are undertaken for the cross section samples of 2001 and 2012 in Chapter 5. 16
For the distributional analysis undertaken in this thesis using panel data (Chapter 7), the potential problem of endogeneity in labour market and sector self-selection will not be accounted for. One method of addressing this issue is to estimate a fixed effects quantile regression model as proposed by Koenker (2004). Koenker's (2004) method allows for unobserved individual fixed effects to be differenced out and omitted from the estimation in linear models. However, unlike linear models, the time-invariant unobserved individual heterogeneity in a quantile regression needs to be estimated along with the coefficients. This is very difficult to implement with a large number of parameters, as is the case in the analysis of this thesis. Further, fixed effects models can only be used to resolve such endogeneity problems if the labour market participation and sector selection are a result of the time-invariant unobserved individual heterogeneity. However, when labour market participation and sector selection are dependent on the time-varying unobservables, fixed effects models cannot be used to resolve the endogeneity problem and instrumental variables are required. As a result of the difficulty associated with identifying appropriate instruments, this potential distributional endogeneity problem is left for future research. 2.1.4 Fixed effects Prior Australian studies have estimated the gender wage gap while accounting for observed labour market characteristics. This is limited as unobserved heterogeneity can influence earnings and the gender wage gap. The analysis in Chapter 6 of this thesis will contribute to this gap in the Australian literature by estimating the gender wage gap while accounting for time-invariant unobserved individual heterogeneity. Further, the analysis in Chapter 7 presents fixed effects estimates of the gender wage gap in the private and public sector of employment. A small number of Australian studies have accounted for fixed effects while estimating the gender wage gap by incorporating firm-specific heterogeneity (Meng 2004; Meng & Meurs 2004). Using the 1995 Australian Workplace Industrial Relations Survey, Meng (2004) measured and decomposed the mean gender wage gap while accounting for firmspecific fixed effects. The author found that at the sample mean, firm-specific wage policies reduced the gender wage gap. Further, the author found that firms with smaller 17