Inter-plot interference and nearest-neighbour analysis of field experiments

Abstract

SUMMARYResults obtained from the plots of a field experiment may be subject to’ representational errors'; that is, they may not exactly reproduce the results that would be expected if the same treatments were applied to whole fields. In particular, the treatment applied to one particular plot often has the potential to affect the adjacent plots; such an effect may be positive or negative according to circumstances, and its magnitude cannot easily be estimated. This paper shows, on theoretical grounds and by simulations, that nearest-neighbour (NN) analysis cannot be relied upon to mitigate the effects of such inter-plot interference, even when the residual variance of the NN analysis is substantially, and by normal criteria significantly, smaller than that of conventional (randomized block) analysis. In certain circumstances, NN analysis of data affected by interference produces greater average bias in estimated treatment differences than conventional analysis.