Application Research On Multilevel Multinomial Logistic Model For Unordered Categorical Responses

Objective: This study is to explore the application of multilevelstatistical model that is to analyze the hierarchical structure data．The datafrom the family health questionnaire of health service survey in WesternChina in2008were descriptively analyzed with the multilevel multinomiallogistic model for unordered categorical responses. This study is toinvestigate the methodology of this model in practical application and toprovide reference for further studies．Methods: Outpatients’ choice of selecting care provider was recruitedas the response variable, including private practices, village clinics andtown hospitals plus level. Age, gender, educational level, marital status,employment status, economic status, medical insurance and health status ofthe interviewees (Chongqing rural patients,15years old or above) were theexplain variables. Based on the results of descriptive study and univariatestudy, town level and individual level were considered to fit the two-levelmultinomial logistic model for unordered categorical responses. Solutionsof fitting problems were presented, including the numbers of the response categories, the reference category’s selection, the set and introduction ofindependent variables, the judgment of data hierarchy, parameter estimationand hypothesis testing methods, trade-offs of the model’s fitting goodnessand complexity, the inspection of the multinomial distribution, etc..Results: There was hierarchy in choices of selecting care provider ofChongqing rural patients by the two-level multinomial logistic model forunordered categorical responses. They had significance among differenttowns．Clustering existed at the level2and selecting multilevel statisticalmodel was appropriate. The results showed that economic status, medicalinsurance and health status were the main influencing factors. This modelconsidering the hierarchical structure and clustering could analyze the fixedeffects and random effects availably and estimate parameters accurately.DIC of the2-level model with explanatory variables was43.622lower thanthe zero model’s, so the fitting goodness raised. We considered thehypothesis of the multinomial distribution was correct with both reasons(the same fixed coefficients and similar random coefficients in our analysesand the results of Yang’s simulation study).Conclusions: The multilevel multinomial logistic model for unorderedcategorical responses was adopted in the field of health service in this studyto analyze the influencing factors of rural residents’ choices of selectingcare providers. The fitting of the model was satisfying. Multilevel statisticalmodel is a good choice in the study of health service. Being a new access to analyze the hierarchical structure data, this method is moreapplicable．Multilevel statistical model will be adopted in more fields withthe development of the theory and MLwiN. Furthermore, the model test,diagnosis and residual analysis will be explored more deeply.