From prior knowledge to informed conclusions: Teaching the Bayesian workflow in problem-solving
Bayesian statistics is conceptually simple. But this conceptual simplicity does not necessarily translate into practical simplicity, mainly for beginners.
Designing an appropriate model for a given data analysis task requires both statistical expertise and knowledge of the application domain, and
is generally carried out as an iterative process involving repeated testing and refinement.
This process can be formulated as a Bayesian workflow that, in addition to the inference process itself,
involves other steps that include specifying a priori distributions, validating models, comparing models,
performing diagnostics of numerical inference methods, presenting the results to a specific audience, etc.
In this work, we propose researching and developing pedagogical strategies for effectively teaching the Bayesian workflow..
Semi-automatic specification of prior distributions
A priori distribution specification refers to the process of transforming domain knowledge into well-defined probability distributions.
Currently, there are no truly general and widely applicable methods to assist in this process.
In this work, we propose the creation of new computational methods that operate,
in a complementary way, in the space of parameters and observations.
The methods will be semi-automatic and will seek to alleviate the user's cognitive load
while facilitating the incorporation of domain knowledge.
Workflow for Approximate Bayesian Computing (ABC)
Simulation is central to many scientific problems.
ABC methods combine the use of Bayesian methods with simulations.
We propose to develop new computational methods for diagnosis and model comparison
specifically for ABC. We also propose to study and disseminate
good practices for ABC modeling.