Background It is often of interest whether changes across the life course influence later disease risk. For example, it has been suggested that small birth size and rapid growth in infancy may be associated with higher blood pressure later in life. However, existing methods to address such questions can give biassed estimates or involve complex models. We aimed to compare a novel method, based on path analysis techniques, with two existing approaches, to assess whether it may overcome these issues.
Methods For 7330 children in the Avon Longitudinal Study of Parents and Children, we fitted multilevel linear spline models to length measurements from birth to 12 months of age (mean 4 measures per child). We compared three methods of estimating the associations of the growth parameters: birth-length, growth 0–3 months and growth 3–12 months with blood pressure at age 7 years, conditional on earlier size/growth. These included a two-stage approach, a bivariate multilevel model and the path analysis-based method, in which we constructed a path diagram and from this derived total effects of the growth parameters on blood pressure. To maximise the comparability of estimates we did not adjust for any covariates. We then simulated 1000 datasets for 1000 individuals, representing this data, and formally estimated the bias and coverage.
Results All methods gave similar estimates of the association of growth from 3–12 months with blood pressure (mean difference (95% confidence interval): 4.60 (3.19, 6.00), 4.57 (3.16, 5.98) and 4.60 (3.19, 6.00) mmHg per cm/month growth for the two-stage, bivariate and path analysis approaches respectively) and simulation showed that these estimates were unbiased with good coverage. However, the associations of growth 0–3 months with blood pressure differed between the two-stage (2.97 (1.91, 4.03)) and bivariate and path analysis approaches (0.95 (–0.30, 2.20) and 0.91 (–0.36, 2.18) respectively). In simulations, the two-stage approach gave biassed estimates of associations of birth-length and growth 0–3 months with blood pressure, with low coverage, but the bivariate and path analysis approaches were unbiased with good coverage.
Conclusion The path analysis approach performs as well as the bivariate model, but only requires that a univariate multilevel model be fitted. This reduces the likelihood of convergence problems and enables model estimates to be used in many subsequent analyses considering different outcomes. However, a limitation is that we have only considered a continuous outcome; the path analysis approach is not expected to be unbiased for binary health outcomes.
- Lifecourse Epidemiology
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.