The Task Content of Occupational Transitions over the Business Cycle: Evidence for the UK

Transcription

1 The Task Content of Occupational Transitions over the Business Cycle: Evidence for the UK Aspasia Bizopoulou Rachel Forshaw Preliminary Draft - Please do not circulate Abstract Using quarterly data from the U.K. Labour Force Survey, which we match to the O*NET dictionary of tasks for the period 1997q1-2016q2, we view occupations as bundles of tasks of varying skill levels and adapt a measure from the mismatch literature in order to understand the task and skill content of occupational transitions over the business cycle. We show that the degree of task and skill change between occupations in employment-to-employment transitions is large and pro-cyclical, even when controlling for composition effects. Furthermore, the burden of re-adjustment over the cycle falls on cognitive, rather than manual tasks within occupations. Keywords: occupational mobility, tasks, business cycles, mismatch, manual versus cognitive tasks. JEL Codes: J62, E32 We would like to thank our advisors Mike Elsby and Maia Güell. We also thank Philipp Kircher, Eva Pocher, Jean-Marc Robin, Anna Salomons, Ludo Visschers, as well as participants from the Edinburgh Reading Group, the Scottish Government Economics Seminar Series, the SGPE conference, and the IAB Nuremberg conference for their helpful comments and suggestions. Any remaining errors are our own. University of Edinburgh, 30 Buccleuch Place, EH8 9JT Edinburgh, UK; a.bizopoulou@sms.ed.ac.uk University of Edinburgh, 30 Buccleuch Place, EH8 9JT Edinburgh, UK; R.J.Forshaw@sms.ed.ac.uk 1

2 1 Introduction As employers demands change over the business cycle, a well-functioning labour market should efficiently reallocate workers of differing skills into the occupations to which they are best suited. While there has been a large amount of research focused on the outcomes of occupational mobility in the labour market, less is known about the task and skill composition of such occupational changes. In this work, we study the cyclicality of both task and skill changes among occupational switchers. We find that the degree of task and skill change between occupations is large and pro-cyclical, even when controlling for composition effects. We also show that the burden of re-adjustment over the cycle falls on cognitive, rather than manual tasks within occupations. Occupational mobility has proven to be an important tool in the study of wage progression, sorting, mismatch and the consequences of the business cycle 1. Our own approach adds to the existing discussion by looking at the mobility between occupations in terms of their task and skill content. In characterising the occupation as a vector of tasks, we follow an already established line of work (e.g Autor et al. (2003); Autor and Handel (2013); Ingram and Neumann (2006); Poletaev and Robinson (2008); Gathmann and Schönberg (2010); Yamaguchi (2012)) This approach has been adopted ever since the literature on job polarisation has highlighted the importance of studying the task composition of occupations, given that technological change has unequally affected the demand for cognitive versus manual tasks. 2 Previous studies taking the task approach have focused on the specificity of human capital across occupations and its impact on earnings. Our first contribution is to add to this literature by studying whether the specificity of human capital is affected by the business cycle. Existing research focuses on longer-term life-cycle effects rather than business cycles when looking at task distance of occupational moves. We find that the business cycle has an economically significant effect on the quality of observed occupational moves. Our second contribution is to show that the burden of adjustment in tasks falls overwhelmingly on cognitive tasks, rather than manual tasks. Our third contribution is in extending the concept of occupational distance to UK data - 1 See respectively Kambourov and Manovskii (2009); Groes et al. (2014); Fredriksson et al. (2016); Carrillo-Tudela et al. (2016) 2 See for example Autor et al. (2003); Autor et al. (2008); Goos and Manning (2007); Goos et al. (2009); Cavaglia and Etheridge (2017) 2

3 previously it has only been studied for the US and Germany. Using UK data allows us to study the effect of the 2008 financial crisis on the task content of transitions - something that would be much more challenging with German data, since the unemployment rate did not rise during that period. A recurring debate within the study of the effect of business cycles is whether they have a sullying or cleansing effect on the labour market. In one line of argument, the frictions that accumulate during expansions are cleansed during recessions by the speeding up of the process of reallocation of workers (e.g. Lilien (1982); Mortensen and Pissarides (1994); Jaimovich and Siu (2014)). An alternative view, put forward by Barlevy (2002), is that since employment-to-employment (E2E) transitions are pro-cyclical it is during economic expansions and not during recessions that labour reallocates better. Thus, in the second view, recessions have a sullying effect on worker reallocation. The evidence on 1-digit occupational moves for the UK by Carrillo-Tudela et al. (2016) is in line with Barlevy (2002). They find that recessions have a sullying rather than cleansing effect since they prevent workers from changing 1-digit occupations at a wage gain. While we do not directly study wages, our approach allows us to say something about the effect of the recession on the realisation of riskier hires and the extend of up-/down-skilling of new hires. Concerning the former, we find during recessions new hires tend to come from more similar occupations to what they were doing before, relative to in good economic conditions, which is in line with Carrillo-Tudela et al. (2016). At the same time, we find that the there is a significant decrease in the change of skills in occupational transitions during recessions. The rest of the paper is organised as follows: section 2 reviews the literature related to this study. Section 3 describes the data used. Section 4 provides an overview of the measures used to ascertain occupational distance in terms of tasks and skills. Section 6 details the methodology behind our reduced form estimation. Section 5 presents summary statistics of the sample used. In section 7 we present the results. Finally, section 8 concludes with some avenues for future research. 3

4 2 Related Literature Characterising an occupation as a vector of tasks is a relatively recent but already well-established practice in the literature studying job transitions. Among applied papers, Poletaev and Robinson (2008) are one of the first to map occupational titles to tasks from the US Dictionary for Occupational Titles. Using factor analysis, they group the tasks into four major categories and subsequently rank them by the intensity that they are used in each occupation. They study occupational switches for displaced workers, which they define as the situation when the new occupation employs the previous occupation s main skill with much lower or much higher intensity. Using this definition of occupational moves, they find that wage losses are closely associated with switching skill portfolio, in particular a decrease in the skills. In a similar vein, Gathmann and Schönberg (2010) also construct a measure of occupational distance based on tasks, which we use in this paper. Using German administrative data, they find that individuals tend to switch to occupations with similar task requirements, and task distance of occupational moves tends to decrease over time. Our own paper differs from these in three ways: first, rather than studying the stylised facts of longitudinal occupational moves, we focus on cohorts and study how the task distance as well as the direction of skill of moves changes when a cohort is hit with a recession. Second, rather than using US or German data, we generate a working dataset incorporating a mapping between O*NET tasks and UK occupations within the UK Labour Force Survey 3. Using UK data allows us to dive into the effects of the recession, since unemployment did significantly rise for the UK during the 2008 financial crisis, while it did not in Germany. Finally, we augment the measure provided by Gathmann and Schönberg (2010) to include not only occupational distance in terms of tasks, but also the direction of the move in terms of the overall skill level. Contemporaneously, Cortes and Gallipoli (2017) also use the concept of task distance but within a different context. Instead of taking the task distance at face value, they interpret it as the cost of occupational mobility. The main idea is that the larger the task distance between two occupations, they greater the cost of moving from one to the other occupation and, as such, the smaller the 3 The mapping of O*NET tasks and UK occupational codes (SOC codes) is available from the CASCOT software developed by the University of Warwick 4

5 ratio of movers to stayers. They borrow a gravity model from the trade literature, where job-to-job flows are aggregated at the 2-digit level and are assumed to behave similar to bilateral trade. The traditional geographical distance is replaced with the task distance, while destination and origin country fixed effects are now destination and origin 2-digit occupational fixed effects. Thus, their analysis is aggregated at the occupational level, and they find that the ratio of movers to stayers is negatively affected by greater task distance. Our approach is different in that we want to understand the effect of business cycles on the distance of occupational moves at the individual level. The second relevant strand of literature is on the cyclicality of occupational mobility. Carrillo- Tudela et al. (2016) look at the propensity for individuals to change 1-digit occupations over the cycle and find that the probability of a career change co-moves positively with the cycle. In addition, they find that outside of recessions movers receive higher wages than those who do not change careers. Their conclusion is that the recession likely has a sullying effect, in that it potentially prevents individuals from making advantageous moves. Our own paper departs from the notion of career change and instead looks at an occupation as task bundle and studies the cyclicality of task distance. Until now, work on the cyclicality of tasks has been relatively sparse. Devereux (2000) offers an early study of the cyclicality of task assignment within the firm. Using US data he finds that firms tend to re-assign individuals to tasks of lower quality during recessions. His work focuses on task reassignment within the same employer, while we instead look at the cyclicality of task distance between different employers. Using Canadian data, Summerfield (2016) shows that an increase in the unemployment rate leads to an increase in the share of tasks that are classified as manual. Our own results for the UK point to a different direction: we find that during recessions individuals tend to move to jobs with more similar tasks to what they were doing before than outside of recession, and this is driven by smaller changes in cognitive tasks rather than manual tasks. 3 Data We use the UK Quarterly Labour Force Surve (LFS) for the years 1997q1-2016q2, which we match to the US O*NET, a dictionary of the task content of occupations. Our aim is to obtain a detailed task 5

6 profile for each occupation and to subsequently measure the distance between different occupations based on task similarity. 3.1 UK Labour Force Survey (LFS) In the LFS the respondents are followed over five quarters, and each quarter a fifth of the sample is replaced by an incoming group. The sample comes in two formats: individuals are either followed over 2 quarters or 5 quarters. Ideally, we would have used the 5 quarter dataset, since it follows individuals over longer periods of time - however, the 5 quarter dataset is much smaller compared to the 2 quarter and raises concerns over attrition bias. As soon as individuals change address in the LFS, they are dropped from the longitudinal sample, which could introduce bias since the individuals that change address are unlikely to be randomly selected. We are primarily interested in individuals occupational transitions and career changes, so we are able to use the LFS 2Q instead of the 5Q. The advantage of the 2Q sample over the 5Q is that we have a much larger number of observations per quarter, approximately 60,000 individuals. The survey contains a large amount of data, including information on individuals employment status, employment SOC (standard occupation classification code), employer tenure, gender, education, marital status & children. The series are weighted using census population weights provided by the Office for National Statistics (ONS) Harmonisation of SOC codes in the UK LFS Over the period of study, the LFS modified its occupational categories twice. One set of occupational definitions runs from 1997q1-2000q4, the second from 2001q2-2010q4, and the third from 2011q1-2016q2. The codes were updated periodically to account for changing requirements within SOC occupation classifications. Since the SOC classifications are integral in the mapping to task and skill data, the change in the definitions could lead to spurious results. Figure 1 shows a mapping of the SOC1990 to SOC2000 and the SOC2000 to SOC2010, which highlights the changing SOC classifications over time. We can see that in the 90s, the codes only had up to 3-digit numbers, while in the 2000s and 2010s the digits increased to four. The move from the 1990s to the 2000s involved splitting up occupations that were previously under the same code into a greater number of classifications, while the move to the 2010s involved a major re-organisation of occupations within 6

7 the codes. In the latter recoding, several codes split into finer occupations in the 2010s, while other codes disappeared, highlighting the redundancy of certain occupations over time SOC code SOC90 SOC2000 SOC2010 SOC classification Figure 1: Mapping between all SOC90, SOC2000 and SOC2010 codes For robustness, we explore a number of different approaches for harmonising the series across time. Previous literature uses the minimum common denominator of occupational codes and applies it to the entirety of the series. The main idea of the approach is as follows: if a SOC2000 code split into two SOC2010 codes, code the two separate SOC2010 codes as one single occupation for the whole series. A similar approach has been used by Dorn (2009) for the US Current Population Survey. Unfortunately, we are not able to apply this harmonising technique, due to the fact that the mapping between SOC code crosswalks is many-to-many 4. For our main analysis, we utilise a tool developed by the Warwick Institute For Employment Research, CASCOT (Computer-Assisted Structured Coding Tool) 5. CASCOT is a computer program designed to make a semantic match between occupational titles and standard occupational 4 More details about why this procedure does not work can be found in Appendix A. 5 More information available at 7

8 codes. Using CASCOT, we are able to exploit the data already available in the US O*NET dictionary of tasks in order to classify the task content of UK occupational codes. This mapping is performed by comparing text descriptions of UK SOC2010 occupations to text descriptions of O*NET occupations for the 2000s and the 2010s, and DOT occupations in the 1990s. It creates task mapping that is internally consistent for the 1990s, the 2000s and the 2010s, without the need to find matching 2010 codes which preserves the contemporaneous task profile of occupations. After the application of CASCOT, the problem is reduced to a series of visible breaks in the data at the time when the different SOC codes where introduced. To overcome the destabilising effect of the breaks in the analysis, we control for them by a adding a set of macro dummies to the regression. As a robustness check we use a simple averaging process, which converts SOC1990 and SOC2000 codes into their matching SOC2010 codes. This works as follows: if a SOC2000 code split into two SOC2010 codes according to the definitions provided by the SOC (as shown in figure 1), take a simple unweighted average of the task vectors associated with the two SOC2010 codes. Results are qualitatively unchanged by using this different measure. 3.2 Occupational transitions over time: the years post-2010 in the LFS Looking at the series from , we observe an unusual drop in the number of transitions in the early 2010s. This drop covers 6 quarters from 2011q1-2012q2. Figure 2 plots the estimated probability of career change at the 1-, 2-, 3-, or 4- digit level. Following Carrillo-Tudela et al. (2016), the probability of career change at the k-digit (k=1,2,3,4) is estimated as: Prob Career Change k = E2Ek m E2E k s The number of individuals that made an employment-to-employment (E2E) move (m), defined as changing k-digit codes divided by the number of of individuals that transitioned employment-toemployment (E2E) but stayed (s) within the same k-digit code. For example, an Economist is in category 2 at the one-digit level. If she changed occupations to become a Florist (category 5 at the 8

9 one-digit level), this would be classed in the numerator as an E2E move. If, however, she became a management consultant (category 2), this would be classed in the denominator as a one-digit stay. Figure 2 clearly shows that the drop in the series occurs at all occupation classification levels from least (1) to most (4) granular. At the time of writing, previous papers using the UK LFS to study occupational transitions stop the analysis in 2010, the year when the SOC code change was introduced. We look at the probabilities of career change at the 1-digit, 2-digit, 3-digit and 4-digit level. Up until 2010, the 1-digit data follows a similar pattern to Carrillo-Tudela et al. (2016). Extending the series beyond the 2010s, using 2010-denominated SOC codes, we see that the sharp drop in the probability of changing careers is present in all different digit denominations. It seems clear that this is not a real phenomenon, but a data anomaly. Our suspicions were raised by the fact that the probabilistic matching of SOC2000 and SOC2010 codes that is provided by the LFS is only available from 2012q2 onwards, which also coincides with the exact moment that the occupational transition series stabilises again. However, the official switch from SOC2000 and SOC2010 happened in 2011q1. Thus, during the 6 quarters between 2011q1 and 2012q2, there is no official harmonisation provided for how to interpret occupational transitions based on SOC2010 codes. We believe that these six quarters straight after the switch to SOC2010 should in fact still be interpreted in SOC2000 and harmonised backwards. More details on how we correct for this data anomaly are provided in section US O*NET The U.S Department for Labor s O*NET dataset provides us with a detailed picture of the tasks that are used in occupations in the US. The O*NET contains task profiles for 974 occupations, which can then be mapped onto the 374 SOC2010 occupations of the UK LFS 6. As discussed above, the mapping between O*NET and SOC codes is completed using CASCOT. As can be seen from the number of US and UK occupational categories, the mapping is not one-to-one, but one-tomany. In order to get a single task vector for each SOC code, we use a confidence-weighted average over all matching O*NET occupations, where the confidence weights are provided by the CASCOT 6 There are 374 SOC 2010 occupations. The number is 352 for 1990s SOC codes and 372 for 2000s SOC 9

10 q1 2000q2 2000q3 2000q4 2001q1 2001q2 2001q3 2001q4 2002q1 2002q2 2002q3 2002q4 2003q1 2003q2 2003q3 2003q4 2004q1 2004q2 2004q3 2004q4 2005q1 2005q2 2005q3 2005q4 2006q1 2006q2 2006q3 2006q4 2007q1 2007q2 2007q3 2007q4 2008q1 2008q2 2008q3 2008q4 2009q1 2009q2 2009q3 2009q4 2010q1 2010q2 2010q3 2010q4 2011q1 2011q2 2011q3 2011q4 2012q1 2012q2 2012q3 2012q4 2013q1 2013q2 2013q3 2013q4 2014q1 2014q2 2014q3 2014q4 2015q1 2015q2 2015q3 2015q4 2016q1 2016q2 1 digit 2 digit 3 digit 4 digit Figure 2: Probability of Career Change at 1-,2-,3- and 4-digit Probability of career change as estimated as the ratio of E2E movers that changed a) 1-digit, b) 2-digit, c) 3-digit, d)4-digit occupation to those E2E that stayed within the respective occupational digit. software. Each O*NET occupation has scores for each of the 147 tasks in terms of the level of a given skill needed to perform a job (possible scores are in the range 0-7) and the importance of that skill in that occupation (possible scores are in the range 0-5). Figure 3 shows a simple example of how the averaging process works. Taking the example of the SOC2010 code that covers occupation Economist, we see that code 2425 also covers a number of other occupations, including Actuary and Bioinformatician. Taking an average over all of the Oral Comprehension scores for the different O*NET occupations gives a single score for this task in the the SOC2010 code, which we then repeat for all tasks. 4 Measuring occupational distance 4.1 Task Distance Commonly-used measures in the mismatch literature require data on either skills or education of individuals that is unfortunately unavailable in the LFS. Examples are by Guvenen et al. (2015) who have skills data of individuals obtained from cognitive tests as well as the same O*NET task data that we use; Fredriksson et al. (2016) who have individual skills data from cognitive tests; Lindenlaub (2014) who also uses the O*NET task data combined with the subject matter of individuals postsecondary education; Bizopoulou (2017) uses the PIAAC dataset which has both cognitive skills of 10

11 Figure 3: Example of mapping the SOC2010 to the O*NET The SOC2010 code that covers occupation Economist is 2425 and also covers a number of other occupations, including Actuary and Bioinformatician. Code 2425 maps to multiple O*NET occupations. Taking an average over all of the Oral Comprehension scores for the different O*NET occupations gives a score for the SOC2010 code. individuals and job tasks in the cross-section. To measure the task distance between two occupations we use the measure of angular separation of Gathmann and Schönberg (2010), a measure which has also been used in the innovation literature (Jaffe (1986)). The measure is as follows: angsep o,o = 1 T t=1 (q t,o q t,o ) [ ( T t=1 q2 t,o) ( ] 1 T t=1 q2 t,o ) 2 [0, 1] (1) where o, o is a pair of different occupations, t is tasks, q t,o represents the importance-weighted skill score of a task t within occupation o. Intuitively, the task distance between a set of occupations o and o are compared by measuring the angle between their respective vectors. The scores in the vectors range between 0 (this task is not used in an occupation) and 7 (this task is used at the highest level), which we normalise within [0, 1], so that the entire measure is within [0, 1]. Each occupation is represented by a vector of equal length dimension and each element of the vector gives a task score, i.e. the intensity with which the task is used in the given occupation. Some of the elements of the vector are zeros, since occupations do not use all available tasks. 11

12 4.2 Skill distance To capture the degree of up- or down-skilling between occupations, we propose the measure skillscore which takes into account the differences in magnitude between two vectors: [ T skillscore o,o = t=1 (q 2 t,o) ] 1 2 [ T ] 1 (qt,o 2 ) / T [ 1, 1] (2) t=1 2 Equation 2 calculates the relative length of two occupation task vectors and has range [ 1, 1]; -1 means that moving occupations results in complete deskilling in every task, 0 means two occupations are equally skilled in every task, 1 complete upskilling in every task. The measure is therefore symmetric, i.e. skillscore AB = 1 skillscore BA. 1.0 A Task B C Task 1 Occupation move angsep skillscore A B C A A D A A Figure 4: An example of Angular Separation and Skill Score with 2 tasks and 5 occupations D Figure 4 shows an example in which there are a total of five different occupations (A, B, C, D and E) which comprise two tasks, task 1 and task 2. Moving from occupation A, which is highly 12

13 skilled in task 1 and task 2, to B, which is lowly skilled both tasks gives an angular separation of 0, since the tasks are still used in the same proportion. The skillscore of reflects the fact that occupation B is much lower skilled than A. Moving from occupation C to A represents both a change in tasks and upskilling, whereas the change in tasks from A to D constitutes downskilling. Finally, moving from and to the same occupation A results in zeros for both measures. 4.3 Empirical series Figures 5 and 6 show angular separation and skill score averaged within quarters over our sample. The summary statistics reported are averaged over the entire time period, 1997q1-2016q2. The blue lines in each of the figures are the raw unadjusted series, which display 3 structural breaks created by the different definitions of SOC codes between the 1990s, 2000s and 2010s as discussed in section 3.1.1, and the unusual drop in transitions over time as discussed in section 3.2. The red series show the adjustment as achieved by using a simple time dummy to shift the mean of the series in line with the data in the 2000s angsep Mean SD Min 0.0 Max q1 2000q1 2005q1 2010q1 2015q1 date (mean) angsep_cascot (mean) angsep_cascot_adj Figure 5: Angular Separation averaged within quarters, pre- and post-adjustment 13

14 q1 2000q1 2005q1 2010q1 2015q1 date skillscore Mean SD Min Max 0.22 (mean) skillscore_cascot (mean) skillscore_cascot_adj Figure 6: Skill Score averaged within quarters, pre- and post-adjustment 5 Summary statistics for E2E transitions A consideration with the design of our study is that the type of individuals that transition during and outside of recessions are significantly different, leading to our results being driven entirely by differences in composition. In table 1 we outline a set of summary statistics of observable characteristics for the individuals observed during and outside of a recession. The sample of study is made up of individuals experiencing an E2E transition, without any unemployment spells in between. Most differences in the means of characteristics are not statistically significant - however, there are a number of exceptions. During recessions there are fewer women and fewer highly educated individuals among our sample, while at the same time there are more lower-educated individuals. The duration spell of the previous job also tends to be shorter among those who transition in a recession and there are fewer individuals transitioning after being fired in a recession. In our regressions, we control for all the variables listed in table 1. While most of them significantly affect the task distance of occupational moves, we see from Table 1 that most are similar for the recession and non-recession samples. This is reassuring, since it suggests that the results are driven by more than compositional effects alone. 14

15 Table 1: Difference in means between E2E during and outside of a recession Variable No recession Recession Difference female ** (0.00) (0.00) (0.006) age (0.10) (0.11) (0.146) married ** (0.00) (0.00) (0.006) spell duration ** (0.02) (0.02) (0.029) full-time (0.00) (0.00) (0.005) temp (0.02) (0.02) (0.031) self-emp (0.00) (0.00) (0.002) self-emp * (0.00) (0.00) (0.003) public (0.00) (0.00) (0.004) public (0.00) (0.00) (0.004) high edu *** (0.00) (0.00) (0.006) med edu (0.00) (0.00) (0.006) low edu *** (0.00) (0.00) (0.004) job centre (0.00) (0.00) (0.003) ads (0.00) (0.00) (0.005) direct app (0.00) (0.00) (0.002) family/friend (0.00) (0.00) (0.002) other (0.00) (0.00) (0.002) quit (0.00) (0.00) (0.005) fired ** (0.00) (0.00) (0.006) other (0.00) (0.00) (0.006) N Significance levels: < 10% < 5% < 1% 2 Standard errors in parentheses

16 6 Econometric Model 6.1 The Effect of Business Cycles on the Task Distance and Skill Distance of Occupational Moves We use a reduced form Tobit model to test for the effect of the business cycle on the size of the task distance of occupational moves as well as absolute level of skill distance. We choose to use a Tobit, since our measures of task and skill distance are continuous on (0, 1], but with a large portion of the sample (approximately 40%) censored at zero. Therefore, using a simple OLS model would lead to biased results. In practice, an angular separation or skill score of zero only occurs where an individual experiences an E2E transition but remains within the same SOC code classification. Below we summarise the reduced form model. y i = max(0, β 1 agg urate t + k β k X k,i + j α j Region j + e i + u t ) (3) where y i is the dependent variable, which is either angsep i or abs(skillscore i ). Here, i represents individuals, t is time, j are the number of regional dummies and k are the number of individual controls. The variable agg urate t is our main independent variable, the aggregate unemployment rate which best captures the effects of business cycles on labour markets. Increases in agg urate t are interpreted as labour market tightness. We add a set of controls commonly found in the literature of occupational transitions. We first add a set of demographic characteristics, namely age and age squared, marital status, gender, level of education 7. We also add a set of variables related to the individual s previous job: the duration of the previous employment and whether the separation was voluntary/involuntary or related to retirement. Controls for the current job include whether the job is temporary, whether it is part- or full-time, self-employed, and in the public or private sector. Finally, we have a set of controls for the method by which the individual search for new jobs: through a job centre, ads, direct applications, family/friends, or some other method. 7 We split education into low, medium and high. In the low category we only include individuals with no qualifications whatsoever; in the middle we include those with at least an Entry Level Qualification and at most A levels (a UK pre-requisite for university entry); and in the high we include all those with any qualification above A levels. 16

17 We also add a set of macro controls. As discussed in sections 3 and 4.3, there are three breaks in the series, two of which are due to SOC code name changes, when the ONS updated the codes between 2000 and 2001 and between 2010 and 2011, while the third is a data entry mistake. We include a set of three dummies to account for these breaks. Our strategy involves taking the series in the 2000s as our reference period and adding a set of dummies for each the other periods, i.e. one for before 2001q1 (first SOC code change), one for between 2011q1 and 2012q2 (second SOC change) and one for after 2012q2 (no SOC code change). The rest of the macro controls are a set of dummies marking quarters to control for seasonality as well as a set of regional dummies to capture regional differences within the UK. Finally, we test for the separate effect of the recession on routine and non-routine as well as cognitive and manual tasks. This is achieved by categorising all tasks in the O*NET as cognitive or manual, according to the definitions by Autor et al. (2003) Interpreting Tobit coefficients While in an OLS regression the marginal effect is simply Ey X j be written as follows: = β j, in Tobit the marginal effect can E(y x) = P (y > 0 x) x j E(y x, y > 0) x j } {{ } A P (y > 0 x) + E(y x, y > 0) x j }{{} B = P (y > 0 x)β j }{{} C The above decomposition, proposed by McDonald and Moffitt (1980), has the interpretation that the left hand side is the marginal effect. Term A is the change in the dependent variable, y, of those above zero, weighted by the probability of being above the zero limit, while term B is the the change in the probability of being above zero, weighted by the expected value of y. Using a simple OLS regression would leave A in the error term, leading to bias. It can be shown that this entire expression simplifies to term C, which has a very simple interpretation: it is the marginal effect of x i on y, weighted by the probability that y > 0 conditional on the independent variables. Looking at term C, we can see that estimated betas from the regressions will not be an accurate 8 Autor et al. (2003) further separate cognitive and manual tasks into routine/non-routine, but we find only the difference between cognitive and manual tasks to be important for the business cycle. 17

18 estimate of the marginal effect - we still have to weight them by P (y > 0 x). Assuming we have estimates for β and σ we can get an estimate for this term using the average partial effects (APE) factor: Ê(y x) x j = 1 N Φ(x i ˆβ/ˆσ) ˆβ j N i=1 }{{} APE scale factor Thus, for the purpose of interpretation of marginal effects, all beta coefficients have to be multiplied by the APE scale factor which has the interpretation ˆP (y > 0 x). 7 Results Table 2 details the regression results based on the regression specification of equation 3. The first row shows the regression coefficient of interest: on the aggregate unemployment rate. For task changes, an increase in the unemployment rate of one percentage point leads to a reduction in task change ( times the APE factor, , first column of table 2). This effect is significant at the 5% level, and economically significant, representing approximately one and a third standard deviations in angular separation. This suggests that in recessions, individuals make much smaller task changes when changing occupations in E2E transitions than during good economic times, even when controlling for composition effects. In terms of skill changes, the effect is in the same direction and of a similar magnitude. An increase in the aggregate unemployment rate results in a ( times the APE factor, ), fourth column of table 2. This suggests that individuals make smaller skill moves when changing occupations in E2E transitions than during good economic times, even when controlling for composition effects. Table 2 also details the effects of recessions on the cognitive and manual components of occupations. In both task changes and skill changes, recessions lower the distance of moves in cognitive occupations by and respectively. However, manual components of occupations are unaffected by recessions in both task and skill changes, as neither coefficient is significantly different from zero. 9 This suggests that changes in cognitive tasks and skills take the burden of 9 Note that because both the Tobit regression and the measures of task and skill change, equations 1 and??, are 18

19 adjustment in recessions. Control variables are, for the most part, significant and of the sign predicted by theory. For example, the longer an individual spent in their previous occupation (spell durat), the smaller task and skill changes they make, suggesting that individuals become more specialised over a longer tenure. The effect is in the same direction for high- and medium educated workers relative to loweducation workers (H edu and M edu), for those moving to a full-time occupation, relative to a part-time position (ft job), and for those who did not voluntarily leave their previous occupation (invol). Task moves for females relative to males (female) are larger, as are for those moving to temporary positions (temporary), occupations in the public sector (public) or moving to selfemployment (selfemp), suggesting that these types of workers must be more flexible in terms of tasks and skill differences when moving occupations. Finally, it seems that all methods of search 10 increase the task and skill distance of occupational moves, relative to not searching. nonlinear we should not expect the coefficients on cognitive and manual changes to add to the overall effect, labeled All in Table Methods of search are: job center: via a job centre, Ads: applied to adverts, Direct app: applied directly to the employers, family/friend: asked family or friends, other method: other method of search. 19

20 Table 2: Tobit Regression Results Angular Separation Skill Scale All Cognitive Manual All Cognitive Manual agg urate (-2.10) (-2.68) (-0.93) (-2.37) (-2.31) (-1.25) reg agg urate (-1.01) (-1.06) (-0.30) (-0.46) (-0.61) (-1.83) dummy (-8.65) (-6.67) (-9.67) (-6.49) (-6.12) (-7.46) dummy (-5.56) (-5.50) (-5.40) (-5.43) (-5.26) (-5.71) dummy (-1.18) (-1.05) (-1.67) (-0.24) (0.03) (-0.88) female (2.91) (1.23) (9.40) (6.18) (5.32) (1.12) age (-11.51) (-12.24) (-10.03) (-11.22) (-11.02) (-10.11) age sq e e e e e e-05 (9.33) (9.98) (8.17) (9.04) (8.97) (7.97) married (-1.53) (-2.01) (-1.14) (-1.76) (-1.81) (-0.57) H edu (-9.26) (-10.56) (-1.33) (-7.86) (-8.85) (-7.16) M edu (-3.49) (-4.40) (1.84) (-2.56) (-3.31) (-1.26) spell durat (-7.05) (-7.58) (-6.38) (-5.98) (-6.23) (-5.89) ft job (-2.50) (-3.78) (-1.12) (-5.68) (-6.24) (-0.62) temporary (5.25) (5.01) (6.21) (3.82) (4.56) (4.78) public (5.37) (2.75) (4.73) (4.34) (2.71) (3.48) selfemp (6.57) (5.85) (7.37) (3.85) (4.78) (5.87) invol (-2.01) (-1.57) (-3.24) (-2.05) (-1.47) (-1.96) other (-4.18) (-3.81) (-4.30) (-3.88) (-4.01) (-3.75) job center (1.49) (1.90) (1.79) (2.80) (2.61) (1.78) Ads (11.15) (10.31) (10.37) (9.31) (9.54) (10.31) Direct app (1.44) (1.54) (0.57) (2.09) (2.74) (1.13) family/friend (2.95) (3.50) (1.42) (2.05) (2.05) (1.11) other method (3.37) (3.82) (3.12) (4.11) (3.86) (2.52) Quarters Yes Yes Yes Yes Yes Yes Regions Yes Yes Yes Yes Yes Yes N APE pseudo R Log llik p < 0.10, p < 0.05, p < 0.01

21 8 Conclusion In this paper we extend the concept of task distance to UK occupational transitions and show that the degree of task and skill change between occupations in employment-to-employment transitions is large and pro-cyclical, even when controlling for composition effects. We show that the burden of re-adjustment over the cycle falls on cognitive, rather than manual tasks within occupations. These findings are in line with the notion that recessions have a sullying effect on labour markets. In future work we plan to extend this research to study the effect of recessions on the task distance of occupational transitions that originate from unemployment or inactivity. A further interesting extension would be to understand the relation between task and skill distances and wages. 21

22 References Autor, D. H. and Handel, M. J. (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. (2008). Trends in U.S. Wage Inequality: Revising the Revisionists. Review of Economics and Statistics, 90(2): Autor, D. H., Levy, F., and Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4): Carrillo-Tudela, C., Hobijn, B., She, P., and Visschers, L. (2016). The extent and cyclicality of career changes: Evidence for the u.k. European Economic Review, 84: European Labor Market Issues. Cavaglia, C. and Etheridge, B. (2017). Job Polarization, Task Prices and the Distribution of Task Returns. Cortes, G. M. and Gallipoli, G. (2017). The Costs of Occupational Mobility: an Aggregate Analysis. Journal of the European Economic Association, 00(January):1 41. Devereux, P. J. (2000). Task assignment over the business cycle. Journal of Labor Economics, 18(1): Dorn, D. (2009). Essays on Inequality, Spatial Interaction, and the Demand for Skills. Fredriksson, P., Hensvik, L., and Skans, O. N. (2016). Mismatch of Talent: Evidence on Match Quality, Entry Wages and Job Mobility. Gathmann, C. and Schönberg, U. (2010). How General Is Human Capital? A TaskBased Approach. Journal of Labor Economics, 28(1):1 49. Goos, M. and Manning, A. (2007). Lousy and Lovely Jobs: The Rising Polarization of Work in Britain. The Review of Economics and Statistics, 89(1): Goos, M., Manning, A., and Salomons, A. (2009). Job Polarization in Europe. American Economic Review PP, 99(2):

23 Groes, F., Kircher, P., and Manovskii, I. (2014). The u-shapes of occupational mobility. The Review of Economic Studies, 82(2): Guvenen, F., Kuruscu, B., Tanaka, S., and Wiczer, D. (2015). Multidimensional skill mismatch. Working Paper 21376, National Bureau of Economic Research. Ingram, B. F. and Neumann, G. R. (2006). The returns to skill. Labour Economics, 13(1): Jaffe, A. (1986). Technological opportunity and spillovers of rd: Evidence from firms patents, profits, and market value. American Economic Review, 76(5): Jaimovich, N. and Siu, H. E. (2014). The trend is the cycle: Job polarization and jobless recoveries. Technical report, National Bureau of Economic Research. Kambourov, G. and Manovskii, I. (2009). Occupational Specificity of Human Capital. International Economic Review, 50: Lilien, D. M. (1982). Sectoral shifts and cyclical unemployment. Journal of Political Economy, 90(4): Lindenlaub, I. (2014). Sorting Multidimensional Types: Theory and Application. McDonald, J. F. and Moffitt, R. A. (1980). The Uses of Tobit Analysis. The Review of Economics and Statistics, 62(2): Mortensen, D. T. and Pissarides, C. A. (1994). Job creation and job destruction in the theory of unemployment. The Review of Economic Studies, 61(3): Poletaev, M. and Robinson, C. (2008). Human Capital Specificity : Evidence from the Dictionary of Occupational Titles and Displaced Worker Surveys, Journal of Labor Economics, 26(3): Summerfield, F. (2016). Matching Skills and Tasks: Cyclical Fluctuations in the Overqualification of New Hires. Yamaguchi, S. (2012). Tasks and Heterogeneous Human Capital. Journal of Labor Economics, 30(1):

24 Appendix A. Occupational codes in the UK LFS Occupational categories in the LFS do not remain constant over time. The UK records up to 4-digit occupational categories in the two-quarter longitudinal LFS, starting with 3-digit codes in 1997 and going up to 4-digit codes from the 2000s onwards 11. Occupational codes have been through two major re-organisations. Between the 1990s and the 2000s, the LFS disaggregated the codes from 3-digit to 4-digits. The transition in the data is not smooth, in the sense that the dictionary of transitions is not accompanied by a probabilistic mapping between the SOC90 and SOC00. In the data, SOC90 is used up until 2000q4 and SOC00 is used starting from 2001q2. Thus, there is a gap in 2001q1, where all occupational information is missing. Between the 2000s and 2010s, the transition is made easier for the researcher, since in addition to the dictionary of transitions, a probabilistic mapping is provided for 2012q2. There are several ways to standardise the occupational series over time, so as to be able to make valid comparisons, which are a crucial element in this chapter. The simplest strategy would have been to take the minimum common denominator. For example, if we observe that a single occupation in 1990 splits into two different occupations in 2000, and these in turn split into 3 more occupations in the 2010s, our strategy would have been to use the 1990 codes as a reference category and merge any follow up occupational splits. However, this particular fix would mean that we would lose most of the variation in the data. Out of the 375 SOC2000, only 69 occupations have a one-to-one match with the new SOC2010 codes. The rest of the codes are many-to-many matches. Applying this to the data would mean that all many-to-many matches would end up becoming one single very large occupation, leaving us with a total of only 70 occupations from which to observe angular separation (69 one-to-one matches + the 1 large occupation from the many-to-many matches). 11 The longitudinal series started in 1992, yet between 1992 and 1997 only 1-digit major occupational codes were recorded. Although 3-digit occupational codes have been recorded for the cross-sectional element of the LFS since 1975, these were not transferred to the longitudinal element before