Tracer Bayes Factor output table

[JS]

Hi,
I have calculated Bayes Factor (BF) in Tracer but I am having trouble
understanding the output table.
Is the sample listed on the top most row the denominator or numerator
of the BF equation number 5 in Suchard et al (2001)?
Thank you,
Joana

[joe parker]

hi all,

i have a similar problem

i'm not sure how to interpret the pairwise BF scores in tracer 1.4
either. by hand i get similar, correlated, but different values...

if we:
* number models under test Hi from (H1...Hn),
* calculate the Bayes' factor B supporting Hi compared to Hi+1 as
either (Hi+1 / Hi) or if using log-likelihoods, (Hi+1 - Hi)

following Suchard eA (2001), then compare to the output in the tracer
table, I find that the values I get are correlated with (R-squ by
nasty excel == 1) those given in the pairwise table; but different in
magnitude.. what's going on (always scaled by 2.3)? I can't seem to
convert my set into the Tracer set by logging or raising to 2 or e or
10 ((deci?)bits or (deci?)bans) ..?


i can however get sort of close to the tracer values as:

tracerBF_Hi = myBF_Hi * (log(base Hi harmonic mean
likelihood)*myBF_Hi)

where myBF_Hi = harmonic mean likelihood Hi+1 - harmonic mean
likelihood Hi


any help anyone? i've reached the limit of my guestimathing!
sample data below. given that it doesn't change the conclusions
regarding model selection i'm not too worried, but it would be nice to
know what the units in beast are.

thanks in advance, joe



sample data (only 1st column of pairwise comparison):

model log harmonic mean likelihood (from tracer) my comparison (hi+1 -
hi for i=1:n) pairwise tracer output for h1 vs. hn factor (mine/
tracer)
GGISB.log -6779.791 - - -
GGISC.log -6780.276 -0.485 -0.211 2.298578199
GGISE.log -6780.336 -0.545 -0.237 2.299578059
GGIUB.log -6742.329 37.462 16.27 2.302519975
GGIUC.log -6742.315 37.476 16.276 2.302531334
GGIUE.log -6742.633 37.158 16.137 2.302658487
GGxSB.log -6791.475 -11.684 -5.075 2.30226601
GGxSC.log -6791.641 -11.85 -5.147 2.302312026
GGxSE.log -6791.672 -11.881 -5.16 2.30251938
GGxUB.log -6753.976 25.815 11.211 2.302649184
GGxUC.log -6753.456 26.335 11.437 2.302614322
GGxUE.log -6754.325 25.466 11.059 2.30273985
GxISB.log -6828.435 -48.644 -21.126 2.302565559
GxISC.log -6828.817 -49.026 -21.292 2.30255495
GxISE.log -6828.553 -48.762 -21.177 2.302592435
GxIUB.log -6793.119 -13.328 -5.789 2.302297461
GxIUC.log -6792.732 -12.941 -5.62 2.302669039
GxIUE.log -6793.203 -13.412 -5.825 2.30248927
GxxSB.log -7075.298 -295.507 -128.337 2.30258616
GxxSC.log -7076.137 -296.346 -128.702 2.302574941
GxxSE.log -7075.617 -295.826 -128.476 2.302577913
GxxUB.log -7039.18 -259.389 -112.652 2.302568973
GxxUC.log -7039.329 -259.538 -112.716 2.302583484
GxxUE.log -7039.34 -259.549 -112.721 2.302578934
HGISB.log -6792.945 -13.154 -5.713 2.302468055
HGISC.log -6793.517 -13.726 -5.961 2.302633786
HGISE.log -6793.322 -13.531 -5.877 2.302365152
HGIUB.log -6755.857 23.934 10.394 2.30267462
HGIUC.log -6755.172 24.619 10.692 2.302562664
HGIUE.log -6755.428 24.363 10.581 2.302523391
HGxSB.log -6805.121 -25.33 -11.001 2.302517953
HGxSC.log -6805.535 -25.744 -11.18 2.302683363
HGxSE.log -6805.538 -25.747 -11.182 2.302539796
HGxUB.log -6767.437 12.354 5.365 2.302702703
HGxUC.log -6767.544 12.247 5.319 2.30250047
HGxUE.log -6768.207 11.584 5.031 2.302524349
HxISB.log -6840.393 -60.602 -26.319 2.302595083
HxISC.log -6840.208 -60.417 -26.239 2.302564884
HxISE.log -6839.993 -60.202 -26.146 2.302531936
HxIUB.log -6805.837 -26.046 -11.312 2.302510608
HxIUC.log -6805.689 -25.898 -11.247 2.302658487
HxIUE.log -6806.229 -26.438 -11.482 2.30256053
HxxSB.log -7095.653 -315.862 -137.177 2.302587168
HxxSC.log -7095.758 -315.967 -137.223 2.302580471
HxxSE.log -7095.356 -315.565 -137.048 2.302587415
HxxUB.log -7060.666 -280.875 -121.982 2.302593825
HxxUC.log -7060.371 -280.58 -121.855 2.30257273
HxxUE.log -7060.745 -280.954 -122.017 2.302580788







On May 21, 1:50 am, JS <joanaza...@gmail.com> wrote:
> Hi,
> I have calculatedBayesFactor(BF) in Tracer but I am having trouble

[Marc Suchard]

Hi!

The difference that you are observing is because the log-base you are
using is different that what Tracer is outputing. I must add a
disclaimer -- I did not write the Tracer code, so I am guessing here.

Tracer outputs numbers in two log-bases: base E or base 10. So, must
likely, Tracer is reporting "log_E marginal likelihood" estimates.
Then, the differences that you are taking are reporting the "log_E
Bayes factor"; while the differences that Tracer is reporting are
"log_10 Bayes factor".

When converting a number X_1 = log_E Y in base-E to X_2 = log_10 Y in
base-10, we see that X_1 / X_2 = log_E 10 = 2.3.

I personally prefer reporting log_10 Bayes factors, as a value of 1
means one model is 10^1 = 10 times more likely, a value of 2 means one
model is 10^2 = 100 times more likely, etc.

best, Marc



On Jun 6, 3:44 pm, joe parker