Some problems in the design and analysis of field experiments subject to inter-plot interference

Abstract

Results from field experiments with mobile pests and air-borne pathogens are subject to bias as a result of inter-plot interference. Serially balanced designs (SBDs) allow interference to be estimated but other designs may be better for decreasing such effects. To investigate this, systematic replicated designs, comparing sprays applied at different times to control powdery mildew of spring barley, were sited in 2 years alongside SBDs testing similar treatments. Yields of grain and assessments of mildew on the leaves were analysed. Results from the balanced designs provided strong evidence of interference in both years but not in a third (when the systematic design was omitted). Estimates of treatment effects from the systematic designs were often, but not consistently, greater than corresponding estimates from the SBDs. A method of analysis from Draper & Guttman (1980) was adapted to produce estimates of the differences between treatments as if applied to all plots of an experiment; this showed larger differences between treatments than the conventional analysis in 1975 and 1976 (when there was appreciable interference), but failed in 1977 when interference was slight. This method fails when applied to the systematic designs; SBDs (which are a subset of all designs in randomized blocks) are probably optimal for this type of analysis. The difficulties of analysing data in the form of percentages or proportions (with consequent non-orthogonality) are discussed.