We use newly collected individual-level hunger recall information from the China Family Panel Survey to estimate the causal effect of undernourishment on later-life health. We develop a two-sample instrumental variable (TSIV) estimator that can deal with heterogeneous samples. We find a nonlinear relationship between mortality rates, a commonly used famine indicator, and the individual hunger experience. The nonlinearity in famine exposure may explain the variation in the famine's effect on later-life health found in previous studies. We also find that exposure to hunger early in life leads to worse health among females 50 years later. This effect is much larger than the reduced-form effect found in previous studies. For males, we find no impact.