Linear models and splines

The investigators analyzed data collected in the Avon Longitudinal Study of Parents and Children, which is a prospective birth cohort study based at the University of Bristol in the United Kingdom. It is designed to investigate the long-term health and development of children. Initially, 14,000 pregnant women with an estimated delivery date between April 1991 and December 1992 were recruited. Their offspring continue to be followed today. At the time of recruitment, all were living in the Avon region of England—Bristol and adjacent areas.

Data were included for women who had a singleton live birth, no history of hypertension outside of pregnancy, and at least 1 available weight and blood pressure measurement. This left a total of 12,522 women in the analysis. The authors used multinomial regression to assess the relationship between GWG up to 18 weeks’ gestation and hypertensive categories (normotensive, gestational hypertension, and preeclampsia) while controlling for confounders; the normotensive women served as the comparison group. In addition, multinomial model estimates were utilized to estimate the predictive probabilities of hypertension and preeclampsia for a range of prepregnancy weights and levels of early GWG. Multinomial logistic regression is appropriate with unordered categorical data, has minimal assumptions, produces a single set of coefficients, and in general, yields estimates of standard error that are smaller than those generated by running multiple logistic regression models.

In order to determine whether weight gain precedes the development of hypertension in pregnancy, the authors turned to linear splines, which are values obtained through a form of interpolation. Using multivariate modeling, linear splines could assess the rates of change over time in weight gain and blood pressure. This tactic allowed them to establish a trajectory in weight gain for each of the predetermined categories: normotensive, gestational hypertension and preeclampsia. It was a useful method for setting up a generalized trend for comparison over a large dataset. The weakness is that the data are generalized and exclude individual factors that may contribute but are not obvious as the linear trend is established.