First of all, TN93 is just one level of complexity lower than the GTR general time reversible model.  For nearly all data set I have observed, ModelTest will select the GTR model if the alignment is more than 1,000 bases long.  TN92 (Tamura-Nei 1992) is more complex than Kimura two-parameter and slightly less complex than TN93.

It is pretty easy to build trees using several different models and compare them to see what difference the different models make.  Also you can use different methods (neighbor-joining, Maximum likelihood, Bayesian, parsimony) with the same models and see what difference that makes.   In general, changing the input data such as using longer sequences, mitochondrial or chloroplast DNA vs nuclear genes, including an outgroup or not, has a FAR larger impact on the tree, than changing any model or method.

With many phylogenetic packages, such as IQ-tree with has a web service at

http://iqtree.cibiv.univie.ac.at/

you can run the GTR model and look in the output log for the full matrix of rates that were calculated form your data, and then decide whether or not the values for C to T vs G to A transitions for example are different enough to require a more complex model than Kimura (transitions and transversions different, but all transitions get the same weight).

See attached JPG screenshots (marked up) for examples.

Hello Brian,

Thank you for your advice, it is much appreciated! ...actually, you are absolutely correct, I mistyped earlier and I used Mega4 for model selection on this dataset (I'd intended to use modeltest but had difficulty getting this particular dataset into it). Considering your comment, I should perhaps persist with modeltest to see if/how the model selection differs (it sounds like it will).

My alignment is 1824bp long, so perhaps the model selection will make little difference on topology, but I did want to select a model that I could justify. I'd intended to build trees using both ML and Bayesian methods, then compare results using a Shimodaira-Hasegawa test. It might also be worth trying a model that differs in complexity from the 'best fit' model to see how much of an impact it would have.

Thank you for pointing me toward IQ-Tree. I hadn't been aware of that, and will definitely take a look.

Once again, thank you for the input!

Best wishes,

martin

Including or not including rate variation between sites (invariant sites plus a gamma distribution of rates among variable sites) makes a larger difference than model changes for protein coding regions or structural RNA regions. 

http://www.hiv.lanl.gov/content/sequence/findmodel/findmodel.html

Rob Lanfear's partitionfinder program is also worth using for larger data sets. 

http://www.robertlanfear.com/partitionfinder/

Data partitions allow you to use a different set of model criteria for different subsets of your data.  One common way of partitioning protein-coding DNA data is to treat the third codon position (often "silent" as far as amino acid change goes) differently than first and second codon positions.

Thank you Brian, I really appreciate your guidance on this.