This paper investigates the causal effect of education on earnings in Brazil by employing a new method proposed by Klein and Vella (2010) that obtains identification on the presence of conditional heteroskedasticity. In contrast to traditionally used IV methods, this approach yields unbiased estimates in the absence of instruments. Results indicate that the average return to education in Brazil was relatively stable at around 14% from 1995 to 2003, declined afterwards reaching 10.7% in 2014, but has bounced back to 11.8% after the economic crisis in 2015. The results suggest that the OLS estimations are downward biased and we interpret this as a sign of under-education premiums that are likely to occur in environments where the more talented ones are dropped from school and moved into the labor market earlier in life.