2 min read

When is Bayesian introductory statistics better?

For the sort of statistics taught in introductory courses, competent Bayesian and frequentist analysis are going to agree – point and interval estimates will be similar, and similar conclusions will be drawn.  Computation isn’t seriously hard for either approach, though prepackaged pointy-clicky software is more available for frequentist inference. 

There are going to be pedagogical differences.  The big one, in favour of Bayesian statistics, is not having to explain p-values. Another benefit of the Bayesian approach is that it gives you a good reason not to talk about rank tests. Against this are some disadvantages.

Firstly, you are more or less forced to work with simple parametric models for the data – you could just do what we do with the t-test and ANOVA and say that group means are Normally distributed, but I’ve never seen this approach used. Of course, a lot of frequentist introductory courses use simple parametric models for the data, but that doesn’t stop it being a somewhat dodgy idea, and it is well known to lead to students behaving as if non-Normal distributions don’t really have means. 

Secondly, and more importantly, it’s harder to explain why random sampling and randomization are important. The reason is to do with independence between exposures and unmeasured confounders, but you usually wouldn’t include the unmeasured confounders in your model at an introductory level. 

In practice, though, the introductory course is a service course and the deciding issues is that the consumers wouldn’t like it to be Bayesian.