This paper evaluates an almost universally distributed supplementary feeding program. The use of simple binary program may not reveal sufficient variation to identify program effect. So, taking advantage from detailed information on program implementation in the data set, this paper uses proportion of child's life exposed to the program to reveal variation in program intensity. This enables us to proceed further to deal with endogenous program placement: excluding the non-treated children and focusing estimation of program effect on treated children. The main findings follow. First, although the program was almost universally distributed, there was high variation in program intensity across communities. In addition, the distribution of program intensity appeared to be non-random as indicated by the importance of several observed community characteristics as well as regional unobserved heterogeneity. Second, program appeared to be effective in maintaining nutritional status of children including those with worst initial nutritional status. Its effectiveness however benefited only some segments of the group.