The decomposition of spatially offset Raman spectra for complex multilayer systems, such as biological tissues, requires advanced techniques such as multivariate analyses. Often, in such situations, the decomposition methods can reach their limits of accuracy well before the limits imposed by signal-to-noise ratios. Consequently, more effective reconstruction methods could yield more accurate results with the same data set. In this study we process spatially offset Raman spectroscopy (SORS) data with three different multivariate techniques (band-target entropy minimization (BTEM), multivariate curve resolution and parallel factor analysis (PARAFAC)) and compare their performance when analysing a spectrally challenging plastic model system and an even more challenging problem, the analysis of human bone transcutaneously in vivo. For the in vivo measurements, PARAFAC's requirement of multidimensional orthogonal data is addressed by recording SORS spectra both at different spatial offsets and at different anatomical points, the latter providing added dimensionality through the variation of skin/soft tissue thickness. The BTEM and PARAFAC methods performed the best on the plastic system with the BTEM more faithfully reconstructing the major Raman bands and PARAFAC the smaller more heavily overlapped features. All three methods succeeded in reconstructing the bone spectrum from the transcutaneous data and gave good figures for the phosphate-to-carbonate ratio (within 2% of excised human tibia bone); the PARAFAC gave the most accurate figure for the mineral-to-collagen ratio (20% less than excised human tibia bone). Previous studies of excised bones have shown that certain bone diseases (such as osteoarthritis, osteoporosis and osteogenesis imperfecta) are accompanied by compositional abnormalities that can be detected with Raman spectroscopy, the utility of a technique which could reconstruct bone spectra accurately is manifest. The results have relevance on the use of SORS in general.