It looks like this is a fairly well understood problem because in the
real world it's also often cheaper to speak to people in a small
geographic area or time interval too. These wikipedia pages sound
interesting and have some external references:
http://en.wikipedia.org/wiki/Cluster_samplinghttp://en.wikipedia.org/wiki/Multistage_sampling
I suspect the hard part will be characterising the nature of the
non-uniformity in the sample generated by taking a whole block. Some
of it may come from how the rows were loaded (e.g. older rows were
loaded by pg_restore but newer rows were inserted retail) or from the
way Postgres works (e.g. hotter rows are on blocks with fewer rows in
them and colder rows are more densely packed).
I've felt for a long time that Postgres would make an excellent test
bed for some aspiring statistics research group.
--
greg