| In this dissertation, A bayesian approach was used to model multinomial ordinal outcomes. Responses without and with misclassification were considered. Transition probability with possible dependence on covariates and misclassification probability without covariate were modeled by multinomial logit models. Metropolis-Hasting algorithm was implemented to generate samples from posterior distribution, and then to get posterior estimates, standard deviations and confidence intervals. First models were tested by simulated data, posterior estimates were compared to the original values used to simulate data, and Deviance Information Criterion (DIC) was calculated and compared for goodness of fit. Lastly, multinomial logit models with race and age covariates were applied to the National Heart, Lung, and Blood Institute Growth and health Study (NGHS). SAS program was written mainly in PROC IML for Metropolis-Hasting algorithm and DIC calculation. |
