Clustering in the Creative Industries: Insights from the Origins of Computer Software

Abstract

We use several different sources (a 1970 Roster of Organizations in Data Processing and the 1960 and 1970 Censuses of Population) to study patterns of geographic clustering at the very origins of the software industry. We find a strong trend toward clustering of the industry in a few metropolitan areas. Furthermore, we uncover a tendency in the early software industry to agglomerate in close proximity to some of its main customers. This tendency holds even after controlling for region-specific heterogeneity and for the potentially endogenous nature of the software customers' location decisions. We explore the factors that may have driven the observed clustering patterns and suggest directions for further research.

Keywords:

• Industry studies (services)
• information and Internet services
• computer software

Acknowledgements

The authors received extremely helpful comments from four anonymous reviewers and from Dr Mark Lorenzen, the chief editor of this journal. The opinions expressed in this paper are exclusively the authors and do not necessarily coincide with those of the institutions with which they are affiliated.

Notes

¹ Rosenthal and Strange (2004) provide a comprehensive survey of econometric approaches and findings in this literature, whereas Duranton and Puga (2003) cover some of the main theoretical issues.

² In recent years this journal has published a number of articles on topics related to our study. See, for example, Dahlander et al. (2008), Di Maria and Finotto (2008), Kaiser and Muller-Seitz (2008), Lange et al. (2008), Lazzeretti et al. (2008), Staber (2008), Kesidou et al. (2009) and Visser (2009).

³ Our study focuses on the origins of the software industry in the USA. For an international perspective see the studies in Mowery (1996).

⁴ The acceleration of entry into software is, we believe, clearly associated with the advent of System/360. Among the software companies in the 1970 Roster that provide information on the computing hardware for which they produce/sell software, IBM System/360 is the hardware of choice for 57 per cent of companies (including those that report IBM System/360 as well as another hardware system).

⁵ The IND 1990 codes are available at http://usa.ipums.org/usa-action/codes.do?mnemonic = IND1990 (last accessed 30 July 2009). We used code # 732 for the software industry, code # 292 for ordnance, code # 322 for the manufacturers of computer hardware, code # 352 for aircraft and parts, code # 700 for banking, code # 711 for insurance, code # 890 for accounting, code # 892 for management and code # 932 for national security. One caveat is worth mentioning: code # 732 includes companies that supplied computer software and computer services. The 1970 Roster reveals that most companies that supplied computer software also delivered some sort of computer service, and most companies that supplied computer services were also involved in the production and/or sale of software. Apparently there were, however, more companies supplying services but not software than there were companies supplying software but not services. Thus, we raised the question: could we be introducing biases in our calculations because of the existence of some companies that delivered computer services but not computer software? The 1970 Census does not allow us to explore this issue, but the 1970 Roster does. We sorted the metropolitan areas in two ways: first, we ordered them by the total number of people employed in companies that supplied services generally (including those employed in companies that supplied both services and software); then, we ordered them by the total number of people employed in companies that supplied software (regardless of whether they also supplied a computer service or not). The correlation coefficient between these two orderings was about 0.99. In short, if there are any biases in our calculations arising from our use of code # 732 for classifying software-industry companies, they are trivial.

⁶ Some of the discrepancies may arise from two sources. For one, the 1970 Roster is a sample of establishments whereas the 1970 Census is a sample of households. Furthermore, location quotients involve scaling, in such a way that the size of the region (in terms of employment) is taken into account.

⁷ The results, not reported here for space reasons, are available from the authors upon request.

⁸ All the regressions reported in this study were estimated with robust standard errors. For every model we estimated via OLS we also estimated a TOBIT version to account for censoring in the left-hand-side variable (the location quotient for the software industry). The TOBIT results, however, were not substantially different from the OLS numbers reported in the relevant tables or text. In addition we explored different functional form specifications. For example, we included the location quotient squared for each one of the heavy demanders on the right-hand side, but this term turned out to be not statistically significant in most cases. Furthermore, we tried a log–log specification (that is, we estimated a model for the log of the location quotient for the software industry on the log of the location quotient for the heavy demanders), but the results were not substantially different from those reported in the core of this study.

⁹ These results are available from the authors upon request.

¹⁰ These findings are available from the authors upon request.

¹¹ We also estimated a different model (not reported here for space reasons), which suggests that the clustering of one additional heavy demander in a region added between 0.25 and 0.28 units to the location quotient of the software industry in that region. These results are available from the authors upon request.

¹² Strictly speaking, in both cases there is an endogeneity problem. In the first case, the location patterns of both the software industry and the heavy demanders would be endogenous vis-à-vis some region-specific factor that is potentially unobserved by the econometrician. In the second case, the location patterns of the heavy demanders would be endogenous vis-à-vis those of the software industry.

¹³ Data for college/university employment are available for 1970 but not for 1960.

¹⁴ The full results are available from the authors upon request.

¹⁵ Since in both cases we relied on 1960 data, we were forced to estimate the models on the basis of the geographic units identified in the 1960 Census, which are considerably less than those identified in the 1970 Census. This is reflected in the degrees of freedom reported in the relevant tables. In any case, the fact that the coefficients on the combined heavy-demander sector estimated in this context reveal a substantial and statistically significant effect, even with a considerably smaller number of observations, suggests that the relationships we have uncovered are quite robust.

¹⁶ The advent of the computer timesharing industry (Campbell-Kelly and Garcia-Swartz, 2008a) surely made it easier to develop software remotely, but there were limits to the kinds of software that could be produced in this way.