A coherent approach to analysing heterogeneity in network data

Introduction

The overarching goal of this project is to develop a coherent set of econometric methods to deal with unobserved heterogeneity in the analysis of social interactions between agents. Such heterogeneity is well recognized to be important.

Importance of Heterogeneity

It is often of great interest to document the degree of heterogeneity, evaluate its impact, and uncover the existence and form of any complementarities that may exist between agents. With the growing availability of network data, questions of this kind are increasingly being asked in applied work.

Current Challenges

The development of appropriate econometric tools to answer them has, however, not followed suit. If anything, recent theoretical work has pointed at substantial difficulties with the so-called fixed-effect approach currently serving as the workhorse tool.

Proposed Solution

This project recognizes the potential of taking a random-effect view. For settings where agents interact in pairs, such a view has received some attention in the literature. However, to date, it struggles with issues of identification, estimation, and computation.

New Nonparametric Approach

We will develop a new nonparametric approach that provides a solution to each of these three issues.

Extension to Larger Groups

We will next venture forward and extend this framework to situations where agents interact in larger groups. Both collaborative and non-collaborative settings will be considered, thereby covering:

1. Team production
2. Competition
3. Peer effects

Special Considerations

Special attention will be given to recovering treatment effects in the presence of social interactions, where interference on unobservable confounders is an issue.

Addressing Data Limitations

For situations where data limitations prevent a fully nonparametric approach, instrumental-variable methods that build on flexible functional form restrictions will be developed.

Implementation and Illustration

The statistical properties of the proposed estimators will be derived, software implementation will be provided, and empirical illustrations will be presented to highlight the usefulness of the methods.