The use of a most likely path (MLP) formalism for protons to account for the effects of multiple Coulomb scattering has improved the spatial resolution in proton computed tomography (pCT). However, this formalism assumes a homogeneous medium and a continuous scattering of protons. In this paper, we quantify the path prediction error induced by transverse heterogeneities to assess whether correcting for such errors might improve the spatial resolution of pCT. To this end, we have tracked protons trajectories using Monte Carlo simulations in several phantoms with different heterogeneities. Our results show that transverse heterogeneities induce non Gaussian spatial distributions leading to errors in the prediction of the MLP, reaching 0.4 mm in a 20 cm wide simulated heterogeneity and 0.13 mm in a realistic phantom. It was also shown that when the spatial distributions have more than one peak, a most likely path, if any, has yet to be defined. Transverse heterogeneities also affect energy profiles, which could explain some of the artifacts described in other works and could make the energy cuts usually performed to exclude nuclear events less efficient.