Skip to main content
Fig. 1 | BMC Medicine

Fig. 1

From: Association of the functional ovarian reserve with serum metabolomic profiling by nuclear magnetic resonance spectroscopy: a cross-sectional study of ~ 400 women

Fig. 1

Directed acyclic graphs (DAGs) of the association of ovarian reserve with cardiometabolic health. a Our main analyses, in which we a priori considered factors that were known to, or highly plausibly influenced ovarian reserve and cardiometabolic health, as confounders to be controlled for. bd Further consider whether we should or should not control for PCOS. In all of these, we assume that there are underlying unmeasured factors that generate an association between ovarian reserve and PCOS (U1). For example, it has been suggested that common genetic and/or intrauterine factors affect both of these. b We assume that PCOS is not causally related to cardiometabolic health (no arrow from PCOS to cardiometabolic health). c PCOS does causally influence cardiometabolic health and that it is on a confounding path between ovarian reserve and cardiometabolic health via U1. d Assume that PCOS causally influences cardiometabolic health but that it is a mediator between ovarian reserve and cardiometabolic health. Deciding whether we need to adjust for PCOS therefore depends on evidence for a causal link between PCOS and cardiometabolic health and if there is a link evidence as to whether PCOS is likely to be a confounder or a mediator. In relation to point 1, whilst there have been several observational studies showing an association of PCOS with cardiometabolic health, recent Mendelian randomisation studies suggest that PCOS does not causally influence type 2 diabetes, coronary heart disease or stroke (making b a plausible scenario) [31]. If the scenario depicted in b is correct, then controlling for PCOS is not necessary, but if done, it should have no impact on the results. It is not uncommon in observational epidemiology to have a risk factor for an outcome and be unsure which is more plausible—that it is a confounder or a mediator. If PCOS is on the confounding path between unmeasured factors (U1) as shown in c, we would definitely want to adjust for it as this would be the only way to block this confounding (given U1 variables are unmeasured). However, if PCOS is a mediator (d), then we would not want to adjust for it. Primarily, this is because we want to know the ‘total’ potential effect of ovarian reserve on cardiometabolic health and not remove any of that going via mediation. It is also possible that adjusting for a mediator can introduce what is known as collider bias [32]. If there are unmeasured confounders of the mediator (PCOS) and cardiometabolic health (shown by U2), then adjusting for PCOS could generate spurious associations between ovarian reserve and cardiometabolic health

Back to article page