Soil erosion by water (rain and wash) is a discrete process, and there are several questions relating to this that have yet to be answered. For instance, for how long should measurements be taken in order to obtain realistic mean erosion rates, and what is the effect of the largest daily events on soil erosion, are questions still not fully answered. To make progress on these issues, detailed information is needed, but usually not available. This work has analysed the USLE database compiled by the United States Department of Agriculture. A total of 27 857 daily erosion events were examined, monitored over 310 erosion plots representing 3195 plot‐years of soil erosion measurements. Periods of measurement varied between plots, ranging from 2 years to 32 years. Data have been analysed by calculating the percentage of soil eroded by the largest to the tenth largest daily events over the entire length of the record. The percentage of soil eroded during the n‐largest events follows a power function y = x−b, where y is the percentage of soil eroded in the selected n‐largest event, and x the total of measured daily events. Results showed that the top 10% of total daily erosive events produce a mean of 50% of eroded soil. Soil erosion measured over short periods, e.g. typically less than 5 to 10 years, is compressed into a few daily events, whatever their magnitude; therefore, mean erosion rates estimated under such small time frames cannot be taken as a good descriptor of the real processes. Analysis suggests that to weight the dependence of soil erosion on the largest daily values, a minimum number of events should be considered. In general, if it is wished to reduce the effect of the largest events to less than 50% of total soil erosion, the mean number of recorded daily events should be between 75 and 100. They represent, on average, 10 years of measurements for the database used here. Copyright © 2009 John Wiley & Sons, Ltd.