Linearity of basin runoff and peak response as a function of watershed scale was examined for a set of 29 nested semiarid watersheds within the U.S. Department of Agriculture-Agricultural Research Service Walnut Gulch Experimental Watershed, located in southeastern Arizona. Watershed drainage areas range from 1.83¿103 to 1.48¿108 m2 (0.183--14800 ha), and all stream channels are ephemeral. Observations of mean annual runoff, database-derived 2- and 100-year peak runoff rates, ephemeral channel area, and areal rainfall characteristics derived from 304 events were examined to assess the nature of runoff response behavior over this range of watershed scales. Two types of distributed rainfall-runoff models of differing complexity were applied to a subset of the watersheds to further investigate the scale-dependent nature of the collected data. Contrary to the conclusions of numerous studies in more humid regions, it was found that watershed runoff response becomes more nonlinear with increasing watershed scale, with a critical transition threshold area occurring roughly around the range of 3.7¿105 to 6.0¿105 m2 (37--60 ha). The primary causes of increasingly nonlinear response are the increasing importance of ephemeral channel losses and partial storm area coverage. The modeling results indicate that significant error will result in model estimates of peak runoff rates when rainfall inputs from depth area-frequency relationships are applied beyond the area of typical storm coverage. For runoff modeling in Walnut Gulch and similar semiarid environments, explicit treatment of channel routing and transmission losses from channel infiltration will be required for watersheds larger than the critical drainage area. ¿ 1997 American Geophysical Union