Essays on assessing methods for modelling the distribution of healthcare costs

摘要: This thesis comprises three essays on assessing methods for modelling the distribution of healthcare costs. Chapter 2 extends literature cost data by applying generalised beta second kind (GB2) to English hospital inpatient data. A quasi-experimental design, estimating models a sub-population and evaluating performance another sub-population, is used compare this with its nested limiting cases. While, these data, (B2) gamma (GG) outperform GB2, our results illustrate that GB2 can be as device choosing among competing parametric distributions data. In Chapter 3, we conduct quasi-Monte Carlo comparison recent developments in semi-parametric regression costs, both against each other standard practice. The population NHS episodes financial year 2007-2008 (summed patient: 6,164,114 observations total) randomly divided into two equally sized sub-populations form an estimation set validation set. Evaluating out-of-sample using set, conditional density approximation estimator shows considerable promise forecasting means, performing best accuracy amongst four (of sixteen compared) bias goodness-of-fit. model linear square root transformed dependent variable, while link function Poisson performs terms Commonly utilising log are shown perform badly relative considered comparison. Chapter 4 examines full costs. Understanding generating process behind costs remains key empirical issue. Although much research date has focused prediction mean cost, potentially miss important features such tail probabilities. We experiment 14 approaches costs: nine which parametric, have commonly been fit five others designed specifically construct counterfactual distribution. Our indicate no one method clearly dominant there tradeoff between precision probability forecasts. find distributional demonstrate significant potential, particularly larger sample sizes where variability predictions reduced. Parametric log-normal, found estimate probabilities high precision, but varying depending upon threshold being considered.