Spill over or Spill out? – A multilevel analysis of the cluster and firm performance relationship

ABSTRACT

Regional clusters have become an inseparable component of modern economies. Spurred by the idea that clusters unrestrictedly encourage firm innovativeness, the cluster approach has particularly gained attention among politicians. Nevertheless, due to a lack of holistic consideration of different influencing variables, the scientific results about the effect of clusters on firm innovative performance are highly contradictive. Consequently, this paper aims to empirically investigate the conditions through which companies can gain from being located in clusters, focussing thereby particularly on moderating variables that relate to possible knowledge spillovers. Therefore, three different levels of analysis are considered separately and interactively. By analysing a unique multilevel dataset of 11.889 companies in Germany evidence is found that being located in a cluster has a positive impact on firm innovativeness. However, the results also indicate that firms benefit unequally within the cluster environment, depending on the specific firm, cluster and market/industry conditions.

KEYWORDS:

• Knowledge spillovers
• cluster effect
• firm performance
• multilevel analysis
• innovation

Acknowledgments

The author would like to thank the Wissenschaftsstatistik GmbH of the Stifterverband für die Deutsche Wissenschaft for providing access to the data (in particular Barbara Grave, Thu-Van Nguyen and Julia Angenendt for their continuous support) as well as the participants of the 4th Geography of Innovation conference 2018 in Barcelona and the research symposium 2018 in Kolding for useful comments on earlier drafts. Furthermore, the author would also like to thank Thomas Brenner for the lively exchange and support in the creation of the cluster index. Moreover, the author would like to thank the Associate Editor Andrea Morrison as well as two anonymous reviewers for their constructive feedback. Furthermore, the author gratefully acknowledges financial support from the Federal Ministry of Education and Research [grant number 03INTBF05A].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ A comprehensive overview about the financial budgets of different cluster programs in Europe is provided by Zenker et al. (2019).

² Although not directly measured (see limitations), the considered moderating variables refer more to the underlying notion of knowledge spillovers.

³ In accordance with Barney (1991) resources are here defined as ‘all assets, capabilities, organizational process, firm attributes, information, knowledge, etc. controlled by a firm that enable the firm to conceive of and implement strategies that improve its efficiency and effectiveness.’ (Barney 1991, p. 101).

⁴ In accordance with Hervas-Oliver et al. (2018) the terms absorptive capacity and innovation capability are used interchangeably.

⁵ According to Dyer and Singh (1998) relational rents are defined as: ‘(…) supernormal profit jointly generated in an exchange relationship that cannot be generated by either firm in isolation and can only be created through the joint idiosyncratic contributions of the specific alliance partners.’ (Dyer and Singh 1998, p. 662).

⁶ Important exceptions are in this context Knoben et al. (2015) as well as Rigby and Brown (2015).

⁷ For example, clusters can be found in the ceramic tile and fashion industry (e.g. in north-east Italy), in the software industry (e.g. in Bangalore in India) and in the automotive industry (e.g. in south-western Germany).

⁸ However, it should be noted that there are also conflicting findings on the impact of clusters on the emergence of radical innovations (e.g. Hervas-Oliver et al. 2019), so further investigations in this new research field are needed.

⁹ For a good overview of the most common distance measures and functions, please see Brenner (2017).

¹⁰ For a good visualisation, please review Figure 2 in Grashof, Hesse, and Fornahl (2019).

¹¹ As a first robustness check, for the firm´s centrality within the corresponding cluster core a dummy variable has additionally been calculated based on the above median of the cluster index (equal to a value of 2.83). The results remain thereby the same and can be provided upon request.

¹² Despite the importance of R&D collaborations (e.g. Aschhoff and Schmidt 2008), as on the firm-level, other types of relations, such as business relations, are not considered. Due to the high relevance of R&D collaborations for the knowledge transmission, the used proxy is, however, argued to be in line with the original idea of McCann and Folta (2008).

¹³ Weighted number of patents means in this context that the corresponding applicant share is used. For example, on a patent with two applicants, each of them get an applicant share of 0.5.

¹⁴ The applied distance decay function is thereby based on travel times, where 45 minutes represent the limit for close geographical distance (Brenner 2017; Scholl and Brenner 2016).

¹⁵ In line with the procedure of the European Cluster Observatory a value of 2 is applied as the corresponding cluster threshold, indicating whether a firm is located in a cluster or not (European Cluster Observatory 2018; European Commission 2008).

¹⁶ To also control for the knowledge diversity within clusters (see model 7 in Table 4), the average spread of the company´s average internal R&D expenditure for the 50 different product categories within clusters is calculated.

¹⁷ Due to data constraints, the average number of research institutes within clusters could not be calculated for a longer time period.

¹⁸ For the final analysis, however, only the highest organisational level of the research institutes, e.g. universities and not their working groups, are considered.

¹⁹ For further information on this please review the terms of use available under https://www.stifterverband.org/fdz.

²⁰ This is, however, a rather typical problem when using individual firm data (e.g. Mairesse and Mohnen 2010).

²¹ Indeed, a combination between AMADEUS and ORBIS databases is used in order to cover preferably all listed firms in Germany.

²² In more concrete terms, the Token, N-Gram, Soundex and Token Soundex algorithms were used.

²³ Due to some changes in the questionnaire of the Stifterverband the dependent variable can only be averaged for the time period between 1997 and 2013. By only using the average it is argued that the results will be unaffected. As a control and a further robustness check the independent variables were also averaged for exactly the same time period as the dependent variable. The corresponding results of this robustness check are in line with the original results and can be provided upon request.:

²⁴ The standard errors are clustered by regional cluster.

²⁵ The corresponding descriptive statistics for all main variables can be provided upon request.

²⁶ Since the focus of this paper is actually on the conditions through which firms can gain from being located in a cluster, only the influence of the cluster affiliation is explained here. Other selected influences from Table 3 are, however, described if they differ from the results of the moderating variables presented in Table 4.

²⁷ Following the remarks by McCann and Folta (2011), it is argued that the concern that the most innovative firms choose to locate in clusters, creating a potential selection bias, cannot be justified empirically nor theoretically. This can also be confirmed by the mean and the standard deviation of firm innovativeness within clusters. The corresponding results can be provided upon request.

²⁸ More and stronger competition, promoted through knowledge leakages, will likely reduce the share of the focal firm´s product innovations in its sales.

²⁹ As a further sensitivity test, the effect of the degree of cluster external relations instead of the share of cluster external relations is additionally investigated. The results remain thereby however the same (positive and insignificant) and can be provided upon request.

³⁰ To verify the results, model 6 has been replicated without the share of cluster external relations. Instead, a new dummy variable has been added, indicating whether firms own a moderate share of cluster internal (external) relations between 0.3 and 0.7. The corresponding results remain insignificant und can be provided upon request.

³¹ In line with the approach by Hainmueller, Mummolo, and Xu (2019), the underlying LIE assumption (linear interaction effect) is tested. For all interaction terms the corresponding results of the suggested tests confirm that the LIE assumption is correct.

³² Here calculated as a dummy variable, where 1 means that the pace of technology evolution is equal or higher than the corresponding 75^th quantile. The 75^th quantile has been chosen, as at the very end of the distribution (from the 90^th quantile upwards) the number of observations for a fast technology evolution becomes too small for a further analysis. Nevertheless, in light of other contributions using for example simply the mean as a threshold (e.g. Audretsch and Feldman 1996), it is argued that the 75^th quantile represents a reasonable threshold for industries with a fast pace of technology evolution.

³³ To simplify the interpretation of the corresponding interaction term, a dummy variable is used for the firm´s central position within the cluster core (calculated based on the above median of the cluster index; equal to a value of 2.83). However, by directly using the cluster index (discrete variable) the results remain the same and can be provided upon request.

³⁴ Results can be provided upon request.

³⁵ A summary of all main results can also be found in Table A1 (appendix).

³⁶ The results of all robustness checks can be provided upon request.

³⁷ Since the underlying survey questions of the dependent and independent variables are only inconsistently answered over the years (e.g. Mairesse and Mohnen 2010), a regression on a balanced panel database would result in a significant loss of observations, thereby creating a potential bias. In more concrete terms, in the underlying panel sample of this study 50% of the firms answer only four times (between 1997 and 2015) the corresponding survey questions.

³⁸ The results of the carried out omitted variables tests, however, suggest that such a bias is not a mayor concern in this study.

³⁹ The procedure by Rigby and Brown (2015) can in this context serve as a valuable starting point, even though, following Krugman (1991), they also argue that knowledge spillovers cannot adequately be measured in a direct way (e.g. through patents).