|
Abstract:
|
A common neuroscience topic is to detect contrast between two stimuli, and is often studied via a logistic model called a psychometric function. These studies are often interested in making inferences at the group level (age, gender, etc.) and at an individual level. Conventional practice is to use simple models that are easy to fit, but inflexible and vulnerable to fitting issues in the situation of complete separation. Bayesian multilevel models are flexible, efficient, and easy to interpret, yet are not broadly adopted due to unfamiliarity among practitioners. We describe a model selection process in a visual workflow, including specifying priors and implementing adaptive pooling. Then we propose and develop specialized quantities of interest (such as the point of subjective simultaneity) and study their operating characteristics. In the development of our model we conduct simulation studies into these proposed quantities of interest that provide insights into experimental design considerations. We discuss in detail a case study of real and previously unpublished data. Finally we provide a user-friendly R package for fitting multilevel psychometric functions.
|
