Two-dimensional, nonlinear, finite-difference analyses were used to investigate the stability of an embankment supported by spatially variable soil–cement columns. This study evaluated the improvement/merit of employing stochastic modeling approaches, such as spatially correlated random fields, relative to deterministic analysis, which commonly neglects spatial variability. The spatial variability, which is specified by mean, coefficient of variation (COV) and autocorrelation distance (
), can significantly influence the performance of the treated in situ soils and the overlying structure. Reliability analyses with three types of spatial variable soil–cement columns were conducted using a COV value of 0.6 for the soil–cement strength, and the effects of the autocorrelation distance were evaluated. Gaussian spatially correlated random field models (stochastic models) with 100 realizations for each case were used in a strength-reduction analysis to determine the failure surface and safety factor. Results of the models with uniform material properties were compared with each of the stochastic model realizations. ANOVA was used to analyze the statistical difference between the means of safety factors. The stochastic distributions of safety factors and consequences of spatial variability on the failure mode of columns and embankment, and implications for engineering practice were discussed.