Bayes Watch!

I’m re-developing a subject called “Bayesian Statistical Methods” (BAY), an advanced elective in the Master of Biostatistics taught entirely online through the Biostatistics Collaboration of Australia.

Students are typically mature-age “career changers” or clinicians / medical researchers who are learning to do their own statistics. They will not have encountered Bayesian statistics previously, which offers a very different philosophical approach to statistical inference compared to the dominant “Frequentist” paradigm with which students will be familiar (Amrhein et al. 2019).

Bayesian statistics was propelled into the mainstream of statistical practice with the appearance of the BUGS (Bayesian Inference Using Gibbs Sampling) software in 1989, which made Bayesian data analysis available to a wide audience. It was the primary motivation for the development of this subject, which has run biennially since 2004.

There are two groups of learning objectives that centre around the logic of Bayesian statistical inference (mathematics) and the practice of Bayesian data analysis (computation).

BAY must now be overhauled for two reasons (1) the model for online delivery is out of date; and (2) the BUGS software was replaced by the Stan platform in 2012.

Updating the model for online delivery

The diagram below shows the workflow for a fortnightly module in the existing BAY subject. Note the linear path through modules notes, exercises and online Q&A to an assessment task, with little opportunity for discussion and to share ideas interactively. This mode of delivery is missing the cognitive and social presence required to create a Community of Inquiry within a Community of Practice (Garrison 2009).

The workflow for a fortnightly module in the existing BAY subject

The hub for the subject will be the Canvas LMS. It will hold a collection of curated material around foundational ideas from which students can choose a format that most appeals to them. These resources will include pre-recorded video lectures, online tools and demonstrations (see video below), suggested reading from the Module Notes (an in-house textbook), published textbooks and research papers, and a selection of quiz questions.

A very preliminary demonstration of iterative sampling using MCMC made by Lyle Gurrin

The design framework that prompted these is the “flipped classroom” (Mazur 1997, Crouch and Mazur 2001), since information transfer and knowledge attainment is expected to take place ahead of the asynchronous online text exchanges and live “Zoom” tutorials each week, during which knowledge can be assimilated and consolidated.

Updating to state-of-the-art software

Stan is a sophisticated platform where users need at least a basic understanding of imperative programming, libraries of functions and algorithms, compiling C++ code and the use of development tools. In 2018, students of BAY were required to use Stan, but responses to student surveys reflected that most had a poor experience in the subject:

Student Experience Surveys for six MBiostat subjects in the BCA, Semester 2, 2018: BAY comes last!

The design of the new online educational environment for BAY takes its inspiration from Laurillard’s Conversational Framework and her six learning types (Laurillard 2012, 2013). A related taxonomy of Conole and Fill (2005) proposed that different media forms have different affordances that provide a different levels of support for learning and gradually broaden the scope of the tasks the student is expected to complete.  Paraphrasing their media forms and applying them to software tools: 

1. Narrative tools are textbooks, screenshots or videos illustrating software in action.
2. Interactive tools respond in a limited way to what the learner does (e.g. change simple settings like number of samples, input a different example dataset).
3. Adaptive tools are changed by what the learner does (e.g. add variables to model, change the form of the model, monitor different output streams).
4. Productive tools allow the learner to produce something (e.g. write code to implement a new model, consolidate component variables to generate desired output streams).

The diagram below maps each of a series of statistical computing tasks with relevant software tools and names the type of tool from the list above.

The level of support afforded by combinations of computing task and software platform

These software tools and online repositories for students’ programming portfolios  complement the curated materials for each fortnightly module and the communication tools such as “Zoom” and “Piazza” to form the Ecology of Resources.

The ideal process to evaluate the new approach would be a design-based research project to determine whether we can (1) effectively triage students to their level of statistical computing knowledge, skills and experience; and (2) get them to set goals for progressing from interactive tools through adaptive to productive and creative platforms.

References

Amrhein V, Greenand S, McShane B. (2019). Retire statistical significance. Nature, 567, 305 – 307.

Crouch C, Mazur E. (2001). Peer Instruction: Ten Years of Experience and Results. American Journal of Physics, 69, 970–977.

Conole G, Fill K. (2005). A learning design toolkit to create pedagogically effective learning activities. Journal of Interactive Media in Education, 1, http://doi.org/10.5334/2005-8

Garrison DR. (2009). Communities of Inquiry in Online Learning. Chapter 52 in Encyclopedia of Distance Learning (2^nd Edition), 352 – 355.

Laurillard D. (2012). Teaching as a Design Science: Building Pedagogical Patterns for Learning and Technology. New York and London: Routledge

Laurillard D. (2013). Rethinking University Teaching: A Conversational Framework for the Effective Use of Learning Technologies (2nd ed.). London: Routledge Falmer.

Mazur E. (1997). Peer Instruction: A User’s Manual. Series in Educational Innovation. Prentice Hall, Upper Saddle River, NJ.