(583) Methods of confounder selection in recent obstetrics and gynecology studies: A review

Objective: Adjusting for confounding is critical in observational studies that estimate the effect of an exposure on an outcome. Some methods for confounder selection, such as significance testing across exposure groups or including/excluding confounders based on whether they change the estimate of interest, can lead to bias by excluding important confounders and erroneously including factors affected by the exposure. Bias can be mitigated by careful selection of confounders a priori using prior knowledge supplemented with a causal diagram such as a Directed Acyclic Graph. We compared the prevalence of different approaches to confounder selection in recent obstetric and gynaecology studies.

Study Design: We reviewed all full-length original research articles from January 2022 through June 2022 in three leading obstetrics and gynecology journals (AJOG, BJOG, Obstetrics & Gynecology). We excluded experimental studies and studies that did not investigate the relationship between an exposure and a health outcome. We calculated the proportion of studies that selected all or some confounders using significance testing or a change-in-estimate approach, the proportion that reported selecting all confounders a priori, and the proportion that reported using a causal diagram.

Results: Of 252 studies, 158 met inclusion criteria. Fifty-three percent (84/158) of studies did not specify how they selected confounders. Twenty percent (32/158) used an outdated technique to select confounders—18% (29/158) used significance testing and 3% (4/158) used a change-in-estimate approach. Twenty-six percent (41/158) of studies reported selecting all confounders a priori, but only 4% (7/158) reported using a causal diagram and 2% (3/158) included their diagram.

Conclusion: Outdated methods of confounder selection are common in recent obstetrics and gynecology studies, and many studies do not specify how confounders were selected. Investigators should emphasize prior knowledge when selecting confounders, justify the selection of confounders, and transparently report their modelling assumptions.