Martyn Plummer edited the most recent issue of R News, and it’s focused on Bayesian computation. Jouni and Sam and I have 2 articles there (one on Bayesian software validation and one on Bayesian data analysis (as opposed to simple Bayesian inference) in R). But here I’ll give some quick comments on the other articles in the issue.
I like all the articles. The below comments might be picky, but I hold articles in a computational forum such as R News to a higher standard.
1. I’m a little disappointed in some of the graphical displays. For example, on page 4, there’s an unwieldy table with numbers presented to 4 or 5 decimal places (e.g., 67.0208 with an sd of 11.08133)! FIgure 3 on page 6 has the same problems.
2. I’d also like to see some clearer guidance on practical issues. For example, for Bayesian inference it doesn’t make much sense to simulate 11,000 iterations, keep the last 10,000, and thin by 1 (also in the example on page 4). If you really needed 11,000 to converge (which I doubt you did, but it’s hard to tell based on simulating just one chain), then I think you’d better discard more than the first 1000. Conversely, if you have approximate convergence after 1000 iterations (i.e., that’s all you need to discard from the start of the chain), then I’d think 2000 would be enough, not 11,000. Obviously in this particular example, maybe 11,000 won’t bust your computational budget, but I’ve seen people running things overnight–and really reducing their ability to do flexible data analysis–so if 100 iterations are enough, why do so many more?
3. On page 5, there’s a pretty picture, but it’s hard to identify which case is which. Given tha tthe distributions are all basically Gaussian, I think they’d be more clearly displayed as intervals, which would free up one of the axes to display another variable (e.g., year when the justice was appointed) and would also allow each justice to be labeled directly.
I have a similar comment about Figure 1 on page 16: once again, with these densities practically Gaussian, it would be much more informative, I believe, to make a plot whose x-axis goes from 1 to 6, showing the posterior medians and intervals for each of the components of alpha. This would take up much less space (basically 1/6 the space of the current graph) and allow the relevant direct comparison between the alpha’s. In terms of data analysis, I just don’t see what’s gained by having all these curves.
4. The trace plot on page 9 is not maximally informative. It would help to display 2 or 3 parallel chains to get a better sense of mixing.
5. I’m disappointed that the example on page 18 uses the discredited dgamma(0.001,0.001) prior density.
OK, that’s all. Sorry for being such a curmudgeon. I really did like all the papers and I hope these comments will be useful.