2.3 Statistical analysis
Statistical description was made by use of percentage for categorical variables, and mean, along with standard deviation, for continuous variables. The group difference was examined using Chi-squared test, variance analysis, or Kruskal-Wallis rank test where appropriate, and the pairwise comparisons were further checked.
The linear mixed models were used to test the fetal growth difference by varied chorionicity and mode of conception. Both the overall longitudinal change including trajectory curve and growth velocity across the pregnancy and week-specific discrepancy were taken into examination. For the purpose of this study, the difference was compared between any two groups. For those ultrasound measurements which showed statistically different either between SC-MCDA and SC-DCDA twins or between SC-DCDA and ART-DCDA twins, week-specific comparison were further conducted after adjusting for maternal age, height, weight, gravidity, and parity5, 6. Moreover, all of the three groups were in comparison with uncomplicated singletons, which delivered in the same hospital at the same time.
The GAMLSS was applied to fit the growth curve of each biometric parameter in relation to gestational age. The GAMLSS is highly flexible as it is capable of modeling not only mean (or location) but also other parameters (such as standard deviation, skewness and kurtosis)10, 21. Box-Cox t (BCT) distribution was used for modeling fetal biometric parameters as non-parametric cubic spline functions of gestational age. Model selection was based on generalized Akaike Information Criterion (GAIC), and the model with the smallest value of GAIC was selected. Worm plots were used for visual inspection of the fit of the smoothed curves. Centile curves for each biometric parameter were constructed according to placental chorionicity and mode of conception.
All analyses were performed with the Statistical Package for the Social Sciences version 24 (SPSS Inc, Chicago, IL) and the GAMLSS and lme4 package for R statistical software (version 3.5.1). Significance was defined as a 2-tail probability value of <0.05.