How to calculate the mean of two log-likelihood values.

#1
Hi All,

I'm confused how to calculate the mean of two (natural) log-likelihood values. The software I'm using (MrBayes) does this for me automatically... sometimes. But in situations when it does not, I need to calculate it myself.

Here is an example:
log-likelihood A: -104520.13
log-likelihood B: -104528.54
software's mean: -104520.83

Here is a different example:
log-likelihood A: -104506.97
log-likelihood B: -104523.28
software's mean: -104507.66

Clearly the mean is not calculated by (A+B)/2.

Thanks much for the advice!
-Steven
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
Does the software have documentation and yes there seems to be some type of weighting going on. Are the inputs for the two log-likelihoods different?
 
#3
Does the software have documentation and yes there seems to be some type of weighting going on. Are the inputs for the two log-likelihoods different?
Hi! Thanks for the reply.

There is plenty of documentation. MrBayes is free software to do phylogenetics. The documentation is here: http://mrbayes.sourceforge.net/download.php

Unfortunately, I see no report on how they get this mean.

These log-likelihoods are values which are summed from the stepping-stone technique used to estimate marginal likelihoods from an mcmcmc run. With two independent runs with 50 stepping stone samples each... the two sums are reported always, and the mean is calculated for you sometimes (not when you are doing more complex tests).

Any thoughts are greatly appreciated.

Cherers,
Steven
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
Well they are logged values but seem off from the harmonic mean. Could there be rounding error, probably not. I would contact the software providers if you can figure it out.