Untitled Document

MCMC Workshops in Evolutionary Biology

I have given two workshops on MCMC methods in population genetics and phylogenetics. The first was held in 2006 at the Monterey Bay Aquarium Research Institute (MBARI) in Moss Landing, CA, and the second in 2009 at Ewha Womans University in Seoul, Korea. The 2006 workshop was a half-day workshop and the 2009 workshop was a full-day workshop with an additional computer lab session devoted to applying the lecture material to MCMC output from the phylogenetic program MrBayes. Course notes and supporting materials from both of these workshops can be found below.

Please contact me if you are interested in holding this workshop or a derivation of it at your institution.

MCMC workshop held at Ewha Womans University, Seoul, Korea, in 2009

Course Notes (pdf, 5.7MB)
Downloaded >3500 times since June 2009. If you have found this e-book useful, please feel free to email me and let me know.

Supporting material (MCMC data files, slides from lab, R scripts in zip file, 9.8MB)

Course Evaluation (Word Document)

The workshop held in Seoul was attended by 59 participants from 14 institutions. Participants ranged from graduate students to senior researchers with attendees having broad scientific backgrounds. The workshop format included morning and afternoon lectures in English with time set aside for Korean translation (translator: Yong-Jin Won).

Institutions with participating members
Busan University
Chungbuk National University
Chungnam National University
Ewha Womans University
Incheon University
Inha University
Korea Ocean Research & Development Institute (KORDI)
Sangmyung University
Seoul National University
Sogang University
Sungshin Women's University
Welleslen college
Yonsei University
Zoo of Seoul

MCMC workshop held at MBARI in 2006

Course notes (slides):
Session 1
Session 2

Course notes (conversational):
Session 2

Supporting material (MCMCthin and R scripts)
Software

Workshop Flyer

---

You are invited to attend a workshop:
Application of MCMC methods in population genetics (case study: IM)

Location: Pacific Forum, MBARI
Date: February 10, 2006
Times: Session 1: 1pm - 3:00pm
Session 2: 3:30pm - 5pm
Instructor: C. R. Young
email: young@mbari.org

This workshop is designed to enable participants to apply Markov chain Monte Carlo (MCMC) methods to population genetic data. Session 1 covers relevant background material. Session 2 places emphasis on convergence issues that arise in any MCMC method, but are particularly troublesome in cases where we must integrate across genealogies and the model state space is complex. Session 2 covers the use of the CODA package in R to monitor convergence of MCMC analyses. The methods are generally applicable to any MCMC analysis, so whether you use IM or not, you may find the information covered in this workshop to be of value.

Session 1: Background material
Review of conditional probability
Statistical distributions
Model vs. observational uncertainty
Coalescent theory
Bayes theorem
Monte Carlo sampling
Markov chains

Session 2: MCMC convergence
MCMC: Convergence and tuning of IM
Some brief background is given on Markov chains and MCMC methods in general in hopes of building some intuition about the behavior of IM's MCMC algorithm. I give some suggestions on formally and informally monitoring convergence with R using the output from MCMCthin. Advice is given regarding tuning of the MCMC algorithm. The companion program/script to these notes are MCMCthin and "Rcode.txt". MCMCthin is a simple thinning utility that reads IM surface files. Rcode is a compilation of R functions that aids in analyzing the output of IM filtered through MCMCthin. This script can be cut and pasted in full into the R console and automatically generates various graphical files as well as numerical descriptions of the MCMC dataset.

The script performs:
· Parameter traces over the length of the chain
· Cumulative quantile plots (0.025%, 50%, and 97.5%)
· Marginal kernel density estimates (KDEs)
· Joint KDEs
· 2D contour plots of joint KDEs
· Summary statistics of marginal distributions (including cross correlations and autocorrelations)
· Effective sample size estimates
· Geweke's convergence diagnostic
· Raftery and Lewis's diagnostic
· Gelman-Rubin diagnostic
· Heidelberger-Welch diagnostic

Hypothesis testing in IM
IM_hypotest is a program that performs various hypothesis tests using the surface files from IM as input.
The tests include:
1) Different effective population sizes (q1 and q2 )
2) Different immigration proportions (m1 and m2 )
3) Different effective numbers of immigrants (M1 and M2 )
I will talk about how to use MCMC data files to rescale parameter estimates and to perform hypothesis tests (again generally applicable to any MCMC analysis).

Comments from student evaluations

Clarity of the presentation of the material: 1(poor)-5(excellent). 5 – one of the best I’ve seen.
Thanks a lot for the workshop! It clarified a lot of things for me!

Very good!! I am really glad I came down and got some valuable nuggets of info concerning MCMC. I also heavily appreciate the Slides and the R code!!!

Thank you so much for taking the time to prepare and give this workshop! On the whole, the MCMC background material was very clear and straight forward. I have never used IM or MIGRATE so I had a little trouble following the second half of the course. I do feel like I have a much better idea of the how the program works and the notes from the course will be a great reference if/when I do use those software packages. This part of the course was still very worthwhile for me if only to better understand papers using this analysis.

Thanks again!

Great job. I think the main suggestion is that more time is needed to get through all the material. And more examples in the concrete world. And to lay out the components of the program in baby terms. Maybe another time, 3 hours in the morning, 3 hours in the afternoon. This will allow longer breaks, more questions, and more time to get through it all.

How would you rate your understanding of this material before the course: 1(never thought about it) – 5(expert)? Before the course I would have said 4 or 5. During and after the course I was shocked by the number of alternative strategies for evaluating convergence. Many of the methods discussed have not been applied to Bayesian phylogenetic analysis yet – although I’m not sure that they could be. This section was terrific.

Just a terrific course! My only advice was again the short intro to IM. Thanks a million!

...I understand many of the concepts much better and feel like I could perform some runs all by myself... Thanks for all your hard work, next one I am pretty sure you need to make people pay for! Great job!

Thanks for a terrific workshop. You really did a dynamite job with it..

Great job! I learned a ton. Not only with respect to IM, but the information applies to other areas too (phylogenetics).

For someone at my level (user of MCMC programs, comfortable with terminology but not sure where to go next), the practical aspects of the course were particularly helpful (the tuning & convergence bits for example). I was a little perplexed about how to implement some of the stuff discussed in the IM context into other MCMC programs, but I feel like at least now I have an idea of what is important to do.

Great workshop. A lot of information, it would be worthy to split it in two days, and include more empirical examples.

How would you rate your understanding of this material after the course: 1(more confused)-5(major strides)? 5 – Major strides for sure – I was naive to IM so this material was all new. I’m used to thinking about Bayes Factors rather than mass of posterior probability on alternative parameter estimates – very cool!

I think the workshop was really useful, thank you so much for that!
Although I think the "public" was really heterogenic: some like me are now starting to think in IM and other programs that will be useful in the future (I don't even have data to apply now), but in the other hand other persons are much more advanced and probably they were looked for solutions for very specific problems.
So, the introductory part was excellent for me, and I've learned a lot, but the Tuning and convergence part were too specific for me, because I never thought about that.