THE ORDINAL LOGISTIC REGRESSION MODEL WITH SAMPLING WEIGHTS ON DATA FROM THE NATIONAL SOCIO-ECONOMIC SURVEY

Abstract

Ordinal logistic regression is a method describing the relationship between an ordered categorical response variable and one or more explanatory variables. The parameter estimation of this model uses the maximum likelihood estimation having assumption that each sample unit having an equal chance of being selected, or using simple random sampling (SRS) design. This study uses data from the National Socio-Economic Survey (SUSENAS) having two-stage one-phase sampling (not SRS). So, the parameter estimation should consider the sampling weights. This study describes the parameter estimation of the ordinal logistic regression with sampling weight using the pseudo maximum likelihood method, especially in SUSENAS sampling design framework. The variance estimation method uses Taylor linearization. This study also provides numerical examples using ordinal logistic regression with sampling weight. Data used is 121,961 elderly spread over 514 districts/cities. Testing data (20%) is used to obtain the accuracy of the prediction results. The variables used in this study are the health status of the elderly as the response variable, and nine explanatory variables. The results of this study indicate that the ordinal logistic regression model with sampling weights is more representative of the population and more capable to predict minority categories of the response variable (poor and moderate health status) than is without sampling weights.