Multicriteria decision analysis (MCDA) is rightly receiving increasing attention in health technology assessment. Nevertheless, a distinguishing feature of the health domain is that technologies must actually improve health, and good performance on other criteria cannot compensate for failure to do so. We argue for two reasonable tests for MCDA models: the treacle test (can a winning intervention be incompletely ineffective?) and the smallpox test (can a winning intervention be for a disease that no one suffers from?). We explore why models might fail such tests (as the models of some existing published studies would do) and offer some suggestions as to how practice should be improved.