Effects of automation on the gender pay gap: the case of Estonia

Abstract

This paper investigates how investments in automation affect the gender pay gap. The evidence of the effects of automation on the labor market is growing; however, little is known about the implications of automation for the gender pay gap. The data used in this paper are from a matched employer–employee dataset incorporating detailed information on firms, their imports, and employee–level data for Estonian manufacturing and service employers for the period of 2006–2018. Through the use of the imports of automation goods as a proxy for the introduction of automation at the firm level, this paper estimates the effect of automation using simple Mincerian wage equations. The causality of the effect is further validated using propensity score matching (PSM). We find that introducing automation enlarges the gender pay gap, and PSM confirms that this also has a higher causal effect on the wages of male employees than female employees. The results imply that a higher representation of women in higher-paid positions does not guarantee a reduction in the gender pay gap in the presence of automation, and appropriate measures in education and retraining are needed to tackle the effect of automation on gender inequality.

1. Introduction

The research object of this article, the effect of introducing automation and its impact on labor dynamics, has been analyzed in several previous studies (Calvino and Virgillito, 2018; Graetz and Michaels, 2018; Dosi and Mohnen, 2019; Acemoglu and Restrepo, 2020; Aksoy et al., 2021; Domini et al., 2022; Mondolo, 2022). Studies on the relationship between automation and the gender pay gap are still limited, which restricts policymakers from acting (Domini et al., 2022). However, the studies that are available show evidence of an increased gender pay gap due to automation (Aksoy et al., 2021). At the same time, Domini et al. (2022) report that most of the wage dispersion in the French economy exists due to differences among workers belonging to the same firm rather than differences between sectors, firms, and occupations and that within-firm inequality and the gender pay gap are unaffected by an automation event. This article includes an evaluation of the gender pay gap with respect to automation technologies implemented in Estonia, taking into account previous studies related to the Estonian market and innovation (Masso and Vahter, 2020; Masso et al., 2022). Indeed, country-specific aspects, such as institutional and societal settings, can have relevant effects on the introduction of new technologies (Mondolo, 2022). Moreover, we observe in detail how the effects of automation can change depending on employee positions and that different types of automation provoke different results. Hence, the authors formulate the following research question: how does the introduction of automation at the firm level affect the gender pay gap in Estonia?

Estonia is a good example for investigating the effects of automation and the gender pay gap. Estonia has one of the fastest-growing information technology sectors among European countries,¹ which develops respective technological innovations. Additionally, in recent decades, all industries in Estonia have been successful at modernizing technologies and making companies competitive by, primarily, introducing automation and technology transfer from abroad (Kalvet, 2004). In Europe, the exposure of employees to industrial robots in 2016 was 19 percentage points higher than the exposure of workers in the USA (Chiacchio et al., 2018). Additionally, Estonia has had the largest gender pay gap among European Union (EU) Member States (up to 30%, Anspal, 2015a; Vahter and Masso, 2019), although a long-term downward trend can be observed (Meriküll and Tverdostup, 2020; Masso et al., 2022), Estonian figures still place at the top end at 21.7%, while the average gender pay gap is still approximately 14%, with some variation between countries (Eurostat series earn_gr_gpg, earn_gr_gpgr2 in 2019). Additionally, previous research has provided evidence that a large part of the gender pay gap is related to firm-level factors (Masso et al., 2022). Being a post-communist economy and a small and fast-developing country, Estonia provides a high level of female employment compared to the average in other European countries (Tverdostup and Paas, 2017). Yet, despite the high gender pay gap being actively investigated and debated by researchers (Tverdostup and Paas, 2016; Vahter and Masso, 2019), the majority of the gap remains unexplained.

This paper aims to provide evidence of the relationship between the introduction of automation and the gender pay gap using Estonian data. This paper uses the import of automation goods as a proxy for the introduction of automation at the firm level and measures the estimated effect of automation on the gender pay gap. Our work uses similar definitions of automation technologies and methods as in Acemoglu and Restrepo (2018) and Domini et al. (2022). Moreover, this paper takes the research a step further by focusing on the size of the gender pay gap and how it is affected by automation. To understand how the effects of automation work, it is also critical to study the data of employees, and so we link annual automation costs (using data about the imports of automation goods) to the Estonian matched employer–employee data. The data are provided by Statistics Estonia and cover the period of 2006–2018. Finally, the work differentiates itself from the paper by Blanas et al. (2019) by using firm-level data instead of industry-level data. Indeed, other previous literature has also stressed how the effects of technology can depend on specific firms or industries (Calvino and Virgillito, 2018; Dosi and Mohnen, 2019). Moreover, the paper also differs from Masso and Vahter (2020), who underlined evidence of the gender pay gap increase in firms with technological and non-technological innovation but did not include the introduction of automation and how that is connected to the gender pay gap.

To analyze the effects of automation on the gender pay gap in Estonia, we use Mincerian wage equations and observe the possible variation of the results with the introduction of innovation variables, as in Masso and Vahter (2020). Mincerian wage equations are standard tools employed in similar studies by Vannutelli et al. (2022) and Damiani et al. (2020). Furthermore, Mondolo (2022) and (Montobbio et al., 2021) show the effect of automation on different occupations and that different genders can select or be assigned to positions that are affected by automation differently, such as positions that are more easily automated. Likewise, we also examine the variations of the effects of automation on the gender pay gap in different occupations. Finally, we perform a standard propensity score matching (PSM) analysis (Rosenbaum and Rubin, 1983) considering not only automation, in general, but also the different kinds of automation (e.g. welding machines, industrial robots, etc.) as well as the level of education of different workers. In doing so, we make a note of how workers are assigned to different tasks and positions is part of the work organization (Cetrulo et al., 2020) and that male and female workers can be in positions where some automations are more relevant than others.

Our findings show how firms investing in capital goods that are intensive in automation technologies affect the gender pay gap in Estonia. The results show a strong positive effect of the import of automation goods on the gender pay gap; however, they also show that the effect varies over the years. In addition, the introduction of innovation variables from the Community Innovation Survey (CIS) does not substantially affect the analysis results. Furthermore, certain occupations, such as managers, technical professionals, and support workers, show a greater increase in the gender pay gap due to automation. Finally, the results of the PSM are in accordance with the results of the Mincerian wage equations and illustrate how different kinds of automation can lead to different levels of the gender pay gap. In addition to contributing to the literature on the effects of automation on wages (Bessen, 2016; Lankisch et al., 2017; Domini et al., 2022), this paper also contributes to the literature that evaluates the various drivers of the male–female wage gap (Blau and Kahn, 2000; OECD, 2012; Mondolo, 2022). Finally, the paper adds insights into the recent contributions in the field of labor economics (Brussevich et al., 2018; Aksoy et al., 2021).

It must be noted that the representation of males and females across various industries differs in terms of occupations, organizational social scales, or roles in the organization of work (Nedelkoska and Quintini, 2018; Mondolo, 2022). Additionally, various analyses of job-related tasks (Brussevich et al., 2018; Aksoy et al., 2021) show that female workers execute fewer assignments requiring interpersonal and analytical skills or physical labor and perform tasks that are characterized by a lack of job variability, with limited opportunities for learning and development. In addition, the respective differences in wages may depend on the fact that, in many countries, women are under-represented in higher-level positions (e.g. in executive-level positions), which influences their possibility of acquiring higher wages. Considering the aforementioned facts, analyzing the relationship between automation and increases in the gender pay gap is a topical issue for further research. Despite the wide discussion about the adoption of technologies in general and automation in particular, there are still limited evidence and explanations concerning the links between the implementation of automation technologies and the gender pay gap.

In Section 2 of the paper, we present the literature review related to the topic. In Section 3, we illustrate the data and the models used. In Section 4, we show the results and their interpretation. Section 5 summarizes the authors’ conclusions.

2. Literature review

The development of technologies and innovation brings automation to different jobs and areas of modern business processes. Implemented technologies may vary, from robotics to various artificial intelligence applications (Calvino and Virgillito, 2018; Mondolo, 2022), and find use in a broad range of economic sectors. Previous research underlines that the current speed of automation and robotization might bring widespread effects and changes in terms of job displacement, reallocation, and polarization (Frey and Osborne, 2017; Calvino et al., 2018; Calvino and Virgillito, 2018; Bessen et al., 2019; Fan et al., 2021; Mondolo, 2022). As such, the possibility that automation will change industries, displace workers, and transform the labor market is a topic of much current research (Acemoglu and Autor, 2011; Benzell et al., 2015; Acemoglu and Restrepo, 2018, 2019; Mondolo, 2022). Theories reflect evidence that automation can lead to the displacement of workers from jobs when newly implemented technologies demand a different set of skills from what was required before (Acemoglu and Restrepo, 2018). At the same time, empirical literature states that large-scale automation should not cause the displacement of occupations but rather a reallocation of labor to new emerging occupations and industries (Autor and Salomons, 2018). In general, research has found that automation likely causes an increase in employment in the respective industries if industry demand is sufficiently elastic (Acemoglu and Restrepo, 2018; Bessen, 2018).

The different theoretical frameworks related to different aspects of automation can be divided into three areas of analysis. The first strand of the literature stresses the relationships between automation and worker skills. As Mondolo (2022) explains, this approach considers how new technologies affect low-skilled and high-skilled workers differently. The effects of automation on the wages of high-skilled and low-skilled workers and, thereby, on the evolution of wage inequality are analyzed by Lankisch et al. (2017) and Bessen (2016). The following tendency is taken into account: despite economic growth in developed countries over the past decades, the median real wage stagnated and the real wages of low-skilled workers have even decreased since the 1970s (Acemoglu and Autor, 2012; Murray, 2016). Concurrently, the wages of high-skilled workers with a degree have grown, revealing an increase in skill premium and a higher dispersion of wages in general. This rise in wage-related inequality is one of the driving incentives behind the rise in overall income inequality observed since the 1980s (Piketty and Saez, 2003; Milanovic, 2016). The active development of international trade and outsourcing has also supplemented skill-biased technological change in its effect on the wage differential (Autor et al., 2016). However, this first theoretical framework has limitations; it is not able to explain the increasing disappearance of middle-skill occupations (Barbieri et al., 2020). Nor does it thoroughly examine the skills used in different occupational tasks and instead focuses on educational levels.

The second theoretical framework focuses more on the tasks that the workers employ internally in the different firms and sectors. A baseline division of tasks refers to routine manual tasks, non-routine manual tasks, routine cognitive tasks, and non-routine cognitive tasks. This approach is based on the routine-biased technological change hypothesis (Mondolo, 2022), wherein the main idea is that automation can replace routine tasks more easily. The limits of this approach can be identified in the difficulties in determining a precise classification of routine and non-routine tasks. Moreover, the routine-biased technological change hypothesis does not consider factors such as institutional structures and work organization. A further part of this second theoretical framework can be found in Acemoglu and Restrepo (2020), and Acemoglu and Autor (2011), where the process of robotization is observed more carefully and an analysis of the complex relationship between humans and machines is included. It is further important to note that some technologies can be more labor augmenting/labor friendly than others (Staccioli and Virgillito, 2021) and bring a more uniform increase in workers’ productivity (Mondolo, 2022). These last considerations are the bases for the third approach in the literature that refers to the evolutionary theory of technical change and the capability-based theory of the firm (Mondolo, 2022). These theoretical frameworks study the relationship between technology, workers, and the organization of their routines inside the firms, as the dichotomies between skilled and unskilled, and routine and non-routine tasks cannot completely identify the transformations due to automation (Cirillo et al., 2021). Indeed, as Cetrulo et al. (2020) and Cirillo et al. (2021) explain, the hierarchy and the degree of autonomy of workers can influence how technologies affect the workers and how the tasks are assigned.

Despite economic literature widely discussing topical technological changes, the majority of studies examine automation from a narrow framework. Literature on automation primarily focuses on the introduction of industrial robots and enlarging robotization, as well as the potential effect this may have on the labor market, which provides limited evidence of the effects of other forms of automation at aggregated levels (Acemoglu and Restrepo, 2018; Graetz and Michaels, 2018; Bonfiglioli et al., 2020). Concurrently, the studies deliver mixed results. Acemoglu et al. (2022) conclude that wages and employment have decreased in the US regions exposed to automation by robots, while Dauth et al. (2018), relying on the empirical design by Acemoglu et al. (2022), find evidence of a positive effect on wages and the absence of changes in total employment in regions of Germany, which is also supported by Graetz and Michaels (2018) when analyzing a panel of countries and the various industries affected. Using French firm-level data, Bonfiglioli et al. (2020) find that robot adoption and employment growth are positively correlated and, at the same time, an increase in robotization intensity is followed by job losses, especially for those who are involved in production. Other recent papers show that even when there is no obvious effect from robotization, employment losses in some sectors may be hidden by the offset of employment gains in others (Bessen et al., 2019). The effects of newly introduced technologies and automation may differ across the different categories of workers. This can cause, for example, some categories of workers (young professionals, females) to change their specialization (i.e. the field of education) and to start looking for other positions or in other industries in case, employment is affected by the introduction of robots (Bessen et al., 2019).

Automation can be related to the innovation process in a company. Calvino and Virgillito (2018) suggest that the positive effects of innovation on employment can vary depending on firms and industry. In terms of product innovation, Dosi and Mohnen (2019), Calvino and Virgillito (2018), and Ugur et al. (2018) agree on its positive effects on employment. Several theoretical contributions are relevant for explaining the channels through which automation, being part of process innovation, affects wage inequality in general or the gender wage gap in particular. Apart from the evidence of direct discrimination (Becker, 1957), the gender wage gap differences between firms introducing innovations and automation and non-innovators can be due to the segregation of males and females in the labor market and the firm or due to productivity differences of otherwise similar employees in a firm. The structural effects can affect the aggregate gender pay gap through changes in demand for certain skills (such as skills that are complementary to the introduction of automation and technological change), tasks or occupations (Acemoglu, 2002), or the organization of work in the firms (Calvino et al., 2018; Calvino and Virgillito, 2018). Recent evidence on automation in the labor market suggests that automation is associated with increasing gender gaps in the labor market (Brussevich et al., 2019), including a higher gender-based wage gap in Europe (Aksoy et al., 2021). These effects have been discussed as a result of the different distributions of routine and non-routine tasks among jobs for males and females (Brussevich et al., 2019; Aksoy et al., 2021). Women tend to look for more routine-based jobs that are, in this sense, more replaceable by automation (Brussevich et al., 2019; Mondolo, 2022). Automation and robotization decrease the relative demand for services in those labor groups that engage more in routine tasks (Blanas et al., 2019; Brussevich et al., 2019). As Montobbio et al. (2021) show, automation can affect different occupations differently.

The earlier discussion shows that analyzing the impact of automation and robotization on the gender pay gap is fundamentally essential. The gains women received due to the introduction of policies aiming to enlarge the number of women present in the paid workforce, along with corresponding equal remuneration, can deteriorate if the process of automation disadvantages women (Brussevich et al., 2018; Aksoy et al., 2021). Such disadvantages could include a lack of specific knowledge needed for certain types of automation, the relatively lower presence of women in occupations or sectors where automation positively affects salaries and country-specific issues.

When discussing labor market dynamics, the issue of the gender pay gap requires special attention. Despite the considerable narrowing of the gender wage gap in developed countries within recent decades, a significant gap remains and is a relevant topic for analysis and has policy implications (Goldin, 2014; Kunze, 2018). It should also be mentioned that several studies pay attention to supply-side explanations, such as gender variances in human capital, psychological characteristics, negotiation behavior, or occupations (Aksoy et al., 2021; Blau and Kahn, 2017; 2000). At the same time, demand-side factors (such as automation) lack scientific discussion and the provision of evidence regarding the effects on the pay gap (Ngai and Petrongolo, 2017). When discussing the respective demand-side factors, only a few papers mention the effects of computerization on gender, indicating that the increase in computer use contributes to narrowing the gender pay gap (Weinberg, 2000; Bessen et al., 2019). Regarding the impact of automation on the gendered labor market, Brussevich et al. (2018) explore the fact that female workers are at a significantly higher risk for displacement or biased attitudes induced by automation than male workers. There is also an indication that the probability of automation having consequences is lower for younger cohorts of women and those in managerial positions (Aksoy et al., 2021). Furthermore, recent data from the US indicate that automation and robotization may have lowered the gender gap in labor force participation and pay (Annelli et al., 2019; Ge and Zhou, 2020). Hence, the presented literature overview provides sufficient evidence that automation and its effects on the gender pay gap are important subjects of scientific discussion. Nevertheless, direct empirical evidence on the impact of automation on workers at the firm level is still scarce, though growing.

3. Data and methodology

3.1. Data and descriptive statistics

In our analysis, we use the imports of automation goods as a proxy for the introduction of automation at the firm level. Following Domini et al. (2022), as well as Acemoglu et al. (2022), we first define the harmonized system (HS) codes (Table 1) related to automation and use these to calculate the value of imports of automation goods among all imports of Estonian firms using data from the firm-product-market level exports and imports dataset elaborated in Masso and Vahter (2019) and Masso et al. (2015). The data on firm-level imports are taken from the international goods trade dataset of Statistics Estonia.² The term “automation goods” includes industrial robots, numerically controlled machines, automatic machine tools, and other automatic machines. As such, purchases can be counted as an investment in tangible assets, and the advantage of this approach is the availability of information on automation over a long period.

The disadvantage of this approach is that it may allow for missing instances of automation when such goods are purchased from companies within Estonia³ or are simply not imported. As mentioned earlier, introducing automation and technology transfer from abroad has been the main source of technological catch-up for Estonian companies (Kalvet, 2004). Other relevant firm-level variables not available in the goods and services imports and exports datasets are taken from the Estonian Commercial Registry (Äriregister) data on annual financial reports.

In addition, we merge the above-mentioned datasets with individual-level wage information via the Estonian Tax and Customs Office dataset on individual monthly payroll tax payments. Social security tax in Estonia is applied to all employees at 33% of the gross wage; thereby, tax payments allow us to identify an employee’s gross wage and employment status for each firm, every year. Additionally, information on the gender and age of workers is provided in the database. Other characteristics (e.g. education and occupation) are obtained from the Population and Housing Census 2011 data, Structure of Earnings Surveys from 2014 and 2018, and the Estonian Population Registry data for 2010–2020 (on education). Thus, to investigate the determinants of the gender wage gap, we combine these firm and individual-level datasets and create a matched employer–employee dataset. The datasets are then merged through the company registration numbers and anonymized personal identification numbers. The wages are transformed into real wages and are converted from Estonian kroon to the euro for the period before Estonia entered the Eurozone (before 2011). Wages are subsequently considered in their logarithmic transformation. Previous studies on the gender wage gap in Estonia have used different datasets: the labor force survey (LFS) data (Anspal, 2015 (a, b); Krillo and Masso 2010), the online job search portal CV Keskus dataset (Meriküll and Mõtsmees, 2017), Programme for the International Assessment of Adult Competencies (PIAAC) data (Tverdostup and Paas, 2016, 2017), and linked employer–employee data (Masso et al., 2022) to provide proof of consistently high gender wage gaps. So far, the principal reasons identified for the gender wage gap in Estonia were the employment industry and the employee’s position (occupation).

In Table 2, we illustrate the dynamics of the share of firms that import automation goods compared to the total number of firms that import. In contrast, Table 3 presents more detailed results for different groups of firms and various automation goods. It can be noticed that the percentage of firms importing automation goods over firms that import in Estonia is stable and around 20% for the period of 2006–2018. The fact that there is no evidence of an increase in the imports of automation goods in the studied period between 2006 and 2018 may seem surprising, given the relatively low share of firms importing automation goods. For instance, Bessen et al. (2019) found that the firm-level automation costs per employee increased over time in the Netherlands, especially in the 95^th percentile of the indicator. We calculated imports of automation goods per employee for our data, but, similar to the other indicators, no growth trend can be noted, which could be regarded as somewhat puzzling. It may raise the question of whether the automation goods are purchased through intra-country business-to-business transactions instead of being imported. However, previous studies have also pointed out the relatively low levels of automation and robotization in Estonia using survey data (Azzopardi et al., 2020).

However, even if the number of firms (companies) that import automation goods is relatively small, 1.8% of all the economically active firms, their turnover, and employment shares in all active firms in Estonia are relevant.

The firms that imported automation goods represent a stable 25% share of employment across the total number of active firms The percentage of firms that imported automation goods is much higher in manufacturing (6.9%), especially for high and medium high-tech manufacturing (16.5% and 15%, respectively). In the services sector, the share of firms that have imported automation goods is higher in less knowledge-intensive services than in knowledge-intensive services (0.5 and 2.6%). Figure 1 shows the top automation goods imported to Estonia in the period under consideration. Electronic integrated circuits and hand tools of base metals⁴ are the most common automation goods imported into Estonia. Processors and controllers and parts and accessories for machine tools, machines, apparatuses, and mechanical appliances are also relevant. However, the dynamics have changed in recent years; for example, hand tools have lost importance, while electronic integrated circuits and machines, apparatuses, and mechanical appliances are gaining relevance.

Previous studies on automation have established that automation expenditures and investments display the typical spiky behavior of investment variables (Bessen et al., 2019; Domini et al., 2021). Thus, we have replicated the basic statistics calculated by Domini et al. (2021). First, automation is relatively rare across firms in Estonian data, as not more than 20% of importers annually import automation goods (Domini et al., 2021; approximately 15%, in the case of French manufacturing firms). Second, automation is also relatively rare within firms—among the firms that imported automation goods, 43.7% only did so once (Domini et al., 2021; approximately 30%). The spike events—the years with the most significant imports of automation goods—constituted a relatively high share of the total import of automation goods over the studied period, 74.1% (Domini et al., 2021; approximately 70%). Thus, we can firmly state that the findings on the spikiness of the automation expenditures also hold in the Estonian firm-level data.

Turning next to the association between automation and the gender wage gap, Figure 2 shows the Kernel density distributions of average wages for males and females in different firms with and without an automation variable. The lower level of wages for females compared to males can be observed more or less throughout the wage distribution (upper left graph in Figure 2) and does not only relate to the differences in average or median wages. Our proxy for firms with automation is, at the same time, clearly associated with higher wages among both males and females. When looking at the data for all sectors (upper left graph in Figure 2), higher wages in firms with automation can be observed especially among males, while the differences in wages between firms with and without the import of automation goods are smaller among females. The latter findings seem to indicate a higher gender pay gap in firms introducing automation, and, indeed, such evidence is apparent in the bottom right graph in Figure 2, where the gender pay gap is larger among firms importing automation goods. When looking at manufacturing industries only (bottom left in Figure 2), the positive association between the wage level and automation is even more visible, and in manufacturing, average female wages are positively associated with automation. Finally, the upper right graph in Figure 2 shows that the previously described wage differences are also visible (albeit smaller) when the variable studied is instead the firms’ average wages minus the 2-digit industry average wage level. These findings suggest the need for an investigation into the issue of wage gaps and automation in econometric analysis.

3.2. Methodology

First, we estimated the Mincerian equation taking into account a female dummy, an import dummy, an automation dummy, the interaction of the female dummy with imports and automation, and other regressors of potential changes in wages (first column of Table 4). Mincerian regressions have been used frequently in similar studies on the firm-level determinants of the gender wage gap, such as by Vannutelli et al. (2022), Damiani et al. (2020), and Masso and Vahter (2020). The goal was to examine the effects of automation goods on wages at the individual level. The Mincerian equation is as follows:

where |$\ln {W_{i,j,t}}$| is the logarithm of the real wage of employee i, in firm j, at time t. |${\rm{Femal}}{{\rm{e}}_{i,t}}$| identifies if individual i, in firm j, is a female. |${\rm{Im}}{{\rm{p}}_{i,t}}$| is the dummy that describes whether firm j, at time t, is importing capital goods, while |${\rm{AutoIm}}{{\rm{p}}_{i,t}}$| signals whether firm j, at time t, is importing automation capital goods. Age_i,_t is the age of the different individuals, R_i,t represents a vector of other time-variant individual-level control variables, and |${Z_{j,t}}$| is a vector of firm-level control variables.⁵  |${\lambda _t}$| is the vector of dummies for different years of the sample, and ν_i are firm-fixed effects. The |${e_{i,j,t}}$| figure is the error term with zero mean and constant variance assumed to be normally distributed. Imports were included in the list of explanatory variables, whereas previous studies have shown that company internationalization, in the form of either exports or foreign direct investment (FDI), may increase the gender pay gap (Vahter and Masso, 2019). Thus, it seemed to be crucial to control for that in case the automation expenditures are proxied by the imports of the automation goods.

When estimating the effects of automation on the gender pay gap, consideration that automation is just a particular kind of process innovation is required. Using the CIS data, Masso and Vahter (2020) showed that innovation is associated with a higher gender pay gap across various innovation output and input indicators. Therefore, to measure the possible effects of innovation, we introduce different kinds of innovation variables and their interactions with the female dummy (Table 5, excluding model 1), including whether the significance of the automation dummy disappears or persists with the presence of the other innovation indicators in the wage equation. While the default dataset covers the whole population, the innovation variables are available for the companies included in the CIS, approximately 1500 companies in each wave. These models with an extended set of explanatory variables can be described as an extension of equation 1 as follows:

where |${\rm{Inno}}{{\rm{v}}_{j,t}}$| is the innovation variable considered in the different models. The estimations are repeated for three different periods (2006–2009, 2010–2013, and 2014–2018) to observe the evolution of the effect of automation on the gender pay gap over time. Finally, the paper considers employee wages for different 1-digit International Standard Classification of Occupations (ISCO) occupations. The latter estimations are made due to the findings of Aksoy et al. (2021), who describe how higher-level occupations should reduce the gender pay gap. The estimations are the same as in equation 1 but for individual wages in different occupations. Table A1 presents the descriptive statistics of the variables used in the regression analysis.

As the second stage of the analysis, the PSM is estimated (Rosenbaum and Rubin, 1983). This paper also checks for the effect of automation on employee wages in the different firms after importing automation goods compared to the same in firms that did not acquire automation goods—which is the counterfactual. The treatment variable “automation” is a dummy variable that assumes the value of “1” in the treatment period, and the unit of analysis is the firm.

A probit model for importing automation goods⁶ is estimated in the first step of the PSM estimation. The independent variables in the probit model are measured 1 year before the acquisition of automation technologies. The list of control variables included follows the literature (e.g. Masso and Vahter, 2019 in the study of the effects of foreign acquisitions on the gender pay gap): firm size, firm size squared, firm age, firm age squared, liquidity ratio, log of capital intensity (considering the possible importance of the firm’s financial conditions and existing capital intensity regarding decisions on automation), and the location of the company in northern Estonia (the capital region). All the mentioned controls are calculated from one year before the treatment (introduction of automation). The control variables also include the lagged values of the outcome variables, log real wages, and log real wages squared.

The probit models summarize the information from various control variables affecting automation to derive the propensity score. Using the propensity score, each firm j, is matched with the five best counterfactual firms, i.e. using the nearest neighbor matching with five neighbors (the standard approach in the literature). As a robustness check, an evaluation was included of the nearest neighbor matching with two neighbors that gave similar results. Finally, the average treatment effect on treated (ATT) firms is estimated to derive the effects of automation on all employee wages, as well as male and female wages. The ATT can be described as follows:

where |${\rm{ATT}}_{{\rm{PSM}}}^s$|is the ATT companies at the period over years s, considering the PSM. The figure |${\Delta ^s}{\pi}_{t + s}^{{\rm{treated}}}$| is the mean growth of the average wage at the firm level for the treated firms at time t + s, and the second term represents the mean growth of the average wage in the control group.

The time of the automation event is defined with t. Thus, the probit model is estimated with the control variables measured at time t − 1, and the ATT effects of the acquisition of automation technologies are estimated at time t, t + 1, and t + 2. The outcome variables examined in the firm-level estimations are, thus, the firm’s average wage, the average wage of male employees, and the average wage of female employees. Furthermore, we also estimated the same outcome variables considering different levels of education, the share of females, the share of females among managers, and the number of employees (with or without tertiary education) as supplementary outcome variables.⁷ This allows us to check the effects of the import of automation goods on the workforce structure and its eventual contribution to wages. To understand which automation goods have more impact on wages, the estimations are repeated with all the subcategories of automation goods that could be derived from the datasets.

4. Results and discussion

4.1. Mincerian wage regressions

Table 4 presents the results from the estimation of the Mincerian wage regressions for the period of 2006–2018. The estimated wage equations included, in addition to the automation variable, different innovation indicators from the CIS, whereas automation can be considered a particular kind of process innovation. While the baseline model (model 1) includes only the automation in the firm (import of automation-intensive goods), the next models sequentially introduce process innovation (model 2), product innovation (model 3), organizational innovation (model 4), and marketing innovation (model 5). Model 6 replicates model 1 on a smaller sample in which the innovation variable is available as a robustness check. As expected, the female dummy has a strong negative association with wages that is statistically significant in all specifications at the 1% level. It ranges from −0.168 (model 1) to −0.230 (model 4). These numbers broadly correspond to the size of the conditional gender pay gap estimated in earlier studies in Estonia using various datasets (e.g. Masso and Vahter, 2020). In all the estimated regressions, the female dummy variable is similar to the value observed in earlier Estonian datasets that employed matched employer–employee data (Masso and Vahter, 2019, 2020).⁸ Considering that the dependent variable is measured in natural logs, by applying exponential transformation of the coefficients, the baseline model (model 1) reports that the conditional gender pay gap as the wage level of females minus the wage level of males divided by the wage level of males is 18.3%.

The dummy variable for firm-level imports shows a positive and statistically significant association with wages ranging from 3.5% (model 6) to 6.6% (model 1), demonstrating the positive effects of internationalization on wages. It is a natural control variable for wages in our model, whereas the imports of particular goods are used as a proxy for automation. The interaction terms between the female dummy and the import dummy are statistically significant at the 1% level and also economically significant with an estimated size of −3.3% to −8.6%, indicating from another angle in addition to FDI (Vahter and Masso, 2019) and exports (Masso and Vahter, 2020) the importance of internationalization for the gender wage gap.

Based on the respective coefficients and interactions, males earn 1.6% more when working in firms that introduce automation in their operations (compared to firms that do not introduce automation). At the same time, female employees working in firms that import automation-intensive goods earn just 0.3% more, i.e. their gain is less by 1.3 percentage points. It can be noticed that introducing innovation variables from the CIS into the model alters the magnitude of the variables’ coefficients of interest compared to the baseline model, yet these always remain statistically significant. For example, in the model with technological innovation, automation is associated with 1% higher wages for males, yet no gain for females due to the interaction term is −2%. In model 2, the specification with the process innovation, male wages increase by 1% and female wages decrease by −0.3%. In the model with product innovation (model 3), the increase in male wages is equal to 1.2% and the decrease in female wages is equal to −1%, with the difference being equal to 2.2%. In model 4, with organizational innovation, male workers gain 1.1% more and female workers −1% (the difference being 2.1%), while in model 5, with marketing innovation, male wages increase by 1.8% and female wages −0.5% (the difference is equal to 2.3%). Finally, the last specification, model 6, replicates model 1 on the smaller sample in which the innovation variables are available to show robust results.

Considering that the introduction of automation may take time to change individual wages, Table 5 presents estimations with lagged automation variables. The coefficients of the imports are stable across the estimations, with imports being associated with approximately 6% higher wages. At the same time, the automation terms lagged 1–3 years have a much stronger association with the wages, e.g. 5.2% at lag 3, as compared to 1.6% in the baseline model. The interaction term between females and automation is also larger in the case of the lagged automation variable, especially at lag 3, with females having a −3.6% lower wage gain due to automation compared to males. Finally, the last column of the table indicates that the qualitative results are the same when using a continuous variable instead of the automation indicator variable, that is, the log automation costs per employee.

Table 6 presents the results from the stepwise estimation of the Mincerian wage regressions as one of the further robustness checks. The aim was to observe if only using gender and firm type as predictors of wages would preserve similar results as in the baseline model (model 4 in Table 6). The results show that the female dummy changes from −12.5% in model 1 (including the automation and import variable) to 18.3% in model 4 (including all individual and firm-level variables but excluding occupation). The estimated coefficients show that, according to first step estimation, when only considering automation, males gain 6.4% more when working in firms that introduce automation, while females gain −0.7%. While adding stepwise further controls variables, the positive association of automation with wages is reduced yet remains always positive (e.g. 2.6% in model 6 with the most extended list of control variables). At the same time, across all specifications, females gain significantly less from automation, bringing their net gain close to zero.

The additional sensitivity of the results can be considered when splitting the sample of observations by educational level, people with a tertiary education vs those with non-tertiary education, and age groups of employees and performing the respective estimations separately for these groups (Table 7). Regarding employees with tertiary education, males gain a 1.5% wage premium when working in firms that introduce automation. In comparison, females gain −0.8%, amounting to a difference in the automation effects on wages between males and females of 2.3%. For employees with non-tertiary education, males gain 2%, and females −0.5%. Concerning the different age groups, the gain of the males from automation is higher in the prime age group and among older employees (and absent until the age of 30 years), but for females, it is very close to zero in all studied age groups. All these indicate that the results are relatively stable across the different groups of employees.

To observe the effects of automation on the gender pay gap throughout the period of analysis in a similar way as in Fan et al. (2021), we conducted additional estimations for different sub-periods: from 2006 to 2009, from 2010 to 2013, and from 2014 to 2018, using the same set of control variables (Table 8). The results indicate that the positive effect of automation on the wages of males is decreasing over time and the effect on the wages of females changes from positive to negative. There is no evidence of an increasing contribution to the gender pay gap over time, which might be somewhat surprising given the general evidence, from Estonia in particular, on the growing importance of firm-level factors for the gender pay gap (Domini et al., 2022; Masso et al., 2022). On the other hand, this result is in line with the long-term downward trend of the pay gap over the last 30 years in Estonia (Meriküll and Tverdostup, 2020).

Following equation 1, we also undertake estimations across various occupation groups as defined at the ISCO 1-digit level from managerial positions to elementary occupations (Table 9).⁹ The estimations indicate the importance of considering the heterogeneity of the effects across the broad occupational groups. While the automation dummy was found to be positive in all previous estimations, it is statistically insignificant in some occupational groups. The highest positive effects can be seen in the case of professionals, +5.8%, which is not unexpected. The interaction terms between females and automation are also statistically significant and negative in most cases, except in the case of professionals, indicating that female professionals may gain 3% more from automation than males. The latter may suggest that female professionals may have complementary skills to automation. The gender–automation interaction term being the most negative in the case of the managers is analogous to the negative effect of company internationalization on female managers due to higher commitment requirements when working in an international environment that may especially hurt females (Vahter and Masso, 2019).

The most evident results of the analysis across the occupations show how different positions and occupations play a significant role in defining whether the gender pay gap increases or decreases and to what magnitude. Different causes may explain the increase in the gender pay gap in some occupations: (i) the tasks given to the female managers by the work organization are more sensitive to automation, (ii) specific positions chosen by female workers are more replaceable, and (iii) the competencies of the female workers resent of a different education received by the education system. The first cause can be related to the findings of Cetrulo et al. (2020), where routine and non-routine differentiation fails to identify the use of knowledge in both the two typologies of tasks, and where instead, issues such as the degree of autonomy, teamwork, and social interactions can play a role independent of the positions’ titles. The second cause relies not just on the occupation itself but also on the tasks assigned to the specific position. Nedelkoska and Quintini (2018) suggest that female workers often do more automatable tasks than their male workers, even if the occupation is the same. For example, it could happen that a female manager is working on specific tasks that are easily automated. After automation, these employees may lose the contractual bargaining power and the reallocation into new positions less automated or where automation is labor augmenting is more difficult due to family duties that are still not sufficiently shared by male workers (Mondolo, 2022).

Table 10 shows the results of the wage regressions in the case of different kinds of automation goods. As can be observed, the effects of automation are almost all negative but differ in magnitude and significance. Certain variation in effects is expected, whereas technologies can be either labor-augmenting or labor-saving (Staccioli and Virgillito, 2021). Hence, different machines may impact the wages of female workers differently, especially if considering that females tend not to reallocate toward positions with less replaceable tasks (Mondolo, 2022).

4.2. PSM results

In the following step, we use the PSM method to view the possible effects on the wages of males and females in firms that introduced automation to explore the possible effect on the gender pay gap. Thereby, firms that did not introduce automation are used as the control group. In addition to performing the impact evaluation for the general automation variable (i.e. imports of any automation goods), it is also performed for firms that have introduced various types of automation (e.g. the introduction of industrial robots, welding machines, and others). Given that automation may be regarded as one kind of process innovation, estimations are also run for the effects of process innovation from the CIS survey for comparison purposes. In the first step for estimating the propensity score for introducing various kinds of automation, probit models were evaluated, with the dependent variable (treatment variable) being equal to 1 if the firm without automation at t − 1 introduced automation at time t. After composing the propensity score, the control group for both males and females is compiled separately. The ATT is estimated based on equation 3. When checking for the matching quality, the t-test related to the matching showed good results, meaning that the control variables are not statistically different in the case of the treatment and control groups after the matching (with just a few exceptions in some estimations; the results are available under request). It is worthwhile to stress (once more) that the use of foreign trade (imports of automation goods) to proxy automation means that the counterfactual is not non-automation but the non-import of automation goods. Suppose that some of the firms in the control group had acquired automation goods via inter-country transactions. In that case, we have probably underestimated the true effects of automation on the studied outcome variables.

Table 11 presents the obtained ATT estimates on the entire economy for males and females 2 years after the actual introduction of automation in a firm.¹⁰ Across the different kinds of automation, the positive effect on wages is, in most cases, larger in the case of males compared to females. For overall automation, the positive effect on the wages of males is 7.1 percentage points and for females, 6.2 percentage points. These effects are larger compared to the effect of the process innovation variable from the CIS, which is only 2.1 and 2.2 percentage points and both estimates are statistically insignificant. Concerning the variation of effects across various automation goods, the largest positive effects can be seen from the introduction of conveyors, 19.1 percentage points for males and 15.4 for females. Rather, large effects can also be seen from the introduction of dedicated machinery, +14.2% for males and 10.3% for females, and welding machines, 13.0% for males and 8.9% for males. Nevertheless, some of the effects are statistically insignificant because the number of treatments is quite small (e.g. for industrial robots, there are only 25 treatments available for evaluation—generally, based on experience, around 50 can be considered as a minimum to have robust results). As one robustness check, we considered whether the size of the automation expenditure matters for the size of the effect. In particular, we conducted estimations separately for the value of imported automation goods per employee with a value greater or less than 5000 euros per employee. The results indicate expectedly that the effects are much larger in the first case, 10.1% and 5% wage gains, respectively. The effect on males exceeding that on females is visible only in the case of the larger automation investment per employee.

Table A2 shows the results of the supplementary outcome variables for the total economy to understand the aforementioned effects further. For instance, some of the effects on the average wage level of the company’s employees were found to exceed the effects on either males or females. However, this might be explained by the increasing share of male employees following automation. Indeed, almost all the automation variables affect the percentage of female employees negatively. In the case of automation in general, the effect is −3.7%. As an exception to that, the introduction of other dedicated machinery and weaving and knitting machines had positive effects on the female share, 3.4% and 9.9%, respectively. The impact on the percentage of female managers follows similar patterns. Negative and significant effects on the share of female managers are present with any automation goods (−2.4%), conveyors (11.3), dedicated machinery including robots (−8.5%), regulating instruments (−9.5%), and welding machines (8.1%). Positive effects can be observed with industrial robots (14.7%), other dedicated machinery (6.7%), and weaving and knit machines (16%). The effects on the number of employees at the company are statistically insignificant in the case of automation in general, but the number of employees is negatively affected by other dedicated machinery (−14.4%) and especially by weaving and knit machines (45.7%). The number of employees with tertiary education is positively and significantly affected by general automation dummy (5.4%), conveyors (22.3%), dedicated machinery including robots (13.4%), regulating instruments (0.9%), tools for industrial work (8.7%), and welding machines (10.2%). The positive effects on the number of employees with tertiary education can benefit females, given their higher level of education (as is the case in many countries).

5. Conclusions

Estonia has had, for many years, the highest gender pay gap among the EU countries. However, in the past decade, the gender pay gap has decreased in Estonia, similar to other developed countries (Meriküll and Tverdostup, 2020), yet the relative importance of firm-level factors has increased. This paper analyzed Estonian data from 2006 to 2018 to study the effects of automation on the gender pay gap. The regression analysis showed a negative and significant association of automation with the gender pay gap, implying that automation has an effect contrary to the observed reduction of the gender pay gap. These results were robust even after controlling for the more general technological and non-technological innovation variables in wage regressions. However, by dividing the timeframe into different intervals, the results indicate that the positive effect of automation on the wages of males decreases over time and the effect on the wages of females changes from positive to negative. Previous literature provides evidence that women acquire new skills more compatible with automation faster than men (Blanas et al., 2019). However, our findings are in contrast to studies that illustrate how automation is decreasing the gender pay gap in other countries, such as the USA (Anelli et al., 2019), or having no effects on the gender pay gap, such as in France (Domini et al., 2022). Understanding why the Estonian labor market does not pursue a similar linear path requires more studies and evaluations. One possible explanation can be related to the Estonian education system. Kindsiko et al. (2020) stress how teaching information and communication technology (ICT) skills in pre-university institutions and schools still faces issues in being more interesting for young girls. This can leave females vulnerable to automation through their lack of useful competencies in positions where automation plays a role.

It can be observed that automation may positively affect the gender pay gap, especially in managerial positions. This suggests that being employed in higher positions alone does not assure better salaries for women, as Aksoy et al. (2021) argue, as the effects of automation would still be associated with lower wages for females compared to males. Another novelty of our results is the strong evidence that the impact of automation on wages and the pay gap varies strongly across the different types of automation. All these relate to female workers’ difficulties in reallocating to new, less replaceable positions and focusing on more routine tasks due to the education level obtained (Nedelkoska and Quintini, 2018; Kindsiko et al., 2020; Mondolo, 2022).

When making estimations using the PSM method, it can be observed that male employees receive higher gains and wages than female employees due to automation. This holds across different categories of automation goods, albeit with insignificant estimates in some subcategories due to the small number of treatment units. These results allow us to observe not only the effects of automation on the gender pay gap in different occupations and firms but also how different kinds of automation can affect the gender pay gap in employer–employee data. Indeed, introducing particular types of automation affects the gender pay gap differently, even if they are consistent with the general results. Further studies should try to understand how the combination of different kinds of automation can replace certain occupations rather than others and how this can be connected with the organization of work and the capabilities of male and female workers (Cetrulo et al., 2020; Mondolo, 2022).

Concerning the supplementary outcome variables, we found that the share of female employees and managers was negatively affected by any automation variable except industrial robots and weaving and knitting machines. This would explain the negative effect of automation on the gender pay gap via the lower demand for female labor. More research would be necessary to understand why the allocation of female employees and managers is so unbalanced at firms with automation (Mondolo, 2022).

Naturally, the limitation of all the aforementioned results is that these were obtained by proxying automation with the imports of automation goods and neglecting the possibility that the automation goods are purchased instead from domestic companies. This limitation could be, to some degree, overcome by using the value added tax registry data. Still, given that the latter does not include any information on the goods and services traded, serious limitations would also remain with that approach.

Regarding the policy implications, both men and women must upgrade their skills to be compatible with new technologies (Kindsiko et al., 2020). However, this should be tailored to the occupation, sector, and typology of automation employed by different firms A higher representation of females in higher-paid positions does not guarantee a reduction in the gender pay gap in the presence of automation. Indeed, more appropriate education/training can be necessary for female workers in managerial positions where automation has significant effects. The effects of the different kinds of automation can indicate which competencies female workers can apply to overcome the gender pay gap. Firms aiming to close the gender pay gap should provide these specific forms of training based on the automation needed for their operations. Possible public subsidies to enterprises for favoring this training can be considered. Alternatively, enterprises importing automation goods should consider the effects on relative wages and male and female workers’ current education and skills and allocate their tasks accordingly. Even if this could reduce the gender pay gap, there is a risk of the polarization of female workers in less remunerative positions. There is a need from the part of statistical institutions to classify the characteristics of the occupations more precisely. Relying on simple classifications such as “managers” fails to understand the increased gender pay gap due to automation. In contrast, categories such as the ones of Cetrulo et al. (2020), in which factors such as power, cognitive and manual dexterity, ICT knowledge, creativity, and teamwork are considered, can provide valuable contributions. Moreover, reallocation in different positions and training in the work environments should be encouraged. Firms and public institutions should provide help to support women in need. Lastly, as suggested by Kindsiko et al. (2020), the education system needs to work to help girls increase their interest in ICT skills.

Supplementary Data

Supplementary materials are available at Industrial and Corporate Change online.

Footnotes

References

Appendix