Please activate JavaScript!
Please install Adobe Flash Player, click here for download

ISCB2014_abstract_book

38 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsMonday, 25th August 2014 • 16:00-17:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C18.4 Bayesian analysis of zero-inflated beta regression models with application to quality of life and functional outcomes L Sharples1 , C Jackson2 1 Clinical Trials Research Unit, Leeds, United Kingdom, 2 MRC Biostatistics Unit, Cambridge, United Kingdom Methods for zero-inflated Poisson and negative-binomial distributions are established and incorporated into standard software. Zero-inflated beta models have received less attention although they arise in several con- texts. For example, renal transplantation recipients listed for re-transplant can be classed as unreactive (frequency=0) or reactive to a proportion of the potential donor pool (frequency Є (0,1)). In clinical trials quality of life and utility are often measured on a scale bounded above and below, with a substantial proportion of patients exhibiting a floor (or ceiling) effect. In both these cases interest surrounds estimation of model parameters in- cluding patient and treatment effects. Although the Beta distribution is very flexible its coverage does not in- clude zero or one, and so it cannot be used to model zero-inflated data.We show how the response can be modelled using a mixture in which (i) the non-zero responses arise from a (suitably-parameterised) Beta distribu- tion, (ii) zero observations are Bernoulli random variables with probability p. Covariates can be included for both the probability of a zero response and the level of (non-zero) response, using regression. Multiple responses per individual can be incorporated using random effects. MCMC imple- mentation is straightforward and flexible enough to accommodate miss- ing data under a missing at random assumption. However, the Bayesian paradigm requires careful specification of priors. Competing models can be compared using the DIC and goodness of fit assessed by comparing observations with model-predictions. We discuss these issues through applications to renal transplantation and quality of life outcomes in RCTs. C18.5 Count data analysis in nutrition clinical trials Y Yavuz1 , S Swinkels1 1 Danone Nutricia Research, Biometrics, Utrecht, The Netherlands In Early Life Nutrition division of Danone Nutricia Research, the analysis of count data such as the number of infections, number of hospital visits, number of doctor-diagnosed diarrhea, is of interest in many clinical trials. Poisson regression models would be the simplest standard framework for the analysis of such data. However, in real life, count data do not always meet the assumption of equal variance-mean relationship induced from Poisson distribution lead- ing to over-(or under-)dispersion. The source of the over- (or under-)dis- persion could be due to a higher than expected occurrence of zero counts. A toddler may have no infection either because of his/her resistance to the infection, or simply because no disease spores have landed on him/her. This is the distinction between structural zeros, which are inevitable, and sampling zeros, which occur by chance. Another source of over-dispersion might be the fact that having an infection might make individuals more vulnerable for a second one. We demonstrate the use of four different models for over-dispersed count data with certain levels of zero inflation: Poisson, negative binomial, zero- inflated Poisson and zero-inflated negative binomial models. We discuss the performance of the models using data from a clinical trial on the num- ber of infections in toddlers during a 12 month period.

Pages Overview