The development of human abilities stems from a complex interplay between genetic predispositions and environmental factors. Numerous studies have compared musicians with non-musicians on measures of musical and non-musical ability, frequently attributing musicians' superior performance to their training. By ignoring preexisting differences, however, this view assumes that taking music lessons is akin to random assignment. In the present longitudinal study, the musical ability of 5- to 10-year-olds was measured at Time 1 with a test of music perception and cognition. Five years later, at Time 2, the children took the same test and a second test designed for older listeners. The test-retest correlation for aggregate scores was remarkably high, r≈0.7, and remained strong when confounding variables (age, cognitive abilities, personality) were held constant. At both time points, music training was associated with musical ability, but the association at Time 2 became nonsignificant when musical ability at Time 1 was held constant. Time 1 musical ability also predicted duration of subsequent music training. These data are consistent with results from genetic studies, which implicate genes in all aspects of musical behavior and achievement, and with meta-analyses, which indicate that transfer effects from music training are weak. In short, early musical abilities significantly predicted later abilities, demonstrating that individual differences are stable over time. We found no evidence, however, to suggest that music training predicted musical ability after accounting for prior ability. The results underscore the importance of considering preexisting abilities in any type of learning.