For noninvasive quantitative analysis of blood biochemical components, near infrared (NIR) subtracted blood volume spectrometry is applied. However, the optical path lengths corresponding to in vivo blood spectra obtained are uncertain. In order to study the effect of optical path length difference on the prediction accuracy of calibration model, 30 serum samples are simulated. A variable path-length cell is used to produce optical path difference, and spectra of two sample sets, with equal and different path-lengths respectively, are measured by Nicolet 6700 Fourier transform infrared (FT-IR) spectrometer. As to serum albumin, the calibration accuracies of the two groups are compared. The results show that the samples with uncertain optical path would reduce the calibration accuracy, as the root mean square error of cross validation (RMSECV) of the model increases from 110.0 mg/dL to 156.0 mg/dL. By applying multiplicative scatter correction, RMSECV of the uncertain optical path set reduces to 98.1 mg/L. It indicates that appropriate pretreatment method can effectively correct the error of spectrum caused by uncertain optical path length, thus improve the accuracy of the prediction model.