Nonmonotonic relationship between human capital and unemployment: an exploratory study with empirical evidence on Singapore

Abstract

We examine empirically the role of human capital, together with global and domestic technologies, in the evolution of unemployment in Singapore. Our results show a threshold level of human capital, below (above) which increases (decreases) in human capital will reduce (raise) unemployment. This nonmonotonic relationship is caused by a greater substitution from labour toward technology feasible at higher levels of human capital.

Notes

¹See, for example, the Wall Street Journal, 30 September 2002.

²This is due to the unavailability of Singapore R&D data prior to 1978. The earliest year of collection for other data sets is 1974.

³Source of data: Singapore Department of Statistics (www.singstat.gov sg).

⁴Source of data: Computed from National Survey of R&D Expenditure and Manpower, National Survey of R&D in Singapore, Report of the Labour Force Survey of Singapore and Report on the Labour Force in Singapore, various issues.

⁵Such assumptions are often employed in growth modelling. See, for example, Howitt (1999) and Zeng (2003).

⁶Source of data: Computed from time series data obtained from Science and Engineering Indicators (Volume 2 – Appendix Tables), National Science Board, USA.

⁷Computations by Howitt and Mayer-Foulkes (2002) revealed that 80% of the global R&D expenditure can be attributed to five countries, this figure rising to 95% with the inclusion of another siz countries.

⁸We have also run regressions that exclude US_R&D_j , but the quadratic term for HC_t , was found to be statistically insignificant. For purposes associated with allowing for consistent comparisons throughout when we exclude the 15–19 and 15–29 age groups in our later regressions, we only consider two lags of US_R&D_{t − j} , with RSE_P_EMP_t included as a regressor and three lags with this regressor excluded.

⁹The test results are based on the existence of a deterministic trend in the data, with an intercept and no deterministic trend in the cointegrating equation.

¹⁰Assuming that the first-differenced leads and lags of the regressors are uncoirelated with the error term, large sample properties associated with OLS and GLS will still hold, which will then allow us to perform the usual t- and F-tests.

¹¹Studies by Abeysinghe and Ian (1999) found that estimates differ widely across various techniques for estimating cointegrating regressions when the sample size is small.

^12,13Computed from data obtained from the Report on the Labour Force in Singapore, 2003.