Abstract (EN)

A type of multilevel model, called a “multiple membership model”, can be specified and applied to situations where there is a single level of social network connections, and interest focuses in variations in an individual level response variable at the individual and network levels, possibly in the context of groups such as schools or areas (Tranmer, Steel and Browne, 2012). In this situation, the social network is seen as a dependency, and we may be interested in the extent to which people who know one another have similar values of their response variables. The data in this presentation comprise a set of French cancer researchers. We know which cancer researchers know one another. In addition, we know the way in which the laboratories in which the cancer researchers work are connected. Hence, we have two levels of network: networks of individuals, and networks of laboratories. In addition, we have information about the cancer researchers: age, speciality etc., and about the laboratories in which they work: size, funding etc.. For each of the cancer researchers, we have an individual level response variable at two time points: their “impact factor” scores as researchers. We explain how the multiple membership model may be extended to allow for two levels of network connection. Thus, we show how variations in the impact factor scores of the cancer researchers may be assessed at three levels: the individual level, the networks of individuals, and the networks of laboratories. We discuss the substantive issues in model selection and the interpretation of the model results for this example, and explain how the model framework developed and applied to the French data could be used more generally to study dependencies in an individual level variable at several different network levels.