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


ISCB 2014 Vienna, Austria • Abstracts - Poster Presentations 89Monday, 25th August 2014 • 15:30-16:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Monday, 25th August 2014 - 15:30-16:00 Poster session P1 P1.1 Bayesian methods in biostatistics P1.1.1 Coverage properties of Bayesian 95% probability intervals for odds ratio and relative risk S Aghlmandi1 , M Zwahlen1 1 Institute of Social and Preventive Medicine (ISPM), Bern, Switzerland   Fagerland & Newcombe1 recently presented different methods to calcu- late 95% confidence intervals for the odds ratio (OR) and relative risk (RR) and advocated the use of the inverse hyperbolic sine interval transforma- tion while adding different pseudo-frequencies.We assessed the coverage probability of Bayesian 95% probability intervals for the same situation us- ing the 2.5% and 97.5% quantiles of the posterior distribution. We defined pi as the probability of an event occuring in group i (i=1, 2), ni+ the total number of patients and ni1 the number of events in group i. For the Bayesian analysis, we used beta priors for pi which allow a beta- Binomial conjugate analysis for the pi . We used Monte-Carlo simulations to generate a 105 sample from the pos- terior distributions of p1 and p2 to then approximate the posterior distribu- tion of the OR and RR.We used the same simulation scenarios as Fagerland & Newcombe1 : n1+ =n2+ =20, true OR=2.41, RR=1.43, p1 ranging from 0.01 to 0.99 in steps of 0.01 with 5000 data sets created for each p1 . Finally, we compared the coverage probability of Bayesian 95% probability inter- vals with the coverage Fagerland & Newcombe1 obtained with their“best” methods, the inverse sinh formula with and without pseudo-frequencies. Results for the OR show that the Bayesian 95% probability intervals have a mean coverage near 95%, very similar to Fagerland’s best results. Both approaches have a coverage above 95% when p1 <10% or >90%. We ob- tained similar results for RR. Reference:1-Fagerland&Newcombe,StatisticsinMedicine2013;32:2823- 2836.   P1.1.17 Comparison of two methods for futility analysis in vaccine efficacy trials FP Bailleux1 , E Bassily2 , AJ Dunning2 1 Sanofi Pasteur, Marcy l’Étoile, France, 2 Sanofi Pasteur, Swiftwater, United States   Aims: Vaccine efficacy (VE) trials involve a large number of subjects and considerable costs/resources. If the VE is not promising it may be benefi- cial to stop the trial. Methods: The statistic to evaluate the VE in the interim or final analyses is derived from the number of cases in vaccine group, which follows a bino- mial distribution conditional on the total number of cases. Two methods for futility analysis were evaluated. The first method tests the null hypothesis that the VE is not greater than a predefined bound. This test is evaluated using the upper bound of confi- dence interval of VE at a fixed alpha level (independent of the alpha level used at the different interim efficacy analyses). The second method is based on the Bayesian predictive power. At each futility analysis the Bayesian predictive probability to conclude at the end of trial is calculated, if this probability is lower than a cut-off then the trial stopped for futility. Results: Methods are compared using different trials design with various theoretical vaccine efficacies, lower bounds and with a limited number of futility/interim analyses. While the Bayesian predictive power permits stopping earlier in case of poor VE, the method based on the upper bound of CI of VE permits better control of the risk of stopping for futility in case of good VE. Conclusion: Various methods exist to perform futility analyses in vaccine VE. Choice of the method is crucial to insure an optimization of the risk the sponsor will want to control. P1.1.20 Choosing a gold standard: support of Bayesian inference methods for diagnostic accuracy of new biomarkers in pediatric urinary tract infection S Bastide1,2 , P Landais1,2 , S Leroy1,2 1 Department of Biostatistics and Epidemiology, Nîmes Hospital, Nîmes, France, 2 EA2415 Biostatistics Research Unit, Montpellier 1 University, Montpellier, France   Background: Acute pyelonephritis (APN - kidney infection) is a common pediatric bacterial infection. Biomarkers-based strategies (e.g. procalcito- nin) aimed at promptly diagnosing APN, and were compared with DMSA scan, whose gold-standard quality raises concerns. We used for the first time Bayesian methods to estimate the diagnostic accuracy of procalci- tonin and DMSA. Methods: We used a Bayesian approach to explore disease prevalence and tests properties, using an independent model and both fixed and random effects models with conditional dependence between tests. Two levels of prior distribution were defined: one informative obtained from a published meta-analysis of individual patient data (1011 patients, 61% APN) and pediatrician beliefs for DMSA, and one non-informative. Standard procedures were used to achieve MCMC convergence, for model checking and for a sensitivity analysis. Results: The fixed model yielded for procalcitonin: Se 74% [71-77], Sp 70% [66-74], and for DMSA: Se 94% [87-98], Sp 90% [80-97] with informative prior; whereas, with the non-informative prior, it achieved: Se 72% [59- 90], Sp 75% [53-94] for procalcitonin, and for DMSA: Se 77% [64-92], Sp 74% [50-94] . Given the important amount of the additional information contained in prior samples, the non-informative prior seemed sounder. The same discordance between the priors was similarly observed with the independent model. A random effect model was completed to further ex- plore this result. Conclusion: A Bayesian approach allowed showing that the gold-stan- dard test for APN, DMSA, was not perfect despite clinical beliefs. Support of Bayesian inference methods for diagnostic accuracy of new biomarkers should be fostered. P1.1.21 Methodological review of Bayesian inference methods used in clinical decision rules and diagnosis studies S Bastide1,2 , P Landais1,2 , S Leroy1,2 1 Department of Biostatistics and Epidemiology, Nîmes Hospital, Nîmes, France, 2 EA2415 Biostatistics Research Unit, Montpellier 1 University, Montpellier, France   Background: In many clinical conditions, the existence of a gold standard is often disputable, or not available for all patients. The introduction of Bayesian methods appeared valuable in situations where other approach- es failed. We aimed at studying the use of Bayesian inference methods dedicated to diagnosis decision making based on clinical decision rules (CDR) and diagnosis studies. Methods: We searched for all articles using Bayesian methods for diagno- sis and/or CDR studies in electronic databases using key words in MEDLINE and citations search of the cornerstone papers in ISI Web of Science. Eligibility was assessed on title/abstract, and finally inclusion on full-text.

Pages Overview