This study develops a stochastic inverse analysis methodology to probabilistically estimate site-specific soil modulus from geophysical test measurements by accounting for the uncertain spatial variability of the soil deposit, any measurement uncertainty and uncertainty due to limited data. Hypothesizing the soil modulus to be a three-dimensional, heterogeneous, anisotropic random field, the methodology first formulates and solves a forward model that mimic a geophysical experiment using a stochastic collocation approach to characterize the effect of spatially variable, uncertain soil modulus on the model response variables, for example, accelerations at the sensor locations. The stochastic collocation approach utilizes recently developed non-product quadrature method, conjugate unscented transformation, to accurately estimate statistical moments corresponding to the model response variables in a computationally efficient manner. The methodology then employs a minimum variance framework to merge the information obtained from the model prediction and the sparse geophysical test measurements to update the statistical information pertaining to the soil modulus. The methodology is illustrated using synthetic data from a fictitious geophysical experiment.