For practical considerations, it is in some case impossible to simulate MAS models at population level. The current paper shows that MAS models applied to samples with heterogeneous costs of interactions between agents have biased results. Heterogeneous costs of interactions in MAS models can come from the spatial dimension in MAS models or from fixed costs per interaction. The paper presents two correction procedures to remove the sampling bias and to increase the reliability of the outcome. The correction procedures can be very promising for future applications of MAS models because it becomes possible to deploy more complex models without bias on more detailed datasets that are only available at sample level, which will be the case for country- or EU-wide MAS applications.