An inductive typology of egocentric networks with data from the SOEP
Bastian Laier, Marina HennigIn the social sciences, we are often interested in typologies to reduce the complexity of social reality. Typologies are also useful when it comes to communicating knowledge to the public, as they often represent archetypes that are easily understood by everyone.
We apply Random Forests to data from egocentric networks from a large German survey, the Socioeconomic Panel (SOEP). The application of the Random Forests method to derive an inductive typology of egocentric networks was a direct response to the paper from Gianella and Fischer (2016) in which they were the first to present a detailed step-by-step application of machine learning algorithms to egocentric network data. Our goal was to validate the approach of Gianella and Fischer (2016) with a different sample and slightly different variables to derive a typology of egocentric networks.
We started with 44 descriptors of 8341 respondents and reduced this number by combining part of them to seven indeces and five single items representing the basic dimensions of egocentric networks. Since many of the descriptors of egocentric networks in the SOEP are at least to some extent comparable to those of Gianella and Fischer (2016), we found similar dimensions: (1) support from kin, (2) support from nonkin, (3) support from coworkers, (4) index of qualitative variation within respondent’s networks, (5) social activities, (6) cultural activities, and (7) membership in various organizations. The single items consist of (I) network size, (II) proportions of kin living abroad and (III) kin living close to the respondents, (IV) the proportion of nonkin in networks as well as (V) frequency of church visits. These descriptors are then used in hierarchical clustering to derive clusters that can be predicted with Random Forests to evaluate the reliability of this approach.
Although the underlying variables are not directly comparable with those of Gianella and Fischer (2016), we obtain qualitatively similar clusters at the end of our analysis. Furthermore, there are parallels to the basic dimensions of egocentric networks, which we would like to discuss.