A comparison of some methods for detection of safety signals in randomised controlled trials

Carragher, Raymond (2015) A comparison of some methods for detection of safety signals in randomised controlled trials. In: Society for Clinical Trials 36th Annual Meeting, 2015-05-17 - 2016-03-20, Virginia. (Unpublished)

[thumbnail of Carragher-SCT-2015-methods-for-detection-of-safety-signals-in-randomised-controlled-trials]
Preview
Text. Filename: Carragher_SCT_2015_methods_for_detection_of_safety_signals_in_randomised_controlled_trials.pdf
Accepted Author Manuscript

Download (197kB)| Preview
[thumbnail of Carragher-SCT-2015-methods-for-detection-of-safety-signals]
Preview
Text. Filename: Carragher_SCT_2015_methods_for_detection_of_safety_signals.pdf
Download (239kB)| Preview

Abstract

The occurrence, severity and duration of patient adverse events are routinely recorded during randomised controlled clinical trials. This data may be used by a trial’s Data Safety Monitoring Committee to make decisions regarding the safety of treatments and in some cases may lead to the discontinuation of a trial if real safety issues are detected. Consequently the analysis of this data is a very important part of the conduct of any trial. There are many different types of adverse event and the statistical analysis of this data must take into account multiple comparison issues when performing statistical tests. Unadjusted tests may lead to large numbers of false positive results, but simple adjustments are generally too conservative and risk compromising the power to detect important treatment differences. Mathematically there are a number of different approaches to analysing safety data with general error controlling procedures, recurrent event analysis, survival analysis and other direct modelling approaches (both Bayesian and Frequentist) all being used. Recently a variety of classical (Mehrotra and Adewale, 2012) and Bayesian (Berry and Berry, 2004; DuMouchel, 2010) methods have been proposed to address this problem. These methods use possible relationships or groupings of the adverse events. We implement and compare by way of a simulation study of grouped data some of these more recent approaches to adverse event analysis and investigate if the use of a common underlying model which involves groupings of adverse events by body-system or System Organ Class is useful in detecting adverse events associated with treatments. All of the group methods detect more correct significant effects than the Benjamini-Hochberg or Bonferroni procedures for this type of data. In particular the body-system as described by Berry and Berry (2004) looks to be a worthwhile structure to consider for use when modelling adverse event data.