Comment on “Bayesian astrostatistics: a backward look to the future” by Tom Loredo, arXiv:1208.3036
This short note points out two of the incongruences that I find in the loredo12 comments on andreon2012bayesian, i.e. on my chapter written for the book “Astrostatistical Challenges for the New Astronomy”. First, I find illogic the Loredo decision of putting my chapter among those presenting simple models, because one of the models illustrated in my chapter is qualified by him as “impressing for his complexity”. Second, Loredo criticizes my chapter at one location confusing it with another paper by another author, because my chapter do not touch the subject mentioned by loredo12 critics, the comparison between Bayesian and frequentist fitting models.
loredo12 paper, “Bayesian astrostatistics: a backward look to the future” is, as indicated in his Comment line, a lightly revised version of a chapter in “Astrostatistical Challenges for the New Astronomy”, the inaugural volume for the Springer Series in Astrostatistics. This volume also includes a chapter written by me (andreon2012bayesian), and one by March et al. My chapter aims to show the simplicity of performing a Bayesian statistical analysis with two examples, the first one is original and has been first presented in andreon2010scaling, the second example is drawn from March11. In Sect. 4.3 of my chapter I also consider a model with increased complexity by inserting new nodes in the March11 model to check model adequacy.
I disagree with several loredo12 comments on my chapter, but I will restrict my attention here to two of them.
loredo12 paper111At the time of this writing there is one single version of the paper, arXiv:1208.3036v1, and this is the version which I’m referring to. splits chapters of the volume in two groups according to the complexity of the used models. The first one is composed by chapters using simpler models and my chapter belongs to this group according to loredo12. The other group is composed by contributions using more complex models, and includes the March et al. chapter. The March et al. model is commented by him as “impressing for his complexity” (his pag 20).
Unfortunately for loredo12, the March et al. model is also the model adopted in my second example, and thus my chapter unambigously deals with complex models, and must be put among those dealing with complex models. To be pedant, in Sect. 4.3 of my chapter I add complexity to March et al. model, and in Sect. 3 I present one more model, developed by myself & Hurn, of similar complexity.
About the second incongruence, loredo12 paper, after referring a few times to my chapter as “Andreon’s” chapter, topic, contribution or just “Andreon’s”, comments: “In the context of nonlinear regression with measurement error –Andreon’s topic– Carroll et al. (2006) provides a useful entry point to both Bayesian and frequentist literature, incidentally also describing a number of frequentist approaches to such problems that would be more fair competitors to Bayesian MLMs than the approach that serves as Andreon’s straw man competitor.”
Unfortunately for loredo12, in my chapter (andreon2012bayesian) there is no comparison between a Bayesian and fit or any other frequentist fitting method. There are no frequentist approaches to regression in my chapter. There is no competition at all between methods, because only a single method (Bayesian) is used to fit the data. loredo12 is criticizing a comparison between frequentist and Bayesian fitting models, but there is no trace of it in my chapter, simply and plainly he is confusing my chapter with another paper, although my chapter get his criticisms.
To sum up, I find illogic the loredo12 decision of putting my chapter among those presenting simple models, given his claim that one of models illustrated in my chapter is “impressing for his complexity” and the other one has similar complexity. Second, loredo12 criticizes my chapter at one location confusing it with another paper by another author, because my chapter do not touch the subject criticized by Loredo, the comparison of frequentist and Bayesian fitting models.