Objective: To test the association between several social networks variables reflecting both structural characteristics and quality of relationships with the risk of dementia and Alzheimer's disease 5 and up to 15 years later. Methods: The study sample is gathered from the Paquid cohort, a French population-based study of 3,777 elderly people evaluated at baseline and regularly revisited during a 15-year interval. The sample consisted of 2,089 subjects who completed the social network questionnaire and were free of dementia at the time of enrollment and also at the next two follow-ups to minimize the problem of reverse causality. The questionnaire collected at baseline included marital status, number of ties, nature of social network, satisfaction, perception of being understood/misunderstood, and reciprocity in relationships. Results: The incident cases of dementia considered were those diagnosed at 5-year and subsequent follow-ups, resulting in 461 dementia and 373 Alzheimer's disease cases. The multivariate Cox model, including the six social network variables and adjusted for numerous potential confounders, showed significant associations with satisfaction and reciprocity in relationships. Participants who felt satisfied with their relations had a 23% reduced dementia risk. Participants who reported that they received more support than they gave over their lifetime had a 55% and 53% reduced risk for dementia and Alzheimer's disease, respectively. Conclusion: The only variables associated with subsequent dementia or Alzheimer's disease were those reflecting the quality of relationships. The delay between social network assessment and dementia diagnosis was from 5 up to 15 years, thus minimizing the problem of reverse causality.