Andreas Dzemski

Thematic seminars
Economic theory seminar

Andreas Dzemski

University of Gothenburg
An empirical model of dyadic link formation in a network with unobserved heterogeneity
Venue

VC Salle A

Centre de la Vieille-Charité - Salle A

Centre de la Vieille Charité
2 rue de la Charité
13002 Marseille

Date(s)
Thursday, November 24 2016| 12:00pm to 1:15pm
Contact(s)

Ugo Bolletta: ugo.bolletta[at]univ-amu.fr
Mathieu Faure: mathieu.faure[at]univ-amu.fr

Abstract

In this paper I present a new model of dyadic link formation for directed networks that extends the classical model by Holland and Leinhardt 1981. Agents are endowed with unobserved effects that govern their ability to establish links (productivity) and to receive links (popularity). The unobserved effects are modelled by a fixed effects approach allowing for arbitrary correlation between the observed homophily component and latent sources of degree heterogeneity. The model can be estimated by conditional ML but inference is non-standard due to the incidental parameters problem (Neyman and Scott 1948). I consider estimation of the parametric part of the linking model as well as estimation of a measure of network transitivity. Moreover, I suggest a specification test for the dyadic model based on predicted transitivity. My approach overcomes the incidental parameters problem by using explicit correction formulas based on an asymptotic approximation that sends the number of agents to infinity. My simulation design suggests promising finite sample performance. A linking model neglecting unobserved sources of degree heterogeneity predicts an insufficient amount of transitivity. This effect is proven for a stylized model and its empirical relevance is confirmed using data on favor networks in Indian villages. In this application, a transitivity statistic changes sign when unobserved agent effects are added. This suggests that, in the real world, unobserved heterogeneity may be a primary driver of local clustering behavior.