We estimate the causal effect of exposure to math during high school on university major choice and performance, using a unique administrative dataset of 1,396 twins extracted from the entire student population enrolled between 2011 and 2021 at an Italian university. We apply a Twin Fixed Effect (TFE) estimator to account for unobserved factors like shared family background. We find that attending a low-math high school reduces the likelihood of enrolling in STEM majors by 32.6 percentage points and improves university performance, by increasing the likelihood of on-time graduation by 11.7 percentage points and boosting grades by 0.139 standard deviations. Leveraging a high school reform that expanded the math content in traditionally low-math curricula, we show that the added math background further reduces STEM enrollment for treated students, while it drives their improvement in performance. Our results suggest that, while increased math exposure does not necessarily boost STEM enrollment, it equips students with skills that help them improve their university outcomes. Compared with TFE, Ordinary Least Squares estimates of the effect of math exhibit a downward bias. The same applies to Difference-in-Differences estimates of the effect of the reform obtained using the entire student population.