We adopt realized covariances to estimate the coefficient of risk aversion across portfolios and through time. Our approach yields second moments that are free from measurement error and not influenced by a specified model for expected returns. Supporting the permanent income hypothesis, we find risk aversion responds to consumption-smoothing behavior. As income increases, or as the consumption-to-income ratio falls, relative risk aversion decreases. We also document variation in risk aversion across portfolios: risk aversion is highest for small and value portfolios.