Non‐linear trajectories were explored by subjecting age to a restricted range of fractional polynomial transformations (x−2, x−1, x–0.5, ln(x), x0.5, x1, x2 and x3), which permit the modelling of monotonic and non‐monotonic relationships between alcohol consumption and age 17. The fit of each transformation was assessed according to the Bayesian information criterion (BIC) 18. Owing to a lack of convergence when some non‐linear transformations were applied, BICs were calculated for simplified models that constrained to zero any covariance between repeated measures. An improvement in fit relative to a linear model was defined as any reduction in the BIC greater than or equal to a value of 10, which is described as a strong indicator of an improvement to model specification 19. The best‐fitting trajectory for each baseline category was then plotted allowing random effects and an unstructured covariance matrix, as per the primary linear models.
