Predictive validity entails the comparison of test scores with some other measure for the same candidates taken some time after the test has been given. For tests that are used for university selection purposes, it is vital to demonstrate predictive validity. The research reported here uses data collected from three cohorts of students enrolled at Florida State University. The data includes information about each student's performance at high school, ethnicity, gender and first year GPA. Multilevel modelling has been applied to the data using the statistical software package MLwiN to investigate the relationships between the variables, and in particular to determine which are the best indicators of academic success at university, whilst taking into account the effects of individual high schools. Issues relating to choice of predictive and university success measures, intervening variables, controlling for selection bias, data and measurement, and choice of research model are discussed in the context of one American university.