- Aims: (1) Understanding how the relationship between species richness and its determinants depends on the interaction between scales at which the response and explanatory variables are measured. (2) Quantifying the relative contributions of local, intermediate and large-scale determinants of species richness in a fragmented agro-ecosystem. (3) Testing the hypothesis that the relative contribution of these determinants varies with the grain size at which species richness is measured. Location: A fragmented agro-ecosystem in the Southern Judea Lowland, Israel, within a desert–Mediterranean transition zone. Methods: Plant species richness was estimated using hierarchical nested sampling in 81 plots, positioned in 38 natural vegetation patches within an agricultural matrix (mainly wheat fields) among three land units along a sharp precipitation gradient. Explanatory variables included position along that gradient, patch area, patch isolation, habitat heterogeneity and overall plant density. We used general linear models and hierarchical partitioning of variance to test and quantify the effect of each explanatory variable on species richness at four grain sizes (0.0625, 1, 25 and 225 m2). Results: Species richness was mainly affected by position along a precipitation gradient and overall plant density, and to a lesser extent by habitat heterogeneity. It was also significantly affected by patch area and patch isolation, but only for small grain sizes. The contribution of each explanatory variable to explained variance in species richness varied with grain size, i.e. scale-dependent. The influence of geographic position and habitat heterogeneity on species richness increased with grain size, while the influence of plant density decreased with grain size. Main conclusions: Species richness is determined by the combined effect of several scale-dependent determinants. Ability to detect an effect and effect size of each determinant varies with the scale (grain size) at which it is measured. The combination of a multi-factorial approach and multi-scale sampling reveals that conclusions drawn from studies that ignore these dimensions are restricted and potentially misleading.