Basically, the idea is this: suppose you are trying to calibrate to a peak, where the peak height is proportional to the concentration of whatever causes the peak. That's easy to do. But NIR is an especially tricky region because all kinds of things absorb there. Suppose that something else also absorbs at the same wavelength as the substance you are measuring (and calibrating to). Could be water, could be air, could even be the substance itself because there are also lots of overtone bands and combination bands that can interfere. Call the substance you are measuring/calibrating A and the interferant B. If contribution to the peak height due to A scales linearly with the concentration of A and the contribution to the peak height due to B scales linearly with the concentration of B, then in principle these could be easy to separate by just varying either A or B while holding the other fixed. The problem is that if the contribution of the peak height due to A depends linearly on A and also to some unknown extent on the concentration of B (and vice-vera), it is hard to deconvolute the signals just by varying A or B, since they are co-dependent. If this codependence is strong enough, it can introduce significant error into your calibration and is also hard to deconvolute because you may not know the degree of the dependence. Statistical methods are needed in this cases.