ABSTRACT
Correlation can be used to detect tr eatment effects in ways other than by correlating treatment levels and data. In fact, the tr eatment levels need not even be quantitatively or dered. Pitman (1937) described a randomization test for correlation for a hypothetical agricultural experiment involving
n
treatments applied to blocks A and B, each of which was subdivided into
n
plots. With independent random assignment of the
n
treatments to the
n
plots within the blocks, ther e are (
n
!)
possible assignments, but for pairing A and B plots administered the same treatment there are only
n
! distinctive pairings. (Each of the
n
! distinctive pairings could be manifested in
n
! ways of permuting treatment designations among the pairs.) Under H
, the measurements for the ordered first, second,
…
n
th plots within A and also within B were fixed r egardless of the assignment, so getting a high corr elation between A and B plots administered identical treatments would be the result of happening to assign the treatments in such a way as to pair similar yields.
