Dear Hajar,

jmodeltest  tests a huge range of possible models. Most of them you'll never going to use or see being used. I would rather recommend you use MrModeltest which tests for all models possible in MrBayes, which also includes the most commonly used models and you can decide based on the AIC or BIC criterion which model best fits (mostly both criteria suggest the same model). The nice thing is that you also directly get the code for implementing  the best model in MrBayes. Here is a short outline on how to use the program:

https://rodybio.wordpress.com/2014/02/07/mrbayes-paup-mrmodeltest-on-windows/

Anyways, in most cases you anyways have an GTR or HKY model as most suitable for your data. What does jmodeltest suggest? And what do you mean with inconsistencies? When you run it several times for the same alignment?

Best,

Belinda

I like Belinda's answer, but if you get two different models, I advise you to go with the more parametric one. Yes, that may increase variance, but under parametrization (is that a word? haha)  may increase your error in accuracy. 

Thanks Silvia, though I would rather take the model with less parameters since as you said, every additional parameter included might increase your overall error (estimating parameters is inevitably associated with some errors). Thus, you always try to use the least complex model that still best fits your data.

I agree with Artur that you should in the end decide what model is most reasonable. Still, I don't see why you cannot use MrModeltest when you in the end go for a Max Likelihood approach? MrModeltest is just performing model testing and gives you finally some model suggestions according to the AIC or BIC. If you would like to do ML and go for one of the most commonly used softwares RaxML, than you can anyways only use a modified GTR model.

I was also working with COI sequences before and I usually got GTR (+G+I) or HKY as most suitable model, which fits to what other people use in their analyses of COI. Of course, the most suitable model always depends on the organisms you are working with and the mutation rate, but for example for insects the mutation rate of COI across many taxa is quite similar.

Hi Everyone,

Many thanks for your clear and thorough explanations. As I mentioned before, I have just started using JModeltest and was expecting to get the same best fit regardless of the criterion (AIC/BIC/DT). But, what happened is that each criterion returns a different best fits. So in total I have got 3 models to choose from. Is this inconsistent? or I should expect to get the same best fit. this is what I am getting:

DT suggested TrN+G

BIC suggested HKY+G

AIC suggested TIM3+I+G

Many thanks for your help

I found MEGA to be very useful in this regard, the model can be quickly ascertained. 

Following which you can run ML without much hassle. 

Hi Hajar,

thats what I meant, you also got suggestions for models (TrN, TIM3) that are hard to implement in the tree reconstruction (MrBayes for example) and also hardly beeing used. How different are the different models TrN and TIM3 from each other and to HKY? I mean, what parameters do they consider? If these models are not too different from each other and HKY is also among the top 3 models according to AIC and DT I would go for that one.

Hi Artru,

Sorry for the late reply and thank you for these useful tips.

Regards

Hajar