The paper studies the effect of health on work among older workers using data on individuals' subjective probabilities of working to specified future horizons under alternative health states. It constructs within-person differences in work expectations across health states, interpretable as Subjective ex ante Treatment Effect (SeaTE) within potential outcomes, and document large SeaTE heterogeneity across individuals. Using standard discrete choice dynamic programming to derive health-contingent values of working longer, it shows that the revealed heterogeneity in taste for work biases regression-based estimates. Finally, the health-contingent working probabilities are shown to strongly predict work given realized health after two years. (joint with Matthew D. Shapiro)