The prevalence of Escherichia coli O157 displays striking variability across the Scottish cattle population. On 78% of farms, in a cross-sectional survey of 952, no shedding of E. coli O157 was detected, but on a small proportion, approximately 2%, very high prevalences of infection were found (with 90-100% of pats sampled being positive). We ask whether this variation arises from the inherent stochasticity in transmission dynamics or whether it is a signature of underlying heterogeneities in the cattle population. A novel approach is taken whereby the cross-sectional data are viewed as providing independent snapshots of a dynamic process. Using maximum-likelihood methods to fit time-dependent epidemiological models to the data we obtain estimates for the rates of immigration and transmission of E. coli O157 infection - parameters which have not been previously quantified in the literature. A comparison of alternative model fits reveals that the variation in the prevalence data is best explained when a proportion of the cattle are assumed to transmit infection at much higher levels than the rest - the so-called super-shedders. Analysis of a second dataset, comprising samples taken from 32 farms at monthly intervals over a period of 1 year, additionally yields an estimate for the rate of recovery from infection. The pattern of prevalence displayed in the second dataset also strongly supports the super-shedder hypothesis.