Bayesian estimation of cost‐effectiveness from censored data

Abstract

Cost-effectiveness models are commonly utilized to determine the combined clinical and economic impact of one treatment compared to another. However, most methods for sample size determination of cost-effectiveness studies assume fully observed costs and effectiveness outcomes, which presents challenges for survival-based studies in which censoring exists. We propose a Bayesian method for the design and analysis of cost-effectiveness data in which costs and effectiveness may be censored, and the sample size is approximated for both power and assurance. We explore two parametric models and demonstrate the flexibility of the approach to accommodate a variety of modifications to study assumptions.

Cost-effectiveness in clinical trials: using multiple imputation to deal with incomplete cost data

“…In this example, all patients had died so the survival times were known for all patients and all costs had been incurred, i.e. uncensored, however, specific methods for dealing with censored survival and cost data are available [21].This paper demonstrates that extending costeffectiveness analysis to MI data is reasonably straightforward and allows the known effectiveness data in the larger sample to be included. Further development work looking into the inclusion of all trial patients (that is, to include patients who were not part of the cost sub-study) in a cost-effectiveness analysis could also be considered.…”

Addressing the issues that arise in analysing multicentre cost data, with application to a multinational study

“…For example, drug costs set by pharmaceutical companies might have additive rather than multiplicative effects, and models with some covariates acting additively and some acting multiplicatively might have to be considered (Basu and Rathouz, 2005). The gamma distribution has commonly been used to model cost and resource use data (Austin et al, 2003;Heitjan et al, 2004;Lee et al, 2005), because it is a simple two-parameter family of positive-valued skewed distributions. In empirical studies it can often fit cost data quite well (Basu et al, 2004;Barber and Thompson, 2004;Grieve et al, 2005;Dodd et al, 2006), but may not always be appropriate (Kilian et al, 2002).…”

“…The distributions of the cost data were positively skewed in almost all the centres. Because cost data are skewed, distributions other than the normal are usually more appropriate (Austin et al, 2003;Heitjan et al, 2004;Lee et al, 2005). The gamma and log-normal distributions are often applied to cost data for this reason, and may yield different conclusions than when assuming a normal distribution (O'Hagan and Stevens, 2003;Nixon and Thompson, 2004).…”

Bayesian statistics in medicine: a 25 year review

“…Survival data with a non-susceptible fraction and dual censoring mechanisms were used in Huntingdon's disease, where not everyone inherits the gene [516] Discrete time survival modelling was used for registry data on haemodialysis patients [517]. Estimation of costeffectiveness from censored data was applied in cardiovascular disease [518]. A neural-Bayesian approach was used to improve predictions from Cox survival analysis [519] The clinical course of bone metastases in women with advanced breast cancer was investigated using models for multivariate interval censored recurrent events [520].…”