Governments try to discourage risky health behaviours, yet such behaviours are bewilderingly persistent. We suggest a new conceptual approach to this puzzle. We show that expected utility theory predicts that unhappy people will be attracted to risk-taking. Using US seatbelt data, we document evidence strongly consistent with that prediction. We exploit various methodological approaches, including Bayesian model selection and instrumental variable estimation. Using road accident data, we find strongly corroborative longitudinal evidence. Government policy may thus have to change. It may need to improve the underlying happiness of individuals instead of, or in addition to, its traditional concern with society's risk-taking symptoms.