Are UK labour markets polarising?

1. Are UK labour markets polarising? Craig Holmes National Institute of Economic and Social Research, London, April 24th 2012

2. Introduction • Part of SKOPE’s ESRC funded research programme (2008-13) • Main research question: what does the development of the “hourglass” labour market mean for: • Earnings and job quality • Mobility and mobility barriers • Skills policy • This presentation draws on this research • Main issues: • Why is occupational polarisation not clearly observed in wage distributions? • In what ways is the change in occupational structure actually important?

3. Structure of talk • Polarisation in occupations and wage distributions • Re-evaluation of theory • Decomposition of changes to wage distributions • Wage mobility

4. Polarisation of occupations • Routinisation hypothesis (Autor, Levy and Murnane, 2003): • Price of computer capital has fallen since late 1970s • Computer capital replaces labour engaged in routine tasks • Non-routine tasks may be complementary to computer capital (e.g. management, skilled professionals) • Result: growth in non-routine occupations due to changes in demand (complementarities) and supply (displaced routine workers) • Polarisation hypothesis (Goos and Manning, 2007) • Routine occupations found in middle of income distribution • Non-routine occupations found at top and bottom of distribution • Managers, skilled professionals at the top • Non-routine ‘service’ occupations at the bottom e.g. hairdressers, cleaners

5. Polarisation and occupations • Following Goos and Manning (2007), hourglass effect shown through changes in employment share of groups of occupations ranked by (initial) average wages – each of approx. 10% of labour supply. • Data : • New Earnings Survey 1986 (ranking wage) • Labour Force Survey 1981-2008 (employment shares) • Hours rather than headcount

6. Polarisation of occupations • Growth in employment share, by ranked occupational group, 1981-2008

7. Polarisation of occupations • Similar evidence found: • US (Autor, Katz and Kearney, 2006; Autor, 2011) • Germany (Spitz-Oener, 2006; Oesch and Rodríguez Menés, 2011) • Spain and Switzerland (Oesch and Rodríguez Menés, 2011) and across Europe (Goos, Manning and Salomons, 2009). • What does this mean for wage distributions?

8. Polarisation and wage distributions • Wage distributions: • Rising upper and lower inequality (Machin and Van Reenen, 2007) • Increasing proportion below low-paid threshold (Lloyd, Mason and Mayhew, 2008) • Does this mean the middle of these distributions is disappearing? • Density functions: • New Earnings Survey 1986-2002 • Labour Force Survey 1995-2008

9. Polarisation and wage distributions • New Earnings Survey:

10. Polarisation and wage distributions • Labour Force Survey – 1995-2008

11. Polarisation and wage distributions • Has employment in the middle declined? • Using same datasets, look at changes in employment across the distribution: • Log gross hourly wage distribution standardised (0.5th percentile up to 99.5th percentile) • Wage range divided into ten groups • Look for changes in employment at different wage levels on this scale • Polarisation would be reflected by growth in low-paying and high paying jobs, and decline in middle-wage jobs

12. Polarisation and wage distributions • New Earnings Survey – (1) 1986-1997 and (2) 1997-2002

13. Polarisation and wage distributions • Labour Force Survey – 1995-2008

14. Polarisation and wage distributions • Family Expenditure Survey – 1987-2001

15. Discussion of results • Why is there a difference between two approaches? • Goos and Manning (2007) demonstrated a compositional effect • Leads to polarised wage distribution if wage structure of occupations remains constant • There may be wage effects: • Between-groups effects (Autor, Katz and Kearney, 2006) • Within-groups effects • Observable differences (e.g. educational composition) • Unobservable differences (e.g. ability)

16. + increased between - + increased within - Initial wage structure Change: Composition group inequality group inequality wage Occupational group – area represents employment share = interquartile range A model of polarisation

17. Discussion of results • Wage distributions seem to exhibit an increase in spread in the upper half • A small proportion have experienced very high wage growth. • By comparison, many good jobs earnings become relatively closer to middle. • A growing lower end in the “lovely” jobs? • Polarisation of knowledge workers (Brown, Lauder and Ashton, 2011) • Less increase in earnings variation at bottom end – minimum wage effect?

18. Decomposing wage distributions • Given above patterns, would like to understand why wage distributions have changed as they have. • Biggest issue with analysing changing distributions is separating out all effects: • Wage determination process: • yt = gt(x) • Composition effects come through changes to x • Wage effects come through changes to g • These may be different at different points in the wage distribution

19. Decomposing wage distributions

20. Decomposing wage distributions • Number of approaches to measuring changing distributions, usually involving some form of quantile regression: • Usually conditional on explanatory variables, like OLS regressions • We need to look at unconditional distributions • Conditional regressions do not aggregate to unconditional quantile regressions, unlike OLS • Firpo, Fortin and Lemieux (2007) – henceforth FFL: • Counterfactual distribution estimated by reweighting • Composition effects: initialcounterfactual • Wage effects: counterfactualfinal • Estimates individual contribution of covariates to both • Similar to Blinder-Oaxaxa decomposition of the mean

21. Data • Family Expenditure Survey (1957-2001) • Two surveys for sample: 1987 and 2001 • Covers period of routinisation • Has wages and education attainment (unlike LFS and NES) • Variables: • Hourly wage = gross weekly wage / (basic hours + usual overtime) • Age finished full-time education – convert this into dummies for degree (21+), post-compulsory education (18-20) and high school education (16-17) • Experience = age – age left FT education • Dummies for gender, union membership, type of work • No variables on racial background or industry.

22. Data • The 1987 survey also has a narrower occupational coding. • 351 groups • KOS (pre SOC90) classification • The 2001 survey uses SOC2000 classification • 353 groups at 4-digit level • 81 groups at 3-digit level • Manual conversion using 1987 descriptions into SOC2000 4-digit equivalent • Changed into 3 digit category – prevents losing 1987 occupations which fit into two closely matched SOC2000 categories • Used in this presentation

23. Data • Creates larger occupational groups: • Professional • Managerial • Intermediate • Admin • Manual routine • Manual non-routine • Service High skill non-routine Routine occupations Low skill non-routine

24. Results: reweighting

25. Results: reweighting • Change in log real gross hourly wage, 1987-2001

26. Results: composition effects • Estimated individual composition effects, 1987-2001

27. Results: composition effects • Large impact of declining unionisation at bottom and middle. • Increased female participation has a negative effect (through initial gender pay gap) • Increase in part-time work has small negative composition effect • Expansion of higher education has impact even on low wage jobs • Largest effect at top of distribution • Occupational compositional effect negative at bottom and positive at top • Not as large as education or union at respective ends?

28. Results: wage effects • Marked fluctuations in occupational premia (relative to administrative occupations):

29. Results: wage effects • Stable graduate premium over majority of distribution:

30. Results: wage effects • Other effects • Small effects through changing union premia and wage penalties to part-time work • Large positive wage effects through narrowing gender pay gap – (between 5% and 7% increase in wages across percentiles except at top decile of distribution) • General ‘shift’ very high at bottom end – possibly the result of minumum wage introduced in 2001.

31. Within-group effects • Earnings within growing occupations should become more varied. • May reflect differences in educational attainment. • If educational attainment has increased too, may reflect varying wage premia across the distribution • May reflect unobservable differences: • General productive ability • Specific skills in certain occupations

32. Within-group effects • NCDS earnings data on managerial workers, based on occupation five years before:

33. Within-group effects • NCDS earnings data on intermediate workers, based on occupation five years before:

34. Within-group effects • Is this the result of observable differences between the two? • Ordered logit model: • Dependent variable, Y – earnings group • Y = 1,…,20 • Include qualifications and demographics • Observed wages in 1991, 1999 and 2004. Observed occupations five years before each date.

35. Conclusion • Little previous work on evidence of polarisation in UK earnings distributions. • We define middle relative to overall spectrum of wages • In some cases, evidence that the middle has expanded – although with different occupational titles • Occupational polarisation  wage polarisation if there are no wage effects • Other determinants of earnings have changed as well as occupational stucture • May be unobservable within-group effects

36. Contact Details Craig Holmes ESRC Centre on Skills, Knowledge and Organisational Performance (SKOPE), Department of Education, Norham Gardens, Oxford Email: craig.holmes@education.ox.ac.uk

37. Appendix • Methodology slides for FFL • Managerial wage distributions

38. Decomposing wage distributions • Data: • N observations, N0 from initial distribution, N1 from final distribution • Ti = 1 if from final distribution, i = 1,...,N. Pr(Ti) = p • Yi and Xi observed • Yi = Yi0 (1 – Ti) + Yi1 Ti where Yit = gt(Xi, ei), t = 0,1 • Data can be reweighted to find the (unobserved) counterfactual distribution. • Counterfactual is wage distribution that would have arisen given initial wage determination process but final explanatory variables

39. Decomposing wage distributions • Reweighting: • where p(X) = Pr (T=1 | x = X) • Calculate p(X) using logistical regression • This counterfactual can be used to decompose wage and composition effects of a distributional statistic: • Let this statistic be represented by functional v(F) – e.g. percentile • Δv(F) = ΔvW + ΔvC

40. Data • Descriptive statistics (mean values):

41. Results: reweighting

42. Results: reweighting

43. A quantile regression approach • FFL’s second contribution is to find a linear approximation of each distributional functional, conditional on the explanatory variables • An influence function, IF, of v(F) is a measure of sensitivity to outliers, where E(IF) = 0 • A recentered influence function, RIF = v(F) + IF, so E(RIF) = v(F) • RIF’s can be conditional on X • Assume a linear projection of RIF onto X: • where j = {0, C, 1}

44. A quantile regression approach • FFL show that: • ΔvC = E(X|T=1) γC - E(X|T=0) γ0 • ΔvW = E(X|T=1) (γ1 – γC) • Moreover, if expectation of RIF is linear, γC= γ0. • Composition effects are sum of change in composition of each explanatory variable, multiplied by wage return in initial distribution • Wage effect is sum of change in wage returns between counterfactual and final distribution, multiplied by final composition of each explanatory variable. • This is a more general case of the Blinder-Oaxaca decomposition, where v(F) is the mean.

45. A quantile regression approach • Our approach looks at quantiles across distribution • j = {0, C, 1} • τ = 0.05, 0.1, 0.15,...,0.95 • Estimate fi(qτ) using kernel density methods

46. A quantile regression approach • FFL’s second contribution is to break wage and composition effects into individual components e.g. occupation, education etc. • Method found in final paper, omitted here for time. • Idea is to find a linear approximation of each statistic in each distribution using explanatory variables: • Composition effects are sum of change in composition of each explanatory variable, multiplied by wage return in initial distribution • Wage is sum of change in wage returns between counterfactual and final distribution, multiplied by final composition of each explanatory variable.

47. Results: individual contributions • Decomposition by wage and composition

48. Wages at the top • Gross weekly earnings data from UK Labour Force Survey • 1990s: • Increased employment in higher wage jobs across all good, non-routine occupations • Long tail: some of growth occurred a long way from the median • 2000s: • Some increase in low wage employment – despite increasing graduatisation • Some increase in very high wage employment  A hollowing out of the middle of the distribution • Differences by sector of employment (manufacturing, retail, financial intermediation and real estate/business activity)

49. Wages at the top • Managerial occupations:

50. Wages at the top • Professional occupations: