Landslide susceptibility mapping is essential for risk management and mitigation. Traditional multivariate models have achieved high nominal accuracy in predicting landslide occurrence, but the underlying methods rarely account for spatial dependence in the input data. We test how spatial Hierarchical Generalized Linear Models (HGLMs) and spatial autoregressive models might enhance both accuracy and reliability of landslide susceptibility. We estimate the frequency of landslides in a catchment, drawing on a catalogue of 10,837 landslides and several predictors (e.g., rainfall, elevation, slope) in the northern Colombian Andes. Our HGLMs integrate the effects of spatial dependency and heterogeneity through Markov Random Field (MRF) models, i.e. Intrinsic Conditional Autoregressive (ICAR), Besag-York-Mollié (BYM), and Leroux models. Our results show that landslide frequency is significantly influenced by spatially-dependent unobserved factors, likely representing contiguous geological formations or shared soil properties, which we therefore incorporate as latent variables. Even after accounting for known covariates, the HGLMs capture spatial dependencies that non-spatial models fail to address. Incorporating spatial structure in the data improves model performance, judging from model selection metrics such as the Deviance Information Criterion (DIC) or the Watanabe–Akaike Information Criterion (WAIC). By accounting for latent spatial effects, spatial HGLMs produce smoother and more reliable susceptibility maps. This approach overcomes a key limitation of traditional models: the underestimation of landslide frequency in high-density areas where unobserved, spatially-structured factors are most influential. Our findings highlight the importance of integrating spatial dependence and heterogeneity in landslide susceptibility models to achieve enhanced predictive performance and reliability.