Introduction to WinBUGS for Ecologists: Bayesian approach to regression, ANOVA, mixed models and related analyses

If you are an R user and a non-statistician (i.e., a professional researcher but without rigorous Math/Stats background) and want to learn Bayesian methods: get this book! It is one of the single best statistics books I've ever read for applied researchers.Most Bayesian method books start off fine with Bayes' Rule, followed by reasonable coverage of the binomial distribution, and then plunge off the deep end with formulas for likelihood and details on how to write MCMC samplers. When I'm in a bad mood, I suspect those books reflect disdain for non-statisticians and a desire to keep Bayesian methods a secret fraternity. If you've felt the same and been stymied, and you do classical stats in R, Kery has the perfect book to learn Bayesian approaches.Other reviewers have commented on how approachable the book is (true), how you need to know R first because it is too much heavy lifting to learn R and Bayes and WinBUGS all at the same time (I agree), and that it is light on Bayes theory and math (a very positive thing, in my opinion).What you might also wonder is: (a) can you use it with a non-Windows computer? and (b) does it only work with WinBUGS, or could one use openbugs or JAGS instead? Answers: yes and yes. I've been working through it with JAGS running on a Mac and using the "rjags" package. The primary changes that are required involve the fact that commands to call JAGS from R (in the rjags package) are different from those to call WinBUGS.I don't want to get too technical in a review, but the rjags approach is simple, works on Macs and Linux as well as Windows, and is not a big stretch from the book. For reference, here is rjags syntax for the first live example ("y1000" model in chapter 5.4; starting after the data setup and model file creation): "test.jags.mod <- jags.model(file="model.txt", data=test.jags.data) # no need to init in JAGS" + "test.jags.out <- coda.samples(test.jags.mod, c("pop.mean","pop.sd"), n.iter=1000) # run MCMC" + "summary(test.jags.out)" or "HPDinterval(test.jags.out)" to get the credible interval. Once you figure out a few of the key commands -- e.g., jags.model() and coda.samples() -- then it is very easy to adapt the book to JAGS. Of course, that is for someone who generally knows his or her way around R.One surprise and extremely nice feature of the book is that Kery takes the reader painstakingly through the generalized linear model. Even after two decades of doing applied research, I learned a few things and clarified a couple of misperceptions and shaky understandings. That in itself would be a good thing, even apart from the boost in my Bayes abilities.The main thing that could be improved is the explanation of the BUGS language. Kery goes through the glm statistical models in close detail, but spends little time on explaining the BUGS syntax and concepts, mostly teaching that by example. That's not a huge problem since it's relatively clear in context, but specific attention such as a special chapter mid-way through the book, would be welcomed.Finally: no, you do not need to be an ecologist to benefit from it! I'm a social scientist but the models are so close (comparing groups, hierarchical models with group level covariates such as location, etc) that there is no difficulty in translation. And I'm learning a little bit about a different field which is interesting in its own right.Is it for you? Applied researcher + using R + use linear/general linear models? Then Yes. Professional statistician? No, too simple. Want math? No, won't satisfy. Don't know R? No, go learn R first. Cheers!

I find this to be a very good introductory text for someone who wants to know WinBUGS, who already has a firm background in non-Bayesian statistics. I would consider this to be an upper-level undergraduate to graduate level text. I find that it does a very good job of getting the reader into WinBUGS analyses within a short time, and with pertinent examples. The breadth of examples and the construction of the simulated data make this book quite appealing for someone with a stats background, but no WinBUGS experience.At the same time, however, I do have some minor issues. First, for someone who is not familiar with R, this text requires one to concurrently learn parts of R at the same time as learning WinBUGS. I ran into several problems while working my way through this text, and had I not already known R, I think I would have gotten quite frustrated. So, while I truly appreciate the R-wrapper, I think a better approach would be to more-explicitly separate the R code and the WinBUGS code and focus more on the WinBUGS code. Second, because it is a strong example-based text, there is sometimes a dearth of explanation about why certain approaches are taken. I think some explanation of why certain priors are being used and what the alternatives are would be better. Third, the simulations and the analyses get quite complex quite early. I truly think more time could be spent on easier models under different designs, and that would prove more helpful than quickly ramping up to difficult models so quickly. And finally, the supposed strength of Bayesian analyses is the ability to bring knowledge to the table with informed priors. However, throughout this text all priors are uninformative, and the case is made about how alike these analyses are to classical analyses. But if that's the case, then why do Bayesian? I would have found it useful to include examples where priors are informative and the data do, and do not, reinforce the prior. A discussion about how to create reasonable informative priors, as well as what to do in the case when the data are either non-informative or contradictory would be a useful addition to this text.

This is a great book for using WinBUGS through R with the R library R2WinBUGS. It is actually also a pretty good book for performing classical linear modeling in R. All the analyses are performed in both R and WinBUGS. Much of the data is simulated but realistic, and the author shows you how he generated the data, which is also useful. Solutions to the exercises are available at the book's web site.There is no reason to be an ecologist to use this book. The examples translate very well to other fields.Despite this book being very useful to me, I gave it 4 instead of 5 stars for a few reasons. Very little attempt is made to explain the theory (to be fair, he says this at the outset, and it is a book about WinBUGS, not Bayesian statistics). The expected understanding of statistics and R is somewhat uneven throughout. For example, the author in one chapter shows you how to load libraries in R and other basic housekeeping tasks, but a few chapters later he shows more advanced model specification code in R's lm function without explaining it. Expect to spend some time in the R manual if you want to understand it all. Similarly, he repeatedly says that much of the statistics behind the code is too advanced for most ecologists, which might annoy me if I were an ecologist, but then he tends to assume a lot of the theoretical statistics is already well understood by the reader.There is a quick introduction to Generalized Linear Models which I found helpful. Basically, this is a great practical book but you will need to look elsewhere for mathematical understanding. I like Peter Hoff's "A First Course in Bayesian Statistical Methods."

For a behavioural ecologist trying to get to grips with Bayesian methods, this book comes as a breath of fresh air. It's a practical guide and uses ecological examples I can relate to, but best of all it tackles the same datasets with both classical frequentist and Bayesian methods, so it is easy to compare the outcomes from the two approaches. The author is a population ecologist at the Swiss Ornithological Institute with a wealth of experience in analysing messy biological data on temporal and spatial distributions of birds, and this is very much a 'hands on' guide. I'm a behavioural ecologist and ornithologist myself, with a side interest in statistics and statistics teaching, and I'm always buying stats books for general 'self improvement' and in the hope I'll find one the perfect book for my students. Much as I like Gelman and Hill (Data Analysis Using Regression and Multilevel/Hierarchical Models (Analytical Methods for Social Research) as an authoritative but accessible text, and Albert (Bayesian Computation with R) got me started on WinBUGS, I'll be telling all my students and friends to buy Kery's book. It is worth pointing out that the book is also useful for teaching the logic of linear modelling within the classical frequentist framework, explaining how analyses from t-tests through ANOVA to mixed models can be expressed within the same framework. Kery's tactic is to encourage the reader to generate and use artificially datasets in addition to real data. Not only does this allow you to check the results against the 'real' answer but, because you create the artifial data using linear equations plus random error, the relationship between fitted linear models and the assumptions being made about the modelled data becomes explicit. Another advantage of the book is that all the code to generate the data and run the analyses is provided, the book being based around 'R', the free, open source, programming language of choice for many statisticians. Although WinBUGS (also free) is the software engine for the Bayesian analyses, it is driven from within R, so you don't need to learn new software and 'staying within R' simplifies the process of comparison with familiar classical tests. Kery's book assumes you already know the basics of R, and it will not teach you the mathematical theory underlying the methods, but it doesn't claim to and it certainly delivers as practical guide for biologists wanting to make the daunting journey from familiar frequentist stats to the world of Bayes. Highly recommended.

Teaches statistics in a whole new way. I recommend this book, not just for the bayesian, but also for the unique approach to reviewing frequentist stats.

ベイズ統計とはどういう考え方かということも含めて基礎から知りたいと思い、同じ著者の前書きを読んでまずこちらを購入しました。一例一例、わかりやすく記載されており、ベイズ的な考え方について次第になじんでくるという意味ではいい本だと思います。ただ、理論的なところや、そもそもWinbugsでなければならないのかといったあたりの統計的なところが知りたい場合は物足りないかもしれません。

Winbugsを使って解析をした人なら十分に理解できるだろう。初学者は、マッカーシーの「生態学のためのベイズ法」あたりで、知識を整理してから読んだほうが無難だろう。