Bayesian inversion techniques for assessing reservoir performance uncertainties involve generating multiple reservoir models conditioned to the available field data. This process requires sampling in a high-dimensional parameter space to identify good data-fitting models. Therefore, the robustness of the inversion technique strongly depends on the efficiency of the sampling methods used to generate the reservoir models. In order to improve the robustness, the factors affecting the uncertainty estimations must be identified. This paper aims to investigate the effect of sampling strategies on the prediction uncertainty estimations. A synthetic reservoir model has been studied to compare the uncertainty estimations based on different sampling outcomes obtained using Genetic Algorithms (GA) and Neighbourhood Algorithm (NA). The main differences in the sampling outcomes from GA and NA are the degree of exploration and the number of good data-fitting regions identified in the parameter space. We show that different sampling strategies may result in significantly different uncertainty estimates. We also demonstrate that the predictive capability of the history-matched models can be used as an indicator for the spread of the posterior probability distribution. Copyright © 2007, Institut français du pétrole.