Model Choice in Understanding the Changing Covariates of Infant Mortality: An Illustration Using Kenya Data

Sam Wafula, Population Council

Covariate estimates of infant mortality is fraught with measurement and choice issues. Standard regression models fail to account for censoring while survival models such as Cox regression may assume the proportionality assumtpion and fail to test for it. Statisticians are also divided on which parametric regression model to use when the proportionality assumption is violated. Some advocate that method choice should be guided by the level of log likelihood in a nested model after the Akaike Information Criteria (AIC). Others argue that method choice should be based purely on theory of the event under study. This study uses the merged Kenya Demographic and Health survey of 2003-2008 to shed light on the best method choice that can yield robust estimates using infant mortality as an illustration. Findings show that that theory driven model selection produces better estimates than models based on the lowest values of the AIC log likelihood estimates.

  See extended abstract

Presented in Poster Session 7