Confounders or intermediate variables? Testing mechanisms for the relationship between depression and smoking in a longitudinal cohort study.