Group sequential and adaptive group sequential tests in sensory and consumer studies

Abstract

Sequential methods have been suggested in sensory and consumer science before but have seen limited applications so far. This can be attributed to the fact that only sequential probability ratio tests have been described in detail, which are impractical as data are rarely collected and analyzed one data point at a time. As an alternative, group sequential methods collect and analyze data in batches, which better reflects common practice in consumer and sensory studies. Scarcer budgets require more efficient use of resources, and (adaptive) group sequential methods can play a significant role to achieve this. An abundance of related methods has been proposed primarily in the biometrical area, ranging from equal group sizes over different group sizes to adaptive designs which allow for modification of sample sizes or other aspects for subsequent study stages based on interim analysis results. Also, early stopping for futility is an option in case initial data are not promising; it is particularly useful when little is known about the (relative) performance of products upon planning the study.

We briefly review key concepts and methods, and we illustrate these for the simple but important example of paired preference tests with the objective to show superiority. Other applications in sensory and consumer research like, e.g., sequential monadic designs can follow the same concepts but might be technically more complex. Entertaining an interim analysis generally comes at the cost of lower power or increased maximum sample size. The benefit lies in the potentially significant savings when early data are clear enough either to support the desired outcome of the study, or to indicate that there is little hope for the study to succeed. Adaptive designs also offer the opportunity to fine-tune the final sample size based on early results, such as to limit the effort to what is really needed, or to expand on the effort to save a study from likely failure. We show how and why the different approaches are technically robust and what benefits they can bring to consumer and sensory studies.