This paper relaxes some restrictions of previous twin-based estimates of the effects of education on earnings. First, it estimates the earnings premiums associated with different educational levels. Second, it estimates a piecewise linear relationship between the natural logarithm of annual earnings and years of schooling. Third, the measurement error corrections are based on a less restrictive, 'non-classical', measurement error model. The estimation strategy implies that ability bias can be investigated separately in different parts of the educational distribution. The linear relationship between the logarithm of annual earnings and years of schooling is rejected. Furthermore, the results in the sample of identical (MZ) twins indicated both that the ability bias could be of different signs and of different magnitudes in different parts of the educational distribution. The twin-based estimates in the sample of fraternal (DZ) twins did not display any marked differences as compared to the cross-sectional estimates. Finally, the results indicated that the error-corrected twin-based estimates of the average return to years of schooling that rely on a classical measurement error model are upwards biased by approximately 30%.