Transition Probability Matrix

The bottom row of the matrix illustrates strong persistence for regions with an unemployment rate below 0.6 times the European average: by 1996 81% remained below 0.6 times the European average, while the remaining 19% had increased their relative rates to between 0.6 and 0.75 times the European average. The next row up shows that of the 23 regions starting with a relative unemployment rate between 0.6 and 0.75, 26% remained in that range while 52% fell to below 0.6. The top row tells a similar story, with strong persistence among the regions with the highest unemployment rates. However, it is the regions that in 1986 had rates between 0.75 and 1.3 that experienced the greatest mobility in the distribution, with a tendency to deviate towards the extremes of high or low unemployment.

The transition probability matrix illustrates the evolution of the distribution over time, but it requires the choice of a specific discretization which could possibly distort the true picture of the transition. The authors resolve this problem by plotting the transition kernel from the 1986 distribution to the 1996 distribution. The contour plot derived from this graph confirms the polarization process suggested by the transition probability matrix.

Of course, the polarization of regional unemployment can stem from both changes in regional employment opportunities and changes in the structure of the labour force. Overman and Puga show that the process has in fact been driven by changes in the pattern of employment opportunities. Indeed, as would be expected, the labour force tended to move so as to offset any polarization, with those regions with below average unemployment seeing the greatest increases in their labour forces while regions with above average unemployment witnessed smaller rises or actual falls. Previous research has shown that between 1960-and 1980, regional labour force adjustments were able to offset changes in regional employment, leaving differences in unemployment rates reasonably stable. Yet the paper shows that, since 1986, labour force adjustments have been unable to keep pace with employment changes and have been insufficient to prevent polarization.

What factors caused the polarization? Regions differ in the sectoral composition of their employment; in the age, sex and skill structure of their populations; and in their geographical location within the EU. Regions that initially specialized in agriculture or manufacturing may have seen their unemployment rates rise as the EU production structure moved away from those sectors. Similarly, regions with a high proportion of low-skilled workers may have seen their unemployment rates rise as production shifted from low-skilled to high-skilled employment. Over the last decade, the Member States of the EU have pushed ahead with ever-closer economic integration. Recent theoretical developments suggest that such a process can be associated with the emergence of spatial concentrations of employment and that with falling barriers to trade these may extend across national borders. If regional labour forces do not fully adjust to such employment changes, then geographical location may be important in explaining the increased polarization of unemployment rates.

Unemployment Clusters

Overman and Puga use two complementary techniques, one parametric, one non-parametric, to examine these alternative explanations. The underlying idea behind the non-parametric approach is to track how closely the evolution of each region's unemployment rate has followed that of some group of regions which would be expected to behave similarly. The regions are grouped by the following criteria: regions in the same country; regions that are geographical neighbours; regions with similar sectoral compositions; and regions with similar proportions of low-skilled.

The results for the same country group suggest that regions do not tend to move strongly with their country over time. Yet even though national borders seem to have become less important in determining regional outcomes, geographical location may still matter at levels below the nation state. The authors test this idea by analysing the evolution of each region's unemployment rate in relation to that of its immediate geographical neighbours (i.e. bordering regions). Here, each region's outcome is much closer to the outcomes of neighbouring regions than to the outcomes of other regions in the same country. In fact even foreign neighbours (i.e. bordering regions that are in a different country) are more closely related than non-bordering regions in the same country. The similarity of outcomes across neighbours could be driven by neighbouring regions having similar characteristics that are important determinants of unemployment rates. The final two criteria test this idea by examining such determinants.

The period 1986-96 saw the continuation of an ongoing shift in European employment from agriculture and industry into services. If labour force adjustment is slow, then regions with high initial specialization in declining sectors may have seen their unemployment rates rise and not recover. Could this be driving the polarization of unemployment in Europe? And can the importance of neighbours be justified by the fact that regions with heavy industrial or agricultural employment often occupy the same areas? The authors' results suggest the answer to both questions is no: regions with very similar industrial specifications have experienced very different outcomes. This is probably due to the fact that the largest drop in agricultural and manufacturing employment had already taken place before the period in consideration - between 1971 and 1986 the share of manufacturing in European employment fell from 41% to 33%; between 1986 and 1996 it fell from 33% to 30%. A final explanation for the neighbour effect could be that neighbouring groups of regions possess similar skill compositions. However, the authors’ results suggest skill composition is only a minor predictor of unemployment.

The ordinary least squares results from the parametric analysis broadly echo the non-parametric conclusions. Using the same types of explanatory variables as the non-parametric approach, the authors show that unemployment rates are only weakly explained by industrial structure and skill level variables. However, neighbouring regions' unemployment rates have a very strong and significant effect. Interestingly, the regressions show that foreign neighbours are as important as domestic neighbours: the authors are unable to reject the hypothesis that the coefficient on both domestic and foreign neighbours are identical. This result can be seen in the map on the previous page, where the high and low unemployment clusters that have emerged over the last decade show little respect for national borders. The use of instrumental variables in the regression does not alter any of these results.

What is driving this emerging pattern of cross-border unemployment clusters? Overman and Puga surmise that the patterns may be the result of firm location and relocation decisions, reflected in an agglomeration of activity over geographical areas somewhat larger than NUTS2, but somewhat smaller than nation states. In contrast to the divergence of unemployment rates across European regions, differences in regional incomes per capita are narrowing. Yet while inequalities in income per capita exhibit a core-periphery gradient, unemployment clusters are more localized and are emerging in both Europe's core and its periphery. Recent location theories suggest that the self-reinforcing nature of agglomerations will make these clusters hard to break once they become established. Given that the unemployment clusters observed are not very large and are scattered across Europe, the authors conclude that it may be more efficient to implement policies that accept some clustering and larger mobility within a neighbourhood.