In cohort studies, researchers usually recruit participants randomly from general population to make sure of the generalisability of their research findings. In prospective breastfeeding studies, pregnant women are normally invited to participate in studies and information is collected since pregnancy up to early infancy. One of the breastfeeding outcomes generally measured is breastfeeding duration. However, this outcome measure is conditional on that mothers initiate breastfeeding after delivery as some mothers do not initiate breastfeeding after delivery (feeding babies with infant formula instead). As such, in the consequent statistical analysis, selection bias would be introduced, which may lead to spurious results. The aim of this project is to examine the structure of the selection bias using causal diagrams and investigate the solution to correcting the selection bias quantitatively with future applications in other similar settings.