Linear mixed models’ construction
Since multiple marker measurements were taken within each subject at
different time points, the usual regression model is inappropriate as it
assumes independence among observations (14). Therefore, we analyzed the
data using linear mixed models that offer personal risk estimation,
introducing a random component to the prediction.
We did not include patient data (maternal age, weight, comorbidities,
etc.) in the model. The Wald chi-squared test was used to decide the
inclusion of covariates in the model applying a backwards elimination
process with p< 0.05. The gestational age at measurement was
retained to control potential confounding effects.
The open source, freely available R statistical software (R version
3.1.1) (15) was used to conduct all the statistical analyses. Linear
mixed model construction and paired area under the curve (AUC)
comparison were performed using the lmer (16) and the pROC(17) R add-on packages, respectively.
Comparisons between groups were performed using the Mann-Whitney U test
for quantitative variables and Pearson’s chi-squared test and McNemar’s
test for unpaired and paired proportions, respectively.