Species distribution within plant communities results from both the influence of deterministic processes, related to environmental conditions, and neutral processes related to dispersal limitation and stochastic events, the relative importance of each factor depending on the observation scale. Assessing the relative contribution of environment necessitates controlling for spatial dependences among data points. Recent methods, combining multiple regression and Moran's eigenvectors maps (MEM), have been proved successful in disentangling the influence of pure spatial processes related to dispersal limitation, pure environmental variables (not spatially structured) and spatially structured environmental properties. However, the latter influence is usually not testable when using advanced spatial models like MEM. To overcome this issue, we propose an original approach, based on torus-translations and Moran spectral randomizations, to test the fraction of species abundance variation that is jointly explained by space and seven soil variables, using three environmental and tree species abundance data sets (consisting of 120, 52 and 34 plots of 0·2 ha each, located along 101-, 66- and 35-km-long transect-like inventories, respectively) collected in tropical moist forests in southern Cameroon. The overall abundance of species represented by ≥30 individuals, and 27% of these species taken individually, were significantly explained by fine-scale (<5 km) and/or broad-scale (5–100 km) spatially structured variations in soil nutrient concentrations (essentially the concentration of available Mn, Mg and Ca) along the 120-plots area. The number of significant tests considerably decreased when investigating the two smaller data sets, which mostly resulted from low statistical power rather than weaker floristic and/or edaphic variation captured among plots. Synthesis. Our results provide evidence that tree species turnovers are partly controlled by spatially structured concentrations in soil nutrients at scales ranging from few hundreds of metres to c. 100 km, a poorly documented subject in Central African forests. We also highlight the usefulness of our testing procedure to correctly interpret the space-soil fraction of variation partitioning analyses (which always accounted here for the most important part of the soil contribution), as this fraction was sometimes relatively high (R2 values up to c. 0·3) but nearly or not significant.