In a large-scale longitudinal study, 11-year-old schoolchildren were given a control-of-variables task, and their scores in this were related to their progress in learning about science at school over the next 3 years. The aim of the control-of-variables task was to measure children's understanding that in a properly controlled scientific comparison one variable is tested at a time, while other variables are held constant. There were 2 kinds of question in the task. In one, the pupils were asked to judge whether particular comparisons between 2 situations were valid ones, in which all variables apart from the one being tested were held constant. In the other, they were asked to set up a valid comparison themselves. The pupils' scores for both kinds of item successfully predicted their progress in science at school later on, even after controls for the effects of differences in age and IQ. Their success in setting up valid comparisons was a better predictor in the long term than the choices that they made in judging whether given comparisons were valid or not.