Accelerated water erosion is a major soil degradation process that affects soil and water quality. In Cuba, specifically, more than 40% of agricultural lands are affected by severe erosion problems. Estimating accurate erodibility values is a crucial step for the calibration and proper application of erosion models. Several equations have been developed to estimate erodibility from soil properties; however, these are often soil- or site-specific, limiting their application. This study aims to (1) identify soil properties governing the erodibility of tropical soils from western Cuba, (2) find suitable regression models to estimate erodibility from these properties, and (3) test widely applied erodibility equations. To achieve these goals, rainfall simulation experiments were conducted on runoff plots, and erosion-related physical, chemical, and mechanical soil properties were determined for 19 different soils. The main results indicated that good correlations between erodibility and certain soil properties were achieved after clustering soils based on their cation exchange capacity (CEC) values and clay content. Soils characterized by more than 30% of clay and 40 cmol+ kg−1 of CEC were excluded from the main analysis. Generally, clay content controls the erodibility of these tropical soils, exhibiting an inverse relationship. However, in the excluded soils, the clay fraction showed a positive relationship with erodibility. Soil water retention at the lowest matric potentials demonstrated the strongest correlation with soil erodibility, as this variable encompasses compound information related to clay, mineralogy, and organic matter. A new regression model to estimate erodibility based solely on the volumetric water content at 1500 kPa is presented. The optimal fitted logarithmic model accounts for 64% of the predictand variability in the studied soils. When testing known erodibility models, the nomograph was found to best mimic the erodibility trend of these soils, although it exhibited marked uncertainty and underestimation biases.