Statistical analyses
We undertook causal mediation analyses to examine the effects of SGA on
caesarean delivery with and without TOL. The analyses were performed for
all subjects in toto , but we also performed separate analyses for
women with and without one prior caesarean delivery. The results of
these stratified analyses (eTables 3-8 in the Supplement) were
similar to the overall analysis, so we have presented the overall
analyses here.
The effects of exposure and mediator on caesarean delivery were
estimated from a multiplicative scale (log-linear Poisson regression
models with a log-link function and robust variance, with risk ratio
[RR] as the effect measure) and an additive scale (log-linear
Poisson regression models with an identity-link function and robust
variance, with the risk difference [RD] as the effect measure). We
undertook a causal mediation analysis based on a counterfactual
framework (11, 20) to disentangle the total effect of the SGA-caesarean
association that could be attributable to a pure mediation effect by
gestational age, to an interaction effect (by gestational age), and no
interaction or mediation effects (21, 22). This method decomposes the
total effect of SGA on caesarean delivery (exposure-outcome association)
in the presence of gestational age (mediator) with which the SGA
(exposure) may interact. The resulting 4-way decomposition includes: (i)
effect of SGA on mode of delivery in the absence of the mediator (i.e.,
at 39-40 weeks [reference]); (ii) interaction effect when the
mediator is left to what it would be in the absence of the SGA exposure
(i.e., among AGA births); (iii) mediated interaction; and (iv) pure
effect mediated through gestational age. Stated differently, these 4
effects correspond to the portion of the SGA-caesarean delivery
association that is neither due to mediation or interaction with
gestational age (also referred to as the controlled direct effect
[CDE]), just to interaction but not mediation (referred to as the
reference-interaction [INTREF]), to both mediation
and interaction (referred to as the mediation-interaction
[INTMED]), and to just mediation but not interaction
(referred to as the pure indirect effect [PIE]). The sum of these 4
decomposed effects amount to 100% on the RD scale.
To facilitate easier interpretation of causal effects, we focus on the
proportion of the SGA-caesarean delivery association that is directly
attributed to exposure to SGA (i.e., the CDE) and two other measures in
the primary analysis: proportion of the SGA-caesarean delivery
association that is mediated through gestational age and the proportion
attributable to an interaction between SGA and gestational age (22). The
full analysis of the 4-way decomposition is shown in eTables
3-8 in the supplement. 95% confidence intervals (CI) for all of causal
estimates were derived from the bootstrap resampling method with 2000
replications.