Cluster vs Cluster Split

I am currently processing MiSeq data and I noticed something interesting. I ran 20 samples using the traditional cluster method and then calculated bray Curtis between them. I also did this same process at the class, order, and species level with cluster.split (taxlevel 3,4,7 respectively). As expected, when going from traditional methods (what i presume is at the sequence level) to class level, the dissimilarities between libraries decreased. What I did not expect was that as you go to higher tax levels using the cluster split command, I was seeing decreasing dissimilarities, ending with the species level having the lowest dissimilarity values and the traditional cluster method having the highest. Can someone explain why this is happening? Am I losing information if I use the cluster.split command exclusively?
Thank you!

Can you post the exact commands you are entering?

For the traditional method, I use the following two commands:

mothur > dist.seqs(fasta=FILE.fasta, cutoff=0.30, processors=8)

mothur > cluster(column=FILE.dist, name=File.names, method=average)

For the cluster split methods I use:

mothur > classify.seqs(fasta=FILE.fasta, template=silva.file.fasta,

mothur > cluster.split(fasta=File.fasta, count=File.count_table,, splitmethod=classify, taxlevel=4, cutoff=0.03,processors=8)

With tax levels at 3, 4,and 7. Are those correct for the class, order, and species level?

I wouldn’t use the silva.bacteria.fasta taxonomy files - instead, i’d use the newer silva.nr_119 files that i’ve posted to the wiki or the rdp or greengenes files. I also wouldn’t use a cutoff in the cluster.split step.

for cluster.split here are the levels…

No database has very reliable species-level names.