Advance directives have been widely endorsed as a method to match medical interventions to patients' preferences. However, applying advance directives to unspecified future decisions may be problematic. We wondered how well scenario-based treatment choices in advance directives can be used to infer other choices specified by the same patients. We asked 495 out-patients from the Massachusetts General Hospital to state 11 treatment choices in each of four illness scenarios. We calculated likelihood ratios (LRs) to see how well one choice could predict another. Predictions within a given scenario were strong and followed a pattern that reflected the invasiveness of the treatment. Thus, decline of antibiotics predicted decline of major surgery, in the same scenario (LR 36.0-108.3). Requests for major surgery strongly predicted requests for antibiotics (LR 90.4 - 244.1). Requests for major surgery and decline of antibiotics were the two most predictive preferences; more traditionally considered decisions about cardiopulmonary resuscitation and mechanical ventilation ranged from the third to ninth most predictive among the 11 interventions. Predictions between scenarios were weaker, but still potentially useful, and followed a pattern that reflected the prognoses of the scenarios. Declining treatment in the best prognosis scenario (coma-with-a-chance) predicted declining the same treatment in other scenarios with LRs of 2.5-6.1. Requests for treatment in the worst prognosis scenario (dementia-with-terminal-illness) predicted the same request in the other scenarios with LRs of 5.2-30.5. These data suggest that patients' advance scenario-based treatment choices can provide potentially useful information to apply to unspecified decisions if inferences follow considerations of prognosis and treatment invasiveness. For maximal predictive power, treatment directives should use a range of illness scenarios and include choices, among several others, on antibiotic use and major surgery.