There is increasing emphasis in medical research on fetal and childhood antecedents of chronic disease risk, and how these interact with other exposures throughout the lifecourse to influence later-life conditions (1). Answering questions about the relative importance of speed, magnitude and timing of growth, behaviour and health status for longer-term outcomes requires appropriate analyses of longitudinal data. For example, to understand how maternal weight gain during pregnancy (gestational weight gain, GWG) influences later cardiovascular disease risk for the child, we might seek to describe relationships between GWG and BMI at age 7, and how any relationships are mediated through birthweight and changes in weight during childhood.

Analysis of lifecourse data poses several statistical problems (2). Analysis of a repeated outcome must account for dependencies between multiple observations on the same person: methods to do this (e.g. multilevel models, (3) (4)) are now widely available in standard statistical software packages (e.g. Stata (5)). Within-individual variation may vary over time (e.g. absolute measurement error in weight will be larger in later childhood than at birth) and there will usually be dropout due to non-response, death, illness, emigration, etc. Where several repeated measures are used as exposures in one regression model for a later-life outcome, standard regression models may be affected by their multicolinearity. Disentangling the genuine associations between GWG, birth weight and later outcomes is important; a negative correlation between birth weight and later cardiovascular risk would imply a public health focus on increasing average birth weight, whereas if the relationship were largely between subsequent growth and later cardiovascular risk, this would imply a focus on minimising excess weight gain in children (6).

We consider methods for relating growth (exemplified here by GWG) to later outcomes (here, BMI at age 7). Methods to be used include: SITAR (7); multivariate multilevel growth models (8); latent class models (9); mediation models (10)including structural equation models (11) and lifecourse models (12). We propose to use the associations between GWG, birthweight, childhood growth and BMI at age 7 in ALSPAC as an example dataset on which to compare these statistical methods.