The variations in the preparation of biological samples (mainly tropical grasses and legumes) by a modified Dumas combustion method were studied to investigate their effect on the reproducibility and precision in estimating δ13C levels. The variables studied included plant species, evacuation of tubes, combustion temperature, amount and form (i.e., particle size) of oxygen donor, type of catalyst and the effect of mixing the ingredients in the combustion tube. The evacuation of the tubes prior to sealing and the use of 2.0 g of copper(II) oxide (CuO)(compared with smaller amounts) contributed more to improving the over-all precision than any other factor. The use of 2.0 g of CuO also led to more complete combustion of some samples. For pasture samples and samples of extrusa from fistulated cattle, there was little to choose between the two combustion temperatures (i.e., 550 and 850 °C). However, for faeces and hair samples, combustion at 550 °C resulted in a positive shift in the δ13C values and a marked decrease in precision compared with results from combustion at 850 °C. Possible reasons for this effect are discussed. Significant interactions of temperature with CuO form and catalyst occurred, but the magnitude of these effects was small. On the basis of the results obtained, the preferred method for the preparation of dried samples, ground to a fine powder (Tema mill), was to use 4–8 mg of sample, 2 g of microanalytical reagent-grade wire-form CuO and 3 cm of 0.2 mm diameter silver wire in a 228 mm long, standard wall, 6 mm o.d. Vycor tube evacuated to 30 × 10–3 Torr and sealed with an H2– O2 flame. The ingredients in the sealed tube were thoroughly mixed by shaking, combusted at 850 °C for 2 h and allowed to cool to room temperature before being removed from the furnace. Storage of the combusted plant samples for 11 months had no effect on the δ13C values. The total preparation time for 40 samples using the recommended procedure was approximately 7 h and the analytical time on the mass spectrometer was 12 ± 2 min per sample. The standard error of the mean of three samples was 0.065‰ with a confidence interval, at the 95% level of probability, of 0.057–0.076‰. These values are lower than those published for the dynamic combustion method.