In the geotechnical determination of the cohesion c and the angle of internal friction φ of a soil from shear tests, a linear regression model is fitted to normal and shear stress data, and confidence bounds are computed. The applicability of standard linear regression is limited by the physical requirement of non-negative cohesion and the statistical requirement of normality. We propose two methods from computational statistics that are able to overcome both obstacles: a bootstrap resampling method in case the experimental data set is sufficiently large, and a Bayesian approach for small samples. The methods are demonstrated at the hand of a real data set for glacial silt.