No, I do at least a two part filtration. First, I drop anything with counts below my rarefaction depth because those are deemed “bad quality” samples for thsi study’s definition of bad quality.
Then, I double check my samples in PCoA space and may drop samples which do not cluster, period. I may also filter at this step to spit or remove samples that aren’t relevant to my current analysis. (For example, sometimes people will send samples about both and but only want to look at so then we filter.)
Third, before feature-based analysis, I filter my table to get rid of anything with less than (1/rarefaction depth) in less than 10% of my communities. My suggestion here, since the joint filtering isnt implemented in qiime (yet… its on my list), to first filter out any feature present in fewer than 10% of your samples and see where that gets you.
This decreases the over all number of features you test while discarding things that are likely either noise or underpowered.