مهارت ها، تقسیم کار و عملکرد در اختراعات جمعی: شواهدی از نرم افزار منبع باز

This paper investigates the skills and the division of labor among participants in collective inventions. Our analysis draws on a large sample of projects registered at Sourceforge.net, the world's largest incubator of open source software activity. We test the hypothesis that skill variety of participants is associated with project performance. We also explore whether the level of modularization of project activities is correlated with performance. Our econometric estimations show that skill heterogeneity is associated with project survival and performance. However, the relationship between skill diversity and performance is non-monotonic. Design modularity is also positively associated with the performance of the project. Finally, the interaction between skill heterogeneity and modularity is negatively associated with performance.

Collective inventions among profit-seeking individuals and organizations have become popular in the economics literature since the seminal paper of Robert Allen (1983) on the iron district of Cleveland in the nineteenth century. More recently, collective inventions have come to the forefront of economists' attention because of the diffusion of open source software (OSS hereafter). OSS can be viewed as a ‘virtual’ community of practice made up of inventors who voluntarily contribute to multiple collective inventions. OSS offers expert developers the opportunity to participate in innovation networks which are, to some extent, reminiscent of the communities of users in the early age of computing (Steinmueller, 1996 and Torrisi, 1998) or other user-centered innovation processes such as those analyzed by von Hippel (1988). Most studies have attempted to explain why a growing number of independent developers (‘hackers’) voluntarily disclose their inventions. Several theoretical works seek to understand not only the motivations for disclosure of the source code, but also the social norms and the patterns of collaboration among distributed developers, and the implications for efficiency and social welfare (e.g. Raymond, 1999, von Hippel, 2001, Lerner and Tirole, 2002, Johnson, 2002, Harhoff et al., 2003 and Dalle and David, 2005).

Table 1 summarizes the variables used in the empirical analysis and provides some descriptive statistics for the sample of 9076 projects. For about 26% of these projects (2372) data on skills are missing or not reported. Given the importance of this variable in our analysis, we compared these two categories of projects and checked for statistically significant differences in the two distributions. About 29.13% of sample projects that report data on skills are active while the same percentage is equal to 24.83% among projects for which data on skills are not available. Moreover, we found significant differences in the average size, the number of CVS commits and the number of subprojects. For instance, the average number of CVS is 59 (SD = 258.46) for projects with missing skills as against 40 (SD = 398.94) for projects that report data on skills. Projects reporting skill data are also characterized by more file releases (mean = 0.476, SD = 1.768) than those that do not report data on skills (mean = 0.328, SD = 1.333). Finally, projects reporting skill data are bigger in size (mean = 4.26, SD = 4.166 vs. mean = 2.82, SD = 1.719) and have more subprojects (mean = 0.83, SD = 1.818 vs. mean = 0.37, SD = 1.067). A statistically significant difference between the two subsamples is registered in each of the dimensions above reported (Pearson's chi-squared test, p-value < 0.01 for all comparisons). Overall, projects with missing data on skills are less likely to be active, are smaller in size, are less productive in terms of CVS commits and file released and have a limited number of subprojects when compared to projects with data on skills. To avoid coefficients of skill-related variables being biased we introduced dummy variables for observations with missing data (see Hall and Ziedonis, 2001 for a similar treatment of potential selection on key regressors).