It is common to define the structure of interactions among a population of agents using the social network metaphor. Most agent-based models were shown to be highly sensitive to that network, so the relevance of simulation results depends directly on the descriptive power of that network. When studying social dynamics in large populations, this network cannot be collected, and is rather generated by algorithms which aim to fit general properties of social networks. However, more precise data is available at a country scale in the form of socio-demographic studies, census or sociological studies. These “scattered statistics” provide rich information, especially on agents’ attributes, similar properties of tied agents and affiliations. In this paper, we propose a generic methodology to bring together these scattered statistics with Bayesian networks. We explain how to generate a population of heterogeneous agents, and how to create links by using both scattered statistics and knowledge on social selection processes. The methodology is illustrated by generating an interaction network for rural Kenya which includes family structure, colleagues and friendship. That network is constrained by data available from statistics and field studies.