Project #1 update

It's been brought to my attention that some of you are getting inconsistent results from Structure. Specifically, the folks from Beach Hall report that some of them find the following for the "Estimated Ln Prob of Data":

• K=7: -2725.6


• K=8: -3093.4

Others get

• K=7: -2723.3


• K=8: -2698.4

This illustrates a property of MCMC implementations of Bayesian analyses I'd hoped to avoid for awhile. Let me give a brief explanation here, and we'll talk more about it when we return to estimating migration rates (with Structure) in a week or two.

The calculations that Structure does have a random element. If you run the analyses long enough, that random element will disappear. In this case, the close agreement for the K=7 results suggests that a longer run will give the same results. (You could try this and check it to see.) The disagreement between the K=8 results suggests that a longer run will give results different from one or both of the ones you've tried. My bet is that a long run will settle on the one with P(data) = -2698.4.

For purposes of Project #1, you may want to try longer sample runs than what I initially suggested (if you have time), but I won't require that. Just make sure that you include the numerical results that caused you to chose the K you did, so that I can verify that you're doing that part of the analysis correctly.

Posted by Kent at February 16, 2004 3:52 PM | TrackBack