Developing mortality patterns: statistical and neural network approach

We study a possibility to assess the health status of the populations of the Federal Republic of Yugoslavia (FRY) through infant mortality rate. This time series has been internationally accepted as a sensitive indicator of social problems and socioeconomic development. In this respect this work can be thought of as a specific study of the influence of several factors, including war and United Nations economic sanctions on some aspects of system behavior, where the system includes the FRY as a whole along with its parts: Vojvodina, Central Serbia and Kosovo & Metohia. A quantitative measure of system behavior is based on mortality rate modeling by a set of different methods: from classical statistical techniques including general linear models, Box-Jenkins ARIMA to change detection analysis by modified generalized likelihood ratio and neural network based predictors. In order to measure the similarity between different regions of the FRY, we develop a new technique which includes a new neural network based similarity measure combined with classical multidimensional scaling. Neural models with exogenous variables representing investigating factors, showed best accuracy and reliable model structure for assessment of their significance and mutual relationship. Our methodology and quantitative results can serve as a base for further investigations of the most significant factors responsible for the health status of one nation, especially in transient economic and war environments.