Bayesian analysis of a patient dataset using prior information from normal volunteers and from another patient group
I. Knutsson (1), L. Aarons (2), S.Callies (3)
(1); School of Pharmacy, University of Manchester (2); Ely Lilly (3)
In drug development, researchers have been using available information for designing trials, analysing data and, ultimately, making informed decisions. In particular it is desirable to use this information to maximise the gain from newly obtained data since the sampling frequency of PK/PD responses often lessens as the development progresses.
In this talk, we address common issues regarding Bayesian analysis of a patient dataset where relevant prior information is mainly available from normal volunteers and from patients belonging to another subpopulation. A simulated patient PK dataset is generated for a primarily-renally-excreted antibiotic based on relevant population PK analysis results that include patients with degrees of renal impairments. The knowledge gathered from PK analyses of normal volunteers and from another patient group, coupled with currently available methods to predict patients’ renal function is utilised to aid the aforementioned data analysis.
First, the set-ups of numerical priors and population PK models for the analysis are defined to avoid retrospectional bias, while further sets of priors and/or models can be cautiously considered at a later time. The suitability of the analysis itself is then assessed in respect of model-based exchangeability; the assessment is also taken after the analysis. Additionally, we assume that a separate frequentist analysis of the patient dataset is already completed and that the results are available at this stage. A number of analyses of the dataset are then carried out using WinBUGS, adopting a vague prior and informative priors in conjunction with models that together take account of potential heteroscedasticity across the populations. The models are able to allow and/or detect potential differences across the populations so that, in conjunction with other related information, further systematic improvements in the modelling can be achieved. The performance of each analysis is examined in three ways: via a posterior predictive check that examines the general quality of the fit; via conditional predictive ordinates that aim to look at details of the fit; and via common assessment plots. Comparisons across these analyses are also made via plots. Finally, potential applications of the results as a whole are discussed, which focus especially on consequences related to the PD and on whether or not, and how, to address questions arisen from the analyses.
In conclusion, Bayesian analyses using a set of priors with varying degrees of informativeness and of models allowing moderate dissimilarity between different populations enable us to project truer state of drug investigation at any given point of time in a quantitative unified manner.