1) Interdisciplinarity

We typically have students that come from various disciplinary backgrounds. I emphasize this for two main reasons. The first reason is for you to realize that interest in Bayesian statistics pervades multiple scientific disciplines (i.e., it is not a tool that is restricted to say ecology). The second reason is for you to be patient and opened to learning when hearing about problems in other fields. You’d be surprised by how problems in disparate fields are related and have similar solutions. I particularly believe that there is a lot to gain in seeing how other scientific fields have approached and solved certain problems.

2) Toy examples and standard statistical models

Students often wonder why would anybody perform a Gaussian linear regression in a Bayesian framework when one can obtain similar results using a single R command within a frequentist framework. Similarly, why do we spend so much time on models that are clearly too simple to be useful in the real world?

I employ toy examples and standard statistical models to illustrate the basic concepts of Bayesian statistics. By using standard models, I am hoping to build on modeling ideas that are already familiar to the student, allowing us to focus on the nitty-gritty details of how to implement it. I view these standard models and toy examples as building blocks. The real advantage of Bayesian statistics is to be able to go way beyond these standard models by combining these basic ideas in more interesting ways, leading to much more complex models.

3) Why don’t we just use JAGS? Why do we have to learn to implement our own Gibbs sampler in R?

Bayesian statistics involves two distinct but interrelated tasks:

1. being able to formulate the statistical model in a way that makes biological sense (statistical modeling); and
2. being able to implement the model in the form of an algorithm (computer programming).

JAGS is useful because it allows us to focus on task (a). Unfortunately, JAGS has several drawbacks. First, JAGS can be very slow when you have lots of data. Second, some things are (to be the best of my knowledge) impossible to implement in JAGS. For instance, JAGS does not work very well for Bayesian model selection procedures. Third, when you do develop your own Gibbs sampler, you have full control over the algorithm and it provides a level of understanding (e.g., helping to identify when the model is misspecified or when the prior is too influential) that is hard to get otherwise.

For this course, I use JAGS as a starting point to then show how we can develop the same model directly in R.

Back to main menu