Multilevel analysis of health and family planning data

Many populations in the social sciences have a hierarchical structure. For example, individuals are often nested within communities which themselves lie within larger units. In such cases, the characteristics of the context in which a person lives are likely to influence their behaviour and thus responses for individuals in the same community will tend to be correlated. Hierarchical data structures can also arise in longitudinal studies where observations over time are nested within an individual, and responses for the same individual may be correlated. One example of longitudinal data is an event history, where an individual is observed until the event of interest occurs or the observation period ends. Multilevel modelling techniques, which take into account these intra-unit correlations, have been developed to analyse hierarchical data. A multilevel approach can also be used as a convenient way of allowing for the effects of omitted covariates, or unobserved heterogeneity, in discrete-time event history models.In this thesis, multilevel modelling techniques are used to analyse a variety of hierarchical population structures in the areas of health and family planning. Four empirical studies are presented. In the first study, a multilevel multinomial model is used to analyse variations in contraceptive choice in Bangladesh between districts, and within districts between clusters. The analysis shows that a large proportion of the district-level variation in modern method use can be explained by differences in religious practice and literacy. Another study uses a two-level event history model to allow for unobserved heterogeneity in women's risks of contraceptive discontinuation in China. This is extended to a four-level model to analyse the extent of extravariation at the district, cluster and woman level in contraceptive method switching in Bangladesh. The results from these studies provide strong evidence of unobserved heterogeneity between women in contraceptive behaviour.Multilevel models are also applied in the area of child health to study immunisation uptake in rural Bangladesh. The results show that even after controlling for a range of child-, parental-, and household-level characteristics, there remains substantial variation in immunisation rates due to unobserved factors at the household and village level. (DX 192, 795).