Accurate prediction and mapping of soil nutrient levels are essential for implementing variable rate technology. For the prediction of soil potassium (K), we evaluated the performance of the inverse distance weight of powers 1, 2, and 3, ordinary kriging, cokriging, multiple linear regression assuming independent error, and multiple linear regression with autocorrelated error structure. Two forms of ordinary kriging were evaluated: kriging the residuals from a trend surface regression (geographic locations only as predictors) and kriging the residuals from a regression of K on geographic location and other soil property predictors (soil pH and apparent electrical conductivity, ECa). The autocorrelated error model as implemented in the Statistical Analysis System (SAS) mixed linear model was employed to adjust for autocorrelated error structure in the regression models used for prediction. For cokriging, either soil ECa or soil pH was used as a secondary soil property to predict K. The root mean square error (RMSE) and mean error (ME) calculated from an independent validation data set (n= 68) were used as comparison criteria. The best result was obtained with the methods that incorporated geographic locations, other soil property predictors, and the correlated error structure. This investigation demonstrated the flexibility of the regression-based autocorrelated error model for spatial prediction compared with other methods. Further, the results of this study have important implications for screening economically acceptable soil and site characteristics that can be used to improve prediction of soil nutrients at unsampled locations within a field.