Human social networks are inherently multiplex, comprising overlapping layers of relationships. Different layers may have distinct structural properties and interpersonal dynamics, but also may interact to form complex interdependent pathways for social contagion. This poses a fundamental problem in understanding behavioral diffusion and in devising effective network-based interventions. Here, we introduce a new conceptualization of how much each network layer contributes to critical contagion pathways and quantify it using a novel metric, network torque. We exploit data regarding sociocentric maps of 110 rural Honduran communities using a battery of 11 name generators and an experiment involving an exogenous intervention. Using a novel statistical framework, we assess the extent to which specific network layers alter global connectivity and support the spread of three experimentally introduced health practices. The results show that specific relationship types - such as close friendships - particularly enable non-overlapping diffusion pathways, amplifying behavioral change at the village level. For instance, non-redundant pathways enabled by closest friends can increase the adoption of correct knowledge about feeding newborns inappropriate chupones and enhance attitudes regarding fathers' involvement in postpartum care. Non-overlapping multiplex social ties are relevant to social contagion and social coherence in traditionally organized social systems.