Near infrared reflectance spectroscopy (NIRS) has been receiving increasing attention for the rapid and inexpensive determination of soil properties, total carbon (Ct) and nitrogen contents (Nt) especially. However, methodological aspects such as sample grinding and drying or replication have not been addressed extensively. The objectives of the paper were thus to assess how NIRS predictions of Ct and Nt were affected by sample grinding (2 mm sieving vs. 0.2 mm grinding), drying (air-drying vs. oven-drying at 40°C during 24 h), and replication (use of one to six subsamples to determine average spectra). This was performed on a range of tropical soils that differed widely in mineralogy (low and high activity clay soils, allophanic soils) and texture (sandy to clayey). The accuracy of NIRS predictions of Ct and Nt was higher with oven-dried compared to air-dried samples, and more markedly, with 0.2 mm ground compared to 2 mm sieved samples. Replication had a positive effect on NIRS predictions when 2 mm sieved samples were used, especially for air-dried samples, but this effect was not clear with 0.2 mm ground samples. Thus the most accurate predictions of Ct and Nt were obtained with oven-dried finely ground samples, with limited response to sample replication. Accurate predictions were, however, also obtained with four replicates on oven-dried 2 mm sieved samples. Acceptable and less tedious results could thus be achieved when replacing fine grinding by replication. Even with this procedure, the r² between predicted (NIRS) and measured (reference) values was 0.9 and the ratio of standard error of prediction to mean (CV%) was 20%, considered satisfactory for the heterogeneous sample set under study.