This paper develops an innovative approach to measuring the effect of health on retirement. The approach elicits subjective probabilities of working at specified time horizons fixing health level. Using a treatment-effect framework, within-individual differences in elicited probabilities of working given health yield individual-level estimates of the causal effect of health (the treatment) on working (the outcome). We call this effect the Subjective ex ante Treatment Effect (SeaTE). The paper then develops a dynamic programming framework for the SeaTE. This framework allows measurement of individual-level value functions that map directly into the dynamic programming model commonly used in structural microeconometric analysis of retirement. The paper analyzes conditional probabilities elicited in the Vanguard Research Initiative (VRI)—a survey of older Americans with positive assets. Among workers 58 and older, a shift from high to low health would on average reduce the odds of working by 28.5 percentage points at a two-year horizon and 25.7 percentage points at a four-year horizon. There is substantial variability across individuals around these average SeaTEs, so there is substantial heterogeneity in taste for work or returns to work. This heterogeneity would be normally unobservable and hard to disentangle from other determinants of retirement in data on realized labor supply decisions and health states. The paper’s approach can overcome the problem that estimates of the effect of health on labor supply based on behavioral (realizations) data can easily overstate the effect of health on retirement whenever less healthy workers tend to retire earlier for reasons other than health.