Animal responses to global climate variation might be spatially inconsistent. This may arise from spatial variation in factors limiting populations' growth or from differences in the links between global climate patterns and ecologically relevant local climate variation. For example, the North Atlantic Oscillation (NAO) has a spatially consistent relation to temperature, but inconsistent spatial relation to snow depth in Scandinavia. Furthermore, there are multiple mechanistic ways by which climate may limit animal populations, involving both direct effects through thermoregulation and indirect pathways through trophic interactions. It is conceptually appealing to directly model the predicted mechanistic links. This includes the use of climate variables mimicking such interactions, for example, to use growing degree days (GDD) as a proxy for plant growth rather than average monthly temperature. Using a unique database of autumn body mass of 83331 domestic lambs from the period 1992–2007 in four alpine ranges in Norway, we demonstrate the utility of hierarchical, mechanistic path models fitted using a Bayesian approach to analyse explicitly predicted relationships among environmental variables and between lamb body mass and the environmental variables. We found large spatial variation in strength of responses of autumn lamb body mass to the NAO, to a proxy for plant growth in spring (the Normalized Difference Vegetation Index, NDVI) and effects even differed in direction to local summer climate. Average local temperature outperformed GDD as a predictor of the NDVI, whereas the NAO index in two areas outperformed local weather variables as a predictor of lamb body mass, despite the weaker mechanistic link. Our study highlights that spatial variation in strength of herbivore responses may arise from several processes. Furthermore, mechanistically more appealing measures do not always increase predictive power due to scale of measurement and since global measures may provide more relevant “weather packages” for larger scales.