Page - (000362) - in Knowledge and Networks

360 absolute difference between the ages of the actors (measured as the length of time since their first patent application). Our age variable was also assumed to capture the effect of firm size because the age and the size of the firm are usually highly correlated. Estimation Strategy The choice of a pair of partners to cooperate was modeled as the probability of observing the realization of a link (coopi,j,t taking the value 1) contingent on the explanatory variables we have discussed in this section. The decision to collaborate in the form of a copatent is a binary one (see Fig. 16.2). We therefore estimate the following logistic model (see Kennedy, 2009). We included all realized and potential i, j combinations over the period from 1983 to 2010. To prevent potential biases from confining our sample to collabora- tive actors only, we included all possible combinations between the focal firms and all actors who had patented at least once. However, inclusion of combinations with all potential actors in the sample (even those that have never collaborated) intro- duces a source of bias due to unobserved heterogeneity. That is, control-group dyads that were never realized might differ systematically in unobserved factors from dyads that were realized at least once. These differences in unobserved characteris- tics might account for systematic differences in the general propensity of actors to collaborate. Furthermore, other specific factors that are not observable and that therefore cannot be included in our model might have caused the formation of each dyad (Gulati & Gargiulo, 1999; Heckman, 1981). To account for pair-specific het- erogeneity, we applied a random-effects panel model by including a random inter- cept for each pair. We thereby assumed that the unobserved differences in the dyads were the results of a random process. However, this method also comes with the strong assumption that the unobserved factors are not correlated with any of the explanatory variables. This assumption is hard to test empirically. Conversely, the fixed-effects estimator would remove these time-invariant factors but would dra- matically shrink the size of the sample. This change would come at a cost: The number of observations would drop from more than 300,000 to 501. Moreover, random-effects estimation allows the model to include additional time-invariant variables, such as DStatus. Given these considerations, we preferred the random- effects over the fixed-effects model. Another issue that arises in the analysis of network data is the dependence of observations. The observations are not completely independent; individual actors might be part of multiple dyads. Consequently, the estimates are consistent, but the standard errors might be underestimated (Kennedy, 2009). Because we could not make any distributional assumption, we obtained robust standard errors by resorting to bootstrapping methods for panel data. We calculated the standard errors from the empirical distribution that was drawn by resampling the original dataset in 1000 iterations. Another form of bootstrapping commonly used to analyze dyadic data is U. Cantner et al.