Statistical analyses
Data from continuous variables were expressed as medians and interquartile ranges and from categorical data as n (%). Comparison of the maternal characteristics between the outcome groups was by the χ2-square test or Fisher’s exact test for categorical variables or Mann-Whitney U-test for continuous variables, respectively. The observed measurements of UtA-PI were expressed as a multiple of the normal median (MoM) after adjustment for maternal and pregnancy characteristics as previously described15. The values of EFW were expressed as Z-scores using The Fetal Medicine Foundation population charts2. We used Bayes’ theorem to combine theprior joint distribution of Z and GA according to the history model with the likelihoods of EFW Z score and UtA-PI MoM to obtain a pregnancy specific posterior distribution; this was used to produce patient specific risks according to the competing risks model for SGA. The distributions of patient-specific risks were used to estimate detection rates (DR) and FPR from analysis of receiver operating characteristic (ROC) curves. Similarly, patient-specific risks were estimated using our previously reported logistic regression model for placental dysfunction related stillbirth 1. The predictive performance for stillbirth of the competing risks model for SGA7 and the stillbirth-specific logistic regression model for placental dysfunction 1 was compared. McNemar’s test was used to compare detection rates of stillbirth achieved from the application of the RCOG guideline and those resulting from the competing risks model for SGA, at the same screen positive rate as that determined from the use of the RCOG guideline.
The statistical software package R was used for data analyses.16