Background Different hypotheses may be proposed regarding the association between exposures, measured repeatedly over the life course, and a later-life outcome. The ‘structured approach’ to life course hypotheses allows competing hypotheses to be compared in terms of their support by observed data. This approach has previously been limited to binary exposures, we aim to extend it to include continuous exposures.

Methods We developed a new method using least angle regression (LARS) to identify which of a small set of hypothesised models explains most of the observed outcome variation. The choice of most-supported hypothesis may be made objectively, by selecting the first hypothesised model found by LARS, or subjectively, using an elbow plot or lasso covariance test, depending whether the investigation is confirmatory or exploratory. All analyses were carried out using the R statistical computing language.

We compared our method with other methods, based on F-tests, for structured approaches through simulation. We simulated three binary exposure measurements and a continuous outcome based on different life course models, and measured the proportion of times that each method selected the correct model. We varied the coefficient of variation in the simulations and varied the sample size between n = 400 and n = 2,500.

As an example of continuous exposures, we used data from the Avon Longitudinal Study of Parents and Children (ALSPAC). We used routine measurements of weight in pregnancy, and birthweight, from 11,499 mother-offspring pairs. The hypotheses considered, for a life course association between gestational weight gain and birthweight, included a pre-pregnancy critical period, sensitivity to change (including Institute of Medicine thresholds), and interaction between them.

Results In simulation, LARS selected the correct model in at least 90.2% of simulations in which the coefficient of variation was at least 100/n, outperforming the methods based on F-tests.

Our structured approach identified an interaction hypothesis, between pre-pregnancy weight and total gestational weight gain, as the best explanation for variation in birthweight after adjusting for gestational age and maternal height.

Conclusion A structured approach using LARS can be used to effectively assess competing life course hypotheses involving continuous exposures. Unlike methods based on F-tests, the LARS approach does not require the identification of a saturated model and need not rely on p-values for interpretation. It is harder for structured approaches to identify life course hypotheses when the exposure measurements are highly correlated.