Statistical Analysis

All statistical analyses were performed using SAS version 9.3 (SAS Institute, Cary, North Carolina). Differences between term delivery and PTD were determined using the χ² test for categorical variables and the t test for continuous variables. Cross-sectional analyses were performed to compare mean sFlt1 and uterine artery RI values at each of the 5 time intervals for the term and preterm groups, and sFlt1 values were log transformed for normality; then, an independent samples t test was performed at each time interval. For uterine artery RI, no transformation was required before conducting t tests. The risk ratios for preterm delivery were explicitly calculated for sFlt1 and uterine artery RI values at each of the 5 time intervals using a log-binomial model.

Inspection of cross-sectional sFlt1 values indicated that in both the term and the preterm groups, the values remained relatively stable through the fourth visit (22-26 weeks of gestation) and then increased sharply by the fifth visit (32-36 weeks of gestation). The change in log(sFlt1) over gestation was therefore modeled using a simple piecewise model in which values were allowed to change linearly over time until the 21 3/7 weeks of gestation, at which point a quadratic term was allowed. Model fit was analyzed using both likelihood-based fit statistics (eg, the Akaike information criterion) and an examination of the residual values from the model. This piecewise model had better fit than simple linear and quadratic models, cubic models, and models fit using restricted cubic splines. Multiple points were examined as possible origins of the quadratic increase in mid to late gestation; 21 3/7 weeks was chosen both for its superior model fit and for its simplicity. Qualitative results did not change substantially among these similar models for sFlt1. A repeated measures model with an unstructured covariance matrix was used to account for correlation between sFlt1 values across visits within a participant.

Inspection of the cross-sectional uterine artery RI values showed that RI decreases over gestation in both the term and the preterm groups, with a faster decrease in early gestation than later. A quadratic model fit this pattern well for both the term and the preterm groups. A repeated measures model with an unstructured covariance matrix was used to account for correlation between uterine artery RI values across visits within a participant; allowing the covariance matrix to differ between term and preterm participants did not improve model fit for uterine artery RI, so nesting within preterm groups was not used for covariance estimation.

The effect of sFlt1 on uterine artery RI was investigated assuming there was a lag in the effect of sFlt1 on uterine artery RI so that the effect of sFlt1 from the prior study visit on the uterine artery RI from the current visit was estimated. The effect of sFlt1 on uterine artery RI was investigated in cross-sectional analyses by fitting a linear model for the uterine artery RI value at that visit on the log(sFlt1) value from the previous visit.

Finally, we use a multivariable repeated measures model to examine whether the effects of sFlt1 from the previous visit on uterine artery RI from the subsequent visit (a “lagged” effect) differ between term and preterm participants. The regression equation used for this model is given by:

where GA is the gestational age at time t, UTRI is uterine artery RI at time t, and lagSFLT is the maternal serum sFlt1 concentration at time t − 1. This repeated measures model was also fit with an unstructured covariance matrix; the covariance matrix was estimated separately for the term and preterm groups.