The clustering of childhood mortality risk by family in Northeast Brazil is investigated. The extent of family clustering is estimated before and after controlling for observed child and family demographic, reproductive, and socioeconomic characteristics and unobserved community characteristics. The study also investigates the extent to which the coefficient estimates and standard errors are altered with better controls for unobserved heterogeneity at the family level and at the community level. These controls are achieved through the use of a multilevel nested frailty model for survival data. The variance of family frailty is overstated by a factor of four when unobserved community effects are omitted. The family-level variance is not statistically significant in the multilevel model, although it is highly significant in the model that includes only the family frailty effect. Furthermore, ignoring the full design effects results in systematic bias in parameter estimates and standard errors.