OBJECTIVES: The values assigned to EQ-5D health states may be inconsistent with their logical severity. A logical inconsistency occurs when a health state is logically more severe than another but receives a lower value. The objective of this research was to determine whether logical inconsistencies are related to differences in the severity of paired health states. METHODS: One hundred twenty-six Dutch university students used visual analog scales to value the 243 EQ-5D health states. Logical inconsistency rates were estimated for 55 pairs of health states over 1,000 randomly generated vectors of 11 states. Multiple logistic regression was used to model the logical inconsistency of paired health state values as a function of between-state differences in the severity of specific dimensions. RESULTS: Over the 1,000 health state vectors, mean inconsistency rates for paired low-, moderate-, and high-severity states were 0.011, 0.220, and 0.045, respectively. Average inconsistency rates for paired moderate-severity states differed from the average rates for paired low-severity states (p <0.001) and high-severity states (p <0.001). For each EQ-5D dimension, the probability of logical inconsistency was positively related to the similarity of the paired states with respect to health problems. The probability of logical inconsistency was higher when both states had moderate or extreme problems in a given dimension than no problems. CONCLUSIONS: The probability of logically inconsistent valuations increases with the similarity of the paired states and is higher when both states are moderately severe than otherwise. These findings have implications for the selection of health states in population-based valuation studies.