Job polarisation in the UK: An assessment and implications for skills policy
Craig Holmes and Ken Mayhew
ESRC Centre on Skills, Knowledge and Organisation Performance,
University of Oxford
Abstract
Our ongoing research is focused on mobility and segmentation in the labour market during a period where the nature of work and occupational composition are changing rapidly. We are particularly interested in two interlinked phenomena: routinisation and polarisation. Our work is still in progress, but this paper details a number of findings which have informed our approach. Predominantly, it provides an assessment of Goos and Manning (2007) and similar work and of the concept of the polarised or 'hourglass' labour market. It uses data taken from two waves of the National Child Development Study to study changing patterns of employment and of labour market outcomes over the period 1981 to 2004. The main proposition is that while the dataset used here does demonstrate changes in employment of the sort consistent with the routinisation hypothesis, it is not clear that a polarising labour market is the inevitable conclusion. We show that wage distributions do not appear to have changed dramatically over the time period.This paper points to the future direction of our research, as well as discussing some implications of the findings for education and skills policy.
1 Introduction
The determinants of and limitations to occupational and wage mobility are a long-studied and much debated topic. Our work has focused on two interlinked phenomena: routinisation, the process whereby labour engaged in routine task based occupations is replaced by increasingly cheap computer capital, and polarisation, the alleged growth of jobs at the top and the bottom of the quality spectrum, and the decline, or hollowing out, of those in the middle.
Our work is set in the broader context of reconsidering the theory and evidence for segmented labour markets. In the original literature (see Leontaridi, 1998, for a survey), a number of different explanationswere put forward for possible segmentation, ranging from the rise of internal labour markets to more radical capitalist conspiracy theories. Our starting hypothesis differs from these and other approaches and is, as far as we are aware, new to the literature:wepropose that segmentation in labour markets may result from polarisation, caused by to the decline of employment in middling jobs that may previously have acted as transitional steps towards better paying or higher quality occupations.
As is typical in studies of intragenerational mobility, the approach taken in our work employs a longitudinal study from the UK to assess lifetime occupational mobility during the period most commonly associated with routinisation. This paper details a number of important findings which inform the future direction of our research. Our greatest concern comes from looking at wage distributions over this period - these distributions do not appear to have changed dramatically over the time period. The largest number of jobs continue to be in the middle of the wage spectrum, leading us to question to the conventional wisdom that UK labour markets have polarised.
Our explanation for this disconnect between the received literature and our analysis is that the existing evidence of polarisation rests on an assumption about occupational wage structure that appears unreasonably strong. As a result, we argue thatwhile employment patterns have changed in a way that is consistent with routinisation, a polarising labour market may not be an inevitable result. This paper argues that the existing methodology (for example, Goos and Manning, 2007; Autor, Katz and Kearney, 2006) is inappropriate for evaluating the polarising phenomenon and that a new approach is needed.
Despite this finding and the implications it may have for our starting hypothesis, the dataset we are using shows clearly that some occupations are in decline, creating a group of displaced routine occupation workers. One future direction for our work will be to look at the mobility of these workers and, returning to our interest in segmentation, to investigate whether there are certain occupational moves which are significantly restricted. Another related question that needs to be addressed is the potential for differences in the experiences of subsequent cohorts. A recent paper by Autor and Dorn (2009) has shown that generally the oldest workers remain in occupations with declining employment, potentially due to a greater degree of skill specificity. Later entrants into the labour market may be increasingly polarised, with implications for their future mobility.
The paper is arranged as follows. Section 2 will describe the theoretical background to our interestin routinisation and polarisation. Section 3 discusses the data and the limitations that have been encountered. Section 4 presents our initial results, broken down into the two areas of interest: employment by occupational group and wages distributions. We limit the analysis in this section to the most important highlights, with more extensive results presented in Holmes (2010).Section 5 concludes and discusses the future direction of the work, including a discussion of the implications it may have for skills policies.
2 Background
This section outlines the theoretical motivation for our work, in particular, the interlinked theories of routinisation and polarisation. The routinisation hypothesis, usually attributed to Autor, Levy and Murnane (2003), is a refinement of the idea of skill-based technical change. In the latter, technological advance increases the relative demand for skilled labour which is complementary to, for example, increased utilisation of advanced technology. The routinisation hypothesis argues that technological advances, especially computer capital, provide a substitute for tasks with a clear set of regular instructions and so will decrease firms' demand for labour to perform these tasks. At the same time, non-routine occupations may be complementary to computer capital and so firms will demand more labour to perform these occupations. Non-routine task-based jobs fall into two categories: skilled professional and managerial jobs, which are the most complementary to technical progress, and unskilled manual tasks or services (e.g. cleaning). The latter is not generally directly affected by technical change, but the impact of technology on other parts of the labour market is likely to lead to a rise of employment in these jobs.
Goos and Manning (2007) use this task categorisation to show that non-routine occupations tend to fall at the top and bottom of the wage spectrum, with routine occupations somewhere between. This observation leads to the polarisation hypothesis, with increasing employment at high-paying and low-paying jobs and falling employment for middle-income jobs.Goos and Manning then look at changes in employment for occupations of varying quality, proxied by initial median wage (in 1979). They find there has been employment growth in wage deciles at both ends of the pay spectrum and declines in the middle deciles. Similar results are found by Autor, Katz and Kearney (2006) for the United States and Spitz-Oener (2006) for Germany.
We are interested in what these changes in the structure of the labour market have meant for mobility. Existing work on polarisation and routinisation shows what has been happening at the macro level, in terms of the share of employment in different occupations and the effect on wage inequality. However, little work has been done on the impact at the individual level. There are currently unanswered questions about the labour market experiences of displaced workers in declining routine occupations, the factors that may determine these moves and the relative experience of later generations once the labour market has routinised or polarised. The literature provokes questions about the impact on overall mobility for low wage workers if the number of middling jobs above them become more scarce. It is this idea which has motivated our focus on segmented labour markets.
3 Data andissues
The data used in this paper is taken from the National Child Development Study (NCDS). The members of the NCDS study were all born in a single week in March 1958 in Great Britain. Data has been collected on these members in a series of waves. The most useful waves for our present purposes are the fourth and seventh waves, taken in 1981 and 2004-5 respectively. The fourth wave is the first one taken after the school leaving age (respondents were aged 23) and records early labour market experiences. The seventh wave was completed in 2004-5 (respondents were aged 46-47) and has the most recent data on wages, employment and education.
Our approach, in line with the existing literature, uses Standard Occupational Classifications as the basis for considering jobs. A recurring issue with labour market data over time is the changing classification of occupations. The 1981 wave uses the KOS (Key Occupations for Statistical Purposes) system of job title classification, which categorises occupations within the 18 CODOT (Classification of Occupation and Directory of Occupational Titles) major groups, while the 2004 wave uses the most recent SOC2000 classification. The SOC2000 coding system of occupations has a four level classification system, from major group (first digit) to unit group (fourth digit). To make data comparable across the two waves, we have derived a conversion system between KOS and SOC2000 codes, using the descriptions of occupations provided for each group[1].The conversion is not always perfect (see Holmes, 2010, for a discussion). In some cases a category in SOC2000 could apply to several categories under KOS (and vice versa) and subjective judgements have been made. Total exclusions on this basis account for 12.43% at the minor group (third digit) level.
As has been indicated in the introduction, one of the central findings of this paper is that it is not immediately apparent from wage distribution data that employees in this cohort have polarised, leading to an examination and critique of the established methodology to demonstrate the polarisation effect.An obvious question that has been raised is over the validity of using longitudinal data to analyse this. The most pressing concern is that the data will be biased due to career advancement or promotion. For example, an increase in employment of managerial occupations for a single cohort will, at least in part, be due to individuals progressing to better jobs due to experience.
Our justification is as follows. Firstly, we can use evidence from cross-sectional data to support our argument. In the annexe to Holmes (2010), wage distributions from the Family Expenditure Survey between 1981 and 2000 are presented to show a similar lack of decline in the middle of the income distribution. We acknowledge that further work on this is still need and we intend to make use of the larger New Earnings Survey to complete the analysis.
It could be questioned why we did not simply start with cross-sectional data. However, our overall focus in this work is on mobility and the effect that the phenomena of routinisation and polarisation may have had on this. It is therefore essential that evidence of both are found within a dataset that is more appropriate for looking at issues of mobility. Later we show that our dataset do indeed exhibit the expected growth in non-routine occupations and decline in routine occupations. Therefore, it becomes a valid question as to why this has not lead to observable polarisation in the resulting wage distribution.
This leads on to the final and, from our perspective, most important point, which is that the critique is one of the prevalent methodology, regardless of the data that has provoked its examination. The implicit assumption in existing studies is that the relative position of occupations in the wage distribution has remained approximately constant over time, so that it is appropriate to use initial wages as a proxy for job quality over the entire period being studied. We will argue that it is not appropriate to do so, certainly without an empirical justification that has not as yet been presented.
One final issue that we have considered is the use of average wages as a proxy for job quality. There are obvious concerns about this (see Holmes, 2010), but for which adetailed examination falls outside of the scope of this paper. As a result, we will continue to use average wages as a proxy for job quality.
4Results
Occupational structure
This section replicates, for a single cohort, the Goos and Manning methodology used to asses the extent to which the labour market has become more polarised. To show changes to the occupational structure, we used SOC2000 occupational minor group, which gives 70 occupational categories, both those reported in the 2004 wave of NCDS interviews and those derived for the 1981 wave using the conversion described in the above section. These occupational categories have been ranked by mean wage of the occupation based on 1981 wages and divided into ten groups, each with approximately 10% of total employment, with group one having the highest paying occupations and group 10 the lowest paying occupations. With one exception, all occupations fall into one group. This lumpiness means thatgroup sizes vary from 9.24% to 11.07%. The one occupation which was split was Administrative Occupations: Records, which has an employment share of 18.90% in 1981. The group is split over three deciles, the middle of which had exactly 10% of employment and contained only workers from this occupation.
Figure 1: Changes in employment share of occupations between 1981 and 2004, by decile (1981 mean wages)
Source: NCDS data (1981 and 2004), own calculations
Figure 1shows the change in employment share of each of the ten groups. It shows that the initially highest and lowest paid occupations grew more than the middle earning occupations, which experience a severe decline. This finding is consistent with that found in Goos and Manning (2007), Autor, Katz and Kearney (2006) and Spitz-Oener (2006).
Figure 1shows the change in occupational employment share for one cohort who start in 1981 aged 23 and finish in 2004 aged 46. Of course, some of these occupational transitions are the result of individual career progression. It remains an ongoing challenge to isolate the effect on employment shares that derives entirely from routinisation and polarisation, in the absence of a earlier comparable longitudinal study.
Wage distributions
The evidence presented in the previous section suggests support for the routinisation hypothesis and indicates that the occupations that have grown tended to be at top and bottom of the wage spectrum, based on initial earnings. However, perhaps the more important outcome when discussing a polarising labour markets is not changes in the relative size of different occupations, but the consequential change in wage distributions. This section will present our main contention: that while routinisation has led to a fall in employment of several types of occupation, it may not necessarily have caused a polarisation of the labour market in terms of wages.
Figure 2: the effect of polarisation on wage distributions
Our starting point is to look at wage distributions before and after a period of routinisation as a way to better describe the overall consequences. A polarising workforce has a greater number of individuals earning low and high wages and a decreasing number of individuals earning middling wages. Consider an initial wage distribution, represented in Figure 2 as a cumulative distribution function. Holding everything else constant, employment increases in the lowest paying jobs will steepen the lowest section of the distribution function. The lowest deciles of workers covers a smaller range of wages as there are more workers earning the lowest wages. A similar steepening would be seen at the top of the distribution as employment increases in well-paid jobs. At the same time, falling employment in the intermediate jobs flattens the distribution function in the middle, as the range of wages covering the middle deciles increases with lower employment in the middle. The end result is show in Figure 2.
Given the evidence in the previous section that suggests that occupational employment patterns within the NCDS data have followed similar patterns found in the polarisation literature, we might expect to see evidence of the shift shown in Figure 2 within the wage distributions.Figure 3 and Figure 4 present hourly and weekly log gross wage distributions from the two NCDS waves. Although hourly wages are the preferred measure, the lack of hours of work data in the later wave limits the sample size. Looking at weekly wages increases the sample size, although this not a reliable basis for comparing distributions over time if patterns of work have changed. To correct partially for this, I present weekly wage distributions for just those in full time employment, as well as for all the available data as a comparison.
Figure 3: Log hourly wage distributions, 1981 and 2004 (all workers and fulltime workers only)