Controlling for Attrition Bias in Child Mortality Analysis Using a Two-Stage Semi-Parametric Proportional Hazard Model

Philippe Bocquier, University of the Witwatersrand

All event history analyses make the explicit assumption of independence between censoring and event. If not, then the results suffer from potential bias. This paper presents a way to deal with non-independent censoring. We follow the rationale of two-stage regression models controlling for endogeneity. First, we model the attrition risk using available independent variables, including an instrumental variable. Second, we derive an individual, time-dependent propensity for attrition. And last, we insert this propensity in the main equation, as a time-varying variable. The model is illustrated with longitudinal data on child mortality in two African sites – one situated in Nairobi slums, the other in a South African rural area – where migration is high. The results show high attrition effect and mortality estimates that are more consistent with the poor living conditions in the study areas. This confirms that child migration is often a response to health hazards.