Is Technology Widening the Gender Gap? Automation and the Future of Female Employment
  • 1 https://isni.org/isni/0000000404811396, International Monetary Fund
  • | 2 https://isni.org/isni/0000000404811396, International Monetary Fund

Contributor Notes

Using individual level data on task composition at work for 30 advanced and emerging economies, we find that women, on average, perform more routine tasks than men?tasks that are more prone to automation. To quantify the impact on jobs, we relate data on task composition at work to occupation level estimates of probability of automation, controlling for a rich set of individual characteristics (e.g., education, age, literacy and numeracy skills). Our results indicate that female workers are at a significantly higher risk for displacement by automation than male workers, with 11 percent of the female workforce at high risk of being automated given the current state of technology, albeit with significant cross-country heterogeneity. The probability of automation is lower for younger cohorts of women, and for those in managerial positions.

Abstract

Using individual level data on task composition at work for 30 advanced and emerging economies, we find that women, on average, perform more routine tasks than men?tasks that are more prone to automation. To quantify the impact on jobs, we relate data on task composition at work to occupation level estimates of probability of automation, controlling for a rich set of individual characteristics (e.g., education, age, literacy and numeracy skills). Our results indicate that female workers are at a significantly higher risk for displacement by automation than male workers, with 11 percent of the female workforce at high risk of being automated given the current state of technology, albeit with significant cross-country heterogeneity. The probability of automation is lower for younger cohorts of women, and for those in managerial positions.

I. Introduction

There is growing policy concern about the implications of digitalization, artificial intelligence, and machine-learning for job displacement. Many jobs involving low- and middle-skill routine tasks are already being eliminated through labor-saving automation. Automation is hardly a novel phenomenon; traditional sectors such as agriculture and manufacturing have experienced large substitutions of labor with machine capital in the past. But computerization of white-collar services in advanced economies has accelerated in recent years (Acemoglu and Restrepo, 2018). At the same time, progress in machine learning is further expanding the set of activities that can be performed more efficiently by computers than humans (Brynjolfsson, Mitchell, and Rock, 2018), suggesting a significantly broader scope for task automation over the medium-term. What are the gender implications of this changing nature of work? How vulnerable are women’s jobs to risk of displacement by technology? What are the intergenerational implications of automation?

This paper uses a task-based approach to investigate the exposure to automation, focusing on gender differences in labor market outcomes. We use individual-level data on task composition at work from the OECD’s Programme for the International Assessment of Adult Competencies (PIAAC) for 30 advanced and emerging economies to identify threats and opportunities from technological progress for workers across sectors, occupations, and countries. Following the seminal task-based framework of Autor, Levy, and Murnane (2003), we first document the relative exposure of men and women to routine, abstract/analytical, and manual tasks.

Second, to quantify the potential impact on labor demand, we derive the likelihood of automation for male and female workers using detailed information on worker characteristics and task composition at work. We derive probabilities of automation at the individual level based on worker characteristics, including age, education, gender, literacy and numeracy skills, and a broad subset of task characteristics included in the PIAAC data set. This allows us to evaluate differences in the probability of automation across different demographic groups unlike most of the existing literature that distinguishes workers only by occupation. Specifically, we estimate the proportion of the female working population that is at risk of being displaced by automation given the current state of technology. Our individual-level estimates of probabilities of automatability also allow for a more detailed assessment of gender gaps across sectors, occupations, and age groups.

We find that women, on average, perform more routine or codifiable tasks than men across all sectors and occupations―tasks that are more prone to automation. Moreover, women perform fewer tasks requiring analytical input or abstract thinking (e.g., information-processing skills), where technological change can be complementary to human skills and improve labor productivity. Self-selection by women into specific occupations explains most of these differences, but women’s exposure to routine job tasks varies significantly across countries. Within Europe, women in Eastern and Southern Europe perform more routine tasks than men, while the gender routineness gap (the difference in the routineness index values for women and men) is lowest in Scandinavian and Central European countries. For our sample as a whole, the gender routineness gap is highest in Japan, the Slovak Republic, Singapore, and Estonia.

Our results indicate that, given the current state of technology, 54 million male and female workers in 30 countries (28 OECD member countries, Cyprus, and Singapore) are at a high risk of being displaced by technology (i.e., facing higher than 70 percent likelihood of being automated) within the next two decades.1 A larger proportion of the female workforce is at a high risk for automation than the male workforce (11 percent versus 9 percent across the entire sample of countries), with 26 million female jobs potentially at stake in these countries. However, differences in proportions of women at high risk of automation vary substantially across countries—in Japan, for instance, women are three times more likely than men to fall into a high-risk group. Less well-educated and older female workers (aged 40 and older), and those in clerical, service, and sales positions are disproportionately exposed to automation.

There are some positive trends. Since the 1990s, shifts in the labor force away from clerical and elementary occupations towards service sector and professional jobs has been more pronounced for women than men in our sample. This suggests that women are increasingly selecting into jobs that are more insulated from displacement by technology. A cross-sectional analysis of age cohorts shows that aggregate gender gaps are primarily driven by older cohorts of women. Gender automation gaps between men and women are smaller for younger cohorts even among workers facing the highest risk of automation (e.g., less-well educated, in clerical and sales positions).2

Our paper is related to several strands of literature. First, our paper is related to studies that use US-based routineness estimates at the occupational-level, to show that routine jobs are more susceptible to substitution of labor with information and communication technology (ICT) (Autor, Levy, and Murnane 2003; Goos, Manning, and Salomons, 2014; Das and Hilgenstock, 2018).3 Conversely, workers who perform more analytical tasks, or tasks requiring interpersonal communication (“abstract” tasks) are less likely to be displaced.4 In contrast to these studies, we use individual-level data for a large sample of countries, which allows us to relax two crucial assumptions: (i) workers perform identical tasks within occupations across countries; and (ii) workers have access to the same technologies across countries to perform their tasks. In this respect, our paper is similar to De La Rica and Gortazar (2016) who use the PIAAC data to show that ICT adoption at work explains over 6 percent of the cross-country differences in job deroutinization in 22 OECD countries. However, these studies do not examine the implications of automation and technological progress separately for men and women.

Our paper is also related to studies that use occupational and task-level data to examine how women’s labor market outcomes differ from those of men. Using occupational data from the US, Bacolod and Blum (2010) and Cortes, Jaimovich, and Siu (2018) find that women are more likely than men to be engaged in occupations requiring greater cognitive and interpersonal skills. These studies find that the recent increase in the returns to cognitive and interpersonal skills relative to motor skills has resulted in a narrowing of the gender-wage gap in the US. Black and Spitz-Oener (2010) show that relative increases in non-routine analytic or interactive task performed by women explains nearly 50 percent of the gender wage convergence in Germany5. Beaudry and Lewis (2014) show that the decline of the gender wage gap in the US is coincident and causally driven by the rate of computer adoption in local labor markets, where technology adoption increases the returns to brains relative to brawn. Much of this literature is limited to single country case-studies, predominantly the US, owing to the lack of standardized data across multiple countries on occupational characteristics.

Finally, our paper is related to studies that quantify gender differences on potential job losses in the future. The World Economic Forum’s 2016 report on The Future of Jobs uses perceptions-based surveys of Chief Human Resource Officers of leading employers globally to identify workforce disruption by industry, country and industry-gender gaps. Using 2022 projections on occupations’ growth from data for the US, the report concludes that, worldwide, “men will face nearly 4 million job losses and 1.4 million gains…, whereas women will face 3 million job losses but only 0.55 million gains— more than five jobs lost for every job gained” (WEF, 2016). The International Labour Organization employs a similar methodology of stakeholder surveys to predict workforce disruption in ASEAN economies (ILO, 2016). One of the main drawbacks of these studies is their application of a universal index of how likely an occupation will be automated for both men and women. The WEF report applies predictions based on a US survey of occupations to a broad set of countries. However, men and women can perform different sets of tasks even within the same occupation. Further, existing evidence suggests that occupational tasks vary non-trivially across countries.

The remainder of the paper is organized as follows: Section II outlines our measure of routineness and documents gender gaps in routineness across countries, occupations and sectors. Section III presents estimates for the probability of automation and evaluates gender gaps in exposure to automation. Section IV elaborates on intergenerational differences in the gender automation gap, highlighting differences in educational attainment and occupational and sector differences. Section V concludes.

II. Measuring Exposure to Routinization: The RTI Index

A. Measuring Routineness

The standard measure of job routineness―an index of routine task intensity (RTI)―developed by Autor, Levy, and Murnane (2003) quantifies the extent of codifiability of tasks performed on the job and serves as a proxy for substitutability of workers and machines. Jobs with a higher share of tasks that can be performed by following a defined set of rules, and are thus easily codifiable, score higher on the RTI index. Jobs requiring analytical, communicational, and technical skills score low on the RTI index. The construction of the RTI index therefore requires detailed information on task composition at work. To date, the literature on routinization has predominantly relied on the Dictionary of Occupational Titles (DOT) data produced by the US Department of Labor for occupation in the US, limiting the scope for cross-country comparisons.

We use the OECD’s PIAAC database to develop a new cross-country index of job routineness from task composition at work. The PIAAC survey covers adults 16 to 65 years of age and collects detailed information on task composition, task frequency, and extent of ICT use in the workplace for 28 OECD member countries, as well as Cyprus and Singapore.6 In addition, the survey contains demographic information and measures of literacy, numeracy, and problem-solving skills for each respondent.

To construct the routine task intensity (RTI) index, we follow the method outlined in Autor and Dorn (2013), modified by De La Rica and Gortazar (2016), to match the content of the PIAAC survey. We calculate the RTI index at the individual-level using task composition at work of each survey respondent. The RTI index evaluates the relative importance of abstract skills, such as reasoning and interpersonal communication and non-routine manual skills against routine tasks which can be easily automated. Specifically, we calculate RTI for each individual worker i as follows:

A01eq1

in which Routinei, Abstracti, and Manuali are index values of routine, abstract, and non-routine manual skills, respectively. The RTI index ranges from zero to one, with higher values indicating that a worker engages in more routine activities.

Table 1 lists the variables used for construction of the RTI index. Most country questionnaire items contain five responses indicating the frequency at which a task is performed: never, less than once a month, less than once a week but at least once a month, at least once a week but not every day, and every day. The “abstract” component consists of analytical and interpersonal tasks (e.g., writing reports, solving complex problems, and negotiating with people). We consider two types of manual tasks. These include routine tasks involving hand and finger dexterity and nonroutine physical work associated with caregiving and operating construction-related equipment. Since the PIAAC survey provides information on only two types of manual tasks, we classify them as routine and non-routine based on previous work by Autor, Levy, and Murnane (2003) and De La Rica and Gortazar (2016). The “routine” component consists of limited flexibility and limited on-the-job learning, low values of which indicate the repetitive nature of work and tasks that can be easily codified.

Table 1.

Questionnaire Items Used to Construct RTI

article image
Sources: PIAAC survey.

We perform a principal component analysis (PCA) to derive an index for each RTI component.7 For the routine component of RTI, we first perform a PCA on variables describing flexibility and learning on the job separately. Using an inverse of the resulting flexibility and learning indices, along with the routine manual component, we use PCA to construct a composite routine component. Each subcomponent (abstract, routine and manual) is standardized to lie between zero and one. Finally, we construct the RTI index by subtracting abstract and manual components from the routine component and standardize the final index score to lie between zero and one.

B. The Gender RTI Gap

The RTI index, on average, is 13 percent higher for female workers across our sample of 30 countries (Table 2), a result that is statistically significant. This gender RTI gap ―the ratio of the female to the male RTI level―is driven by female workers performing fewer tasks requiring analytical and interpersonal skills or physical labor, and more tasks that are characterized by lack of job flexibility, little learning on the job, and greater repetitiveness. This implies that female workers are more exposed to automation than male workers, on average, across our sample.

Table 2.

Gender Differences in RTI Components

article image
Sources: PIAAC survey; and IMF staff estimates.Note: The RTI index ranges from zero to one, with higher values indicating that a worker engages in more routine activities. The t-test evaluates whether there is a significant difference between the mean values for males and females.

The sample average masks significant cross-country heterogeneity (Appendix Table A2). The level of female RTI is highest in Eastern and Southern European countries and lowest in Scandinavian and Central European countries. For instance, the RTI index level for female workers in Lithuania is 36 percent higher than in Norway. There are also marked differences in gender RTI gaps, which capture the relative routineness of female versus male jobs. Within Europe, while Estonia and the Slovak Republic have the highest gender RTI gaps, Greece exhibits much lower gender RTI gaps. Outside Europe, Japan and Singapore exhibit the highest gender RTI gaps in our sample. This heterogeneity is indicative of countries’ positions along the automation path as well as cross¬country differences in selection of women into the labor force. It can also reflect differences in the structure of production (e.g., manufacturing vs. services sectors which require interpersonal communication), technologies adopted, and labor market flexibility, and thus differential distribution of workers across sectors and occupations (IMF 2018; Das and Hilgenstock 2018).8

We find that the gender RTI gap is correlated with the female labor force participation (FLFP) in a country (Figure 1). Gender RTI gaps are smaller in countries with higher FLFP, suggesting that a more equal representation of women and men in the workplace is associated with more equal division of tasks between men and women. We also find that gender RTI gaps are positively associated with the size of manufacturing sector (as a percentage of GDP). In other words, countries with larger manufacturing sectors tend to exhibit larger gender RTI gaps. This relationship is supported by existing research which shows that services tend to be more gender equal in employment than manufacturing (Weinberg, 2000; Borghans, Ter Weel, and Weinberg, 2014; Ostry and others, 2018).9 Hence, structural transformation away from manufacturing and towards a large service-based economy not only implies greater labor force participation of women on the “extensive margin,” but can also lead to a more equal division of tasks between genders.

Figure 1.
Figure 1.

Relationship Between Female Labor Force Participation and Size of Manufacturing Sector and the Gender RTI Gap

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: International Labour Organization; PIAAC survey, World Bank, World Development Indicators; and IMF staff calculations.Note: RTI = routine task intensity. RTI gap= Female RTI/Male RTI.

C. Decomposing the Gender RTI Gap

To better understand the determinants of the gender RTI gap, we decompose the RTI gender gap into contributions from individual and job characteristics (e.g., age, numeracy and literacy skills, etc.), worker’s education, and occupational and sectoral choices. In particular, we estimate the following specification:

RTIi=β0+β1Femaleic+ΣindβindXicind+ΣabilityβabilityXicability+ΣjobβjobXicjob+α0+σs+τc+εic,

where β0 is a constant; Femaleic is a binary (0/1) indicator for female workers; Xicind are individual controls including age, level of education, presence of a partner and children, and immigrant status; Xicability are numeracy, literacy, and problem-solving test scores; Xicjob include experience, on-the-job training, and part-time status; α is occupation fixed effects; σs is sector fixed effects; τc is country fixed effects; and εic is a normally distributed error term, with standard errors clustered at the country level.

For the decomposition, we first estimate the coefficient on the female indicator without any additional controls to derive an unconditional gender gap in the RTI index. We then estimate the full model to pin down the conditional gender RTI gap and examine how much of the change in the unconditional gender gap can be attributed to different control variables, using the decomposition method outlined in Gelbach (2012). The decomposition results show that almost 13 percent of the unconditional gender RTI gap is explained by occupational choice (Table 3).

Table 3.

RTI Decomposition: Drivers of RTI Gap

article image
Sources: PIAAC survey; and IMF staff calculationsNote: RTI = routine task intensity. Statistical significance levels: p<0.01; p<0.05; p<0.1.

These results indicate that women’s job positions within occupations and therefore the types of tasks that they perform (“the intensive margin”) are the main driver of gender disparities in routineness in the work place.

Since our decomposition results indicate that women’s occupational choices are the largest contributor to the gender RTI gaps, we analyze within-occupation gender RTI gaps in Figure 2 (Table A3 in Appendix).10 We find that gender RTI gaps persist across all occupations but are lower among professionals and managers―occupations which also exbibit a lower average level of routineness of tasks. These occupations, however, make up a smaller share of women’s overall employment, indicating that women are potentially less insulated from automation owing to their occupational choices. For instance, while over 30 percent of the female labor force works in professional occupations (relative to 25 percent of the male labor force), only 8 percent are managers, legislators and senior officials, relative to 15 percent among males. On the other hand, 20 percent of working females in our sample hold clerical positions relative to only 8 percent of males, with clerical occupations having the second largest gender RTI gap.

Figure 2.
Figure 2.

Gender Differences in RTI by Occupation

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: PIAAC survey; and IMF staff calculations.Note: Note: RTI = routine task intensity. Bars represent respectively average male and female RTI indices for each occupation. The occupational classification corresponds with the respondent’s 2-digit ISCO code. Statistical significance levels on the occupations reflect t-tests of the difference between male and female means of the RTI index. Statistical significance levels: p<0.01; p<0.05; p<0.1.

Table A4 in the Appendix shows the sectoral distribution of female workers. With the exception of retail trade and services, a sector which employs a significant proportion of the female labor force and has a large gender RTI gap (corresponding female-to-male ratio of RTI is 1.19), other sectors that employ large proportions of the female labor force (e.g., health and education) also tend to exhibit lower gender RTI gaps. This suggests that greater participation by women on the “extensive margin” in a sector can potentially lower their relative exposure to job automation.

III. Risk of Automation and The Future of Work for Women

A. Estimating the Probability of Automation

In this section we analyze how differences in task composition of men’s and women’s jobs translate into their relative likelihoods of being automated by estimating the probabilities of automation at the individual level.

The starting point of our analysis is the probability of automation estimates constructed by Frey and Osborne (2017). These estimates are based on the occupational classification and job task descriptions from ONET and rankings created by a panel of machine learning researchers of job automatability given “state of art computer-controlled equipment” and availability of big data.11 To construct a measure of an occupation’s susceptibility to automation, Frey and Osbourne (2017) use a two-stage process. In the first stage, a subset of occupations is hand-labeled into two categories, “automatable” or “non-automatable,” by machine learning experts. In the second stage, a probabilistic model is used to impute the probability of automation from the hand-labeled sample to the full sample of occupations using nine specific job task characteristics contained in the ONET data that are deemed to constitute bottlenecks in automatability. The resulting dataset contains 702 occupations and their associated probabilities of automation on a continuous scale between 0 and 100 percent.

Probabilities of automation provide an intuitive and forward-looking measure of the likelihood that an occupation can be codified and performed by a computer. However, there are two key drawbacks of assigning probabilities of automation solely based on the individual’s occupation. First, estimating automation probabilities at the occupational level assumes that all tasks within an occupation are automatable. This is likely to overstate automation probabilities, given that many occupations will entail a mix of tasks, not all of which can be automated at the current level of technology. Second, this method assumes that differences between men’s and women’s exposure to automation only arises from different modes of participation in occupations. This assumption obscures the fact that gender differences in exposure to automation can also arise from variation in task assignment at the workplace. To remedy this, we apply the method developed by Arntz, Gregory, and Zierahn (2017) to impute the probabilities to worker characteristics and job task descriptions from the PIAAC dataset.12

To relate automation probabilities at the occupational level to individuals, each observation in the PIAAC data is matched to the occupational codes in ONET for which we have estimates of automation probability from Frey and Osbourne (2017). Since the PIAAC data only contains 2-digit ISCO codes for occupations, each observation can be mapped to multiple occupations in the Frey and Osbourne (2017) estimates. Therefore, in the spirit of Arntz, Gregory, and Zierahn (2017) we use the Expectation Maximization algorithm and estimate an individual-level regression:

pij=Σn=1NβnXin+ϵij,(3)

where i denotes individuals, j denotes duplicates of these individuals when multiple probabilities are associated with one individual due to differences in the aggregation level of occupations, pij is the probability of automation, and Xin contains N individual, job, and task characteristics. βn are parameters which capture the impact of the regressors on probability of automation, which is restricted to the interval 0 to 100 percent.

We use a weighted Generalized Linear Model (GLM) for our estimation, with equal initial weights for all duplicates j for individual i For each iteration of the regression, we compare the prediction from our estimated model with the actual probability of automation from Frey and Osbourne (2017) and recalculate the weights as per Ibrahim (1990):

wij=f(p^pij|xin,βn)Σn=1Nf(p^pij|xin,βn),

where f(.) is the standard normal density and p is the predicted probability. Once weights converge and best fit is achieved, we use the estimated parameters βn to calculate the predicted probabilities of automation based on individual worker and job task characteristics. We estimate four models with minor variations in the set of regressors contingent on data availability.13 Table A5 in Appendix 2 shows model estimates based on variables that are common for all countries.

To test for robustness of our results, we use an alternative measure of automatability at the occupational level developed by Brynjolfsson, Mitchell, and Rock (2018). The “suitability for machine learning” (SML) index measures the degree to which machine learning has a potential to eliminate tasks within an occupation. There are two key differences between the SML index and probability of automation estimate by Frey and Osborne (2017). First, the SML index does not explicitly measure the likelihood of automation and its values exceed one. Therefore, interpretation of the SML index is less intuitive. Second, a more elaborate rubric for each set of tasks, or “direct work activities” (DWA), is used to assign subjective measures of automatability. Thus, the SML index is entirely based on subjective rankings.14 Higher values of the SML index indicate higher capability of machine learning to substitute a worker in a given occupation. We apply the Expectation Maximization algorithm described above to map the occupation-level SML index to the individual level.

B. Gender Differences in Probability of Automation

Cross-country Differences

As can be seen from Table 4, women face a higher probability of automation than men. The average probability of automation among women in our sample is 40 percent, 2 percentage points higher than the average probability of automation among men―a difference that is statistically significant. The right panel of Table 4 focuses on workers at the highest risk of automation, defined as those with at least 70 percent probability of being displaced by current technologies in the next two decades. Women are 1.2 times more likely to be in the high-risk group—around 11 percent of women and 9 percent of men face a probability of automation higher than or equal to 70 percent.

Table 4.

Gender Differences in Probability of Automation and SML

article image
Sources: Frey and Osbourne (2017); Brynjolfsson, Mitchell, and Rock (2018); PIAAC survey; and IMF staff estimates. Note: The probability of automation and SML are estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. High automatability is defined as having probability of automation >= 0.7 or SML>=3.5. Gender ratio = female measure/male measure. Statistical significance levels on ratios reflect t-tests of the difference between male and female means or proportions of workers at high risk for a given measure. Statistical significance levels: p<0.01; p<0.05; p<0.1.

The second row of Table 4 contains average predicted SML values by gender as well as the share of female and male workers at high risk automation (the 70th percentile of the SML index).15 The results are consistent with our baseline findings for both average values of SML as well as shares of female and male workers at high risk of automation. In both cases, women have a higher exposure to machine learning than men, and these mean differences are statistically significant.

The difference between shares of men and women most exposed to machine learning is substantially larger when measured with the SML index compared with our baseline results. In particular, 27 percent of women compared to only 7 percent of men are at high risk of displacement by machines. These results provide further supporting evidence for our baseline estimates of women’s susceptibility to automation.

How do the estimates for probability of automation compare to the RTI results? As foreshadowed in the previous section, the likelihood of automation increases with the RTI level. Our estimates for the probability of automation line up with Autor, Levy, and Murnane’s (2003) task framework, with the probability of automation having a strong, statistically significant, positive relationship with the degree of routineness of work tasks and a statistically significant negative relationship with the abstract and manual components of the RTI index (Table 5).

Table 5.

Correlation of RTI Components and Probability of Automation

article image
Sources: Frey and Osbourne (2017); PIAAC survey; and IMF staff calculations.Note: Routine task intensity (RTI) index is calculated at the individual level using information on routine, abstract, and manual tasks. Abstract index describes analytical and interpersonal tasks; manual index describes long hours of physical work (non-routine); routine index describes lack of job flexibility, little learning on the job, and repetitive tasks. Subcomponents of the routine index (lack of flexibility, lack of learning, manual routine) are also individually significant with a positive sign. The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.

Country rankings of the gender gaps in probability of automation, both in levels and in the shares of high-risk workers (Table A6 in Appendix 2), are similar to the ranking of gender RTI gaps. In Japan, Singapore, and South Korea, women face a higher probability of being displaced by technology compared to men. However, this result is reversed in Sweden and Finland, where men face a greater likelihood of automation. These differences are statistically significant for both average probabilities of automation and relative shares of workers at high risk of automation.16 In France, the UK, and the US―countries with large service-dominated economies―the proportion of men and women in the workforce who face a high risk of automation is roughly similar.

Individual Differences

In this subsection, we explore the effects of individual and job characteristics on the susceptibility to automation (see Table A5 in Appendix 2). Workers spending a larger share of their time on selling, adhering to manuals, and using fingers and hands to complete the tasks, face a higher probability of automation. On the other hand, tasks such as trainings others, solving complex problems, and writing articles are associated with a lower risk or automation.

The likelihood of automation is decreasing in education, income level, and firm size.17 The risk of automation is less than 1 percent among workers who have a bachelor’s degree or higher. The most disadvantaged group is women with lower secondary education or less, with nearly 50 percent at high risk of automation, relative to less than 40 percent of men with the same level of education.18 Interestingly, our analysis predicts that workers in small and medium enterprises (SMEs) have a higher risk of being automated as compared to workers in large enterprises (greater than 1000 workers). This may be a result of SMEs lagging behind large firms in their adoption of digital technologies, resulting in their workforce being more substitutable and less complementary with these technologies.19 Moreover, while female workers are significantly more likely than male workers to face risk of automation in SMEs, they are significantly less likely to face risk of automation in large firms relative to their male counterparts. This could reflect differences in firm policies regarding hiring or retention, or the selection of preferred workers into large firms.20

Sectoral and Occupational Differences

There is significant heterogeneity in the likelihood of automation across sectors and occupations (Figures 3 and 4). This evidence is consistent with the earlier finding that sectoral and occupational choices are important determinants of the gender RTI gaps. Overall, our analysis suggests that accommodation and food services, retail trade, and transportation face a high risk of automation (Figure 3 and Table A7 in Appendix 2). While retail trade and accommodation and food services employ roughly similar proportions of the male and female workforce, men are disproportionately represented in the transport sector, and therefore more exposed to risk of automation owing to their sectoral choice.21 Women however, are vastly overrepresented in education, health, and social services―sectors which are at a relatively low risk of being automated. However, gender gaps in the risk of automation varies within sectors (Figure 3). For instance, despite the large share of female workers in education, women face a higher likelihood of being automated than men. This suggests that the extent to which female workforce is at risk of being automated depends not just on sectoral choices but also the job composition within a sector.

Figure 3.
Figure 3.

Gender Gap in High Risk of Automation Across Sectors

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: Frey and Osbourne (2017); PIAAC survey; and IMF staff estimates.Note: The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. Bars represents respectively the proportion of male and female workers at high risk of automation. High automatability is defined as having probability of automation >= 0.7. Dots (and associated percentages) reflect the proportion of the male and female labor force employed in each sector. Statistical significance levels on the sectors reflect the t-test of the differences between proportions of male and female workers at high risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.
Figure 4.
Figure 4.

Gender Gap in High Risk of Automation Across Occupations

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: Frey and Osbourne (2017); PIAAC survey; and IMF staff estimates.Note: The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. Bars represents respectively the proportion of male and female workers at high risk of automation. High automatability is defined as having probability of automation >= 0.7. Statistical significance level on the occupations reflects the t-test of the differences between proportions of male and female workers at high risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.

The risk of automation also varies across occupations (Figure 4 and Table A8 in Appendix 2). Overall, elementary occupations, clerks, and service workers are most exposed to risk of automation, on average. Women are overrepresented among service workers and face a higher risk of automation than their male counterparts.22 Legislators, managers, and professionals are relatively insulated from the threat of displacement by automation, with less than 1 percent of the workforce (male and female) in these occupational categories being at high risk of automation. Among professionals while the threat from automation is overall low and women are at a lower risk of automation than men on average, they are also twice as likely as their male counterparts to be among the group of workers at highest risk for automation (Figure 4 and Table A8 in Appendix 2). In retail trade, where the overall risk of automation is very high, female workers are significantly less likely than male workers to perform abstract tasks—reflecting the fact that there are fewer women in managerial positions. These findings suggest that the assumption of a homogeneous distribution of tasks across different workers within occupations is not supported by the data.

IV. Narrowing Gender GAPS Across Generations

The previous sections highlight women’s exposure to automation, but there are some positive trends. The distributional shift of women towards technical and professional occupations has accelerated in the last two decades. A shift-share analysis of jobs between 1994 and 2016, suggests that most job growth in OECD countries has been on the high-skill end, and that women have benefited from this trend more than men.23

Figure 5 indicates an overall shift away from clerical occupations towards service and retail workers, technicians, and professionals―a trend that is more pronounced for women. Overall, women appear to be increasingly opting into occupations that are relatively more insulated from the risk of automation.24 However, there appears to be no significant increase in the numbers of women among legislators, senior officials and managers suggesting that while women may be entering more professional occupations, they have yet to increase their representation in leadership positions.

Figure 5.
Figure 5.

Changes in Occupational Shares by Gender (1994–2016)

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: International Adult Literacy Survey; PIAAC survey; and IMF staff calculations.Note: Shift in occupation is calculated as the difference between share of female (male) workers in occupation X and country Y in 1994, and the share of female (male) workers in the same occupation and country in 2016. Top and bottom values of the intervals represent country-specific maximum and minimum values for occupational share differences.

We find that among both men and women, younger people are at the highest risk of losing their jobs to automation given current technologies. This is potentially driven by the selection of less-educated workers into the labor force, given the high returns from human capital accumulation through education at this age. In addition, even highly educated youth tend to occupy more junior-level positions, which are more prone to automation, before moving to senior positions requiring more abstract skills. Moreover, automation can reduce employment, not merely by replacing workers, but by potentially reducing new-job creation in some sectors or creating new high-skilled jobs, which is more likely to affect young entrants as opposed to older incumbents.

To further analyze intragenerational gender difference in the risk of automation, we divide our sample into a younger cohort (ages 20 to 39) including individuals born between early 1970s and mid-1990s, and an older cohort (ages 40 to 65) including individuals born between mid-1940s and early 1970s. Overall, we find that older cohorts of working women (older than 40) are at significantly higher risk for automation than men in the same age cohorts, suggesting greater disadvantage of women among older age groups. We further document a generational shift towards higher educational levels which are relatively insulated from risk of automation. Figure 6 shows that a higher share of women (1 percent) in younger generations received a tertiary degree than their older counterparts (0.4 percent). Moreover, while gender differences in tertiary education groups are not statistically significant for either older or younger generations, gender gaps in risk of automation have narrowed for less-well educated women, especially those with upper secondary education.

Figure 6.
Figure 6.

Gender Gaps in High Risk of Automation by Generation and Educational Level

Citation: IMF Working Papers 2019, 091; 10.5089/9781498303743.001.A001

Sources: PIAAC survey; and IMF staff calculations.Note: The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. Bars represent the gender difference in automatability = (Share of females at high risk for automation) / (Share of males at high risk of automation). High automatability is defined as having probability of automation >= 0.7. Individuals between ages 20 and 39 are defined as younger generation. Individuals between ages 40 and 65 are defined as older generation. Statistical significance levels on bars reflect t-tests of the differences between proportions of male and female workers at high risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.

Table 6 documents a shift away from the clerical occupations and towards professional occupations for younger cohorts of women as well as narrowing of the gender gap within both occupations, consistent with the shift-share analysis presented in Figure 5. While the older cohort of women face a higher probability of automation than men in both clerical and professional occupations, this difference becomes statistically insignificant for younger generations and even reverses in the case of professionals. The automatability gap between genders also reverses in the case of legislators and senior managers across cohorts.25 A less encouraging trend pertains to the large automatability gap and rising share of women in service and retail occupations, which face the highest risk of automation.

Table 6.

Gender Gaps in High Risk of Automation by Generation and Occupation

article image
Sources: PIAAC survey; and IMF staff calculations.Note: The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. High automatability is defined as having probability of automation >= 0.7. Individuals between ages 20 and 39 are defined as younger generation. Individuals between ages 40 and 65 are defined as older generation. Gender difference in automatability = (Share of females at high risk of automation) / (Share of males at high risk of automation). Female share of occupation is defined as the share of occupation in total female employment within a generation. Statistical significance levels on ratios reflect t-tests of the differences between proportions of male and female workers at high risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.

Finally, we compare inter-generational trends across sectors. Table 7 shows an across-the-board decline in women’s relative exposure to automation for younger cohorts. Gender differences in the degree of automatability are statistically significant in almost all sectors for older generations. However, gender differences are substantially lower (four-fold in cases of education and financial services) and lose their statistical significance for younger cohorts of women. Accommodation and food services—one of the sectors most exposed to automation—is the only exception to this trend.

Table 7.

Gender Gaps in High Risk of Automation by Generation and Sector

article image
Sources: PIAAC survey; and IMF staff calculations.Note: The probability of automation is estimated using an expectation-maximization (EM) algorithm that relates individual characteristics (age, education, training, among other) and job task characteristics to occupational level risk of automation. High automatability is defined as having probability of automation >= 0.7. Individuals between ages 20 and 39 are defined as younger generation. Individuals between ages 40 and 65 are defined as older generation. Gender difference in automatability = (Share of females at high risk of automation) / (Share of males at high risk of automation). Female share of sector is defined as the share of sector in total female employment within a generation. Statistical significance levels on ratios reflect t-tests of the differences between proportions of male and female workers at high risk of automation. Statistical significance levels: p<0.01; p<0.05; p<0.1.

The overall positive trends displayed by the inter-generational analysis suggests that there is room for optimism about the future of work for women. An important caveat, however, is in order. Since our dataset does not allow us to track workers over time, these results could also reflect widening gender gaps in routinization and thus risk of automation over the women’s lifecycle. Kleven, Landais, and Sogaard (2018), for instance, document that a large part of gender inequality in the labor market can be explained by a “child bearing penalty.” Thus, our findings suggest an important role of policies for smoothing transitions of younger women workers and ensuring adequate safety nets for older, displaced workers.

V. Conclusion

Using individual-level data on task composition at work, this paper finds that women, on average, perform more routine tasks than men across all sectors and occupations―tasks that are most prone to automation. Given the current state of technology, we estimate that 26 million female jobs in 30 countries (28 OECD member countries, Cyprus, and Singapore) are at a high risk of being displaced by technology (i.e., facing higher than 70 percent likelihood of being automated) within the next two decades. Female workers face a higher risk of automation compared to male workers across all occupations, albeit with significant heterogeneity across sectors and countries.

There are some positive trends. Women’s educational attainment has increased markedly in the last few decades in many developed economies. Further, women appear to be increasingly opting into occupations that are relatively more insulated from the risk of automation. The increase in women in professional roles and away from clerical and service positions is particularly encouraging in this regard, but more needs to be done.26 Across sectors and occupations, underrepresentation of women in managerial positions places them at high risk of displacement by technology.

Our estimates for the likelihood of automation assume current levels of technology and prevailing bottlenecks in the use of computer-controlled equipment. As such, our analysis presents a lower bound for the potential impact of automation. Given the speed of technological advancement in recent years, further improvements in the state of computing could result in more tasks being automated than predicted by the current level of technology. Our estimates are also based on the technological feasibility of automation as opposed to the economic feasibility: tasks could be automated given the current state of technology, but the costs of automation may be prohibitive relative to the prevailing cost of labor. Moreover, job losses could potentially be offset by new work opportunities created by technology and higher output potential owing to falling costs and prices. Finally, our analysis does not capture the emerging “gig” economy― employment in flexible, independent work arrangements – which could make it easier for women to combine paid work with family responsibilities, potentially improving their labor market outcomes.

Given considerable cross-country heterogeneity, ongoing work examines the drivers of gender RTI gaps across countries and evaluates differences in the probability of automation across sectors and countries. Further work on the gig economy is also warranted to assess the attendant opportunities and challenges for the future of work for women.

Is Technology Widening the Gender Gap? Automation and the Future of Female Employment
Author: Mariya Brussevich, Ms. Era Dabla-Norris, and Salma Khalid
  • View in gallery

    Relationship Between Female Labor Force Participation and Size of Manufacturing Sector and the Gender RTI Gap

  • View in gallery

    Gender Differences in RTI by Occupation

  • View in gallery

    Gender Gap in High Risk of Automation Across Sectors

  • View in gallery

    Gender Gap in High Risk of Automation Across Occupations

  • View in gallery

    Changes in Occupational Shares by Gender (1994–2016)

  • View in gallery

    Gender Gaps in High Risk of Automation by Generation and Educational Level