Although emission models have been designed using vehicle data over driving cycles of a few minutes, they are often applied at large scale to estimate total emission (inventories). In between, there is a range of scales in use in traffic and environmental studies (road sections, sub-areas, etc.). Coupling a traffic microsimulation with COPERT emission factors at different scales reveals scaling biases. We compare network fuel consumption (FC) and nitrogen oxide (NOx) emissions resulting from emission calculations based on different spatial decompositions. The results show that for an area of Paris covering 3 km2, the differences due to the aggregation scale for emissions range from 5 to 17% depending on the pollutant, spatial partitioning and traffic conditions. These discrepancies can be reduced using a distance-weighted mean speed, which is not a scale-consistent definition of mean travel speed. They can almost be cancelled by using a correction term derived analytically in this paper, thus consistency can be guaranteed between emissions assessed at different scales. Finally, a case study shows that it is possible to evaluate FC and NOx emissions on a large-scale network from a sample of traffic data (probes), and obtain the corrective term to be applied to remove scaling bias. The most critical step is the accurate estimation of the total travel distance. The gaps were successfully reduced to a maximum of 8% in congestion for a penetration rate of about 20%.