I've always wondered if there are others like me who run multiple iterations with gradiently increasing/decreasing trimming parameters.
Basically, I do a bash loop to combine all possible trimming parameters by an increment of 5. So something like forward reads trimmed (250, 255, 260, and so on) each combined with reverse reads trimmed (180, 185, 190, and so on, usually much shorter).
Then my loop exports the content of the stats.qza file and grabs a few samples from the table and merges them into one final table output to compare retention rates. There are usually two very clear thresholds: the lower one, where the retention drops to nearly 0%, due to overtrimming and not enough overlap (the 12bp overlap minimum from dada2) and the sweet spot, where nearly 50-70% of merged reads are retained (exact percentage depends on the sequencing quality of the run), after which a slow, gradual reduction of the retention rate begins, because longer fragments=lower avg. quality. For example, 65% (highest) at 260-180 (420bp insert + 12bp overlap), 64.1% at 265-175, or 63.9% at 265-180, and 0.5% at 255-175 (too short, no overlap, dada2 throws those out).