Do previous birth interval and mother's education influence infant survival?: a Bayesian model averaging analysis of Chinese data

We examine the effect of socio-economic covariates on infant mortality in China in the 1980s, particularly the role of previous birth interval and mother's education, using an event history approach with data from the 1988 Two per Thousand Fertility Survey. We use a Bayesian model averaging strategy that takes account of model uncertainty as well as parameter uncertainty. A standard stepwise logistic regression analysis finds no statistically significant relationship between the preceding birth interval and infant survival after controlling for socio-demographic factors, but this finding is reversed when the Bayesian model averaging approach is adopted. However, the method finds less support than a standard stepwise approach for the role of mother's education. We consider the model- fitting criterion of predictive power when applied to out-of- sample observations, and show that Bayesian model averaging outperforms the stepwise approach. We conclude that, even with large sample sizes, the interpretation of results can vary substantially according to model selection and fitting criteria.