We expand on prior qualitative research about mathematics identity development using data from a large national US survey of 10,437 students in 336 college calculus classes. Multinomial logistic regression models find that a stronger mathematics identity predicts higher student interest in pursuing certain STEM careers when compared with non-STEM careers, particularly those in physical and computer sciences, engineering, and mathematics, or in science or mathematics teaching. Multiple linear regression models identify that certain instructional practices employed by high school mathematics teachers predict higher levels of students' mathematics identity. These include a high amount of interaction within the classroom, a focus on mathematics connections, and activities involving conceptual learning. Surprisingly, the role of the textbook, ways of organizing students (individual, small group, whole class), forms of assessment, and use of calculators or computers did not significantly predict students' mathematics identity.