Analysis
All analyses were conducted using STATA 17 (College Station, TX). We
assessed whether the proportions of participants who reported
miscarriages or stillbirths were similar across study arms using
Chi-squared or Fisher’s Exact tests. Further, we compared baseline
participant characteristics between those missing and not missing birth
outcome data using Chi-squared or Fisher’s Exact tests for dichotomous
variables and Student’s t-test or the Wilcoxon-Mann-Whitney tests for
continuous variables.
We provided descriptive statistics of the study sample by intervention
and standard-of-care arm assignment. We assessed the balance between the
arms using Chi-squared or Student’s t-tests. To address potential
clustering and unmeasured confounding, we assessed results from three
statistical models. First, we utilized a multivariable logistic
regression model. Next, we estimated a fixed effects model in an attempt
to purge some effects of unobserved clinic-level characteristics that
may influence birth outcomes.(18) We also estimated a random effects
model, with the clinic specified as a random effect. As the intraclass
correlation coefficient was 0.0017, suggesting that only 0.17% of the
variance in the outcome was due to variations across clinics, we decided
not to estimate a generalized estimating equation. For each model,
covariates of interest (described above) were entered into a full model
and then excluded one at a time, starting with the variable with the
largest p-value until all remaining independent variables had p-values
<0.20. Then, covariates were readded one at a time in the
order they were dropped and retained if they had a p-value of
<0.20 or the odds ratio (OR) of the primary exposure was
changed by 15%. We also assessed the potential interaction between
clinic and intervention assignment. Models were run for the primary
outcome (preterm or low birth weight), as well as preterm birth only
(includes low birthweight if preterm) and low birth weight only
(includes preterm if low birth weight) individually. When the clinic
site variable was included, we assessed the models with and without the
clinic that did not crossover.
Further, we conducted post-estimation predictive margins analysis to
estimate the predicted prevalence of the composite outcome of low birth
weight or preterm birth in the intervention and standard of care arms as
well as the adjusted risk ratios. In other words, following the logistic
regression, all participants were set to both exposure values
(standard-of-care and intervention) and the logistic regression
coefficients were used to calculate the predicted prevalence for
everyone at their observed confounder pattern and assigned exposure.
Confidence intervals were calculated using the delta method.(19)
Thereafter, the effect of intervention versus the standard-of-care was
calculated as an adjusted risk ratio.
In another attempt to balance confounding, we conducted post hoc
analysis that stratified the sample by nulliparous and multiparous
women.