In the developing world, breastfeeding is a major determinant of infant morbidity, infant mortality an birth spacing. In the literature on the determinants of breastfeeding duration and the effect of breastfeeding on infant mortality, data quality is a constant concern. Much of the concern is stimulated by the clear evidence of heaping in retrospective breastfeeding data. Retrospective responses show pronounced spikes at 12, 24, and 36 months. Since equivalent drops in the hazard are not present in current status data, the conventional wisdom is that the heaping is a form of recall bias. The existing literature contains several ad hoc approaches to these apparent data quality problems. Work by Palloni suggests that the magnitude of the estimated relations is sensitive to the treatment of the heaped responses. We propose to use the interview-reinterview component of the Malaysian Family Life Surveys (MFLS) to jointly estimate the determinants of breastfeeding durations, the effect of breastfeeding on infant mortality, and the process determining the recorded duration of breastfeeding and age at child death conditional on the true duration/age. In 1976, MFLS-1 collected retrospective breastfeeding durations for about 5,000 children. Four months later, a second retrospective response was collected for about 1,000 of those children. In 1988 (twelve years later), MFLS-2 collected retrospective breastfeeding durations on about 10,000 children including most of the 5,000 collected in 1976. This interview-reinterview data provides a unique opportunity to evaluate the quality of recall data over both short and long intervals and its effects on estimated behavioral relations. Exploiting these rich data, we will develop and implement a new statistical model which jointly estimates the determinants of the true behavior and the determinants of the response conditional on the true behavior. This work extends preliminary work on breastfeeding duration which found that explicitly modeling the recall process increased regression coefficients by about fifty percent. Furthermore, the estimated time trends change from U- shaped when ignoring the recall bias; to monotone increasing when the recall process is modeled. The proposed project will extend these preliminary results in several ways. First, it will specify and estimate a richer model of the true behavior and conditional response. Second, it will explore the efficacy of these models using interview-reinterview data of the type which could be collected in a post-enumeration survey (e.g. several months later). Third, it will extend the results from breastfeeding, to amenorrhea, and infant mortality. Finally, the separate models for breastfeeding and infant feeding will be merged to estimate the effect of breastfeeding on infant mortality, recognizing that all of the responses are subject to recall bias.
