Confounder selection and use of directed acyclic graphs and sensitivity analyses in COVID-19 vaccine effectiveness studies: a methodological review — The Association Specialists
  1. Schedule
  2. Oral Presentations - Advancing Epidemiological Methods 1
  3. Confounder selection and use of directed acyclic graphs and sensitivity analyses in COVID-19 vaccine effectiveness studies: a methodological review

Confounder selection and use of directed acyclic graphs and sensitivity analyses in COVID-19 vaccine effectiveness studies: a methodological review (23612)

Brett P Dyer 1 , Mari Takashima 1 2 3 , Robert S Ware 1 4
  1. Griffith University, Southport, QLD, Australia
  2. School of Nursing, Midwifery and Social Work, The University of Queensland, Brisbane, QLD, Australia
  3. Centre for Children's Health Research, Children's Health Queensland Hospital and Health Service, Brisbane, QLD, Australia
  4. Wesley Research Institute, Brisbane, QLD, Australia

Background: In observational research, appropriate control for confounding is essential for estimating vaccine effectiveness, since inadequate or inappropriate control can substantially bias effect estimates. Directed Acyclic Graphs (DAGs) are therefore encouraged to transparently communicate assumed causal relationships among the exposure, outcome, and covariates, and to identify confounders while avoiding overadjustment bias. This methodological review aimed to summarise (i) the use and reporting of DAGs, (ii) the consistency of adjustment sets, and (iii) the use of sensitivity analyses for unmeasured confounding and overadjustment biases, in COVID-19 vaccine effectiveness studies.

Methods: Observational studies from an existing systematic review of COVID-19 vaccine effectiveness against infection, hospitalisation, and mortality in adults were included. Information relating to confounder selection and adjustment, DAG use and reporting, and the use of sensitivity analyses for unmeasured confounding and overadjustment bias was extracted. Adjustment sets were compared to an a priori DAG-informed adjustment set identified by the review team.

Results: All 52 included studies accounted for potential confounders, but none reported using a DAG. Thirty-seven (71%) did not specify how potential confounders were identified. Both the number and specific confounders identified varied substantially. The median number of potential confounders accounted for was 6 (25th-75th percentile: 4-9; range 3-14). All studies adjusted for fewer variables than the 15-variable DAG-informed adjustment set derived by the review team. Six (11.5%) studies conducted sensitivity analyses for unmeasured confounding, and three (5.8%) studies repeated analyses with alternative adjustment sets to assess the impact of adjustment-set misspecification.

Conclusions: These findings highlight the need for improved transparency in reporting confounder selection methods to support valid causal inference in observational studies and inform public health decision-making. Sensitivity analyses are currently underutilised and should be routinely used to assess the impact of unmeasured confounding and misspecification of adjustment sets.