Technological change, routinization and job polarization: evidence from a middle-income country

Transcription

1 Technological change, routinization and job polarization: evidence from a middle-income country Tiago Fonseca Francisco Lima Sónia Pereira February 1, 2014 Abstract Technology has been thought to favor skilled labor due to its complementarity. However, evidence from the 1990s shows technology benefiting both high-skilled and low-skilled, while middle-skilled labor is disappearing job polarization. Polarization has been identified in several OECD countries. We address polarization in an European country were firms have access to the same technologies than EU counterparts, but has lower wages and lower skilled labor: Portugal. Using a matched employer-employee dataset, we explore the evolution of employment and wages from 1986 to Using task measures computed from O*NET dataset, we test the routinization hypothesis technology intensive capital displaces workers in routine jobs. Our results show a polarization trend, especially during , in both wages and employment. Using measures of routine jobs, makes clear those jobs present a different nature regarding wages and employment. Wages and employment estimations show a rise in the relative employment and wages for abstract, while for routine cognitive employment decreases, but not wages. We argue that technological change interacts with changes in the firms internal organization, which can lead to this result, presenting a nuanced version of the routinization hypothesis, while exploring firm size and sectors dimensions. Keywords: technological change, routinization, job polarization, wages JEL codes: J24, J31, O33 Instituto Superior Técnico, Universidade de Lisboa, and CEG-IST Bernard College, Columbia University 1

2 1 Introduction In this paper we study job polarization in Portugal and its causes. Job polarization is a fairly recent phenomenon characterized by employment growth at both the bottom and the top of the income distribution. Job polarization has been observed in the U.S. (e.g., Autor et al. (2006) in the late 1980s and 1990s Acemoglu and Autor (2011)) and in the U.K. for the period between 1975 and 1999 (Goos and Manning, 2007). Polarization has also been documented in West-Germany (Spitz-Oener, 2006; Dustmann et al., 2009), Northern European countries (Asplund and Barth, 2011). In addition, Michaels et al. (2013) find polarization for several OECD countries. 1 Goos et al. (2009, 2010) also find evidence of polarization in Europe as a whole, but not in every single country 2. Using the European Labor Force Survey they found that in all countries the middle paying jobs decreased their employment share and the low-paid jobs were, for all countries (except Italy), growing or experiencing a small decrease. High-wage jobs increased their employment shares in all countries studied except in Finland, Ireland and Portugal. The decrease in the high-wage jobs employment share in these countries is not consistent with both job polarization or skilled biased technological change. While the reasons behind job polarization are still subject to some debate, the main candidate is technology and the so called routinization hypothesis (first put forward by Autor et al. (2003)) in which technological improvements allow machines to replace workers performing routine tasks. In particular, the increase in computer use in the workplace resulting from the declining price of computer capital has lead to a decrease in the demand for workers performing routine tasks, which tend to be middle skilled. According to this hypothesis, while computers substitute for middle skilled workers, they complement high skilled workers whose productivity increases with computer capital investment. The majority of workers displaced from middle skilled occupations, unable to secure high paid jobs requiring higher skills, take low paid jobs previously taken by low skilled workers, leading to a polarization of jobs in the economy. Several reasons could explain why certain countries may fail to show evidence of job polarization, despite experiencing technological change and computerization. First, institutional differences in pay setting could act as counteracting forces in the wage distribution which in turn may affect employment growth of different types of jobs. These include minimum wages and collective bar- 1 The study covers Austria, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Spain, the U.K. and the U.S.. 2 The countries studied are: Austria, Belgium, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Luxembourg, Netherlands, Norway, Portugal, Spain, Sweden and the U.K. 2

3 gaining agreements and are unlikely to largely affect the high paid jobs. Second, countries with a large public sector may also present a wage (and occupational) structure less permeable to market forces. Third, if preferences are non-homothetic, differences in the level and distribution of income across countries may result in differences in the occupational structure of employment (Goos et al., 2010). For example income elasticity of demand for services is thought to be greater than unity (see for example Clark (1957)). In the case of Portugal, its lower GDP per capita relative to the rest of Europe could result in a smaller service sector, which could translate into a different job structure. The timing of technology dissemination and job polarization may also differ across countries. For example, Continental Europe appears to have a lag of one decade in experiencing the labor market patterns of the Anglo-Saxon countries Spitz-Oener (2006); Dustmann et al. (2009), which could explain some of the differences found. Differences in the skill composition across countries could also lead to different job structures. For example a country with a highly skilled workforce could conceivably see most of its workers moving up from middle paid to high paid jobs as a result of computerization. A country with a less skilled workforce like Portugal, on the other hand, may have low skilled workers in some of its middle and even higher paid jobs to begin with, which could result in less clear cut job polarization, even if computers displace workers performing routine jobs. Another reason for lack of observed job polarization despite computerization is the direction of job offshorability. Countries that are recipient of offshorable jobs may experience different changes in their job structure than remaining countries. In view of the above discussion and the current state of the job polarization literature, the study of job polarization in Portugal is of particular interest. Are the market forces created by computerization likely to generate job polarization even in middle income countries with lower skills and wages and a less developed service sector? In this paper we use Quadros de Pessoal - the Portuguese firm census, to show that despite its lower wages, GDP per capita and share of the service sector, Portugal has experienced job polarization from the mid-1990s. This provides important evidence that job polarization is not a phenomenon exclusive of richer economies, but applies to middle-economies as well. Decomposing changes in employment between- and within-industries, we show that employment shifts within industries point to the role of technology as the main culprit behind job polarization in Portugal. In fact, using routine and non-routine task measures derived from the O*NET dataset, we show that the routinization hypothesis is a robust explanation for the employment and wage patterns observed. The rest of the paper is organized as follows. Section 2 reviews the extant literature on job polarization and discusses what different models predict in terms of employment and wage 3

4 changes and how those can be tested. Section 3 describes the data. Section 4 presents descriptive statistics for employment and wages between 1986 and Section 5 presents regressions results and section 6 concludes. 2 Literature review Since the early 1960s, the U.S. has seen the stock of capital equipment rapidly increase alongside a growing trend in the relative demand for skilled workers, suggesting the complementariness between capital equipment and skilled labor (Krusell and Ohanian, 2000). The skill premium - the wage ratio between skilled and unskilled workers - rose particularly rapidly since the late 1970s despite the increase in the supply of skills (Juhn et al., 1993). Moreover, the rate of within-industry skill upgrading was more concentrated in the most computer intensive sectors of the economy (Autor et al., 1998), suggesting that skilled labor and capital are complements. During the 1980s - the decade of the microelectronics revolution - less-skilled workers suffered a decrease in relative wages and demand for the less-skilled decreased in developed countries (Berman et al., 1994, 1998). 3. The increase in the demand for skills seems to have been pervasive in developed countries, with documented changes in the wage structure in many countries, especially the U.K. and U.S. (Machin and Reenen, 1998). The literature suggests that the changes in wages and employment between the 1960s and the 1980s are related to skills and technological change (Katz and Murphy, 1992; Levy and Murnane, 1992; Bound and Johnson, 1992; Juhn et al., 1993). In particular, computer capital seems to have a positive effect on the wages of those workers who use computers in the workplace (Krueger, 1993; DiNardo and Pischke, 1997). These facts have lead to the idea that recent technology developments have been biased towards the most skilled workers in the form of higher employment and wages - the skill-biased technological change hypothesis (SBTC hereafter). 4 When Autor et al. (2006) uncovered job polarization in the U.S. and (Goos and Manning, 2007) did the same for the U.K., it became clear that SBTC hypothesis could not explain the recent labor market patters alone (Card and DiNardo (2002) describe additional issues with the SBTC hypothesis). Autor et al. (2003, 2006) proposed a nuanced view of the impact of technical change in the labor market according to which technology changes the tasks performed by workers at their 3 There is some evidence of institutions explaining changes in the bottom half of the wage distribution (e.g., DiNardo et al. (1995)) 4 See, e.g., Krueger (1993), Berman et al. (1998), Machin and Reenen (1998), Autor et al. (1998), Acemoglu (1998), Bresnahan (1999),Krusell and Ohanian (2000). 4

5 jobs which in turn changes the demand for human skills. In their model technology and skills labor are still complements, but computer capital substitutes for workers performing routine tasks, i.e., tasks that can be programmed into a machine because they follow a set of well-determined rules. They show that the adoption of computers is associated with reduced labor of routine tasks (both cognitive and manual) and increased labor input of non-routine cognitive tasks within industries, occupations and education groups. 5 Building on Autor et al. (2003), Autor et al. (2006) propose a task-based model with three categories of tasks: abstract, routine and manual, where college workers perform abstract tasks and high school workers can substitute between routine and manual tasks. Routine tasks and computer capital are perfect substitutes. Because of the falling price of computer capital - exogenous to the model -, computers substitute routine workers, causing a reduction on its employment and, by definition, wages. 6 Also, non-routine cognitive task workers see their employment and wages rise as computer capital is, for those, a complement and therefore increases their productivity. Finally, non-routine manual task workers employment rise because of the influx of displaced workers from routine tasks, and the wage ratio between routine task workers and non-routine manual and interpersonal wages drops. Routinization hypothesis provides two main observable implications: changes in wages and in employment. But it does not impose any type of sorting of wages throughout the three tasks. Wages depend on the labor supplies and on the price of computer capital. Evidence shows that routine jobs are not at the bottom of the wage distribution, because those jobs require a certain amount of skills Goos and Manning (2007). Therefore we expect ceteris paribus decreasing employment in the middle of the wage distribution, as routine jobs tend to be substituted by technology. While in the US wage polarization has occurred hand in hand with job polarization Autor et al. (2006), in the UK, Goos and Manning (2007) failed to find wage polarization, despite strong evidence of job polarization. While they found that wages at the top of the distribution increased relative to the median, wages at the bottom did not rise relative to the median. A possible explanation given is that when hit by demand shocks, job changes do not occur randomly. In particular, Goos and Manning (2007) hypothesize that if those displaced from the middling jobs are less skilled, the average skill of those who remain increases, which could explain the observed wage patterns. In line with their hypothesis, they find evidence of greater educational upgrading in middle than in lower paid jobs. Selection of workers in the process of labor market polarization 5 See Spitz-Oener (2006) for similar evidence for Germany. 6 By definition, the price of computer capital is equal to the wage of routine workers. 5

6 renders the straightforward wage predictions in Autor et al. (2006) model empirically hard to test. Acemoglu and Autor (2011) propose an empirical strategy which uses demographic variables to construct skill groups and regresses the wage changes of these groups on their task measures at the beginning of the period. Their premise is that if the market price of the tasks in which a skill group holds comparative advantage declines, such as routine tasks, the relative wage of that skill group should decline, whether workers in that group change tasks or not. In this paper we use a similar empirical strategy to provide evidence of wage trends that are consistent with a task based explanation for polarization. 3 Data The main data source used is the Quadros de Pessoal (QP). The QP is a matched employeremployee dataset created by the Portuguese Ministry of Labor in the 1980s. Every, employers answer a survey with information on personnel and firm characteristics. The dataset covers Portuguese firms with at least one employee and excludes agriculture, military, public administration, institutionalized or self-employed workers. Using the QP, we have access to a longitudinal dataset, covering a period of 21 s, from 1986 to However, the ministry did not release the worker level information for 1990 and Firm level characteristics include: industry (NACE five-digit code), annual sales, location, legal form, the structure of capital (percentage of private, public and foreign capital), initial equity and starting date. For workers, we have gender, age, schooling, tenure, promotion dates, job level and occupation (ISCO five-digit codes), hours of work (regular and extra), wages (base wage, regular and irregular bonuses and payments for extra hours of work), type of contract (if permanent or fixed-term). Despite being a rich dataset, there are some problems of data inconsistency. In order to reduce data error, we impose some restrictions. We restrict our sample to workers aged between 16 and 65, both male and female. Workers must earn at least the minimum wage, where wage is the sum of base wage plus regular and seniority related bonuses. 7 We also restrict to full-time workers. All wages are deflated to the 1985 using the Consumer Price Index (CPI). Finally, we drop workers on both measures of occupation and industry associated with the primary sector, as the survey is optional for those. Also, we exclude workers in public administration 7 Portuguese minimum wage is set monthly. 6

7 and defense and education sector, because the model is suited for the corporate private sector. 8 Therefore, we excluded the industries (NACE) A, B, C, L, M and Q, and the occupations (ISCO) 11, 23, 33, 61 and Trends in employment and wages Employment and wage changes are the observable effects of polarization. Figure 1 shows the evolution in wages, measured at percentiles 10, 50 and 90 across time. We can observe a rise in inequality especially in the 1990s, as the top percentile rose sharply, while the middle and bottom of the wage distribution had a moderate increase.. Since 2000 the change in the distribution was much less pronounced, with an only slightly larger growth in the 90th percentile than in the other two. The fact that top jobs wages increased much more than middle and bottom ones is consistent with job polarization. [Figure 1 about here] Using employment data, we find clear evidence of polarization. Figure 3 plots the percentage change in employment during by job quality. Job quality can be measured in several ways. Ceteris paribus, a worker will arguably prefer higher wages and hence wages are a potential measure for job quality (Meisenheimer, 1998). We use the mean wage of each job in the base (1986). 9 To measure employment, we use the industry-occupation-region cells as representative of each job. Services comprise more jobs (more circles) when compared with manufacturing, consistent with the shift of the economy from manufacturing to services that occur in the mid 1990s. Also for services, middle (wage) jobs decrease consistent with the routinization hypothesis observed also for manufacturing, though not so pronounced.. The economy presents the same U-shaped curve, job polarization. [Figure 3 about here] 8 Education in Portugal is mainly public, in the dataset there are only the private schools. Furthermore, the health sector is also heavily public, but there are private hospitals, small clinics and diagnostic centers that do not belong to the National Health Service. 9 A similar procedure is used by Goos and Manning (2007), Goos et al. (2010) and Acemoglu and Autor (2011). 7

8 5 Occupational groups The discussion of job polarization and its causes, immediately creates the need for occupational level discussion. As so, we consider seven occupational groups. Top waged occupational groups: 1. Managerial and health professions, 2. Technical and professional; middle waged occupational groups: 3. Office clerks, 4. Operators; bottom waged occupational groups: 5. Sales, ticket clerks and other services, 6. Personal and protective services and 7. Routine operators. 10 In Figure 4 we plot the evolution the the real wage of top and middling occupations and bottom occupations. The vertical line identifies the change in the periods (and occupational codes). We can see top occupations in the first periods increase their wages as we expect, however in during the 2000s Technical and professional curve flatten and Managerial and health professionals even decline. The decline is attributed mainly to Managers. There mean wages decrease from 2000 to 2002 sharper than the remaining occupational group, anticipating 2003 depression. This is specially driven by managers wages, which decrease is substantial. Bottom occupations, present almost null real wage growth. There is a timid increase in wages of Sales, ticket clerks and other services and Routine operators, contrasting with a small decrease on Personal and protective services groups. It is not clear if Routinization hypothesis plays a role in this patterns. As so, we need to use more sophisticated methods in order to address this changes in wages. Regarding employment, Figure 5, the top occupations employment share presents a up warding pattern, specially in the second period. Office clerks employment share increased slightly in the first period, however in the second period the pattern is clearly down warding. Operators employment share slope is negative overall. Looking for bottom occupations, there is a sharp decline in Routine operators, while the remaining bottom occupations are increasing their employment shares. The results are in line with the Routinization hypothesis, the top occupations that are mostly nonroutine cognitive are increase, while routine intensive occupations are decreasing there shares (e.g., office clerks, routine operators). Occupations that are non-routine and non-cognitive those that require flexibility and in-person contact are increasing there employment shares. However, some of this results can be driven by changes in industrial composition. In next section, we decompose this changes according to between and within industry composition effects. [Figure 4 about here] 10 See Appendix Table A1 for more details on the groups. 8

9 [Figure 5 about here] 6 Industrial decomposition The industrial structure shifts can contribute to the observed polarization. If industry shifts in a way where industries that are more intensive in manual (e.g., construction workers) and analytical/interpersonal (e.g., managerial) occupations grow and routine (e.g., office clerks) task intensive industries decrease, job polarization will occur. Hence, industry changes can actually be shaping the results. Also, we are looking for within-industry changes, rather than between industry shifts. To test whether industrial change can be pointed out as an explanation for job polarization, we perform a standard shift-share decomposition (Acemoglu and Autor, 2011). 11 We decompose the total change in employment share of each occupation j over time interval t ( E jt ) into two parts: the first is the change in occupational employment share due to changes in the industry shares or between industry ( Ejt B ); the second captures the change in employment share result due within-industry shifts ( E W t ). Equation 1 express this relationship. E jt = E B jt + E W jt (1) The changes in employment from equation 1 can be computed as expressed in equations 2 and 3. E B jt = k E kt λ jkt (2) E W jt = k λ jkt E kt (3) In equation 2, E kt = E kt1 E kt0 being the change in employment share for industry k over time interval t = {t 0 ; t 1 } and λ jkt = (λ jkt1 + λ jkt0 )/2 is the average employment share in industry k for occupation j over the same time interval. Similarly, for equation 3, λ jkt = λ jkt1 λ jkt0 is the variation of industry k s employment share 11 An alternative approach to calculate shifts in the industry can be found in Berman et al. (1994). 9

10 for occupation j during the time interval t and E kt = (E kt1 + E kt0 )/2 is average employment share of industry k in the given time interval. We decompose employment changes for two time periods: and (Figure 6). We are particularly interested in changes that occur within industry, as between industry changes can mask the effect of technology in firms. Overall between changes are mostly related with the economy shift to services. Manufacturing occupations are declining, while services related occupations are growing. Top occupations shares rise in both periods, especially in the second. Bottom occupations office clerks and operators see their within shares decrease. Even though office clerks overall share increase in the first period, the result is due to between industry changes. For bottom occupations, the changes in employment varies according to the period. In the first period, overall within changes are positive, especially routine operators. Conversely, in the second period, the more routine intensive task occupations decrease the share: sales, ticket clerks and other services and routine operators. Personal and protective service are increase their within share. This occupational group is mainly intensive in manual tasks. The within and between industry employment changes are broadly consistent with job polarization due to technology shifts. [Figure 6 about here] 7 Task measures 7.1 O*NET database We use is the Occupational Information Network (O*NET) database as an additional source of information concerning the tasks performed by workers. The O*NET database is the primary project of the O*NET program promoted by the U.S. Department of Labor. The database provides information about occupations in several dimensions. The framework behind O*NET is its Content Model that establishes what are the most important pieces of information regarding work, defining which set of measurable variables (descriptors) has to be collected. Data are collected by several federal agencies (e.g., Bureau of Labor Statistics). The collected information is then processed and several descriptors are created for each occupation. The use of the O*NET database is particularly relevant to establish the task measures that we will use. The O*NET database (version 15.0) codify 854 occupations using SOC codes which we had to convert to ISCO codes, as we only have occupations in ISCO (2-digits). We aggregate the 10

11 854 SOC codes into 27 ISCO codes. Our conversion and aggregation of the codes was done using the 2007 U.S. employment data from the U.S. Bureau of Labor Statistics as weights and a O*NET SOC to ISCO crosswalk. 12 In the end we have the corresponding O*NET descriptors for each ISCO occupation. 7.2 Occupational task measures definition In order to establish the task content of each occupation s measures as well, we use the same O*NET descriptors as Acemoglu and Autor (2011). we select the O*NET descriptors that have importance and context scales, both continuous between 1 and 5 (none used are represented in both scales). 13 We apply principal components, by task type, to the O*NET descriptors to reduce the dimension of the descriptors. The use of principal components enables us to lose as minimum information as possible. By construction, the principal components maximize the total variability of the original data. Each occupation is now represented by a set of task measures. Table 1 summarizes the results obtained, displaying the standardized principal components (mean 0 and standard deviation 1) for each task measure and the total variability of the original data explained by the principal component. As expected, managerial, science and health-related occupations are the ones that require the most analytical skills. Physical, mathematical and engineering science professional are the most analytical-intensive occupation, with 1.58 scale points. There is also high intensity of interpersonal task in the top occupations, especially in managers (1.31 points) and life science and health professionals (1.61). By contrast sales and services elementary occupations (e.g., housekeeping) are the ones that require less analytical skills (-1.89 points), but require the most manual skills (2.95 points). At the bottom of interpersonal skills importance measure are the elementary occupations (-0.97 points). Office clerks have the with a higher measure in routine cognitive importance (1.02), while machine operators and assemblers are, as expected, the ones with the highest importance in routine manual (2.15 points), as result of their repetitive work. Most occupations with low intensity in 12 The U.S. employment dataset can be found in BLS data are only available for the 6-digits SOC, while O*NET uses 8 digits. To overcome the problem, we aggregate for each occupation (reduce to 6-digits) the O*NET measure by its median. SOC to ISCO crosswalk was kindly provided by Maarten Goos, Alan Manning and Anna Salomons to whom we are thankful. The original crosswalk was just updated for some newer O*NET SOC codes U.S. employment data were used as weights for a mean in order to aggregate the data, as we do not have Portuguese employment data in SOC codes. Using a weighted mean, we guarantee that an occupation with fewer workers weights less than a large occupational group. The weighted mean allows us to aggregate the 854 SOC O*NET descriptors into just 27 ISCO O*NET descriptors. 13 A detailed list of the descriptors by task type is in Appendix Table A2. 11

12 manual task also have low values in routine manual. When the routine manual measure approaches the mean (0 points) the relationship is not as straightforward. For instance, metal, machinery and related trade workers are highly intensive manual tasks (1.81 points - the second highest value), but has a negative value on routine manual tasks (-0.02). [Table 1 about here] 8 Task importance across time The constructed task measures constitute a vital piece of information to test the routinization hypothesis. In order to assess its importance along the s, we compute the mean task points by each. The results are plot in Figures 7. The results show a sharp reduction in the importance of manual and routine manual. Also, routine cognitive importance exhibits a decreasing in its importance. By turn, analytical and interpersonal task importance as increased strongly along the s, in particular interpersonal in the 2000s. Overall the results are in line with routinization hypothesis: routine tasks being decreasing their importance, while the remaining task increase. Manual task importance does follow routine trends, however this can be explained as consequence of economy shifting from manufacturing to services. More importantly, routine cognitive importance, that is close to the services, contradicts the pattern, exhibiting a clear decreasing pattern. [Figure 7 about here] There is not a direct correlation between every task and the wages. We can show that by looking for each task and the wage distribution. Figure 11 plots the result of changes in task intensity by wage deciles, which corresponds to the bars, and the task intensity in 1986 by each wage decile. So, adding the bar and the dot corresponds to the task intensity in The results are very much what our intuition would propose as a result. Analytical tasks are more intensive for workers at the top of the wage distribution. Moreover, the intensity has change in the way, where the top waged are strong in doing such tasks. Interpersonal should be more intensive at the top (e.g., managers) and at the bottom (e.g., sellers). The result on interpersonal shows precisely that, emphasizing that this polarization has been growing. Regarding routine tasks, routine cognitive wages are, in the beginning, more concentrated at the middle-top of the wage distribution. However, the top decile is decline, as well as the bottommiddle. This can be due to selectivity, the most able workers (and most waged) remain, while 12

13 the rest is substituted by computer capital or cheaper labor. Contrary to routine cognitive tasks, routine manual are concentrated at the bottom of the wage distribution. The moving picture for routine manual is very different. First there is a strong decline in routine manual intensity along the wage distribution. The results is in line with the the routinization model: as technology prices decline, the relative price of routine labor also decrease. Manual changes are very similar to routine manual: decrease along all wage deciles. However, much of manual more intensive task are concentrated at the middle of the wage distribution. The output is very much in line with the routinization hypothesis. Increase of workers at the top of the wage distribution due to analytical and interpersonal (cognitive) increases. Increase of workers at the bottom of the wage distribution due to interpersonal (non-cognitive) and squeezing prices of routine tasks, while the while middle is depopulated as results of routine tasks middle waged prices decrease. [Figure 11 about here] 9 Regression results We consider until here a continuous approach to tasks, where each occupation is endowed with a set of tasks. Although this classification can be very useful examining the evolutions of tasks in the labor market, it does not provides the most tractable way of dealing with tasks. For our regressions we assign tasks to occupations. To allocate task into occupations we look for the most intensive task within an occupation. Thus we assign a task i to an occupation j if the task measure i is greater than task measure k, with k i. 14 In addition this, we aggregate analytical and interpersonal in just one category as they exhibit similar patterns and are closely related. We call to this task Abstract, which is similar to Autor et al. (2006) terminology. The finally aggregation can be find in Appendix Table A3. Using allocation of occupations to tasks, we plot employment share and log wages evolution for our four broad occupation groups based on tasks: abstract, routine manual, routine cognitive and routine manual. We do this for for the economy, manufacturing and services (Figure 9). For the economy, noticeably most workers are manual (approximately 50%), while the second group is routine cognitive. Those two groups employment shares do not change pronouncedly during 14 The process is straightforward for all tasks, only Life science and health associate professionals (ISCO 32) has to be assign as interpersonal by us. 13

14 the considered period, but the same does not happen with routine manual and abstract. The employment share for abstract is rising sharply, especially in the 2000s, while routine manual is losing its share at a similar pace. Looking for the graphs for the manufacturing and services, we can understand that routine manual decrease is pushed by the manufacturing firms, while the rise of abstract is especially sharp in the services. Moreover, there is a decline in routine cognitive for the manufacturing, which is balanced out by a slightly increase in the services - services share of routine cognitive is substantial larger than in the manufacturing. Regarding wages. Abstract is rising more than the remaining tasks, widening the gap between abstract and the rest. I worth of note a decrease in manual relative to routine manual (and consequently the remaining tasks), which contradicts what the routinization model predictions. However, confounding factors can always distort the results, so that we have to account for those. [Figure 9 about here] The same graphs are plotted by firm size in Figure 10. As before, manual dominates the remaining tasks in share, despite the firm size. Manual occupations grow from firms with less than 50 workers, while for firms with more than 50 workers the trend is negative. Routine manual employment presents a sharp decrease for firms with more than 50 workers, while for other firm sizes the trend is almost flat. The other routine workers, routine cognitive, have seen their employment grow or stabilize for firms with less than 249 workers, but larger firms (more than 249) routine cognitive employment is decreasing at a similar pace as routine cognitive. The abstract rise observable to firm sizes, however for the larger firms (more than 249 workers) the up warding trend slope is quite low. For wage patterns are similar to those found in the earlier Figures: abstract rising slightly more than the rest, while manual decreases in the 2000s for the larger firms. [Figure 10 about here] Additionally to the graphs, some descriptive statistics are presented in Table 2 for the pool of workers between 1986 and We group the statistics by sector (all, manufacturing and services) and by firm size (1 to 9, 10 to 49, 50 to 249 and more than 249 workers). The Table shows that the services employ more females (44%). Larger firms have on average older employees and higher wages. The average wage is also higher for services firms. Finally, it is worth of note the very low schooling level of the Portuguese workers. Only 8% have a college degree, while most (60%) have less than 9 s of schooling. [Table 2 about here] 14

15 With tasks assigned to occupations, we build a simple empirical model based on cells to test if routinization hypothesis can explain the observed changes in wages and employment. We consider cells defined by workers demographics, industry and firm size (when applicable). Using cells defined as such can mitigate the problem of self-selection of workers into occupations and consequently into tasks. Additionally, cells can net out the compositional effects, thus bolding the associations between employment, wages and task composition. Given the cell as the unit, let the natural logarithm of employment in a cell or natural logarithm of the mean wage in a cell be denoted by Ω. Our research focus relies on what is the impact of task composition T on Ω over time t. Thus, the empirical model is given by: Ω i = T it0 θ + T it0 t i θ + t i α + X i β + ɛ (4) Where t represents a set of time dummies and X i is a fixed-effect for cell i. Task composition T ij measures the share of workers in task j for cell i for the base t 0 We generate gender-age-education-region-industry cells. For each cell we compute the share of workers allocated to each task (abstract, routine cognitive, routine manual and manual). As being shares we must omit one, in this case, we omit manual. Our interest variables are the interaction effects, so that we can capture what is the results on employment and wages for the individuals (defined as a cell) that hold comparative advantage in 1986 in a given task, netting out compositional effects and between industry movements. We start by estimating the model without the interactions effects and not fixing tasks, so that the formulation is given by: Ω i = T i θ + t i α + X i β + ɛ (5) The estimations are summarized in Table 3 for employment and Table 4 for wages. The results are segmented by economy (all firms), manufacturing, services and firm size (four categories). The omitted task is manual. Overall results show that for the economy manual employs most workers, followed by routine manual. Abstract and routine cognitive are very similar employment wise, when we net out cell s compositional effects. For the services, routine manual is the most important task regarding employment, while this holds true to all firm sizes, expect for firm with 50 to 249 workers. For wages results confirm that abstract workers earn more in average, followed 15

16 by routine cognitive. This holds for sector or firm size. Routine manual workers wages are bellow manual workers wages. [Tables 3 and 4 about here] The next step is naturally run the model from Equation 4. For simplify interpretation, we plot the interactions coefficients in Figures 12 by sector and 13 by firm size. The full results can be find in Appendix Tables YYYYY. The results by sector (Figure 12) show a clear rise in employment for those individual who had comparative advantage in 1986, relative to those that are in manual (the omitted group). By the same token, routine cognitive is decreasing for all sectors, hypothesized by routinization models (e.g., Autor et al. (2006)). Routine manual presents a slightly up warding pattern, specially for the manufacturing, which contradicts the routinization hypothesis. Regarding wages, abstract exhibit an upward trend for all specifications, with some decline after 2003 crisis. The same does not holds for the remaining tasks. Routine cognitive workers have seen their unchanged in the manufacturing, while climbing in the services. By turn, despite the growing trend in employment for those workers that possess comparative advantage in routine manual in 1986, there wages are declining, specially for those in manufacturing firms. This results are robust to specifications with tasks not fixed or with firm size in the cell. For the results by firm size (Figure 13), the decline of routine cognitive is pervasive for all firm sizes. The same hold for the rise in abstract employment and wages. Again routine manual almost mimics abstract workers employment wise. However, for the wages the results are different. Those workers that had comparative advantage in routine cognitive tasks have seen their wages unchanged relatively to manual counterparts, unless those workers are performing their jobs in the smaller firms. The workers in routine manual also have seen wages are dropping, specially if they are not working in large firms (more than 249 workers). Once again, our results are robust to changes in cells and tasks not fixed in Our results are in line with what routinization hypothesis predicts, however two questions that are not explained by the model emerge. The first is why is the relative price of routine cognitive to manual rising in the services; and the second is why the same price is rising for smaller firms (1 to 9 workers). Figure 11 can give us some support to explain the first question. We find that routine cognitive tasks are becoming more intensive in the top wages, while routine manual or manual do not. Thus, as routine cognitive task are substituted by computer capital, the remaining workers will be the better ones and as consequence more waged, because still computer capital cannot substitute some highly trained routine cognitive works. Specially in the services, divisional manager is very 16

17 much associated which those tasks. So, if indeed there is reorganization of the workplace due to nonneutral technological change (Lindbeck and Snower, 2000) and elimination of middle management (flatter firms) (Rajan and Wulf, 2006; Guadalupe and Wulf, 2010). Middle middle management (intermediate managers) are substituted by computer capital and consequently divisional managers get more responsibility and span of control, which by turn results in an increase in wages. The same reasoning can be applied to office clerks: the best ones will address non routine problems that computer capital is not able to solve. In addition, top managers span of control increases, while divisional managers gain more responsibility. The sharp increase in abstract wages reflects just that, bolding the complementarity between those workers and computer capital. The technological argument hold for smaller firms. As smaller firms have less incentives to invest in technology that substitutes routine tasks (Cohen and Klepper, 1996), those firms will continue to employ the most productive workers (better paid) that can perform the routine tasks that are essential for their businesses. Thus employment tend to decrease, but the most routine task performers remain and get more rents from their employers, because of their comparative advantage. 10 Conclusion In this paper we have shown that since late 1980s, the Portuguese labor market exhibits job polarization since late 1980s. In addition, we show that job polarization is consistent with the routinization hypothesis,where technology changes replace routine tasks and complement abstract tasks. The importance of both non-routine cognitive skills (analytical and interpersonal) are increasing, but routine and manual skills importance is being reduced. We also add a new result to the literature on job polarization, by showing that employment and wage changes took different forms in smaller versus larger firms, with larger firms contributing stronger to the polarization of the Portuguese labor market. This result is likely to have significant implication for policy. 17

18 References Acemoglu, D. (1998). Why do new technologies complement skills? Directed technical change and wage inequality. Quarterly Journal of Economics 113 (4), Acemoglu, D. and D. Autor (2011). Chapter 12 - Skills, Tasks and Technologies: Implications for Employment and Earnings. Volume Volume 4, pp Elsevier. Asplund, R. and E. Barth (2011). Polarization of the Nordic Labour Markets. Finnish Economic (2), Autor, D., L. Katz, and A. B. Krueger (1998). Computing inequality: have computers changed the labor market? Quarterly Journal of Economics 113 (4), Autor, D., F. Levy, and R. Murnane (2003). The Skill Content of Recent Technological Change: An empirical exploration. Quarterly Journal of Economics 118 (4), Autor, D. H., L. F. Katz, and M. S. Kearney (2006). The polarization of the US labor market. American Economic Review 96 (2), Berman, E., J. Bound, and Z. Griliches (1994). Changes in the Demand for Skilled Labor Within U.S. Manufacturing: Evidence From the Annual Survey of Manufacturers. Quarterly Journal of Economics 109 (2), Berman, E., J. Bound, and S. Machin (1998). Implications of Skill-Biased Technological Change: International Evidence. Quarterly Journal of Economics 113 (4), Bound, J. and G. Johnson (1992). Changes in the structure of wages in the 1980 s: an evaluation of alternative explanations. American Economic Review 82 (3), Bresnahan, T. F. (1999). Computerisation and Wage Disperson: an Analytical Reinterpretation. The Economic Journal 109 (456), Clark, C. (1957). The Conditions of Economic Progress. London: Macmillan. Cohen, W. and S. Klepper (1996). Firm Size and the Nature of Innovation Within Industries: The Case of Process and Product R&D. Review of Economics and Statistics 78 (2), DiNardo, J., N. Fortin, and T. Lemieux (1995). Labor Market Institutions and the Distribution of Wages, : A Semiparametric Approach. Econometrica 64 (5), DiNardo, J. E. and J.-S. Pischke (1997). The Returns to Computer Use Revisited: Have Pencils Changed the Wage Structure Too? Quarterly Journal of Economics 112 (1),

19 Dustmann, C., J. Ludsteck, and U. Schönberg (2009). Revisiting the German Wage Structure. Quarterly Journal of Economics 124 (2), Goos, M. and A. Manning (2007). Lousy and lovely jobs: The rising polarization of work in Britain. The Review of Economics and Statistics 89 (1), Goos, M., A. Manning, and A. Salomons (2009). Job polarization in Europe. American Economic Review 99 (2), Goos, M., A. Manning, and A. Salomons (2010). Explaining Job Polarization in Europe: The Roles of Technology, Globalization and Institutions. CEP Discussion Papers (1026). Guadalupe, M. and J. Wulf (2010). The Flattening Firm and Product Market Competition : The Effect of Trade Liberalization on Corporate Hierarchies. American Economic Journal: Applied Economics 2 (4), Juhn, C., K. M. Murphy, and B. Pierce (1993). Wage Inequality and the Rise in Return to Skill. Journal of Political Economy 101 (3), Katz, L. F. and K. M. Murphy (1992). Changes in Relative Wages, : Supply and Demand Factors. Quarterly Journal of Economics 107 (1), 35. Krueger, A. B. (1993). How Computers Have Changed the Wage Structure: Evidence from Microdata, Quarterly Journal of Economics 108 (1), 33. Krusell, P. and L. Ohanian (2000). Capital-skill complementarity and inequality: A macroeconomic analysis. Econometrica 68 (5), Levy, F. and R. Murnane (1992). U.S. Earnings Levels and Earnings Inequality: A Review of Recent Trends and Proposed Explanations. Journal of Economic Literature 30 (3), Lindbeck, A. and D. Snower (2000). Multitask learning and the reorganization of work: from Tayloristic to holistic organization. Journal of Labor Economics 18 (3), Machin, S. and J. V. Reenen (1998). Technology and changes in skill structure: evidence from seven OECD countries. Quarterly Journal of Economics 113 (4), Meisenheimer, J. (1998). The Services Industry in the Good versus Bad Jobs Debate. Monthly Labor Review 121, 22. Michaels, G., A. Natraj, and J. V. Reenen (2013, April). Has ICT Polarized Skill Demand? Evidence from Eleven Countries over 25 Years. Review of Economics and Statistics,

20 Rajan, R. and J. Wulf (2006). The flattening firm: Evidence from panel data on the changing nature of corporate hierarchies. Review of Economics and Statistics 88 (4), Spitz-Oener, A. (2006). Technical change, job tasks and rising educational demands: Looking outside the wage structure. Journal of Labor Economics 24 (2), Cumulative log wages change th percentile 50th percentile 90th percentile Figure 1: Evolution of cumulative changes in wage percentiles, 1986 to

21 Relative log wage change Wage Quantile Figure 2: Relative to the median changes in wage distribution, 1986 to 2007 Manufacturing Services Total Change in log employment ( ) log mean real wage (1986) Figure 3: Change in log employment over the period from 1985 to 2007 by job quality Job quality defined as the log median wage for each occupation-industry-region cells (generating 2,919 cells). The circles represent the initial employment share for each cell and the line corresponds to the quadratic-fit weighted by the initial employment. 21

22 Mean real wage Top and middling occupations Mean real wage Bottom occupations Technical and professional Managerial and health professionals Personal and protective services Sales, ticket clerks and other services Office clerks Operators Routine operators Figure 4: Wages by occupational group, 1986 to 2007 Top and middling occupations Bottom occupations Employment share (%) Employment share (%) Technical and professional Managerial and health professionals Personal and protective services Sales, ticket clerks and other services Office clerks Operators Routine operators Figure 5: Employment share by occupational group, 1986 to

23 Broad Group Managerial and Health Professionals Technical and Professional Office Clerks Operators Sales, ticket clerks and other services Personal and Protective Services Routine operators Between Within Figure 6: Industrial decomposition for the economy by occupational groups Note: Each bar represents the change in employment share (in percentage points) of the corresponding occupational group. Mean task measure Analytical Routine Cognitive Manual Interpersonal Routine Manual Mean Offshorability Figure 7: Mean task points by Figure 8: Mean offshorability points by 23