Jerry suggested
>as a general rule, the
>power of the test improves through a sample of 25 (bees, light bulbs,
>etc.).  Then you hit the old issue of diminishing returns.  The amount of
>improvement in the power or reliability of the test begins to fall off
>rapidly between 25 and 30, and sample sizes over 30 don't add much for the
>time invested.

        As a scientist who ruthlessly evaded statistics in his education, I
nevertheless make bold to suggest a handy rule of thumb:
                the  power or reliability of the test improves with,
roughly, the square root of the sample number.

        e.g.  in order to achieve a ten-fold decrease in your uncertainty
you have to take 100-fold larger sample.

R