Goldin’s Last Chapter on the Gender Pay Gap: An Exploratory Analysis Using Italian Data

Introduction

While the gap in schooling between men and women has narrowed over the last century, there are still considerable gender gaps in pay levels (Organisation for Economic Co-operation and Development (OECD), 2002). According to a 2019 report on equality between women and men in the European Union, women earn on average 15% less per hour than men (European Commission, 2019). Significant heterogeneity can be observed among EU Member States: the gender pay gap varies from 3% in Romania to 23% in Estonia. Italy has one of the bloc’s lowest total gaps (5%) but a high unexplained (or residual) gap, which is often taken as a measure of gender pay discrimination. ¹ Despite the high institutional attention on this phenomenon, research on it in Italy has been relatively scarce. Notable contributions include Addabbo and Favaro (2011), Mussida and Picchio (2014a, 2014b) and Piazzalunga (2018). Addabbo and Favaro (2011) and Mussida and Picchio (2014a) find evidence of a substantial penalty for both low- and highly educated women. Mussida and Picchio (2014b) show that this unexplained gap increased over time, particularly at the top end of the pay distribution. Piazzalunga (2018), focusing on graduate workers, finds that the unexplained gap is higher where consensus with traditional gender norms about work–life balance is stronger, while it is reduced by greater availability of childcare. All these studies focus on the relationship between the unexplained gap and the educational attainment of workers but neglect the role of the occupational characteristics of jobs. A partial exception is provided by Addabbo and Favaro (2011), who utilise information about the principal activity performed and degree of responsibility of workers. ² In this article, drawing upon Goldin (2014), we expand upon these considerations and show that analysing pay determination within occupations sheds further light on the factors behind the residual pay gap in Italy. This effort will also provide a common explanation for some results of the above quoted studies.

Goldin (2014) highlights a relatively disregarded reason for the residual gender pay gap, contending that firms have an incentive to disproportionately reward individuals who work long and particular hours. If some workers want the amenity of workplace flexibility, firms may find it convenient to provide this in exchange for a lower wage, not least to incentivise other workers. A non-linearity (more-than-proportional relationship) ensues between earnings and hours worked, leading in turn to a gender pay gap not explained by differences in human capital or in other workers’ measurable characteristics. Indeed, if the elasticity of earnings with respect to hours worked is greater than one, employees who want fewer hours and more flexible employment experience an earnings penalty; as such employees are typically women, a gender gap consequently emerges.

This framework performs well empirically when applied to US data in Goldin (2014). However, to the best of the authors’ knowledge, it has not been applied outside the United States. This article addresses this gap by applying Goldin’s framework to Italian data. Goldin (2014: 1091) states explicitly that although her article deals with US evidence, her approach should have broader applicability. Italy is an interesting field of research, given its rather large residual pay gap and the above noted lack of attention in the literature about the occupational characteristics of jobs. When undertaking this effort, however, one must pay attention to the widely acknowledged differences between the Italian and US labour markets.

First, the EU Working Time Directive may affect the average working time of men and women in Italy more than the Fair Labor Standards Act (FLSA) does in the United States. Indeed, the FLSA places no limit on the number of hours an employee aged 16 years or older may work in any work week, while the Working Time Directive imposes, at least in principle, a limit of 48 hours per working week. Landivar (2015) finds that the gap in work hours between spouses is lower in countries with shorter maximum weekly work hours. Unfortunately, her study’s data do not allow a direct comparison of Italy with the United States. The OECD Statistics Portal reveals that the two countries have similar gender gaps in the work week of full-time workers for the reference year 2014 (5.3% in Italy vs 5.7% in the United States). ³

Second, the United States and other Anglo-Saxon countries have been characterised in recent decades by a stark decrease in unionisation and union coverage of collective bargaining. Albeit less radically, wage-setting institutions have also changed in a number of continental European countries, including Italy (OECD, 2004: ch. 3). These countries are now characterised by a system in which single-employer bargaining has developed alongside multi-employer bargaining. This setting should give scope for compensating wage differentials, such as envisaged by Goldin’s hypothesis, especially at the higher end of the earnings distribution, where unionisation and union coverage matter less (Destefanis and Naddeo, 2018; Lucifora and Meurs, 2006).

These features imply that, notwithstanding the clear differences in the two labour markets, Goldin’s framework has some potential for explaining the residual gender pay gap in Italy. ⁴ They also suggest the potential importance of considering different worker groups in the analysis, a point to which this article returns at the end of Section 2. The hypothesis to be tested is whether, in Italy, as in the United States, the unexplained component of the gender gap is a function of the elasticity of earnings with respect to work hours in different types of occupations. To measure this unexplained component, this study follows both Goldin’s own procedure and the well-known Oaxaca–Blinder (OB) decomposition method, the latter providing a robustness check of the analysis in Goldin (2014).

This study uses a large dataset of Italian responses to the 2014 European Structure of Earnings Survey (henceforth ‘SES’), which provides accurate and harmonised data on earnings, and matched information on the individual characteristics of employees and of their employer. Moreover, to assess whether some qualitative characteristics of occupations are related to the unexplained component of the gender gap, the study uses data from two other surveys: the 2013 Italian Sample Survey on Professions (henceforth ‘ICP’) and the 2014 Occupational Information Network (henceforth ‘O*Net’) online survey. Both surveys use the same questionnaire and collect the same kind of information for each respective country. However, they have different methods of disclosing collected information, with the O*Net survey reporting frequency distributions and the ICP reporting averaged responses. As the ICP’s approach leads to a potential loss of explanatory power, it is advisable to use both sources of information.

The rest of the article is structured as follows. Section 2 briefly surveys the literature on the gender gap and provides a more detailed account of Goldin’s (2014) framework. Section 3 presents the empirical procedures and data. The results, including several robustness checks, are reported and discussed in Section 4. Finally, Section 5 concludes the article.

The gender pay gap and the Goldin hypothesis

The literature on the determinants of gender pay gaps provides an extensive set of theories to explain the persistence of this phenomenon. This article examines the residual gender pay gap – the portion not explained by gender differences in human capital or other determinants of labour productivity. As Goldin (2014) and Bertrand (2018) recognised, the explained portion of the gender pay gap has decreased over time as differences in years of education and of labour market experience between men and women have gradually narrowed. ⁵ Consequently, the relative importance of the residual gender pay gap has increased over time.

Among the competing theoretical explanations of this residual gap, the role of psychological attributes has come to the fore, since Babcock and Laschever (2003) evidenced that women bargain less fiercely, and Gneezy et al. (2003) argued that women are less interested than men in competing in the labour market. These views are still controversial, with findings that support them being provided by Niederle and Vesterlund (2007; see also Bertrand, 2011; Croson and Gneezy, 2009), and evidence to the contrary coming from Manning and Saidi (2010) and Artz et al. (2018). At any rate, considerable uncertainty remains over the extent to which these psychological traits quantitatively explain the gender pay gap (Blau and Kahn, 2017). Moreover, these approaches do not explain why the residual gender pay gap tends to increase with age, or why this gap narrows decisively for women without children, and they downplay the role of social norms in shaping gender-related preferences. These points are picked up again below in this section.

Goldin (2014) also notes that the gender pay gap cannot be satisfactorily explained by approaches based on the distribution of men and women between occupations, and therefore focuses on explaining the residual gap through an analysis within occupations. ⁶ Goldin’s main idea is that the residual gender pay gap relates to how firms reward workers with different propensities for workplace flexibility. The concept of workplace flexibility ‘incorporates the number of hours to be worked and also the particular hours worked, being “on call,” providing “face time,” being around for clients, group meetings, and the like’ (Goldin, 2014: 1094). Individuals differ in how they value workplace flexibility, and firms face different costs in providing this amenity. Consequently, individuals accept lower earnings in exchange for more flexible workplaces. If workers were perfect substitutes for one another, there would be no premium on earnings with respect to the number or timing of work hours, and earnings would move linearly with respect to hours. However, as workers are not in fact perfectly substitutable, those whose working hours are relatively shorter and/or not manipulable by the firm produce a loss for the organisation that is deemed to justify a penalty. Hence, a non-linear (more-than-proportional) relationship arises between earnings and work hours.

This simple idea has important consequences. First, it is women who typically want more workplace flexibility, and the gender gap that emerges is residual in the sense that it cannot be explained by workers’ characteristics. Second, the residual gender gap is decisively linked to the existence of a non-linear relationship between earnings and work hours within each occupation. A less clear-cut issue is whether this phenomenon is found mainly at the higher end of the earnings distribution. Goldin (2014) focuses on highly educated workers but nonetheless shows that, for individuals at the lower end of the distribution, there is still an earnings premium for people who work longer hours. Another explanation of the link between a non-linear relationship between earnings and work hours and the gender pay gap is proposed by Blau and Kahn (2017): under the traditional division of labour by gender in the family, women anticipate shorter and more discontinuous work lives because of their family responsibilities, and so may be less interested to invest in firm-specific human capital, particularly by committing to long and particular hours of work.

Goldin’s hypothesis is taken up by Bertrand (2018) and Cortés and Pan (2019), who highlight some possible developments of the approach. Bertrand (2018) validates Goldin’s finding that the elasticity of earnings to hours worked is larger in the higher-paying occupations. Moreover, she finds that the reduction of the gender wage gap over time is hampered by the growth in women’s commitment to work in occupations that reward long and particular hours worked. Cortés and Pan (2019) highlight that the relaxation of constraints hindering women from supplying long and particular hours (for instance through a higher availability of caregivers) increases the relative earnings of highly skilled women in the higher-paying occupations.

Both Bertrand (2018) and Cortés and Pan (2019) deal with US data. Naturally, when applying Goldin’s framework to other countries, its key features (also relating to sociological implications) must be critically appraised. As suggested by Blau and Kahn (2017), Goldin’s approach can be seen as a compensating differential equilibrium, with female workers placing a higher value on types of work that do not necessitate long and particular work hours, and firms internalising in the wage schedule the costs associated with these preferences. There is recent evidence showing that women place higher value on jobs featuring more flexibility or shorter working hours (Mas and Pallais, 2017; Shahriar, 2018). There is also a large amount of literature on the ‘motherhood pay gap’, ⁷ whose relevance for her approach was noted by Goldin (2014: 1094).

An important point that must be stressed here is that these gender-related preferences are largely shaped by social norms (this remark also applies to the above-discussed psychological traits related to bargaining and competition). Kleven et al. (2019) find suggestive evidence that the motherhood pay gap is transmitted across generations through the influence of the family and social environment on gender identity. Matteazzi and Scherer (2021) likewise stress the role of social norms; their study finds that women’s domestic work reduces their own earnings while increasing the earnings of their partners. A related point is that employers’ preferences may also be shaped by social and institutional norms. For instance, Parry et al. (2005), Lanfranchi and Narcy (2015) and Powell and Cortis (2017) document that not-for-profit organisations, often unwilling or unable to use monetary rewards, may be readier than private firms in offering flexible work-time schedules to attract or retain their workforce. ⁸ This issue must be placed within the wider literature of work–life balance, which has recently come to the fore amid the impact of the COVID-19 pandemic on flexible working. This literature also stresses the role of socially based norms in shaping individual preferences.

Clawson and Gerstel (2014) argue that in societies with deeply rooted traditional gender norms, flexible working helps to fulfil these norms. In studies that specifically examine the expansion of flexible working hours from a gender perspective, men have been found more likely to expand their working hours than women (Glass and Noonan, 2016; Lott and Chung, 2016), whereas women tend to exploit increases in flexible working by expanding their care and household activities (Hilbrecht et al., 2013; Singley and Hynes, 2005; Sullivan and Lewis, 2001).

Very interestingly for the present study, Lott and Chung’s (2016) analysis of longitudinal data from Germany reveals that mothers given the opportunity to work flexibly apparently accept longer overtime with no monetary reward for additional hours worked. In a similar vein, Chung and Van der Horst (2018) analyse the relationships between various forms of flexible working and unpaid overtime hours for men and women, parents and non-parents in a UK longitudinal dataset. They find that increases in unpaid overtime hours, a strong factor in achieving promotion and higher pay in the long run, are largely driven by professional men and women without children, especially if flexible working is introduced as part of high-performance work practices. Working mothers, by contrast, are unlikely to profit from this access to the career ladder.

We draw the following considerations from the above overview. First, the relevance of social norms to preferences at the base of Goldin’s compensating differentials framework highlights the importance of considering different groups of workers separately in any empirical analysis. This point also follows from the earlier introduction of the Italian and US labour markets, particularly concerning the role of unions. Indeed, Goldin herself followed the approach of disaggregation, focusing only on the lower end of the earnings distribution in Goldin (2015). Given the characteristics of this study’s dataset (presented in the following section), separate analyses should be performed for graduate and non-graduate workers. Second, compensating differentials may not be the only explanation of how a non-linear relationship between earnings and work hours is linked to the gender pay gap. This nexus could also arise from the relative lack of interest among women to invest in firm-specific human capital by working long and particular hours. Finally, the deep connection between Goldin’s analysis and the work–life balance literature means that the analytical and empirical relationships highlighted by this approach are very relevant for assessing the impact of flexible working on the gender gap.

Empirical framework

Data

Our main dataset is from the 2014 SES, which details earnings and individual characteristics of employees (sex, age, occupation, length of service, highest educational level attained, type of contract, public or private sector, and work hours) matched with the individual characteristics of their employer (branch of economic activity, size and location). ⁹

The baseline sample includes workers aged 20 to 60+ with earnings between the first and ninth percentiles of the earnings distribution, working full-time (30+ hours) and with a contract for three months or more. The key variables from this sample are gross earnings from labour for the reference month and the number of paid hours for the reference month. ¹⁰ To compare the behaviour of different segments of the labour force, the empirical analysis is based on two subsamples, labelled ‘Full-time graduate’ and ‘Full-time non-graduate’. Intuitively enough, the first subsample includes individuals who have at least an undergraduate degree. Part-time workers are excluded to enable focus on a more homogeneous set of workers and earnings. Proper modelling of part-time work would require a dataset containing more information on determinants of the labour supply. Moreover, as females in the workforce are usually segregated in part-time jobs, considering the latter may be more relevant for the between-occupation gender gap, whereas this article focuses on the within-occupation gender gap. ¹¹

Our subsamples are of comparable size and can be broken down into occupational cells containing sufficient numbers of workers for analytical purposes. The numbers of observations in each subsample are reported in Table 1. In total there are 47,515 observations in the full-time graduate category (37% of the overall sample) and 79,546 in the full-time non-graduate category (63%).

OPEN IN VIEWER

Table 1.

Observations for each subsample by gender and education category.

	Full-time graduate		Full-time non-graduate
	Number	%	Number	%
Female	21,910	46.11	30,321	38.12
Male	25,605	53.89	49,225	61.88
Total	47,515	100	79,546	100

Source: Own calculations based on 2014 SES.

To analyse the role of occupation in explaining the gender gap, the SES observations are decomposed into the three-digit occupations of the International Standard Classification of Occupations (ISCO; reported in online Table A1). Of the 111 occupations reported in the original dataset, 66 occupations are selected in which the total number of observations in any subsample numbers at least 25 individuals for each gender. Below this number of observations, estimates of the residual gender gap for an occupational cell are unreliable.

To control for relevant qualitative characteristics of the occupations, this study uses data from the 2013 ICP survey and the 2014 O*Net survey. The ICP survey employs the same structure as the O*Net survey, examining occupational characteristics in seven sections (knowledge, skills, attitudes, generalised work activities, values, work styles, and working conditions). However, whereas the O*Net survey discloses frequency distributions for all questions, the ICP survey only reports averaged responses. This leads to a potential loss of explanatory power, which makes it advisable to draw on both sources of information. Assuming the characteristics of a given occupation are similar in Italy and in the United States, the following analysis of the residual gender gap also relies on the O*Net selected frequencies. ¹² The Appendix provides further details on the surveys and the construction of the following five occupational indicators:

Time pressure: the percentage of individuals required to meet strict deadlines once a week or more;

Contact with others: the percentage of individuals who have contact with others either most of the time or constantly;

Structured versus unstructured work: the percentage of individuals who have some or high discretion in determining their tasks, priorities, and goals (more discretion makes the worker less substitutable);

Freedom to make decisions: the percentage of individuals with some or high freedom in decision-making, and consequently more responsibilities;

Establishing and maintaining interpersonal relationships: the average on a 1–5 scale of the importance to job performance of establishing and maintaining interpersonal working relationships.

Table 2 presents means and standard deviations for each of the five variables across the ISCO three-digit occupations. In both surveys, the highest values were found for individuals who reported having contact with others either most of the time or constantly. However, differences across the two surveys were evident when looking at the averages, in particular for the time pressure indicator.

OPEN IN VIEWER

Table 2.

Mean ICP and O*Net values for selected occupation characteristics.

Indicator (ISCO classification)	ICP (averages)		O*Net (selected frequencies)		O*Net (averages)
	Mean	SD	Mean	SD	Mean	SD
Time pressure (H54)	70.819	9.792	66.139	12.227	3.823	0.358
Contact with others (H6)	85.174	8.027	88.004	9.384	4.491	0.291
Less structure, more discretionality (H52)	67.927	13.044	80.463	12.423	4.153	0.326
Freedom in making decisions (H48)	65.630	15.336	81.639	11.903	4.202	0.330
Interpersonal working relationships (G28)	68.009	9.387	3.869	0.320	3.869	0.320

Note: The variable G28 is recorded as a mean as there are no frequencies for it; the O*Net variables are coded with values from 1 to 5, the ICP with values from 0 to 100.

Source: Own elaborations based on ICP (2013) and O*Net (2014).

Note that in the O*Net survey the average and selected frequency values for interpersonal working relationships are the same, as this variable is computed as an average on a 1–5 scale.

Results

A first point to be made is that, using the same kind of within–between decomposition proposed by Goldin yielded a within gap of either 60% or 56% for the full-time graduate subsample, and hence a between gap of either 40% or 44%. For the full-time, non-graduate subsample, these figures became 74% or 70% for the within gap and 26% or 30% for the between gap (calculations are detailed in the online Appendix). Somewhat surprisingly, Goldin’s hypothesis is not only of interest for the Italian data but also potentially relevant to the lower end of the earnings distribution.

Turning now to the analysis, Table 3 illustrates the impact on the residual gender gap of work hours and the Occupation dummies.

OPEN IN VIEWER

Table 3.

Estimates of the residual gender gap.

Sample	Specification	(a) Estimates from Eqs (1)–(3)			(b) OB decomposition
		Female	SE	R 2	Unexplained gap	SE
FTG	(1)	−0.199	0.004	0.323	−0.203	0.004
FTG	(2)	−0.190	0.004	0.333	−0.198	0.004
FTG	(3)	−0.127	0.003	0.549	−0.133	0.004
FTNG	(1)	−0.140	0.003	0.293	−0.145	0.003
FTNG	(2)	−0.141	0.003	0.293	−0.147	0.003
FTNG	(3)	−0.116	0.002	0.468	−0.121	0.003

Notes: Specification (1) only includes basic covariates; (2) includes basic covariates and work hours; (3) includes basic covariates, work hours and Occupation dummies. Eqs (1)–(3) are detailed in the online Appendix. FTG: full-time graduate; FTNG: full-time non-graduate.

Source: Own calculations based on 2014 SES.

In panel (a), for the full-time graduate subsample, the Female coefficient was initially −0.199, meaning that the monthly earnings of a female are 18% lower than those of a male. After adding the Occupation dummies to the estimate, the coefficient on Female became −0.127. A similar, although less stark, reduction occurred in the full-time, non-graduate subsample (the coefficient on Female finally becoming −0.116). Earning variability across occupation groups was an important source of gender inequality, as reinforced by variations in R² across specifications. This behaviour is confirmed for the OB decomposition in panel (b): after adding the Occupation dummies, the unexplained gender gap fell from −0.203 to −0.133 for the full-time graduate subsample, and from −0.145 to −0.121 for the full-time, non-graduate subsample. These values for the residual gap comfortably fall within the range of those obtained in the recent Italian empirical literature (Addabbo and Favaro, 2011; Mussida and Picchio, 2014a; Piazzalunga, 2018). This indicates that the present analysis draws upon statistical information highly consistent with that used in previous studies and can enrich the literature about the determination of the residual gender pay gap in Italy.

According to Goldin (2014), a large part of the relationship between the residual gender gap and the Occupation dummies is linked to the non-linearity between earnings and work hours. To further understand this phenomenon, Figure 1 plots the residual gender pay gap for each occupation (the sum of coefficients γ₁ and γ₅ from Eq. (4) in the online Appendix) against the elasticity of earnings with respect to work hours for each occupation (the sum of coefficients δ₃ and δ₆ from Eq. (5) in the online Appendix). This exercise was carried out for both graduates (panel (a)) and non-graduates (panel (b)). In both cases, occupations with a higher elasticity had wider (more negative) gender pay gaps, in line with Goldin’s hypothesis. The same results were found when the residual pay gap was obtained through the OB decomposition (Figure 2). Clearly, Goldin’s hypothesis applies to both the higher and the lower segments of the Italian labour market.

OPEN IN VIEWER

Figure 1.

Residual gender pay gap on elasticity of earnings with respect to work hours for each occupation. (a) Graduate; (b) Non-graduate.

OPEN IN VIEWER

Figure 2.

Unexplained gender pay gap from OB decomposition on elasticity of earnings with respect to work hours for each occupation. (a) Graduate; (b) Non-graduate.

The graphical analyses in Figures 1 and 2 also make clear that these relationships were affected by considerable noise, highlighting the existence of many potential outliers. Indeed, a non-negligible number of occupations were characterised by strongly negative elasticity of earnings with respect to work hours. In light of these considerations, robust correlation techniques were adopted to appraise the strength of the relations shown in Figures 1 and 2. In addition to the usual Pearson coefficient, the robust Spearman rank correlation coefficient was calculated and the Pearson coefficient was recalculated for a sample in which potentially influential outliers had been detected using the Andrews–Pregibon (1978) method and excluded (details of this method and the nature of the anomalous observations are provided in the online Appendix).

Table 4 reports the size and significance of the correlation coefficients between elasticity of earnings with respect to work hours and: (a) the Female coefficients from Eq. (4); ¹⁴ and (b) the unexplained gap from the OB decomposition (both already considered in Figures 1 and 2).

OPEN IN VIEWER

Table 4.

Correlation coefficients between residual gender pay gap and elasticity of earnings with respect to work hours.

Full-time graduate	Female coeff. a (Eq. (4))		OB unexplained gap
	Coeff.	T-stat	Coeff.	T-stat
Pearson	−0.28	−2.35	−0.12	−0.93
Spearman	−0.21	−1.68	−0.06	−0.44
Pearson (no outliers)	−0.29	−2.22	−0.21	−1.66
Full-time non-graduate	Coeff.	T-stat	Coeff.	T-stat
Pearson	−0.35	−3.03	−0.17	−1.35
Spearman	−0.37	−3.16	−0.32	−2.72
Pearson (no outliers)	−0.40	−3.30	−0.65	−6.45

Note: ^aFemale coeff. stands for γ₁ + γ₅ from Eq. (4) in the online Appendix.

Source: Own calculations based on 2014 SES.

Table 4 shows that occupations with greater elasticity had wider gender pay gaps across all gap measures for both graduate and non-graduate workers. This relationship was particularly significant for the Female coefficients from Eq. (4) and for non-graduate workers.

As explained in previous sections, Goldin’s analysis relates the residual gender gap (or the elasticity of earnings with respect to work hours) to the characteristics of an occupation that prompt firms to impose earnings penalties on workers who seek more flexibility. To measure these characteristics, the 2013 ICP and the 2014 O*Net surveys were used. Figures 3 to 5 respectively present graphical analysis of the residual gender earnings gap (Female coefficients), the OB unexplained gap and the elasticity of earnings with respect to work hours for each occupation against the arithmetic mean of selected frequencies for key O*Net characteristics chosen following Goldin (2014).

OPEN IN VIEWER

Figure 3.

Residual gender earnings gap for each occupation on the arithmetic mean of selected frequencies for key O*Net characteristics. (a) Graduate; (b) Non-graduate.

OPEN IN VIEWER

Figure 4.

Unexplained gender earnings gap from OB decomposition for each occupation on the arithmetic mean of selected frequencies for key O*Net characteristics. (a) Graduate; (b) Non-graduate.

OPEN IN VIEWER

Figure 5.

Elasticity of earnings with respect to work hours for each occupation on the arithmetic mean of selected frequencies for key O*Net characteristics. (a) Graduate; (b) Non-graduate.

The relationships under scrutiny appeared to be stronger for graduates but, once again, were affected by potential outliers. Hence, Table 5 reports the same correlation analysis adopted in Table 4 linking the synthetic occupational indicators with the Female coefficients, the OB unexplained gap and the elasticity of earnings with respect to work hours. Including the earnings elasticity in this correlation analysis logically follows from Goldin’s hypothesis and elucidates the robustness of the whole approach. A positive relationship was expected between elasticity and the occupational characteristics that, according to Goldin, prompt firms to impose earnings penalties on flexible workers.

OPEN IN VIEWER

Table 5.

Correlations between the means of five occupational characteristics and the residual gender earnings gap, the OB unexplained gap and elasticity of earnings with respect to work hours.

Occupational characteristic	Female coeff. a (Eq. (4))		OB unexplained gap		Elasticity of earnings with respect to work hours
Full-time graduate	Coeff.	T-stat	Coeff.	T-stat	Coeff.	T-stat
ICP normalised averages
Pearson	−0.24	−1.99	0.00	0.01	−0.10	−0.82
Spearman	−0.13	−1.02	0.07	0.52	0.10	0.78
Pearson (no outliers)	−0.12	−0.96	0.00	0.01	0.27	2.28
O*Net normalised averages
Pearson	−0.17	−1.34	0.03	0.23	0.00	0.01
Spearman	−0.17	−1.40	−0.03	−0.21	0.25	2.05
Pearson (no outliers)	−0.20	−1.60	0.01	0.09	0.51	4.72
O*Net selected frequencies
Pearson	−0.20	−1.66	0.01	0.05	0.06	0.48
Spearman	−0.20	−1.67	−0.06	−0.51	0.32	2.65
Pearson (no outliers)	−0.26	−2.16	−0.26	−2.12	0.56	5.34
Full-time non-graduate	Coeff.	T-stat	Coeff.	T-stat	Coeff.	T-stat
ICP normalised averages
Pearson	−0.02	−0.14	−0.13	−1.09	−0.03	−0.20
Spearman	−0.07	−0.56	−0.18	−1.44	0.03	0.23
Pearson (no outliers)	−0.02	−0.13	−0.11	−0.86	−0.00	−0.02
O*Net normalised averages
Pearson	0.02	0.19	−0.10	−0.82	−0.05	−0.42
Spearman	−0.07	−0.55	−0.15	−1.20	0.06	0.44
Pearson (no outliers)	0.03	0.21	−0.08	−0.61	−0.01	−0.05
O*Net selected frequencies
Pearson	−0.01	−0.06	−0.13	−1.04	0.01	0.10
Spearman	−0.08	−0.67	−0.17	−1.36	0.13	1.04
Pearson (no outliers)	−0.00	−0.03	−0.10	−0.81	0.07	0.54

Note: ^aFemale coeff. stands for γ₁ + γ₅ from Eq. (4) in the online Appendix.

Source: Own calculations based on 2014 SES, 2013 ICP and 2014 O*Net surveys.

The results for graduate workers revealed the expected relationships, especially for the Female coefficients and the elasticity of earnings with respect to work hours. However, only the correlations with the mean of selected frequencies of the O*Net characteristics were significant, as in Goldin (2014). The means of the normalised averages of the ICP and O*Net characteristics were much less consistently significant. The irretrievability of original frequencies for the Italian data likely led to a loss of potentially relevant information. Even more importantly, no significant correlations emerge from Table 5 for non-graduate workers. Apparently, for these workers the relationship between the residual gap and the earnings elasticity that was found in Table 4 does not depend on the compensation differential mechanism highlighted in Goldin (2014).

The evidence can be summarised as follows. Analysing pay determination within occupations proved capable of shedding further light on the factors behind the residual pay gap in Italy. This gap increased as a function of the elasticity of earnings with respect to work hours. This relationship held true for both graduates and non-graduates; somewhat surprisingly, it was much stronger for the latter. As a robustness check of Goldin’s framework, her preferred measure of the residual gender pay gap was substituted with the classic OB measure of gender pay discrimination. Overall, the same results were obtained. However, the similarity between the results from the two gap measures extended to a feature of this study’s empirical analysis that Goldin’s approach cannot easily explain. For non-graduate workers, the evidence did not support a significant relationship of the residual gap and the elasticity of earnings with the occupation characteristics that prompt firms to impose an earnings penalty on workplace flexibility. This discrepancy suggests that the relationship between the residual gap and the elasticity of earnings with respect to work hours does not depend for non-graduate workers on the compensating differential mechanism singled out by Goldin. In this segment of the labour market, the relationship between the residual gap and the earnings elasticity could potentially be explained by the relative lack of incentives for women to invest in firm-specific human capital through longer working hours. The present article’s data do not allow testing of this hypothesis. A concomitant explanation is that, for non-graduate workers, whose wages are more heavily determined by collective wage agreements (Destefanis and Naddeo, 2018; Lucifora and Meurs, 2006), the scope for individual or firm-level bargaining is more limited.

Concluding remarks

This article attempts to unravel the unexplained gender pay gap in Italy in light of Goldin’s (2014) hypothesis that this gap is crucially linked to firms’ incentive to disproportionately reward individuals who work long and particular hours. The study uses Italian data from the 2014 SES, and employs the O*Net and ICP survey datasets to measure important occupational characteristics. The results reveal a sizeable gender earnings gap within occupations for both (full-time) graduates and non-graduates. Furthermore, there is a positive relationship between the unexplained gender pay gap and elasticity of earnings with respect to work hours for both graduate and non-graduate workers. However, a relationship between the pay gap, earnings elasticity and occupational characteristics considered as disamenities is only found for graduate workers.

Analysing pay determination within occupations offers novel insights on the factors behind the residual pay gap in Italy. The relationship between this gap and the elasticity of earnings with respect to work hours helps to explain why Mussida and Picchio (2014b) find that this gap increased over time. As pointed out by Bertrand (2018), earnings penalties on workers seeking more flexible work schedules are likely to be gaining strength in Western societies. Furthermore, the relationships between the gap and traditional gender norms or childcare availability found in Piazzalunga (2018) may arise because these factors make it, respectively, more and less difficult for women to supply the long and particular hours of work eventually demanded by employers. Both these insights are relevant for the policy implications of this article that are discussed below.

As was highlighted in Section 2, not-for-profit organisations may offer flexible work-time schedules more readily than private firms. Future research on the consequences of this behaviour for the gender pay gap could be relevant for the Italian labour market. ¹⁵ The analysis of presenteeism could give rise to another potentially interesting extension of this article. In the Italian case, however, the quantitative importance of this phenomenon is relatively limited (Kwon, 2020). In future work, the authors also plan to elucidate the different behaviour found for graduates and non-graduates. Ideally, this investigation should use datasets that include information on firm-specific human capital, as this is the key variable highlighted by the alternative explanation proposed for the relationship between the unexplained gender pay gap and elasticity of earnings with respect to work hours for non-graduate workers. At the very least, these datasets should include past work histories: this information could be a proxy measure for firm-specific human capital, while past spells of inactivity due to motherhood or more general family responsibilities are important determinants of female workers’ preferences. It would be useful for these datasets to include countries other than Italy or the United States. Testing Goldin’s approach in other countries is of high potential interest given that the OECD average gender gap in the work week has risen from 4.3% in 2014 to over 6% in recent years (data sourced from the OECD Statistics Portal). Using data from other countries would also provide information about how collective bargaining and social norms influence the process of wage determination.

Acquiring further knowledge about the relevance of collective bargaining, the economic environment and social norms is also critically important for using evidence from this article to formulate policy prescriptions. Bertrand (2018) depicts the current situation as one in which countervailing forces are at work: while technology has lessened the time load of homecare, thus reducing the pressure for women to demand more flexible work schedules, the factors prompting firms to impose earnings penalties on workers seeking more flexibility also appear to be gaining strength. Given this potential deadlock, policy action aimed at weakening traditional gender norms may be decisive. In this sense, Bertrand (2018) highlights the potential importance of dedicated paternity leave – a quota of parental leave for the father, which is lost if not taken up. She perceives that other family-friendly policies would increase female participation in the labour force without reducing the potential penalty for demanding workplace flexibility. However, in light of this article’s evidence, policies helping women to stay in the labour force, notwithstanding surges in family demands, may incentivise their investment in firm-specific human capital and thus narrow the gender pay gap associated with longer work hours. Likewise, the literature surveyed in Section 2 makes clear that in the presence of deeply rooted views on gender roles and responsibilities, the relationship between flexible working and work–life balance is likely to have different outcomes for men and women (Chung and van der Horst, 2018; Chung and van der Lippe, 2020; Lott and Chung, 2016; Sullivan and Lewis, 2001). However, if flexible working helps mothers to rebuild their working hours after childbirth (Chung and van der Lippe, 2020) and remain in human capital-intensive jobs in times of heavy homecare burden (Fuller and Hirsh, 2019), the evidence in this article suggests that flexible working could reduce the gender pay gap, particularly for non-graduate workers. A more general conclusion is that changes in gender norms, such as those potentially associated with dedicated paternity leave, are relevant for both graduates and non-graduates.

Summing up, the evidence from this article supports Goldin’s intuition that the gender pay gap is related to the elasticity of earnings with respect to work hours. However, the Italian case sheds doubt on the idea that earnings penalties for requesting workplace flexibility may explain this link for all workers and in all institutional setups. This novel evidence is important for guiding policy action, implies that the impact of different policies may vary across segments of the labour force (depending, for instance, on the relative strength of collective bargaining) and accordingly suggests some avenues for future research.

Supplemental Material