We only considered sampling location with at least 10 samples,
reducing the collapsed community matrix to 83 rows. The sampling
effort per location varied over three orders of magnitude, from the
low-end cut-off of 10 samples to over 1000 samples in the Gulf of
Finland, the Baltic Sea. To account for this vastly different sampling
effort, we averaged the biomasses over 20 randomly drawn samples
per location (or all samples when a station had less then 20
samples). This could create another bias, with locations with high
sampling effort seriously under-sampled. As a compromise we
repeated the random sampling five times, giving better representation
to locations with high sampling effort, but still retaining
equal sample size of the basic analysis units. The final community
matrix had 415 rows, made up of five strata corresponding to the
five random draws, each stratum with 83 locations.