Agent-based models and hypothesis testing: an example of innovation and organizational networks

Fagiolo et al. (2007) point out conventional hypothesis testing procedures also struggle with the relationship between theory and empirical data, but neoclassical economics has a consensus about testing procedures that has not yet emerged in AB modeling.

By restricting agents to a single play in each round, a simple snapshot of the distribution of decisions after each round of play captures every agent's decision. Removing this constraint results in a few more individual decisions in each round but has virtually no impact on the distribution of decisions given in Tables 1 and 2 or on the implications of the model.

For another example a computational/analytical anchor, see Wilhite (2006a, 2006b).

There is a vast literature on the evolution of networks, but those studies usually focus on attributes of the emergent network. See Jackson (2005) and Vriend (2006) for overviews and Chang and Harrington (2005) for an application.

Because most of the agents in a tree network have a single edge (all those agents out on the end of the branches have only one neighbor), edges were only added during its rewiring. In addition, rewiring the complete network usually just severed edges (the proposed new link already exists).

While the randomness introduced by rewiring increases the survivability of innovative agents, the effect diminishes with increased rewiring. In Table 2, most of the increased flexibility brought forth by rewiring occurs by the time r = 0.25. Additional rewiring continues to increase the chance that the firm will develop an innovative culture, but the marginal impact of additional rewiring falls.

Perhaps the most commonly used stand-in for innovativeness is patent awards and/or patent applications. However, patent information is unreliable for international firms and patents do not reflect the incremental innovations firms introduce but do not patent. Because we are interested in all types of innovation, patent applications are not the appropriate measure for this study.

A referee suggests that firms in business for only a few years may have a higher percentage of new products simply because they themselves are new. This conjecture seems to be warranted, but the impact on innovation is unclear. Omitting firms less than 15 years old reduces the sample by about 100 firms and the coefficient on age become insignificant in the first regression. However, the coefficient on age becomes stronger (significant at the 0.05 level) in the project-level innovation regression when these younger firms are omitted. Age seems to have an impact but deserves more study.