Development practitioners have long advocated for targeted beneficiary communities to participate in decision-making about how to provide local public goods and services. However, community participation in decision-making may have disadvantages as well as benefits, and the balance between the benefits and disadvantages may vary across different decisionmaking processes, interventions, and contexts. Previous studies of community participation in decision-making indeed report mixed results, but the underlying reasons for this remain uncertain. This study demonstrates that context matters. We investigate the heterogeneity of impacts in a field experiment conducted in two regions of rural Bangladesh. The experiment randomly assigns different decision-making processes to villages that receive otherwise identical interventions in the form of a program to increase access to safe drinking water. We show that a deliberative, consensus-based approach to community participation in decisionmaking has strongly heterogeneous effects compared to either a top-down approach or community decision-making without rules about how decisions are made. The consensusbased process doubles the program's impact in one region but barely increases it in the other. We use machine learning to identify the baseline characteristics and mechanisms that correlate most strongly with impact. The results suggest that the consensus-based process yields better outcomes when there is more at stake: specifically, when the community has fewer pre-existing safe sources of drinking water and thus more to gain from the intervention. The results are consistent with the view that inclusive participatory approaches to decisionmaking can increase program impacts but fully engaging in these processes is costly. When less is at stake, communities may not fully engage, and the advantages of inclusive participatory decision-making may not be realized.