An inductive typology of egocentric networks with data from the Socio-Economic Panel

Abstract

In this article we apply Random Forests to data from the German Socio-Economic-Panel (SOEP), creating an inductive typology of egocentric networks. Using an earlier application of the machine learning algorithm as a guideline, we are putting the wider applicability of the method to the test by using data not exclusively constructed for network analysis and focusing on core networks of respondents. Applying Random Forests promises more differentiated compositional typologies than currently used. Overcoming the prevailing approach to typologies by creating a larger number of finer-grained types can improve network-oriented life course analysis. We started our analysis by selecting 44 descriptors of 8341 respondent networks and reduced this number by combining most of them to indices representing basic dimensions of egocentric networks. The descriptors are used in hierarchical clustering to derive clusters that are then predicted with Random Forests to evaluate the reliability of this approach. We can identify stable core differentiators underlying typology construction but are also confronted with variation in the resulting types. While some of this can be traced back to the distinct socio-demographics of the respective populations, some is the result of measurement disparities. Against the backdrop of the presented results, we discuss the relationship between a deductive approach of network sociology, dedicated to theory-testing and inductive typologies of egocentric networks within a framework of life course analysis.