Bayesian Core under fire!

A pretty harsh criticism of Bayesian Core appeared on Amazon in early December but I only saw it a few days ago. Its entry is clearly setting the tone: “There are many first rate texts on Bayesian statistics. Unfortunately Bayesian Core is not one of them.” Aie, aie, aie!

Everyone is entitled to his/her (own) opinion and we cannot expect our book to satisfy all readers. So I perfectly understand for someone to be unhappy with our book and to think that there are other (much) better books at a similar level.. The published reviews reproduced on our webpage also note some possible improvements for the book, albeit in a much more balanced way!, and in particular that it may be aiming at a higher level than the one we claim.

The only strong reservation I have about those comments deals with the final sentences: “I am surprised that Springer published Bayesian Core. From every point of view this book is a disaster and deserves zero stars. I think the authors have pulled off a huge con-job.” Since “con-job” is defined as

1. an act or instance of duping or swindling.
2. an act or instance of lying or talking glibly to convince others.

I do object to this accusation of dishonesty! Bayesian Core was tested in courses here and there before being published and we did get a fair return: the course given in Paris Dauphine was intended for fifth year students completing a professional degree and they did benefit from the course enough to put Bayesian statistics into practice when they started working. Similarly, the course I gave at the University of Canterbury to third year students induced one student to engage into a Master and then a PhD in Bayesian statistics! These examples are just given to state that (a) we (and others) are actually teaching from the book and (b) we truly got an enthusiastic response from students when using this book. So how and where are we lying or duping the reader..? Bayesian Core was regularly assessed by Springer Verlag through (a) two anonymous academics whose suggestions were mostly implemented before publishing the book and (b) a professional copy-editor who corrected our syntax and style.

Now, the “swindle” may be that the “book is not “self-contained” and that it is not the “practical approach” referred to in the title“. I faced a similar reaction last summer during a one-week course in Brisbane when some students came with high expectations but no background: after the first hour, they objected to my customary use of the t density and of too many Greek symbols on my slides… This was unfortunate as the requested background in probability and statistics had been advertised on the advertising flier (as well as on the backcover of Bayesian Core) and the slides were available before the registration. So “self-contained” is a rather slippery notion that one interprets from one’s own viewpoint. For us, it is meant as “no prior knowledge about Bayesian statistics, nor computational statistics”…

The point about the book not being “practical” is a matter of opinion. For us, Bayesian Core is practical in that (a) it does not cover the theory behind the tools [as noted in the comments, “don’t expect expect any statistical theory“] or even philosophical motivations for the Bayesian approach, but on the opposite it provides a single set of priors for each model analysed there and (b) it supports each chapter with at most two datasets that are realistic (if not always real) and diverse enough to motivate readers. The choice made by using simple datasets requesting short descriptions can be argued as being far from true practice, but this is in line with the terse style we adopted. This is also a pedagogical choice in that more advanced applications require some details about the area of application that may not appeal to readers from other application domains.

The vision that Bayesian Core is a compulsory setwork for the authors’ hapless students with compulsory exercises (…) with no hints or answers, interspersed with formulae with little or no discussion, very few worked examples and no theorems.” is immensely reductive and, in short, wrong, in that it gives the impression that the book was built around the exercises (and that our students were unhappy!)… This is obviously not the case. When designing the book, we found that having compulsory exercises had a positive impact on the students. This may be a cultural difference with other audiences in that our French students do not spontaneously study problems on their own, but having them hit compulsory exercises in the middle of the text is a pedagogical device that worked for us and for them. Providing answers to the exercises is not possible when considering that those exercises are used as weekly homeworks. We did write a complete set of answers available for instructors, though (when looking hard enough, you may alas find a pirated version on the Web). Each chapter being centred on a dataset, there are few different examples in each chapter but they are worked out to address the issues central to the current topic. Having “no theorems” is a deliberate choice and in line with the “practical” orientation of the book (check the backcover, it’s there!). Theoretical references are provided whenever necessary but in limited numbers because we think readers should practice on the datasets using the tools we provide. Obviously, this pedagogical choice is arguable, but it keeps the length under control (250 pages!) and the discourse as linear (as opposed to branching) as possible.

The subsequent comments in the review oppose Bayesian Core to “four excellent texts” from Bolstad, Leonard and Hsu, Albert, and Gill. I cannot launch here into a comparison with those−indeed excellent−books, but I think the most relevant reference is Jim Albert’s “Bayesian Computation with R“, Jeff Gill’s 752 pages Bayesian Methods for social scientists being closer to Bayesian Core scope, bulk and audience.

Now about details: The comment that “conjugacy is mentioned in exercise 2.10 on page 22 with no discussion” is mistaken since conjugate priors are introduced on page 21 and discussed all the way to the end of page 23. Similarly, improper priors are defined on page 24, including Jeffreys’ prior, instead of having “a terse and unhelpful half page entry on page 34” (the index under the entry “Prior, Jeffreys'” correctly points out to page 24). ! I do not see why a book for practitioners does require a “full discussion of Jeffrey’s prior, its historical antecedents, theory“…. Indeed, Jim Albert’s book adopts the same perspective as ours on improper priors (and on theory in general) with a very terse introduction backed up by a few standard cases. That the “Jeffreys-Lindley paradox on page 33 is not explained” is a matter of opinion: the fact that the Bayes factor always goes to 1 when the prior variance goes to infinity illustrates the difficulty in using improper priors. (In the next edition, we should avoid relating this fact with the “Jeffreys-Lindley paradox“…) The criticism that “the six pages [after page 36] of poor discussion and two impenetrable examples [on the Monte Carlo method] bear testimony to this cop-out” [that it is straightforward] is also surprising given that the Monte Carlo method is indeed straightforward. I agree however that we could change Example 2.3 into a more natural problem given that the integral I on page 40 based on a Student’s t density is artificial (but not “impenetrable”). But the point in showing that poor choices of importance functions lead to poor Monte Carlo estimates is made. Similarly, the “almost incomprehensible entry for the Gibb’s Sampler” page 72 would have benefited from a clearer criticism. The choice we made is as usual to skip the theoretical justifications for MCMC methods and to illustrate the point through examples. Example 3.2 is a standard completion example that I have been using for decades, with no negative feedback so far. The generalisation of the two-stage Gibbs sampler to the generic Gibbs sampler on pages 76-77 can be seen as too terse, but it is further illustrated in Section 3.5.4 for model selection. Again, this pedagogical choice of backing a generic tool on a specific model supported by a dataset is open to criticism.

As posted earlier, we are planning writing the new edition later this year so we are definitely welcoming detailed comments and specific criticisms from readers to make the second edition as improved as possible. Of course, if they could be more constructive than the ones just discussed, it would help! And if you have an opinion on Bayesian Core that you want to share with others, feel free to write an Amazon review!

7 Responses to “Bayesian Core under fire!”

  1. […] thus seems like a compulsory element of publishing books with exercises. (As we discovered a wee too late for Bayesian […]

  2. […] to provide graded homeworks, as I do in Dauphine. However, this policy of ours induced  repeated criticisms and urgent requests from readers unable to complete some of the essential exercises. Since some […]

  3. […] like homeless, heroin addicts or prostitutes…) I was thus quite surprised to hear from many that the capture-recapture chapter in Bayesian Core was hard to understand. In a sense, I find it […]

  4. […] time as well as make [some of] them only available to instructors), even though this may attract criticisms. (The mcsm package can [and need] also be upgraded at will, given the rapidity with which the CRAN […]

  5. Critics are just that CRITICS. But like any literature author will ask WHERE WERE ALL THE CRITCS WHEN I WAS WRITING? Mark you most critics are never authors and all they do is lurk in corners with selfishly sharpened blades in wait to slay and mutilate the hard work of those selfless enough to advanceknowledge. Opinions are free but facts remain sacred, so may all critics stick to facts and keep their opions to themsleves. We are far mush better in acadmia with fewer or no alarmist.

  6. […] the very critical review discussed in this post, two more balanced reviews of Bayesian Core have now appeared on amazon.com, including one by the […]

  7. […] post, we are obviously interested in any feedback from instructors and students about teaching from Bayesian Core in order to improve further the second […]

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.