The dataset consists of an alignment of 63 Hepatitis C sequences sampled in 1993 in Egypt (Ray et al., 2000). This dataset has been used previously to test the performance of skyline methods (Drummond et al., 2005; Stadler et al., 2013).

Choosing the dimension for the Bayesian Skyline can be rather arbitrary. If the dimension is chosen too low, not all population size changes are captured, but if it is chosen too large, there may be too little information in a segment to support a robust estimate. When trying to decide if the dimension is appropriate it may be useful to consider the average number of informative (coalescent) events per segment. (A tree of n n n taxa has n − 1 n-1 n − 1 coalescences, thus N e N_e N e ​ in each segment is estimated from on average n − 1 d \frac{n-1}{d} d n − 1 ​ informative data points). Would this number of random samples drawn from a hypothetical distribution allow you to accurately estimate the distribution? If not, consider decreasing the dimension. If the difference is less than 2, the hypotheses may not be distinguishable – in terms of Bayes factors, are barely worth mentioning. Is NS faster than path sampling/stepping stone (PS/SS)? Thanks also to all the speakers for agreeing to take part and making the long journey to New Zealand!

Say, we have two models, M1 and M2, and estimates of the (log) marginal likelihood, ML1 and ML2, then we can calculate the Bayes factor, which is the fraction BF=ML1/ML2 (or in log space, the difference log(BF) = log(ML1)-log(ML2)). If BF is larger than 1, model M1 is favoured, and otherwise M2 is favoured. How much it is favoured can be found in the following table (Kass & Raftery, 1995): Figure 1: Bayes factor support. The choice of the number of dimensions can also have a direct effect on how fast the MCMC converges ( Figure 14). The slower convergence with increasing dimension can be caused by e.g. less information per interval. To some extent it is simply caused by the need to estimate more parameters though. Figure 14: The ESS value of the posterior after running an MCMC chain with 1 0 7 10 An NS analysis produces two trace log files: one for the nested sampling run (say myFile.log) and one with the posterior sample ( myFile.posterior.log). It has already been more than two weeks since the second Taming the BEAST workshop took place on Waiheke island in New Zealand. In practice, we can get away much smaller sub-chain lengths, which you can verify by running multiple NS analysis with increasing sub-chain lengths. If the ML and SD estimates do not substantially differ, you know the shorter sub-chain length was sufficient. How many particles do I need?

Nested sampling stops automatically when the accuracy in the ML estimate cannot be improved upon. Because it is a stochastic process, some analyses get there faster than others, resulting in different run times. Why are the ESSs so low when I open a log file in Tracer? So, the main parameters of the algorithm are the number of particles N and the subChainLength. N can be determined by starting with N=1 and from the information of that run a target standard deviation can be determined, which gives us a formula to determine N (as we will see later in the tutorial). The subChainLength determines how independent the replacement point is from the point that was saved, and is the only parameter that needs to be determined by trial and error – see FAQ for details. To change the number of segments we have to navigate to the Initialialization panel, which is by default not visible. Navigate to View > Show Initialization Panel to make it visible and navigate to it ( Figure 7).

Sign up to our newsletter

SCOTTI Tutorial: NEW Reconstruct transmission trees using within-host data with an approximate structured coalescent. Kass, R. E., & Raftery, A. E. (1995). Bayes factors. Journal of the American Statistical Association, 90(430), 773–795. replace the point with a new point randomly sampled from the prior using an MCMC chain of subChainLength samples under the condition that the likelihood is at least L min The third edition of Taming the BEAST took place last week at the London School of Hygiene and Tropical Medicine. I had a lot of fun at the workshop in London and also learned a lot. Thanks to a great bunch of participants who came out to learn about BEAST2 with us, I hope you enjoyed the workshop! Note that since BEAST 2.7 the filenames used here are the default filenames and should not need to be changed!)

The workshop organisers and participants outside of the London School of Hygiene and Tropical Medicine. There are two ways to save the analysis, it can either be saved as a *.pdf for display purposes or as a tab delimited file.

NS works in theory if and only if the points generated at each iteration are independent. If you already did an MCMC run and know the effective sample size (ESS) for each parameter, to be sure every parameter in every sample is independent you can take the length of the MCMC run divided by the smallest ESS as sub-chain length. This tend to result in quite large sub-chain lengths. In this tutorial we will estimate the dynamics of the Egyptian Hepatitis C epidemic from genetic sequence data collected in 1993. Press OK to reconstruct the past population dynamics ( Figure 11). Figure 11: Reconstructing the Bayesian Skyline plot in Tracer. Marginal likelihood: -12428.557546706481 sqrt(H/N)=(11.22272275528845)=?=SD=(11.252847709777592) Information: 125.94950604206919

There are descendants of the coalescent skyline in BEAST that either estimate the number of segments (Extended Bayesian Skyline (Heled & Drummond, 2008)) or do not require the number of segments to be specified (Skyride (Minin et al., 2008)), but instead makes very strong prior assumptions about changes in N e N_e N e ​ . Exploring the results of the Coalescent Bayesian Skyline analysis However, the only informative events used by the Coalescent Bayesian Skyline plot are the coalescent events. Thus, using a maximally-flexible parameterization with only one informative event per segment often leads to erratic and noisy estimates of N e N_e N e ​ over time (especially if segments are very short, see Figure 6). Grouping segments together leads to smoother and more robust estimates.We will be using R to analyze the output of the Birth-Death Skyline plot. RStudio provides a user-friendly graphical user interface to R that makes it easier to edit and run scripts. (It is not necessary to use RStudio for this tutorial). It may be tempting to specify the maximum dimension for the model (each group contains only one coalescent event, thus N e N_e N e ​ changes at each branching time in the tree), making it as flexible as possible. This is the parameterization used by the Classic Skyline plot (Pybus et al., 2000), which is the direct ancestor of the Coalescent Bayesian Skyline plot.