AIMS: To examine the longitudinal associations in both directions between mental health and substance use in adolescence. DESIGN: Three-year longitudinal cohort. SETTING: Britain (nationally representative sample). PARTICIPANTS: 3607 youths aged 11-16 years at baseline. MEASUREMENTS: Externalizing and internalizing mental health problems were measured using brief questionnaires (parent-reported Strengths and Difficulties Questionnaire) and diagnostic interviews, including clinician-rated diagnoses of mental disorder. Substance use was measured by youth self-report, and included regular smoking, frequent alcohol consumption, regular cannabis use and ever taking other illicit drugs. FINDINGS: Externalizing (specifically behavioural) problems at baseline independently predicted all forms of substance use, with a particularly strong effect on smoking. In all cases this association showed a dose-response relationship. In contrast, although internalizing problems had a strong univariable association with smoking, this disappeared after adjusting for comorbid externalizing problems. There was little or no evidence that baseline substance use predicted mental health at follow-up. CONCLUSIONS: Externalizing problems predict adolescent substance use, and adjusting for comorbid externalizing problems is vital when investigating the effects of internalizing problems. A dose-response effect of externalizing problems is seen across the full range. Programmes seeking to prevent adolescent substance use by reducing externalizing problems may therefore wish to consider population-wide interventions rather than targeting individuals only at the negative extreme.