Firm heterogeneity in productivity across Europe: evidence from multilevel models

Abstract

This paper analyses the Total Factor Productivity (TFP) heterogeneity of a sample of manufacturing firms operating in seven EU countries (Austria, France, Germany, Hungary, Italy, Spain and the UK). TFP data refer to 2008. The empirical setting is based on the multilevel modeling which provides two main results. Firstly, we show that TFP heterogeneity is largely due to firm-specific features (85% of TFP variability in the empty model). Interestingly, we find that some key-drivers of firm performance (size, family management, group membership, innovations and human capital) are significantly related to TFP, but do not, on the whole, absorb much of firm TFP variance, implying that differences in productivity are due to notable yet unobservable firm characteristics. Secondly, as far the role of localization is concerned, we demonstrate that the country effect is more influential than region effect in explaining individual productivity. Net of the country effect, the localization in different European regions explains about 5% of TFP firm heterogeneity. When considering the case of three individual countries, France, Italy and Spain, location in different regions explains 5.3% of TFP heterogeneity in Italy, while this proportion is lower (3.6%) in France and higher (9.9%) in Spain.

Keywords:

• TFP heterogeneity
• firm behaviour
• localization
• European countries
• multilevel model

JEL classifications:

• C30
• D22
• L60
• R15

Acknowledgements

The authors thank Giuseppe Albanese, Carlo Altomonte, Tommaso Aquilante, Antonio Aquino, Matteo Bugamelli, Paola Cardamone, Marco Cuccurelli, Sergio Destefanis, Giorgio Fazio, Arzdar Kiraci, Janna Smirnova, Marco Vivarelli and three anonymous referees for valuable suggestions on an earlier version of the paper. They also are grateful to the participants of the 54th ERSA congress (St Petersburg, 26–26 August 2014), the 55th Annual Scientific Meeting of the Italian Economic Society (23–25 October 2014), the workshop ‘Explaining Economic Change’ (Rome, 12 November 2014) and the seminar at the University of Salerno (18 December 2014) for their helpful comments. Grateful acknowledgements are also due to Bruegel (www.bruegel.org) for making available the EFIGE data set in the extended version including the elaborations of Total Factor Productivity and for the easy access and use of the data set at Brussels. Editorial assistance by Kevin O'Connell and John Richard Broughton is also acknowledged. Usual disclaimer applies.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Francesco Aiello http://orcid.org/0000-0001-7533-4927

Notes

1. Altomonte and Colantone (2008) calculate several compositional effects of multinational enterprises and demonstrate that the regional disparities observed in Romania over the 1995–2001 period depend on the interaction between firm-level dynamics and the initial market conditions. Aiello, Mastromarco, and Zago (2011) used a panel of Italian firms to decompose the output growth into factor accumulation, technological change, efficiency and scale effects over the 1998–2003 period. They found that efficiency change (technological catch-up) explains much of the output growth observed in Italy, as a whole, and in the two macro-areas (North and South) of the country, separately. The connections between micro and aggregate industry productivity have been surveyed by Foster, Haltiwanger, and Krizan (2001) and Van Biesebroeck (2003).

2. As regards the choice to use TFP, it is worth pointing out that a vast literature demonstrates how the economic divide observed across countries and regions is mainly due to differences in TFP instead of differences in physical and/or human capital deepening. This issue has been initially demonstrated by the seminal studies of Hall and Jones (1999), Klenow and Rodriguez-Clare (1997) and Caselli (2005).

3. The sampling design has been structured following a three-dimensional stratification: industry (11 NACE-CLIO industry codes), region (at the NUTS-1 level of aggregation) and size class (10–19; 20–49; 50–250; more than 250 employees). Given their importance in aggregate competitiveness dynamics, but their relatively light weight in standard stratification of the population of firms, large firms have been oversampled. In computing the correlation over time (2001–2009) between some variables in the EFIGE data set (aggregated with proper weighs) and the national statistics provided by EUROSTAT, Altomonte and Aquilante (2012) show that the correlations are 0.82 for labour productivity, 0.71 for labour cost, 0.52 for revenues and 0.61 for workers. Correlations increase to 90% when considering the countries (France, Italy and Spain) with a good quality of balance sheet data. For details on the EFIGE data set, see Altomonte and Aquilante (2012) and Barba Navaretti et al. (2011).

4. As a by-product of the EFIGE project, the survey data have been integrated with firms' balance sheets of Amadeus database managed by Bureau van Dijk. The survey data set is available in different versions, depending whether the user has an active licence with Bureau van Dijk. In this paper, we mainly refer to the version which is freely downloadable from the Bruegel website, plus the TFP array released by Bruegel after presenting a research proposal. We complement the study on TFP by using labour productivity and labour costs (Note 16 and Table A2). A potentially important data limitation of original ‘free’ data set is that it includes just randomised regional and industry identifiers. This is a sensitive data-related issue that we address by running all regressions at Bruegel, in Brussels.

5. Hungarian data on TFP seem surprising, given that the GDP pro-capita in this country is far below the level of the other countries of the EFIGE sample (it was 40% lower than the 2012 EU-28 average). While the understanding of this country-specific evidence goes beyond the objective of the study, in the econometric section of the study we perform some robustness checks aimed at controlling for any potential bias due to outliers.

6. It is well known that a multilevel approach is not the only way to address the hypothesis of residuals independence. Spatial econometrics has made important advances in this respect, even though the interest is confined to single-level relationships (firms and regions), without treating the micro–macro interactions as multilevel does. Some methodological attempts to combine multilevel models and spatial econometrics are in Corrado and Fingleton (2012).

7. For Equation (5), VPC coincides with the intra-class correlation that measures the expected degree of similarity between responses within a given cluster (e.g. region). This equivalence will not hold in more complex models, such as those including random coefficients (Leckie 2013).

8. In the multilevel approach, a key issue to be addressed concerns the sample size at any level of analysis. Indeed, the requirements of precise measurement of between-group variance impose a ‘sufficient’ number of clusters. Although there are some, albeit very different from each other, rules of thumb, a clear indication does not exist in this respect (Richter 2006). Some authors suggest that 20 is a sufficient number of groups (Heck and Thomas 2000; Rabe-Hasketh and Skondal 2008), others 30 (Hox, 2002) or 50 (Mass and Hox 2004). In addition, it is worth noting that in random-effects models the clusters must be sized with at least two observations. The alternative is a fixed-effects approach in which the number of groups is not important, although their dimension then becomes crucial as the estimated group effect is unreliable for small-sized groups. These numbers condition our empirical setting: the preferred specification is a two-level random-intercept model where firms and regions are treated as source of randomness and countries and sectors are modelled with dummy variables.

9. When considering sectors a source of randomness, the estimations have been made through the model allowing for random intercepts for sectors and regions and augmenting this specification with the interaction region–sector (as Equation (8) briefly highlights)

10. The contribution of country effect is calculated by comparing the total TFP variance (0.03) explained at the regional level in the empty model (column 1 of Table 3) and the variance (0.011) obtained when this model is augmented by country dummies (column 1 of Table 4), that is: [(0.03–0.011)/0.03] (cf. Note 21).

11. In the remainder of Table 4, the country effect is modelled with dummies, whereas sectors act as random instead of fixed effects. In other words, these estimations replicate all the models used in Table 3, with the inclusion of country dummy variables. As can be seen, the results suggest that the proportion of TFP variance explained by the region-random effect is 4.4% in model 5, and 3.5% in model 6. Sectors contribute to explain about 11.5% of TFP variance. The evidence in columns 4–6, however, suffers from a small number of sector groups, and should thus be treated with caution.

12. The results on the capability of regions to explain the TFP heterogeneity are robust to the potential bias due to outliers (cf. Section 2). Indeed, the evidence holds when regressions for the EU7-EFIGE sample are estimated when excluding (a) Austria and Hungary (columns 1 and 2, Table A3), (b) the 739 firm observations falling in the first and last 5% TFP distribution (column 3, Table A3) and (c) Austria and Hungary and the 739 potential outliers (column 4, Table A3). As we can see, regions always explain less than 5% of TFP heterogeneity.

13. The contribution of sector effect is calculated by comparing the total variance (0.03) explained at the regional level in the empty model (column 1 Table 3) and the variance (0.025) obtained when this model is augmented by sector dummies (column 2 of Table 4), that is 16.7% [(0.03–0.025)/0.03] (cf. Note 21).

14. In order to check the robustness of location-effect at the regional level, we complement the analysis on TFP by considering the labour productivity and the labour costs. Results are displayed in Table A2. As far as the empty model is concerned, the location across the regions of EU7-EFIGE countries contributes to explain 25.4% of firm labour productivity (column 1). This proportion drops to 5.5% when the empty model is augmented with the country and the sectoral dummies (column 2). When attempting to explain labour costs heterogeneity, the role of regions is 4.8% in the empty model and just 0.8% in the more extended model. As in the analysis of TFP, these checks confirm that the country effect is more important than the region effect in explaining the heterogeneity in performance across European firms.

15. This is why foreign-controlled enterprises benefit both from being part of a global group and from the advantages of vertical and/or horizontal integration. They gain from factor price differentials, global economies of scale, outsourcing and the knowledge transfers from parent companies and flows among subsidiaries. This makes them more productive than firms which are not part of a foreign group (see, e.g. Griffith (1999) for evidence on the UK, Benfratello and Sembenelli (2006) for Italy and Weche Gelübcke (2013) for Germany).

16. Two hypotheses about the positive correlation between export activity and productivity have been extensively investigated. The first hypothesis is that the most productive firms self-select into foreign markets because they can overcome sunk costs associated with foreign sales (ISGEP 2008; Melitz 2003). The second hypothesis raises the possibility of ‘learning by exporting’. Firms participating in international markets acquire knowledge and technology with positive feedback as regards knowledge and technology. Furthermore, firms which are active in world markets are exposed to more intensive competition than firms which only sell their products domestically.

17. Because of the cross-section nature of Equation (9), one should be cautious in interpreting coefficients as a result of causal relationships. This is due to the potential reverse causation that Equation (9) may suffer. Hence, Table 6 should be viewed as a convenient way of summarizing statistical regularities among variables. Phrased differently: estimates must be read as associations rather than causality. However, some controls for level 2 endogeneity have been carried out. In fact, the endogeneity may occur at level 2 when the random effects are correlated with level 1 covariates. As shown by Snijders and Berkhof (2007), the correlation between the lower-level predictor variables and higher-level error terms can be removed by including the group-level means of the lower-level variables, a procedure known as the Mundlak (1978) correction. Estimations with Mundlak correction are displayed in Table A3. As can been seen, the results are qualitatively the same as those discussed throughout the paper.

18. Data for Spain come from the Spanish National Institute of Statistics, while for the other countries the source is the Eurostat Regional Statistics database.

19. The idea that agglomeration economies affect firms productivity can be traced back to the Marshallian tradition, according to which location within any given area allows firms to benefit from labor market pooling, the development of specialized intermediate goods and from knowledge spillovers. Localized knowledge externalities have gained the most prominence in the empirical literature (Feldman 1999). The main argument is that there may be geographic boundaries to knowledge spillovers since tacit knowledge cannot be easily transferred over large distances or bought via the market. Empirics (Audretsch and Feldman 1996; Jaffe 1989) have shown that there are knowledge spillovers and innovative activities tend to cluster spatially (for a wide overview, see Döring and Schnellenbach 2006). The importance of agglomeration externalities is also underlined by the new theories on economic growth, geographical economics and evolutionary economic geography (cf. Introduction). However, the investigation of agglomeration-externalities effect is beyond the scope of this paper. This explains why we just refer to a general measure of agglomeration with the restricted aim to control, in our regressions, for any potential agglomeration effect.

20. The coefficient of determination for the two-level model is given by

where N stands for the null model and M for the model of interest.The proportional reduction in each of the variance components can be calculated separately. The proportion of the level 2 variance explained by the covariates is

, and the proportion of the level 1 variance explained is

21. For Italy, Benfratello and Sembenelli (2006) find that only firms owned by USA corporations tend to be more productive than national-owned firms.

22. Performance measures are Tobin's Q and ROA in Barontino and Caprio (2006) and Maury (2006) and the market value in Pindado, Requejo, and Torre (2008). Barontino and Caprio find that performance is significantly higher in founder-controlled corporations and corporations controlled by descendants who sit on the board as non-executive directors. When a descendant takes the position of CEO, family-controlled companies are not statistically distinguishable from non-family firms.

23. For Italy, Cucculelli et al. (2014) show that family management has a negative effect on TFP but not for older firms: family-managed firms become more efficient as they mature. As for France and Spain, previous research focuses on profitability and the role of family ownership by considering the generation of family management and the effect on firm. Sraer and Thesmar (2007) find that French family-managed firms, first or later generation, outperform non-family firms. For Spanish firms, the relationship between ownership concentration and performance is significant only in the first-generation family firms and it is positive at a low level of ownership concentration and negative at a high level (Arosa, Iturralde, and Maseda 2010).

24. Firms are defined ‘internationally active’ when they have been involved in at least one international activity such as exports, imports of materials or services, active or passive outsourcing, production in another country via direct investment (Altomonte, Aquilante, and Ottaviano 2012).

25. Crozet, Méjan, and Zignago (2011) argue that the exporter productivity premium could be due to omitted variables, correlated to the probability to export as, for example, belonging to a foreign group. Barba Navaretti et al. (2011) show that firms belonging to a foreign group are more likely to be exporters and this finding may suggest a cost reduction effect stemming from belonging to a foreign group.

26. Cassiman, Goloso, and Martinez-Ros (2010) suggest that one potential underlying mechanism for the selection of more productive firms in the export market could be the fact that successful innovation improves the firm's productivity and, hence, these more productive firms became exporters. As a result, the omission of an innovation variable from the analysis may lead to the overestimation of the productivity-export association. Using a panel of Spanish manufacturing firms for the period 1990–1998, they find support for their hypothesis. However, as far as French firms are concerned, Bellone, Guillou, and Nesta (2009) show that the introduction of innovation does not significantly alter the size of the export premium.

27. For reference, we also estimate Model 1 of Table 6 by running a standard OLS regression. In so doing, we have treated regions and sectors as fixed effects and clustered standard errors at regional and sectoral level. The results are displayed in the Table A4. As expected, OLS estimations and the significance of firm-specific factors do not qualitatively differ from those reported in Table 6, although hierarchical modelling has the advantage of discerning different sources of heterogeneity even in its most basic specification.