Bayesian models for small area estimation based on unequal probability sampling of binomial and multinomial responses

Abstract

In this research a Bayesian Method of Small Area Estimation (SAE) has been developed based on binomial and multinomial response variables using Susenas data obtained from unequal probability sampling. Case study was carried out to predict education level of the population measured by literacy rate and mean years of schooling in sub-district level in East Java Province. The SAE model for binomial response was developed with two methods, i.e. using weighted logit normal mixed model and involving the probability of sampling selection model as exponential function into the SAE model. A simulation study was carried out by implementing 100 times sampling selection into population data. Penalized Quasi Likelihood (PQL) and Restricted Maximum Likelihood method (REML) was used to parameter estimation of SAE model. Based on the simulation result, we found that the weighted logit normal mixed model gave the best estimate. In application, the weighted logit normal mixed model also provided good prediction of literacy rate in Sumenep and Pasuruan regency.For the multinomial respons, we applied the weighted logit multinomial mixed model. MSE estimation was used Jackknife method and it gave very small MSE of about 1,14 x10-7.