A variety of formulas have been published for predicting GA in the second and third trimesters. We assessed these formulas for the presence of systematic errors. Our study population consisted of 1036 obstetrical sonograms at 14 to 42 weeks' gestation in patients with highly accurate dating based on a prior first trimester scan. Most formulas had mean errors (mean value of the difference between predicted and true GA) of no more than 1 week between 14 and 38 weeks' gestation. At 38 to 42 weeks, however, all formulas had large systematic underestimation biases, with mean negative errors in the range of 1.6 to 3 weeks. By regressing ln(GA) against one or more measurements, we developed new formulas that eliminate this underestimation bias, with mean errors of no more than 0.6 week throughout the entire 14 to 42 week period. As with current formulas, however, our formulas have wide 95% confidence ranges of +/- 3 weeks or greater at 38 to 42 weeks as a result of inherent biological variability. We recommend the use of these formulas in clinical practice, as they may prevent errors in diagnosis late in pregnancy.