Abstract: A improved modeling method for selection of variables, i.e., CARSiPLS was proposed by combining the methods of competitive adaptive reweighted sampling (CARS) and internal partial least square regression (iPLS) and applied to the modelling in NIRS determination of moisture and volatiles in bitumite. PLS regression models for NIRS determination of moisture and volatiles in bitumite were established by stepwise selection of those variables related to the measurements from each interval by CARS. It was shown that as compared with PLS and iPLS, the number of variables was significantly reduced in CARSiPLS and the prediction performances of the models were also improved. In establishment of the models for moisture and volatiles, the number of variables was reduced from 1557 to 317 and 15 respectively. Values of MAPE and RMSEP were reduced remarkably from 0.031 5 to 0.018 4 and from 0.010 8 to 0.006 7 for volatile determinations, and from 0.188 4 to 0.094 6 and from 0.005 0 to 0.002 8 for moisture determinations.