By combining two unique Taiwanese datasets, this paper investigates how birth weight affects grades at age 15 years. To tackle the endogeneity problem caused by omitted variables, we first compare birth weight and grade variation within twins. We find that birth weight does increase grades but only when both twins weigh less than 3000g at birth, which indicates that the effect is non-linear, and when the weight difference between the twins is larger than 200g. Furthermore, twin fixed effect estimates are similar to the ordinary least squares (OLSs) ones. We then use the public health budget and the number of doctors in the county where the children were born as instrumental variables for the children's birth weight. We found that instrumental variable estimates are significant only for the less educated (<9 years) and young (<25 years) mothers. We conclude that the effect of birth weight is real and non-linear and its effect on less educated and young mothers is the most severe. Furthermore, the bias produced by OLS may not be large if the correct subgroup of the population has been identified.