Visible and near-infrared (VNIR) diffuse reflectance spectroscopy has potential in site-specific measurement of soil properties. However, previous studies have reported VNIR estimates of plant-available soil phosphorus (P) and potassium (K) to be of variable accuracy. In this study, we used a database of over 1500 soil samples to investigate what factors influenced P and K estimation accuracy. Specifically, the effects of classifying soil samples by Major Land Resource Area (MLRA), cation exchange capacity (CEC), or organic matter (OM) were investigated. Additionally, calibrations using only those samples within the approximate range of interest for fertilizer application to field crops (P from 0 to 27 mg kg -1 and K from 0 to 192 mg kg -1 ) were compared to calibrations using the full range of soil samples. Pretreatments of log 10 (1/reflectance) plus mean normalization plus median filter smoothing with or without direct orthogonal signal correction (DOSC) were investigated. Models were developed using partial least squares regression (PLSR) with leave-one-out cross-validation. Reasonable estimates of P and K were obtained for soil samples from two Missouri MLRAs (109 and 115B) out of the eight analyzed. Model estimates were poor when soil samples were grouped by CEC or OM; however, there was some indication that VNIR estimation of P and K might be possible for soils low in OM. Accuracy was maintained when analyzing a reduced wavelength range of 1100 to 2450 nm, suggesting that this narrower sensing range might be used for lower-cost on-the-go sensors. Inclusion of the DOSC pretreatment reduced P and K estimation errors by 25% to 39%. The results of this research provided some insight into the factors affecting the accuracy of P and K estimation by VNIR models, but additional research is needed to determine if these findings can lead to P and K estimations sufficiently accurate to guide variable-rate fertilization.