Logical inconsistencies in survey respondents' valuations of hypothetical health states - represented by the EQ-5D, for example - present a conundrum as to whether or not their responses ought to be included for estimating social 'tariffs'. A 'logical inconsistency' occurs when a state that 'in logical terms' is unambiguously less severe than another is assigned a lower value. Excluding such responses is defensible on data quality grounds but puts at risk the representativeness of the estimated tariff, given it is meant to represent the preferences of 'society'. This paper explores the rationale for and effect of excluding, to varying degrees, responses distinguished by the number of pairwise inconsistencies they contain, and reports equations for two tariffs that arise from contrasting approaches. The data are from a random sample of adult New Zealanders whose visual analogue scale valuations for a selection of EQ-5D states were collected in 1999 via a postal survey to which 1360 people responded (a 50% response rate). We conclude that there is no simple, generalisable 'rule' to guide exclusions and therefore researchers ought to explore the sensitivity of their estimated tariffs (and ultimately QALY estimates) to alternative treatments of logically inconsistent responses