On the robustness of the Arrow-Pratt risk aversion measure