OBJECTIVES: To compare conceptual models, empirical measurement, and results of alternative methods for measuring patients' willingness to trade off severe adverse event risks for specified health gains. METHODS: We define and compare the theoretical foundations of standard gamble (SG) and multi-attribute conjoint analysis (CA) methods. SG derives from von Neumann-Morgenstern expected-utility theory, while CA applies McFadden random-utility theory to hypothetical choices. We define theoretical conditions under which the two methods provide equivalent health-preference measures, including linearity, separability, and risk-neutrality. We evaluate accepted empirical methods used in SG and CA studies and propose methods for incorporating risks as CA treatment attributes. We then compare empirical maximum acceptable-risk estimates from CA studies of multiple sclerosis and Crohn's disease patients with and without restrictive SG assumptions, as well as with published SG estimates from other disease interventions. RESULTS: We find that SG can be used to estimate MARs for specific health outcomes only by imposing more restrictive assumptions on patient preferences than CA methods require. We show that CA methods can be used to test various theoretical restrictions imposed by the SG assumptions and find that risk neutrality and linearity are rejected statistically in most cases. By imposing SG assumptions on CA results and by comparing CA results to published SG estimates for chronic conditions, we find that SG assumptions increase MAR estimates by 20% to 150% relative to those obtained by CA methods. CONCLUSIONS: CA methods can be used to replicate SG tradeoff tasks and to test the restrictions required to interpret SG estimates as risk-preference measures. CA offers a more flexible and conceptually rigorous method than SG as conventionally applied for measuring treatment preferences and risk-benefit tradeoffs. Most importantly, multiattribute CA methods can more realistically simulate clinically relevant risk-benefit tradeoff choices to improve the validity and reliability of preference estimates.