A study in 3 laboratories has demonstrated that accuracy of infrared analysis of fat in milk can be improved relative to results by AOAC method 16.083-16.092 by measuring IR absorption by CH2 and CH3 groups in fat molecules instead of or in addition to measuring absorption by carbonyl groups. Theoretical implications for alternative choices of measurements for fat analysis are discussed, and data are presented to indicate the agreement between theory and practice. Mathematical formulas developed from the Beer-Lambert law are presented to describe mechanisms for dual wavelength, double beam, and single beam optical arrangements. A mathematical description of circuitry for these arrangements is included. Methods for multiple linear regression estimates of calibration parameters are discussed. Comparison data indicate optimum situations regarding frequency of calibration and random selection of calibration samples. With optimum control of calibration parameters, mean bias for randomly chosen check samples was reduced to less than 0.01%. Standard deviation of difference for fat estimates was reduced to range from 0.026 to 0.034%. Quality control procedures and calibration techniques for individual instruments are described. The use of the Fat B filter and a combination of the Fat A and Fat B filters has been adopted official first action as a revision in method 16.083-16.092.