Abstract
1605
Objectives: Data analysis in nuclear medicine often requires working with several candidate models. The model which is supported best by the data can be determined using various criteria. This work compares two model selection criteria namely the F-test and the AICc for the example of sparse data. Model selection frequency of the “true” model and the determination of relevant parameters were examined using Monte-Carlo simulations.
Methods: Data were generated with four different Gaussian errors by a sum of two exponentials, which parameters were determined by fitting to serum time activity data. The sampling times of the synthetic data were chosen according to a recommended sample design for radioimmunotherapy (ICRU Report67). The generated data were fitted to each model of a set of candidate models (sums of exponentials) and the best model was selected on the basis of the F-test and the AICc. Subsequently, the clearance, the steady state volume of distribution and the mean residence time were calculated.
Results: The AICc and the F-test selected the true model with increasing error of 100% to 79% and 96% to 90% of the converged replications, respectively. The AICc tends to select a model of lower dimension for higher Gaussian error while the F-test tends to overfit for all Gaussian errors. Variability and bias depend on the estimated parameter but are of the same magnitude for both approaches. Model averaging using the AICc reduced bias and variability.
Conclusions: Good scientific practice requires working with a small set of candidate models rather than one arbitrary model. We advocate the use of AICc for data analysis in nuclear medicine, as the AICc has proved to be an adequate method for model selection and inference with sparse data.
Research Support: German Research Foundation (DFG GL 236/6-1,2 and GL 236/7-1).
- Society of Nuclear Medicine, Inc.