The Impact of Automation on Job Polarisation

The impact of automation on job polarisation. An empirical investigation.

Abstract

This paper examines the effect of automation on changes in task contents and returns to tasks within occupation, which receives less attention in the literature as it typically assumes occupation tasks to be static. Using British Skill Survey 1997-2012, my results support the routinisation hypothesis firstly proposed by Autor, Levy and Murnane (2003). Individual shifts towards abstract tasks and away from routine tasks within occupation due to computerisation. I also find a negative premium on routine tasks and positive premium on abstract tasks. Return to managerial tasks has increased over time, indicating the growing importance of tasks that cannot be automated.

Word count:

1. Introduction

The race between technology and labour has long been a debate since the Industrial Revolution. With the rise of automation and artificial intelligence, the concern on whether computer can eventually replace human has reemerged recently. A widely cited study by Frey and Osborne (2013) estimate 47 percent of total US employment is at high risk of being replaced by computers, while the Bank of England estimates 15 million UK jobs are under threat due to automation in the next twenty years.

Autor et al. (2003) are the first to propose the routinisation hypothesis, by incorporating the ability of technology in automating some tasks more than others. As price of computer has decreased significantly, they are increasingly used to substitute routine tasks, which follow explicit rules and easy to automate. On the other hand, the greater intensity of computer usage complements abstract tasks in which high-skilled workers have comparative advantages in. Occupations at the bottom of the wage distribution are least affected by technology and absorb the displaced workers. Automation therefore decreases demand for routine tasks that are usually performed by middle-skilled workers and raises demand for high-skilled and low-skilled, generating the U-shaped job polarisation pattern emerged in developed countries since 1990s[1]. This routine-biased technological change (RBTC) framework can also explain changes in the wage distribution, as routinisation alters task prices and affects returns to different occupations over time.

Previous research has found evidence of employment polarisation at the occupation, regional, industry and country level[2]. This paper instead focuses on the effect of technology on job contents and returns to tasks within occupation (intensive margin), which receive less attention in the literature. In an occupation, workers often conduct several tasks, some more routine than others. With technological progress, automation replaces the occupation partially by substituting the more routine parts, but not entirely. Workers reallocate their skills to nonroutine tasks that they have comparative advantages in compare to computers, and receive productivity gain if those tasks are complemented by technology. Analysing changes within occupation is of great interest, as assuming tasks contents to be static underestimates the amount of task reallocations taking place and overestimates jobs automobility[3].

The novel aspect of this paper is to use pseudo-panel to remove endogeneity due to self-selection. I group together individuals with the same occupation, education and age to form a cohort, as they have similar comparative advantages and cohort fixed effect can remove time-invariant unobserved heterogeneity. Much of the existing literature looks at employment or wage separately but the aim of this paper is to examine both aspects. Using British Skill Survey 1997-2012, I firstly consider if greater computer intensity substitutes routine tasks and complement nonroutine tasks. My results show there has been a shift towards abstract tasks and away from routine tasks. Other supply side factors such as education upgrade and changes in age or gender composition cannot explain this finding. Secondly, I analyse the impact of routinisation on task prices and find evidence that workers receive a positive premium if they perform more abstract tasks and a negative premium if they perform more routine tasks. The main results are robust to alternative specifications of OLS, interaction terms and industry fixed effects. I also go further by exploring recent trends, and find evidence suggesting technology is increasingly capable of substituting routine tasks, and is also starting to replace some abstract tasks. Returns to tasks that do not require human interaction has fallen, while return to managerial tasks in which human still has comparative advantages has risen.

The paper is organised as follows: Section 2 provides a literature review on the use of task framework in explaining job polarisation and wage inequality. Section 3 presents the theoretical model. Section 4 describes the data and recent trends in UK labour market. Section 5 presents results and Section 6 considers different robustness tests. Section 7 concludes.

2. Literature review

The development of routine-biased technological change (RBTC) is motivated by the inconsistency of skill-biased technological change (SBTC) in explaining recent changes in labour markets[4]. Autor et al. (2003, 2006) are the first to consider a task framework by highlighting the difference between tasks and skills. They argue technology affects demand for workers heterogeneously depending on the tasks they perform. Using job description from Dictionary of Occupational Titles (DOT), they find evidence that computerisation is associated with reducing in routine task inputs and increasing nonroutine task inputs within industries, within industry-education and within occupation. With the observation of declining middle-skilled occupations that have high routine content have declined.  Similar conclusions are reached by Goos and Manning (2007, 2009) by using UK and Europe data.

Since the influential work by Autor et.al (2003), there has been a growing literature that established routinisation to be driving labour market changes. Using hours work as a measure of labour demand in 16 European countries, Goos et al. (2014) show that routinisation shifts employment away from industries that use routine employment more intensively, leading to within and between industry job poliarsiation. A distinct feature in job polarisation is the growth of low skills job, which is examine by Autor and Dorn (2013). They consider variation between initial task composition across 722 US commuting zones, and find that commuting zones that initially specialised in routine tasks have greater reduction in routine task intensive jobs, more likely to have low skill labour reallocate from routine into service occupation, and strong evidence of employment and wage polarisation. The reallocation in service sector in particular in driven by xxx. As with Goos et al (2014), they also consider alternative explanations including offshoring, incomeinequality and find routinisation to be the most significant in driving observed changes in labour market.

The major limitation with these studies is they implicitly equate tasks with occupation by assuming task contents within an occupation are static over time. This paper goes one step further by studying RBTC within occupation. Although Autor et al. (2003) consider changes within occupation, their analysis is limited by availability of data, as DOT are not updated regularly. Apart from British Skill Survey, the only other dataset that contains detailed occupational information is German Qualification and Career Survey, which is employed by Spitz-Oener (2006). By regressing computer indicators on job tasks using first difference, her results support RBTC that changes in task contents are most rapid in occupation that has rapid computerisation. This paper follows a similar approach, but use fixed effects for more efficient estimates and control for cohort time-invariant heterogeneity using pseudo-panel. I also go beyond merely investigating changes in job contents, by considering how technology affect returns to tasks.

Despite employment and wages are closely linked, the literature provides no conclusive evidence on the effect of automation on wages under the task framework. Whilst there is a clear pattern of wage polarisation in the US (Autor et al. 2006), this is not found in any other countries (GOos and Manning, 2007; Dustmann et. al 2009). This has led to criticism of RBTC (Salvatori 2015, Mishel et al., 2013) who argue there is little relationship between changes in employment and wage structure. The main difficulty theoretically is to differentiate between returns to skills and returns to tasks. Simply regressing wages on tasks following the Mincer equation will lead to biased results, as individual will move endogenously into different occupation to maximise their earnings when task prices change. This critic can be reconciled by recognising the difference between returns to skill and returns to task. Acemoglu and Autor (2011) extend the task framework to a general equilibrium and consider the dynamics between tasks, skills and wages. In this Ricardian-type model, workers allocate skills to tasks endogenously based on comparative advantage, and this allocation is subjected to change when there is a shift in task prices. RBTC does not necessarily lead to wage polarisation, as it depends on tasks complementarity with technology, demand elasticity and labour supply (Autor 2013). On one hand, middle skill workers receive lower wage as they compete with technology in providing routine tasks, or become less productive when they shift to other tasks. However, it is also possible for their wage to increase, since only the most productive workers can compete with automation.

Using initial occupation specialisation as proxy for comparative advantage, they show that as price of routine tasks decrease due to computerisation, wages of workers with initial comparative advantage in routine tasks decrease while wages for those with initial comparative advantage in abstract or manual tasks increase. As automation replaces routine tasks, workers that initially perform routine tasks reallocate to tasks they have lower comparative advantage in, leading to a relative wage decline. On the other hand, the greater intensity of use of the newly automated tasks due to lower costs complements the remining tasks performed by labour, leading to a relative wage increase for these groups of workers.

Earlier research that investigates impact of computer on wages often run into endogeneity problems. While Krueger (1993) find significant wage premium of 10-15 percent associated with computer usage, DiNardo and Pischke (1997) find a similar magnitude of wage premium on the use of pencil and argue the computer premium only reflects unobserved ability. Using individual fixed effects, Entorf et.al (1999) confirms this argument, as they find that computer users earn a higher wage even before they start using computers. However, these results do not mean computers have no effect on wages. Computers alter workplace organisation and demand for various skills. Even computer itself do not carry any wage premium, it increases demand and therefore returns of tasks that are complements of computer. This is address in this paper by examining the effect of computerisation on wages using the task framework. A number of researches have consider alternative methods in testing the relationship between technology and wages.  Acemoglu and Autor (2011) use initial occupation specialisation as proxy for comparative advantage, and show that as price of routine task decreases due to computerisation, wages of workers with initial comparative advantage in routine tasks also decrease. Firpo et al. (2011) argue that if task prices decrease, the mean and variance of occupational wage that specialised in that task also decrease. Their decomposition analysis finds evidence that changes in task measures due to technology is significant in explaining occupation wage structure in the US 1980-1990.

There is a growing literature in estimating returns to different tasks. Weinberg (2014) analyses the wage premia of social skills by comparing labour market outcomes for two high school cohorts of 1972 and 1992. Her research reveals a positive and growing premium in occupations that requirw both high level of cognitive and social skills has grown. Borghan et al. (2014) and Demning (2016) also find similar trends using more recent data. In contrast, Beaudry et al. (2016) argue demand for cognitive tasks since 2000 by developing a model in which SBTC causes a boom and bust in the demand for cognitive tasks. In analysing returns to tasks and recent trends, this paper follows a similar approach as Dickerson and Green (2004) in the use of pseudo-pane, but extend their analysis by testing the routinisation hypothesis more directly and use more recent data 1997-2012.

3. Theory

The model presents below is based on Autor et al. (2006) and Acemoglu and Autor (2011), but I relax the assumption of law of one price by not restricting wages to equalise across occupation condition on skills to allow flexibility. In contrast to their models, I focus on the intensive margin, in which workers shift to perform different tasks within the same occupation. This narrows down the effects of automation on task contents and task prices.

Tasks are defined as work activities. Following the routinistion literature, tasks are split into three broader groups: abstract, routine and manual, based on three assumptions that are commonly observed in the workplace:

1. Computers are more substitutable for routine tasks than abstract or manual tasks.
2. The three tasks are imperfect substitute. Routine tasks are complement to abstract tasks, and less so for manual tasks.
3. The three tasks are q-complements. Greater intensity in one task increases the marginal productivity of the other two tasks.

In a static economy, each individual i equipped with a number of skills and chooses occupation j that consists of a bundle of tasks based on comparative advantage. Each occupation produces an occupation-specific output

Yj, which are then combined to produce a unique final good according to the production function:

Q=F (Y1,…, Yj)

In each occupation,

Yjis produced by using three tasks: abstract (A), routine (R) and manual (M). Assume a Cobb-Douglas production function,

Yj=AαRβMγ,        α+β+γ=1

There are two factors of production in the economy, labour and capital. Assume abstract and manual tasks can only be produced using labour, while routine tasks can be produced using labour or computer capital that are perfect substitute. Production function can be written as,

Yj=LAα(LR+C)βLMγ

where

LKrepresents labour input in the three tasks

K∈A,L,Mand C is computer capital, all measured in efficiency units.

Under perfect competition, each task is paid according to its marginal product, shown by the first order condition:

πA= αAα-1RβMγ

πR= βAαRβ-1Mγ=p

πM= γAαRβMγ-1

where

πKis the price per efficiency unit of task k and p is price of computer capital. As computer is a perfect substitute to routine task,

πR=p.

Supply of factors

Computer capital is supplied perfectly elastically at price p. Labour supply is characterized by the Roy model. Each worker is endowed with a vector of three skills

Ei= φiA, φiR, φiM, where

φiKis the efficiency of individual i at task K that are assumed to be exogenous in the model. For simplification, they can only choose one task to specialise within occupation at any one point in time. They will therefore perform tasks that maximise wages given comparative advantage:

W=Max φiAπA,φiRπR, φiMπM

The indicator function is given by:

IA=1 [φiAπA>φirπR ,    φiAπA>φiMπM]

IR=1 [φirπR>φiAπA ,    φiRπR>φiMπM]

IM=1 [φiMπM>φiAπA ,    φiMπM>φiRπR]

Impact of routinisation

The exogenous variable in this model is technology advancement represent by a decline in p. This reduces

πRone for one as they are perfect substitutes. As task demand curve is downward sloping, decline in p raises demand for routine tasks. This additional demand is only supplied by additional computer capital, as workers that used to supply R will now move to supply A or M as

πRfalls.

Assume it is easier for these workers to move to M compare to A as they need special training to be efficient in supplying A. In other words, the substitutability between A and R is higher than between R and M. This will lead to an unambiguous increase in

πAthrough two q-complementarity channels: increase R from computer capital and increased labor supply to M. This offsets the small increase in the supply of A.

The effect on

πM  is ambiguous, as q-complementarity between computer capital and M that increases

πM   is offset by increasing supply. It is possible for both

πM  and

πRto fall when p declines, but the relative wage (

πM/

πR) unambiguously rises.

Hypothesis

1. As price of IT falls steadily, substitution effect reduce demand for routine tasks and complementarity effect raised relative demand for abstract tasks. Computerisation does not directly affect demand for manual tasks.

2. Returns to routine tasks decrease while returns to abstract tasks increase. The effect on manual task is ambiguous depending on its complementarity with computer and the increase in labour supply from routine workers that are replaced by automation.

4. Data

4.1 British Skill Survey

This paper uses British Skill Survey (BSS) conducted in 1997, 2001, 2006 and 2012, which collects individual data on the importance of different tasks performed in a job, demographic characteristics and workplace conditions. In contrast to DOT, BSS uses self-reported information and does not suffer bias from external panel assigning scores to characterise task contents of a job. The four comparable waves allow me to analyse changes in task contents within occupation over time, that are often assumed to be fixed in the literature. However, the major limitation is this dataset only started in 1997, while most studies show 1980-1990 as a period with fastest technological progress[5]. Rather than comparing long run trends, I focus on analysing recent changes.

Sampling and response weights are used throughout (See Felstead et al., 2015). The four surveys have sample size of 2,467, 4,470, 7,787 and 3,200 respectively. I exclude individuals who are aged 60 or above and those live in Northern Ireland, as these samples are only collected from 2006 onwards. I also exclude individuals that are self-employed and if their information on task contents and wages are missing, yielding 13,222 samples overall. Hourly wages are deflated using 2010 inflation index.

4.2 Construction of Task Measures

Respondents were asked to evaluate the importance of 32 tasks using Likert scale from 1 to 5 (with 1 representing “essential” and 5 representing “not important at all”). I use two methods to form task indicators, firstly by categorising them into abstract, routine and manual following Autor et. al (2003, 2006) and Spitz-Oener (2006), and secondly by using principal components analysis (PCA) following Dickerson and Green (2004). Details can be found in Appendix A with correlation matrix shown in Table A3.

4.2.1 Routine Task Intensity

A major empirical challenge is to classify tasks in a consistent and objective way. Although some tasks such as analysing complex problems or thinking of solutions are easily defined as abstract, other tasks such as selling or listening to colleagues are more general and difficult to put into one category. The literature offers no conclusive way of defining tasks, as each research uses different classification. I manually assign 32 work activities into three broad categories: abstract, routine and manual. Task intensity are calculated by averaging the scores within each category and converted into an increasing cardinal scale from 0 to 4 (with 0 representing “not important at all’ and 4 representing “essential”). While the theoretical model assumes workers only perform one task at a time, an occupation typically comprises multiple tasks. I therefore create a Relative Routinisation Index (RRI) as the ratio of routine tasks to all the tasks.

4.2.2 Generic Task Indexes

In addition to classifying tasks into three broad categories, I use PCA[6] to reduce 32 activities into 7 task indexes: Analytical, Managerial, Social, Planning, Physical, Numeracy and Literacy. This removes the subjectivity of assigning tasks into routine or nonroutine, and adds another dimension in analysing what types of tasks are affected by computerisation. I retain all factors that have eigenvalues above one, and the resulted 7 factors capture 61 percent of variance. Varimax rotation is used but the construction of factors is robust to other types of rotations. For interpretation purposes, task indexes are constructed by averaging scores from work activities that correlates highly with the composite factor (with correlation above 0.3).

4.3 Computer intensity

Computer intensity index is generated by averaging the scores from computer importance and complexity. The survey asks respondents to rank the importance of computer (from 1 representing “not at all” to 5 representing “essential”) and complexity of computer usage (from 1 representing “straightforward” to 5 representing “advance”) in their job.

4.4 Aggregate trends

Prior to regression analysis, I present summary statistics and aggregate trends in employment and wages from 1997-2012. Similar to the pattern found in other developed countries, Table 1 provides evidence of job polarisation in the UK. Column 2 shows the employment share at the top and bottom end of the skill distribution increases by 10.7 and 3.5 percent respectively, while occupations that require middle skills decreases by 14 percent.  In contrast to wage polarisation observed in the US, changes in real hourly wages in the UK follow a monotonic pattern, as those with the highest skills experience greatest rise in wages. Turning now to task contents, the high- paying occupations have relatively low RRI as expected. The most routine occupation is elementary trade, plan and storage, while the least routine occupation is business and public services professionals.

Figure 1 shows the aggregate trends in task contents and computer intensity from 1997 to 2012. Computer intensity has increased significantly since 1997. RRI decreases mainly due to increasing intensity of abstract tasks. There is only a small decrease in routine task intensity from 1997 onwards, as one would expect the greatest decline to occur during 1980-1990, when technology experienced a major breakthrough in workplace.

The next section presents detailed empirical analysis. Section 5.1 investigates the relationship between automation and changes in task contents and Section 5.2 analyses changes in returns to tasks.

5. Empirical Results

5.1 Hypothesis 1: Automation and task contents

5.1.1 Identification

To explore changes in task contents following computerisation, I form pseudo-panels using repeated cross-section data by grouping individual into cohorts with similar characteristics due to lack of panel data (Deaton 1985). Changes in cohort average are used as a proxy for changes within individuals over time. A crucial assumption is cohort have similar set of characteristics that stay constant from one survey to the next such as education and age. Although fixed effect estimates from pseudo-panel are less precise, it is less biased comparing to OLS and large number of individuals in each cohort can be used to solve measurement errors.

I construct two pseudo-panels. In the first panel, I aggregate individual data to occupation level using 3-digit SOC2000 code (refer to occupation fixed effects hereafter). Due to unobserved variables within occupation that are changing over time, I form a second panel with 108 cohorts by grouping individual with the same occupation, education and age range (9 occupation groups x 4 age groups at 10-year intervals x 3 education groups, refer to occupation fixed effects hereafter). An example of a cohort would be a manager aged between 20 and 30 with high education. A critical assumption is that individuals in each cohort have similar characteristics and these characteristics do not change over time. I estimate the following regressions for each task category:

A̅ct=βAC̅ct+X̅’ct δ +∑t=13γt+u̅c+ε̅ct

R̅ct=βRC̅ct+X̅’ct δ +∑t=13γt+u̅c+ε̅ct

M̅ct=βMC̅ct+X̅’ct δ +∑t=13γt+u̅c+ε̅ct

where

A̅ct,

R̅ct,

M̅ctrepresent the three tasks: abstract, routine, manual performed by cohort c in period t respectively. The overbar represents averages within cohort.

γtis the time dummies to control for underlying trends and

X’ctdenotes all the controls. The error term can be decomposed into two parts: a time-invariant variable

(u̅c)which can be eliminated using fixed effects and the other changing over time

(ε̅ct). The variable of interest

βis unbiased if strict exogeneity holds. The theory predicts automation leads to an increase in demand for abstract tasks (

β>0), decrease for routine tasks

(β<0), and a limited impact on manual tasks.

Estimates are weighted by the number of observation in each cohort, and cluster robust standard errors are used to mitigate heteroschedasticity. In forming cohort groups, there is a tradeoff between the number of individuals in each cohort and the number of overall cohorts in each period.  By constructing fewer cohort groups and averaging more individual within each group, there are more variation within cohort but a smaller overall sample size. To avoid biasedness, I drop cohorts that have less than 5 observations and construct a cohort in different ways as robustness tests.

5.1.2 Results

Table 2 estimates the relationship between automation and shifts in job contents based on equation 1.  Column 1 and 3 estimate the impact of computer on tasks using occupation fixed effects. The coefficient on abstract task is positive as predicted. The sign on routine task is positive but statistically insignificant. This can be driven by omitted variable bias, as the regression has not controlled for characteristics within occupation that may affect task contents such as education and experience.

Column 2 and 4 add further restriction by considering cohort fixed effects. Coefficients for abstract task is lower after controlling for cohort fixed effects, and estimates for routine task is now negative and statistically significant. 1 unit increase in computer intensity leads to 0.16 unit increase in abstract task intensity and 0.20 decrease in routine task intensity. Considering abstract tasks changes by 0.2 and routine tasks change by -0.01 from 1997 to 2012, these estimates are economically significant. The effect on manual tasks is insignificant shown by Column 5, suggesting computer is neither a direct substitute or complements to manual tasks. Overall, the negative coefficient on RRI in Column 6 supports the routinisation hypothesis. Individuals who have relatively large increase in computer usage in their job, experience larger increase in abstract tasks and reduction in routine tasks.

5.1.3 Alternative explanations

The impact of computerisation on job tasks may vary depending on individual skill level due to differences in comparative advantages. Table 3 considers the heterogeneous effect by education groups. Consistent with the theoretical model, the shifts towards abstract task and away from routine task are observed across education group. This shows changes in task contents are not merely a reflection of education upgrades. Column 3 shows that the increase in abstract tasks as a result from computerisation is greatest amongst highly educated workers, as they typically have comparative advantages in abstract tasks. The biggest decline in routine tasks due to computer intensity are among those with middle education, as they typically perform the most routine tasks.

For robustness test, I use three other measures of computer intensity in testing the hypothesis: a binary response of whether the individual uses a computer in their job, the proportion of employers using a computer in the workplace and if new communication technologies have been installed in the workplace. All results confirm that there is a shift towards abstract tasks and against routine tasks due to computerisation.

To test the result is not driven by reverse causality, I regress changes in task inputs with lagged changes in computer intensity using first difference. Lagged change in computer intensity having smaller but positive effect on abstract task. The effect on routine tasks is negative but insignificant.

Consistent with Autor et al. (2003) and Spitz-Oener(2006), this section establishes that variation in task contents within occupation is systematically related to computerisation. This paper extends their findings by considering a wider range of occupation over consistent time interval, and control for cohort fixed effects.

5.2 Hypothesis 2: Automation and returns to tasks

5.2.1 Identification

The previous section provides evidence of routinisation from a demand perspective. To understand how workers reallocate skills to tasks following automation, it is necessary to examine returns to tasks. The theoretical model predicts that as computer prices decrease exogenously, return to routine task should also decrease while return to abstract task increases due to enhancing productivity from q-complementarity. Return to manual task depends on two offsetting effects, with increasing supply from displaced routine workers cancelling out the productivity enhancing effects.

Returns to tasks cannot be analysed using Mincer equation, as tasks are not durable investment like education that earns a positive rate of return to compensate for foregone earnings. There are also endogeneity issues as occupation choice and tasks allocation are not random. Workers self-select into occupations based on comparative advantages. My identification strategy is based on the model from Autor and Handle (2013). Skill endowment is defined as a vector of three task efficiencies

Ei= φiA, φiR, φiM, where

φiKis the efficiency of individual i at task K.

Each individual i in occupation j produce occupation specific output

Yijusing K tasks, with

uidenoting worker-specific error:

Yij=eaj+∑Kλjkφik+ui

In a competitive market, workers are paid their marginal product. Log wage of worker i in occupation j is represented by:

wij=aj+∑Kλjkφik+ui

Workers self-select into occupation by choosing occupation j that maximises their output and earnings. At equilibrium, workers are employed in the occupation that gives them highest earnings to their bundle of tasks. Task returns in this model are occupation specific:

∂W∂φk=λjk.

Using this framework, I estimate task returns within occupation and within cohort using pseudo-panel a discussed in section 5.1.1. As workers are likely to have similar characteristics if they are in the same occupation, fixed effect can eliminate time-invariant characteristics that drives self selection. More specifically, I estimate the following equation:

w̅ct=a̅c+∑Kλjkφ̅ik+δt+ui

where

w̅ctis the wage of cohort c in time t and

δtdenotes the time dummy. In this analysis, K represents abstract, routine and manual tasks, and later extends to seven generic tasks. The variable of interest is

λjk, which shows how much wages increase in percentage when a task intensity increases by one unit.

5.2.2 Results

Table 4 shows the impact of tasks on wages using OLS. After adding individual characteristics such as education and experience in Column 3 and firm characteristics in Column 4, the estimated effect of tasks on wages reduces compare to Column 1 which do not have any controls. This is expected due to upward ability bias. One unit increase in abstract task intensity increases wages by 16 percent and similar increase in routine task intensity decreases wages by 9 percent. Manual task intensity has a small negative effect on wages at 2 percent, which suggests the negative effect due to increasing supply outweigh the productivity enhancing effect. Although estimates in Table 4 are consistent with the hypothesis, it is likely to be biased due to self-selection. High ability individuals are more likely to choose occupations that have higher abstract task intensity, leading to an upward biased in the coefficient. I next consider fixed effects to reduce biasedness.

Table 5 shows fixed effects which yields similar estimates as OLS. Column 1 shows occupation fixed effects while Column 2 shows cohort fixed effects by adding additional restriction to consider variation within individuals with the same occupation, education and age range.  As shown in Column 2, workers performing one more unit of abstract task earn 18 percent more compare to others in the same cohort, and earn 14 percent lower if they perform one more unit of routine task. The coefficient on manual task is not statistically significant. Although pseudo-panel eliminates potential bias due to unobserved time-invariant variables, the estimates are less precise resulting in higher standard errors compare to Table 4. To avoid potential biasedness due to small observations in some cohorts. I examine sensitivity of the results using different formation and find them to be robust.[7]

Returns to tasks may differ across education group due to differences in efficiency endowments, which is tested in Column 4-6 by regressing tasks on wages for three education groups separately Returns on routine tasks is negative and significant across education groups, reflecting routinisation is occurring across education groups. Those with high and middle education have higher returns to abstract tasks than low education. This suggests productivity gains for them are more significant given they have more comparative advantages in abstract tasks.

5.2.3 Interaction with computer intensity

To further test the hypothesis that changes in wages is a result of technology rather than proxying individual ability, I interact task contents with computer intensity in Column 6. This allows effects of task on wages to be heterogeneous depending on computer intensity. However, lack of variation can be problematic using fixed effects, as occupations that have high computer intensity are typically those that also require higher abstract task intensity. OLS in Column 7 is therefore my preferred specification in testing interaction effect, especially since the previous section shows that biasedness is likely to be small after adding relevant controls.

As expected, the interaction term between routine tasks and computer index is negative and significant, driven by competition between routine workers and computers.  Given the same amount of routine tasks input, Column 7 illustrates that one unit increase in computer intensity from the mean reduces wages by 2 percent. The interaction between manual and computer index is individually and jointly insignificant. For abstract tasks, Column 6 shows the interaction term to be negative and insignificant, which is likely to be driven by small variation. My preferred specification in Column 7 shows one unit increase in computer intensity from the mean increases wages by 13 percent, which is more in line with the prediction of theory.

Column 6 and 7 are also consistent with the literature that criticises Krueger’s (1993) finding of a positive premium with computer usage. After controlling for tasks, computer does not have significant effect on wages. Using computer by itself does not carry any wage premium, as the effects depend on what tasks computers are used for.

5.2.4 Returns to generic tasks

A more objective way to classify tasks is using PCA, in which I reduce 32 tasks surveyed in BSS into 7 indexes. Column 1 supports the routinisation hypothesis. Analytical and managerial tasks that have high abstract content have positive returns, while tasks that have more routine component such as physical and numerical tasks receive negative returns. The negative and significant sign on social tasks seems to contradict existing literature (Borghans et.al, 2014; Deming, 2016), as they find positive returns to social skills. This inconsistency can be explained by the difference in constructing the index, as the social skills indicator defined by Deming (2016) [8] is more closely related to the definition of managerial tasks in this paper, as both include tasks such as persuading and influencing others. As Column 1 shows return to managerial tasks as positive, my result is consistent with Deming’s.

Column 2 shows the results after interacting 7 generic tasks with computer intensity. Managerial and analytical have positive interaction term, with the complementarity between managerial tasks and computer intensity highest amongst all the generic tasks. Interaction is negative for physical and numerical task.

Overall, Table 5 confirms the hypothesis that due to the decreasing price of computers over time, returns to tasks that are easily substitutable by technology is negative. Tasks that cannot replace have positive returns, as increasing computer intensity increase productivity in performing these tasks.

5.3 Trends over time

In this section, I consider the effect of automation in shifting task contents over time and analyse trends in returns to different tasks. Table 6 suggests the demand for abstract tasks have slowed over time as the interaction term becomes more negative. On the other hand, the substitution effect on routine tasks have strengthened as the interaction term get increasingly negative. A possible explanation is that recent innovation, particularly robotics and artificial intelligence have increased the capacity of technology in substituting routine tasks, as well as some abstract tasks that require analytical skills in which computer becomes increasingly more efficient in. One would expect this to affect returns to analytical tasks to decrease over time, which is confirmed in Table 7.

Table 7 summarises the returns to tasks over time, with Column 1 representing the return to tasks in 1997 and Column 2-4 showing the interaction term between task intensity and three time dummies. In Panel A, the positive return on abstract task decreases through time compare to the base year 1997. This is consistent with Beasley (2016) findings, as he documents demand and returns of cognitive tasks have reversed since 2000. The return to routine task is consistently lower than 1997 and remain negative, consistent with the demand side explanation that automation is increasingly substitute routine tasks.

Panel B estimates the returns to 7 generic tasks over time. Considering only interaction terms that are jointly significant, Column 2-4 show that returns to analytical tasks has decreased over time, consistent with Beasley (2016). In contrast, return to managerial task has increased over time, which corresponds to Deming’s(2016) findings that return to social skills are growing over time. This highlights the growing importance in tasks which human have comparative advantages over computers, as managerial tasks require human interaction and organizational skills that cannot be replaced by technology. As routine part of the job is increasingly substitute by technology, workers focus more on managerial tasks and technology raises the productivity due to complementarity. This relates to Bresnahan et al. (2002), who discuss the rising importance of social skills in the work place as ICT alters firm organisation structure.

6. Robustness test

6.1 Industry level

In addition to OLS and cohort fixed effects, I also aggregate individuals to industry level and consider variation within industry. According to the model, industry with the greatest increase in computer intensity should experience more significant increase in abstract tasks and decrease in routine tasks. Results from Table B1 and Table B2 confirms this prediction. It also provides evidence that abstract tasks receive a positive return while routine task receive negative return. These results are also robust after considering industry-education fixed effects.

6.2 Alternative explanation 1: Offshoring

Several papers argue offshoring as an explanation for job polarisation (Grossman and Rossi-Hansberg, 2008; Firpo, Fortin and Lemieux, 2011; Blinder and Krueger, 2013). A task is considered offshorable if it can be performed without face-to-face interaction or at specific geographical location. Since many routine tasks have these characteristics, offshoring can potentially contribute to job polarisation and changes in wage structure, by replacing these tasks with cheaper labour abroad.

Following Blinder’s (2006) approach, I construct offshorability index by averaging the scores of dealing with people and counselling, and interact it with an indicator of whether the job is required to do at a single workplace away from home. In Table B2, I analysis the impact of offshoring on wages within industry (Column 1) and within industry-education (Column 2), as one would expect offshoring to have an impact within industry instead of within occupation. The result shows that although offshoring on its own has negative impact on wages, the coefficient becomes insignificant after adding task intensity. The effects of abstract and routine tasks remain significant and as expected.

6.3 Alternative explanation 2: Supply side

Other supply side phenomenon such as education upgrades and change in age composition are unlikely to be an explanation, as the pseudo-panel is constructed to analyse individual within the same occupation-education-age group. I have also control for gender composition by adding percentage of male as a control throughout the regressions, with the coefficient being insignificant.

6.3 Different measure of task

A major limitation to fixed effects is small variation within cohort, reducing the precision of estimates. I test whether the result is sensitive to construction of task measures. Rather than averaging scores, I use an alternative method following Spitz-Oener (2006). Task measures are defined as:

Taskijt=number of activities in category j performed by i at time ttotal number of activities in category j at time t ×100

where t is the four periods that they survey was conducted, j represents three categories of tasks: abstract, routine and manual[9]. Table B3 shows that results is not sensitive to alternative formation of task measures, as estimates are consistent to the theoretical predictions.

7. Conclusion

The literature typically focusses on the effect of technology in changing employment structure and wage across occupation. This paper provides new evidence to support RBTC hypothesis within occupation, and makes two contributions to the expanding literature of explaining changes in labour market using task framework. My empirical analysis shows that workers reallocate their skills towards abstract task and away from routine task within occupation in response to automation. Secondly, by controlling self-selection using pseudo-panel, I find that tasks that are easily substitutable by computer receive a negative premium, while abstract tasks receive a positive premium. Although there is evidence that technology is increasingly capable in substituting routine tasks and starting to substitute some abstract task, return to managerial tasks has increased since 1997, indicating its growing importance in the labour market.

There is no doubt that technological advancement will continue to drive changes in employment and wage structure. Assuming task content to be fixed within occupation will overstate the extent to which automation is substituting for human labour, as workers can shift between tasks within occupation. Rather than being pessimistic that workers are losing the race to technology, human still has comparative advantages in tasks that require adaptability, judgement and creativity. Further research in demand and returns to these tasks would be beneficial. A major policy implication is to equip workers with the rights skills using education and training, so that they can shift to perform new tasks when technology automates the routine parts of their jobs.

References

Acemoglu, D. and Autor, D., 2011. Skills, Tasks and Technologies: Implications for Employment and Earnings. In Ashenfelter O. and Card D. E. (eds) Handbook of Labor Economics, Vol. 4B, Amsterdam: Elsevier, pp. 1043–1171.

Akcomak, S., Kok, S. and Rojas-Romagosa, H., 2013. The effects of technology and offshoring on changes in employment and task content of occupations. CPB Discussion Paper No. 233. The Hague, CPB Netherlands Bureau for Economic Policy Analysis.

Autor, D., 2013. The “task approach” to Labor Markets: An Overview. Journal for Labour Market Research, 46, 185–199.

Autor, D., 2015. Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29, 3–30.

Autor, D. and Dorn, D., 2013. The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market. American Economic Review 103 (5), 1553-1597.

Autor, D. and Handel, M., 2013. Putting Tasks to the Test: Human Capital, Job Tasks and Wages. Journal of Labor Economics 31 (2), S59-S96.

Autor, D. H., Katz, L. F. and Kearney, M. S., 2006. The Polarization of the U.S. Labor Market. American Economic Review 96(2), 189–194.

Autor, D. H., Katz, L. F. and Kearney, M. S., 2008. Trends in U.S. Wage Inequality: Revising the Revisionists. The Review of Economics and Statistics 90(2), 300–323.

Autor, D. H., Levy, F. and Murnane, R. J., 2003. The Skill Content of Recent Technological Change: An Empirical Exploration. Quarterly Journal of Economics 118, 1279– 1333.

Beaudry, P., Green, D.A. and Sand, B.M., 2016. The Great Reversal in the Demand for Skill and Cognitive Tasks. Journal of Labor Economics 34 (1), S199-S247.

Black, S. and Spitz-Oener, A., 2010. Explaining Women’s Success: Technological Change and the Skill Content of Women’s Work. The Review of Economic Statistics 92(1), 187-194.

Blinder, A. S. and Krueger, A. B., 2013. Alternative Measures of Offshorability: A Survey Approach. Journal of Labor Economics 31(2), S97-S128

Borghans, L., Ter Weel, B. and Weinberg, B. A., 2014. People skills and the labor-market outcomes of underrepresented groups. Industrial & Labor Relations Review 67(2), 287–334.

Bresnahan, T., Brynjolfsson, E. and Hitt, L., 2002. Information technology, workplace organization, and the demand for skilled labor: Firm-level evidence. Quarterly Journal of Economics, 339–376.

Deaton, A., 1985. Panel data from a time-series of cross-sections. Journal of Econometrics 30, 109- 126

Deming, D., 2015. The Growing Importance of Social Skills in the Labour Market. NBER Working Paper No. 21473, Cambridge, MA: National Bureau of Economic Research

DiNardo, J. and Pischke, J., 1997. The Returns to Computer Use Revisited: Have Pencils Changed the Wage Structure Too? Quarterly Journal of Economics 112, 291–303.

Dickerson, A. and Francis. G., 2004. The growth and valuation of computing and other generic skills. Oxford Economic Papers, Vol. 56, No. 3, 371–406.

Dustmann, C., Johannes L and Schonberg, U., 2009. Revisiting the German wage structure. Quarterly Journal of Economics 114, 843–82.

Entorf, H., Gollac, M. and Kramarz, F., 1999. New Technologies, Wages, and Worker Selection, Journal of Labor Economics 17(3), 464–491

Felstead, A., Gallie, D., Green, F. and Zhou. Y., 2007. Skills at work, 1986 to 2006. ESRC Centre on Skills, Knowledge and Organisational Performance, University of Oxford.

Firpo, S., Fortin, N. M. and Lemieux, T., 2011. Occupational Tasks and Changes in the Wage Structure. IZA Discussion Paper 5542, University of British Columbia.

Frey, C. B. and Osborne, M., 2013. The Future of Employment: How Susceptible are Jobs to Computerisation? Oxford Martin School Working Paper No. 7.

Goos, M. and Manning, A., 2007. Lousy and Lovely Jobs: The Rising Polarization of Work in Britain, Review of Economics and Statistics 89(1): 118–133.

Goos, M., Manning, A. and Salomons, A., 2009. Job Polarization in Europe. American Economic Review 99(2): 58–63.

Goos, M., Manning, A. and Salomons, A., 2014. Explaining Job Polarisation: Routine-Biased Technological Change and Offshoring, American Economic Review. 104 (8), 2509-2526

Grossman, G. M. and Rossi-Hansberg, E., 2008. Trading Tasks: A Simple Theory of Offshoring. American Economic Review 98(5), 1978–1997.

Katz, L. and Autor, D., 1999. Changes in the Wage Structure and Earnings Inequality. In O. Ashenfelter and D. Card (eds), Handbook of Labor Economics, Elsevier Science, Amsterdam, pp. 1463–1555.

Krueger, A., 1993. How Computer have Changed the Wage Structure: Evidence from Microdata, 1984-1989. Quarterly Journal of Economics 108(1), 33–60.

Michaels, G., Natraj, A. and Van Reenen, J., 2014. Has ICT Polarized Skill Demand? Evidence from Eleven Countries Over Twenty-Five Years. The Review of Economics and Statistics, 96 (1), 60-77

Mishel, L., Shierholz, H. and Schmitt. J., 2013. Don’t Blame the Robots: Assessing the Job Polarisation Explanation of Growing Wage Inequality, EPI-CEPR Working Paper.

Nordhaus, W., 2007. Two Centuries of Productivity Growth in Computing, Journal of Economic History 67(1), 17–22.

Spitz-Oener, A., 2006. Technical Change, Job Tasks, and Rising Educational Demands: Looking Outside the Wage Structure”, Journal of Labor Economics 24(2), 235–270.

Vivarelli, M., 2013. Technology, employment and skills: an interpretative framework, Eurasian Business Review 3(1), 66–89.

Vivarelli, M., 2014. Innovation, employment and skills in advanced and developing countries: a survey of economic literature, Journal of Economic Issues 48(1), 123–154.

Weinberger, C., 2014. The Increasing Complementarity between Cognitive and Social Skills, The Review of Economics and Statistics 96(5), 849-861.

Appendix

Table A1 Description and classification of tasks

Table A2: Factor loading

Notes: Factor analysis are applied on all observations using PCA and have been orthogonlally rotated. I have decided to retain 7 factors as they have eigenvalue above 1, supported by scree plot analysis. Kaiser-Meyer-Olkin test shows a value of 0.91, suggesting sampling is adequate for factor analysis.

Factor loadings larger than 0.3 are bolded. Generic tasks are constructed by averaging tasks that have factors loading larger than 0.3. Means are calculated adjusting for survey weights. Cronbach’s Alpha measures internal consistency of the factors.

Table A3: Correlation matrix

Appendix B: Robustness test

[1] Job polarisation has been documented in the US (Autor et al. 2008, Acemoglu and Autor, 2011), the UK (Goos and Manning, 2007), Germany (Spitz-Oener, 2006, Dustmann, et al., 2009), Japan (Ikenaga and Kambayashi, 2010), and wider Europe (Goos, Manning and Salomon, 2009; Michaels, Natraj, and Van Reenen, 2014).

[2] Autor and Dorn, 2013; Goos et al., 2014; Michaels et al., 2014

[3] Arntz et al. (2016) argue only 9 percent instead of 47 percent of US jobs face high risk of automation if one assumes tasks rather than entire occupation can be replaced.

[4] Earlier research attributes SBTC as the primary reason causing the rapid rise in wage inequality since the 1970s, driven by increasing demand for high skill workers at a faster pace than supply as technology is factor augmenting. See Katz and Murphy (1992); Katz and Autor (1999); Acemogu and Autor (2011) for review on this literature.

[5] Nordhaus (2007) argues 1985-1995 as the period with most rapid pace of computer improvement, with computer price declines averaging over 70 percent per year.

[6] PCA is used instead of factor analysis, to maximise variance subject to components being uncorrelated with one another.

[7] Results are robust when I increase the minimum cell size from 5 to 10, but with larger standard errors. I also test the sensitivity of results to different composition of cohort, by using 1-digit and 2-digit SOC2000 occupation code, forming age group in 5-year intervals and separating female and male in forming the groups. Results are robust to these alternative constructions of cohort.

[8] Deming (2016) construct social skills by averaging four variables: social perspective, coordination, persuading and negotiating

[9] I have catergorised the 32 activities surveyed in BBS into 9 abstract activities, 3 routine activities and 3 manual activities in total. For example, if an individual indicates two out of nine abstract activities are “essential” or “very important” in her job, the she gets a abstract task score of 22 (2/9*100).