Pathotypes of Erysiphe graminis f.sp. hordei were monitored at fortnightly intervals in pure and mixed stands of spring barley during the course of mildew epidemics in two field trials. Mixtures were composed of cultivars with Arabische (gene Mla12), Laevigatum (Ml(La)), and Monte Cristo (Mla9) resistance, respectively. The three-way mixtures were either random or, in 1989, laid out as one-row mixtures (i.e., regularly alternating rows of different genotypes) or three-row mixtures (i.e., regularly alternating three-row strips of different genotypes), respectively. In 1990 only random mixtures and six-row mixtures were compared with pure stands. The virulence complexity (i.e., the average number of virulence factors per isolate with reference to Mla12, Ml(La), and Mla9) was always maximal in the random mixtures. In 1989, linear regression of complexity on mildew generations gave significant b-values (slopes) of 0·049, 0·031, and 0·025 in the random mixture, one-row mixture, and three-row mixture, respectively; the b-value from pure stands was not significant. In 1990, another sampling technique allowed selection to be observed on each genotype in the mixtures separately. In the random mixture b-values were 0·048, 0·064 and 0·017 (not significant) on Mla12, Ml(La), and Mla9 cultivars respectively. In six-row mixtures and in pure stands, there was no significant increase in complexity (b > 0) on any of the mixture components. Although the frequency and relative fitness of complex pathotypes were higher in all types of mixtures than in pure stands, selection towards complex races was much less intense in row mixtures than in random mixtures in both field trials.