Causal model (DAG) - PCOS and the vaginal microbiota in Bukavu, DR Congo

This deposit contains the directed acyclic graph (DAG) used to derive the adjustment strategy of the study "Vaginal microbiota and polycystic ovary syndrome in Bukavu: a case-control study". It is released before data collection so that the adjustment set is fixed in advance and cannot be selected post hoc. Contents: - dag_model.txt : the model in DAGitty syntax. Paste into -adjustment sets.R : R script reproducing the adjustment sets with the `dagitty` package -DAG PCOS and vaginal microbiota :Rendered graph Causal question: Exposure: polycystic ovary syndrome (PCOS). Outcome: composition of the vaginal microbiota. Result Minimal sufficient adjustment sets obtained by the back-door criterion: 1. BMI,gut microbiome, diet, gravidity, socio-economic status2. BMI,gut microbiome, diet, gravidity, period products, vaginal douching3. BMI,gut microbiome, diet, gravidity, period products, ethnicity Two consequences follow: The gut microbiome must be measured: It opens a back-door path to the outcome that no other variable blocks. If it is treated as unmeasured, no sufficient adjustment set exists and the effect of PCOS is not identifiable. This is why the study sequences both compartments. BMI is adjusted for but never used for matching: BMI and PCOS share an unmeasured common cause (drawn as a bidirected edge), so BMI bahaves as a collider; conditioning on it opens spurious paths through its other parents: diet, gravidity, socio-economic status and the gut microbiome, all of which must therefore be included alongside BMI. The menstrual cycle is a descendant of PCOS (oligo and/or anovulation is a Rotterdam criterion) and is therefore excluded from the adjustment set. How to reproduce -R : install.packages("dagitty") then use source("adjustment sets.R") -Or paste dag_model.txt into and read the adjustment sets from the interface. Assumptions worth challenging - The BMI-PCOS relationship is modelled as an unmeasured common cause rather than a directed effect. Modelling it as an PCOS-\>BMI changes the adjustment set to diet,gut microbiome and estimates the direct rather than the total effect. Both must be reported in the sensitivity analyses.- The graph assumes all relevant common causes are represented. Unrepresented environmental exposures acting on both microbiota would remain a source of residual confounding; an E-value must be reported on the primary estimate.