In the sensing of blood glucose by the near-infrared spectroscopy, building a robust and effective model is the precondition to obtain an accurate and reasonable prediction result of glucose concentration. In the chemometrics analysis, training set should be representative, reasonable distribution and cover the scope of prediction set. So the experiment designs became one of most difficult challenges for the noninvasive glucose sensing, especially for the in vivo experiments. In this paper, the oral glucose tolerance tests of two diabetics were carried out. The transcutaneous diffuse reflectance spectra were collected by a custom-build spectrometer and the glucose reference were measured by an invasive portable glucose meter. Then the influence of different experiment designs including the error in the references, the time delay between glucose in blood and interstitial fluid, the change in physiological temperature and different validation methods were analyzed. The result showed that, the error induced by the uncertainty in the reference was lower than that by the time delay, which could be up to 15.4%. And the proportion of error induced by temperature change is more than 50%, which is the most significant. Furthermore, the prediction error was restricted by the validation set selection and the way to change the blood glucose concentration.