Cognitive scientists use mathematical models to describe behavior. These models can be fit to behavioral data to make inference about latent variables. Normally, multiple subjects perform the same experiment and Maximum Likelihood (ML) is used to find the set of parameters that best describe the observed data of each individual subject. However, this method has several disadvantages (discussed in more detail below). One major drawback is that the estimation does not take into account that there are likely to be a similarities between subjects and what we learn about one we can apply to others. One alternative is to fit one model to all subjects however ignoring individual differences. Ideally, we want a method that allows us to portray the structure of our data as closely as possible onto the model.
The Solution
Hierarchical Bayesian modeling offers a remedy to this problem: Each subject parameter is constrained by a group parameter. Also, the full posterior distributions get estimated which allows to see how probable other parameter values are and allows for direct hypothesis testing that does not rely on t-tests or F-tests.
Kabuki
kabuki is a Python library that builds on top of PyMC, a general purpose library to do Bayesian inference using MCMC. What kabuki provides is a flexible implementation that allows users to easily construct their own hierarchical models. It features a common design where individual subject parameter distributions are themselves draws from a group parameter distribution.
Moreover, kabuki comes equipped with some models of ubiquitous use such as a Bayesian ANOVA model that allows estimation and hypothesis testing of different treatment groups and makes traditional statistical tests such as t and F-tests obsolete.
Product's homepage
Requirements:
· Python
Find us on Google+
