Diffusion MRI provides a unique opportunity to study the brain tissue architecture at a microscopic level. More specifically, it allows to infer biophysical properties of the axons in the white matter in-vivo. Microstructural parameters are widely used in multi-subject studies to track pathological processes, follow normal development and aging, or investigate behaviour. This thesis aims to identify and potentially address the limitations and pitfalls in voxelwise comparison of diffusion MRI parameters across subjects. To allow for accurate matching of brain structures across subjects, non-linear transformation that spatially aligns the data is required. We demonstrate that using advanced registration methods, we can outperform the standard registration-projection approach both in terms of sensitivity and specificity. The coarse resolution of the data typically causes partial volume effects that bias the diffusion parameters and potentially mislead the interpretation of a group study outcome. We provide evidence that these effects can be addressed by constraining the diffusion model parameter space, which leads to marginally lower sensitivity, but allows an accurate interpretation of the results. Additionally, we suggest that additional information inferred with a data driven approach might mitigate the loss in sensitivity. Finally, we design an original tract-specific modelling framework that enables to estimate microstructural parameters unbiased by the presence of foreign fibre populations or tissues. We demonstrate the sensitivity of our method in a study relating microstructure and behaviour.