Abstract
This paper classifies 49 manufacturing and service industries according to their educational workforce composition. Statistical cluster techniques are applied to data for the USA, Germany, France, the UK and Austria. Industries are first classified separately for each country, providing an appropriate tool for the analysis of national micro-data. Later, the paper proposes a common consensus classification, enabling comparative international studies. Validation of the cluster solution reveals considerable robustness to variations over time and between countries. Finally, regression analyses and ANOVA decompositions on various measures of sector performance confirm a significant tendency towards “education-biased structural change” between industries.
This is a preview of subscription content, log in via an institution to check access.
Similar content being viewed by others
Notes
See Hamermesh (1993) for an elaborate discussion of the demand for heterogenous labor under a variety of production and cost functions.
For example, Machin and Van Reenen (1998, p. 1232) report “a significant association between skill upgrading and R&D intensity”, whereas Kahn and Lim (1998) find a positive relationship between skill intensity and TFP growth. Berman et al. (1998) show that among developed countries, the within-sector upgrading of labor skills has been largely concentrated in the same industries, while Haskel and Slaughter (2002) additionally claim that SBTC itself has a sector bias, demonstrating that rising skill premiums are often caused by technological change that is concentrated in skill-intensive sectors, and vice versa. For further examples, see Krueger (1993), Autor et al. (1998), Falk and Seim (2001), or Chun (2003).
An often cited example of “de-skilling” is the introduction of assembly lines, e.g., in the production of automobiles, which substituted low-skilled factory workers for high-skilled craftsmen. In contrast, the introduction of new computer technologies is the standard example of a positively “skill-biased technological change” (SBTC), where the technology-induced growth in demand for better educated personnel accounts for a simultaneous rise in both the employment and wages of skilled labor.
Other contributing institutions are The Conference Board, the Groningen Growth and Development Center (GGDR), the Centre d‘Etudes Prospectives et d‘Information Internationaux (CEPII), the Zentrum für Europäische Wirtschaftsforschung (ZEW), and the Austrian Institute of Economic Research (WIFO).
Thus, “the results are less likely to be an artifact of a single method of analysis and more likely to provide a reliable summary of any class structure that is present in the data” (Gordon 1999, p. 184).
For example, the communications sector is comprised of two industries that are likely to differ substantially in their educational requirements. We therefore attempted a more detailed breakdown into 3-digit industry codes. But only the US and the UK provided separate data for “post and courier activities” (ISIC 641) and “telecommunications” (ISIC 642). Since they rely on original data, both are in ordinary typeface. In contrast, the italic numbers in parentheses for France, Germany, and Austria signal that I had to insert the value derived from a corresponding, more highly aggregated sector (ISIC 64).
The median is preferred to the mean, because the latter would require interval-scaled data. In all but four cases the unambiguous median would be equivalent to the integer of the mean anyway. Furthermore, when individual industries exhibit identical cluster identification, this number is inserted for the higher aggregated sector (e.g., ISIC 17–19 in Germany and Austria). But when the sub-sectors exhibit different cluster identifications, the record of the higher aggregate abides as “missing” (for instance, ISIC 71–74).
For balanced panels ANOVA generally produces best quadratic unbiased estimators of the variance components (Baltagi 2001, p. 162). It is, however, also based on the assumptions of normal distribution and homoscedasticity (equal variances). Even though it is generally considered to be relatively robust with respect to violations of the first assumption, heteroscedasticity can pose a serious problem, especially when the sample sizes differ, as is the case for the new sector types. As a consequence, for all reported regresssions, I also ran a series of non-parametric tests. In addition to the Kruskal Wallis and Median tests, which can only tell whether the overall taxonomy discriminates significantly, I apply the Kolmogorov–Smirnov test for equality of distribution functions and the Wilcoxon rank-sum (“Mann–Whitney”) test for each pair of industry types. The outcome generally confirmed the sign and significance of the coefficients reported in Tables 4, 5, and 8.
References
Acemoglu D (1998) Why do new technologies complement skills? Directed technical change and wage inequality. Quart J Econ 113:1055–1089
Acemoglu D (2002) Technical change, inequality, and the labor market. J Econ Liter 40:7–72
Acemoglu D, Zilibotti F (2001) Productivity differences. Quart J Econ 116:563–606
Akerlof GA, Kranton RE (2000) Economics and identity. Quart J Econ 115:715–753
Akerlof GA, Kranton RE (2002) Identity and schooling: some lessons for the economics of education. J Econ Liter 40:1167–1201
Arrow KJ (1973) Higher education as a filter. J Public Econ 2:193–216
Autor DH, Katz LF, Krueger AB (1998) computing inequality: have computers changed the labor market? Quart J Econ 113:1169–1213
Baltagi BH (2001) Econometric analysis of panel data, 2nd edn. Wiley & Sons, Chichester, New York
Baron JN, Kreps DM (1999) Strategic human ressources. frameworks for general managers Wiley & Sons, New York
Becker GS (1964/1975) Human capital. a theoretical and empirical analysis, with special reference to education. NBER and Columbia University Press, New York
Berman E, Bound J, Machin S (1998) Implications of skill-biased technological change: international evidence. Quart J Econ 113:1245–1279
Bresnahan TF, Brynjolfsson E, Hitt LM (2002) Information technology, workplace organization, and the demand for skilled labor: firm-level evidence. Quart J Econ 117:339–376
Brewer DJ, Eide ER, Ehrenberg RG (1999) Does it pay to attend an elite private college? Cross-cohort evidence on the effects of college type on earnings. J Hum Resour 34:104–123
Card D, Krueger AB (1992) Does school quality matter? Returns to education and the characteristics of public schools in the United States. J Political Econ 100:1–40
Caroli E, Van Reenen J (2001) Skill-biased organizational change? Evidence from a panel of British and French establishments. Quart J Econ 116:1449–1492
Caselli F (1999) Technological revolutions. Am Econ Rev 89:78–102
Chun H (2003) Information technology and the demand for educated workers: disentangling the impacts of adoption versus use. Rev Econ Stat 85:1–8
Falk M, Seim K (2001) The impact of information technology on high-skilled labor in services: evidence from firm-level panel data. Econ Innov New Technol 10:289–324
Forbes KJ (2001) Skill classification does matter: estimating the relationship between trade flows and wage inequality. J Int Trade Econ Develop 10:175–209
Gordon AD (1999) Classification, 2nd edn. Chapman & Hall, Boca Raton
Hamermesh DS (1993) Labor demand. Princeton University Press, Princeton
Haskel JE, Slaughter MJ (2002) Does the sector bias of skill-biased technical change explain changing skill premia? Eur Econ Rev 46:1757–1783
Kahn JA, Lim J-S (1998) Skilled labor-augmenting technical progress in U.S. manufacturing. Quart J Econ 113:1281–1308
Krueger A (1993) How computers have changed the wage structure: evidence from micro data, 1984–1989. Quart J Econ 108:33–60
Lazear EP (1998) Personnel economics for managers. Wiley & Sons, New York
Machin S, Van Reenen J (1998) Technology and changes in skill structure: evidence from seven OECD countries. Quart J Econ 113:1215–1244
Peneder M (2005) Creating industry classifications by statistical cluster analysis. Estudios de Economia Aplicada 23(2):451–463
Peneder M (2003) Industry classifications. aim, scope and techniques. J Indust Competit Trade 3(1–2):109–129
Peneder M, Kaniovski S, Dachs B (2003) What follows tertiarisation? Structural change and the role of knowledge-based services. Service Indust J 23:47–66
Romesburg HC (1984) Cluster analysis for researchers. Waldsworth Inc., Belmont
Schultz ThW (1960) Capital formation by education. J Political Econ 68:571–583
Schultz ThW (1961) Investment in human capital: reply. Am Econ Rev 51:1035–1039
Sharma S (1996) Applied multivariate techniques. Wiley & Sons, New York
Sianesi B, van Reenen J (2002) The returns to education: a review of the empirical macro-economic literature. IFS Working Paper 02/05, London
Spence M (1973) Job market signalling. Quart J Econ 87:355–374
Spence M (2002) Signalling in retrospect and the informational structure of markets. Am Econ Rev 92:434–459
Wolff EN (2003) Skills and changing comparative advantage. Rev Econ Stat 85:77–93
Acknowledgements
This paper has benefitted from numerous conversations, helpful comments and invaluable assitance with the national data. Martin Falk, Edward Lazear, Mary O’Mahony, Laurence Nayman, Bart van Ark, Michela Vecchi, Thomas Zwick, and an anonymous referee deserve especial thanks. Naturally, none of them bears any responsibility for eventual errors or omissions. The research was partly funded by the 5th Framework Programme of the European Commission with additional support provided by the Austrian Ministry of Education, Science and Cultural Affairs.
Author information
Authors and Affiliations
Corresponding author
Appendix A: cluster identification for the US national taxonomy
Appendix A: cluster identification for the US national taxonomy
The data set for the USA is comprised of 39 sectors covering two-year averages from 1979 until 2000, amounting to a total of 429 observations. Following the two-step approach already outlined, we begin with the k-means method, where the set of observations is divided by a pre-defined number of clusters k. Cluster centers are computed for each group, which are the vectors of the means of the corresponding values for each variable. The objects are then assigned to the group with the nearest cluster center. After this, the mean of the observations are recomputed and the process is repeated until convergence is reached. This is the case, when no observation moves between groups and all have remained in the same cluster of the previous iteration.
With this method, a critical choice is the initial number of clusters k, for which I consistently applied the following self-binding rule-of-thumb: “Choose the lowest number k that maximizes the quantity of individual clusters l which include more than 5% of the observed cases”. Running the k-means algorithm on a dissimilarity matrix made up of Euclidean distances between any pair of observations for all values of k ranging from 2 to 35, the lowest number which fulfils the above rule turns out to be k = 11 with l = 10. This partition has thus produced 11 surprisingly compact clusters, of which only one comprises less than 5% of total observations.
In the second step, the cluster centers from the first partition described above are entered as individual observations in the hierarchical analysis. I use all the four (dis)similarity measures mentioned before (Euclidean- and city-block distance, angular separation, and correlation coefficient), each of which is applied to the three agglomeration algorithms (i.e., average-, complete-, and single linkage method). This process thus established twelve different cluster trees, which were then compared.
Figure A.1 presents the hierarchical cluster tree for the USA, using the average linkage method in combination with each of the aforementioned (dis)similarity measures. The branches on the bottom of the chart represent one entity each, while the root on top represents the entire set of objects. As we move upwards on the chart, the degree of association between objects is higher, the sooner they are connected by a common root. Conversely, objects or groups are the more dissimilar, the longer they remain disconnected. As is frequently the case with statistical cluster techniques, the most stable patterns appear at the extreme ends of the distribution. Clusters 10 and 11 show the highest degree of educational intensity, with shares of both university and associate degrees far above the average. They are consistently grouped together or located as immediate neighbors in all twelve dendrograms. But the graphical representations also reveal that they are still very distinct among themselves. Cluster 11, which is comprised of sectors such as education, research and development, or computer services, by far exceeds the other sectors in terms of university degrees held by members of its workforce. We consequently label this cluster “industries with very high educational intensity”. In comparison, cluster 10 comprises industries with a lower but still high educational intensity, including for example, financial services, insurance, or the group of other business services.
USA—hierarchical clustering, average linkage method
Clusters 6 and 7 exhibit an extremely robust association and are grouped together in all twelve dendrograms. The lack of any pronounced deviations from the average educational composition is the most characteristic observation with respect to those industries with an intermediate educational intensity. Examples are mechanical engineering, motor vehicles or retailing. Clusters 8 and 9 exhibit an above average educational profile, but with more emphasis on associate degrees and fewer university graduates. They are most closely associated in six of the 12 charts, although cluster 9 also appears repeatedly with cluster 10 and cluster 8, together with the intermediate group of clusters 6 and 7. We therefore attach the label “intermediate-to-high educational intensity” to this category. Among others, it comprises chemicals, instrument engineering, electrical machinery, and air transport.
The remaining clusters 1–5 all exhibit below average educational intensity. Cluster 2 is the most consistent outlier; its workforce has a very low educational profile. Clusters 4 and 5 have the least clear patterns, including among others, food, drink and tobacco, rubber and plastics, basic and fabricated metals, and the group of other inland and water transport. These industries are only consistently close to each other when the correlation coefficient or angular separation are used as similarity measures. The same applies to the city block measure in combination with the complete linkage method. In all other graphs, they are disjoint. Nevertheless, the two are close in 6 of 12 dendrograms. Since the other graphs do not suggest a convincing alternative pattern, we classify them together as industries with intermediate-to-low educational intensity.
Clusters 1 and 3 appear together in all 12 cluster trees. They encompass industries such as construction and vehicle maintenance and repair. Their educational profile is distinctly low, characterized by above average shares of labor without any formal qualifications, some college (but no degree) or only high school graduation. Finally, cluster 2, which consists of agriculture, forestry and fishing, textiles and clothing, as well as hotels and catering, appears most often in an outlying position. It displays a very low educational profile, wherein the share of labor without any formal degrees by far exceeds the other groups.
Rights and permissions
About this article
Cite this article
Peneder, M. A sectoral taxonomy of educational intensity. Empirica 34, 189–212 (2007). https://doi.org/10.1007/s10663-007-9035-2
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10663-007-9035-2