Technology, routinization and wage inequality: gender differences in the case of Uruguay

As is well known, gender differences in employment rates remain, although female participation in the labor market has been steadily increasing. This implies that, even conditional on observable characteristics, the wage gap that emerges from the comparison between male and female workers could be misleading. The wage gap may be influenced by this bias, due to the fact that women that participate are those who expect to have higher remuneration. This is known in the literature as the ‘selection bias’ which is necessary to address when performing estimations of the wage gap. Olivetti and Petrongolo (2008) analyze the median gender gap in developed countries, and use a novel procedure to correct for selection bias. To this end they impute a variable that indicates whether the wage of the unemployed would be above or below the median value of wages. Here we follow the same idea but using the multiple stochastic imputation technique which allows us to correct our estimates for selection bias across the whole distribution. The multiple stochastic imputation technique, firstly proposed by Rubin (1987), consists of three stages: a) imputation, in which K > 1 complete (simulated) databases are generated from a given imputation model; b) analysis with complete data, calculating the indicators of interest in each of the K bases constructed and c) aggregation of the parameters of interest estimated following the so-called ‘Rubin’s rule’. The procedure take into account the additional uncertainty arising from missing data. As in Olivetti and Petrongolo (2008), we impute the wages of those who are unemployed, and leave the wages of those who do not participate in the labor force as missing. To do this, separated by gender and period, we estimate mean linear regression models where log hourly wage is the dependent variable. From these estimates, we generate a simulated value for each unemployed person’s wage by randomly drawing it from the estimated distribution of coefficients and residuals and then, impute the value by taking the mean of the four closest values within the set. This procedure is repeated K = 10 times to generate our multiple imputed data sets.

In all models we include as predictors: education, experience, industry and geographical location dummy variables. We consider flexible specifications that introduce interactions between these covariates. All covariates are observed for the unemployed except industry. However, the industry of their last job is collected by the survey and we use it as a predictor.

The consistency of our imputation-based estimation is mainly based on two assumptions. First, if the unemployed were working, they would be working in the sector of their last job. The same assumption is made to define the information and automation task content measures that correspond to them. Second, and more important, after controlling for observable characteristics, the observation is missing at random. This implies that the wage that an unemployed person would earn is equivalent on average to that earned by an employed person with identical observable characteristics. This assumption is not too restrictive in the current Uruguayan context, given that a large share of the unemployed belong to sectors and occupations where wages are set according to collective bargaining rules, with individual wage bargaining being less frequent within these occupations.

A final consideration regarding the use of multiple imputed data sets is how to obtain point estimates and their variance. To do so, we use a procedure proposed in Schomaker and Heumann (2018) that first obtains the point estimate and bootstrapped standard errors on each simulated dataset and then, applies Rubin’s rules to obtain the estimates of the parameters and their variance. Schomaker and Heumann (2018) show that this procedure yields the correct coverage of the estimated statistics, in particular when K > 5.

In this section we present the RIF-regression decomposition method introduced by Firpo et al. (2009, 2011a, 2018) (FfL from here on). In what follows we present the main ideas and refer to the authors for further details. A RIF-regression is a regression where the dependent variable, Y, has been replaced by the recentered influence function (RIF) of the statistic of interest v(F). In the particular case of quantiles, the RIF-regression is known as unconditional quantile regression (UQR) since, unlike conditional quantile regressions, its coefficients reflect the partial effect of changes in the covariates over the unconditional quantile of the variable of interest (UQPE). RIF-regressions in the case of quantiles can be expressed as, (1) E(RIFτ(y) | X)=X′γ(τ) E\left( {{RIF_\tau }\left( y \right)\,|\,X} \right) = X^\prime\gamma \left( \tau \right) where RIF_τ(y) is the RIF at quantile τ of variable y (in our case log hourly wages), X is a vector of observable covariates and γ(τ) are the UQPE of the covariates at quantile τ.

An advantage of RIF-regressions is that they allow identifying non-monotonic effects, like explanations regarding changes on wage inequality that affect specific points of the distribution. For instance, the automation of routine jobs proposed by ALM tends to affect the middle and lower-middle of the distribution. As RIF-regressions can be applied to quantiles and other distributional statistics, they represent a good methodological alternative to go beyond the mean to better understand changes in wages inequality.

Once the RIF-regressions are estimated it is straightforward to decompose the overall change of the statistic of interest (quantile, variance, Gini index, etc.) performing Oaxaca-Blinder. The idea is to use the RIF for the statistic of interest, instead of the outcome variable, as the left hand side variable in a regression. The estimated coefficients of the RIF-regression can be used to perform the detailed decomposition in the same way as a standard Oaxaca-Blinder decomposition (Firpo et al. (2011a)).

The proposed method presents several advantages, being easier to interpret and less computational intensive than other decomposition methods like Chernozhukov et al. (2013) or Machado and Mata (2005). Another important advantage is that the detailed decomposition of wage structure and composition effects is path independent since it is possible to isolate the effect of each covariate by introducing all covariates in one step.

As in the case of the standard Oaxaca-Blinder decomposition, performing a decomposition based only on the RIF-regression may have a bias problem if the linear specification used in the regression is inadequate (Firpo et al. (2018)). To solve this problem, Firpo et al. (2011a, 2018) recommend a two-step procedure to estimate the different elements of the decomposition. In the first stage, distributional changes are divided into a structure effect and a composition effect. This stage is based on a reweighting procedure to cope with potential non-linearities in the true conditional expectation. The second stage further divides the structure and the composition effects into the contribution of each covariate, and is based on the estimation of RIF-regressions.

The aggregate decomposition consists of dividing the overall change of a given distributional statistic (in our case quantiles) (∆^τ_O) into the effect of changes in coefficients (structure effect, (∆τ_S)) and in characteristics (composition effect, (∆^τ_X)). The sample analogous of these parameters can be defined as, (2) Δ^Oτ=Δ^Sτ+Δ^Xτ=wage structure effect︷(X¯B ι)′(γ^A(τ)−γ^B(τ))+composition effect︷(X¯A−X¯B)′γ^A(τ) \hat \Delta _O^\tau = \hat \Delta _S^\tau + \hat \Delta _X^\tau = \overbrace {{\rm{wage}}\,{\rm{structure}}\,{\rm{effect}}}^{\left( {{{\bar X}_B}\,\iota } \right)^\prime\left( {{{\hat \gamma }_A}\left( \tau \right) - {{\hat \gamma }_B}\left( \tau \right)} \right)} + \overbrace {{\rm{composition}}\,{\rm{effect}}}^{\left( {{{\bar X}_A} - {{\bar X}_B}} \right)^\prime{{\hat \gamma }_A}\left( \tau \right)} where A,B denotes two different groups, or the same group at two different points in time, and ι is a vector of ones.

The structure effect reflects the change on the conditional distribution (F(Y/X)) of the variable of interest and the composition effect reflects the effect of changing the distribution of the covariates (X).

In the literature the composition effects are usually referred to as the explained effects while the structure effects as the unexplained effects.

The detailed decomposition allows to disentangle the aggregated wage and structure effects into the contribution of each individual covariate (or group of covariates). This lets us compare the contribution of changes in the returns to occupational tasks to other explanations such as changes in the labor market returns to general skills (experience and education), which have been the most common explanations to changes in wage distribution.

In this paper we use this approach to decompose the observed wage changes over a decade (A corresponds to 2005 and B to 2015) and the gender gap in each period (A corresponds to men and B to women). We use analogous ideas to decompose the change in the gender wage gap over the decade, as we explain in Section 6.

We construct our task content measures following Firpo et al. (2011b) and Jensen and Kletzer (2010) who focus on the ‘Occupational Requirements’ of occupations. In the spirit of Autor et al. (2003) to measure routine versus non routine and cognitive versus non cognitive aspects of occupation, we construct two variables thought to be positively related to technology: ‘Information content’ and ‘Automation/routinization’. As for Uruguay there are no studies nor a systematic database of task content of occupations, we use for this purpose the O*NET 15.0 data available from the National Center for O*NET Development. The O*NET content model organizes the key features of an occupation into a standardized, measurable set of variables called ‘descriptors’.

The Information Content Index seeks to identify occupations with high information content that are likely to be positively affected by ICTs, and within the Generalized and Detailed Work Activities subdomain we consider the following work activities: ‘Getting information’, ‘Processing information’, ‘Analyzing data or information’, ‘Interacting with computers’ and ‘Documenting/Recording information’.

For the construction of the Information Content Index, the O*NET provides information on the importance and level of the required work activity and on the frequency of five categorical levels of the work context. Following Firpo et al. (2011b), we give a Cobb-Douglas weight of two-thirds to importance and one-third to level by using a weighted sum for work activities. In addition, we performed a robustness check in which we calculated this index by giving one-third to the former and two-thirds to the latter.

The Automation Content Index is constructed using the Work Context subdomain, to reflect the degree of potential automation, including: ‘Degree of automation’, ‘Importance of repeating same tasks’, ‘Structured versus unstructured work (reverse)’, ‘Pace determined by speed of equipment’, and ‘Spend time making repetitive motions’.

We normalize the task measures by dividing them by their maximum value observed over all occupations, so that they range between zero and one. This gives us a ranking of occupations for each of the two dimensions. We use these indexes to assess the impact of technological progress on changes in wages and their influence on the gender wage gap.

As it is observed in Figures 3 and 4, alike the results reported by Firpo et al. (2011b) using US data, Professional, managerial and technical occupations have the highest score in terms of their use of information, and a relative low score for automation. On the other hand, Production workers and operators have a low score in terms of their use of information and the highest score for automation. According to the routinization hypothesis, technological change is expected to have an adverse impact on wages in this last group of occupations, which tends to be the more intensely subject to machine displacement, while benefiting those with a more intense use of information technology.

Figure 3

Figure 4

In Table 3 and 4 we report the average value and standard deviations of the measures of task content indexes for five major occupational groups. As it can be seen, the distribution of both indexes among occupational categories by gender shows significant differences. The average Information Index of female occupations is higher than that of male occupations. However, when consider particular sectors, female information task index is only higher in clerical and service occupations in 2005 and in Agricultural, construction and transport and service occupations in 2015. The opposite is the case when considering the average Automation Index which is higher for male than for females. However, only in the case of manager, professionals and technicians sector men automation index is higher than that of women.

Since the task content indexes are not readily interpretable, we compute quartiles of each of the two task content measures that are included in our regressions. In Table 5 we report the percentage of workers, by gender, in five major occupational groups that rank in the top quartile of each of our O*NET task content measures in 2005–2006 and 2014–2015. As expected, the highest percentage of workers in the top quartile of our information content measure is found among professional, managerial and technical occupations and among them, men are more likely than women to be in the top quartile of our information content measure. On the other hand, more than half of production workers are in the top quartile of our automation/routine measure, and this is even clearer in the case of women.

In the overall, the percentage of men in the top 25 percent of information is smaller than the percentage of women, while in the case of automation is the other way around. However, within each occupational group women are more likely than men to perform tasks more likely to be automated, except for service occupations. Nevertheless, an important part of the variation in the occupational task measures comes from production operators, primary, construction and transport occupations and 53 percent of men but only 8 percent of women are in these occupations, which have very low scores for information and very high scores for automation.

In the group of professional, managerial and technical occupations, which is the group with the highest information index, a smaller percentage of women than of men are placed in the top quartile, although in both cases the percentage is quite large. On the other hand, in the case of automation, in the group of production and operators occupations, where only 6 percent of women are working, 69 percent of women work at occupations at the top 25 percent rank of the automation index. Besides, among clerical and sales workers, which gather 36 percent of women, more than half of them were at the top quartile degree of automation in 2005, but this percentage reduced to 30 percent in 2015. However, women are still concentrated in the service sector (42 percent of women), which after professional and technical occupations is the group with the smallest index of automation task content.

Given these dissimilarities between genders in the distribution of occupations and their degree of automation and information task content, technology is expected to impact men and women employments and salaries in quite a different way. What is more, it is interesting to notice that in the period of analysis, women employment increased mostly among professionals and technicians and in clerical and sales occupations. The first ones demand high educated workers, and the last ones require middle and low educated workers. Although a big share of this employment still demands a high degree of automate tasks, the percentage of workers in this sector at the top 25 per cent of automation had significantly diminished. Nevertheless, it still remains high, and women are a major part of them. Therefore, they are more vulnerable to being substitute by a machine in the future, since we expect ICTs to enhance tasks involving the processing of information performed by high skilled workers while substituting those tasks that can be automated and generally performed by middle skilled workers.

We first estimate the RIF-regressions for different quantiles of the log hourly wage distribution. Covariates include dummies for each quartile of our two measures of occupational tasks. In addition, we include standard controls for education (five dummies) and potential experience (nine groups)

Potential experience is measured as age minus years of education minus six.

We estimate RIF-regressions in each period for the whole sample (in which case we add the female covariate) and for each gender separate. The base group are married formal employees, working at the private sector, with six or less years of education, 15 to 20 years of potential experience, at the bottom quartile of each of the task content measures and living in Montevideo.

The omitted group education and experience categories were chosen based on the modal of each category.

Tables 6, 7 and 8 report the Unconditional Quantile Partial Effects at 10^th, 50^th, and 90^th quantiles of the whole sample, male and female wage distribution, along with their bootstrapped multiple imputation standard errors. We consider two periods 2005/06 and 2014/15. Additional estimates for the 10^th to the 90^th quantiles are also reported in Figures 5 and 6 for some variables of interest.

Figure 5

Figure 6

The first thing to note is that the gender wage gap widens, in the period covered, except at the top of the distribution. Indeed, the female dummy variable increases in absolute terms from 2005 to 2015 to the 80^th percentile (see Figure 8).

Regarding task measures, as expected, we find that the effects of information and automation content of task diverge (Figure 5). This is true no matter if we consider all cases, or just males or females. The information task content premium is positive for male and substantially increases along the whole wage distribution. For women, that premium is also positive but a little hump-shaped: it increases until the 70^th percentile but reduces the slope from that point on. Such patterns are observed in both periods, but premiums fall for both genders between 2005 and 2015; in particular at the top half of male wage distribution. Those results are in line with those presented by Firpo et al. (2011b) for the US during the nineties. It is also remarkable that the premium to the information task content is larger for females than for males, in particular in 2015.

On the other hand, the return to the automation task content is positive in the first half of the wage distribution but turns negative in the top half. This result is driven by returns to female workers. For males, it is not significantly different from zero at the bottom half and negative at the top half. For women, however, it is positive and significant until the 60^th percentile and negative at the top tail. Such results would not be consistent with ALM’s routinization hypothesis, which postulates that workers at the middle of the distribution are more likely to experience negative wage changes as the routine tasks they perform are more likely to be executed by machines. What is more, in the case of women returns to automation increased at the middle of the distribution between 2005 and 2015, although they diminished at the top end.

It is interesting to notice, the different impact of automation by gender at the bottom of the distribution. It could be explained by the predominant presence of women in the service sector, which is characterized by low wages and a relative low degree of automation. That is, the returns to automation capture the fact that women at the bottom of the wage distribution earn a positive premium if their work imply some degree of automation and, contrary to men, at the lower end of the distribution - which perform task more subject to automation - are less subject to displacements effects.

Education premium varied along the wage distribution and between education levels. The premium for tertiary education (16 and more years) relative to the base group (6 or less years of schooling) is strongly increasing over the wage distribution for both genders, but substantially larger for male than for women at the top tail. Therefore, as it is well documented, despite women being on average more educated than male, skill premia received by women at the top are significantly less than those payed to male. An interesting feature is that premiums at the top showed a substantial reduction between 2005 and 2015 for men, but not for women. For 10 to 12 years of education, premiums are increasing along quantiles for men but they are almost flat for women. Also, premiums for men experienced a decrease in the period under analysis. Those results are in line with the evidence of other studies regarding skill premia in Latin American and Uruguay (de la Torre (2012); Yapor (2018)) (See Figure 6). As we show in the next section, this different evolution of education premia for males and females worked in the direction of reducing the wage gender gap in the period, specially at the top end of the distribution.

To sum up, in the case of Uruguay, as well as in other developing countries, the effect of automation and the consequences predicted by the routinization hypothesis seem to be displaced towards the right side of the wage distribution. Such a shift may be linked to the relative cost of labor to technology, divergences of occupations’ shares; and also to the different degree of automation of tasks, compared with those of developed countries’ labor markets. That is, in developing countries, labor tasks subject to substitution by technology would rather be placed at an upper level of the distribution of wages than at the middle; since wages at the middle are relatively low and therefore there are fewer incentives to substitute labor by technology. Furthermore, contrary to the routinization hypothesis predictions, Apella and Zunino (2017) find that between 1995 and 2015 (and between 2005 and 2015) there was an increase of the relative importance of routine cognitive task in the labor market. Since this task are generally performed by workers at the middle of the distribution, this increase in the demand of routine cognitive task at the labor market could explain the positive effect of automation in returns at the lower half of the distribution.

To cope with the first approach, we add covariates that capture labour attributes into the equation (3) such that the model to estimate is: (4) E(RIFτ(y) | t Male X)=α(τ)+β(τ)t+δ(τ)Male+φC(τ)t*Male+π(t, Male, τ)X E\left( {{RIF_\tau }\left( y \right)\,|\,t\,{\rm{Male}}\,X} \right) = \alpha \left( \tau \right) + \beta \left( \tau \right)t + \delta \left( \tau \right)Male + {\varphi _C}\left( \tau \right)t*Male + \pi \left( {t,\,{\rm{Male}},\,\tau } \right)X Where the covariates included in vector X are, education, experience, region, whether the worker is formal or not, marital status and our information and automation task content indexes. We use the function π(t,Male,τ) to indicate that the specification allows returns to each attribute at every percentile τ to vary between periods, and between males and females within each period.

It should be noted that the difference between φU(τ) estimated from the ‘raw’ wage gap equation (3) and the φ_C(τ) considering the observables in the equation (4) can be interpreted as the contribution of the attributes to the change in the gender wage gap over the period of analysis.

Figure 9.a plots these estimates and shows that the attributes considered drive an increase in the gender wage gap in the top half of the distribution, since the specification that does not include them is above the one that does. Such results could reflect changes in the mean of the covariates or in their returns, but this approach is not able to disentangle these features.

Figure 9

Regarding our measures of the technological content of tasks (Information and Automation), we find that technology acts in the direction of increasing the wage gap over the period. This is because the estimated coefficients of the specification that does not include these measures are above those of the one that includes all covariates.

Figures 9 (c) and (d) replicate a similar exercise but consider each variable separately. They show that both automation and information content increased the gender wage gap at the extremes of the distribution. However, when the information and automation task content measures are considered separately, their influence, after correcting for selection, is not significant at the 5 per cent confidence level, except at the top of the distribution. It should be noted, however, that without controlling for selection bias, the differences go in the same direction and are statistically significant (see Figure A2 in the Supplementary Appendix).

Education also plays a role in changing the gender wage gap, but in the opposite way to technology, i.e. by reducing the gap, as we show in figure 9e. This conclusion also holds when the estimates do not correct for selection.

In summary, our first methodological approximation to analyze the change of the gender wage gap, indicates that the full set of attributes, as well as the technology content measures, drove the gender wage gap up, while education drove it down. However, it should be noted that, while this approach provides valuable evidence, it also has some shortcomings. The first one is that it does not allow for an assessment of whether the influence of attributes comes from changes in mean characteristics and/or their returns, as mentioned above. Furthermore, although the consideration of alternative specifications provides insight into how job attributes have influenced the gender wage gap, these results suffer from the well-known problem of ‘path dependence’. That is, the method is not able to disentangle precisely the contribution of each covariate. In the following subsection, we address these problems with another approach.

In this section, we provide further evidence on the influence of job attributes on the change in the gender wage gap. To do so, we apply similar decomposition techniques to those used previously to analyze wage change between years. However, we now have to decompose a double difference. Our approach is based on the estimation of the following equation (analogous to equation 1), (5) E(RIFτ(y, t) | g, X)=X′γ(τ)tg E\left( {{RIF_\tau }\left( {y,\,t} \right)\,|\,g,\,X} \right) = X'\gamma \left( \tau \right)_t^g where t = 0,1 stand for years 2005 and 2015 respectively and g is M (Male), and F (Female). Given this model, the decomposition of the wage gap at percentile τ in each period t can be expressed as: (6) Δ^t, M Fτ=wage structure effect︷(X¯tFι)′(γ^(τ)tM−γ^(τ)tF)+composition effect︷(X¯tM−X¯tF)′γ^(τ)tM \hat \Delta _{t,\,M\,F}^\tau = \overbrace {{\rm{wage}}\,{\rm{structure}}\,{\rm{effect}}}^{\left( {\bar X_t^F\iota } \right)'\left( {\hat \gamma \left( \tau \right)_t^M - \hat \gamma \left( \tau \right)_t^F} \right)} + \overbrace {{\rm{composition}}\,{\rm{effect}}}^{\left( {\bar X_t^M - \bar X_t^F} \right)'\hat \gamma \left( \tau \right)_t^M}

Our first step is to separately estimate the aggregated decomposition of the gender wage gap in each period (figure 10.a). We find that in 2005 both, the composition and the structure effects, contribute to generate a gender wage gap, where males hourly wages are larger than those of females. The composition and structure effects are similar up to percentile 30, but from this point onwards the composition effect decreases while the structure effect increases, and at the 90th percentile the former becomes even negative (2 p.p.) while the latter is always positive scaling up to 30 p.p. at the top quantile. These results indicate that both average characteristics and returns can explain the gender pay gap in 2005, but returns play a major role at the half top of the distribution.

Figure 10a also shows that, in 2015, the composition effect decreases and the structure effect increases compared to 2005, almost in parallel along the distribution (except for the structure effect at the 90^th percentile). Moreover, the composition effect becomes negative from the 30^th percentile onwards. This means that, contrary to 2005, in 2015 women’s job attributes were, on average, better than men’s, the gender wage gap against women became larger.

Figure 10

We move forward to analyze the variation in the gender wage gap over time, which can be defined as: (7) Δ^Δt,M Fτ=Δ^1, M Fτ−Δ^0, M Fτ \hat \Delta _{\Delta t,M\,F}^\tau = \hat \Delta _{1,\,M\,F}^\tau - \hat \Delta _{0,\,M\,F}^\tau

It is important to note that the simple difference between the wage structure (composition) effect between periods 0 and 1 does not correctly reflect how changes in returns (characteristics) have contributed to the variation of the gender wage gap, because such differences mix shifts in returns and characteristics.

There are many alternatives to decompose the change of the gap over time, in this paper, we decompose it in three terms: structure, composition and interaction: (8) Δ^Δt,M Fτ=Δ^Δt(γ)τ+Δ^Δt(X)τ+Δ^Δt(γ)Δt(X)τ \hat \Delta _{{\Delta _t},M\,F}^\tau = \hat \Delta _{{\Delta _t}\left( \gamma \right)}^\tau + \hat \Delta _{{\Delta _t}\left( X \right)}^\tau + \hat \Delta _{{\Delta _t}\left( \gamma \right){\Delta _t}\left( X \right)}^\tau where (omitting the argument τ);

The wage structure is: (X¯0M−X¯0F)(γ^1M−γ^0M)+X¯0F((γ^1M−γ^1F)−(γ^0M−γ^0F)) \left( {\bar X_0^M - \bar X_0^F} \right)\left( {\hat \gamma _1^M - \hat \gamma _0^M} \right) + \bar X_0^F\left( {\left( {\hat \gamma _1^M - \hat \gamma _1^F} \right) - \left( {\hat \gamma _0^M - \hat \gamma _0^F} \right)} \right)

The wage structure effect captures how changes in returns contribute to narrowing/widening the gender wage gap in the period covered, under the assumption that average characteristics do not change over time. In other words, as if the characteristics remain as the beginning of the period. Conversely, the composition effect is capturing the influence of the variation in female characteristics and in the characteristics gap, assuming that returns do not change, and is measured at returns in period 0. Notice that the constant only enters into the structure effect, through its second term. Finally, the interaction effect reflects the difference between the observed gap variation and the sum of the wage structure and composition effects, since both characteristics and returns change over time.

Figure 10b shows the aggregate decomposition of the change in the gender wage gap into the three terms defined above. The composition term is negative, while the structure term is positive; both are almost flat in the distribution up to the 80th percentile, but the former increases and the latter decreases from this point onwards. In turn, the interaction term is close to zero along the distribution. This result is consistent with the previous finding, which indicates that women’s job attributes improved over the period, but the market remuneration for these attributes deteriorated compared to that of men. The opposite conclusion is reached when looking at the upper part of the distribution, where the composition term approaches zero and the structure term becomes negative. The latter indicates that the difference in returns at the 90th percentile moderates compared to 2005.

An advantage of our approach is that it allows us to proceed with the detailed decomposition of the influence of any variable of interest. We do that to estimate the influence of our measures of technological task content and education on the change of the gender wage gap.

Although we concentrate our analysis on these covariates, decomposition considers all attributes mentioned above.

Figure 11

Concerning the overall influence of the information task content we find that it acts in the direction of reducing the gender wage gap, in particular at the top half of the distribution (Figure 11.a). The composition term is negative, mainly due to the fact that in 2015 women were employed in jobs with a higher information task content than in 2005, while in 2005 the returns to these tasks were higher for women than for men, except at the top of the distribution. As for the structure effect, it is not significant up to the 80th percentile, but then becomes negative, indicating that changes in information content returns acted at the top of the distribution in the sense of reducing the gender wage gap.

In contrast, the overall influence of the automation task content drives an increase of the gender wage gap, it is U-shaped and its effect is close to zero at the central part of the distribution (Figure 11a). The contribution of the composite effect is negative but increasing, turning positive at the top of the distribution. This could be explained by the reduction in the degree of automation in employments performed by women at the top of the distribution over the period (Table 5). On the other hand, the structure effect is positive along the whole wage distribution, which can be explained by an increase in the gap in returns to automation task content.

With respect to education, the composition effect is zero at the bottom and increases monotonically up to 0.06 at the top. This effect is positive because the return gap is positive in 2005 and women education level increased between 2005 and 2015. At the same time, the increase in men education level was a little bit higher than that of women. The structure term clearly predominates and is negative and almost flat across the distribution indicating that changes on the return to education reduced the gender wage gap. This result, is mostly due to a reduction in the gap of education returns over the period and a reduction in returns to education observed during the period, specially at the top of the distribution. This reduction on the wage gap at the top could be explained by the effects of the tax and the health reforms. As noted before, both reforms contributed to reduce inequality over the whole wage distribution, given that they reduced in a higher proportion the actual wage received by workers mainly at the top of the distribution. Since in 2005, men wages at the top of the distribution were higher than women wages, in other words the wage gap was positive, the reforms implemented after 2008, affected more men than women wages and in that sense contributed to reduce the wage gap.

To sum up, the gender wage gap of private employees increases over the period of analysis, except at the top of the distribution. Gender wage gap is almost flat along the distribution until 80^th percentile, as well as its change between 2005 and 2015. Meanwhile, at the 90^th percentile the wage gap is significantly larger than at lower percentiles in 2005, while in 2015 it is not. The rise in the gender wage gap over the period covered was due to changes in the returns gap, which were partially offset by changes in the composition effects, except at the very top of the distribution. Regarding covariates, it is interesting to highlight that technology favored the increase of the gender wage gap while education worked in the opposite way.