Serial versus sparse sampling in toxicokinetic studies

Sparse sampling in rodent toxicokinetics usually involves the collection of a single blood sample on a given study day from each animal in a treatment group. The samples are allocated to different time points, often allowing some replicates, and statistical inferences are then made about the concentration-time behavior of the test compound. The present study compared the results of one such analysis with those obtained from serial sampling as might be applied using satellite animals. Ten rats each received a single oral dose of tritium-labeled compound X. Blood concentrations in each rat at 10 time points post-dose were determined by liquid scintillation counting. Individual peak concentrations of blood radioactivity (Cmax) and peak times (tmax) were recorded, and area-under-curve (AUC) values were calculated by trapezoidal rule. The mean AUC and Cmax of all 10 animals over all 10 time points, referred to as AUCtrue and Cmax,true, were used as points of reference. These values were then estimated using subsets of the data that simulated satellite-animal or sparse-sampling designs. First, several different sampling schedules of 5 bleeding times were stimulated by taking subsets of the full data set. For analysis using satellite animals, serial blood concentrations from subsets of 3 or 4 rats were used to calculate point and confidence-interval estimates of AUCtrue and Cmax,true by standard methods; all possible subsets of 3 or 4 of the 10 rats were considered. For sparse data analysis, a single concentration from each of the 10 rats was used to calculate both point and confidence-interval estimates of AUCtrue by the Bailer-Satterthwaite method, and point estimates of Cmax,true, according to several different designs of replication. Animals were randomly assigned to time points, and 1000 of over 50000 possible combinations were evaluated for each bleeding schedule. The average percent absolute errors of the point estimates were computed and, for the 95% confidence intervals, average widths were determined. For point estimates of AUCtrue, sparse sampling yielded average percent absolute errors of 7-13%. Percent errors for 3 and 4 satellite animals were 6-12% and 5-10%, respectively. For 95% confidence intervals, sparse sampling yielded widths of 24-90% of AUCtrue, whereas for 3 and 4 satellite animals widths were 37-50% and 24-34%, respectively. For Cmax,true point estimates from sparse-sampling and satellite-animal approaches had average percent absolute err