Bayesian Computation with R
Author:Jim Albert
Publisher: Springer (2007)

As the title said, accompanying R package LearnBayes is available (clicking the name will direct you for downloading the package). Furthermore, the last chapter is about WinBUGS. (Please, check out resources listed in BUGS for other BUGS, Bayesian inference Using Gibbs Sampling) Overall, it is quite practical and instructional. If an young astronomer likes to enter the competition posted below because of sophisticated data requiring non traditional statistical modeling, this book can be a good starting. (Here, traditional methods include brute force Monte Carlo simulations, chi^2/weighted least square fitting, and test statistics with rigid underlying assumptions).

An interesting quote is filtered because of a comment from an astronomer, “Bayesian is robust but frequentist is not” that I couldn’t agree with at the instance.

A Bayesian analysis is said to be robust to the choice of prior if the inference is insensitive to different priors that match the user’s beliefs.

Since there’s no discussion of priors in frequentist methods, Bayesian robustness cannot be matched and compared with frequentist’s robustness. Similar to my discussion in Robust Statistics, I kept the notion that robust statistics is insensitive to outliers or iid Gaussian model assumption. Particularly, the latter is almost ways assumed in astronomical data analysis, unless other models and probability densities are explicitly stated, like Poisson counts and Pareto distribution. New Bayesian algorithms are invented to achieve robustness, not limited to the choice of prior but covering the topics from frequentists’ robust statistics.

The introduction of Bayesian computation focuses on analytical and simple parametric models and well known probability densities. These models and their Bayesian analysis produce interpretable results. Gibbs sampler, Metropolis-Hasting algorithms, and their few hybrids could handle scientific problems as long as scientific models and the uncertainties both in observations and parameters transcribed into well known probability density functions. I think astronomers like to check Chap 6 (MCMC) and Chap 9 (Regression Models). Often times, in order to prove strong correlation between two variables, astronomers adopt simple linear regression models and fit the data to these models. A priori knowledge enhances the flexibility of fitting analysis in which Bayesian computation works robustly different from straightforward chi-square methods. The book does not have sophisticated algorithms nor theories. It only offers very necessities and foundations for Bayesian computations to be accommodated into scientific needs.

The other book is

Bayesian Core: A Practical Approach to Computational Bayesian Statistics.
Author: J. Marin and C.P.Robert
Publisher: Springer (2007).

Although the book is written by statisticians, the very first real data example is CMBdata (cosmic microwave background data; instead of cosmic, the book used cosmological. I’m not sure which one is correct but I’m so used to CMB by cosmic microwave background). Surprisingly, CMB became a very easy topic in statistics in terms of testing normality and extreme values. Seeing the real astronomy data first from the book was the primary reason of introducing this book. Also, it’s a relatively small volume book (about 250 pages) compared other Bayesian textbooks with the broad coverage of topics in Bayesian computation. There are other practical real data sets to illustrate Bayesian computations in the book and these example data sets are found from the book website

The book begins with R, then normal models, regression and variable selection, generalized linear models, capture-recapture experiments, mixture models, dynamic models, and image analysis are covered.

I feel exuberant when I found the book describes the law of large numbers (LLN) that justifies the Monte Carlo methods. The LLN appears often when integration is approximated by summation, which astronomers use a lot without referring the name of this law. For more information, I rather give a wikipedia link to Law of Large Numbers.

Several MCMC algorithms can be mixed together within a single algorithm using either a circular or a random design. While this construction is often suboptimal (in that the inefficient algorithms in the mixture are still used on a regular basis), it almost always brings an improvement compared with its individual components. A special case where a mixed scenario is used is the Metropolis-within-Gibbs algorithm: When building a Gibbs sample, it may happen that it is difficult or impossible to simulate from some of the conditional distributions. In that case, a single Metropolis step associated with this conditional distribution (as its target) can be used instead.

Description in Sec. 4.2 Metropolis-Hasting Algorithms is expected to be more appreciated and comprehended by astronomers because of the historical origins of these topics, detailed balance equation and random walk.

Personal favorite is section 6 on mixture models. Astronomers handle data of multi populations (multiple epochs of star formations, single or multiple break power laws, linear or quadratic models, metalicities from merging or formation triggers, backgrounds+sources, environment dependent point spread functions, and so on) and discusses the difficulties of label switching problems (identifiability issue in codifying data into MCMC or EM algorithm)

A completely different approach to the interpretation and estimation of mixtures is the semiparametric perspective. To summarize this approach, consider that since very few phenomena obey probability laws corresponding to the most standard distributions, mixtures such as (*) can be seen as a good trade-off between fair represntation of the phenomenon and efficient estimation of the underlying distribution. If k is large enough, there is theoretical support for the argument that (*) provides a good approximation (in some functional sense) to most distributions. Hence, a mixture distribution can be perceived as a type of basis approximation of unknown distributions, in a spirit similar to wavelets and splines, but with a more intuitive flavor (for a statistician at least). This chapter mostly focuses on the “parametric” case, when the partition of the sample into subsamples with different distributions f_j does make sense form the dataset point view (even though the computational processing is the same in both cases).

We must point at this stage that mixture modeling is often used in image smoothing but not in feature recognition, which requires spatial coherence and thus more complicated models…

My patience ran out to comprehend every detail of the book but the section of reversible jump MCMC, hidden Markov model (HMM), and Markov random fields (MRF) would be very useful. Particularly, these topics appear often in image processing, which field astronomers have their own algorithms. Adaption and comparison across image analysis methods promises new directions of scientific imaging data analysis beyond subjective denoising, smoothing, and segmentation.

Readers considering more advanced Bayesian computation and rigorous treatment of MCMC methodology, I’d like to point a textbook, frequently mentioned by Marin and Robert.

Monte Carlo Statistical Methods Robert, C. and Casella, G. (2004)
Springer-Verlag, New York, 2nd Ed.

There are a few more practical and introductory Bayesian Analysis books recently published or soon to be published. Some readership would prefer these books of running ink. Perhaps, there is/will be Bayesian Computation with Python, IDL, Matlab, Java, or C/C++ for those who never intend to use R. By the way, for Mathematica users, you would like to check out Phil Gregory’s book which I introduced in [books] a boring title. My point is that applied statistics has become more friendly to non statisticians through these good introductory books and free online materials. I hope more astronomers apply statistical models in their data analysis without much trouble in executing Bayesian methods. Some might want to check BUGS, introduced [BUGS]. This posting contains resources of how to use BUGS and available packages under languages.

[stat.CO:0808.2902] A History of Markov Chain Monte Carlo–Subjective Recollections from Incomplete Data– by C. Robert and G. Casella
Abstract: In this note we attempt to trace the history and development of Markov chain Monte Carlo (MCMC) from its early inception in the late 1940′s through its use today. We see how the earlier stages of the Monte Carlo (MC, not MCMC) research have led to the algorithms currently in use. More importantly, we see how the development of this methodology has not only changed our solutions to problems, but has changed the way we think about problems.

Here is the year list of monumental advances in the MCMC history,

• 1946: ENIAC
• late 1940′s: inception along with Monte Carlo methods.
• 1953: Metropolis algorithm published in Journal of Chemical Physics (Metropolis et al.)
• 1970: Hastings algorithms in Biometrika (Hastrings)
• 1974: Gibbs sampler and Hammersley-Clifford theorem paper by Besag and its discussion by Hammersley in JRSSS B
• 1977: EM algorithm in JRSSS B (Dempster et al)
• 1983: Simulated Annealing algorithm (Kirkpatrick et al.)
• 1984: Gibbs sampling in IEEE Trans. Pattern Anal. Mach. Intell. (Geman and Geman, this paper is responsible for the name)
• 1987: data augmentation in JASA (Tanner and Wong)
• 1980s: image analysis and spatial statistics enjoyed MCMC algorithms, not popular with others due to the lack of computing power
• 1990: seminal paper by Gelfand and Smith in JSAS
• 1991: BUGS was presented at the Valencia meeting
• 1992: introductory paper by Casella and Georgy
• 1994: influential MCMC theory paper by Tierney in Ann. Stat.
• 1995: reversible jump algorithm in Biometrika (Green)
• mid 1990′s: boom of MCMC due to particle filters, reversible jump and perfect sampling (second-generation of MCMC revolution)

and a nice quote from conclusion.

MCMC changed out emphasis from “closed form” solutions to algorithms, expanded our immpact to solving “real” applied problems, expanded our impact to improving numerical algorithms using statistical ideas, and led us into a world where “exact” now means “simulated”!

If you consider applying MCMC methods in your data analysis, references listed in Robert and Casella serve as a good starting point.