Mobility and the changing structure of occupations: a cross-cohort comparison Craig Holmes, Ken Mayhew and Felix Chow1 ESRC Centre on Skills, Knowledge and Organization Performance, University of Oxford Abstract We look at the way occupational mobility may have been affected by technological progress and the process of routinisation – the decline in the relative employment of middle wage routine occupations (Autor, Levy and Murnane, 2003; Goos and Manning, 2007). We focus on two key questions: Where do workers in routine occupations end up if they are displaced from their jobs by routinisation, and does this differs between younger and older workers? Do younger workers recognize that declining routine occupations may affect existing career paths and choose different entry points into the labour market than older cohorts? To compare older and younger workers, we will use panel data on occupations, education and training from 1981 onwards, using the 1958 National Child Development Study (NCDS) and the 1970 British Cohort Study (BCS). As well as comparing differences in initial occupational choices, this paper presents a mobility analysis using logistical regressions. The effect of routinisation on job moves for each cohort is separated out from other driving factors, such as career advancement or job mismatch. Keywords: Class / occupational stratification; Labour markets; Transitions. 1 Correspondence email: craig.holmes@education.ox.ac.uk 1 Introduction Over the previous three decades, technological progress has driven a shift in the occupational structure of many countries, including the UK. Some jobs comprise of a number of tasks which could be replaced by ICT capital. As the price of such technology falls, so too does employment in these jobs. These jobs are referred to as routine in the sense that the tasks performed by workers in them tend to follow a series of instructions, which could be replicated by an appropriately programmed machine. This process is often referred to as routinisation (Autor, Levy and Murnane, 2003). Along with the related phenomena of polarisation – which identified routine occupations as mostly middle-wage jobs – much of the related discussion has been on the implications of these changes for wage inequality (Goos and Manning, 2007; Autor, Katz and Kearney, 2006). However, changes in the occupational structure have potentially important effects on mobility as well, but as yet these effects have not been rigorously analysed. With improving upward mobility often mentioned as an ambition of successive governments, it is important to establish what barriers exists in order to devise policies to overcome them. There are numerous dimensions to the effect routinisation may have on mobility. In a recent paper Holmes (2011a) examined the labour market outcomes of routine job employees displaced by routinisation. A key question is whether these workers are able to move to well-paid non-routine jobs, and if they are, what factors contribute to this upward mobility. Using data from the 1958 National Child Development Study (NCDS), the paper presents a mobility analysis of these routine workers between 1981 and 2004, separating out the effect of routinisation on job moves from other driving factors, such as career advancement or job mismatch. As expected, periods where the employment share of routine jobs fell markedly across the entire economy were periods which witnessed increased the mobility of routine workers towards both high and low wage non-routine jobs. The relationship between routinisation and mobility is mediated through the qualification levels, specific skills and experience of workers. We add to this analysis by comparing the experience of the NCDS cohort with a younger cohort taken from the 1970 British Cohort Study. Members of this group entered the labour market near the end of the 1980s, after a decade of routinisation-driven changes to the structure of occupations. It is realistic to expect that these changes may affect different cohorts in different ways. This paper presents some initial findings from our analysis of two questions, through a comparison of the two cohorts. Firstly, we examine where workers in routine occupations end up if they are displaced from their jobs by routinisation, and whether this differs between younger and older workers. Secondly, we look at whether younger workers recognize that declining routine occupations may affect existing career paths and choose different entry points into the labour market than older cohorts. The paper is arranged as follows. Section 2 provides a brief overview of an economic model of occupational structure and mobility. This model is tested in the remainder of the paper. Section 3 sets out the empirical strategy, including the methodology used in Holmes (2011a) to identify mobility effect of routinisation, compared to other factors. Section 4 discusses the two survey from which the data is drawn. Section 5 and 6 present the results of the two main analyses. Section 7 concludes, and points towards further directions for this research. 2 The determinants of occupational mobility: a theory Holmes (2011b) presents a model of the economy where production depends on the input of four types of labour: professional, managerial, routine and service. This model will be briefly summarised in this section, as it motivates the empirical strategy. Routine labour is a perfect substitute for capital in the production process, so, as this technology becomes cheaper, it drives down routine employment and wages. Employment in professional occupations requires the necessary accreditation or qualifications. Managerial occupations and service occupations have no accreditation requirement, and represent high skill and low skill non-routine occupations respectively. Labour market outcomes, and hence mobility, is driven by the human capital of the individual, which comprises of a number of different elements, including formal education, on-the-job experience and innate productive ability. Workers differ in their levels of education attainment – they may be either university graduates, high school graduates or unqualified. Individuals with higher attainment are able to work in the better jobs, either because these jobs require higher levels of skills developed through more education, or because education acts as a signal in a screening process where the most able invest in more education. University graduates are able to supply their labour to either managerial or professional occupations. High school graduates supply their labour to either managerial or routine occupations. Finally, the unqualified are able to supply their labour to either routine or service occupations. Thus, obtaining higher qualifications can facilitate upward progression. Within each educational category, workers differ in their productivity in each occupation. This relative productivity can reflect a number of different dimensions. Firstly, it can represent differences in general cognitive abilities. For example, some people may be innately better problem solvers, making them more productive managers, or possess better tool-handling skills, making them better process operatives (a routine occupation). Secondly, individuals develop capabilities through experience which allows for progression to better jobs, leading to the creation of a career. For example, working for a time in a workshop is considered necessary experience for promotion to the position of shop foreman. This type of mobility was introduced in a formal model by Sicherman and Galor (1990). It is included in the model by allowing relative productivity in better occupations to increase as workers get older. Thirdly, individuals develop specific skills from working in a particular occupation, where these skills are only useable within that occupation. This is included in the model by allowing the relative productivity of the job currently being performed to increase as the worker spends more time in it. Wages in each occupation are set per effective unit of labour input, so more productive workers earn more for the same time input. In the absence of routinisation (i.e. with stable relative wage rates) mobility would be in the form of career progression, as individuals become relatively more productive in better occupations, and as a result may move occupations as they get older2. A change in the relative wages of the different occupations alters supply patterns and generates additional mobility. For example, an exogenous fall in the wage of routine workers (caused, as in the case of routinisation, by a fall in the price of ICT capital for which routine labour is substitutable) would lead some unqualified workers to move to service occupations while some high school graduates would supply their labour to managerial work3. The model also makes predictions about the outcomes of different cohorts. For example, the younger cohort may be more mobile – skill specificity make older workers less likely to move between occupations and more likely to accept falling wages, whilst young workers are less keen to move into those in decline. Moreover, they may be better placed in terms of transferable or generals skills to move out of routine occupations into those better jobs. At the same time, general career-progressing experience can increase the prospects of upward mobility following routinisation. In addition, as a result of entering at a time where some of these occupational changes were starting to be recognised, the entry points of younger workers might be different. They may have taken into account changes in established career paths that older generations who entered prior to these changes were unable to. If they anticipate a further decline in routine occupations, they are less likely to want to develop specific skills in these occupations. These factors have been discussed by Autor and Dorn (2009), who argue that as a result, routine jobs are getting 'older'. 2 Assuming that this effect outweighs the routine occupation specific skill effect. This is the sort of occupational mobility that a part of the literature tends to focus on, which starts with the assumption that changes in employer or job arise when there is a shift in an individual's information set or preferences, or the external labour market environment, such that an alternative occupation yields a higher expected lifetime utility. 3 3 Methodology We run two analyses in this paper. The first analysis follows Holmes (2011a), which separates out the two major reasons for occupational transitions identified in the model: career progression, and displacement caused by routinisation. Simply looking at the occupational mobility of routine workers over time will capture both. Ideally, we would compare against a counterfactual cohort, looking at the probability of moving between different occupational categories for a workforce unaffected by routinisation, and one that entered the labour market just as routine occupations began to decline. However, an appropriate early cohort study does not exist in the UK. The methodology in Holmes (2011) looks at transitions between occupational groups are examined across a number of periods of time. The factors affecting transitions from routine occupations are estimated using a binominal logistic regression, where the independent variables are education, specific experience, age and two demographic variables (gender and race), running one regression for each final occupation group. As a final explanatory variable, a measure of routinisation is included for each period of transition. This final addition to the model can, at least to an extent, separate out occupation moves caused by routinisation and those caused by general career progression. The logistic regression (or logit model) is set up as follows. Let Yi be a dummy variable which takes value of 1 if the individual i is in the given occupational group Y at the end of the period. Then, the logit model specifies the functional form of the probability distribution over belonging to a certain occupation at the end of the period t as: Pr(𝑌𝑖𝑡 = 1|𝑋𝑖𝑡 , 𝑅) = exp(𝛼+𝛽𝑋𝑖𝑡 +𝜃𝑅𝑡 ) 1+exp(𝛼+𝛽𝑋𝑖𝑡 +𝜃𝑅𝑡 ) = 𝑙𝑜𝑔𝑖𝑡 −1 (𝛼 + 𝛽𝑋𝑖𝑡 + 𝜃𝑅𝑡 ) Rearranging gives the log odds ratio, z, as a linear function: Pr(𝑌 =1|𝑋𝑖𝑡 ,𝑅) ) 𝑖𝑡 =1|𝑋𝑖𝑡 ,𝑅) 𝑧𝑖 = ln (1−Pr(𝑌𝑖𝑡 = 𝛼 + 𝛽𝑋𝑖𝑡 + 𝜃𝑅𝑡 (1) where Xit is a vector of educational attainment, specific experience, age, cohort and other personal characteristics of individual i (including any interaction terms), in period t, the period of transition, and R is a measure of routinisation (or the change in the occupational structure) in that period. The time period may enter the equation in a number of forms, but the baseline model includes it as a linear trend and is interpreted as the accumulation of career-progressing experience. The second analysis looks at early occupational choices. This assesses the hypothesis that, when faced with the prospect of declining routine jobs, different cohorts may enter the labour market at different ways, with less people moving into routine occupations in cohorts that are more able to anticipate the declining occupations. We estimate a logit model for the probability that a worker is in a routine occupation as a starting occupation. A baseline model predicts the probability conditional on qualification levels personal characteristics, and the cohort. This model is extended to interact cohort with qualifications, to see whether people of different education levels systematically start their working lives in different places across the two age groups. 4 Data These models are estimated using the National Child Development Study (NCDS) and the 1970 British Cohort Study (BCS). The members of the NCDS study were all born in a single week in March 1958. Data has been collected on these members in a series of waves. The most useful waves for assessing labour market outcomes over a period where routinisation has taken place are between the fourth and seventh waves, taken in 1981, 1991, 1999-2000 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 experience. The seventh wave was completed in 2004-5 (respondents were aged 46-47), and has the most recent data on wages, employment and education. It is possible to construct an entire working life history over this time period using responses from all four waves, including periods of employment, unemployment, selfemployment and non-participation for a number of reasons such as sickness or further education. As with all longitudinal studies, there are missing data. The sample size is around 12,000 for the fourth wave, and around 10,000 for the seventh wave. Employment data are available for a subset of these observations. Just under 5000 individuals report employment data in both the fourth and seventh waves, though this sample increases if we look at those who were non-employed or unemployment in one of the waves. The BCS is a longitudinal study following 17,200 subjects born in England, Scotland, Wales and Northern Ireland in a particular week in April, 1970. The cohort has been surveyed in series of waves since the study's inception, and until now there have already been eight waves. As the focus of our study is on labour market outcomes, we restrict our sample to the fifth wave onwards, when the subjects are at the ages of twenty-six, twenty-nine, thirty-four, and thirty-eight. The fifth wave, taken in 1996, is the first full wave taken after the school leaving age of the subjects4. The most recent wave is the eighth wave, which was completed in 2008. There is missing data in our dataset. The fifth wave has a sample size of 9,000, and the eighth wave has data on around 8,900 individuals. Approximately 6,500 subjects are reported in both the fifth and the eighth wave. 4 The forth wave was taken in 1986 when the subjects were sixteen. There was a sample-survey in 1992 where only a sample of 1600 of the subjects were surveyed. The data suggests six periods of transition for the first analysis: 1981-1986, 1986-1991, 1991-1995, 19951999, 1999-2004 and 2004-2008. For the NCDS, four of these periods coincided with the actual survey years of the dataset (although we do not include the 2008 NCDS wave), and the remainder (1986 and 1995) fell in the mid-point between two surveys. For the BCS, we start with occupational data in 1995. The remainder of the years coincided with the survey years. While it is possible to construct a full occupational profile for all study members in all years, answers to questions about current jobs would seem more reliable than answers to questions about past jobs. 4.1 Occupational mobility Occupations of employment are measured using the narrowest available occupational coding. One problem with doing this over a long period of time is that the system of coding occupations has changed three times since 1980. 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 1991 and 1999 surveys use SOC90 and 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 between 1981 and 2004, a conversion system was derived between KOS and SOC2000 codes, using the descriptions of occupations provided for each group. 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. In some cases, observations have been dropped because it was not possible to place one KOS code into a single SOC2000 code. In the NCDS, total exclusions on this basis account for 12.43% at the minor group (threedigit) level for the 1981 survey. A similar conversion was created between SOC90 and SOC20005. These two classification systems had a much more overlap in terms of the descriptions of each category. A conversion was made from each SOC90 occupation to a 4-digit SOC2000 category, where descriptions were on a similar level of aggregation. These were then reduced into 3-digit categories which are used in the analysis. Each 3-digit category was assigned to one of the six occupational categories, two more than the four categories discussed in the model, as shown in Table 3. This adds an intermediate occupation category the two high skill non-routine occupations on the basis that such occupations (including associate professionals and technicians) have different entry requirements than professionals and potentially different employment trends than managerial occupations. Moreover, a small category for manual non5 Both conversions are available upon request. routine occupations was added to distinguish this sort of low wage non-routine work from service occupations. The allocation between different occupational categories were based on the wages (using average values from NCDS data in the 1981 and 2004 waves), description and change in employment share (using Labour Force Survey data). Aside from a few obvious categories (such as those which are clearly professional from the descriptions), any occupation which experienced a trend decline in employment share is considered to be a routine occupation. Looking at the wages and descriptions is used as a common sense check – all these occupations have middle range wages and their descriptions suggest the work involves administrative or manual processes which could be replaced by computer technology6. 4.2 Educational attainment Across the four waves of the NCDS used in this paper, there are numerous systems for recording educational achievement, including detailed data on a wide range of vocational courses which have declined in importance in recent years. As a way to bring all of this data together, the highest NVQ equivalent level across time is recorded7. Each individual has two educational variables – a highest NVQ level in academic courses and a highest NVQ level in vocational courses, with both ranging from 0-5. The data are cleaned so that NVQ levels do not decline over time. 4.3 Experience and specific skills General experience is captured by age, or period of transition, which ranged from 1 to 5. Specific experience within a routine occupation is included below using a scale measure from 0 to 4, which counts all the previous periods where a worker was in a routine occupation at both the beginning and end of it. This is likely to be a conservative measure of experience in specific routine occupations. However, it is significantly simpler than checking over all employment histories and counting the total number of years in routine occupations. Specific experience and the time period are positively correlated, as would be expected, with a correlation coefficient of 0.672 in the NCDS study. Within each period of transition, there is sufficient variety in existing routine experience that issues relating to multicollinearity would not be an immediate concern – around 55% of the total variance of specific experience is not accounted for by age. 6 A small number of occupational categories have been updated between this paper and Holmes (2011a). This was based on feedback about the classification initially used. 7 The full table of NVQ equivalent qualifications can be found here: http://www.gos.gov.uk/497745/docs/379399/428699/469541/qualificationsguidance 4.4 Routinisation The measure of routinisation comes from data from the Labour Force Survey (LFS). Data on employment have been collected by the LFS since 1973, on a yearly basis until 1992, then on a quarterly basis. Wage data have been available from the LFS since 1993. The sample is much larger than any UK cohort study, as it looks at around 60,000 households, leading to approximately 120,000-130,000 individual observations. Using these data, the measure of routinisation in each of these time periods is the changes in employment of routine occupations across the whole labour market. shows the change in the different occupation groups over each period. Routinisation had its largest effects in the late 1980 and the early 2000s, and had much less of an effect in the 1990s. The process of routinisation is assumed to directly affect the employment share of routine workers and is exogenous. It seems sensible to look at decline rates rather than absolute declines – a 2% fall in employment share will have a much larger effect on occupational mobility if the initial employment share was 4% than if it were 20%. Thus, the routinisation variable takes the following values: Year 1981-6 1986-91 1991-5 1995-9 1999-2004 2004-8 Period of transition, t, NCDS 1 2 3 4 5 - Period of transition, t, BCS - - - 1 2 3 Change in employment share -4.40% -6.07% -4.27% -1.62% -8.56% -2.31% Rate of decline 7.37% 10.98% 8.67% 3.59% 19.76% 6.74% Routinisation 0.0737 0.1098 0.0867 0.0359 0.1976 0.0674 Source: LFS, own calculations 5 Displacement Table 4 shows the six estimated regressions for transitions, conditional on starting in a routine occupation. These regressions show, across the two cohorts, the same broad patterns found for just the single NCDS cohort as estimated in Holmes (2011a). Some higher qualifications increase the likelihood of moving from routine occupations to higher wage non-routine occupations, however, there in many cases it is not possible to distinguish significant differences in the effect of vocational qualifications below level 4 and academic qualifications at level 2 and 3. They show slightly more evidence that higher qualifications reduce the likelihood of moving to low wage service work. Moreover, the gender effects are similar to those estimated previously – women are significantly more likely to move from routine occupations to service or intermediate occupations, whilst it is more probable that men move to the other occupations or stay in routine jobs, everything else being equal. Routine occupation-specific experience and age reduce mobility. Finally, routinisation is again found to be an important driver of mobility, with periods where more routine jobs lost across the whole economy associated with a much higher probability of transitioning to both high wage and low wage non-routine occupations for both cohorts. The model includes a dummy for cohort (BCS = 1 if the individual was born in 1970), and an interaction term between the measure of routinisation and cohort8. These two variables help identify differences in patterns of mobility across the two cohorts, resulting from both career progression and routinisation. The estimates show that the younger BCS cohort is generally more upwardly mobile and less downwardly mobile, everything else being held constant (including the occupational structure, so that ROUTINISATION = 0). In addition, the BCS cohort is better educated, and so there is not just the relative effect (comparing like for like, the individuals in the BCS cohort are more mobile) but an absolute effect as well (more individuals with mobility-enhancing characteristics in the BCS cohort to begin with). Conversely, the estimation also finds that, after controlling for age, specific experience and qualifications (the human capital drivers of mobility) the BCS cohort is less likely to be displaced, and so the additional mobility generated by routinisation is small for this cohort. To an extent, this is good for the younger cohort, as they generally avoid being displaced to low wage service work. At the same time, however, they do not move to high wage non-routine work to the same degree that the older cohort do. Explaining this is a challenge – the younger cohort has been shown to be generally more mobile in terms of upward transitions from routine occupations, so it is not clear why such workers are less likely to be affected following a decline in these jobs. One possibility is that the younger cohort may achieve systematically better outcomes in these routine occupations than the older cohort, even after controlling for specific and general experience and qualifications. What the source of this advantage might be is unclear in the present analysis. Secondly, this could be evidence that the growth in non-routine occupations is biased in favour those with more general labour market experience, in the way predicted by a career mobility model, but which is not captured by the age variable. To illustrate the size of these mobility effects, Table 1 presents estimates of transition probabilities across the two cohorts, in two cases: the counterfactual world where occupational structure remained constant, and the world where 10% of routine occupations had been lost in a given period. The estimates are for a 8 For brevity, we omit including the Ai and Norton (2003) approach for correctly identifying significant marginal and interaction effects in logit models with interaction terms (see Holmes, 2011, for full discussion of this issue). The estimates here are not substantively affected by the concerns they raise. white male in the second period of transition (28-33 for the NCDS cohort, 29-34 for the BCS cohort), with one prior period of routine occupation experience. The table shows that in both worlds, the younger cohort is more mobile – in the counterfactual world, 82% of younger cohort routine workers remain in this group compared to 93% in the older cohort. However, following routinisation, this gap in mobility declines (as shown by the interaction effect column). Most noticeably, following routinisisation, far more older cohort workers move to higher wage non-routine occupations than before, while the estimated mobility of the BCS sample increases by very little. Table 1: Estimated probabilities of transition from routine occupations 0% ROUTINISATION Marginal NCDS BCS effect 10% ROUTINISATION Marginal NCDS BCS effect PROFESSIONAL 1.21% 9.63% 8.42% 3.55% 8.92% 5.37% -3.05% MANAGERIAL 3.14% 12.75% 9.61% 6.27% 12.95% 6.68% -2.93% INTERMEDIATE 2.98% 5.64% 2.67% 5.42% 7.25% 1.82% -0.84% ROUTINE 93.21% 82.08% -11.13% 85.53% 77.95% -7.59% 3.54% SERVICE 0.17% 0.05% -0.12% 0.34% 0.08% -0.26% -0.14% 6 Interaction effect Entry Given these differences in mobility patterns across cohorts out of routine occupations, an obvious question is whether they affect where workers start their working lives. Table 5 shows the results of a logit regression on occupation of employment at the beginning of the first period of the sample for each cohort (i.e age 23 for the NCDS cohort, and 25 for the BCS cohort). To include interaction between cohort and education (given that the precise role of different qualifications as mechanisms for entry into the labour market may have changed over time), we reduce the latter down to two dummies – one for academic qualifications at level 3 and above, and another for vocational qualifications at the same levels. This reduces the number of interaction terms to a more reasonable level9. The estimation shows that workers in the younger BCS cohort are less likely to start work in routine occupations. This is expected, as there are far fewer routine occupations in 1995 than there are in 1981. 9 We follow the best practice set out in Brambor, Clark and Golder (2006) and include all two-way and three-way interaction terms in the model, even where no significant effects are expected. Keeping a large number of educational qualification dummies would necessitate fourth, fifth and higher interaction terms, which would overcomplicate the analysis. The size of this change is demonstrated in Table 2, which compares the routine occupation employment share in each cohort at the same point in the lifecycle to that of the overall labour force. Encouragingly, the predicted model compares closely to the actual initial employment share of routine workers in each cohort, and shows that the proportion of a cohort initially working in a routine occupation fell by about 20%. This drop is not as large as might be expected – the routine occupation employment share across the entire labour force fell by around 25% between 1981 and 1995. Table 2: Probability of initial employment in routine occupations across cohorts COHORT (predicted) YEAR COHORT POPULATION (LFS data) Pr(ROUTINE=1) N Pr(ROUTINE=1) N Pr(ROUTINE=1) N 1981 50.3% 7430 51.9% 10188 59.7% 84471 1995 40.6% 5083 41.1% 5541 44.9% 60415 That the younger BCS cohort is not avoiding initial employment in declining routine occupations is just one result that does not fit with the predictions of the theory. Faced with a declining number of routine occupations, we would expect more educated workers in the later cohort to be employed in other higher wage non-routine jobs. This should be captured by a negative interaction effect between cohort and education. However, the estimated regression shows no evidence of this. There is no significant interaction effect between academic qualifications and cohort and positive effect of cohort on the effect of vocational qualification levels, rather than a negative one. Again, to given an idea of the magnitude of the effects, 25% of white male workers with higher academic qualifications enter routine occupations in the NCDS cohort, compared to 18% in the BCS cohort. For those without higher qualifications, the probabilities are 58% and 50% respectively. Thus, higher academic qualification are associated with far fewer entrants into routine occupations, but this relationship is not significantly different across the two cohorts. 7 Discussion and conclusion This paper has analysed several aspects of occupational mobility by comparing the experiences two UK cohort studies. We focus on entry early in working life into, and subsequent mobility out of, routine jobs. This has demonstrated a number of unexpected results that suggest a more complicated pattern of working life occupational mobility than suggested by the model discussed at the beginning of the paper. Moreover, it to a more complex version of Autor and Dorn's notion that routine jobs are 'getting older'. The younger cohort may generally be more mobile, have less specific skills and higher qualifications than the older cohort, such that they are less likely to be found in routine occupations over time. However, within this there are a number of issues. For example, while workers from the younger cohort are generally more upwardly mobile than comparable older workers at the same stage of their careers, routinisation-driven mobility tends to be larger for the latter group and not the former. Secondly, a smaller proportion of the younger cohort enter a routine job early in their working life, as many of these jobs had disappeared. However, there is no evidence that the younger cohort disproportionately avoided routine occupations, even though occupation specific skills are likely to be less valuable in the future. In some cases, there is evidence that higher skilled workers, who could have found employment in higher wage non-routine occupations, were relatively more likely to be found in routine occupations, compared to the older cohort. Why this might be is a challenge for further research. Along side this analysis of occupational mobility, it is also important to consider the mobility of wages. The literature on polarisation suggests that there has been a growth in high wage and low wage occupations and a decline in the middle. However, recent work (Holmes, 2010; Holmes and Mayhew, 2010) finds little evidence of that in the wage distributions, with the middle of the distribution still accounting for the majority of jobs. As we differentiate between the effects of routinisation on occupational and wage distributions, it is also sensible to differentiate between occupational and wage mobility to see whether the job moves captured here translate into the expected change in wages. So far, there is an implicit hierarchy of wages for each occupational group, so that we refer to moves from routine occupations to service occupations as downward mobility and to intermediate or managerial occupations as upward mobility. Differences in occupational wages across cohorts might help explain differences in mobility patterns. A second area of future research would be to consider more closely to issue of career paths. One possible explanation of the patterns of mobility found in this paper is that a number of routine occupations act as staging posts, which are attached to mechanisms allowing for career progression into higher wage nonroutine occupations. Hence, workers in the younger cohort take routine jobs early in the working life on the expectation that they will likely progress to growing high wage non-routine occupations in the future. However, this points to non human capital barriers to mobility and entry in the higher wage occupations which would need to be identified. References Ai, C, and Norton, E, (2003), Interaction terms in logit and probit models, Economics Letters, 80(1). Autor, D, and Dorn, D, (2009), This job is “getting old”: measuring changes in job opportunities using occupational age structure, American Economic Review, 99(2). Autor, D, Levy, F, and Murnane, R, (2003), The skill content of recent technological change: an empirical exploration, Quarterly Journal of Economics, 118(4). Brambor, Clark and Golder (2006), Understanding interaction models: improving empirical analyses, Political Analysis, 14(1). Goos, M, and Manning, A, (2007), Lousy jobs and lovely jobs: the rising polarization of work in Britain, The Review of Economics and Statistics, 89(1). Holmes, C, (2010), Job polarisation in the UK: An assessment using longitudinal data, SKOPE Research Paper No. 90, Cardiff: SKOPE ―, (2011a), The route out of the routine: where do the displaced routine workers go?, SKOPE Research Paper No. 100, Cardiff: SKOPE ―, (2011b) A model of occupational choice, unpublished note. Holmes, C, and Mayhew, K, (2010), Are labour markets polarising?, SKOPE Research Paper No. 97, Cardiff: SKOPE Sicherman, N, and Galor, O, (1990), A theory of career mobility, Journal of Political Economy, 98(1). Table 3: Occupational groups using SOC 3-digit categories PROFESSIONAL MANAGERIAL INTERMEDIATE ROUTINE Business and statistical professionals Health professionals Legal professionals Information and communication technology professionals Public service professionals Architects, town planners, surveyors Science professionals Engineering professionals Teaching professionals Librarians and related professionals Therapists Functional managers Production managers Protective service officers Corporate managers and senior officials Financial institution and office managers Managers in distribution, storage and retailing Managers and proprietors in hospitality and leisure services Managers and proprietors in other service industries Transport associate professionals Protective service occupations Artistic and literary occupations Business and finance associate professionals Sales and related associate professionals Public service and other associate professionals Social welfare associate professionals Science and engineering technicians Sports and fitness occupations Health associate professionals Design associate professionals Media associate professionals Administrative occupations: government and related organisations Leisure and travel service occupations Sales related occupations Draughtspersons and building inspectors Administrative occupations: finance Administrative occupations: records Administrative occupations: communications Secretarial and related occupations Electrical trades Printing trades Metal machining, fitting and instrument making trades Metal forming, welding and related trades Building trades Textiles and garments trades Vehicle trades Skilled trades nec Food preparation trades Construction operatives Mobile machine drivers and operatives Plant and machine operatives Process operatives Transport drivers and operatives Assemblers and routine operatives Elementary administration occupations Elementary process plant occupations Elementary goods storage occupations Elementary cleaning occupations Elementary personal services occupations Elementary agricultural occupations NONROUTINE MANUAL Healthcare and related personal services Childcare and related personal services Housekeeping occupations Sales assistants and retail cashiers Hairdressers and related occupations Personal services occupations nec Customer service occupations Elementary security occupations Elementary sales occupations SERVICE Elementary construction occupations Construction trades Agricultural trades Table 4: Results of logit regressions for transitions from routine occupations across cohorts GENDER NONWHITE BCS ROUTINISATION BCS*ROUTINISATION ROUTINE EXP AGE VOC LVL 0 ACAD LVL 0 VOC LVL 1 ACAD LVL 1 VOC LVL 2 ACAD LVL 2 VOC LVL 4 ACAD LVL 4 VOC LVL 5 ACAD LVL 5 CONSTANT PROFESSIONAL MANAGERIAL -0.4163 *** -(3.96) 0.2342 (0.97) 2.1661 *** (7.95) 11.0335 *** (5.52) -11.8695 *** -(5.45) -0.3933 *** -(5.82) -0.2778 *** -(4.48) -0.0719 -(0.49) -1.5142 *** -(5.38) 0.3614 * (1.85) -1.1372 *** -(5.61) -0.2384 -(1.11) -0.6318 *** -(3.86) 0.6512 *** (3.59) 1.1406 *** (6.78) 1.2044 *** (5.54) 1.7489 *** (5.41) -3.4566 *** -(14.30) -0.1347 * -(1.90) -0.1777 -(0.87) 1.5046 *** (8.73) 7.2344 *** (6.09) -7.0572 *** -(5.25) -0.2734 *** -(6.31) -0.1611 *** -(4.06) -0.1203 -(1.29) -0.8399 *** -(4.88) -0.0433 -(0.33) -0.5363 *** -(3.88) -0.4082 *** -(3.04) -0.0353 -(0.29) 0.1925 (1.39) 0.5544 *** (3.94) 0.9097 *** (5.35) 0.7490 ** (2.20) -2.8319 *** -(16.82) INTERMEDIATE 0.3965 *** (5.48) 0.1510 (0.81) 0.6675 *** (3.63) 6.2533 *** (4.66) -3.5802 ** -(2.43) -0.2255 *** -(4.83) -0.2596 *** -(6.12) -0.0541 -(0.56) -1.1310 *** -(6.32) -0.1380 -(1.00) -0.7484 *** -(5.54) -0.1536 -(1.18) -0.2853 ** -(2.49) 0.3141 ** (2.21) 0.2672 * (1.91) 0.0850 (0.37) 0.3033 (0.83) -2.7395 *** -(15.59) ROUTINE -0.3535 *** -(8.33) 0.0621 (0.52) -1.0976 *** -(10.83) -8.4230 *** -(12.73) 5.8310 *** (7.51) 0.3009 *** (11.84) 0.1471 *** (6.32) -0.0593 -(1.01) 0.4801 *** (5.04) -0.0714 -(0.88) 0.3987 *** (4.76) 0.1068 (1.38) 0.0639 (0.85) -0.3575 *** -(3.89) -0.8201 *** -(8.73) -0.9807 *** -(7.81) -1.2644 *** -(4.91) 2.0244 *** (19.55) SERVICE 2.0193 *** (19.34) -0.3873 -(1.48) -1.1782 *** -(5.96) 6.5980 *** (6.26) -2.5963 ** -(1.98) -0.1532 *** -(3.54) 0.0540 (1.29) 0.5285 *** (4.01) 0.7703 *** (3.73) 0.3213 ** (2.02) 0.7148 *** (3.67) 0.4139 *** (2.67) 0.7282 *** (3.97) 0.2159 (0.96) 0.3429 (1.48) 0.3457 (1.03) -0.8524 -(0.83) -6.3050 *** -(24.96) MANUAL NON ROUTINE -2.1778 *** -(10.21) -0.0760 -(0.21) 3.8240 *** (10.87) 6.1810 *** (3.00) -3.1571 -(1.37) -0.2397 *** -(3.24) -0.0180 -(0.24) -0.0532 -(0.32) 1.0699 *** (3.01) 0.1990 (0.85) 0.6951 ** (2.03) 0.1032 (0.49) 0.4550 (1.35) -1.1745 ** -(2.49) -0.3223 -(0.71) 0.3602 (0.75) 0.1767 (0.17) -4.9948 *** -(12.48) Table 5: Results of logit regression on initial occupation across cohorts GENDER NONWHITE BCS ACAD LVL 3-5 VOC LVL 3-5 BCS * ACAD LVL 3-5 BCS * VOC LVL 3-5 ACAD LVL 3-5 * VOC LVL 3-5 BCS * ACAD LVL 3-5 * VOC LVL 3-5 CONSTANT ROUTINE 0.1022 *** (2.71) 0.0843 (1.05) -0.3195 *** (6.71) -1.4466 *** (19.08) -0.4736 *** (7.24) -0.0924 -(0.87) 0.4238 *** (3.44) 0.5921 *** (4.30) 0.1121 (0.40) 0.3225 *** (9.19)