The past and future of evolutionary economics: some reflections based on new bibliometric evidence

Abstract

The modern wave of ‘evolutionary economics’ was launched in 1982 with the classic study by Nelson and Winter. This paper reports a broad bibliometric analysis of ‘evolutionary’ research in the disciplines of management, business, economics, and sociology over 25 years from 1986 to 2010. It confirms that Nelson and Winter's book (An evolutionary theory of economic change, Harvard University Press, Cambridge, MA, 1982) is an enduring nodal reference point for this broad field. The bibliometric evidence suggests that ‘evolutionary economics’ has benefitted from the rise of business schools and other interdisciplinary institutions, which have provided a home for evolutionary terminology, but it has failed to nurture a strong unifying core narrative or theory, which, in turn, could provide superior answers to important questions. This bibliometric evidence also shows that no strong cluster of general theoretical research immediately around Nelson and Winter's classic book has subsequently emerged. It identifies developmental problems in a partly successful but fragmented field. Future research in ‘evolutionary economics’ needs a more integrated research community with shared conceptual narratives and common research questions, to promote conversation and synergy between diverse clusters of research.

Appendix: Bibliometric methodology

Bibliometrics involves the analysis of patterns that appear in the publication and use of documents, to shed light on the nature and development of a discipline.⁸

Citation analysis is a powerful tool for the identification of intellectual bases and underlying research streams (Usdiken and Pasadeos 1995; Pasadeos et al. 1998; Schildt and Mattsson 2006). Citation analyses divide into ‘macro’ approaches that focus on the overall structure of disciplines, and develop principles governing the evolution of science, and ‘micro’ approaches that describe retrospectively the structure and historical development of schools of research and their interdependencies (Gmür 2003). This study fits with the micro-stream of research.

Criticisms of the use of citation analysis concern citation biases, a focus on only published articles and books, and the technical limitations and imperfections of citation indices and bibliographies (Macroberts and Macroberts 1989; Osareh 1996). With improved databases, some of these limitations have been ameliorated (Sillanpää 2006). Important limitations remain, but we have done our best to address possible biases and to remove errors from our extensive database.

Our approach combines co-citation and cluster analysis (Schildt et al. 2006; Sillanpää 2006; Schildt and Mattsson 2006). Co-citation analysis reveals the closeness of two pieces of work in a common discourse. Cluster analysis and network analysis enable further structuring the research field under study. Our approach of highlighting the structure of the field by the means of both cluster analysis and network analysis (which produce highly similar results) mitigates the biases in any individual research method.

1.1 Data

We used data from the Social Science Citation Index (SSCI) of the Thomson-Reuters Web of Science, which is a massive multidisciplinary index to social sciences journals. It indexes over 1,720 journals across 50 social science disciplines; and individually selected, relevant items from over 3,300 of the world’s leading scientific and technical journals.

Within the database, we conducted searches for the word evolution and its derivatives. Further searches confirmed the result of Dachs et al. (2001) that related search words (e.g., Schumpeter, biological, biology, genes) yielded a much smaller number of retrieved articles, compared to ‘evolution’ and ‘evolutionary’. To narrow down the number of hits (over 20,000), and confine our study to business-related issues, we refined the search to cover documents related to the following fields only: management, business, economics, and sociology. The search was further refined to cover articles only, thus excluding book reviews, notes and editorial announcements.

The start date of the searches was 1 January 1986 (the 1st accessible year on the Thomson-Reuters database), and the end date was 31 December 2010. Before 1986, much fewer articles discussing ‘evolution’ were published in the social sciences (Hodgson 1998). To identify changes for the whole period, we retrieved 8,474 articles; 217 were published during 1986–1990, 954 during 1991–1995, 1,637 during 1996–2000, 2,172 during 2001–2005, and 3,494 during 2006–2010. These were all possible citation sources.⁹

Sitkis computer software (Schildt 2004) was used to download data on possible citation objects from the Web of Science to a Microsoft Access database. The articles in the whole period cited another 373,848 texts, of which 24,098 were discarded by the program.¹⁰ The program reported disregarded citations, and all of these were checked manually. Most referred to newspapers, trade journal articles or statistics and were deemed tangential to this analysis. A small number of corrections were made.

Thomson Reuters’ data are not entirely accurate. In the first 5-year sub-period, we went through all the citations manually and made any required corrections. But because the total number of references in other sub-periods exceeded 20,000, going through all of these was impossible. Schildt (2002) argued that correcting citation data for the top 20–50 authors or documents is sufficient to provide reliable and usable results. But we imposed higher standards.¹¹

References made to reprints and book editions were combined as references to one, original article or book. But citations to compiled book editions were left unaltered (Sillanpää 2006).

1.2 Analysis

A co-citation involves a link between two documents that is created by a later document (Griffith et al. 1974). A co-citation measures ‘the frequency with which two documents are cited together’ (Small 1973, p. 265). If two articles are cited in the same text, then they may be closely related to each other either because they are part of the same topic area or because their topic areas are closely connected (Small 1973; Cawkell 1976). Although some co-citations are between unrelated references, a sufficiently large sample of cited articles enables researchers to mitigate this problem (Schildt and Mattsson 2006).

Using Sitkis software, we produced a co-citation network for each sub-period. A threshold level, based on the frequency the citing articles cited the references, was used to exclude references that did not have a serious impact on the study (Schildt et al. 2006). A series of two-dimensional (citer-cited) networks were then produced to determine the best threshold level. In a two-dimensional network, the citing articles were the first dimensions, and the cited texts acted as their affiliations. When the threshold was raised, the number of remaining cited documents decreased, and the number of citing articles also declined. After testing the series of networks, the threshold was set to a point at which lowering the threshold level by one would bring the maximum marginal increase in the number of cited articles. Below this threshold, the heterogeneity of the cited documents increased considerably, leaving additional documents outside the core of the field.¹²

Next, we normalized co-citation data to emphasize proximate relationships between similar references that are not cited as often as the most common references (Gmür 2003). The normalized co-citation strength measure, S, for individual pairs was calculated by means of the Jaccard index (Small and Greenlee 1980). The co-citation link strength S(A, B) between papers A and B is defined as follows:

$$S(A,B) = \frac{a \cap b}{a + b - a \cap b},$$

where a represents the number of citations to document A, b the number of citations to document B, and a ∩ b the number of co-citations of A and B.

We employed cluster analysis to classify objects into clusters that maximize homogeneity within clusters and heterogeneity between clusters (Culnan 1987; Hair et al. 1998).¹³ We employed Johnston’s average-link hierarchical algorithm, as in the Ucinet 6 software (Borgatti et al. 2002), to produce clusters from the co-citation network data. In the average-link algorithm, the distance between two clusters is the average dissimilarity between members (Borgatti et al. 2002). According to Sillanpää (2006), the average-link method produces clusters more continuously than other hierarchical methods.

Ucinet Netdraw software was used to draw network figures from the co-citation network data for the sub-periods. To make reading of the networks easier, we reduced the number of visible links by imposing an arbitrary cut-off level of co-citation strength. The links below the cut-off level were left out of the figures, as well as documents isolated by the procedure.¹⁴ The Netdraw software then arranged the remaining documents according to geodesic distances.

We performed the cluster analysis for documents in co-citation networks for the sub-periods, and for the 1986–2010 period as a whole. As there is no unique way to identify clusters, their identification involves some interpretation: we used similarity levels calculated by the algorithm as guidelines. We set two rules for the identification of clusters from the tree diagrams. First, an independent cluster or sub-cluster must consist of at least two documents. Second, main clusters were separated at a similarity level that produced a moderate number of clearly identifiable clusters.¹⁵