The paper discusses criteria for comparing risk aversion when outcomes are multidimensional. It introduces two concepts: "commodity-specific greater risk aversion," which compares risk premia paid in a specified commodity, and "uniformly greater risk aversion," which compares risk premia regardless of the commodity used for payment. These concepts do not assume that von Neumann-Morgenstern utility functions are ordinally equivalent. The paper shows that nonincreasing consumption-specific risk aversion is sufficient to make randomization undesirable in an agency problem with hidden characteristics. The paper also provides an analogue of Pratt's characterization result for the unidimensional case and discusses the role of premium specificity versus uniformity in risk aversion. An example demonstrates that the comparison of risk attitudes based on commodity-specific risk aversion can be sensitive to the choice of commodity, while uniformly greater risk aversion is independent of the commodity used for payment.The paper discusses criteria for comparing risk aversion when outcomes are multidimensional. It introduces two concepts: "commodity-specific greater risk aversion," which compares risk premia paid in a specified commodity, and "uniformly greater risk aversion," which compares risk premia regardless of the commodity used for payment. These concepts do not assume that von Neumann-Morgenstern utility functions are ordinally equivalent. The paper shows that nonincreasing consumption-specific risk aversion is sufficient to make randomization undesirable in an agency problem with hidden characteristics. The paper also provides an analogue of Pratt's characterization result for the unidimensional case and discusses the role of premium specificity versus uniformity in risk aversion. An example demonstrates that the comparison of risk attitudes based on commodity-specific risk aversion can be sensitive to the choice of commodity, while uniformly greater risk aversion is independent of the commodity used for payment.