European regional productivity: does country affiliation matter?

ABSTRACT

We assess the importance of country affiliation for European regional productivity growth between 1980 and 2012 and reveal that 93% of regional variation in gross domestic product (GDP) per worker is explained at the national level, regional divergence was experienced throughout the period after 1983 and country divergence occurred throughout the period until a reversal around the beginning of the recent recession. Distinct groups of regions have been growing together and although regional groups are generally confined to national borders, there are notable exceptions. The results rehabilitate the importance of national borders for understanding the productivity performance of regions, implying that regional policy-makers should take cognizance of national effects.

KEYWORDS:

• Regional productivity
• multilevel modelling
• group-trajectories

JEL CLASSIFICATION:

• R11
• O47

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary data

Supplementary data for this article can be accessed at https://doi.org/10.1080/02692171.2018.1515899

Notes

1. We conducted a stability check to identify whether qualitatively similar results stand for 1990 onwards. This indeed was the case. A full list of regions included in the analysis is available in the online supplementary material.

2. There are also imbalances (‘missing data’) that will have to be taken into account. Multilevel modelling can deal with such imbalances and hierarchical structure.

3. Orthogonal polynomials avoid multicollinearity problems, which would occur in such models if only the square and cube of time were used (Hedeker and Gibbons 2005). The use of orthogonal polynomials also means that each predictor is on the same scale which allows a direct comparison of the magnitudes of the estimated coefficients.

4. We estimate the growth curve model using the MLwiN software (Rasbash et al. 2004; http://www.cmm.bristol.ac.uk).

5. These details are available on request from the authors and not included here for brevity.

6. There is the need to fit a polynomial that is complex enough to capture major trends but not too complex that it compromises the ability to fit models reliably with convergence.

7. In each group, g, productivity trajectories are assumed to be a cubic polynomial function of year, . Note that when compared to the multilevel time series cross-sectional regression, here we assume that, within a group, only the intercept of the trajectory varies across regions and that the within-region variance is constant over time.

8. It should be stressed that the classification is non-hierarchical, so that when an additional group is created, this does not necessarily mean that only one of the previous groups is split. When three groups replace two, for example, it may well be not only that both of the two groups in the latter solution lose some members to the third group but that there is also some redistribution of members between the first two. Model selection works inductively and in two stages. First, the models are fitted with a third-order polynomial of time for a single group, and the Bayesian information criterion (BIC) is calculated. Then, two groups are specified and the BIC obtained. Next, three groups are fitted and then four, five and so on until there is no further reduction in the BIC. This is then the most parsimonious model in terms of groups with distinctive trends.

9. The profiles represent the growth trajectory of the regions within the club, and the profile’s intercept is based on the average starting value of productivity across this club’s sample. Although all regions within the club grow at a very similar rate as denoted by the profile slope, around this line is a spread of regions that vary in distance from this mean-average trend line.

10. It is possible that these are outcomes of the modifiable areal unit problem; see Openshaw (1983).