This kinda thing drives me crazy! But what causes it? Is it the result of learning different mass matricws for each chain during warmup?

I'm not sure how much value there is reasoning about the marginal distribution of effect sizes over all conducted experiments, but if you do, I think it's kind of reasonable to think that most real world effects are small and yield something a priori kinda Gaussian (but maybe heavy tailed).

Yeah-- I'm not a fan of NHST, in particular because I don't really believe there are "true nulls". I'm just trying to articulate how I think people think about the two groups model (null/alt). I prefer Bayesianism, where I do think you should think about prior distributions over effect sizes.

Ignoring publication selection effects, Z-scores under the null are standard normal and under the alternative "something else". How non-normal the marginal z-scores look depends on the mixture weights (the proportion of randomized experiments that are null) and how non-normal the alts look.

I find his lecture on storytelling particularly awesome: clauswilke.com/lectures/sli...

Goes beyond just graphic design but @clauswilke.com has some great tips for visualizations and presentations, e.g. wilkelab.org/dataviz_shor...

I've been thinking about developing this course for a little while, and the @chelseaparlett.bsky.social post and reaction may be the tipping point for me to finally take it on: bsky.app/profile/chel...

Not saying that I totally buy it, but this post outlines an interesting justification based on modeling the half life of a "fatal mistake" using tools from survival analysis: www.tobyord.com/writing/half...

Meant to say AOS.

I do think that stats depts should be doing much more to emphasize data analysis. One of the best ways to generate idea for stats methods research is to be intimately familiar with data analysis and the challenges that come with it.

In general, I think nowadays most departments would be equally happy with an applied statistician publishing in AOAS, JASA A&CS and JCGS as they would publishing in AOS.

What is most valued at tenure time also depends on what kind of department. E.g. as you allude to, a biostat dept will have very different criteria than a stats dept in say, a business school, which definitely skews more toward AOAS style publications.

I like to think of academic statistics on a continuum: applied, methods and then theory. In my opinion, a "strong" stats dept should have faculty at a range of points on this continuum.