DescriptionThe purpose of this dissertation is to analyze three models in medicine and finance using Bayesian inference with the Markov chain Monte Carlo method. The model in medicine addresses cost-effectiveness analysis using copulas, and the two models in finance include discrete-time asset pricing models and a short-term interest rate model with stochastic volatility.
The first chapter develops the model that allows dependence between cost and effectiveness using copulas in cost-effectiveness analysis. The model was applied with sample of adults from the NHANES I Epidemiologic Follow-up Study, assuming a lognormal distribution for cost and a Weibull distribution for effectiveness as the marginals. Cost-effectiveness analysis is conducted for two types of patients using the estimated posterior densities of parameters regarding the hypothetical intervention for hypertension. A simulation based on Bayesian predictive densities is also performed to analyze cost and effectiveness at an individual patient level. The empirical result indicated a negative dependence between measures of effectiveness and cost.
The second chapter conducts a Bayesian analysis of discrete-time asset pricing model. The chapter particularly discusses the naive discretization problem, which arises from using discrete-time data to estimate continuous-time models. Our results using generated data showed that the naive discretization would not work well when data generating process is unknown, when the data is sampled at low frequency, and averaged data is used.
The final chapter develops a Bayesian analysis of a short-term interest rate model with stochastic volatility. The model was developed based on the CKLS model (Chan et al. 1992). We constructed MCMC algorithms suitable for the model based on the Jacquire, Polson and Rossi(1994) algorithm. The empirical results with the 3-month Treasury constant maturity rate suggested that there was high autocorrelation in volatility of the error terms. Finally, the developed model was compared with the model with a GARCH error, using Bayesian predictive densities. The predictive densities obtained by CKLS with stochastic volatility have wider variance than the ones from CKLS-GARCH,
and the realized value did not fall in the support of the predicted values for the CKLS GARCH model because of the tight variance in prediction.