Capturing Unobserved Correlated Effects in Diffusion in Large Virtual Networks

5 Social networks and social capital are generally considered to be important 6 variables in explaining the diffusion of behavior. However, it is contested 7 whether the actual social connections, cultural discourse, or individual 8 preferences determine this diffusion. Using discrete choice analysis applied 9 to longitudinal Twitter data, we are able to distinguish between social 10 network influence on one hand and cultural discourse and individual 11 preferences on the other hand. In addition, we present a method using freely 12 available software to estimate the size of the error due to unobserved 13 correlated effects. We show that even in a seemingly saturated model, the 14 log likelihood can increase dramatically by accounting for unobserved 15 correlated effects. Furthermore the estimated coefficients in an uncorrected 16 model can be significantly biased beyond standard error margins. 17