OBJECTIVES: To construct an algorithm which converts statistics commonly reported in publications with the SF-12 health status measure to an average SF-6D preference based score. METHODS: We used SF-12 data from the 2002 Medical Expenditures Panel Survey. We presumed commonly published sufficient statistics would include average age, sex, physical component score (PCS), and mental component score (MCS). All combinations of these variables were used as predictors in models built with WinBUGS 1.4. Model fit was evaluated with the Deviance Information Criterion (DIC). The best fit model was also evaluated using R-square for comparison to other algorithms that convert SF-12 summary scores to preference scores. RESULTS: We used all respondents with PCS and MCS scores (n=20,206). The best fit model included age, sex, PCS, and MCS as predictor variables (DIC = -67434). The model was SF-6D = -0.001544 - 0.002173*female + 0.000144*age + 0.008097*MCS + 0.00816*PCS. The R-square of this model (0.88) was substantially better than models that convert to EQ5D summary scores developed by Lawrence et al (0.61) or Franks et al (0.63 and 0.59) or to HUI Mark 3 summary scores by Franks et al (0.51) or Sengupta et al (0.55). Because this model does not include power or interaction terms, knowing the average age, PCS score, MCS score, and the percent who are female in a sample is sufficient to predict an average SF-6D score. The residual from directly calculated SF6D scores drops dramatically as group size increases; the standard deviation of residual size is .046 for 1 subject, 0.014 for 10 subjects, 0.006 for 50 subjects, 0.005 for 100 subjects, and approaches an asymptote of 0.003 with more than 200 subjects. CONCLUSIONS: Commonly reported summary statistics from previously published articles provide sufficient information for estimating an average SF-6D score without accessing individual level data.