Using slope units (SUs) as analysis units can significantly enhance the readability and interpretability of landslide susceptibility maps. To bridge the scale gap between polygon-format SU data and raster-format landslide conditioning factor data, statistical values derived from grid cells must be assigned to individual SUs. Consequently, selecting appropriate statistical values becomes a critical scientific issue. To address this issue, a machine learning-based evaluation was performed to determine optimal factor assignment methods. The analysis indicated that different factor assignment methods generated a maximum difference of 0.0133-0.0198 in the area under the receiver operating characteristic curve and a maximum difference of 2.7%-7.1% in the percentages of historical landslides that fell within high and very high susceptibility zones. For categorical factors, using the fraction of each class within an SU outperformed using the predominant class and is recommended. For continuous factors, relying solely on the mean of grid values within an SU generally yielded poorer performance than using multiple statistical measures, and 10 quantiles of the grid values are recommended as input variables. These results highlight the important influence of selecting appropriate statistical values on improving SU-based landslide susceptibility mapping and are expected to provide a valuable reference for future studies.