Growing Collaborations
Results
Estimated ERGM Specification
We reviewed 22 potential ERGM specifications to find the best-fit model . ERGM fit is determined through calculated AIC and BIC values , which are reported in the estimation output . The lower the value the better the ERGM ’ s fit to the data . We selected the ERGM specification with the lowest AIC and BIC goodness of fit values , using the structure of the informationsharing subnetwork as the dependent variable . This specification was then applied to ERGM estimations using the other three subnetworks as dependent variables . We utilized a geometrically weighted edgewise pair term ( gwesp ) with a non-fixed alpha for transitivity and a geometrically weighted degree distribution term ( gwdegree ), reviewed with a wide range of potential alpha values , for preferential attachment . Estimation terms available for ERGM allow for a consideration of whether homophilic pairings within any given attribute have a different influence on the probability of a link occurring ( Hunter et al , 2013 ). For example , estimation terms allow the researcher to determine if two state-level public agencies are more likely to form a partnership than two non-profit organizations that operate throughout Vermont . We determined that the best specification allowed for the probability of forming a link to vary
by an organization ’ s individual attribute values , homophilic pairings by attribute , and transitivity . Estimation using this specification reported the lowest AIC and BIC values for the information-sharing subnetwork . Models using preferential attachment did not perform as well as those that used transitivity while models including both transitivity and preferential attachment failed to run due to excessive correlation between model terms .
Error Rates in Network Forecasts
We performed nine model runs in each phase of analyzing our network growth algorithm . This includes four runs of our proposed algorithm , four runs of the fixed initial conditions control , and one run of the random change control . The AIC and BIC goodness of fit measures from the selected ERGM analyses for each run are included in tables 3 and 4 . The results from phase 1 , which forecasted new and existing links in the 2012 data from existing links in the 2012 data that are presented in table 3 . The results from phase 2 , which forecasted new and existing links in the 2014 data from the new and existing links in the 2012 data , are presented in table 4 . The results show that our model did not perform as well as the fixed initial conditions control but did perform better than the Erdos-Renyi control . A pattern emerges in the results that shows that our model tended to produce twice the
15