A finite element soil-structure interaction model is coupled with a discrete auto-regressive code, in order to analyze the effect of added spatial variability due to soil improvement in seismic risk analysis. The success of soil improvement techniques is related to the effectiveness of the method-i.e. how much of the soil is being changed-but also to its efficiency in improving the soil behavior-i.e. how much are the liquefaction and liquefaction-induced settlement reduced. As these techniques can add spatial variability to a deposit, it can affect the triggering of liquefaction on the soil and the behavior of the structures above it. In this study, this heterogeneity is modeled as a binary mixture, composed by the original liquefiable sand and the added treated sand. The soil behavior is represented by a fully nonlinear elastoplastic multi-mechanism model. The co-seismic settlements of the structure and the liquefaction of the soil deposit are estimated for different effectiveness levels-measured by mixture fractions-and for different spatial distributions. In general, both very small or very high mixture fractions presented low efficiency as the improvement in the relative settlement was small. Additionally, results show that the interaction between loose and dense deposits is highly dependent on the spatial distribution. Therefore , homogeneous equivalent models will rarely correspond to the average of the heterogeneous response.