Determinants of Regional Economic Growth by Quantile

Abstract

Crespo-Cuaresma J., Foster N. and Stehrer R. Determinants of regional economic growth by quantile, Regional Studies. The robustness of growth determinants across European Union regions is analysed using quantile regression. Using Bayesian model averaging (BMA) on the class of quantile regression models, it is proposed that the set of relevant covariates is assessed, allowing for different effects across growth quantiles. The results indicate that the robust growth determinants differ across quantiles. The set of robust variables includes physical investment when taking country fixed-effects into account, and skill endowment and initial gross domestic product per capita when not. Even when a variable is found to be robust across quantiles, its estimated impact on growth is often found to vary across quantiles.

Crespo-Cuaresma J., Foster N. et Stehrer R. Les déterminants de la croissance économique par quantile, Regional Studies. A partir d'une régression par quantile, on analyse la solidité des déterminants de la croissance à travers les régions de l'Union européenne. Employant un Bayesian Averaging Model (BMA) sur la catégorie de modèles de régression par quantile, on propose une évaluation de la covariance, compte tenu des effets différents suivant les quantiles de croissance. L'ensemble de variables solides comprend l'investissement matériel quand on tient compte des effets spécifiques à un pays, sinon la dotation en connaissance et le produit intérieur brut initial par tête. Même quand une variable s'avère solide à travers les quantiles, l'impact prévu sur la croissance varie souvent à travers les quantiles.

Croissance régionale Bayesian model averaging Régression par quantiles

Crespo-Cuaresma J., Foster N. und Stehrer R. Determinanten regionalen Wachstums nach Quantilen, Regional Studies. In diesem Beitrag wird die Robustheit von Wachstumsdeterminanten in EU-Regionen mittels Quantilsregressionen analysiert. Dabei wird ein Bayesian Model Averaging (BMA) für Quantilsregressionen verwendet, um die relevanten Kovariaten, die unterschiedliche Effekte in den jeweiligen Wachstumsquantilen aufweisen können, zu ermitteln. Die Resultate zeigen, dass die robusten Wachstumsdeterminanten in den jeweiligen Quantilen tatsächlich unterschiedlich sind. Unter Berücksichtigung von länderspezifischen Effekten ist insbesondere die Variable Anlageinvestitionen ein robuster Erklärungsfaktor regionalen Wachstums; ohne Berücksichtigen dieser Effekte sind Humankapitalausstattung und das Pro-Kopf-Einkommen robuste Determinanten. Auch wenn eine bestimmte Variable robust in mehreren oder allen Quantilen ist, sind die ermittelten Effekte auf das Wachstum der Regionen in den jeweiligen Quantilen oftmals unterschiedlich.

Regionales Wachstum Bayesian Model Averaging Quantilsregressionen

Crespo-Cuaresma J., Foster N. y Stehrer R. Determinantes del crecimiento económico regional por cuantiles, Regional Studies. Analizamos la solidez de los determinantes de crecimiento en las regiones de la Unión Europea usando una regresión cuantílica. Mediante el uso de promedios de modelo bayesiano sobre la clase de los modelos de regresión cuantílica, proponemos que se evalúe el conjunto de las covariantes correspondientes teniendo en cuenta los diferentes efectos en los cuantiles de crecimiento. Los resultados indican que los determinantes de un crecimiento sólido son diferentes entre los cuantiles. Si se tienen en cuenta los efectos fijos de cada país, la inversión física es una variable fuerte, de no ser así son variables fuertes la dotación de habilidades y el producto interno bruto per cápita inicial. Incluso cuando se halla una variable que es sólida en varios cuantiles, se observa con frecuencia que el impacto estimado en el crecimiento varía entre los cuantiles.

Crecimiento regional Promedios de modelo bayesiano Regresión cuantílica

Keywords:

• Regional growth
• Bayesian model averaging (BMA)
• Quantile regression

JEL classifications:

• C11
• C21
• R11

Acknowledgement

This paper was based on a background study written for the European Commission's DG Regional Policy within the project ‘Analysis of the Main Factors of Regional Growth: An In-depth Study of the Best and Worst Performing European Regions’ (Contract Number 2007.CE.16.0.AT.209). The authors would like to thank the participants at the WIIW Workshop on Regional Growth, as well as three anonymous referees for helpful comments on earlier drafts of this paper. The usual disclaimer applies.

Notes

For a review of the empirical growth literature, see Temple (1999) and Durlauf and Quah (1999).

Kalaitzidakis et al. (2000) employed the same approach as Levine and Renelt (1992), but allowed for potential non-linearities. They found more variables to be robustly related to growth, emphasizing the importance of non-linearities in the growth process.

Examples using cross-country data include Mello and Perelli (2003), Osborne (2006), Canarella and Pollard (2004), and Foster (2008). All these papers found evidence of heterogeneous effects of some growth determinants across quantiles.

BMA using QR may be also embedded in classes of models which assess spatial correlation across variables or errors explicitly, but this falls outside the scope of this study.

The figures are based on simple bivariate regressions of per-capita GDP growth on each of the growth determinants.

Useful surveys of QR methods include Buchinsky (1998) and Koenker and Hallock (2001). A book-length treatment of the subject is provided by Koenker (2005).

QR coefficients measure the marginal effect of changes in the independent variables on the dependent variable for representative under- and over-achieving countries in terms of growth and not slow- and fast-growing countries per se. This can be contrasted with OLS, which considers the average behaviour of representative countries.

For overviews of BMA, see Raftery et al. (1997), Hoeting et al. (1999), Clyde and George (2004), and Doppelhofer (2007).

This section follows closely the description of Raftery (1995) and Raftery et al. (1997), who provide a fuller description of BMA techniques.

Originally we started with a slightly larger set of variables. Some of these were dropped, however, because of issues of multicollinearity.

Admittedly, endogeneity may still be present in some models despite the (Granger-) causal structure that has been imposed in the specifications by measuring the regressors at the beginning of the period. A more systematic account of the issue of endogeneity in the setting of quantile-BMA falls outside the scope of this research and is proposed as a potentially fruitful avenue for further research. In particular, recent results by Moral-Benito (2009) and Chernozhukov and Hansen (2003) may prove helpful in this respect.

When country effects are controlled for, this is done using the within transformation (that is, subtracting from each observation the country mean of the relevant variable).

The convergence of the MC3 algorithm was checked by computing the correlation between posterior model probabilities based on the Markov chain frequencies and the exact marginal likelihoods (as proposed by Fernández et al., 2001). In all reported results, this correlation was above 0.95.

The prior inclusion probability is taken as the threshold to define robust variables. The intuition for this choice is that it helps one identify variables for which the probability of inclusion in the true model increases after observing the data.

These results are available from the authors upon request. The robustness of the other variables as growth determinants is not affected by the use of these subsamples.

A deeper analysis of the Williamson hypothesis falls outside the scope of this paper. Crespo-Cuaresma et al. (2009b) investigate this issue further.

The full set of results is available from the authors upon request.