Statistical analysis
We considered potential confounders between 3rd and
4th degree tear and VBAC status to be: maternal age
(recorded to the nearest year at prenatal booking and centred at 30
years), mode of birth (as a three level nominal variable: normal vaginal
birth, forceps vaginal birth and ventouse/vacuum vaginal birth),
completed week of term gestational age (37-42 weeks), birthweight
(continuous in grams), episiotomy (binary yes/no), BMI (continuous) and
epidural analgesia (binary yes/no). Raw data for the outcome and
covariates were presented as number (%) or mean (SD) by exposure
status. The proportion of missing data was documented for each covariate
(Table 1).
The distribution of covariates between exposure groups was examined
using univariable logistic regression for categorical data and Wilcoxon
rank-sum test for continuous data, based upon non-missing data. The same
tests were repeated as a sub-analysis, looking at the distribution of
covariates between women who had a VBAC and had a 3rdor 4th degree tear, and women who had a VBAC but no
significant perineal injury.
The adjusted analysis consisted of two parts: multiple imputation of
missing covariate data, followed by regression adjustment analysis based
upon potential outcome means (POM). We performed multiple imputation
consisting of fully conditional specification (FCS) using a predictive
mean model for continuous, logistic model for binary and multinominal
model for unordered categorical covariates. The imputation model
included outcome, exposure and all covariates used in the analysis
model. There were no additional auxiliary variables available for
inclusion. The number of imputed datasets (20) was set above the highest
percentage of missingness and considered adequate if the Monte Carlo
error was less than 10% of regression coefficient standard errors.
Standard imputation diagnostics included: graphical assessment of
convergence; comparison of distributional shape between both imputed
(complete) and observed (non-imputed) datasets; and graphical and
numerical imputed values and observed values.
The analysis was then performed on each of the 20 imputed datasets,
using Regression Adjustment (RA) to derive a pooled estimate of
potential outcome means (POM), Risk difference (RD) and Relative risk
(RR) with associated 95% confidence limits adjusted for imputation
using Rubin’s rules. Regression adjustment estimates the POM separately
for both exposure groups and these are then used to derive RD and RR
estimates. The model will produce unbiased estimates if both covariate
relationships are correctly specified, and no important confounders have
been omitted.
Sensitivity analysis was performed using the same regression adjustment
model for two complete case cohorts (including and excluding BMI
and analgesia in the covariate list). Unadjusted RD and RR are also
presented. Analysis was performed using Stata v16, including the
multiple imputation and teffects suites. The significance level was
two-sided and set at 0.05.