Cohesion and structural equivalence are two competing network models to explain diffusion of innovation. The dispute of which model plays a more influential role has not been resolved. This paper attempts to reconcile this problem in a large network setting - adoption of caller ringback tone (CRBT) in a cellular telephone conversation network. Since this societal scale network is very large, we use a novel technique to extract multiple densely connected and self-contained subpopulations from the network. We found subpopulation size in such million-node network only falls in two levels, 200 and 500, in the extraction step. Using a new auto-probit model with network terms, we then compare the competing influences of cohesion and structural equivalence on each of the subpopulation extracted. Finally we use meta-analysis to summarize the estimated parameters from all subpopulations. The results show CRBT adoption is affected by both cohesion and structural equivalence. The size and direction of network influence both change with the size of group. Structural equivalence has a negative effect on adoption when group size is at about 200, and has a positive effect when group size is at about 500. The effect of cohesion, on the other hand, is consistent.