We use data on Latino children in the United States who have been randomly assigned calculation tests in English or Spanish to check for the so-called bilingual advantage, the notion that knowing more than one language improves individuals' other cognitive skills. After controlling for different characteristics of children and their parents, as well as children's time in the US, we find a bilingual advantage among children who read or write in English and Spanish but not for those who only speak or understand both languages. In particular, bilingual readers or writers perform one fourth to one third of a standard deviation better than monolingual children, equal to learning gains of an additional school year. Applying the Oster test, we find that selection on unobservables would need to be 3-4 times stronger than selection on observables to explain away our results. The bilingual advantage is stronger among children in two-parent households with siblings and for those at the upper end of the ability distribution.