New revelations about unemployment persistence in Spain: time-series and panel data approaches using regional data

Abstract

This article aims to re-examine the persistence of unemployment in Spain. For this purpose, we use time-series and cross-section analysis. From a time-series viewpoint we disaggregate unemployment by regions, and use unit root tests, AR coefficients and fractional differencing parameters as indicators of persistence. For the cross-section approach, we first estimate mean regressions of regional unemployment rates. Then, using a panel of 114 periods and 50 provinces, we estimate pooled, fixed and random effects models. Finally, following some recent developments, we implement several panel data unit root tests. Previous studies had already shown the strong persistence of Spanish unemployment. Our disaggregated analysis extends the finding to reveal that the persistence is greater in the most industrialized regions. The results also suggest that a structural break took place in 1994, implying a decline in the unemployment persistence since then.

Acknowledgements

L. A. Gil-Alana gratefully acknowledges financial support from the Ministerio de Ciencia y Tecnologia (SEJ2005-07657, Spain).

Notes

¹ Arellano and Bond (1991) derived a generalized method of moments (GMM) estimator using lagged levels of the dependent variable and the predetermined variables and differences of the strictly exogenous variables. This methodology assumes that there is no autocorrelation in the error term.

² The ongoing nature of the EPA has allowed the INE, since the 3rd quarter of 1976, to compute homogeneous data on the main characteristics of the labour market. In order to ensure such homogeneity in the results, it was required to carry out retrospective adaptations (from 1994 untill the 2nd quarter of 1987) as well as to incorporate to the survey some improvements applied since 1999. For a detailed description of changes affecting EPA-2002, see: http://www.ine.es/daco/daco42/daco4211/notamet.htm

³ Spain is divided into 50 provinces, which are gathered in 17 regions, with different cultural traditions and political autonomy: Andalucía, Aragón, Asturias, Baleares, Canarias, Cantabria, Castilla-Mancha, Castilla-León, Cataluña, Valencia, Extremadura, Galicia, Madrid, Murcia, Navarra, País Vasco and Rioja.

⁴ See also Cook (2002) for spurious rejections using Dickey–Fuller tests.

⁵ A seasonal AR(1) process can be defined as (1φL^s )u_t = ϵ_t , where s is the number of time periods within a year.

⁶ Though not reported in this article, we also examined the orders of integration of unemployment disaggregating the series by sectors. It was found that the lowest degree of integration was obtained for the Services sector. On the other extreme, Industry had the highest levels of persistence. Tolvi (2003) found that unemployment in Finland is less persistent for females and young people. We leave for future research the analysis of persistence for groups of age and sex in Spain.

⁷ Some major economic and political events in Spain are summarized by Bentolila and Blanchard (1990, p. 238). The reforms of the Spanish labour market are described in European Commission (2005, p. 49).

⁸ First, in 1994 a number of changes took place. Economic circumstances were included in the reasons justifying individual dismissal procedures, which made it easier to occur. Unemployment benefit would come to be considered as taxable income and a range of issues on working conditions were no longer regulated. In that year the principle of causality was also re-established as a general rule for fixed-term contracts. Then, in 1997 a new permanent contract with lower dismissals costs was approved.

⁹ On the other hand, the Sargan test from the one-step homoskedastic estimator rejected the null hypothesis that the over-identifying restrictions are valid, but this could be due to heteroskedasticity.

¹⁰ The presence of first-order autocorrelation in the differenced residuals does not imply that the estimates are inconsistent, while it would be the case if the second-order autocorrelation were found. See Arellano and Bond (1991, pp. 281–2).

¹¹ Note, however, that even with the novel techniques used here, it is still statistically difficult to distinguish between nonstationarity (with a unit root) and stationarity with AR coefficients close to the unit circle.

¹² The estimations based on different k for each individual (instead of enforcing the same k for all the provinces within the same region) yielded very similar results and are not shown.

¹³ The devise of grouping ‘ad hoc’ sub-panels for particular purposes, has been applied even for the case of just a few individuals (c.f. Culver and Papell, 1997). The shorter is the panel, the more difficult is that the cross-section variation helps to reject the unit root null, but whenever it happens, the result is very strong. Note also that LLC (2002) test is to some extent accurate for small panels, since ‘the normal distribution provides a good approximation to the empirical distribution of the test statistic in relatively small samples’.

¹⁴ That is not the case for the majority of the regions, implying that the individual unemployment series included in those regions should be considered I(0). This panel-based test could not be carried out for the case of regions containing a number of provinces smaller that 3.

¹⁵ As stressed by Strauss and Yigit (2003), subtracting cross-sectional means to remove common time specific effects is a quite common procedure but, if there is heterogeneity in the cross sectional correlation, it will only partially reduce the problem. Regarding the LLC test, it requires independence across individuals, assumption that can be somewhat relaxed to allow for some degree of dependence by including time-specific intercepts, which does not affect the limiting distributions of the test (c.f.: LLC, 2002, p. 13).

¹⁶ Note that IPS (2003) pointed out that the tests perform better if the orders of the underlying ADF regressions are correctly chosen or over-estimated.

¹⁷ This can be related to the number of temporary jobs too. Spain is the EU member with the highest level of temporary employment, which affects the duration distribution of unemployment (c.f. Guell, 2003).