Using longitudinal data of the Health and Demographic Surveillance System (HDSS) in Matlab, Bangladesh, covering the time period 1982–2005, and exploiting dynamic panel data models, we analyze siblings’ death at infancy, controlling for unobserved heterogeneity and a causal effect of death of one child on survival chances of the next child. Matlab is a rural area split into two: a “treatment” area where along with standard government services extensive maternal and child health interventions are available, and a “comparison” area where only the standard government services are available. The observed infant mortality rates are 50 per 1,000 live births in the treatment area and 67.4/1,000 in the comparison area. We use separate models for the two areas and analyze the differences in infant mortality between the two areas using several decompositions. Our model predicts that in the comparison area, the likelihood of infant death is about 30% larger if the previous sibling died at infancy than if it did not, and the estimates suggest that, in the absence of this “scarring” effect, the infant mortality rate among the second and higher order births would fall by 6.2%. There is no evidence of such a scarring effect in the treatment area, perhaps because learning effects play a larger role with the available extensive health interventions. We find that distance to the nearest health clinic can explain a substantial part of the gap in infant mortality between the two areas.