The pharmacokinetics of a drug in individual patients can be estimated using plasma samples collected at a limited number of time points. However, different methods for a limited-sampling strategy (LSS) design exist and the optimal method has not yet been defined. Plasma concentration data were available from 27 of 74 courses in a phase I study (dose range, 5-55 mg m-2) of the novel anthrapyrazole DuP-941. Three approaches to LSS development were compared. Firstly, forward stepwise regression (FSR) was used to derive equations to predict the DuP-941 area under the concentration-time curve (AUC) based on plasma concentrations measured at specified times. LSSs were developed using 14 randomly chosen data sets and were validated using the remaining 13 data sets. Secondly, "all subsets" regression (ASR) was used to develop LSSs. A jack-knife technique was also used to allow model development utilising 26 data sets and validation on the 27th data set. Thirdly, an LSS was developed using optimal sampling theory (OST), and the LSS was used in conjunction with a Bavesian algorithm. Selected sampling times for four-point LSSs were 10, 65, 185 and 485 min (FSR) and 10, 45, 200 and 480 min (OST). Ten candidate LSSs were developed using the ASR approach. ASR- and OST/Bayesian-derived four-point LSSs gave more precise (P < 0.05) estimates of AUC [mean absolute percentage of difference (MAD%) +/- SD: ASR, 6.4 +/- 3.7%; OST/Bayesian, 6.8 +/- 4.6%] than did FSR (MAD% = 15.1 +/- 9.9%). The OST/Bayesian approach is recommended because it allows estimation of all model parameters and is more flexible with regard to sample collection time and design variables.