A reliable estimation of primary production of terrestrial ecosystems is often a prerequisite for carrying out land management, while being important also in ecological and climatological studies. At a regional scale, grassland primary production estimates are increasingly being made using satellite data. In a currently used approach, regional Gross, Net and Above-ground Net Primary Productivity (GPP, NPP and ANPP) are derived from the parametric model of Monteith and are calculated as the product of the fraction of incident photosynthetically active radiation absorbed by the canopy (fAPAR) and gross, net and above-ground net production (radiation-use) efficiencies ((epsilon) g, (epsilon) n, (epsilon) an); fAPAR being derived from indices calculated from satellite measured reflectances in the red and near infrared. The accuracy and realism of the primary production values estimated by this approach therefore largely depend on an accurate estimation of (epsilon) g, (epsilon) n and (epsilon) an. However, data are scarce for production efficiencies of semi-arid grasslands, and their time and spatial variations are poorly documented, leading to often large errors on the estimates. In this paper a modeling approach taking into account relevant ecosystem processes and based on extensive field data, is used to estimate sub- seasonal and inter-annual variations of (epsilon) g, (epsilon) n and (epsilon) an of a shortgrass site of Arizona, and to quantitatively explain these variations by these of plant water stress, temperature, leaf aging, and processes such as respiration and changes in allocation pattern. For example, over the 3 study years, the mean (epsilon) g, (epsilon) n, and (epsilon) an were found to be 1.92, 0.74 and 0.29 g DM (MJ APAR)-1 respectively. (epsilon) g and epsilonn exhibited very important inter- annual and seasonal variations mainly due to different water stress conditions during the growing season. Inter-annual variations of (epsilon) an were much less important, while for periods shorter than a growing season, (epsilon) an exhibits very contrasting values from re-growth to senescence. Therefore the calculation of ANPP based on Monteith's approach seems less prone to errors due to environmental effects when computed on an annual basis, whereas for periods shorter than the growing season the computation of either GPP, NPP or ANPP is delicate.