As the population ages, governments and international organizations are trying to lengthen the labor-force participation of older adults. For older adults, health is an important determinant of working decisions. In this paper, I introduce heterogeneity in health dynamics with age and argue uncertainty about health dynamics affects the working decisions of older adults. Using the Health and Retirement Study, I first show evidence of heterogeneity in health profiles with age. Second, I use subjective survival expectations to infer health beliefs in a Bayesian-learning framework. Third, I flexibly estimate how working decisions depend on those beliefs, using a neural-network approach that does not require additional structure. The results show beliefs have substantial negative bias. That is, on average, individuals incorrectly believe their health will deteriorate too fast. Furthermore, eliminating that bias would increase labor-force participation by up to 2 percentage points. In the last part of the paper, I look at a policy that could affect beliefs: the provision of information on blood glucose and cholesterol levels. I take advantage of the randomization in the collection and provision of such information. The results show the information has only small effects on beliefs and working decisions, and consequently, policies with larger effects on beliefs are needed to delay retirement.