negative LRT

[Anna Faddeeva]

Greetings, I perform branch site test in codeml to estimate positive selection in the branches of interest. I have negative LRT in approximatelly 10% of cases, mostly they are below -1 but some of them are extreme negative, does anybody experience this problem?

Is it correct to approximate negative LRT values to 0 and then correct them with qvalues?

Thank you in advance.

Anna

[Anna Faddeeva]

I also runned codeml 3 times and took the MAX value, but still have these cases..

[Ziheng]

The b-s model sometimes suffer from numerical difficulties. If you run each model several times with differnt initial values, the setting the negative lnL difference to zero should be fine,. In theory those negative values should not be very negative. Otherwise you can try to use in.codeml to start the alternative model with initial values a that are very close to the MLE under the null model.
Ziheng.