OBJECTIVE: Several estimators exist when average utility scores are not available for patient populations with multiple disease conditions. The multiplicative estimator is a widespread choice among them. This study is to empirically test the accuracy of the multiplicative estimator and compare it with other estimators. METHODS: Using the pooled Medical Expenditure Panel Survey (MEPS) 2001 and 2003 data, a sample of 40,846 adults with EQ-5D preference-based index scores were categorized into 238 disease condition categories. Due to the MEPS sampling property, comorbid pair categories with less than 100 individuals were excluded, which left us with 760 comorbid pairs in total. The study focus was the bias from the estimators to the observed mean scores for each comorbid pair, with the observed scores presumed to be the true value. The analyses were conducted using both the raw estimators and the rescaled (purified) estimators. Regressions and concordance correlation coefficients were also used to evaluate the agreement between the estimators and the observed scores. RESULTS: Using the rescaled approach, the scores estimated by multiplying the 2 mean scores of the corresponding disease conditions on average had a statistically significantly larger bias (p<0.0001) from the observed ones (-0.043) than simply picking the smaller mean of the 2 paired conditions (minimum estimator, bias=0.027). However, the multiplicative estimator had less bias than other estimators including the additive estimator (bias=-0.054), the larger mean (bias=0.077), the average of the means (bias=0.052), mean of the condition with smaller sample (bias=0.053). Results produced by other analyses, including using the raw scores, all favored the minimum estimator than the multiplicative estimator. CONCLUSIONS: Multiplication is not a good estimate when the average utility score for patients with 2 disease conditions is not readily available. The lower of the 2 utility scores had the least error among those estimators that we compared.