Confounding involves error (bias) in the interpretation of an accurate measurement by attributing it to the wrong cause
- Mill stated that exclusion of all other possible causes (i.e. confounders) is a requirement for establishing causation
Confounding occurs when there is a relation between a certain characteristic or covariate (C) and group allocation (G) and also between this characteristic and the outcome (O). When the occurs the covariate (C) is termed a confounder.
- A confounder is prognostic factor – a factor that predicts the outcome of interest
- Confounders are usually unknown
- must be unevenly distributed between the comparison groups in a study for confounding to occur
- A factor is NOT a confounder if it lies on the causal pathway between the variables of interest
COMPARISON WITH MEDIATORS, MODERATORS, AND COVARIATES
Confounders predict the outcome (dependent variable, response, or effect) by interacting with the exposure (cause or independent variable)
- In kinship terminology, a confounders is a variable that is both an ancestor of the exposure and an ancestor of the outcome (along a path that does not include the exposure)
- Mediators are part of the causal pathway from exposure to outcome
- Moderators are interaction terms that change the size or direction (or both) of the effect of the exposure on outcome
- Covariates are other independent variables that may or may not predict outcomes. A covariate may or may not be confounder.
DEALING WITH CONFOUNDERS
Known confounders can be corrected for by ‘adjustment analysis’:
- adjustment analysis attempts to control for baseline imbalances in important patient characteristics
- e.g. stratifying results according to known confounders
- e.g. correcting for known confounders using statistical methods
- typically afflict observational studies, for which propensity score analysis may be used to try to correct for their presence (it’s utility is debatable and remains inferior to randomisation)
- randomisation, to ensure that confounders are randomly distributed between treatment groups, is the best way to correct for both known and unknown confounders
REFERENCES AND LINKS
- Pourhoseingholi MA, Baghestani AR, Vahedi M. How to control confounding effects by statistical analysis. Gastroenterol Hepatol Bed Bench. 2012;5(2):79-83. [article]
FOAM and web resources
Chris is an Intensivist and ECMO specialist at the Alfred ICU in Melbourne. He is also a Clinical Adjunct Associate Professor at Monash University. He is a co-founder of the Australia and New Zealand Clinician Educator Network (ANZCEN) and is the Lead for the ANZCEN Clinician Educator Incubator programme. He is on the Board of Directors for the Intensive Care Foundation and is a First Part Examiner for the College of Intensive Care Medicine. He is an internationally recognised Clinician Educator with a passion for helping clinicians learn and for improving the clinical performance of individuals and collectives.
After finishing his medical degree at the University of Auckland, he continued post-graduate training in New Zealand as well as Australia’s Northern Territory, Perth and Melbourne. He has completed fellowship training in both intensive care medicine and emergency medicine, as well as post-graduate training in biochemistry, clinical toxicology, clinical epidemiology, and health professional education.
He is actively involved in in using translational simulation to improve patient care and the design of processes and systems at Alfred Health. He coordinates the Alfred ICU’s education and simulation programmes and runs the unit’s education website, INTENSIVE. He created the ‘Critically Ill Airway’ course and teaches on numerous courses around the world. He is one of the founders of the FOAM movement (Free Open-Access Medical education) and is co-creator of litfl.com, the RAGE podcast, the Resuscitology course, and the SMACC conference.
His one great achievement is being the father of three amazing children.
On Twitter, he is @precordialthump.