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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Poster Presentations 97Monday, 25th August 2014 • 15:30-16:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust P1.2.106 Analysis of cluster randomised cross-over trials with binary outcomes KE Morgan1 , BC Kahan1 , RH Keogh2 1 Queen Mary University of London, London, United Kingdom, 2 London School of Hygiene & Tropical Medicine, London, United Kingdom   In cluster randomised cross-over (CRXO) trials clusters are randomised to receive treatments in a particular order. Data from a CRXO trial may have a complex dependency structure, with outcomes correlated within clusters and, potentially, additional correlation within treatment periods. This ad- ditional correlation arises when outcomes in one period are more similar to each other than they are to outcomes in another period in the same cluster. Assessing whether analysis methods adequately account for these correlations is important. We used simulation to compare methods of analysing a binary outcome, including logistic mixed effects models, generalised estimating equations (GEEs) and cluster-level summary methods. Models that ignore additional within-period correlation led to increased Type I error rates. An unweight- ed linear cluster-level summary regression gave nominal error rates in all scenarios, but lost power if extra correlation was present especially for small numbers of clusters. A mixed effects model with random effects for cluster and cluster-by-period interaction gave nominal error rates only in scenarios with large numbers of clusters, if there was additional correla- tion. GEEs did not give correct error rates in any scenarios considered. Assessing whether extra within-period correlation is possible is important when conducting a CRXO trial, and if it is this should be accounted for in both the sample size calculation and the analysis method. Unweighted cluster level summaries are the most robust method of analysis; however with a very large number of clusters a mixed effects model with random effects for cluster and cluster-by-period interaction could also be used.   P1.2.109 Optimal sampling times for pharmacokinetic modelling of a cocktail of phenotyping drugs TT Nguyen1 , H Bénech2 , M Delaforge2 , A Pruvost2 , F Mentré3 , N Lenuzza1 1 CEA, LIST, Data Analysis and Systems Intelligence Laboratory, Gif- sur-Yvette, France, 2 CEA, DSV, iBiTecS, Gif-sur-Yvette, France, 3 IAME, UMR 1137, INSERM - University Paris Diderot, Paris, France   “Cocktail”of drugs is of high interest to determine enzyme activity respon- sible for drug metabolism and pharmacokinetics. Phenotyping indexes can be derived from a few samples using nonlinear mixed effect models (NLMEM) for analyzing drug concentrations and maximum a posteriori estimation (MAP) of individual parameters. We proposed an informative design common to two molecules for a phenotyping study: midazolam (probe for CYP3A activity) and digoxin (P-glycoprotein). Using data of a previous study, NLMEM for midazolam, its 1-OH-metabolite and digoxin were developed in software MONOLIX4.2. Based on these models, we proposed a common design using a compound optimal- ity criterion which is a weighted sum of log determinants of population Fisher information matrix (FIM) for each compound. The resulting design was evaluated for MAP and predicted shrinkages were reported, based on Bayesian FIM, using R function PFIM4.0. Finally, sampling windows were computed around the optimal times, satisfying an expected joint loss of efficacy (evaluated by Monte-Carlo simulations) <5%. The common design was composed of six samples (0.25, 1, 2.5, 5, 12, 48h post-administration) instead of ten samples if considering separately each molecule. Predicted relative standard errors of derived phenotyping indexes were <30%, with shrinkages <40%. The sampling windows pro- vided more flexibility while maintaining 95%-efficacy, compared to the optimal design. By combining NLMEM, compound design and sampling windows based on FIM, we were able to determine sparse samples allowing correct es- timation of parameters for three compounds. This approach can be ex- tended to efficiently design studies with cocktails including more drugs. P1.2.112 Generalization of the big stick randomization rule to more than two treatment groups and unequal allocation rates P Ofner-Kopeinig1 , M Errath1 , A Berghold1 1 Inst Med Info, Stat & Docu, Medical University of Graz, Graz, Austria   The Big Stick randomization rule suggested by Soares and Wu provides good performance according to predictability and balance behavior. Treatments are allocated at random until a tolerance“a”is reached. In case “a”is reached, the under-represented treatment will be allocated next. The Big Stick procedure has so far only been described for two arm studies with equal allocation rates. The idea of using absolute differences to identify treatment imbalances does not work anymore if there are more than two treatments or unequal allocation rates. To allow for a generalization of the Big Stick randomiza- tion rule we developed a generalized measure of treatment imbalance which is based on the differences of observed and expected frequencies. This measure of imbalance takes into account the expected frequencies of the treatments at each randomization step and also treatment weights. Using this generalized measure of imbalance one can expand the Big Stick randomization procedure for unequal allocation rates and more than two treatment groups. To evaluate the generalized Big Stick randomization procedure a simula- tion study using the simulation tool of the “Randomizer for Clinical Trials” was performed. For an accuracy of 1 % and a confidence interval of 95 %, each simulation was performed 10000 times. Treatment imbalances were calculated to show the balance behaviour. To estimate the predictability the probability of correct guessing and the probability of deterministic allocation was calculated. The procedure was compared with complete randomization and permuted block randomization with different block lengths.   P1.2.114 The implications of differential clustering for the analysis of binary outcomes in cluster randomised trials N O’Leary1 , SA Roberts1 , C Roberts1 1 University of Manchester, Manchester, United Kingdom   Introduction: Cluster randomised trials usually assume homogeneity of the clustering effect. In cluster randomised trials of professional behaviour change interventions the clustering effect may differ between interven- tion arms. We examine the robustness of a range of standard techniques for the analysis of cluster randomised trials where the outcome is binary in the presence of between arm differences in the intra-cluster correlation coef- ficient (ICC). Methods: Binary data were simulated for a two-arm cluster randomised trial from a logistic-normal model with different values of the manifest ICC in each arm and a range of proportions. Analysis methods assessed were (1) an adjusted test of proportions, (2) a summary measures analyses, (3) a logistic generalised estimating equa- tion model with an exchangeable correlation structure, and logistic-nor- mal models with (4) a random intercept or (5) random coefficients for each arm.We assessed consistency, small sample bias of estimates and test size. Results: Treatment effect estimates for the random coefficient models were not consistent with estimates from other analyses; over-estimating the effect when proportions were low and under-estimating it when pro- portions were high.

Pages Overview